content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
Inspired by the achievements of convolutional networks (a.k.a, ConvNets) in the field of computer vision, more and more researchers constitute ConvNets for kinds of natural language processing tasks, e.g., text classification \cite{kim2014convolutional}, text regression \cite{bitvai2015non}, short text pair re-ranking \cite{severyn2015learning}, and semantic matching \cite{hu2014convolutional}. \\
\indent For the answer selection task, i.e., given a question and a set of candidate sentences, choose the correct sentence that contains the exact answer and sufficiently support the answer choices. Most of the previous methods constitute Siamese-like deep architectures (like LSTM-RNN, CNN, etc.) to learn the semantic representation for each sentence, and then use cosine similarity or weight matrix to compute the similarity of the pairwise representations \cite{wang2015long}. At the same time, these works mostly adopted shallow architectures for sentence modeling, since deeper nets did not bring better performance. On the contrary, we firmly convinced that one can benefit much more from deep learning strategy.\\
\indent Following the success of RNN-based attentive mechanism designed for machine translation task \cite{bahdanau2014neural}, recently some works attempted two-way attention mechanism for sentence pair matching problems \cite{tan2015lstm,santos2016attentive,yin2015abcnn}. Such soft attention mechanism proves the effectiveness of the interaction between sentence pairs from lexical level to semantic level, yet aggravates much more computations and model complexity.\\
\begin{table
\setlength{\abovecaptionskip}{0.3cm}
\setlength{\belowcaptionskip}{-0.55cm}
\begin{tabularx}{7.7cm}{lX}
Q: & When did Amtrak {\bf{\em \color{red} \underline {begin}}} operations?\\%Where do you put the {\bf {\em \underline {apple}}}?\\
A: & Amtrak has not turned a profit since it was {\bf{\em \color{red} \underline {founded}}} in 1971.\\%{\bf {\em \underline{ Apple}}} is an American company headquartered in Cupertino, California.\\
\end{tabularx}
\caption{An example of QA-pair in TREC-QA.}
\label{tab:example}
\end{table}
\begin{figure*}
\setlength{\abovecaptionskip}{0.0cm}
\setlength{\belowcaptionskip}{-0.35cm}
\centering
\includegraphics[width=16.0cm]{framework.pdf}
\label{fig:thewholeframework}
\caption{Our M$^2$S-Net for sentence pair matching.}
\end{figure*}
\indent All the previous works mentioned above motivate us to construct a network based on pairwise token matching for exhaustive matching learning. However, a vital issue of this constitution is the word similarity measurement.Take Q and A in Table \ref{tab:example} for example, distinguishing the similarity between "begin" with "found: set up" from that between "begin" with "found: discovered" makes a lot of sense. To solve this, we constitute a deep convolutional neural network based on pairwise token matching measured with multi-modal similarity metric learning, named by us M$^2$S-Net, where the learnable multi-modal similarity metric provides a comprehensive and multi-granularity measurement. Experimental results on the benchmark dataset of answer selection task indicate that the proposed model can greatly benefit from deep network structure as well as multi-modal similarity metric learning, and also demonstrate that the proposed M$^2$S-Net outperforms a variety of strong baselines and achieve state-of-the-art.
\section{M$^2$S-Net}
In this paper, we propose a novel learning framework for sentence pair matching, where the pairwise token similarity matrix is computed firstly, and then a deep convolutional network is constructed to learn matching representation exhaustively, finally concatenate the learned pairwise matching representation and additional simple word-level overlap feature to feed into a {\em pointwise} rank loss for end-to-end fine-tuning (please see Fig. \ref{fig:thewholeframework} for better understanding).
\subsection{Multi-Modal Similarity Metric}
As a fundamental component in M$^2$S-Net, it is of vital importance to design an appropriate similarity measurement $f_{match}$ for pairwise token matching. Given a sentence pair $S_1 = \{{\bm w}_1^{i}, i \in [0, l_{1}-1]\}$ and $S_2 = \{{\bm w}_2^{j}, j \in [0, l_{2}-1]\}$, where $l_{1}, l_{2}$ are the word count of sentence $S_1, S_2$, respectively, ${\bm w}_1^i, {\bm w}_2^j$ are from $d$-dimensional word embeddings $W$ which are pre-trained under vocabulary $V$, The similarity matrix $M \in \mathbb{R}^{k \times l_{1}\times l_{2}}$ is formulated as follows:
\begin{comment}
\begin{align}
\begin{small}
m_{i,j} = \left\{
\begin{array}{l@{}c@{}l}
1/({1+||{\bm w}_i^1-{\bm w}_j^2||}) & & ({\textrm {Euc.}}) \\
({{{\bm w}_i^1}^T{\bm w}_j^2})/({||{\bm w}_i^1||\cdot ||{\bm w}_j^2||}) & & ({\textrm {Cos.}}) \label {eq:1}\\
{{\bm w}_i^1}^T{U}{\bm w}_j^2+b_{i,j}& & ({\textrm {Metric.}})\\
\end{array} \right.
\end{small}
\end{align}
\end{comment}
\begin{equation}
M = [m^k_{i,j}] = [{{\bm w}_i^1}^T{U^k}{\bm w}_j^2+b^k_{i,j}]
\end{equation}
where k represents the number of modality that can be tuned, ${U^k} \in \mathbb {R}^{d\times d}$ represents the matrix of the learnable similarity metric, and the corresponding bias term ${B}^k \in \mathbb {R}^{l_{1}\times l_{2}}$. Since the dimension of metric $U^k$ increases exponentially with word embedding dimension, some regularization methods were proposed \cite{shalit2010online, cao2013similarity} to limit model complexity for preventing overfitting. Frobenius norm is adopted here for simplification. For better comparison, we also design cosine and euclidean similarity, which are formulated as follows:
\begin{align}
\begin{small}
m_{i,j} = \left\{
\begin{array}{l@{}c@{}r}
1/({1+||{\bm w}_i^1-{\bm w}_j^2||}) & & ({\textrm {Euc.}}) \\
({{{\bm w}_i^1}^T{\bm w}_j^2})/({||{\bm w}_i^1||\cdot ||{\bm w}_j^2||}) & & ({\textrm {Cos.}}) \label {eq:1}\\
\end{array} \right.
\end{small}
\end{align}
\subsection{Convolution and Pooling}
\label{sect:convandpooling}
The convolution layer in this work consists of a filter bank $ F\in \mathbb {R}^{n \times c \times h \times w}$, along with filter biases $\bm b \in \mathbb {R}^n$, where $n$, $w$ and $h$ refer to the number, width and height of filters respectively, and $c$ denotes the channels of data from the lower layer. More specifically, for the first convolution layer, $c$ equals to the multi-modal parameter $k$, which means convolving across all the similarity modalities to learn the pattern. Given the output $L^{t-1} \in \mathbb {R}^{c\times l_h \times l_w}$ ($L^0$ represents similarity matrix $M$) from the lower layer, the output of the convolution with filter bank $F$ is computed as follows:
\begin{equation}
\begin{small}
\begin{aligned}
L^{t} &= {\tanh}(F * L^{t-1} + {\bm b}) \\
&={\tanh}([{\bm f}_i^T{\bm l}^{t-1}_{c\times(j-h+1:j)\times(l-w+1:l)} + b_i])\\
\end{aligned}
\label {eq:conv}
\end{small}
\end{equation}
where * is marked as the convolutional operation, $i$ indexes the number of filters, $j$ and $l$ indicates the sliding operations for dot production along the axis of width and height with one step size. \\
\indent Typically, there exist two types of convolution: {\em wide} and {\em narrow}. Even though previous works \cite{KalchbrennerGB14} have pointed out that using {\em wide} type of convolution got better performance, we use the {\em narrow} type for convenience. Finally, we get the output of layer $t$ as $L^{t} \in \mathbb {R}^{n \times (l_h-h+1) \times (l_w-w+1)}$.\\
\indent The outputs from the convolutional layer (passed through the activation function) are then fed into the pooling layer, whose goal is to aggregate the information and reduce the representation. Technically, there exist two types of pooling strategy, i.e., {\em average} pooling and {\em max} pooling, and both pooling methods are widely used. However, {\em max} pooling can lead to strong over-fitting on the training data and, hence, poor generalization on the test data, as shown in \cite{Zeiler_stochpooling}. For stability and reproductivity, we adopt the {\em average} pooling strategy in our work.
\subsection{{\em Pointwise} Learning to Rank with Metric Regularization}
\label{ssec:pwrl}
We adopt simple {\em pointwise} method to model our answer selection task, though {\em pairwise} and {\em listwise} approaches claim to yield better performance. The cross-entropy cost deployed here to discriminantly train our framework as follows:
\begin{multline}
\setlength{\abovedisplayskip}{1pt}
\setlength{\belowdisplayskip}{1pt}
C_{\theta}={-\frac {1}{N}\sum_{i=1}^N{[y_i\log p_i+(1-y_i)\log(1-p_i)]}} \\
{\hspace{0mm}}+ \frac{\lambda}{2}{ \Arrowvert U \Arrowvert^2_F}
\label{eq:4}
\end{multline}
where $p_i$ is the output probability of $i^{th}$ sample through our networks, $y_i$ is the corresponding ground truth, and $\theta$ contains all the parameters optimized by the network, i.e., $\theta=\{W;U;B;[F];[{\bm b}]\}$. Frobenius norm is used to regularize the parameter $U$ of the metrics to prevent over-fitting.\\% $\lambda$ is set to be $5e^{-4}$.
\indent We use Stochastic Gradient Descent (SGD) to optimize our network, and AdaDelta \cite{zeiler2012adadelta} is used to automatically adapt the learning rate during the training procedure. For higher performance, hyper-parameter selection is conducted on the development set, and Batch Normalization ({\em BN}) layer \cite{ioffe2015batch} after each convolution layer is also added to speed up the network optimization. In addition, dropout is applied after the first hidden layer for regularization, and early stopping is used to prevent over-fitting with a patience of 5 epochs.
\begin{table}
\setlength{\abovecaptionskip}{0.3cm}
\setlength{\belowcaptionskip}{-0.35cm}
\centering
\small
\begin{tabular}{lrrrl}
\toprule[1.5pt]
Set & \#Question & \#QApairs & \%Correct & Judge\\
\midrule[0.8pt]
Train-All & {1,229} & {53,417} & {12.0\%} & {auto} \\
Train & {94} & {4,718} & {7.4\%} & man \\
Dev & {65} & {1,117} & {18.4\%} & man \\
Test & {68} & {1,442} & {17.2\%} & man\\
\bottomrule[1.5pt]
\end{tabular}
\caption{Statistics of the answer sentence selection dataset. Judge denotes whether correctness was determined automatically (auto) or by human annotators (man).}
\label{tab:datasetsinfo}
\end{table}
\section{Experiments}
\label{sec:exp}
\subsection{Dataset}
\label{ssec:dataset}
In this section, we use TREC-QA dataset to evaluate the proposed model, which appears to be one of the most widely used benchmarks for answer sentence selection. This dataset was created by \cite{wang2007jeopardy} based on Text REtrieval Conference (TREC) QA track (8-13) data\footnote{\url{http://cs.stanford.edu/people/mengqiu/data/qg-emnlp07-data.tgz}}. Candidate answers were automatically retrieved for each factoid question. Two sets of data are provided for training, one is small training set containing 94 questions collected through manual judgement, and the other is full training set, i.e., Train-All, which contains 1,229 questions from the entire TREC 8-12 collection with automatically labeled ground truth by matching answer keys' regular expressions\footnote {\url{http://cs.jhu.edu/~xuchen/packages/jacana-qa-naacl2013-data-results.tar.bz2}}. Table \ref{tab:datasetsinfo} summarizes the answer selection dataset in details. In the following experiments, we use the full training set due to its relatively large scale, even though there exists noisy labels caused by automatically pattern matching.\\
\indent The original development and test datasets have 82 and 100 questions, respectively. Following \cite{wang2015long,santos2016attentive,tan2015lstm}, all questions with only positive or negative answers are removed. Finally, we have 65 development questions with 1,117 question-answer pairs, and 68 test questions with 1,442 question-answer pairs.
\subsection{Token Representation}
\label{sec:tokens}
We use a pre-trained 50-dimensional word embeddings\footnote{\url{http://nlp.stanford.edu/data/glove.6B.zip}} \cite{pennington2014glove} as our initial word look-up table. These word embeddings are trained on Wikipedia data and the fifth English Gigawords with totally 6 Billion tokens. Need to be mentioned here, trading off between model complexity and performance, we do not use the 300-dimensional embeddings, which are trained on much more data and more widely adopted by previous works \cite{santos2016attentive,tan2015lstm}.
\begin{figure}
\setlength{\abovecaptionskip}{0.1cm}
\setlength{\belowcaptionskip}{-0.35cm}
\centering
\includegraphics[width=8.0cm]{comp.pdf}
\caption{Comparison of $M^2$S-Nets with different measurements and network structure.}
\label{fig:comp1}
\end{figure}
\subsection{Experimental Setting}
\label{sec:setting}
Following previous works, we also use the two metrics to evaluate the proposed model, i.e., Mean Average Precision (MAP) and Mean Reciprocal Rank (MRR). The official {\em trec\_eval} scorer tool\footnote{\url{http://trec.nist.gov/trec_eval/}} is used to compute the above metrics.\\
\indent The simplest word overlap features between each question-answer pair are computed, and we concatenate them with our learned matching representation for the final rank learning. This feature vector contains only two features, i.e., word overlap and IDF-weighted word overlap.\\
\indent Experiments of our M$^2$S-Net on three pre-defined similarity measurements are denoted as M$^2$S-Net-Euc, M$^2$S-Net-Cos, and M$^2$S-Net-Metric respectively. All of these models share the same network configuration. To demonstrate the fact that the proposed network can benefit more from deep structure, we compare M$^2$S-Net-Metric with a one-convolutional layer network, namely M$^2$S-Net-Shallow (We found that much deeper construction might bring in randomness which harms the reproductivity of the performance, so we use two-convolutional layer for strict experimental comparison). Furthermore, we list results of M$^2$S-Net-Metric with $k = 1, 2, 4$, respectively denoted as M$^2$S-Net-Metric, M$^2$S-Net-Metric-2 and M$^2$S-Net-Metric-4, to verify the effectiveness of the proposed multi-modal similarity metric. All the networks mentioned here are implemented using Caffe \cite{jia2014caffe} and the code is open now\footnote{\url{https://github.com/lxmeng/mms_answer_selection}}.
\begin{comment}
\begin{table}
\setlength{\abovecaptionskip}{0.3cm}
\setlength{\belowcaptionskip}{-0.35cm}
\centering
\begin{tabular}{lcc}
\toprule[1.5pt]
Model & MAP & MRR\\
\midrule[0.6pt]
M$^2$S-Net-Shallow & .7111 & .8210\\
\midrule[1pt]
M$^2$S-Net-Euc & .7220 & .8172\\
M$^2$S-Net-Cos & .7226 & .8144\\
M$^2$S-Net-Metric & .7256 & .8082\\
\midrule[0.6pt]
M$^2$S-Net-Metric-2 & .7698 & {\bf .8640}\\
M$^2$S-Net-Metric-4 & {\bf .7793} & .8487\\
\bottomrule[1.5pt]
\end{tabular}
\caption{Results of M$^2$S-Nets on the answer sentence selection dataset.}
\label{tab:rst1}
\end{table}
\end{comment}
\section{Results and Discussion}
\label{sec:rst}
We are motivated to use multi-modal similarity metric to solve polysemy of words, and construct thorough matching network between sentence pairs for end-to-end question answering modeling. From Fig. \ref{fig:comp1}, we can see that one-modality metric is slightly better than euclidean and cosine similarity measurement. Increasing the number of modality of measurement greatly boost the performance by 7\%. The comparison between shallow and deep network structure indicate that the proposed M$^2$S-Net benefits much from deep construction. The rank of answer in Table \ref{tab:example} is promoted from top 35 and 26 by using euclidean and cosine similarity measuremen to top 3 by using ours.\\
\indent For comprehensive comparison, we also list the results of prior state-of-the-art methods in literature on this task in Table \ref{tab:rst2}. It can be seen that the proposed method outperforms the most recently published attention-based methods by 1\% in both MAP and MRR metrics.\\
\indent The proposed method could be further improved by upgrading the regularization term to limit the rank of metric, which had been proved by \cite{law2014fantope, cao2013similarity}. Besides, combining the dissimilarity modeled by distance metric learning with similarity mentioned here would be our future work.
\begin{table}
\setlength{\abovecaptionskip}{0.3cm}
\setlength{\belowcaptionskip}{-0.35cm}
\centering
\begin{tabular}{lcc}
\toprule[1.5pt]
Reference & MAP & MRR\\
\midrule[1pt]
\newcite{wang2007jeopardy} & .6029 & .6852 \\
\newcite{heilman2010tree} & .6091 & .6917\\
\newcite{wang2010probabilistic} & .5951 & .6951\\
\newcite{yao2013answer} & .6307 & .7477\\
\newcite{yih2013question} & .7092 & .7700\\
\newcite{yu2014deep} & .7113 & .7846\\
\newcite{wang2015long} & .7134 & .7913\\
\newcite{tan2015lstm} & .7106 & .7998\\
\newcite{severyn2015learning} & .7459 &.8078\\
\newcite{santos2016attentive} & .7530 & .8511\\
\newcite{wang2016sentence} & .7714 & .8447\\
\midrule[0.6pt]
M$^2$S-Net-Metric-2 & .7698 & {\bf .8640}\\
M$^2$S-Net-Metric-4 & {\bf .7793} & .8487\\
\bottomrule[1.5pt]
\end{tabular}
\caption{Results of our models and other methods from the literature.}
\label{tab:rst2}
\end{table}
\section {Conclusion}
\label{sect:ccls}
A novel end-to-end learning framework (M$^2$S-Net) is proposed for answer sentence selection task. Interdependence between sentence pair at lexical level is explored much more by constituting deep convolutional neural network directly on pairwise token matching. To enrich the lexical modality measurement, we adopt multi-modal similarity metric learning. The proposed architecture is proved effective, and surpasses previous state-of-the-art systems on the answer selection benchmark, i.e., TREC-QA dataset, in both MAP and MRR metrics.
|
\section{Introduction}
\label{sec:intro}
The top quark is the heaviest known elementary particle. Its large mass suggests that it may play a
special role in theories of physics beyond the Standard Model (BSM)~\cite{Hill:1994hp,Lillie:2007yh,AXI,ZP,KK}. Such a role could be elucidated via precision tests of the Standard Model (SM) in large data samples of top--antitop quark pair (\ttbar) events collected at the Large Hadron Collider (LHC) in
proton--proton ($pp$) collisions. One such test is the measurement of the charge asymmetry. The production of \ttbar~pairs at hadron colliders is symmetric under charge
conjugation at leading order (LO) in quantum chromodynamics (QCD),
i.e., the probability of a top quark flying in a given direction is the
same as for an antitop quark~\cite{Aguilar-Saavedra:2014kpa}. At
next-to-leading order (NLO) in QCD, an asymmetry arises from
interference between different Feynman diagrams~\cite{AXI}. In particular, interference
between the Born and one-loop diagram of the $q\bar{q} \to t\bar{t}$~processes
and between $q\bar{q} \to t\bar{t}g$~diagrams with initial-state and
final-state radiation (ISR and FSR) processes lead to a charge asymmetry. In the
\ttbar~rest frame, this
asymmetry causes the top quark to be preferentially emitted in the
direction of the initial quark, and causes the antitop quark to be
emitted in the direction of the initial antiquark. The size of the
asymmetry can be enhanced by
contributions beyond the SM, for example, \ttbar~production via the
exchange of new heavy particles such as axigluons~\cite{AXI}, heavy
$Z$ particles~\cite{ZP}, or colored Kaluza--Klein excitations of the
gluon~\cite{KK}.
Inclusive and differential measurements of the \ttbar~asymmetry were
first performed at the Tevatron proton--antiproton collider,
where forward-backward asymmetries were measured. Several
measurements were reported by the CDF
and D0 experiments~\cite{CDF2,CDF1,Aaltonen:2012it,D02,D03,Abazov:2015fna}
in dileptonic and semileptonic \ttbar~events. For these
measurements, the direction of the initial quark can be assumed
to be the direction of the proton, and the direction of the
antiquark that of the antiproton, which yields straightforward
access to the asymmetry. Initial tension between these measurements
and theory predictions have
been reduced with the latest SM calculations at next-to-next-to-leading order
(NNLO) QCD~\cite{Czakon:2014xsa}.
Since the start of the LHC, measurements of $t\bar{t}$ charge asymmetries have
been performed by the ATLAS and CMS experiments. Two features
complicate the measurement of the asymmetry
at the LHC: in proton--proton collisions the initial state is symmetric, so
there is no \ttbar~forward-backward asymmetry,
and the dominant production mechanism is
gluon fusion, which is symmetric under charge conjugation to all
orders in perturbative QCD. However, valence quarks carry on average a
larger fraction of the proton momentum than sea antiquarks, hence top antiquarks
produced through quark--antiquark annihilation are more central than top quarks~\cite{Kuhn:1998jr}.
By using differences between the absolute
rapidity of the top and antitop quarks, ATLAS and CMS performed
measurements of the charge asymmetry in dileptonic and semileptonic
events at $\sqrt{s} = 7$~\TeV~and $8$~\TeV~\cite{atlas7tev,atlas7tevljets,cms7tevljets,cms7tevdil,atlas8tevljets,cms8tevljets,Khachatryan:2015mna,Khachatryan:2016ysn}.
All asymmetry measurements at the LHC show good agreement with the SM
prediction~\cite{Bernreuther:2012sx}, which is approximately an order of magnitude smaller than the predicted
asymmetry at the Tevatron.
In this article, new measurements of the charge asymmetry are presented
using dileptonic \ttbar~events at $\sqrt{s} = 8$~\TeV. The dileptonic channel is characterized by two charged leptons (denoted $\ell=e,\mu$), coming from either a direct
vector boson decay or through an intermediate $\tau$ lepton decay. Two different observables are used, based either on the the selected leptons or the reconstructed $t\bar{t}$ final state. Inclusive and
differential measurements as a function of the invariant mass of
the \ttbar~system ($m_{t\bar{t}}$), the transverse momentum of
the \ttbar~system ($p_{\rm{T},t\bar{t}}$), and the absolute value of the boost of the
\ttbar~system along the beam axis ($\beta_{z,t\bar{t}}$) are performed.
The inclusive and differential measurements are performed in the full phase space as well as in a
fiducial volume based on the detector acceptance and selection
requirements, using particle-level objects.
The measurement in the fiducial region does not rely on extrapolating
to regions of phase space that are not within the detector acceptance, while the
full phase space measurement has the benefit of being comparable to
theoretical calculations at the parton level, including BSM models.
In Sec.~\ref{sec:detector}, a brief description of the ATLAS detector is
given. Section~\ref{sec:montecarlo} describes the data and Monte Carlo (MC) samples,
and Sec.~\ref{sec:selection} the event selection and background estimation.
The observables are described in
Sec.~\ref{sec:observables}. Section~\ref{sec:analysis} outlines the
measurement methods, including a description of
the \ttbar~reconstruction, the definition of fiducial volume, and a
description of the unfolding procedure. In
Sec.~\ref{sec:systematics}, the sources of systematic uncertainties
affecting the measurements are discussed, and results are provided in
Sec.~\ref{sec:result}. Finally, conclusions are given in Sec.~\ref{sec:conclusions}.
\section{ATLAS detector}
\label{sec:detector}
The ATLAS detector~\cite{ATLAS} at the LHC covers nearly the entire solid angle around the interaction
point.\footnote{ATLAS uses a right-handed coordinate system with its origin at the nominal interaction point (IP) in
the center of the detector and the $z$-axis along the beam pipe. The $x$-axis points from the IP to the center of the LHC ring, and the $y$-axis points upward. Cylindrical coordinates $(r, \phi)$ are used in the transverse
plane, $\phi$ being the azimuthal angle around the beam pipe.}
It consists of an inner tracking detector surrounded by a thin superconducting solenoid, electromagnetic and hadronic calorimeters, and a muon spectrometer incorporating
superconducting toroid magnets.
The inner-detector system is immersed in
a $2$ T axial magnetic field and provides charged-particle tracking in
the pseudorapidity\footnote{The pseudorapidity is defined in terms of
the polar angle $\theta$ as $\eta = - \ln[\tan \theta/2]$, while the
rapidity $y$ is defined as $y = - (1/2) \ln[(E + p_z)/(E - p_z)]$.} range $|\eta| <2.5 $. A high-granularity silicon pixel detector covers the interaction region and
provides typically three measurements per track. It is surrounded by a silicon microstrip tracker
designed to provide four two-dimensional measurement points per track. These silicon detectors
are complemented by a transition radiation tracker, which enables radially extended
track reconstruction up to $|\eta| =2.0 $. The transition radiation tracker also provides electron
identification information based on the fraction of hits (typically 30 in total) exceeding an
energy-deposit threshold corresponding to transition radiation.
The calorimeter system covers the pseudorapidity range $|\eta| <4.9 $. Within the region
$|\eta| <3.2 $, electromagnetic calorimetry is provided by barrel and endcap high-granularity
lead/liquid-argon (LAr) electromagnetic calorimeters, with an additional thin LAr presampler
covering $|\eta| <1.8 $ to correct for energy loss in the material upstream of the calorimeters.
Hadronic calorimetry is provided by a steel/scintillator-tile calorimeter, segmented into
three barrel structures within $|\eta| <1.7 $, and two copper/LAr hadronic endcap calorimeters.
The solid angle coverage is completed with forward copper/LAr and tungsten/LAr
calorimeters used for electromagnetic and hadronic measurements, respectively.
The muon spectrometer comprises separate trigger and high-precision tracking chambers
measuring the deflection of muons in a magnetic field generated by superconducting
air-core toroids. The precision chamber system covers the region $|\eta| <2.7 $ with drift tube chambers, complemented by cathode strip chambers. The muon trigger system covers the range of $|\eta| <1.05$ with resistive plate chambers
in the barrel, and the range of $1.05< |\eta| <2.4$ with thin gap chambers in the endcap regions.
A three-level trigger system is used to select interesting events. The Level-1 trigger is
implemented in hardware and uses a subset of detector information to reduce the event rate
to a design value of at most 75 kHz. This is followed by two software-based trigger levels,
which together reduce the event rate to about 300 Hz.
\section{Data and Monte Carlo samples}
\label{sec:montecarlo}
The data used for this analysis were collected during the 2012 LHC
running period at a center-of-mass energy of $8$ \TeV. After
applying data-quality selection criteria, the data sample used in the
analysis corresponds to an integrated luminosity of $20.3$
$\rm{fb}^{-1}$.
For the modeling of the signal processes and most background
contributions, several MC event generators are used. The main
background contribution in this measurement comes from Drell--Yan
production of $Z/\gamma^{*} \rightarrow \ell \ell$, which is estimated
by a combination of simulated samples modified with corrections
derived from data, as described in
Sec.~\ref{sec:selection}. The smaller
contributions from diboson ($WW$, $ZZ$, and $WZ$) and single-top-quark ($Wt$
channel) production are evaluated purely via MC simulations. Further background contributions can arise from events including a jet or a
lepton from a semileptonic hadron decay misidentified as an isolated
charged lepton as well as leptons from photon conversions, together
referred to as ``fake leptons''.
This contribution is estimated using simulated
samples, modified with corrections derived from data. The samples mentioned above together with simulated samples of $t\bar{t}+W/Z$, $t$-channel of single-top-quark production,
$W$+jets, and $W$+$\gamma$+jets are included in the estimation. The estimation
procedure is described in Sec.~\ref{sec:selection}.
The nominal $t\bar{t}$ signal sample is generated at NLO in QCD using \powheg
(version 1, r2330)~\cite{Alioli:2010xd,Frixione:2007vw,Nason:2004rx}
and the CT10~\cite{ct10} parton distribution function (PDF) set, setting the
$h_{\rm{damp}}$ parameter to the top-quark mass of $172.5 \gev$. The
$h_{\rm{damp}}$
parameter is the resummation scale that is used in the damping
function, which is designed to limit the resummation of higher-order
effects at large transverse momentum without spoiling the NLO accuracy
of the cross section.
The parton shower, hadronization, and underlying event are simulated using \pythia (version 6.427)~\cite{Sjostrand:2006za}
with the CTEQ6L1 PDF~\cite{Pumplin:2002vw} and the corresponding set of
tunable parameters (Perugia 2011C tune~\cite{Skands:2010ak})
intended to be used with this PDF.
The $t\bar{t}$ cross section for $pp$ collisions at a center-of-mass energy of $8$~\TeV~ is set to
$\sigma_{t\bar{t}}= 253^{+13}_{-15}$~pb, calculated at NNLO in QCD including resummation of next-to-next-to-leading logarithmic (NNLL) soft gluon terms with top++2.0~\cite{Cacciari:2011hy,Beneke:2011mq,Baernreuther:2012ws,Czakon:2012zr,Czakon:2012pz,Czakon:2013goa,Czakon:2011xx}. The PDF and $\alpha_{\rm{S}}$ uncertainties were calculated using the PDF4LHC prescription~\cite{Botje:2011sn} with the MSTW2008 $68\%$ CL NNLO~\cite{Martin:2009iq,Martin:2009bu}, CT10 NNLO~\cite{Lai:2010vv,Gao:2013xoa}, and NNPDF2.3 5f FFN~\cite{Ball:2012cx} PDF sets, and added in quadrature to the scale uncertainty.
Single-top-quark production in the $Wt$ channel is simulated using \powheg
with \pythia (version 6.426) and the CT10 (NLO) PDF set. The cross
section of $ 22.3 \pm 1.5$~pb is estimated at approximate NNLO in
QCD including resummation of NNLL terms~\cite{Kidonakis:2010ux}.
The parton shower, hadronization, and underlying event are simulated
by \pythia using the Perugia 2011C tune. The Drell--Yan process is modeled using \alpgen (version 2.14)~\cite{Mangano:2002ea} interfaced
with \pythia with the CTEQ6L1~\cite{Pumplin:2002vw} PDF set using
the MLM matching scheme. Its
heavy-flavor component is included in the matrix element calculations
to model the $Z/\gamma^{*} + b\bar{b}$ and $Z/\gamma^{*} + c\bar{c}$
processes. Diboson processes ($WW$, $ZZ$, and $WZ$) are simulated using \alpgen interfaced with \herwigjimmy (version
4.31)~\cite{Corcella:2000bw,JIMMY} with the
CTEQ6L1~\cite{Pumplin:2002vw} PDF set for parton fragmentation~\cite{Mangano:2001xp}.
The only exceptions are the same-charge $W^{+(-)}W^{+(-)}$ samples, which
are simulated using \madgraph (version 5.1.4.8)~\cite{Alwall:2011uj}
interfaced with \pyEight (version 8.165)~\cite{Sjostrand:2007gs}. The
samples are normalized to the reference NLO QCD prediction, obtained
using the MCFM generator~\cite{Campbell:1999ah}.
The associated production of a $t\bar{t}$ pair with a vector boson ($t\bar{t}Z$ and $t\bar{t}W$) is simulated with \madgraph interfaced with \pyEight and normalized to NLO cross-section calculations~\cite{Campbell:2012dh,Garzelli:2012bn}. The $W$+jets events are simulated
using \alpgen interfaced with \pythia and the $W$+$\gamma$+jets process
is simulated using \alpgen interfaced with \jimmy.
To model the LHC environment properly, additional inelastic $pp$
collisions are generated with \pyEight\ and overlaid on the hard
process. All the simulated samples are then processed through a simulation of the ATLAS detector~\cite{Aad:2010ah}. For most of the samples, a full simulation based
on GEANT4~\cite{geant4} is used. Some of the samples used to evaluate the generator modeling uncertainties are obtained using a faster detector simulation where only the calorimeter simulation is modified and relies on parametrized showers~\cite{ATL-PHYS-PUB-2010-013}. The simulated events are passed through
the same reconstruction and analysis chain as data.
\section{Event selection and background estimation}
\label{sec:selection}
In order to enrich the data sample in dileptonic \ttbar~events,
requirements are imposed on reconstructed charged leptons (electrons and
muons), jets, and the missing transverse momentum. Three different final
states are considered in the analysis: events with two electrons in
the final state ($ee$), with one electron and one muon ($e\mu$), and
with two muons ($\mu\mu$).
Electron candidates are reconstructed from an electromagnetic
calorimeter energy deposit matched to a track in the inner
detector and must pass the likelihood-based ``medium'' identification
requirements~\cite{Aad:2014fxa}. They are required to have
transverse momentum $p_T > 25$ \GeV~and must also lie in the region
\mbox{$|\eta_{\rm cl}| < 2.47$}, where $\eta_{\rm cl}$ is the pseudorapidity
of the calorimeter energy cluster associated with the electron,
excluding the transition region between the calorimeter
barrel and endcaps $1.37 < |\eta_{\rm cl}| <
1.52$. Moreover, electrons are required to be isolated from
surrounding activity in the inner detector. The
scalar sum of the track $p_{\rm{T}}$ within a cone of
$\Delta R = \sqrt{(\Delta \eta)^2+(\Delta \phi)^2} = 0.3$ (excluding the track of
the electron itself) divided by the electron $p_{\rm{T}}$ should be less than 0.12.
Muon candidates are reconstructed using combined information from the muon
spectrometer and the inner detector~\cite{Aad:2014rra}. They are required to
have $\pt > 25 \GeV$ and $|\eta| < 2.5$. In addition, muons are
required to satisfy track-based \pt-dependent isolation
criteria. The scalar sum of the track \pt\ within a cone of size \mbox{$\Delta R =
10~\textrm{\GeV}/p_{\rm{T}}^{\mu}$} around the muon (excluding the muon track
itself) must be less than 5\% of the muon \pt\ ($p_{\rm{T}}^{\mu}$).
Both the electrons and muons have to be consistent with the primary vertex,\footnote{The primary vertex is defined as the
reconstructed vertex with at least five associated tracks (of $p_T >$
0.4 \GeV) and the highest sum of the squared transverse momenta
of the associated tracks.} by requiring the absolute value of the
longitudinal impact parameter to be less than 2~mm.
Jets are reconstructed from clustered energy deposits in the
electromagnetic and hadronic calorimeters, using the anti-$k_{t}$~\cite{Cacciari:2008gp} algorithm with a radius
parameter $R = 0.4$.
The measured energy of the jets is corrected to the hadronic scale using \pt-
and $\eta$-dependent scale factors derived from simulation and validated in
data~\cite{Aad:2014bia}.
After the energy correction, the jets are required to have
$\pt > 25 \GeV$, to be in the pseudorapidity range $|\eta| < 2.5$, and to have a jet vertex fraction
$|\textrm{JVF}| > 0.5$ \cite{Aad:2015ina} if $\pt < 50 \GeV$.
The jet vertex fraction is defined as the summed scalar $p_{\rm T}$ of the tracks associated with both the jet and the primary vertex divided by the summed scalar $p_{\rm T}$ of all tracks in the jet.
The jet that is the closest to a selected electron is removed from
the event if their separation is $\Delta R < 0.2$.
After this jet overlap removal, electrons and muons that are within a cone of $\Delta R = 0.4$ around the closest jet are removed.
Jets containing $b$-hadrons are identified ($b$-tagged) using a
multivariate algorithm
(MV1)~\cite{Aad:2015ydr}. This is a neural-network-based algorithm that makes use of track impact parameters
and reconstructed secondary vertices. Jets are identified as $b$-tagged
jets by requiring the MV1 output discriminant to be above a certain
threshold value. This value is chosen such that the overall tagging
efficiency for $b$-jets with $\pt > 20 \GeV$
and $|\eta| < 2.5$ originating from top-quark decays in dileptonic MC
$t\bar{t}$ events is 70\%.
The rejection factor for jets originating from gluons and light quarks
is about 130, while for $c$-quarks it is about 5.
The magnitude of the missing transverse momentum (\met) is calculated from the
negative vector sum of all calorimeter energy deposits and the momenta of
muons~\cite{Aad:2012re}.
The calculation is refined by the application of the object-level
corrections for the contributions arising from identified electrons and
muons.
Events recorded with single-lepton triggers
($e$ or $\mu$) under stable beam conditions with all detector
subsystems operational are considered. The transverse momentum thresholds are $24$ \GeV~for isolated single-lepton triggers and $60$ ($36$) \GeV~for nonisolated single-electron (single-muon) triggers. The nonisolated triggers are
used to select events that fail the isolation requirement at trigger
level but pass it in the offline analysis.
In all three final states, exactly two isolated leptons with opposite
charge and an invariant mass $m_{\ell \ell} > 15$ \GeV~are required,
together with at least two jets.
In the same-flavor channels ($ee$ and $\mu\mu$),
the invariant mass of the two charged leptons is required to be
outside of the $Z$ boson mass window such
that \mbox{$|m_{\ell \ell}-m_Z|>10$ \GeV}.
Furthermore, it is required that $\met > 30$ \GeV~and at
least one of the jets must be $b$-tagged.
These requirements suppress the dominant background contribution from
Drell--Yan production of $Z/\gamma^{*} \rightarrow \ell \ell$ and also
suppress diboson backgrounds.
In the $e\mu$ channel, the background contamination is much smaller
and the background suppression is achieved by
requiring the scalar sum of the $p_{\rm{T}}$ of the two leading jets and
leptons $(H_{\rm{T}})$ to be larger than $130$ \GeV.
The event selection requirements are summarized in Table~\ref{tab:cuts}.
\begin{table}[h]
\centering
\caption{The summary of the event selection requirements applied in different channels.\label{tab:cuts}}
\begin{tabular}{lcc}
\hline
\hline
Requirements & $ee/\mu\mu$ & $e\mu$ \\ \hline
Leptons & 2 & 2\\
Jets & $\ge 2$ & $\ge 2$\\
$m_{\ell \ell}$ & $> 15 $ \GeV & $> 15 $ \GeV \\
$|m_{\ell \ell}-m_Z|$ & $> 10 $ \GeV & -- \\
$\met$ & $> 30 $ \GeV & -- \\
$b$-tagged jets & $\ge 1$ & -- \\
$H_{\rm{T}}$ & -- & $> 130 $ \GeV \\
\hline
\hline
\end{tabular}
\end{table}
The modeling of Drell--Yan events in the
same-flavor channels with $\met > 30$ \GeV~may not be
accurate in simulation due to the mismodeling of the $\met$ distribution.
Moreover, after applying the $b$-tagging requirement, a large
contribution to the background comes from the associated production of
$Z$ bosons with heavy-flavor jets, which is not well predicted by MC
simulation. The first source of mismodeling depends on the reconstructed objects and is therefore different in each channel. The second source is a limitation of the MC simulation and is expected to be the same in both channels. Thus, the normalization of the
inclusive and heavy-flavor component of the Drell-Yan background in the
same-flavor channels is computed simultaneously using
data in two control regions with three scale factors.
Two scale factors are applied to all Drell--Yan events to take into
account the mismodeling from the $\met$ requirement (one in the $ee$
and one in the $\mu\mu$ channel) while another is applied only to
$Z$+heavy-flavor events.
The control regions are defined using the standard selection described
previously but inverting the $m_{\ell \ell}$ cut to be within the $Z$
mass window.
The first control region is defined without the $b$-tagging requirement
while the second is defined with at least one $b$-tagged jet. The simulated $m_{\ell \ell}$ distribution in these control regions is simultaneously fit to the data and the scale factors are extracted.
The scale factors derived in these two regions are $0.927 \pm 0.005$ and $0.890 \pm 0.004$ for the $ee$ and $\mu\mu$ channels, respectively, and $1.70\pm0.03$ for the heavy-flavor component. The $Z \rightarrow \tau\tau$ process in the $e\mu$ channel is estimated using MC simulation only: no data-driven correction is applied since neither the $\met$ requirement nor $b$-tagging requirement are applied to this channel.
The background arising from misidentified and nonprompt (NP) leptons is
determined using both MC simulation and data.
The dominant sources of these fake leptons are semileptonic $b$-hadron decays, long-lived
weakly decaying states (such as $\pi^{\pm}$ or $K^{\pm}$ mesons), $\pi^{0}$
showers, photons reconstructed as electrons, and electrons from photon conversions.
$W$+jets, $W$+$\gamma$+jets, \ttbar, $t\bar{t}Z$, $t\bar{t}W$, Drell--Yan, single-top-quark, and diboson production
are taken into account for the estimation of this background.
Multijet events do not contribute significantly to this background,
since the probability of having two jets misidentified as isolated leptons is very small.
The shapes of the kinematic distributions are taken from simulated events where
at least one of two selected leptons is required not to be matched
with the MC generator-level leptons.
Scale factors are derived from data in order to adjust the normalization.
A control region, enriched in fake leptons, is defined by applying the same cuts as
for the final selection but requiring the two leptons to have the same charge.
The shapes of the distributions for various kinematic variables of
leptons, jets, and \met are checked and found to be well modeled
in the MC simulation.
The scale factors are derived in this region by comparing data and
simulation and are then applied to the simulated events in the signal region.
The scale factor is $1.2 \pm 0.3$ in the $ee$ channel, $1.1 \pm 0.2$ in the $e\mu$
channel, and $3.7 \pm 0.8$ in the $\mu\mu$ channel, where the
uncertainties are statistical. The sources of misidentified muons, such as heavy-flavor decays, are quite different from those of misidentified electrons. The large difference between the scale factor for the $\mu\mu$
and the $e\mu$ channel is mainly due to the $b$-tagging requirement, that is
applied only in the $\mu\mu$ channel. However, the shapes of the distributions of the relevant kinematic variables
in the $\mu\mu$ channel are cross-checked in control regions and found
to be consistent with the distributions from a purely data-driven
method.
The systematic uncertainties of both Drell--Yan background and the
background due to events from misidentified and nonprompt leptons are
discussed in detail in Sec.~\ref{sec:syst_background}.
The numbers of events for both expectation and data after
applying the selection criteria are shown in Table~\ref{tab:evtnumbers}
for the three final states. The uncertainties shown correspond to the total
uncertainty (including the statistical uncertainties from the limited
size of the MC simulated samples, as well as the systematic
uncertainties). The $e\mu$
channel contributes with the largest number of events, followed by
$\mu\mu$ and $ee$. Figure~\ref{fig:dataMC1} shows good agreement
within the systematic uncertainties between data and the predictions
as a function of jet multiplicity, lepton $p_T$ and $\eta$, for all
channels combined.
\begin{table}[htbp]
\begin{center}
\caption{Observed numbers of data events compared to the
expected signal and background contributions in the three decay channels.
The uncertainty corresponds to the total uncertainty in the given
process. Data-driven (DD) scale factors are applied to the $Z+$jets
and the NP \& fake leptons contributions. The
$Z \rightarrow \tau\tau$ process in the $e\mu$ channel is estimated
using MC simulation only. \label{tab:evtnumbers}
}
\renewcommand\arraystretch{1.1}
\centering
\resizebox{0.6\linewidth}{!}{
\sisetup{retain-explicit-plus}
\begin{tabular}{
l
r@{$\,\,$}
c@{$\,\,$}
r@{$\,\,\,\,\,\,$}
r@{$\,\,$}
c@{$\,\,$}
r@{$\,\,\,\,\,\,$}
r@{$\,\,$}
c@{$\,\,$}
r
}
\hline
\hline
Channel & \multicolumn{3}{c}{ $ee$ } & \multicolumn{3}{c}{$\mu\mu$ } & \multicolumn{3}{c} {$e\mu$} \\
\hline
$t\bar{t}$ & 10200 & $\pm$ & 800 & 12100 & $\pm$ & 800 & 36000 & $\pm$ & 2400 \\
Single-top & 510 & $\pm$ & 50 & 590 & $\pm$ & 50 & 1980 & $\pm$ & 170 \\
Diboson & 31 & $\pm$ & 5 & 40 & $\pm$ & 6 & 1320 & $\pm$ & 100\\
$Z \rightarrow ee $ (DD) & 1200 & $\pm$ & 260 & \multicolumn{3}{c}{--} & \multicolumn{3}{c}{--} \\
$Z \rightarrow \mu\mu $ (DD) & \multicolumn{3}{c}{--} & 1520 & $\pm$ & 300 & \multicolumn{3}{c}{--} \\
$Z \rightarrow \tau\tau $ (DD/MC)& 31 & $\pm$ & 15 & 58 & $\pm$ & 25 & 1120 & $\pm$ & 430 \\
NP \& fake leptons (DD) & 62 & & $^{\numRP{+119}{3}}_{\numRP{-29}{2}}$ & 45 & & $^{+36}_{-24}$ & 480 & & $^{+240}_{-220}$ \\
Total Expected & 12000 & $\pm$ & 900 & 14400 & $\pm$ & 800 & 40900 & $\pm$ & 2500\\ \hline
Data & \multicolumn{3}{c}{12785} & \multicolumn{3}{c}{14453} &\multicolumn{3}{c}{42363} \\ \hline \hline
\end{tabular}}
\end{center}
\end{table}
\begin{figure}[!htbp]
\begin{center}
\includegraphics[width=0.49\textwidth]{fig_01a.pdf}
\includegraphics[width=0.49\textwidth]{fig_01b.pdf}\\
\includegraphics[width=0.49\textwidth]{fig_01c.pdf}
\caption{Distributions of the jet multiplicity, lepton $p_T$, and lepton
$\eta$ for data (points) and predictions (histograms) for all
channels combined after event selection. The data/expected
ratio is also shown. The shaded area corresponds to the detector
systematic uncertainty, the signal modeling systematic uncertainty, and the normalization uncertainty in signal and
background. In the lepton $p_{\rm{T}}$ distribution, the last bin includes the overflow. }
\label{fig:dataMC1}
\end{center}
\end{figure}
\section{Observables}
\label{sec:observables}
In dileptonic events, the charge asymmetry can be measured in two
complementary ways: using the pseudorapidity of the charged leptons or
using the rapidity of the top quarks.
The asymmetry based on the charged leptons uses the difference of the absolute pseudorapidity values of the positively and negatively charged leptons, $|\eta_{\ell^{+}}|$ and
$|\eta_{\ell^{-}}|$
\begin{linenomath}
\begin{equation}
\Delta |\eta| = |\eta_{\ell^{+}}|-|\eta_{\ell^{-}}|.
\end{equation}
\end{linenomath}
The leptonic asymmetry is defined as
\begin{linenomath}
\begin{equation}
\label{eq:ac_lep}
A_{\rm{C}}^{\ell\ell} = \frac{N(\Delta |\eta| >0) - N(\Delta |\eta| <0)}{N(\Delta |\eta|
>0) + N(\Delta |\eta| <0)} ,
\end{equation}
\end{linenomath}
where $N(\Delta |\eta| >0)$ and $N(\Delta |\eta| <0)$ represent the number of events with
positive and negative $\Delta |\eta|$, respectively.
The SM prediction at NLO in QCD, including electroweak corrections, is
$\Acll = \Aclltheory$~\cite{Bernreuther:2012sx}, where the uncertainty
includes variations in scale and choice of PDF.
The leptonic asymmetry, that is slightly diluted with respect to the underlying top-quark asymmetry, has the advantage that no reconstruction of the
top--antitop quark system is required. Furthermore, it is also
sensitive to top-quark polarization effects, which occur in some
models predicting enhanced charge asymmetries.
For the \ttbar~charge asymmetry, the \ttbar~system has to be reconstructed and the absolute values of the top and antitop quark rapidities ($|y_{t}|$ and $|y_{\bar{t}}|$, respectively) need to be computed.
Using
\begin{linenomath}
\begin{equation}
\Delta |y| = |y_{t}|-|y_{\bar{t}}|,
\end{equation}
\end{linenomath}
the $t\bar{t}$ charge asymmetry is defined as
\begin{linenomath}
\begin{equation}
\label{eq:ac}
A_{\rm{C}}^{t\bar{t}}= \frac{N(\Delta |y| > 0) - N(\Delta |y| < 0)}{N(\Delta |y|
>0) + N(\Delta |y| <0)},
\end{equation}
\end{linenomath}
where $N(\Delta |y| > 0)$ and $N(\Delta |y| < 0)$ represent the number of events with
positive and negative $\Delta |y|$, respectively. The top (antitop) quarks are identified as those giving rise
to positive (negative) leptons.
The SM prediction at NLO QCD, including electroweak corrections, is
$\Ac = \Actheory$~\cite{Bernreuther:2012sx}.
The measurements of $\Acll$ and $\Ac$ are performed inclusively and
differentially as a function of $m_{t\bar{t}}$, $p_{\rm{T},t\bar{t}}$, and
$\beta_{z,t\bar{t}}$. The fractions of quark--antiquark annihilation
and gluon fusion processes change as a function of
$m_{t\bar{t}}$, and thus an increasing asymmetry for increasing
$m_{t\bar{t}}$ is expected. Since $p_{\rm{T},t\bar{t}}$ depends on the
initial-state radiation, the asymmetry value is
expected to change as a function of $p_{\rm{T},t\bar{t}}$. In particular, the
contribution to the asymmetry from interference of diagrams with initial- and final-state radiation is negative, resulting in decreasing asymmetries
with increasing $p_{\rm{T},t\bar{t}}$. While the initial antiquark is
always a sea quark, the initial quark can be a valence quark. On
average, valence quarks have higher momenta than sea quarks, which can
result in a boost of the $t\bar{t}$ system in the direction of the
incoming quark. This results in an increased charge asymmetry for
increasing $\beta_{z,t\bar{t}}$. The asymmetry is also expected to be
different inclusively and differentially in different BSM models.
\section{Asymmetry measurements}
\label{sec:analysis}
The following measurements are performed:
\begin{itemize}
\item inclusive measurements of the \ttbar and leptonic asymmetries, corrected for reconstruction and acceptance effects to parton level in the full phase space;
\item inclusive measurements of the \ttbar and leptonic asymmetries, corrected for reconstruction effects to particle level in the fiducial region;
\item differential measurements of the \ttbar and leptonic asymmetries as a function of $m_{t\bar{t}}$, $p_{\rm{T},t\bar{t}}$, and $\beta_{z,t\bar{t}}$ in the fiducial region and the full phase space.
\end{itemize}
Particle-level results consider stable particles with a mean lifetime
larger than $0.3 \times 10^{-10}$~s. For the parton-level
measurements, MC generator-level objects are used. The
parton-level top quarks and leptons are selected after radiation.
The leptonic asymmetry can be extracted directly using the pseudorapidities of the measured charged leptons. For the \ttbar~charge asymmetry, the reconstruction of the top and antitop quark four-momenta is necessary. A kinematic method is used for the reconstruction, as described in Sec.~\ref{sec:reconstruction}.
Section~\ref{sec:fiducial} details the definition of the fiducial volume and the particle-level objects used for the fiducial measurement. In order to correct the measured asymmetry distributions for detector and acceptance effects, an unfolding method, described in Sec.~\ref{sec:unfolding}, is used for all asymmetry measurements. Section~\ref{sec:measurement} describes how the various asymmetries are extracted.
\subsection{Top and antitop quark reconstruction} \label{sec:reconstruction}
For the reconstruction of the top and antitop quark four-momenta, a
kinematic reconstruction is used. The reconstruction is
performed by solving the system of equations that relate the particle
momenta at each of the decay vertices in the $t\bar{t} \rightarrow
W^{+}bW^{-}\bar{b} \rightarrow \ell^{+} \nu_{\ell} b \ell^{-} \bar{\nu_{\ell}} \bar{b} $ process.
Two neutrinos are produced and escape undetected. Thus, an
underconstrained system is obtained. This system is solved using the
kinematic (KIN) method~\cite{Abulencia:2006js,Aaltonen:2012tk},
assuming values of $172.5$~\GeV~and $80.4$~\GeV~for the top quark and $W$ boson masses, respectively, which
allows the system of equations to be solved numerically by the
Newton–Raphson method.
If there are more than two reconstructed jets in a given event,
the two jets with the highest $b$-tagging weights (as determined by
the MV1 $b$-tagging algorithm) are used. This improves
the probability of choosing the correct jets, compared to just
choosing the two jets with the highest $p_{T}$,
from about 54\% to about 69\% in the inclusive selected sample. The
experimental uncertainties of the
measured objects (described in Sec.~\ref{sec:systematics}) are taken
into account by sampling the phase space of the measured jets and
$\MET$ according to their resolution in simulation.
The number of sampled points is called $N_{\rm{smear}}$, whose optimization is based on the time and efficiency of the top-pair reconstruction.
The resolution functions, obtained from the $t\bar{t}$ simulated sample, with respect to the jet $p_{\rm{T}}$ (for jets) and the
total transverse momentum in the event (for $\MET$) are used for the sampling.
For each sampling point, up to four solutions can be obtained.
The KIN method chooses the solution that leads to the lowest reconstructed mass of the $t\bar{t}$
system. The reason for this is that the $t\bar{t}$ cross section is a
decreasing function of the partonic center-of-mass energy
$\sqrt{\hat{s}}\simeq m_{t\bar{t}}$, so events with smaller
$m_{t\bar{t}}$ are more likely.
There is also a twofold ambiguity in the lepton and $b$-jet
assignment. The correct assignment to the top and antitop quarks is
chosen to be the one that has more reconstructed trials
$N_{\rm{smear}}^{\rm{reco}}$, i.e., the one that maximizes
$N_{\rm{smear}}^{\rm{reco}}/N_{\rm{smear}}$.
The chosen solution is either the solution found using the nominal
jet energies and measured $\MET$, if available, or the first solution found during
the sampling.
The kinematic reconstruction fails for a given
event if no solution is found in any of the $N_{\rm{smear}}$ sampled
points. This is possible if, for example, the solution does not
converge within a given number of iterations.
The performance of the method is quantified by evaluating the
efficiency of reconstructing $t\bar{t}$ events that pass dilepton
event selection, and the probability of reconstructing the correct
sign of $\Delta |y|$. These probabilities are found to be 90\% and
76\%, respectively. The reconstruction efficiency is consistent
between data and the prediction.
Figure~\ref{fig:dataMC1reco} shows the distributions for data and prediction of the $p_T$, mass, and longitudinal boost of the $t\bar{t}$ system after applying the reconstruction method. Good agreement between data and prediction is found.
\begin{figure}[!htbp]
\begin{center}
\includegraphics[width=0.49\textwidth]{fig_02a.pdf}
\includegraphics[width=0.49\textwidth]{fig_02b.pdf}
\includegraphics[width=0.49\textwidth]{fig_02c.pdf}\\
\caption{Distributions of $\pttt$, $\mtt$, and $|\betatt|$ for data (points) and predictions (histograms)
after kinematic reconstruction. The data/expected ratio is also
shown. The shaded area corresponds to the detector
systematic uncertainty, the signal modeling systematic uncertainty, and the normalization uncertainty in signal and
background. In the
$\pttt$ and $\mtt$ distributions, the last bin includes the
overflow. }
\label{fig:dataMC1reco}
\end{center}
\end{figure}
\subsection{Particle-level objects and fiducial region} \label{sec:fiducial}
A fiducial region is defined in order to closely match the phase space
region accessed with the ATLAS detector and the requirements made on the reconstructed objects.
A fiducial measurement usually allows for MC generator dependencies to
be reduced, since it avoids large extrapolation to the full phase space.
In the fiducial region, only objects defined at particle level are used.
The considered charged leptons (electrons and muons) are required not to originate from hadrons.
Photons within $\Delta R = 0.1$ around the charged lepton are included in the four-momentum calculation.
The $\MET$ is
calculated as the summed four-momenta of neutrinos from the $W/Z$ boson decays, including those from $\tau$ decays.
Jets are reconstructed using the anti-$k_{t}$ algorithm with a radius
parameter $R = 0.4$. The electrons, muons,
neutrinos, and photons that are used in the definition of the selected
leptons are excluded from the clustering. Finally, identification of jets originating from $b$-quarks
is achieved using ghost matching~\cite{Cacciari:2008gn}.
The MC generator-level $b$-hadrons are clustered into the
particle-level jets, with their momenta scaled to a very small value. If a clustered
jet is found to contain a $b$-hadron, the particle-level jet is labelled as a
$b$-jet.
The fiducial volume is defined by requiring at least two particle-level jets and at least two leptons in the event,
both objects with $p_T >25$~\GeV~ and $|\eta| < 2.5$.
Events where leptons and jets overlap, within $\Delta R$ of $0.4$, are rejected. The particle-level jets are not required to be $b$-jets since this requirement is not shared between the three channels in the selection.
Using these objects, the reconstruction of top quarks (known as pseudotops~\cite{Aad:2015eia}) can be performed.
The assignment of the proper jet-lepton-neutrino permutation is chosen by first minimizing the difference between the mass computed from each lepton-neutrino combination and the $W$ boson mass value used in the MC simulation. Then, the difference between the mass of each combination of the chosen lepton-neutrino pairs with a jet and the top quark mass value, used in the MC simulation, is minimized. The $b$-jets are prioritized over the light jets for the proper jet-lepton-neutrino assignment.
The correlation coefficient between \deltay at the parton
and particle levels is found to be 79\%, while for \deltaeta it is 99\%.
The measurements of the asymmetry in the fiducial volume require the
treatment of an additional background contribution,
in which signal events from outside of the fiducial region migrate
into the detector acceptance due to resolution effects.
This nonfiducial background constitutes about 8\% of the expected
\ttbar events after selection,
as estimated by using MC simulation, and it was found to be independent
of the charge asymmetry value of the simulated sample. A bin-by-bin
scale factor derived from simulation is applied to background-subtracted data to estimate the contribution of these events.
\subsection{Unfolding} \label{sec:unfolding}
The measurements are corrected for detector resolution and acceptance effects.
These corrections are performed using the fully Bayesian unfolding (FBU) technique~\cite{Fbu2012arXiv1201.4612C}.
The FBU procedure applies Bayes' theorem to the problem of unfolding.
This application can be stated in the following terms: given an observed spectrum
$\Data$ with $N_r$ reconstructed bins and a migration matrix
$\TrasfMatrix$ with $N_r \times N_t$ bins giving the detector
response to a true spectrum with $N_t$ bins, the posterior probability
of the true spectrum $\Truth{}$ with $N_t$ bins follows the
probability density
\begin{linenomath}
\begin{equation}
\conditionalProb{\Truth{}}{\Data{}}
\propto{}
\conditionalLhood{\Data{}}{\Truth{}}
\cdot{}
\pi{}\left(\Truth{}\right),
\end{equation}
\end{linenomath}
where \conditionalLhood{\Data{}}{\Truth{}} is the
likelihood of \Data{} assuming \Truth{} and \TrasfMatrix{}, and $\pi{}$ is
the prior probability density for the true spectrum \Truth{}.
The selection and reconstruction efficiency, which is the probability
that an event produced in MC generator-level bin $t$ is reconstructed in
one of the $N_r$ bins included in \TrasfMatrix{}, is taken into
account in the likelihood.
An uninformative
prior probability density is chosen, such that equal probabilities
are assigned to all \Truth{} spectra within a wide range.
The background in each bin is taken into account when computing
\conditionalLhood{\Data{}}{\Truth{}}. The unfolded spectrum and its
associated uncertainty are extracted from the posterior probability
density distribution.
The migration matrix is obtained from the nominal
$t\bar{t}$ simulated sample using the top quarks
before their decay (parton level) or pseudotops (particle
level). The combination of the three decay channels is performed by
using a rectangular migration matrix, which maps the reconstructed
distribution of the three channels to the same corrected
distribution.
To validate the method, a linearity test is performed for the inclusive and differential measurements of the charge asymmetry. A given
asymmetry value is introduced by reweighting the samples according to a nonlinear function of
$\Delta|y|$ and $\Delta|\eta|$ based on a BSM axigluon model~\cite{AguilarSaavedra:2009es}. The asymmetry values are in the range of $-6\%$ to $6\%$ in steps of $2\%$. Good agreement between the unfolded values and the injected values is found, and the calibration curves
derived from this test are linear.
For the treatment of systematic uncertainties in the Bayesian inference approach,
the likelihood \conditionalLhood{\Data{}}{\Truth{}} is extended
with nuisance parameter terms. This marginal likelihood is defined as
\begin{linenomath}
\begin{equation}
\conditionalLhood{\Data{}}{\Truth{}} = \int
\conditionalLhood{\Data{}}{\Truth{},\vect{\theta}}\cdot \pi(\vect{\theta}) d\vect{\theta} \, ,
\end{equation}
\end{linenomath}
where $\vect{\theta}$ are the nuisance parameters, and $\pi(\vect{\theta})$ their prior probability densities,
which are assumed to be normal distributions $\mathcal{N}$ with a mean
value of zero and a variance of one.
A nuisance parameter is associated with each of the uncertainty sources.
As is described in Sec.~\ref{sec:systematics}, four categories of uncertainties are considered in this analysis,
but only two are included in the marginalization:
the normalizations of the background processes ($\vect{\theta}_b$),
and the uncertainties associated with the object identification,
reconstruction and calibration ($\vect{\theta}_s$).
While the first ones only affect the background predictions, the latter, referred to as
object systematic uncertainties, affect both the reconstructed distribution for the $t\bar{t}$ signal ($\vect{R}(\Truth{}; \vect{\theta}_s)$) and the total background prediction ($\vect{B}(\vect{\theta}_s, \vect{\theta}_b)$). The marginal likelihood then becomes
\begin{linenomath}
\begin{equation}
\conditionalLhood{\Data{}}{\Truth{}} = \int
\conditionalLhood{\Data{}}{\vect{R}(\Truth{}; \vect{\theta}_s),\vect{B}(\vect{\theta}_s, \vect{\theta}_b)} \cdot
\mathcal{N}(\vect{\theta}_s) \cdot \mathcal{N}(\vect{\theta}_b)\ d\vect{\theta}_s\ d\vect{\theta}_b \ .
\end{equation}
\end{linenomath}
\subsection{Binning optimization and asymmetry extraction} \label{sec:measurement}
For each measurement, the choice of binning for the $\Delta |y|$ and $\Delta|\eta|$ distributions is optimized
by minimizing the expected statistical uncertainty while allowing only a negligible bias in the linearity of the calibration curve.
The optimal binnings are found to be $4$ and $16$ bins in an interval between $-5$ and $5$
for the inclusive measurements of the $\Delta |y|$ and
$\Delta|\eta|$ distributions, respectively. For the differential
measurements, $4$ bins are used for the $\Delta |y|$ and
$\Delta|\eta|$ distributions for each of the chosen
$m_{t\bar{t}}$, $p_{\rm{T},t\bar{t}}$ and $\beta_{z,t\bar{t}}$ ranges.
Due to the limited size of the data sample, only two ranges of values are
considered for the $m_{t\bar{t}}$, $p_{\rm{T},t\bar{t}}$ and
$\beta_{z,t\bar{t}}$ variables.
The charge asymmetry predicted in the SM is expected to increase as a function of
$m_{t\bar{t}}$ while it is expected to be large for low
$p_{\rm{T},t\bar{t}}$ and small and roughly constant for higher $p_{\rm{T},t\bar{t}}$.
The exact boundary between the bins for $m_{t\bar{t}}$ was chosen to minimize
the expected uncertainties in the bins. For $p_{\rm{T},t\bar{t}}$, the boundary was
set at 30 \GeV as a compromise between the uncertainty optimization and
the interest in the $p_{\rm{T},t\bar{t}}$ dependence described above. For $\beta_{z,t\bar{t}}$,
the boundary at 0.6 is motivated by the large difference of the predicted
asymmetry between SM and BSM models in the range (0.6,1.0)~\cite{atlas8tevljets}.
Table~\ref{tab:binning} summarizes the differential bins used in the analysis.
\begin{table}[h]
\renewcommand\arraystretch{1.4}
\centering
\caption{Bins and ranges used for the inclusive and differential
measurements. The binning choices used in the $\Delta|\eta|$ and $\Delta|y|$
distributions are shown. The bins are symmetric around zero.}
\label{tab:binning}
\begin{tabular}{cccc}
\hline
\hline
& & $\Delta|\eta|$ & $\Delta|y|$ \\ \hline
\multicolumn{2}{c}{Inclusive} & $[0.0, 0.3, 0.6, 0.9, 1.2, 1.5, 1.7, 1.9, 5.0]$ & $[0.0,0.75,5.0]$ \\ \hline
\multirow{2}{*}{$m_{t\bar{t}}$} & 0--500 \GeV & $[0.0,0.8,5.0]$ & $[0.0,0.6,5.0]$ \\
& 500--2000 \GeV & $[0.0,1.4,5.0]$ & $[0.0,1.2,5.0]$ \\ \hline
\multirow{2}{*}{$\beta_{t\bar{t}}$} & 0--0.6 & $[0.0,0.8,5.0]$ & $[0.0,0.5,5.0]$ \\
& 0.6--1.0 & $[0.0,1.2,5.0]$ & $[0.0,0.9,5.0]$ \\ \hline
\multirow{2}{*}{$p^{t\bar{t}}_{\rm{T}}$} & 0--30 \GeV & $[0.0,0.7,5.0]$ & $[0.0,0.8,5.0]$ \\
& 30--1000 \GeV & $[0.0,0.7,5.0]$ & $[0.0,0.8,5.0]$ \\ \hline
\hline
\end{tabular}
\end{table}
For the optimized binning choice, more than $50 \%$ of the events
populate the diagonal bins of the migration matrix for the
$\Delta|y|$ distribution, and more than $97\%$ for $\Delta|\eta|$. The
rectangular migration matrix, normalized by row for each channel, used for the
inclusive $t\bar{t}$ asymmetry measurement is shown in
Fig.~\ref{fig:ttasym_responsemat}.
Due to the nonuniform shape of the $\Delta|y|$ distribution, the
matrix is not symmetric around the diagonal.
The migrations are symmetric around zero
and do not affect the asymmetry value. The $\Delta|y|$ and
$\Delta|\eta|$ input distributions used for the inclusive measurement
are shown in Fig.~\ref{fig:ttasym_inputdist}.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.95\columnwidth]{fig_03.pdf}
\end{center}
\caption{Rectangular migration matrix for the $\Delta |y|$
observable in the fiducial volume. The first four columns correspond to the $ee$ channel, followed by $\mu\mu$ and $e\mu$.
The numbers are normalized by row for each channel. }
\label{fig:ttasym_responsemat}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.95\columnwidth]{fig_04a.pdf}\\
\includegraphics[width=0.95\columnwidth]{fig_04b.pdf}
\end{center}
\caption{Input distributions for the inclusive $t\bar{t}$ (top) and leptonic (bottom) asymmetry measurements. The first 4 $\Delta|y|$ and 16 $\Delta|\eta|$ bins correspond to the $ee$ channel, followed by $\mu\mu$ and $e\mu$. The bin boundaries are symmetric around zero and are defined as $[0.0,0.75,5.0]$ and $[0.0, 0.3, 0.6, 0.9, 1.2, 1.5, 1.7, 1.9, 5.0]$ for $\Delta|y|$ and $\Delta|\eta|$, respectively. The data/expected ratio is also shown. }
\label{fig:ttasym_inputdist}
\end{figure}
The asymmetry values are extracted by taking the mean of the posterior probability density obtained during the unfolding procedure. The uncertainty is obtained from the standard deviation of the posterior probability density.
\section{Systematic uncertainties}
\label{sec:systematics}
Four classes of systematic uncertainties affect the measurement of the
charge asymmetry: detector modeling uncertainties, uncertainties
related to the estimation of the backgrounds, signal modeling
uncertainties, and other uncertainties, which involve the top-quark
reconstruction, the bias introduced by the unfolding procedure and the
MC statistical uncertainty.
The first two categories are estimated within the unfolding
through the marginalization procedure where the total uncertainty
includes these systematic uncertainties together with the
statistical uncertainty.
In order to estimate the impact of each source of systematic uncertainty, pseudodata corresponding to the sum of the nominal signal and background samples is used. The unfolding procedure with marginalization is applied to the pseudodata and constraints on the systematic uncertainties are obtained. These constraints are then used to build the $\pm 1\sigma$ variations of the prediction. The varied pseudodata are then unfolded without marginalization. The impact of each systematic uncertainty is computed by taking half of the difference between the results obtained from the $\pm 1\sigma$ variations of pseudodata. Clearly, this is only
an approximate estimate of the individual contribution of
each source of systematic uncertainty within the overall marginalization procedure.
The signal modeling uncertainties are not estimated through the marginalization
procedure. For these uncertainties, the migration matrix is fixed to
the nominal $t\bar{t}$ sample and
distributions obtained with different generators and different
injected asymmetries are unfolded. The unfolded asymmetries are
compared with the injected asymmetries and the calibration curves are
obtained.
The slopes and offsets of the calibration curves are extrapolated to
the measured value in data.
The final category of systematic uncertainties involves different estimation
methods. The uncertainty related to the top-quark reconstruction is
estimated on pseudodata by varying the starting point of the smearing
procedure within the kinematic reconstruction and repeating the
unfolding. The bias introduced by the unfolding procedure is estimated
by propagating the
residual slope and offset of the nominal calibration curve to
the measured value. The MC statistical uncertainty is estimated by
varying the nominal migration matrix within the MC statistical
uncertainty and the unfolding procedure is repeated for each
variation.
All sources of systematic uncertainties are discussed below in detail.
\subsection{Detector modeling uncertainties}
\subsubsection*{Lepton-related uncertainties}
The reconstruction and identification efficiencies of electrons and muons, as well as the efficiency of the
triggers used to record the events, differ between data and simulation. Scale factors, and their uncertainties, are derived using
tag-and-probe techniques on $Z \rightarrow \ell^{+} \ell^{-} (\ell = e,\,\mu)$ in data and in simulated samples to correct the simulation for these differences~\cite{Aad:2012xs,Aad:2014fxa,Aad:2014rra,Aad:2014nim}. Moreover, the accuracy of the
lepton momentum scale and resolution in simulation is also checked using
reconstructed distributions of the $Z \rightarrow \ell^{+} \ell^{-}$
and $J/\psi \rightarrow \ell^{+} \ell^{-} $ masses. In the case of
electrons, $E/p$ studies using $W \rightarrow e\nu$ events are also
used. Small differences are observed between data and
simulation. Corrections for the lepton energy scale and resolution,
and their related uncertainties are
considered~\cite{,Aad:2014fxa,Aad:2014rra,Aad:2014nim}.
The uncertainties are propagated through this analysis and represent a minor source of uncertainty in the measurements.
\subsubsection*{Jet-related uncertainties}
The jet energy scale and its uncertainty are derived combining information from test-beam
data, LHC collision data, and simulation~\cite{Aad:2014bia}. The jet energy scale uncertainty is split
into 22 uncorrelated sources that have different jet $p_{\rm{T}}$ and $\eta$ dependencies
and are treated independently in this analysis.
The total jet energy scale uncertainty is one of the dominant
uncertainties in $\Actt$ and in the differential measurements of $\Acll$. The jet reconstruction efficiency is found to be about $0.2\%$ lower in simulation than in data for jets
below $30$~\GeV~and consistent with data for higher jet $p_T$. All jet-related
kinematic variables (including the missing transverse momentum) are recomputed by removing randomly $0.2\%$ of the jets with
$p_{\rm{T}}$ below $30$~\GeV~and the event selection is repeated. The efficiency for each jet to satisfy the JVF requirement is measured in $Z \rightarrow \ell^{+} \ell^{-}+\textrm{1-jet}$ events
in data and simulation~\cite{Aad:2015ina}. The corresponding uncertainty is evaluated in the analysis by changing the nominal JVF cut value and
repeating the analysis using the modified cut value. The uncertainty related to the jet energy resolution is estimated
by smearing the energy of jets in simulation by the difference between the jet energy resolutions for data and simulation~\cite{Aad:2012ag}. Finally, the efficiencies to tag jets from $b$- and $c$-quarks, light quarks, and gluons in simulation are
corrected by $p_T$- and $\eta$-dependent data/MC scale factors~\cite{Aad:2015ydr,ATLAS-CONF-2014-004,ATLAS-CONF-2014-046,}.
The uncertainties in these scale factors are propagated to the measured value. The impact on the measurement of the jet reconstruction efficiency, jet vertex
fraction, jet resolution, and jet tagging efficiency is minor.
\subsubsection*{Missing transverse momentum}
The systematic uncertainties associated with the momenta and energies
of reconstructed objects (leptons and jets) are also propagated to the $\met$ calculation. The $\met$ reconstruction
also receives contributions from the presence of low-$p_{\rm{T}}$ jets and calorimeter cells not included in reconstructed objects (``soft terms''). The systematic uncertainty of the soft terms is evaluated using
$Z \rightarrow \mu^{+} \mu^{-}$ events using methods similar to those used in Ref.~\cite{Aad:2012re}.
The uncertainty has a negligible effect on the measured asymmetries.
\subsection{Background-related uncertainties}
\label{sec:syst_background}
The uncertainties in the single-top-quark and diboson backgrounds are about 7\% and 5\%, respectively.
These correspond to the uncertainties in the theoretical cross sections used for the normalization of the MC simulated samples.
The uncertainty in the normalization of the fake-lepton background is
evaluated by using various Monte Carlo simulations for each process
contributing to this background and propagating the change into the number
of expected events in the signal region.
In the $\mu\mu$ channel, the uncertainty is obtained by comparing a
purely data-driven method based on the measurement of the
efficiencies for real and fake loose leptons,
and the estimation used in this analysis.
Following a Bayesian procedure assuming constant
a priori probability for a non-negative number of events, the
resulting total relative uncertainties are
$^{+193\%}_{-47\%}$ in the $ee$, $^{+80\%}_{-53\%}$ in the $\mu\mu$, and $^{+49\%}_{-45\%}$ in the
$e\mu$ channel, where the uncertainties correspond to the $68$\% central probability region.
In the case of the Drell--Yan events, the detector modeling
systematic uncertainties described previously are propagated to the scale factors
derived in the control region by recalculating them for all the systematic uncertainty variations.
An additional uncertainty of 6\% is estimated by varying the $Z$ mass window of the control
region used to obtain the scale factors and is added in quadrature to
obtain the final uncertainty in these scale factors.
This category represents a minor source of uncertainty in the measurement.
\subsection{Signal modeling uncertainties}
The uncertainty due to the choice of MC generator is obtained by taking the full difference between the \powheg
and \mcatnlo predictions, both interfaced with \herwig, while the uncertainty from parton showering and hadronization is obtained by
comparing \powheg interfaced with either \pythia or \herwig.
These components are among the dominant uncertainties.
The effect produced by the different amount of ISR and FSR in the
events is estimated as half the difference between the asymmetries
obtained from MC samples with more or less ISR/FSR.
These samples are generated with \powheg interfaced with \pythia
for which the parameters of the generation were varied to span the
ranges compatible with the results
of measurements of $t\bar{t}$ production in association with
jets~\cite{TOPQ-2011-21}.
Finally, PDF uncertainties are obtained by
using the error sets of CT10,
MWST2008 and NNPDF2.3,
and following the prescriptions recommended by the PDF4LHC working
group~\cite{Botje:2011sn}. The impact of the last two uncertainties is small.
\subsection{Other uncertainties}
\subsubsection*{Top-quark kinematic reconstruction}
There is an intrinsic uncertainty of the reconstruction method due to
the randomness in the smearing procedure.
If the smearing starts from a different point it
could lead to a different solution. The uncertainty from this effect
is computed by performing pseudoexperiments on MC events. For each
event, the $t\bar{t}$ system is reconstructed multiple times varying
the starting point of the smearing procedure. Then, for each variation
the unfolding procedure is repeated and the standard deviation of the
asymmetries obtained is taken as the uncertainty. This represents one
of the major systematic uncertainties for the measurements, but it is
still only half of the statistical uncertainty for most of them.
\subsubsection*{Nonclosure uncertainties}
When the calibration curve for the nominal signal \powheg sample is estimated a residual slope and a nonzero offset are observed. This bias, introduced by the unfolding procedure, is propagated to the measured values in the same way as for the signal modeling uncertainties. This source of uncertainty is negligible in all the measurements.
\subsubsection*{MC sample size}
The uncertainty associated with the limited size of the nominal signal \powheg sample is evaluated by performing pseudoexperiments on MC events. The migration matrix is varied within the MC statistical uncertainty and the unfolding procedure is repeated. The standard deviation of the obtained asymmetries is taken as the uncertainty. This uncertainty has a minor impact on the measurements.
\subsection{Summary of systematic uncertainties}
Tables \ref{tab:uncertainties_lepton} and \ref{tab:uncertainties_ttbar} show how each category of uncertainty affects the measurements of the lepton and \ttbar\ asymmetry, respectively. The statistical uncertainty gives the largest contribution to the measurement, followed by the reconstruction and the signal modeling uncertainties. The signal modeling uncertainties are enhanced in the differential measurements by the migrations between the differential bins across the different MC generators used for their estimation. The uncertainty obtained by the sum in quadrature of the individual systematic uncertainties is slightly larger than the total marginalized uncertainty in the measurements.
\begin{table}[h]
\renewcommand\arraystretch{1.4}
\centering
\caption{Absolute uncertainties from the different sources affecting the leptonic asymmetry of the three channels combined in the fiducial and full phase space.}
\label{tab:uncertainties_lepton}
\resizebox{\linewidth}{!}{\begin{tabular}{cccccccp{0.5cm}ccccc}
\hline
\hline
& & \multicolumn{11}{c}{Absolute uncertainties in $A_{\rm{C}}^{\ell \ell}$ } \\ \hline
& & \multicolumn{5}{c}{Fiducial volume} & & \multicolumn{5}{c}{Full phase space} \\ \hline
& & Statistics & Detector & Bkg & Signal modeling & Other & & Statistics & Detector & Bkg & Signal modeling & Other \\ \hline \hline
\multicolumn{2}{c}{Inclusive} & $0.005$ & $0.001$ & $0.001$ & $0.002$ & $0.001$ & & $0.005$ & $0.001$ & $0.001$ & $0.004$ & $0.001$ \\ \hline
\multirow{2}{*}{$m_{t\bar{t}}$} & 0--500 \GeV & $0.008$ & $0.002$ & $0.001$ & $0.005$ & $0.005$& & $0.008$ & $0.002$ & $0.001$ & $0.005$ & $0.006$ \\
& 500--2000 \GeV & $0.012$ & $0.004$ & $<0.001$ & $0.013$ & $0.005$ && $0.011$ & $0.004$ & $<0.001$ & $0.014$ & $0.005$ \\ \hline
\multirow{2}{*}{$\beta_{t\bar{t}}$} & 0--0.6 & $0.007$ & $0.003$ & $<0.001$ & $0.004$ & $0.004$ & & $0.007$ & $0.002$ & $<0.001$ & $0.005$ & $0.005$ \\
& 0.6--1.0 & $0.010$ & $0.005$ & $0.001$ & $0.005$ & $0.004$ & & $0.010$ & $0.003$ & $0.001$ & $0.006$ & $0.004$ \\ \hline
\multirow{2}{*}{$p^{t\bar{t}}_{\rm{T}}$} & 0--30 \GeV & $0.015$ & $0.009$ & $0.001$ & $0.015$ & $0.006$ & & $0.015$ & $0.010$ & $0.001$ & $0.017$ & $0.007$ \\
& 30--1000 \GeV & $0.011$ & $0.004$ & $0.001$ & $0.012$ & $0.005$& & $0.010$ & $0.004$ & $0.001$ & $0.013$ & $0.006$ \\ \hline
\hline
\end{tabular}}
\end{table}
\begin{table}[h]
\renewcommand\arraystretch{1.4}
\centering
\caption{Absolute uncertainties from the different sources affecting the $t\bar{t}$ asymmetry of the three channels combined in the fiducial and full phase space.}
\label{tab:uncertainties_ttbar}
\resizebox{\linewidth}{!}{\begin{tabular}{cccccccp{0.5cm}ccccc}
\hline
\hline
& & \multicolumn{10}{c}{Absolute uncertainties in $A_{\rm{C}}^{t\bar{t}}$ } \\ \hline
& & \multicolumn{5}{c}{Fiducial volume} & & \multicolumn{5}{c}{Full phase space} \\ \hline
& & Statistics & Detector & Bkg & Signal modeling & Other & & Statistics & Detector & Bkg & Signal modeling & Other \\ \hline \hline
\multicolumn{2}{c}{Inclusive} & $0.013$ & $0.008$ & $<0.001$ & $0.007$ & $0.007$ & &$0.011$ & $0.006$ & $<0.001$ & $0.008$ & $0.006$ \\ \hline
\multirow{2}{*}{$m_{t\bar{t}}$} & 0--500 \GeV & $0.030$ & $0.024$ & $0.001$ & $0.016$ & $0.021$ & &$0.028$& $0.021$ & $0.002$ & $0.018$ & $0.020$ \\
& 500--2000 \GeV & $0.018$ & $0.007$ & $<0.001$ & $0.015$ & $0.009$ & &$0.015$ & $0.006$ & $<0.001$ & $0.016$ & $0.008$ \\ \hline
\multirow{2}{*}{$\beta_{t\bar{t}}$} & 0--0.6 & $0.023$ & $0.021$ & $0.002$ & $0.014$ & $0.018$ & & $0.023$ & $0.019$ & $0.002$ & $0.015$ & $0.017$ \\
& 0.6--1.0 & $0.021$ & $0.009$ & $0.001$ & $0.013$ & $0.011$ & &$0.018$ & $0.009$ & $0.001$ & $0.013$ & $0.010$ \\ \hline
\multirow{2}{*}{$p^{t\bar{t}}_{\rm{T}}$} & 0--30 \GeV & $0.035$ & $0.019$ & $0.003$ & $0.018$ & $0.020$ & & $0.031$ & $0.015$ & $0.004$ & $0.019$ & $0.017$ \\
& 30--1000 \GeV & $0.027$ & $0.015$ & $0.003$ & $0.018$ & $0.017$& &$0.025$ & $0.013$ & $0.003$ & $0.014$ & $0.015$ \\ \hline
\hline
\end{tabular}}
\end{table}
\section{Results}
\label{sec:result}
Figures~\ref{fig:sumary_plot_lepton} and~\ref{fig:sumary_plot_ttbar}
show the inclusive and differential results for the leptonic and
$t\bar{t}$ charge asymmetry in the fiducial region and in the full
phase space. All the results are compatible with the Standard Model
predictions~\cite{Bernreuther:2012sx,Alioli:2010xd,Frixione:2007vw,Nason:2004rx}.
Figure~\ref{fig:unfolded_plots} shows the unfolded
distributions of the $\Delta|\eta|$ and $\Delta|y|$ observables for the
inclusive measurement in the fiducial volume.
The distributions are compared with Monte
Carlo predictions at NLO provided by \powheg. The measured inclusive
values in the full phase space are $\Acll = \Acllcomb$ and $\Ac
= \Accomb$.
They are in agreement with
the Standard Model predictions \mbox{$\Acll = \Aclltheory$}
and \mbox{$\Actt = \Actttheory$}~\cite{Bernreuther:2012sx}.
The measurements are consistent with other LHC asymmetry
measurements at 8 \TeV~\cite{cms8tevljets,
atlas8tevljets,Khachatryan:2015mna}.
The statistical uncertainty is in
most cases the dominant contribution to the total uncertainty.
The dominant systematic uncertainties across all the measurements are the signal modeling and the kinematic reconstruction uncertainty.
The signal modeling uncertainties are reduced in most of the cases by performing the measurements in the fiducial region,
since the extrapolation from detector acceptance to the full phase space is avoided.
The statistical uncertainty is slightly larger in the fiducial region than in the full phase space;
this is expected because some reconstructed events fail the fiducial requirements in the fiducial analysis.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.70\columnwidth]{fig_05a.pdf}\\
\includegraphics[width=0.70\columnwidth]{fig_05b.pdf}
\caption{Summary of all the measurements in this paper for the leptonic
asymmetry in the fiducial volume (top) and full phase space
(bottom). The predictions shown in blue are obtained using \powheg
+ \pythia at NLO where the uncertainties are statistical, and the
corresponding theoretical uncertainties are
small compared to the experimental precision. The inclusive
measurement in the full phase space is compared to a NLO + EW
prediction ~\cite{Bernreuther:2012sx}.}
\label{fig:sumary_plot_lepton}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.70\columnwidth]{fig_06a.pdf}
\includegraphics[width=0.70\columnwidth]{fig_06b.pdf}
\caption{Summary of all the measurements in this paper for the $\ttbar$
asymmetry in the fiducial volume (top) and full phase space
(bottom). The predictions shown in blue are obtained using \powheg
+ \pythia at NLO where the uncertainties are statistical, and the
corresponding theoretical uncertainties are
small compared to the experimental precision. The inclusive
measurement in the full phase space is compared to a NLO + EW
prediction ~\cite{Bernreuther:2012sx}.}
\label{fig:sumary_plot_ttbar}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.45\columnwidth]{fig_07a.pdf}
\includegraphics[width=0.45\columnwidth]{fig_07b.pdf}
\caption{Data distribution after the unfolding procedure compared with the \powheg + \pythia prediction at
NLO for the inclusive $\Delta|\eta|$ (left) and $\Delta|y|$ (right) observables in the fiducial volume. The data/expected ratio is also shown.}
\label{fig:unfolded_plots}
\end{center}
\end{figure}
Figure~\ref{fig:2d_map_dy} compares the values of
$A_{\rm{C}}^{\ell \ell}$ and $A_{\rm{C}}^{t\bar{t}}$ from the
inclusive measurements in the full phase space
to the SM predictions and two BSM models~\cite{Aguilar-Saavedra:2014nja}
compatible with the Tevatron results.
Two BSM models with a new color-octet particle that is exchanged in the
$s$-channel are considered. In the model with the light octet, the
new particle's mass ($m = 250$ \GeV) is below the $t\bar{t}$
production threshold and its width is assumed to be $\Gamma = 0.2m$. The model with the heavy octet uses an octet
mass beyond current limits from direct searches at the LHC. The
corrections to $t\bar{t}$ production are independent of
the mass but instead depend on the ratio of coupling to mass, which is
assumed to be $1 \, \rm{\TeV}^{-1}$. The new particles in both BSM models would
not be visible as resonances in the $m_{t\bar{t}}$ spectrum at the
Tevatron or at the LHC.
In the figures, model predictions for different left-handed, right-handed, and axial coupling constants
to top quarks are shown. The ellipses correspond to
the $1 \sigma$ and $2 \sigma$ total uncertainty in the measurements. The correlation between these
two measurements is taken into account. The statistical and detector
systematic uncertainty correlation between $A_{\rm{C}}^{\ell \ell}$ and $A_{\rm{C}}^{t\bar{t}}$
is found to be $30 \%$. The modeling systematic uncertainties are assumed
to be $100\%$ correlated. The resulting correlation between
$A_{\rm{C}}^{\ell \ell}$ and $A_{\rm{C}}^{t\bar{t}}$ is about $48\%$.
The measurements are compatible with the SM and do not exclude the two sets of BSM models considered.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.45\columnwidth]{fig_08a.pdf}
\includegraphics[width=0.45\columnwidth]{fig_08b.pdf}
\caption{Comparison of the inclusive $A_{\rm{C}}^{\ell \ell}$ and
$A_{\rm{C}}^{t\bar{t}}$ measurement values in the full phase space to
the SM NLO QCD+EW prediction~\cite{Bernreuther:2012sx} and to two benchmark
BSM models~\cite{Aguilar-Saavedra:2014nja}, one with a light octet with mass below the $t\bar{t}$ production threshold (left) and one with
a heavy octet with mass beyond the reach of the LHC (right), for various couplings as described in
the legend. Ellipses corresponding to $1\sigma$ and $2\sigma$ combined statistical and
systematic uncertainties of the measurement, including the correlation between $A_{\rm{C}}^{\ell \ell}$ and $A_{\rm{C}}^{t\bar{t}}$, are also shown. }
\label{fig:2d_map_dy}
\end{center}
\end{figure}
\FloatBarrier
\section{Conclusion}
\label{sec:conclusions}
Measurements of the leptonic and \ttbar{} charge asymmetry in the
dilepton channel, characterized
by two high-$p_{\rm T}$ leptons (electrons or muons), are presented. The measurements, corrected for
detector resolution and acceptance effects, are performed
using data corresponding to an integrated luminosity of \lumi of \pp
collisions at $\sqrt{s} = 8$~\tev\ collected by the \atlas
detector at the LHC. The inclusive asymmetries are measured in the full
phase space to be:
\begin{eqnarray*}
\Acll & = & \Acllcomb~\textrm{and}
\end{eqnarray*}
\begin{eqnarray*}
\Ac & = & \Accomb.
\end{eqnarray*}
They are in agreement with the Standard Model predictions $\Acll
= \Aclltheory$ and $\Ac = \Actttheory$.
Differential measurements of the asymmetries as a function
of the invariant mass, transverse momentum, and longitudinal boost of
the $t\bar{t}$ system are also performed and they are found to be in
agreement with the SM predictions, although they have relatively large
uncertainties.
All measurements are also performed in a fiducial region at particle
level where the modeling uncertainties are reduced.
For all measurements, the statistical uncertainty is the
dominant contribution to the total uncertainty. The unfolded distributions of
lepton \deta and \ttbar{} \dy are provided. Good agreement between
the corrected distributions and the predictions of \powheg + \pythia is
observed.
\section*{Acknowledgments}
We thank CERN for the very successful operation of the LHC, as well as the
support staff from our institutions without whom ATLAS could not be
operated efficiently.
We acknowledge the support of ANPCyT, Argentina; YerPhI, Armenia; ARC, Australia; BMWFW and FWF, Austria; ANAS, Azerbaijan; SSTC, Belarus; CNPq and FAPESP, Brazil; NSERC, NRC and CFI, Canada; CERN; CONICYT, Chile; CAS, MOST and NSFC, China; COLCIENCIAS, Colombia; MSMT CR, MPO CR and VSC CR, Czech Republic; DNRF and DNSRC, Denmark; IN2P3-CNRS, CEA-DSM/IRFU, France; GNSF, Georgia; BMBF, HGF, and MPG, Germany; GSRT, Greece; RGC, Hong Kong SAR, China; ISF, I-CORE and Benoziyo Center, Israel; INFN, Italy; MEXT and JSPS, Japan; CNRST, Morocco; FOM and NWO, Netherlands; RCN, Norway; MNiSW and NCN, Poland; FCT, Portugal; MNE/IFA, Romania; MES of Russia and NRC KI, Russian Federation; JINR; MESTD, Serbia; MSSR, Slovakia; ARRS and MIZ\v{S}, Slovenia; DST/NRF, South Africa; MINECO, Spain; SRC and Wallenberg Foundation, Sweden; SERI, SNSF and Cantons of Bern and Geneva, Switzerland; MOST, Taiwan; TAEK, Turkey; STFC, United Kingdom; DOE and NSF, United States of America. In addition, individual groups and members have received support from BCKDF, the Canada Council, CANARIE, CRC, Compute Canada, FQRNT, and the Ontario Innovation Trust, Canada; EPLANET, ERC, FP7, Horizon 2020 and Marie Sk{\l}odowska-Curie Actions, European Union; Investissements d'Avenir Labex and Idex, ANR, R{\'e}gion Auvergne and Fondation Partager le Savoir, France; DFG and AvH Foundation, Germany; Herakleitos, Thales and Aristeia programmes co-financed by EU-ESF and the Greek NSRF; BSF, GIF and Minerva, Israel; BRF, Norway; Generalitat de Catalunya, Generalitat Valenciana, Spain; the Royal Society and Leverhulme Trust, United Kingdom.
The crucial computing support from all WLCG partners is acknowledged gratefully, in particular from CERN, the ATLAS Tier-1 facilities at TRIUMF (Canada), NDGF (Denmark, Norway, Sweden), CC-IN2P3 (France), KIT/GridKA (Germany), INFN-CNAF (Italy), NL-T1 (Netherlands), PIC (Spain), ASGC (Taiwan), RAL (UK) and BNL (USA), the Tier-2 facilities worldwide and large non-WLCG resource providers. Major contributors of computing resources are listed in Ref.~\cite{ATL-GEN-PUB-2016-002}.
\printbibliography
\newpage \input{atlas_authlist}
\end{document}
|
\section{Introduction}
Various cosmological observations support the standard cosmological model: inflation+$\Lambda$CDM model \cite{cmb-review}, including the temperature and polarization anisotropies of the cosmic microwave background (CMB) radiation, the distribution of large-scale structure, Type Ia supernovae, the baryon acoustic oscillation and the cosmic weak lensing. This successful model is based on the following assumptions: (1) The universe is homogeneous and isotropic in large scales (2) Gravity is correctly described by general relativity in all macroscopic scales. (3) Random and nearly Gaussian distributed cosmic anisotropies originated from the quantum fluctuations in the early inflationary stage.
However, the increasing release of precise data, in particular the Wilkinson Microwave Anisotropy Probe (WMAP) and the Planck satellites measurements of the temperature anisotropies of the CMB \cite{wmap,planck}, led to the claim of a number of anomalies in the CMB largest scales: the low quadrupole problem \cite{cobe}, the lack of both variance and correlation on the largest angular scales \cite{lack}, cold spot problem \cite{cold-spot}, power asymmetry \cite{power}, hemisphere asymmetry \cite{hemisphere}, large-scale quadrant asymmetry \cite{quadrant}, alignment of low multipoles \cite{alignment0,alignment,alignment2}, parity asymmetry \cite{parity1}, mirror-parity asymmetry, and so on (see \cite{1001.4613} as the review of WMAP results, \cite{1303.5083,1506.07135} for a review on Planck results, and \cite{review} for a recent review).
Based on these observations, the Planck collaboration stated that ``\emph{The Universe is still weird and interesting}". The origin of these anomalies are still not well understood, and it could hint to the physics of the earliest stage of the universe, preceding the big bang. Other more prosaic explanations are also possible, including foreground microwave emissions from objects that are not yet known and have not been predicted. If the anomalies have a cosmological origin, violating the cosmological principle, the standard model of cosmology should be revised. On the contrary, if they have non-cosmological origins (possible foreground residuals or unsolved systematics), the non-cosmological artifact should be well studied and removed from the data to avoid misleading physical explanations of the universe. So, for any of the above cases, these large-scale anomalies deserve to be well studied.
Among all the anomalies, several ones are direction dependent. For instance, the alignment of the CMB low multipoles, for example, from $\ell=2$ to $\ell=5$ seem to be pointing to a common direction. If the explanation for this anomaly is cosmological, it indicates that there is a preferred direction in our universe, which is a significant violation of the cosmological principle.
Another direction dependent problem is related to the CMB parity asymmetry. This problem has been investigated in the literature \cite{parity1,kim2011} and shows a significant dominance on the CMB power spectrum stored in the odd multipoles over the even ones. This odd parity preference was also confirmed in the recent Planck data \cite{planck2013,planck2015}. By defining various directional statistics, in previous works \cite{zhao2012,zhao2014,zhao2015}, we investigated the directional properties of the CMB parity asymmetry, and found that CMB parity violation favors a preferred direction, which is independent of the choice of the statistics. In this paper, we shall review the directional properties of the CMB parity asymmetry, and search for the preferred directions stored in CMB low multipoles. In particular, we compare this preferred direction with the announced preferred directions in CMB quadrupole and octopole, and find that these directions have strong correlations. The alignment between them is also confirmed at more than $3\sigma$ confidence level. This result indicates that the CMB parity asymmetry should have a common explanation with other anomalies including the alignment problem of CMB low multipoles, the low quadrupole problem, and the lack of large-scale correlation. Most importantly, we find that all these preferred directions are coincident with the direction of CMB kinematic dipole. In addition, the preferred directions were also reported in a number of other cosmological observations: the velocity flows \cite{velocity}, quasar alignment \cite{quasar}, anisotropy of the cosmic acceleration \cite{acceleration,acc2}, the handedness of spiral galaxies \cite{spiral}, and angular distribution of the fine-structure constant \cite{fine}. Even though there are many debates \cite{debate1,debate2,debate3,debate4,debate5}, it was also reported that all these preferred directions seem to coincide with the CMB kinematic dipole.
It is well known that the CMB kinematic dipole is caused by the motion of our local group of galaxies (including the Milky Galaxies) relative to the reference frame of CMB in the direction of the Galactic coordinate ($\theta=42^{\circ}$, $\phi=264^{\circ}$), which is a pure kinematic effect. So, if the preferred direction of any claimed CMB anomaly coincides with the CMB dipole direction, it should have a non-cosmological origin. The explanation should consider the possible CMB dipole-related foreground residuals or systematical errors that should be seriously handled in the future measurements, and which could hint to some unsolved contaminations in the large-scale observations, including CMB, galaxies distribution, Type Ia supernovae, quasar distribution, and so on.
The rest of this article will be structured as follows. In Sec. 2, we briefly summarize the anomaly on CMB parity asymmetry. In Sec. 3, we focus on the directional properties and the preferred directions of CMB parity asymmetry. In Sec. 4, we compare this preferred direction with the preferred directions stored in the CMB dipole, quadrupole and octopole. In Sec. 5, we briefly introduce the other large-scale anomalies and their direction dependence. In Sec. 6, we list some possible explanations for the cosmological anomalies mentioned above. Sec. 7 is contributed as a summary and conclusion of this paper.
\section{CMB parity asymmetry}
The CMB temperature fluctuation on a two-dimensional sphere is a scalar field. According to the coordinate transformation, it can be decomposed as the standard spherical harmonics as follows,
\begin{equation}
\Delta T(\hat{n})=\sum_{\ell=0}^{\infty} \sum_{m=-\ell}^{\ell} a_{\ell m} Y_{\ell m} (\hat{n}),
\end{equation}
where $Y_{\ell m}(\hat{n})$ are the spherical harmonics, and $a_{\ell m}$ are the corresponding coefficients. In the standard inflationary scenario, both primordial scalar and tensor perturbations are random Gaussian fields. In the linear order approximation, the two-dimensional temperature fluctuations also satisfy the random Gaussian distribution, i.e., the amplitudes $|a_{\ell m}|$ are distributed according to Rayleigh's probability distribution function, and the phase of $a_{\ell m}$ with $m\neq 0$ is supported to be evenly distributed in the range $[0,2\pi]$. For the random Gaussian field, the statistical properties are completely described by the second-order power spectrum, namely
\begin{equation}
\langle a_{\ell m}a^*_{\ell' m'}\rangle=C_{\ell}\delta_{\ell\ell'}\delta_{mm'},
\end{equation}
where $\langle ... \rangle$ denotes the average over the statistical ensemble of realizations, and the spectrum $C_{\ell}$ is independent of the magnetic quantum number $m$ for the statistical isotropic field. In the real measurements, it is impossible to directly observe the power spectrum $C_{\ell}$ itself. One has to construct the estimators. For the full-sky map, if the noise is negligible, the best unbiased estimator for $C_{\ell}$ is \cite{grishchuk1997}
\begin{equation}\label{hat-cl}
\hat{C}_{\ell}=\frac{1}{2\ell+1} \sum_{m=-\ell}^{\ell} a_{\ell m} a_{\ell m}^*,
\end{equation}
and the statistical uncertainty is $\Delta \hat{C}_{\ell}=\sqrt{\frac{2}{2\ell+1}}C_{\ell}$, which is the so-called cosmic variance.
According to the symmetry under the transformation of coordinate inversion $\hat{n}\rightarrow -\hat{n}$, CMB anisotropy field can be decomposed as the symmetric component $\Delta T^{+}$ and the antisymmetric component $\Delta T^{-}$ as follows,
\begin{eqnarray}
\Delta T^{+}(\hat{n})=\frac{\Delta T(\hat{n})+\Delta T(-\hat{n})}{2}, ~~
\Delta T^{-}(\hat{n})=\frac{\Delta T(\hat{n})-\Delta T(-\hat{n})}{2}.
\end{eqnarray}
The patterns $\Delta T^{+}(\hat{n})$ and $\Delta T^{-}(\hat{n})$ have even and odd-parity respectively, which can also be written as
\begin{eqnarray}
\Delta T^{+}(\hat{n})=\sum_{\ell m} a_{\ell m} Y_{\ell m}(\hat{n})\Gamma_{\ell}^{+}, ~~
\Delta T^{-}(\hat{n})=\sum_{\ell m} a_{\ell m} Y_{\ell m}(\hat{n})\Gamma_{\ell}^{-},
\end{eqnarray}
where $\Gamma_{\ell}^{+}=\cos^2\left(\frac{\ell\pi}{2}\right)$ and $\Gamma_{\ell}^{-}=\sin^2\left(\frac{\ell\pi}{2}\right)$.
Therefore, significant power asymmetry between even and odd multipoles may be interpreted as a preference for a particular parity of the anisotropy pattern. Fig. \ref{fig1} (left panel) presents the observed low multipole data by WMAP satellite. Comparing with the theoretical predictions, we find that in the lowest multipole range, the odd multipoles are systematically larger than expected, while the even multipoles are systematically smaller than the theoretical ones. This regular pattern significantly violates the random and Gaussian assumption of the CMB field, and strongly indicates the odd-multipole preference, i.e. the antisymmetric preference of the CMB anisotropy.
\begin{figure}[t]
\begin{center}
\includegraphics[width=14cm]{figs/f1.pdf}
\end{center}
\caption{Left panel: The predicted CMB low multipole power spectrum in the $\Lambda$CDM model, and 3-year, 5-year and 7-year WMAP observed data; Right panel: Theoretical values of $P^+/P^-$ compares with the WMAP observed results \cite{1202.0728}.}\label{fig1}
\end{figure}
In order to quantify this asymmetry, we introduce the following statistic:
\begin{eqnarray}
P^{+}=\sum_{\ell=2}^{\ell_{\max}}\frac{\ell(\ell+1)}{2\pi}C_{\ell}\Gamma_{\ell}^{+},
~~P^{-}=\sum_{\ell=2}^{\ell_{\max}}\frac{\ell(\ell+1)}{2\pi}C_{\ell}\Gamma_{\ell}^{-}.
\end{eqnarray}
$P^{+}$ and $P^{-}$ are the sum of the power spectrum for even and odd multipoles, respectively. Therefore, the ratio $P^{+}/P^{-}$ is associated with the degree of the parity asymmetry, where the lower value of $P^{+}/P^{-}$ indicates the odd-parity preference, and vice-versa. Fig. \ref{fig1} (right panel) shows the ratio derived from the WMAP observed data. Comparing with the theoretical values, we find that if the maximum multipole $\ell_{\max}$ is small, i.e., in the low multipole range, the observed results significantly deviate from the model predictions, showing an obvious odd-multipole preference. This is the so-called CMB parity asymmetry anomaly.
In order to quantify the statistical significance of this anomaly, as in general, we can simulate a large number of random Gaussian samples based on the best-fit $\Lambda$CDM model. For each sample, we calculate the ratio $P^{+}/P^{-}$, and count the probability (i.e. the $p$-value) to get the ratio $P^{+}/P^{-}$, which is smaller than the observed value for each $\ell_{\max}$. In Fig. \ref{fig2} (left panel), we plot the $p$-values for various $\ell_{\max}$, from which we find that the $p$-value curves minimize at $\ell_{\max}\sim (20,30)$, and the minimal $p$-value is less than $1\%$ in the multipole range. Recently, Planck collaboration repeated this calculation using the new released data. The corresponding curves for the $p$-values as function of $\ell_{\max}$ are shown in Fig. \ref{fig2} (right panel), from which we find the results derived from Planck data are quite close to those from the WMAP data, and all of them indicate a significant odd-parity preference in the CMB low multipoles, and the minimal $p$-value is much smaller than $1\%$ for all the Planck released data. So, we conclude that the CMB low-multipole data indicate a significant violation of parity symmetry, which conflicts with the prediction of the random Gaussian assumption of the standard cosmological model.
\begin{figure}[t]
\begin{center}
\includegraphics[width=14cm]{figs/f2.pdf}
\end{center}
\caption{Left panel: Probability of getting $P^+/P^-$ as low as WMAP data for multipole range $2\le\ell\le\ell_{\max}$ \cite{1202.0728}. Right panel: Probability of getting $P^+/P^-$ as low as Planck Commander (red), NILC (orange), SEVEM (green), SMICA (blue) data for multipole range $2\le\ell\le\ell_{\max}$ \cite{1506.07135}.}\label{fig2}
\end{figure}
The parity asymmetry problem can also be understood by studying the temperature correlation between any direction and its opposite one. For the two-dimensional spherical CMB map, the two-point correlation function is given be
\begin{equation}\label{C_theta}
C(\Theta)=\langle \Delta T(\hat{n})\Delta T(\hat{n'})\rangle=\sum_{\ell=\ell_{\min}}^{\infty}\frac{2\ell+1}{4\pi} C_{\ell} P_{\ell}(\cos\Theta),
\end{equation}
where $P_{\ell}$ are the Legendre polynomials, and $\cos\Theta=\hat{n}\cdot\hat{n'}$. From this formula, we can easily show the correlations of the largest angular distance,
\begin{equation}\label{C_phi}
C(\Theta=\pi)=\sum_{\ell=\ell_{\min}}^{\infty}\frac{2\ell+1}{4\pi} C_{\ell} (\Gamma_{\ell}^{+}-\Gamma_{\ell}^{-}).
\end{equation}
For the even-parity preference case, we have a positive correlation, and for the odd-parity preference case, we have an anti-correlation. In Fig. \ref{fig4}, we plot the theoretical prediction of the correlation function $C(\Theta=\pi)$ for different $\ell_{\min}$, and also compare them with the results derived from the real WMAP data. Clearly, we find that the model predictions favor the even-parity preference, while the real data seems favor the odd-parity preference. So, this analysis also shows an odd-parity preference of CMB low-multipole data, which is consistent with the conclusion above.
\begin{figure}[t]
\begin{center}
\includegraphics[width=10cm]{figs/f3.pdf}
\end{center}
\caption{Theoretical (blue curve) and observed (red curve) values of $C(\Theta=\pi)$ as a function of $\ell_{\min}$, where the seven-year WMAP power spectrum have been used as the observed data. The error bar indicates the $1\sigma$ confident level caused by cosmic variance \cite{zhao2012}.}\label{fig4}
\end{figure}
\section{Preferred axis of CMB parity violation}
\subsection{Preferred axis in the full-sky maps}
In this paper, we shall mainly investigate the directional properties of the CMB parity asymmetry. In order to realize it, we need a directional dependent statistic. However, as shown in the previous section, all the statistics defined to describe the parity problem are based on the power spectra $C_{\ell}$ and their estimators $\hat{C}_{\ell}$, which are all rotationally invariant. Statistical invariance means that for any rotation of the reference system of coordinate, the power spectrum and the correlation function are invariant. So, any statistic defined by them are all coordinate independent. In order to break this kind of invariance, similar to other works \cite{alignment0,alignment,alignment2}, we should replace the estimator $\hat{C}_{\ell}$ by rotationally variant estimators. First, let us work on the full-sky CMB maps. The simplest one can be defined as following,
\begin{equation}\label{D_l}
\hat{D}_{\ell}=\frac{1}{2\ell}\sum_{m=-\ell}^{\ell} a_{\ell m} a_{\ell m}^*(1-\delta_{m0}),
\end{equation}
where $\delta_{mm'}$ is the Kroneker symbol. From the definition, we know that $\hat{D}_{\ell}$ is also an unbiased estimator for the power spectrum $C_{\ell}$ for any given multipole, i.e. $\langle \hat{D}_{\ell}\rangle=C_{\ell}$. Now, we can study the estimator $\hat{D}_{\ell}$ in any coordinate system. Imagining that the Galactic coordinate system is rotated by the Euler angle $(\psi,\theta,\phi)$, the coefficients $a_{\ell m}(\psi,\theta,\phi)$ in this new coordinate system can be calculated by
\begin{equation}
a_{\ell m} =\sum_{m=-\ell}^{\ell} a_{\ell m'} D_{m m'}^{\ell}(\psi,\theta,\phi),
\end{equation}
where $a_{\ell m}\equiv a_{\ell m}(0,0,0)$ are the coefficients defined in the Galactic coordinate system, and $D_{m m'}^{\ell}(\phi,\theta,\phi)$ is the Wigner rotation matrix \cite{edmond}. Similar to Eq. (\ref{D_l}), we can define the estimator $\hat{D}_{\ell}(\psi,\theta,\phi)$. It is easy to find that $\hat{D}_{\ell}(\psi,\theta,\phi)$ is independent of the angle $\psi$. So in this paper, we only consider two Euler angle $\hat{\rm{\bf q}}\equiv(\theta,\phi)$ and set $\psi=0$. If we consider $\hat{\rm{\bf q}}$ as a vector, which labels the $z$-axis direction in the rotated coordinate system, then $(\theta,\phi)$ is the polar coordinate of this direction in the Galactic system, which relates to the Galactic coordinate $(l,b)$ by $b=90^{\circ}-\theta$ and $l=\phi$. For any given coordinate labeled by $\hat{\rm{\bf q}}$, the components $a_{\ell 0}$ are naturally symmetric around the $z$-axis, i.e. $\hat{\rm{\bf q}}$. \emph{So from the definition of $\hat{D}_{\ell}$, in which the $m=0$ components are excluded, we know, in this coordinate system, that the $z$-axis (i.e. $\hat{\rm{\bf q}}$) is the preferred axis. If we rotate the coordinate system, the preferred axis also rotates.}
Now, we can define the rotationally variable parity parameter $G_1(\ell;\hat{\rm{\bf q}})$ as follows,
\begin{equation}\label{G_1}
G_1(\ell;\hat{\rm{\bf q}})=\frac{\sum_{\ell'=2}^{\ell}{\ell'(\ell'+1)}\hat{D}_{\ell'}(\hat{\rm{\bf q}})\Gamma_{\ell'}^{+}}{\sum_{\ell'=2}^{\ell}{\ell'(\ell'+1)}\hat{D}_{\ell'}(\hat{\rm{\bf q}})\Gamma_{\ell'}^{-}}.
\end{equation}
This statistic stands for the amplitude of the original parity parameter $P^+/P^-$, which is associated with the degree of the parity asymmetry, where a value of $G_1<1$ indicates an odd-parity preference, and $G_1>1$ indicates an even-parity preference. At the same time, due to the rotational variance of $G_1(\ell;\hat{\rm{\bf q}})$, we can study the possible preferred direction, which may reveal hints on the origin of the observed parity asymmetry in the CMB field. For any given $\ell$, the sky map $G_1(\ell;\hat{\rm{\bf q}})$ can be constructed by considering all directions $\hat{\rm{\bf q}}$. In practice, we pixelize the full sky in HEALPix format with the resolution parameter $N_{\rm side}=64$ and set the direction $\hat{\rm{\bf q}}$ to be those of the pixels.
{In previous works, the authors in \cite{kim2011}, using the power spectrum $C_{\ell}$, found that the CMB parity asymmetry is quite significant at the low multipoles, and this tendency extends to the multipole range $\ell<22$. Since the definition of the estimator $\hat{D}_{\ell}$ is similar to that of $C_{\ell}$, one expects the parity asymmetry of the statistic $G_1$ to also extend to this multipole range.} Using the Planck 2013 SMICA map, we compute the directional parity parameter $G_1(\ell;\hat{\rm{\bf q}})$ for any direction $\hat{\rm{\bf q}}$. As we have mentioned, $\hat{\rm{\bf q}}$ labels the $z$-axis direction in the rotated coordinate system, and $(\theta,\phi)$ is just the the polar coordinate of this direction in the Galactic coordinate. We plot the parameter $G_1(\ell;\hat{\rm{\bf q}})$ as a function of $\hat{\rm{\bf q}}$ for $3\le\ell\le22$ in Fig. \ref{fig5} (left panels). From this figure, we find that $G_1(\ell;\hat{\rm{\bf q}})<1$ holds for any direction $\hat{\rm{\bf q}}$ and maximum multipole $\ell$, which is consistent with the discovery of the odd-parity preference in the previous discussions. Smaller $G_1$ value leads to larger parity violation. In addition, we find that for any given maximum multipole $\ell$, except for the case with $\ell=3$, all the $G_1$ maps have quite similar morphologies. In Table \ref{tab2}, we list the preferred direction $\hat{\rm{\bf q}}$, where the parity parameter $G(\ell;\hat{\rm{\bf q}})$ for each maximum $\ell$ is minimized, and find that in all these cases, the preferred directions are nearly the same. Note that, throughout this paper, we do not differentiate the direction $\hat{\rm{\bf q}}$ and the opposite one $-\hat{\rm{\bf q}}$.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=15cm]{figs/f4.pdf}
\end{center}
\caption{Three directional statistics $G_{1}(\ell;\hat{\bf q})$ (left),
$G_2(\ell;\hat{\bf q})$ (middle), and $G_3(\ell;\hat{\bf q})$ (right) as
functions of $\hat{\bf q}\equiv (\theta,\phi)$. Note that,
these results are based on the Planck 2013 SMICA data \cite{zhao2014}.}\label{fig5}
\end{figure*}
\begin{table*}
\caption{The preferred direction $\hat{\bf q}=(\theta,\phi)$,
where the parity parameter $G_i(\ell;\hat{\bf q})$ based on Planck 2013
SMICA data is minimized, compared with the other CMB preferred
axes. In each box, the upper one is the result for the statistic with $i=1$,
the middle one is that for $i=2$, and the lower one is that for $i=3$. In
this table, $\alpha$ is the angle between $\hat{\bf q}$ and the
CMB kinematic dipole, $\langle|\cos\theta_{ij}|\rangle$ is the
quantity defined in Eq. (\ref{ctheta}), and the $\Delta_c/\sigma_c$ value denotes
the number of $\sigma_c$ the observed
$\langle|\cos\theta_{ij}|\rangle$ deviates from the simulations \cite{zhao2014}.}
\begin{center}
\label{tab2}
\begin{tabular}{ |c||c |c |c |c |c| }
\hline
& ~~~~~~$\theta[^{\circ}]$~~~~~~ & ~~~~~~$\phi[^{\circ}]$~~~~~~ & ~~~~~~$|\cos\alpha|$~~~~~~ & ~~~~$\langle|\cos\theta_{ij}|\rangle$~~~~ & ~~~~~~~$\Delta_c/\sigma_c$~~~~~~~ \\
\hline
\hline
\multirow{3}{*}{$\ell_{\max}=3$} & 90.00 & 23.20 & 0.3265 & 0.6066 & 0.90 \\
& 90.00 & 23.20 & 0.3265 & 0.6066 & 0.90 \\
& 90.00 & 23.20 & 0.3265 & 0.6066 & 0.90 \\
\hline
\multirow{3}{*}{$\ell_{\max}=5$} & 45.80 & 281.07 & 0.9767 & 0.9015 & 3.40 \\
& 45.80 & 281.07 & 0.9767 & 0.9015 & 3.40 \\
& 45.80 & 281.07 & 0.9767 & 0.9015 & 3.40 \\
\hline
\multirow{3}{*}{$\ell_{\max}=7$} & 48.19 & 277.73 & 0.9799 & 0.8979 & 3.37 \\
& 47.39 & 279.29 & 0.9782 & 0.8987 & 3.38 \\
& 52.83 & 267.89 & 0.9710 & 0.8915 & 3.32 \\
\hline
\multirow{3}{*}{$\ell_{\max}=11$} & 52.08 & 284.06 & 0.9525 & 0.8744 & 3.17 \\
& 49.77 & 280.54 & 0.9697 & 0.8886 & 3.29 \\
& 53.58 & 226.41 & 0.8679 & 0.8793 & 3.21 \\
\hline
\multirow{3}{*}{$\ell_{\max}=21$} & 52.08 & 285.47 & 0.9479 & 0.8721 & 3.15 \\
& 50.55 & 284.06 & 0.9575 & 0.8804 & 3.22 \\
& 21.32 & 131.90 & 0.5292 & 0.8295 & 2.79 \\
\hline
\end{tabular}
\end{center}
\end{table*}
\begin{table}
\caption{The definitions of six directional statistics considered in the text.}
\begin{center}
\label{tab1}
\begin{tabular}{ |c|c| }
\hline
Number of statistic & Definition \\
\hline
$1^{\rm st}$ & $G_1(\ell;\hat{\bf q})$ with $\hat{D}_{\ell}(\hat{\bf q})$ \\
\hline
$2^{\rm nd}$ & $G_2(\ell;\hat{\bf q})$ with $\hat{D}_{\ell}(\hat{\bf q})$ \\
\hline
$3^{\rm rd}$ & $G_3(\ell;\hat{\bf q})$ with $\hat{D}_{\ell}(\hat{\bf q})$ \\
\hline
$4^{\rm th}$ & $G_1(\ell;\hat{\bf q})$ with $\hat{D}'_{\ell}(\hat{\bf q})$ \\
\hline
$5^{\rm th}$ & $G_2(\ell;\hat{\bf q})$ with $\hat{D}'_{\ell}(\hat{\bf q})$ \\
\hline
$6^{\rm th}$ & $G_3(\ell;\hat{\bf q})$ with $\hat{D}'_{\ell}(\hat{\bf q})$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Independence of the statistics}
In the previous discussion, based on the analysis of the parity parameter $G_1(\ell;\hat{\rm{\bf q}})$, we found the preferred direction of the CMB parity asymmetry, which is independent of the maximum multipole $\ell$. However, an important problem arises: Whether or not the conclusion derived above depends on the definition of the statistic or estimator? In order to cross-check the result, we consider another rotationally variant estimator, which is proposed by de Oliveira-Costa et al. \cite{de2004}
\begin{equation}
\hat{D}'_{\ell}\equiv \frac{1}{2\ell+1}\sum_{m=-\ell}^{\ell} m^2 |a_{\ell m}|^2.
\end{equation}
If our universe is statistically isotropic, the ensemble average of this estimator is related to the power spectrum as follows,
\begin{equation}
\langle \hat{D}'_{\ell}\rangle =\frac{\ell(\ell+1)}{3}C_{\ell}.
\end{equation}
\emph{As we discussed, this estimator has also chosen a preferred direction, i.e. the $z$-axis direction.} In addition, this estimator favors high $m$s and so it works well in searches for planarity. In a quantum-mechanical system, this quantity also corresponds to the angular momentum along the $z$-axis direction. So the statistic defined by this estimator can also be used to search for the preferred axis in the CMB field.
In order to avoid the dependence of the statistic, in addition to the estimator $G_1(\ell;\hat{\rm{\bf q}})$ defined in Eq. (\ref{G_1}), we also consider the following different statistics, which are constructed by using two kinds of estimators $\hat{D}_{\ell}$ and $\hat{D}'_{\ell}$. From the definition of the two-point function in Eqs. (\ref{C_theta}) and (\ref{C_phi}), we can easily define the estimator for the correlation function at the largest angular distance $\Theta=\pi$ as,
\begin{equation}
\hat{C}(\Theta=\pi;\hat{\rm{\bf q}})=\sum_{\ell}\frac{(2\ell+1)}{4\pi} \hat{D}_{\ell}(\hat{\rm{\bf q}})(\Gamma^{+}_{\ell}-\Gamma^{-}_{\ell}).
\end{equation}
So, the natural way to estimate the relative contribution of the even and odd multipoles to the correlation function is to define the statistic
\begin{equation}\label{G_2}
G_2(\ell;\hat{\rm{\bf q}})=\frac{\sum_{\ell'=2}^{\ell}{(2\ell'+1)}\hat{D}_{\ell'}(\hat{\rm{\bf q}})\Gamma_{\ell'}^{+}}{\sum_{\ell'=2}^{\ell}{(2\ell'+1)}\hat{D}_{\ell'}(\hat{\rm{\bf q}})\Gamma_{\ell'}^{-}},
\end{equation}
which follows that $\hat{C}(\Theta=\pi)\propto (G_2(\ell;\hat{\rm{\bf q}})-1)$. $G_2>1$ corresponds to the positive correlation of the opposite direction, and $G_2<1$ indicates the anticorrelation of them. Note that the statistic $G_2$ is different from $G_1$, due to the different factors before $\hat{D}_{\ell}$ in their definitions. Therefore, the relative weights of low multipoles are much higher in $G_2$ than those in $G_1$.
For further investigation, in this paper, we also consider a third statistic to quantify the parity asymmetry, which was first introduced in Ref.\cite{Aluri2012}
\begin{equation}\label{G_3}
G_3(\ell;\hat{\rm{\bf q}})=\frac{2}{\ell-1}\sum_{\ell'=3}^{\ell} \frac{(\ell'-1)\ell'\hat{D}_{\ell'-1}(\hat{\rm{\bf q}})}{\ell'(\ell'+1)\hat{D}_{\ell'}(\hat{\rm{\bf q}})},
\end{equation}
where the maximum $\ell$ is any odd multipole $\ell\ge 3$ and the summation is over all odd multipoles up to $\ell$. This statistic is the measure of the mean deviation of the ratio of power in the even multipole and its succeeding odd-multipole from one.
We apply these two statistics for all the odd multipoles $3\le \ell
\le 21$ to the released Planck 2013 SMICA, NILC and SEVEM data. The results for SMICA
data are presented in Fig. \ref{fig5} (middle and right panels). For all the odd maximum multipoles $\ell$
and directions $\hat{\bf q}$, we have $G_i<1$ for $i=2,3$. These
are also correct for both Planck NILC and SEVEM data. So, we
find that the real CMB data have the odd-parity preference, which
is independent of the choice of the parity statistics.
The other three directional statistics are defined by similar manner, but the estimator $\hat{D}_{\ell}$ is replaced by $\hat{D}'_{\ell}$, i.e.
\begin{equation}
G_4(\ell;\hat{\rm{\bf q}})=G_1(\ell;\hat{\rm{\bf q}})|_{\hat{D}_{\ell}\rightarrow\hat{D}'_{\ell}},
\end{equation}
\begin{equation}
G_5(\ell;\hat{\rm{\bf q}})=G_2(\ell;\hat{\rm{\bf q}})|_{\hat{D}_{\ell}\rightarrow\hat{D}'_{\ell}},
\end{equation}
\begin{equation}
G_6(\ell;\hat{\rm{\bf q}})=G_3(\ell;\hat{\rm{\bf q}})|_{\hat{D}_{\ell}\rightarrow\hat{D}'_{\ell}},
\end{equation}
which are summarized in Table \ref{tab1}.
{As we have emphasized, the estimator $\hat{D}'_{\ell}$ is quite different from $\hat{D}_{\ell}$ due to the factor $m^2$ in the definition. In the statistics $G_4$, $G_5$, or $G_6$, the contributions of the higher multipoles, $\ell\sim \ell_{\max}$, become completely dominant. For this reason, we only apply these statistics to the multipole range in which the CMB parity asymmetry is most obvious. Although the CMB parity asymmetry can extend to multipole ranges up to $\ell \sim 22$, the main contribution comes from the lowest multipoles, i.e., $\ell<10$, which can be clearly seen in Fig. \ref{fig4}. So, we only consider the parity statistics $G_4$, $G_5$, and $G_6$ for the multipoles $\ell < 10$.} We apply these three statistics to the Planck 2013 SMICA, NILC, and SEVEM data and find similar results. In Fig.\ref{fig6}, we present the results of SMICA data, which show that $G_i(\ell,\hat{\bf q})< 1$ ($i=4,~5,6~$) is held for any direction $\hat{\bf q}$. Therefore, we conclude that the odd-parity preference exists even if the estimator $\hat{D}'_{\ell}$ is considered. In Tables \ref{tab2} and \ref{tab3}, we list the preferred directions $\hat{\bf q}$ for six statistics, which are quite similar to each other. So, we conclude that the preferred direction in the CMB parity asymmetry is independent of the definition of statistics.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=15cm]{figs/f5.pdf}
\end{center}
\caption{Three directional statistics $G_4(\ell;\hat{\bf q})$ (left),
$G_5(\ell;\hat{\bf q})$ (middle), and $G_6(\ell;\hat{\bf q})$ (right) as
functions of $\hat{\bf q}\equiv (\theta,\phi)$. Note that
these results are based on the Planck 2013 SMICA data \cite{zhao2014}.}\label{fig6}
\end{figure*}
\begin{table*}
\caption{The preferred direction $\hat{\bf q}=(\theta,\phi)$,
where the parity parameter $G_i(\ell;\hat{\bf q})$ based on Planck 2013
SMICA data is minimized, compared with the other CMB preferred
axes. In each box, the upper one is the result for the statistic with $i=4$,
the middle one is that for $i=5$, and the lower one is that for $i=6$. In
this table, $\alpha$ is the angle between $\hat{\bf q}$ and the
CMB kinematic dipole, $\langle|\cos\theta_{ij}|\rangle$ is the
quantity defined in Eq. (\ref{ctheta}), and the $\Delta_c/\sigma_c$ value denotes
the number of $\sigma_c$ the observed
$\langle|\cos\theta_{ij}|\rangle$ deviate from the simulations \cite{zhao2014}.}
\begin{center}
\label{tab3}
\begin{tabular}{|c||c |c |c |c |c| }
\hline
& ~~~~~~$\theta[^{\circ}]$~~~~~~ & ~~~~~~$\phi[^{\circ}]$~~~~~~ & ~~~~~~$|\cos\alpha|$~~~~~~ & ~~~~$\langle|\cos\theta_{ij}|\rangle$~~~~ & ~~~~~~~$\Delta_c/\sigma_c$~~~~~~~ \\
\hline
\hline
\multirow{3}{*}{$\ell_{\max}=3$} & 88.81 & 23.20 & 0.3109 & 0.5975 & 0.83 \\
& 88.81 & 23.20 & 0.3109 & 0.5975 & 0.83 \\
& 88.81 & 23.20 & 0.3109 & 0.5975 & 0.83 \\
\hline
\multirow{3}{*}{$\ell_{\max}=5$} & 47.39 & 307.86 & 0.8582 & 0.8458 & 2.93 \\
& 47.39 & 310.71 & 0.8408 & 0.8390 & 2.87 \\
& 46.59 & 309.92 & 0.8488 & 0.8442 & 2.92 \\
\hline
\multirow{3}{*}{$\ell_{\max}=7$} & 62.72 & 280.55 & 0.9107 & 0.8306 & 2.80 \\
& 56.49 & 281.25 & 0.9431 & 0.8599 & 3.05 \\
& 55.77 & 281.95 & 0.9443 & 0.8621 & 3.07 \\
\hline
\multirow{3}{*}{$\ell_{\max}=9$} & 32.60 & 236.25 & 0.9451 & 0.9424 & 3.75 \\
& 36.43 & 248.88 & 0.9815 & 0.9418 & 3.74 \\
& 34.89 & 246.06 & 0.9737 & 0.9435 & 3.76 \\
\hline
\end{tabular}
\end{center}
\end{table*}
\subsection{Independence of the CMB masks}
In the CMB observations, various foreground residuals are always unavoidable, especially in the Galactic region. The foreground residuals for the CMB maps released by Planck in 2015 are shown in Fig.\ref{fig7}. Usually one anticipates that the effects of these residuals are small and ignorable in the low multipole range. However, it is still worthy to investigate the cases in which these contaminated data are excluded. The simplest way to exclude the polluted region is to apply the top-hat mask to the data. For each CMB map, the corresponding mask suggested by Planck collaboration is also shown in Fig. \ref{fig7} (lower panels). We find that the masked region in the Commander and SMICA maps are quite similar. While the masked region for the NILC map is quite small, and the information loss in NILC map is expected to be much smaller than in the other two maps. For the masked map, the unbiased estimator for $C_{\ell}$ is not straightforward and a large number of methods to obtain it have been suggested in the literature \cite{method1,method4}. In this paper, we adopt the so-called pseudo-$C_{\ell}$ (PCL) estimator method \cite{method4}. Although PCL estimator is a suboptimal one, it can be easily realized in pixel space using fast spherical harmonics transformation, and has been applied to various CMB observations including to WMAP and Planck data. Considering the window function $W(\hat{n})$, the pseudo coefficients $\tilde{a}_{\ell m}$ can be defined as
\begin{equation}
\tilde{a}_{\ell m}=\int \Delta T(\hat{n}) W(\hat{n}) Y_{\ell m}(\hat{n}),
\end{equation}
which is related to $a_{\ell m}$ by
\begin{equation}
\tilde{a}_{\ell m}= \sum_{\ell_{1}m_{1}} a_{\ell_{1}m_{1}} K_{\ell m\ell_{1}m_{1}}.
\end{equation}
The coupling matrix $K$ is given by
\begin{equation}
K_{\ell m\ell_{1}m_{1}}=\sqrt{\frac{(2\ell_1+1)(2l+1)}{4\pi}}\sum_{\ell_2 m_2} (-1)^{m}(2\ell_2+1) w_{\ell_{2}m_{2}}
\begin{pmatrix}
\ell_1 & \ell_2 & \ell \\
0 & 0 & 0
\end{pmatrix}
\begin{pmatrix}
\ell_1 & \ell_2 & \ell \\
m_1 & m_2 & -m
\end{pmatrix},
\end{equation}
and $w_{\ell m}$ are the coefficients of spherical harmonics expansion of the mask $W(\hat{n})$, i.e.,
\begin{equation}
w_{\ell m}=\int W(\hat{n}) Y_{\ell m}^* (\hat{n}) d\hat{n}.
\end{equation}
The pseudo estimator $\tilde{C}_{\ell}$ is defined analogous to (\ref{hat-cl}) in terms of the multipole coefficients $\tilde{a}_{\ell m}$ as
\begin{equation}
\tilde{C}_{\ell}=\frac{1}{2\ell+1} \sum_{m=-\ell}^{\ell} \tilde{a}_{\ell m} \tilde{a}^*_{\ell m}.
\end{equation}
The expectation value of $\tilde{C}_{\ell}$ is
$\langle \tilde{C}_{\ell} \rangle =\sum_{\ell'}C_{\ell'} M_{\ell \ell'}$,
where the coupling matrix is
\begin{equation}
M_{\ell \ell'}=(2\ell'+1)\sum_{\ell_{2}} \frac {2\ell_{2}+1} {4\pi}
\begin{pmatrix}
\ell'&\ell_2&\ell \\
0 & 0& 0
\end{pmatrix} ^{2}\tilde{w}_{\ell_{2}}
\end{equation}
and ${\tilde w}_{\ell}$ are the following power spectrum,
\begin{equation}
\tilde{w}_{\ell} = \frac{1}{2\ell+1}\sum_{m=-\ell}^{\ell}w_{\ell m}w^{*}_{\ell m}.
\end{equation}
Similarly, the unbiased estimator in the masked sky can be constructed as
$\hat{\mathcal{C}}_{\ell}=\sum_{\ell'} M^{-1}_{\ell \ell'} \tilde{C}_{\ell'}$.
Note that this unbiased estimator $\hat{\mathcal{C}}_{\ell}$ is also rotationally invariant. Actually, the general analyses of the CMB parity asymmetry are always based on estimators of the CMB power spectrum in the masked space \cite{kim2011,planck2013,planck2015}.
In this paper, we focus on the direction dependence of the CMB parity violation. So, the direction dependent estimators are needed in advance. Similar to the discussion above, we can build the direction dependent estimator by excluding the $m=0$ components,
\begin{equation}\label{tilde-dl}
\tilde{D}_{\ell}=\frac{1}{2\ell} \sum_{m=-\ell}^{\ell} \tilde{a}_{\ell m} \tilde{a}^*_{\ell m}(1-\delta_{m0}).
\end{equation}
For each multipole, we have excluded the $m=0$ component, which means that the $z$-direction of the coordinate system is chosen as the preferred direction in the definition. However, the estimators $\tilde{D}_{\ell}$ are not unbiased. The expectation values are given by
$\langle \tilde{D}_{\ell} \rangle =\sum_{\ell'}C_{\ell'} N_{\ell \ell'}$,
where the coupling matrix $N_{\ell \ell'}$ is given by
\begin{equation}
N_{\ell \ell'}=M_{\ell \ell'}-\frac{2\ell' +1}{2\ell} \sum_{\ell_{2} \ell_{2}^{'}m_{1}} \frac {\sqrt{(2\ell_{2} +1)(2\ell_{2}^{'} +1)}} {4\pi}
\begin{pmatrix}
\ell'&\ell_2&\ell \\
0 & 0& 0
\end{pmatrix}
\begin{pmatrix}
\ell'&\ell_2^{'}&\ell \\
0 & 0& 0
\end{pmatrix}
\begin{pmatrix}
\ell'&\ell_2&\ell \\
m_1 & -m_1& 0
\end{pmatrix}
\begin{pmatrix}
\ell'&\ell_2&\ell \\
m_1 & -m_1& 0
\end{pmatrix} w_{\ell_2 m_1} w_{\ell_2^{'} m_1}.
\end{equation}
Based on this relation, we can construct the unbiased estimator $\hat{\mathcal{D}}_{\ell}$ as follows,
\begin{equation}
\hat{\mathcal{D}}_{\ell}=\sum_{\ell'} N^{-1}_{\ell \ell'} \tilde{D}_{\ell'}.
\end{equation}
Similar to $\hat{D}_{\ell}$, $\hat{\mathcal{D}}_{\ell}$ are also the coordinate dependent unbiased estimators for the power spectra $C_{\ell}$, and the preferred direction is also the $z$-direction of the corresponding coordinate system.
For any coordinate system, the direction-dependent unbiased estimator $\hat{\mathcal{D}}_{\ell}(\hat{\rm{\bf q}})$ can be built in the same manner with $\hat{\mathcal{D}}_{\ell}$, being, however, the coefficients $\tilde{a}_{\ell m}$ and ${w}_{\ell m}$ replaced by $\tilde{a}_{\ell m}(\hat{\rm{\bf q}})$ and ${w}_{\ell m}(\hat{\rm{\bf q}})$. The direction-dependent statistic for the CMB parity asymmetry can be defined as
\begin{equation}
\mathcal{G}(\ell;\hat{\rm{\bf q}})=\frac{\sum_{\ell'=2}^{\ell}\ell'(\ell'+1)\hat{\mathcal{D}}_{\ell'}(\hat{\rm{\bf q}})\Gamma^{+}_{\ell'}}
{\sum_{\ell'=2}^{\ell}\ell'(\ell'+1)\hat{\mathcal{D}}_{\ell'}(\hat{\rm{\bf q}})\Gamma^{-}_{\ell'}}.
\end{equation}
Since $\hat{\mathcal{D}}_{\ell}$ are the unbiased estimators for the power spectra $C_{\ell}$, the new statistic $\mathcal{G}(\ell;\hat{\rm{\bf q}})$ also indicates the degree of the CMB parity asymmetry and its direction dependence. Comparing with the ideal case for full-sky map and negligible noise, the use of the mask affects the values of statistic $G$ in two aspects: 1) the CMB information is lost in the masked region, and the values of the unbiased estimators for the power spectrum $C_{\ell}$ and their uncertainties might be influenced; 2) the structure and position of the mask may influence the preferred direction of $\mathcal{G}$-maps by the definition of the directional estimator $\tilde{D}_{\ell}$ in Eq. (\ref{tilde-dl}). If the masked region is small, we expect both effects to be negligible, and the results in the masked case should be very close to the ideal case. So, in this paper, we shall only consider the NILC mask suggested by Planck collaboration (see the lower middle panel in Fig. \ref{fig7}).
Based on the estimators in the masked maps, we plot the $\mathcal{G}$-maps for different maximum multipole $\ell$ and different CMB maps in Fig. \ref{fig8}, from which we find that the morphological structures of the $\mathcal{G}$-maps are nearly the same for all three input CMB maps. In addition, we also find that they are also quite similar with the results without the mask. In Table \ref{tab6}, as an example, we consider the Planck 2015 NILC map and list the preferred directions for both unmasked and masked cases. All these directions nearly align with each other. This result clearly shows that the preferred directions are independent of the used CMB masks, which stabilize our discovery on the directional properties stored in the CMB parity asymmetry.
\begin{figure}[t]
\begin{center}
\includegraphics[width=15cm]{figs/f6.pdf}
\end{center}
\caption{The 2015 Planck temperature anisotropy maps, including Commander, NILC, and SMICA. The lower panels are the corresponding masks suggested by Planck collaborations \cite{zhao2015}.}\label{fig7}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=15cm]{figs/f7.pdf}
\end{center}
\caption{The directional statistics $\mathcal{G}(\ell;\hat{\rm{\bf q}})$ for different maximum multipoles $\ell=9$ (upper) and $\ell=19$ (lower) based on the masked Commander (left), NILC (middle) and SMICA (right) maps. Note that, in the figure, we have applied the NILC mask to all the three maps \cite{zhao2015}.}\label{fig8}
\end{figure}
\begin{table*}
\caption{The preferred direction $(\theta,\phi)$, and the corresponding $|\cos\alpha|$ and $\Delta_c/\sigma_c$ for $\mathcal{G}(\ell;\hat{\rm{\bf q}})$ based on Planck NILC map, where the different maximum multipole $\ell$ is considered. For each $\ell$ case, the upper values denote the results derived from the full-sky analysis, and the lower values denote those derived from the masked case in which NILC mask is applied \cite{zhao2015}.}
\begin{center}
\label{tab6}
\begin{tabular}{ |c||c |c |c |c |}
\hline
& $~~~~~\theta[^{\circ}]$~~~~~ & ~~~~~$\phi[^{\circ}]$~~~~~ & ~~~~~$|\cos\alpha|$~~~~~ & ~~~~~$\Delta_c/\sigma_c$~~~~~\\
\hline
\hline
\multirow{3}{*}{$\ell_{\max}=5$} & $45.82$ & $279.73$ & $0.980$ & $3.42$ \\
& $45.82$ & $279.73$ & $0.980$ & $3.42$ \\
\hline
\multirow{3}{*}{$\ell_{\max}=7$} & $47.41$ & $278.00$ & $0.981$ & $3.39$\\
& $48.21$ & $275.06$ & $0.985$ & $3.40$\\
\hline
\multirow{3}{*}{$\ell_{\max}=9$} & $48.21$ & $276.47$ & $0.982$ & $3.35$\\
& $49.80$ & $272.25$ & $0.985$ & $3.38$\\
\hline
\multirow{3}{*}{$\ell_{\max}=11$} & $49.01$ & $277.17$ & $0.979$ & $3.35$\\
& $49.80$ & $272.25$ & $0.985$ & $3.38$\\
\hline
\multirow{3}{*}{$\ell_{\max}=13$} & $49.01$ & $278.58$ & $0.976$ & $3.34$\\
& $49.80$ & $272.25$ & $0.985$ & $3.38$\\
\hline
\multirow{3}{*}{$\ell_{\max}=15$} & $49.80$ & $282.10$ & $0.965$ & $3.27$\\
& $49.80$ & $272.25$ & $0.985$ & $3.38$\\
\hline
\multirow{3}{*}{$\ell_{\max}=17$} & $50.57$ & $284.21$ & $0.957$ & $3.22$\\
& $49.80$ & $270.84$ & $0.987$ & $3.39$\\
\hline
\multirow{3}{*}{$\ell_{\max}=19$} & $50.57$ & $284.21$ & $0.957$ & $3.22$\\
& $49.01$ & $270.14$ & $0.990$ & $3.42$\\
\hline
\multirow{3}{*}{$\ell_{\max}=21$} & $50.57$ & $284.21$ & $0.957$ & $3.22$\\
& $49.01$ & $270.14$ & $0.990$ & $3.42$\\
\hline
\end{tabular}
\end{center}
\end{table*}
\section{Comparing with the CMB low multipoles}
\subsection{CMB kinematic dipole}
As well known, the CMB has an extremely uniform temperature of 2.725 Kelvin, which is a left over from the period of recombination. The lowest anisotropy is the dipole component with an amplitude of $3.35$ mK from the Doppler shift of the background radiation \cite{dipole} caused by the
peculiar velocity of solar system at about $370$ km/sec relative to the comoving cosmic rest frame the dipole anisotropy defines a peculiar axis in the Universe, which is at $(\theta=42^{\circ}, \phi=264^{\circ})$ in Galactic coordinate system \cite{dipole} (see left panel of Fig. \ref{fig9}) \footnote{Note that, an alternative explanation for the CMB dipole is also discussed in \cite{alternative-dipole}}.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=10cm]{figs/f8.pdf}
\end{center}
\caption{The upper panel is the CMB kinematic dipole with in mK. Middle panel is the quadrupole (temperature range $\pm$ 35 $\mu$K) derived from wiener-filtered SMICA CMB sky, and lower panel is the derived octopole (temperature range of $\pm$ 35 $\mu$K). The plus and the star symbols indicate the axes of the quadrupole and octopole, respectively, around which the angular momentum dispersion is maximized. The diamond symbols correspond to the quadrupole axes after correcting for the kinematic quadrupole \cite{1303.5083}.}\label{fig9}
\end{figure*}
In Figs. \ref{fig4}, \ref{fig5} and \ref{fig7}, we compare the preferred directions $\hat{\bf
q}$ of parity asymmetry with the CMB kinematic dipole and find that
they are very close to each other.
Especially, all these directions are close to the ecliptic
plane. To quantify it, we define the quantity $\alpha$, which is
the angle between $\hat{\bf q}$ and the CMB kinematic dipole
direction at $(\theta=42^{\circ},\phi=264^{\circ})$ \cite{dipole}.
In Tables \ref{tab2}, \ref{tab3} and \ref{tab6}, we list the values of $|\cos\alpha|$ in the corresponding cases, and find that all of them are very close to $1$. For instance, by using the most recent Planck NILC map (see Table \ref{tab6}), in both masked case and unmasked case, we get that $|\cos\alpha|>0.98$ holds for all maximum multipole cases, which means that the angles between the preferred direction $\hat{\rm{\bf q}}$ and the CMB dipole direction are all smaller than $11.5^{\circ}$. In Tables \ref{tab4} and \ref{tab5}, we also list the corresponding values of $|\cos\alpha|$ derived from the other CMB maps based on six different statistics, and find similar results. So, we get a stable conclusion: \emph{The preferred direction in the CMB parity asymmetry strongly aligns with the CMB kinematic dipole, which is independent of the CMB maps, the directional statistics or the used mask.} This coincidence strongly indicates that the anomaly on the CMB parity asymmetry, as well as its directional properties, may be related to the CMB kinematic dipole component.
\begin{table*}
\caption{The values of $|\cos\alpha|$ and $\Delta_c/\sigma_c$ for the
statistics $G_i(\ell;\hat{\bf q})$ based on Planck 2012 year NILC and SEVEM
data. Similar to Table \ref{tab2}, in each box, the upper one is the result
for the statistic with $i=1$, the middle one is that for $i=2$, and the
lower one is that for $i=3$.}
\begin{center}
\label{tab4}
\begin{tabular}{ |c||c |c| |c |c| }
\hline
& ~~~$|\cos\alpha|$ for NILC~~~ & ~~~$\Delta_c/\sigma_c$ for NILC~~~ & ~~~$|\cos\alpha|$ for SEVEM~~~ & ~~~$\Delta_c/\sigma_c$ for SEVEM~~~ \\
\hline
\hline
\multirow{3}{*}{$\ell_{\max}=3$} & 0.3259 & 0.88 & 0.2656 & 0.66 \\
& 0.3259 & 0.88 & 0.2656 & 0.66 \\
& 0.3259 & 0.88 & 0.2656 & 0.66 \\
\hline
\multirow{3}{*}{$\ell_{\max}=5$} & 0.9758 & 3.38 & 0.9802 & 3.42 \\
& 0.9748 & 3.36 & 0.9802 & 3.42 \\
& 0.9748 & 3.36 & 0.9802 & 3.42 \\
\hline
\multirow{3}{*}{$\ell_{\max}=7$} & 0.9822 & 3.37 & 0.9840 & 3.41 \\
& 0.9769 & 3.36 & 0.9812 & 3.39 \\
& 0.9812 & 3.33 & 0.9861 & 3.37 \\
\hline
\multirow{3}{*}{$\ell_{\max}=11$} & 0.8600 & 3.18 & 0.8600 & 3.18 \\
& 0.9697 & 3.29 & 0.9766 & 3.34 \\
& 0.8520 & 3.15 & 0.8600 & 3.18 \\
\hline
\multirow{3}{*}{$\ell_{\max}=21$} & 0.8932 & 3.25 & 0.9529 & 3.19 \\
& 0.9575 & 3.22 & 0.9575 & 3.22 \\
& 0.8522 & 3.13 & 0.5295 & 2.79 \\
\hline
\end{tabular}
\end{center}
\end{table*}
\begin{table*}
\caption{The values of $|\cos\alpha|$ and $\Delta_c/\sigma_c$ for the statistics $G_i(\ell;\hat{\bf q})$ based on Planck 2013 year NILC and SEVEM data. Similar to Table \ref{tab3},
in each box the upper one is the result for the statistic with $i=4$, the middle one is that for $i=5$ and the lower one that is for $i=6$.}
\begin{center}
\label{tab5}
\begin{tabular}{ |c||c |c| c |c| }
\hline
& ~~~$|\cos\alpha|$ for NILC~~~ & ~~~$\Delta_c/\sigma_c$ for NILC~~~ & ~~~$|\cos\alpha|$ for SEVEM~~~ & ~~~$\Delta_c/\sigma_c$ for SEVEM~~~ \\
\hline
\hline
\multirow{3}{*}{$\ell_{\max}=3$} & 0.3094 & 0.78 & 0.2650 & 0.64 \\
& 0.3094 & 0.78 & 0.2650 & 0.64 \\
& 0.3094 & 0.78 & 0.2650 & 0.64 \\
\hline
\multirow{3}{*}{$\ell_{\max}=5$} & 0.8705 & 3.06 & 0.9040 & 3.16 \\
& 0.8705 & 3.06 & 0.8963 & 3.13 \\
& 0.8617 & 3.03 & 0.8963 & 3.13 \\
\hline
\multirow{3}{*}{$\ell_{\max}=7$} & 0.9838 & 3.37 & 0.9585 & 3.14 \\
& 0.9852 & 3.43 & 0.9710 & 3.26 \\
& 0.9782 & 3.38 & 0.9693 & 3.28 \\
\hline
\multirow{3}{*}{$\ell_{\max}=9$} & 0.9097 & 3.73 & 0.9351 & 3.74 \\
& 0.9450 & 3.77 & 0.9713 & 3.77 \\
& 0.9290 & 3.75 & 0.9529 & 3.77 \\
\hline
\end{tabular}
\end{center}
\end{table*}
\subsection{The CMB quadrupole and octopole}
The lowest cosmological anisotropic modes of the CMB fluctuations are the quadrupole and octopole. By defining the directional statistic and applying it to the first year WMAP data, Tegmark et al. found that both CMB quadrupole and octopole point to preferred directions \cite{alignment0}. It was also found that their orientation are strongly aligned with each other. Slightly later, by alternative methods, the alignment was confirmed by several groups \cite{alignment}. In addition, \cite{alignment2} found that the alignment of the CMB low multipole can extend to $\ell=5$, which is a significant violation of the random Gaussian assumption of the primordial fluctuations. Recently, Planck collaboration applied similar analysis to the Planck data \cite{1303.5083}. For each multipole $\ell$, they determine the orientation of the multipoles by finding the axis $\hat{n}$ around the maximized angular momentum dispersion.
\begin{equation}
\sum_{m} m^2 |a_{\ell m}(\hat{n})|^2
\end{equation}
Using the Planck 2013 SMICA map and considering the U73 mask, Planck collaboration constructed the quadrupole and the octopole, determining their preferred directions: $(\theta=13.4^{\circ}, \phi=238.5^{\circ})$ and $(\theta=25.7^{\circ}, \phi=239.0^{\circ})$, respectively. The angular difference between these directions is of only $12.3^{\circ}$, and the significance of the alignment is of $96.8\%$. In Fig. \ref{fig9} (middle and lower panels), we plot the constructed multipole components, and the corresponding preferred directions. The alignment between them is clearly shown. Using different maps, or considering the correction for the kinematic quadrupole (denoted as KQ corrected), the preferred directions of the quadrupole and the octopole are slightly different. They are listed in Table \ref{tab8}. However, in each case, the alignment holds at quite high confidence level.
\begin{table}
\caption{Preferred directions of CMB quadrupole and octopole in the Planck 2013 year data \cite{1303.5083}.}
\begin{center}
\label{tab8}
\begin{tabular}{ |c|c|c|c| }
\hline
CMB map & $(\theta,\phi)$ for quadrupole [degree] &$(\theta,\phi)$ for octopole [degree]&angle between them [degree]\\
\hline
C-R & $(29.7,228.2)$ & $(24.0,246.1)$ & 9.80 \\
NILC & $(12.7,241.3)$ & $(25.8,241.7)$ & 13.1 \\
SEVEM & $(16.2,242.4)$ & $(25.2,245.6)$ & 9.08\\
SMICA & $(13.4,238.5)$ & $(25.7,239.0)$ & 12.3 \\
NILC, KQ corrected & $(20.3,225.6)$ & $(25.8,241.7)$ & 8.35\\
SEVEM, KQ corrected & $(21.7,228.3)$ & $(25.2,245.6)$ & 7.69 \\
SMICA, KQ corrected & $(20.8,224.2)$ & $(25.7,239.0)$ & 7.63 \\
\hline
\end{tabular}
\end{center}
\end{table}
In this paper, we shall investigate whether or not
the alignment of quadrupole and octopole connects with the parity
asymmetry. Following Ref. \cite{anto}, it is
straightforward to evaluate the mean value of the inner product
between all the pairs of unit vectors corresponding to the
following four directions: The preferred directions of quadrupole,
octopole, parity asymmetry and the direction of the CMB kinematic
dipole. So, we define the quantity,
\begin{equation}\label{ctheta}
\langle |\cos\theta_{ij}|\rangle = \sum_{i,j=1,~j\neq i}^{N} \frac{|\hat{r}_i\cdot \hat{r}_j|}{N(N-1)},
\end{equation}
where $N$ is the number of the directions that will be
investigated. First, we shall study the case in the absence of the
parity asymmetry. The alignment between these three axes was also
reported in WMAP data \cite{schwarz}. In this case, we have $N=3$
and $\langle |\cos\theta_{ij}|\rangle=0.9242$ for the real data.
{To evaluate the significance of the alignment, we pixelize the two-dimensional sphere in the HEALPix format with the resolution parameter $N_{\rm side}=256$, which corresponds to the total pixel number $N_{\rm pix}=12\times N_{\rm side}^2$. Then, we randomly generate $10^5$ realizations. For each realization, all three directions are randomly and independently picked on the sky using a uniform distribution between 0 and $N_{\rm pix}-1$,} and the corresponding value $\langle |\cos\theta_{ij}|\rangle$ is calculated directly. Considering all the random samples, we obtain that $\langle |\cos\theta_{ij}| \rangle = 0.500 \pm 0.167$. {To quantify the significant level of the deviation from the random distribution, we define the
$\Delta_c/\sigma_c$, where $\Delta_c$ is the difference between the observed value of $\langle |\cos\theta_{ij}|\rangle$ and the mean value of the simulations, and $\sigma_c$ is the corresponding standard deviation of the simulations. Considering the observed result $\langle |\cos\theta_{ij}|\rangle=0.9242$, and the simulated value $\sigma_c=0.167$, we obtain that $\Delta_c/\sigma_c=2.54$, which indicates that the alignment of these three directions is around $2.5\sigma$ confidence level.}
Now, let us take into account the preferred direction of the CMB
parity asymmetry. For the $10^5$ random realizations of four
random points on the sphere, by a similar analysis, we get
\begin{equation}\label{theta4}
\langle |\cos\theta_{ij}| \rangle = 0.500 \pm 0.118.
\end{equation}
As anticipated, compared with the case of $N=3$, the mean value
stays the same, while the standard deviation $\sigma_c$ significantly decreases. Now, we
calculate the real value of the quantity $\langle
|\cos\theta_{ij}| \rangle$. For the preferred direction of parity
asymmetry, we consider the results of all the six statistics
$G_i(\ell;\hat{\bf q})$ defined in this paper, and list the
corresponding values in Tables \ref{tab1}-\ref{tab5}. For every
case with $\ell>3$, we find that $\langle |\cos\theta_{ij}| \rangle$
is close to $0.9$, and the corresponding significance of the
alignment between these directions increases to $\Delta_c/\sigma_c \gtrsim 3$.
In Fig. \ref{fig10}, we plot all these preferred directions in the Galactic coordinate system (left) and ecliptic coordinate system (right), and clearly present the alignment of them.
We therefore conclude that \emph{the preferred direction of the CMB
parity asymmetry is not only very close to the CMB kinematic
dipole, but also close to the preferred axes of the CMB quadrupole
and octopole, which is nearly independent of the choice of the
parity statistic, and the used CMB map or mask}. Since the quadrupole and the octopole also relate to other CMB anomalies, including the low quadrupole problem and the lack of the large-scale correlation, the coincidence between the preferred directions in the CMB low multipoles also hints: \emph{The CMB parity asymmetry is not an isolated anomaly, it may have the common origin with all these CMB anomalies.}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=13cm]{figs/f9.pdf}
\end{center}
\caption{The preferred directions of the SMICA-based statistics
$G_1(\ell;\hat{\bf q})$ in the Galactic coordinate
system (left) and in the ecliptic coordinate system (right). In both
panels, we have compared them with the CMB kinematic dipole
direction and the preferred directions of the CMB quadrupole and
octopole.}\label{fig10}
\end{figure*}
\section{Preferred axes in other large-scale observations}
In addition to the directional problem in the CMB low multipoles, in other cosmological observations, similar preferred axes were also reported in the literature. In particular, several axes are announced to align with the CMB kinematic dipole and in this section, we briefly list them as below.
\subsection{Alignment of quasar polarization vectors}
Quasars are the most energetic and distant members of a class of objects called active galactic nuclei. Since quasars show very high redshifts, they can be treated as the trackers to study the matter distributions of the Universe. Based on the sample of 170 polarized quasars, in 1998, Hutsemekers studied the distribution of the quasar polarization vectors \cite{quasar}. In general, we naturally expect the quasar polarization angles to be randomly distributed between $0^{\circ}$ and $180^{\circ}$. However, by applying two different statistical tests, the author found that the optical polarization vectors of quasars are not randomly distributed over the sky but are coherently oriented on very large spatial scales. The instrumental bias and the contamination by interstellar polarization in our Galaxy are unlikely to be responsible for these features. Lately, different groups and different analyses \cite{quasar2} also derived similar results, i.e. the quasar polarization vectors are not randomly oriented over the sky with a probability often in excess of $99.9\%$. The alignment effect seems to be prominent along a particular axis in the direction $(\theta=69^{\circ},\phi=267^{\circ})$ in the Galactic coordinate system (see Table \ref{tab7}), which seems coincident with the axis defined by the CMB kinematic dipole.
\subsection{Large-scale velocity flows}
Since the discovery of the cosmic microwave background radiation, people realized that one absolute reference system is defined, in which the CMB photons have no velocity flows. However, the regular matters, including baryons and dark matter, may have bulk flow relative to the CMB reference coordinate. The peculiar velocity field of the galaxies provides an important and robust way to understand the matter distribution and motion of the universe, which can be used to detect possible unobservable sources of gravitational fields. Recently, the peculiar velocity field of the nearby universe has been explored extensively using variant velocity tracers \cite{chen,velocity}. It was found that the flow amplitudes and directions strongly depends on the scales which have been focused on. In particular, on the intermediate scales from $20h^{-1}$Mpc to about $100h^{-1}$Mpc, there is evidence of a bulk flow with amplitude $416\pm 78$ ${\rm km~s}^{-1}$ at the direction of $(\theta=84^{\circ}\pm 6^{\circ},\phi=282^{\circ}\pm11^{\circ})$ \cite{velocity}. In the standard $\Lambda$CDM model, this flows can be generated by the primordial fluctuations with long-wavelength modes. However, the predictive amplitude at this scale is only approximately $100$ ${\rm km~s}^{-1}$, and the probability that a flow of magnitude larger than 400 ${\rm km~s}^{-1}$ is realized is less than $0.5\%$. This is the so-called dark flow problem for the galaxies. However, we should also mention that this problem is still in debate. Some other authors claimed the normal bulk velocities on this scale using separate velocity catalogs \cite{velocity2}.
\subsection{Handedness of spiral galaxies}
Another directional anomaly was found in the spiral galaxies. Using 15158 spiral galaxies with redshifts $z<0.085$ from the Sloan Digital Sky Surveys, in \cite{spiral} the author studied the left-handed and right-handed spirals of these galaxies and their distribution in the sky. Surprisingly enough, it was found that the handedness distribution of these spiral galaxies are not random. A significant dipole component with the amplitude $A=-0.0408\pm 0.011$ was discovered at more than $99.9\%$ confidence level. The observed spin correlation extends out to separations $\sim 210{\rm Mpc/h}$. Defining the axis of the asymmetry along to the direction of $L$ ($\circlearrowright$) excess, the preferred direction lies near $(\theta=158.5^{\circ},\phi=232^{\circ})$ (or the equivalent direction $(\theta=21.5^{\circ},\phi=52^{\circ})$) in Galactic coordinates, which seems to align with the preferred axis of spin vector in our galaxy. This distribution asymmetry and the alignment indicate a parity violation in the overall Universe.
\subsection{Anisotropy of the cosmic acceleration}
Soon after the discovery of the accelerating cosmic expansion from the observation of Type Ia supernova, a number of authors have investigated the anisotropies of this cosmic acceleration. In particular, several groups have applied the hemisphere comparison method to study the anisotropy of cosmic acceleration \cite{acceleration}. By comparing the supernova data and the corresponding cosmic accelerations on several pairs of opposite hemispheres, a statistically significant preferred axis has been reported in literature. Recently, this study is extended by Javanmard et al. in \cite{acc2}. Using the new leased Union 2.1 compilation, the authors singled out the most discrepant direction with the respect to the all-sky data, and built maps with three different angular resolutions to test the isotropy of the magnitude-redshift of supernovae. It was found that the null hypothesis should be rejected at $95\%$-$99\%$ confidence level, and the strongly deviations in the Union2.1 sample occurs at $(\theta=23.4^{\circ},\phi=247.5^{\circ})$, which closely align with the CMB kinematic dipole.
However, we should emphasize that this conclusion seems not to be stable now. By using different data samples, and/or different analysis methods, people have derived quite different results in the recent literature \cite{debate1,debate2,debate3}. For example, in our paper \cite{debate3}, we divided the Unions dataset into 12 subsets according to their positions in the Galactic coordinate system. In each region, we derived the deceleration parameter $q_0$ as the diagnostic to quantify the anisotropy level in the corresponding direction. The dipole component was also found in the $q_0$-map at more $95\%$ confidence level. The direction of the best-fit dipole is $(\theta=108.8^{\circ},\phi=180.7^{\circ})$, which is quite different from the direction found in \cite{acceleration} and \cite{acc2}. So, further analysis and much better data are needed to clarify this debate in the future.
\subsection{Anisotropic distribution of fine-structure constant}
On the sightline from us to the quasar, there are numerous gas clouds that absorbs the quasar radiation. Spectroscopic observations can reveal the absorption spectrum of the atoms, ions and molecules of the intervening clouds. These absorption spectra have been used to search for the space-time variation of the fine-structure constant $\alpha$ in cosmology \cite{fine}. Using more than one-decade data of Keck telescope and Very Large Telescope, Webb et al. reported evolution of $\alpha$ in the Universe, which significantly violate the prediction of the standard model of particle physics. The evolution tendencies of $\alpha$ are also different at different regions in the sky. It was shown that they well fit the angular dipole model. The dipole axis was found to point in the direction ($\theta=104^{\circ},\phi=331^{\circ}$) and the dipole amplitude was found to be $A=(0.97\pm0.21)\times10^{-5}$, which excludes the isotropic hypothesis at more than $4\sigma$ confidence level \cite{fine}. The discovered preferred axis in $\alpha$ evolution is claimed to strongly coincide with the other axes, in particular the preferred axis in the anisotropy of the cosmic acceleration \cite{acceleration}. So, it was suggested that at least these two anomalies have a common origin \cite{fine-common}: the anisotropy of the cosmic acceleration and the fine-structure constant anomaly. However, we also need to mention that there are still some debates regarding this conclusion \cite{debate4,debate5}.
\begin{table}
\caption{Preferred directions in various large-scale observations}
\begin{center}
\label{tab7}
\begin{tabular}{ |c|c|c| }
\hline
observations & ~~~~~~~~~~~$\theta$ [degree]~~~~~~~~~~~ & ~~~~~~~~~~~$\phi$ [degree]~~~~~~~~~~~ \\
\hline
CMB kinematic dipole & 42 & 264 \\
CMB quadrupole & 13.4 & 238.5 \\
CMB octopole & 25.7 & 239.0 \\
CMB parity asymmetry & 45.82 & 279.73 \\
Polarization of QSOs & 69 & 267 \\
Large-scale velocity flows & 84 & 282 \\
Handedness of spiral galaxies & 158.5 & 232 \\
Anisotropy of cosmic acceleration & 23.4 & 247.5 \\
Distribution of fine-structure constant & 104 & 331 \\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Dipole observations with radio galaxy catalog}
The peculiar velocity of the Solar System relative to the frame of distant radio sources were also studied by \cite{radio_sources} using the data of the NRAO VLA Sky Survey (NVSS). In these papers, even though, the direction of the dipole is in agreement with the CMB previous results, its magnitude was found to be much larger than the CMB expectations for the number counts or sky brightness observables. Presently, the NVSS dipole is in disagreement with the CMB predicted velocity dipole at 2.7$\sigma$ C.L. In addition, if the linearly polarised flux density is considered instead, the level of anisotropy increases in comparison to pure number counts or sky brightness. The authors suggest that if the difference between the two dipoles is confirmed, it would imply an intrinsically anisotropic universe, with the anisotropy changing with the epoch. Moreover, the axis of anisotropy would point roughly towards the CMB dipole direction. However, recent results of the NVSS dipole leaded to the conclusion that a large bias factor for radio galaxies at low redshift could explain the NVSS dipole signal. It is important to point out that several other explanations for the NVSS dipole are still considered in literature, and it remains a puzzle.
\section{Possible interpretations}
From the discussions above, we know the preferred axes have been found in a number of large-scale observations. And also, at least some of them seem to align with each other. These coincidences might imply the same origin for these anomalies and their interpretations can be divided into two different kinds: In the first kind of explanations, the anomalies have a cosmological origin. Since the existence of the preferred axis in the universe is a significant violation of the cosmological principle, if these anomalies are due to some cosmological effect, one has to consider an alternative model to replace the standard $\Lambda$CDM model. In the second kind of explanations, the standard cosmological model is correct, while the current observations or data analysis of the large-scale data are still inaccurate. These anomalies may be caused by some unsolved systematics or contaminations. In the following discussion, we will briefly introduce these interpretations, respectively.
\subsection{Non-trivial topology of the universe}
As mentioned before, the standard cosmological model is based on two assumptions: One is that Einstein's general relativity correctly describes gravity, the other assumes the universe as homogeneous and isotropic on large scales. If we believe that the anomalies have a cosmological origin, at least one of these two assumptions will be broken. In \cite{1303.5083,1303.5086}, Planck collaboration investigated several non-trivial topological models in the universes with locally flat, hyperbolic and spherical geometries. Unfortunately, no evidence has been found in the observed data.
Another possibility relies on the Bianchi models. The Bianchi classification provides a complete characterization of all the known homogeneous but anisotropic exact solution to general relativity \cite{ellis}. So, in general, Bianchi models can provide preferred directions in the universe (see for instance \cite{ellip}). Among them, Bianchi VII$_{h}$ models describe a universe with overall rotation, parameterized by an angular velocity and a three-dimensional rate of shear (including the case with zero angular velocity and non-zero shear \cite{john}). In this model, a free parameter is defined to describe the comoving length-scale over with the principal axes of shear and rotation change orientation, and many authors have studied the imprints of these models in the CMB data \cite{jaffe}. In \cite{1303.5086}, it was found that Planck data do provide evidence supporting a phenomenological Bianchi VII$_{h}$ component. However, a physical, anisotropic Bianchi universe is not supported by the data.
\subsection{Alternative gravitational theories}
In order to explain the bulk flow, in \cite{1009.1509} the authors considered that the universe is influenced by large-scale ``wind", and the cosmic matter is drifted by this ``wind". The velocity of the ``wind" takes account for the observed peculiar velocity. When the ``wind" has a privileged direction, the cosmic matter drifts towards the same direction. Actually, the ``wind" picture refers to the Zermelo navigation problem, which is described by the Finsler geometry \cite{bao2000}. Finsler geometry is a natural framework for describing an anisotropic spacetime. The study of the metric is the fourth root of a quartic differential form. Chang et al. found that if Riemann geometry is replaced by Finsler geometry, and Einstein's general relativity is replaced by the general relativity based on Finsler geometry, the observed bulk flow can be naturally explained. In addition, it was found that a matter dominated navigation cosmological model could account for the accelerating expansion of the universe \cite{li2012}, in which the anisotropy of the cosmic acceleration is also possible to be explained \cite{1009.1509}.
Another theoretical explanation of the observed preferred direction is suggested by Yan et al. \cite{deSitter}, which is motivated by the fact that the cosmological constant $\Lambda$ is nonzero. So, the metric of the local inertial reference system in the standard model of cosmology is the Beltrami metric instead of the Minkowski one, and the basic spacetime symmetry has to be from de Sitter's group. To avoid ambiguities caused by the inertial forces, quantum mechanics for spectra in atoms are defined in inertial coordinate systems. The corresponding special relativity is de Sitter special relativity, which is proposed in the literature. In this model, the Minkowski point does exist in the universe, where the Beltrami metric returns to the Minkowski one, and the physics at this point returns to the Einstein's special relativity. The Minkowski point naturally defines a preferred point in the universe, and the direction pointing is the preferred direction. When extending this theory to the general relativity, i.e. de Sitter general relativity, the authors found that the evolution of fine-structure constant $\alpha$, and its dipole structure can be well explained \cite{deSitter}.
\subsection{Particular fluctuation modes or dark energy models}
Actually, the explanation for the directional anomalies with a minimum cost is to consider the possible anisotropic matter component and/or superhorizon fluctuation modes in the universe. In the present stage of universe, the dark energy is the dominant component. If this component is anisotropic, the observed universe may display special directions. In the general scenario, the dark energy is described by the cosmological constant $\Lambda$ or some scalar field. However, it is also possible that the dark energy can be described by a vector field. For instance, in \cite{ym}, we suggested to use the quantum Yang-Mills condensate to describe the dark energy and promote the cosmic acceleration. For the vector fields, the spatial distribution is always anisotropic, which could easily lead to a preferred axis in the large scales. Besides, in \cite{fine-common}, some other special dark energy models have been proposed to explain the anisotropic of cosmic acceleration or the anisotropic distribution of the fine-structure constant, providing another possibility to explain the large-scale anomalies. Among them, we can cite anisotropic dark energy models, inhomogeneous dark energy models, topological quintessence, and spherical dark energy overdensity.
For the large-scale CMB anomalies, the mechanism of Grishchuk-Zeldovich effect is a possible explanation \cite{gz-effect}. It is the contribution to the CMB temperature anisotropy from an extremely large-scale adiabatic density perturbation, considering the standard hypothesis that this perturbation is a typical realization of an homogeneous Gaussian random field. Assuming that one particular large-scale mode of fluctuation, generated in the early inflationary stage, dominates the perturbations of the largest scales in the current stage of universe, then this mode necessarily indicates a preferred direction in the universe. It could also significantly influence the CMB quadrupole and octopole, which might answer the alignment problem of CMB low multipoles.
\subsection{Unsolved systematical errors or contaminations}
On the contrary, some people believe that the standard model of cosmology, based on the cosmological principle and general relativity, is an accurate model to describe the current universe. Considering that the observed large-scale directional anomalies have a non-cosmological origin, being therefore caused by some unsolved systematical errors, calibration errors or contaminations, we list some possibilities.
One possible explanation is related to the contaminations generated by the collective emission of Kuiper Belt objects and other minor bodies in the solar system where the kinematic dipole of CMB is located. Since the emission of the Kuiper Belt objects is nearly independent of the frequency, this contamination is very hard to remove from the CMB data analysis. In Maris et al. \cite{maris} and Hansen et al. \cite{hansen}, it was discussed that this foreground residual could leave significant imprints in the CMB low multipoles and possibly explain the CMB parity asymmetry, as well as the alignment of the CMB quadrupole and octopole.
Another explanation may relate to a deviation measured in the CMB kinematic dipole, which could be due to a measurement error in the dipole direction, a problem in the antenna pointing direction, sidelobe pickup contamination, and so on. In \cite{liu2011}, it was found that this kinematic dipole deviation could generate the artificial CMB anisotropies in the low multipoles. If this is true, these artificial components may account for the CMB large-scale anomalies.
It is also possible that the preferred direction is caused by the tidal field originated from the anisotropy of our local halo. In \cite{wang}, the authors found that the tidal field tends to preferentially align with the orientation and spatial distribution of galaxies, which may also generate some unsolved kinematic or higher order effects, and influence the cosmological observations \cite{wang2}.
\section{Conclusions and discussions}
The standard $\Lambda$CDM model has a great success in explaining the observations of the CMB temperature anisotropies, as well as the galaxies distribution and motion. The standard model of cosmology is based on the assumptions: the validity of Einstein's general relativity, and the cosmological principle. This model can explain most large-scale observations with unprecedented accuracy. However, several directional anomalies have been reported in various observations: the polarization distribution of the quasars, the velocity flow, the handedness of the spiral galaxies, the anisotropy of the cosmic acceleration, the anisotropic evolution of fine-structure constant, including anomalies in the CMB low multipoles, such as the CMB parity asymmetry. Although the confidence level for each individual anomaly is not too high, the directional alignment of all these anomalies is quite significant, which strongly suggests a common origin of these anomalies.
If these anomalies are due to cosmological effects , e.g. the alternative theory of gravity or geometry, the non-trivial topology of the universe, the anisotropic dark energy or the particular large-scale fluctuation modes, they indicate the violation of the cosmological principle. So, one should consider to build a new cosmological model to explain the large-scale data. However, if these directional anomalies arise from non-cosmological reasons, e.g. the unsolved systematical errors or contaminations, we should carefully treat the current data, and exclude the errors in the future analysis to avoid the misleading explanations of the data.
In order to distinguish these two kinds of explanations, we compare the preferred directions in large-scale observations and the CMB kinematic dipole, and found a strong alignment between them. As well known, the CMB dipole is caused by the motion of the Solar System in the universe, which is a purely kinematic effect. The alignment of CMB dipole and the other preferred direction strongly suggest a non-cosmological origin of the large-scale anomalies, which should be caused by some CMB dipole-related systematics or contamination. In future cosmological observations, we suggest to further study these possible errors, and subtract them from the observed data. In addition, we expect that the future measurements on the CMB polarization fields, the cosmic weak lensing, or the distribution of 21-cm line can help us to solve the puzzles.
\section*{Acknowledgements}
We acknowledge the use of the Legacy Archive for Microwave Background Data Analysis (LAMBDA) and Planck Legacy Archive (PLA). Our data analysis made the use of HEALPix
\cite{healpix}. This work is supported by Project 973 under Grant No. 2012CB821804, by NSFC No. 11173021, 11322324, 11421303 and project of KIP and CAS.
|
\section{Introduction}
The identification of individuals with influential positions in a network is of outmost importance for a number of problems in economics and beyond, as for instance regarding the diffusion of epidemics \citep[][]{PastorSatorrasVespignani2002}, the stability of systems of interconnected banks \citep[][]{Gofman2015} and the development of criminal networks \citep[][]{Liuetal2015}. Different contexts often lead to different measures of \emph{centrality} that capture the level of the said influence more appropriately. Degree centrality is found to be important in problems of adoption with word--of--mouth communication \citep[][]{GaleottiGoyal2009} and biases in the perception of social norms \citep[][]{Jackson2016}. Katz--Bonacich centrality \citep[][]{Katz1953,Bonacich1987} is found to be crucial in problems related to criminal behavior \citep[][]{Ballesteretal2006}, whereas eigenvector centrality is found to be important in diffusion processes \citep[][]{Banerjeeetal2013}.
In this article we focus on \emph{Decay Centrality}. This is a measure of centrality in which a node is rewarded for how close it is to other nodes, but in a way that very distant nodes are weighted less than closer ones \citep[see][]{Jackson2008}. It is defined as $\sum\limits_{j\neq i}\delta^{d(i,j)}$, where $0<\delta<1$ is a decay parameter and $d(i,j)$ is the geodesic distance between nodes $i$ and $j$. For low values of $\delta$ decay centrality puts much more weight on closer nodes, thus becoming proportional to degree centrality, whereas for high values of $\delta$ it measures the size of the component a node lies in
It is considered to be richer than other distance related measures, because it captures the idea that the importance of a node for another is proportional to their distance \citep[see for instance][]{JacksonWolinsky1996}. It has been considered important in problems of optimal targeting selection in networks \citep[see][]{Banerjeeetal2013,ChatterjeeDutta2015,Tsakas2016}. In particular, \cite{ChatterjeeDutta2015} and \cite{Tsakas2016} find decay centrality to be the measure that helps selecting the node that can lead to the maximum diffusion of a given action in a social network.
Nevertheless, its use is cumbersome for two main reasons. First, except in very simple structures, the nodes with maximum decay centrality cannot be easily identified, since the measure depends vastly on the exact network topology and the value of the decay parameter. Second, calculating the decay centrality of all nodes and subsequently choosing the one that maximizes it might be computationally costly, since it requires calculating the geodesic distance between each pair of nodes and subsequently summing a function of them.\footnote{For a network with $n$ nodes, the time complexity for calculating degree and closeness centrality are in $O(n^2)$ and $O(n^3)$ respectively \citep[see][]{BrandesErlebach2005}, where for the calculation of shortest paths that is necessary for closeness centrality is used the simple Dijkstra algorithm \citep[see][]{Dijkstra1959}. Once the shortest paths have been calculated, decay centrality requires the calculation of $\delta^{d(i,j)}$ for each pair of nodes $(i,j)$. Hence, the time complexity of calculating decay centrality is in $O(n^5)$.}
The aim of this article is to show the close connection between decay centrality and two well--studied and computationally cheaper measures, namely degree and closeness centrality. The relations are established both analytically and numerically and suggest that the nodes with maximum decay centrality usually belong either to the set of nodes with maximum degree or to the set of nodes with maximum closeness.
In particular, focusing on connected networks, we show that for sufficiently low values of the decay parameter the nodes that maximize decay centrality belong to the set of nodes with maximum degree, whereas for sufficiently high values of the decay parameter the nodes that maximize decay centrality maximize closeness as well. The first proposition is not surprising as it is already known that for low values of $\delta$ decay centrality is proportional to degree. However, the second proposition establishes a novel relationship between decay and closeness centrality for high values of $\delta$, for which so far decay centrality was associated only with the size of the component a node lied in.
Furthermore, we provide two conditions that are sufficient to order a pair of nodes with respect to their decay centrality for all values of the decay parameter. The conditions depend, in an intuitive way, on the distances between the nodes under comparison and all other nodes in the network.
Finally, we provide a more general relation between decay centrality, degree and closeness that extends to intermediate values of the decay parameter. In particular, we provide sufficient conditions for a node with higher degree than another node to also have higher decay centrality for all $\delta\leq 1/2$ and similarly for a node with higher closeness to also have higher decay centrality for all $\delta\geq1/2$.
Based on those theoretical findings, we attempt to obtain a better understanding on these relations for intermediate values of the decay parameter via numerical simulations. We find that in the vast majority of cases the nodes with maximum decay centrality belong either to the set of nodes with maximum degree or to the set of nodes with maximum closeness. When the two sets intersect a node that belongs to their intersection almost always maximizes decay centrality as well, whereas when the two sets do not intersect for low values of $\delta$ decay centrality is maximized by nodes with maximum degree and as $\delta$ increases there is a threshold above which decay centrality is maximized by nodes with maximum closeness. The threshold varies with the network parameters, however a rule of thumb with a threshold at $\delta=0.5$ is sufficient to ensure that the chosen node is ranked among the top nodes in terms of decay centrality, with high probability.
The relation between different measures of centrality has attracted research interest over the years \citep[see][and references therein]{Valenteetal2008}. For instance, \cite{Faust1997} found strong correlation among centrality measures that included degree and closeness, but not decay centrality, in a network of relations between CEOs. Similarly, \cite{Rothenbergetal1995} found strong concordance in the ranking of individuals across different network centrality measures in a network of individuals who participated in activities with the risk of HIV transmission. Finally, \cite{Valenteetal2008} find similar results, although not as strong correlation coefficients as the previous studies, for a variety of networks corresponding to different process. All of these studies show particularly strong connection between degree and closeness, but do not consider decay centrality.
More recently, and closer to the flavor of our analysis, \cite{Blochetal2016} provided an axiomatization of several centrality measures, including decay centrality, and showed that all of them can be described using the same set of axioms. Given these axioms, the measures differ only in that they consider different vectors of data that describe the position of nodes in the network. The authors also provide correlation coefficients between different measures in simulated networks and their results are in line with our findings. The simulations report results on correlation coefficients, similarly to the previously mentioned articles, whereas the emphasis of our analysis is put particularly on nodes that are highly ranked in different measures. To some extent our analysis is related to the problem of identifying networks for which different centrality measures generate the same ranking, which remains largely unexplored for non--tree networks. For degree and closeness centrality \cite{Konigetal2014} show that this is indeed the case for nested--split graphs. Finally, there is a broad connection with the literature related to the axiomatization of centrality measures, although not decay centrality per se, and other related problems \citep[see][]{DequiedtZenou2014,PalaciosHuertaVolij2004}.
\section{Notation and Analytical Results}
Consider a set of nodes $N$, with cardinality $n$, which are connected through a network. A network is represented by a family of sets $\mathcal{N}:=\{N_i\subseteq N\ |\ i=1,\dots,n\}$, with $N_i$ denoting the set of nodes that are directly connected with $i$. $N_i$ is called $i$'s neighborhood and has cardinality, $|N_i|$. We focus on undirected networks, where $j \in N_i$ if and only if $i \in N_j$. A path in a network between nodes $i$ and $j$ is a sequence $i_1,...,i_K$ such that $i_1=i$, $i_K=j$ and $i_{k+1} \in N_{i_k}$ for $k=1,...,K-1$. The \emph{geodesic distance}, $d(i,j)$, between two nodes is the length of the shortest path between them. Two nodes are connected if there exists a path between them. The network is connected if every pair of nodes is connected. Our analysis focuses on connected networks.
For a given network $\mathcal{N}$, a \emph{centrality measure} is a function that maps from the set of nodes to the real numbers, i.e. ${\bf c}: N \to \mathbb{R}$, where $c_i$ is the centrality of node $i$ in the network $\mathcal{N}$.
The \emph{degree centrality} (or simply \emph{degree}) maps from each node to the cardinality of its neighborhood, i.e. $D_i=|N_i|$. We also define the set of nodes with maximum degree, i.e. $I_{deg}=\arg \max\limits_{i \in N} D_i$. Similarly to this, denote by $D_i^l$ the number of nodes that have geodesic distance $l$ from node $i$, i.e. $D_i^l=|N_i^l|$ where $N_i^l=\{j \in N\ |\ d(i,j)=l\}$. The superscript $l$ will often be omitted when referring to $l=1$ (degree). Also, for connected networks it holds by definition that $\sum\limits_{l=1}^{n-1}N_i^l=n-1$.
The {\em closeness centrality} (or simply {\em closeness}) maps from each node to the inverse of the sum of the geodesic distances from each other agent in the network, i.e. $C_i=\frac{1}{\sum\limits_{j \neq i} d(i,j)}$. Notice that closeness centrality measures how easily a node can reach all other nodes in the network. According to this definition, we define the set of nodes with maximum closeness centrality, i.e. $I_{clos}=\arg \max\limits_{i \in N} C_i$.
For some of the analytical results we focus on the inverse of closeness, which is known as \emph{farness} and is defined as $F_i=\sum\limits_{j \neq i} d(i,j)$. This is essentially a measure of discentrality, as it measures how far a node is from all other nodes. Obviously the order of nodes with respect to closeness is the reverse of that with respect to farness, hence central nodes are going to be considered those that minimize farness.
Finally, given a decay parameter $\delta \in (0,1)$, \emph{decay centrality} is a measure that maps from each node to the sum of distances from each other node in the network, adjusted by a decay parameter that makes distant nodes count less than closer ones, i.e. $DC_i^{\delta}=\sum\limits_{j\neq i}\delta^{d(i,j)}$. As in the previous two cases, for each value of $\delta$, we define the set of nodes with maximum decay centrality, i.e. $I_{dc}^{\delta}=\arg \max\limits_{i \in N} DC_i^{\delta}$.
\begin{proposition}
\label{proplowdelta}
Exists $\underline{\delta}$ such that for all $\delta \in (0,\underline{\delta})$ and $i,j \in N$, if $D_i>D_j$ then $DC_i(\delta)>DC_j(\delta)$.
\end{proposition}
The proposition states that for low values of the decay parameter a node with higher degree than another will necessarily have higher decay centrality as well. This result implies immediately Corollary \ref{corlowdelta} which states that for sufficiently low values of the decay parameter the set of nodes with maximum of decay centrality will be a subset of the set of nodes with maximum degree.
\begin{corollary}
\label{corlowdelta}
Exists $\underline{\delta}$ such that for all $\delta \in (0,\underline{\delta})$ holds that $I_{dc}^{\delta}\subseteq I_{deg}$.
\end{corollary}
The reason why the two sets may not be identical is that two nodes with the same degree might have different decay centralities. When the degrees of two nodes are equal, differences arise from the factors corresponding to more distant nodes, which previously had a negligible effect. In fact, again for low values of $\delta$, between two nodes with equal degrees, higher decay centrality has the node with higher number of nodes at distance equal to 2, i.e. $D_i^2$. If these quantities are also equal then we move to $D_i^3$ etc. In order to state this result formally we need an additional definition.
\begin{definition}
Given two vectors of real numbers $\mathcal{A}_i=(A_i^1,A_i^2,\dots,A_i^{n-1})$ and $\mathcal{A}_j=(A_j^1,A_j^2,\dots,A_j^{n-1})$ the first vector is larger than the second one according to the \emph{lexicographic order}, i.e. $\mathcal{A}_i >_L \mathcal{A}_j$, if $A_i^l>A_j^l$ for the first $l$ for which $A_i^l$ and $A_j^l$ differ. The two vectors are equal according to the lexicographic order, i.e. $\mathcal{A}_i =_L \mathcal{A}_j$, if $A_i^l=A_j^l$ for all $l \in \{1,\dots,n-1\}$.
\end{definition}
It is apparent that two vectors that are equal according to the lexicographic order, i.e. $D_i^l=D_j^l$ for all $l$, will also have equal decay centralities. The following result covers the remaining cases.
\begin{proposition}
\label{proplowdeltageneral}
Exists $\underline{\delta}$ such that for all $\delta \in (0,\underline{\delta})$ and $i,j \in N$, if $\mathcal{D}_i >_L \mathcal{D}_j$ then $DC_i(\delta)>DC_j(\delta)$, where $\mathcal{D}_i=(D_i,D_i^2,\dots,D_i^{n-1})$
\end{proposition}
Propositions \ref{proplowdelta} and \ref{proplowdeltageneral} allow the comparison of decay centralities between any pair of nodes for low values of $\delta$. We proceed in a similar way to obtain the respective results for high values of $\delta$.
\begin{proposition}
\label{prophighdelta}
Exists $\overline{\delta}$ such that for all $\delta \in (\overline{\delta},1)$ and $i,j \in N$, if $C_i>C_j$ then $DC_i(\delta)>DC_j(\delta)$.
\end{proposition}
Proposition \ref{prophighdelta} establishes a relation between the order of nodes with respect to decay centrality and closeness for high values of the decay parameter. This also implies immediately Corollary \ref{corhighdelta} which states that for sufficiently high values of the decay parameter the set of nodes with maximum decay centrality will be a subset of the set of nodes with maximum closeness.
\begin{corollary}
\label{corhighdelta}
Exists $\overline{\delta}$ such that for all $\delta \in (\overline{\delta},1)$ holds that $I_{dc}^{\delta}\subseteq I_{clos}$.
\end{corollary}
Similarly to the previous case, Proposition \ref{prophighdelta} does not allow the comparison between nodes with equal closeness centralities. In order to do so it is necessary to define for each node the vector $\mathcal{F}_i=(F_i^1,F_i^2,\dots,F_i^{n-1})$, with typical element $F_i^k=(-1)^{k-1}\sum\limits_{l=k}^{n-1}\binom lk D_i^l$, where $0!=1$ by definition. Observe that $F_i^1=\sum\limits_{k \neq i} d(i,k)$, i.e. the \emph{farness} of node $i$. Note that, the rest of the terms of vector $\mathcal{F}$ are related to the derivatives of farness, which is a different interpretation compared to $\mathcal{D}$.\footnote{The alternating signs appear due to the following result: Let $f,g$ continuously differentiable functions in $\mathbb{R}$ such that $f(1)=g(1)$, $f^{(l)}(1)=g^{(l)}(1)$ for all $l \in \{1,\dots,k\}$ and $f^{(k+1)}(1)>g^{(k+1)}(1)$. Then, there is $\overline{x}$ such that for all $x \in (\overline{x},1)$ it holds that $f(x)>g(x)$ if $k$ is even and $f(x)<g(x)$ if $k$ is odd. $f^{(l)}$ denotes $l$-th order derivative of $f$.}
Based on this definition, one can also define the vector $\mathcal{C}_i=(C_i^1,C_i^2,\dots,C_i^{n-1})$, where $C_i^k=1/F_i^k$ if $F_i^k\neq0$ and $C_i^k=0$ if $F_i^k=0$ for all $k \in \{1,\dots,n-1\}$. Note that $C_i^1$ corresponds to the closeness centrality, thus the overscript will often be omitted, and the following equivalence relations hold always.
\begin{remark}
\label{remarkfarclos}
$\mathcal{F}_i >_L \mathcal{F}_j \Leftrightarrow \mathcal{C}_i <_L \mathcal{C}_j$ and $\mathcal{F}_i =_L \mathcal{F}_j \Leftrightarrow \mathcal{C}_i =_L \mathcal{C}_j$.
\end{remark}
It is important to underline that such an equivalence does not hold for all orders. For instance, it does not hold for the \emph{unsorted dominance order} that is defined and used later in the article.
Having said that, we are ready to provide a result that determines the complete order among nodes in terms of decay centrality.
\begin{proposition}
\label{prophighdeltageneral}
Exists $\overline{\delta}$ such that for all $\delta \in (\overline{\delta},1)$ and $i,j \in N$, if $\mathcal{C}_i >_L \mathcal{C}_j$ then $DC_i(\delta)>DC_j(\delta)$, where $\mathcal{C}_i=(C_i,C_i^2,\dots,C_i^{n-1})$
\end{proposition}
The results so far have established a clear relation between decay centrality, degree and closeness, in the two limits, which provides a more intuitive picture of the characteristics of nodes with high decay centrality for some values of the decay parameter. However, the established relations still leave unanswered what happens for intermediate values of $\delta$.
\begin{definition}
Given two vectors of real numbers $\mathcal{A}_i=(A_i^1,A_i^2,\dots,A_i^{n-1})$ and $\mathcal{A}_j=(A_j^1,A_j^2,\dots,A_j^{n-1})$ the first one is larger than the second one according to the \emph{unsorted dominance order}, i.e. $\mathcal{A}_i >_{UD} \mathcal{A}_j$, if $\sum\limits_{l=1}^k A_i^l\geq\sum\limits_{l=1}^k A_j^l$ for all $k \in \{1,\dots,n-1\}$ and at least one of the inequalities is strict.
\label{defunsorteddominance}
\end{definition}
We use the term \emph{unsorted} because the standard definition of dominance order considers vectors whose elements are already sorted in decreasing order, which is not the case here. Moreover, notice that \emph{unsorted dominance} is a partial order, thus it may not allow the comparison between all nodes. Finally, observe that, under certain assumptions on the vectors under comparison, the unsorted dominance order can be seen as a deterministic analog of \emph{first order stochastic dominance}.
Definition \ref{defunsorteddominance} allows us to establish the following result:
\begin{proposition}
For $i,j \in N$, if $\mathcal{D}_i >_{UD} \mathcal{D}_j$ then $DC_i(\delta)>DC_j(\delta)$ for all $\delta \in (0,1)$.
\label{propintermediatedeltadomdeg}
\end{proposition}
Despite not providing a complete order, the result grasps the important relation between degree and closeness in the calculation of decay centrality. On one hand a high degree is beneficial as it is included in all sums $\sum\limits_{l=1}^k D_i^l$, on the other hand, high closeness implies, though indirectly, that a higher number of nodes will be in a shorter distance from $i$, thus boosting all relevant sums upwards.
A similar result can be established using the relation between $\mathcal{F}_i$ and $\mathcal{F}_j$ as follows:
\begin{proposition}
For $i,j \in N$, if $\mathcal{F}_j >_{UD} \mathcal{F}_i$ then $DC_i(\delta)>DC_j(\delta)$ for all $\delta \in (0,1)$.
\label{propintermediatedeltadomfar}
\end{proposition}
Observe that the subscripts in the first relation are reversed compared to Proposition \ref{propintermediatedeltadomdeg}, meaning that lower values of the elements of $\mathcal{F}_i$ are associated with higher decay centrality. Additionally, as it has already been mentioned, there is no equivalence relation between $\mathcal{F}$ and $\mathcal{C}$ with respect to \emph{unsorted dominance}, thus the result needs to be stated using the vector associated with farness, rather than the one associated with closeness.
The next result, builds upon Propositions \ref{proplowdelta} and \ref{proplowdeltageneral}, as it provides several sufficient conditions that characterize nodes with high decay centrality for values of $\delta<1/2$. Namely,
\begin{proposition}
\label{propdeltalowerhalf}
Consider two distinct nodes $i,j \in N$ and let $A_1=D_i-D_j$ and $A_l=D_i^l-D_j^l$ for all $l \in \{2,\dots,n-1\}$. Then, if $A_1>0$ either of the following four conditions is sufficient to ensure that $DC_i(\delta)>DC_j(\delta)$ for all $\delta \in (0,1/2]$.
\begin{multicols}{2}
\begin{enumerate}
\item $2A_1 \geq (n-1)-D_j$
\item $4A_1+2A_2\geq(n-1)-(D_j+D_j^2)$
\item $A_1\geq \max\{|A_2|,\dots,|A_{n-1}|\}$
\item $A_1 \geq \max \left\{\left|\sum\limits_{l=1}^2A_l\right|,\dots,\left|\sum\limits_{l=1}^{n-2}A_l\right|\right\}$
\end{enumerate}
\end{multicols}
\end{proposition}
The essence of all conditions is that as long as node $i$ has sufficiently many more immediate neighbors than node $j$ (high $A_1$) then node $i$ has higher decay centrality not only very close to the limit, but for a large range of values of $\delta$. The first two conditions provide also a simple illustration of the higher relative importance of closer neighbors compared to more distant ones, which is essential for decay centrality. A closer look at the proof of Proposition \ref{propdeltalowerhalf} reveals that the conditions are relatively demanding and the bounds are not tight, which is to be expected given that they hold for any potential network structure.
The next result establishes similar conditions for high values of $\delta$, where high decay centrality is more closely associated with low farness. Namely,
\begin{proposition}
\label{propdeltaupperhalf}
Consider two distinct nodes $i,j \in N$ and let $B_1=F_i-F_j$ and $B_l=F_i^l-F_j^l$ for all $l \in \{2,\dots,n-1\}$. Then, if $B_1<0$ either of the following two conditions is sufficient to ensure that $DC_i(\delta)>DC_j(\delta)$ for all $\delta \in [1/2,1)$.
\begin{multicols}{2}
\begin{enumerate}
\item $|B_1|\geq \max\{|B_2|,\dots,|B_{n-1}|\}$
\item $|B_1| \geq \max \left\{\left|\sum\limits_{l=1}^2B_l\right|,\dots,\left|\sum\limits_{l=1}^{n-2}B_l\right|\right\}$
\end{enumerate}
\end{multicols}
\end{proposition}
Propositions \ref{propdeltalowerhalf} and \ref{propdeltaupperhalf} provide a connection between decay centrality, degree and closeness for values that extend away from the limits. An important observation is that $\delta=1/2$ is the maximum upper bound for which conditions 3 and 4 of Proposition \ref{propdeltalowerhalf} guarantee the order of decay centralities, given that the remaining conditions hold. Similarly, $\delta=1/2$ is also the minimum lower bound for which conditions 1 and 2 of Proposition \ref{propdeltaupperhalf} guarantee the order of decay centralities, given that the remaining conditions hold. This last observation provides a natural motivation for the rule--of-thumb that is proposed in the next section.
\section{Numerical Results}
The theoretical results have provided several conditions that support the relation between decay centrality, degree and closeness. Nevertheless, some of these conditions may not be easily satisfied in certain networks. For this reason, we perform an extensive set of simulations that intends to identify how strong this relation is in general networks. In particular, we focus on analyzing the extent to which the set of nodes with maximum decay centrality, $I_{dc}^{\delta}$, coincides with either $I_{deg}$ or $I_{clos}$ (or both) for different values of the decay parameter $\delta$.
We simulate undirected \`Erdos-Renyi networks \citep{ErdosRenyi1959}, $G(n,p)$, where $n$ is the network size and $p$ is the link probability. The networks are required to be connected so that geodesic distances are well defined. We consider five network sizes, $n\in\{10,20,50,100,200\}$ and ten link probabilities $p\in\{0.05,0.1,\dots,0.45,0.5\}$ and perform 10000 trials for each combination $(n,p)$.\footnote{The main body of the article contains representative figure and tables. The Online Appendix contains results for $(n,p)=(1000,0.05)$, some additional figures and some extensions. Additional figures are available upon request.}
The first question is how often $I_{dc}^{\delta}$, i.e. the set of nodes with maximum decay centrality, intersects with either $I_{deg}$, i.e. the set of nodes with maximum degree, or with $I_{clos}$, i.e. the set of nodes with maximum closeness. We find that in the vast majority of the cases $I_{dc}^{\delta} \subseteq \left(I_{deg} \cup I_{clos}\right)$ for almost all values of $\delta$ and not only for the limit values.
It is important to mention that $I_{deg}$ intersects with $I_{clos}$ quite often for random networks. The reasons why this occurs are outside the scope of this paper, however it has an apparent effect on our results as it provides a natural connection between the two limit cases explored by theory. In fact, in ``almost all'' cases where there are nodes that belong both to $I_{deg}$ and $I_{clos}$, those nodes also belong to $I_{dc}^{\delta}$. This result cannot be generalized as there are cases in which, for some values of $\delta$, the nodes with maximum decay centrality do not belong to either $I_{deg}$ or $I_{clos}$. Nevertheless, as it becomes apparent from Tables~\ref{table::coinciding} and~\ref{table::coincidingandintermediate} the frequency with which such cases arise is practically negligible.
\begin{table}[h!]
\centering
\begin{tabular}{| c || c | c | c | c | c | c | c | c | c | c |}
\hline
\backslashbox{n}{p} & 0.05 & 0.10 & 0.15 & 0.20 & 0.25 & 0.30 & 0.35 & 0.40 & 0.45 & 0.50 \\
\hline\hline
10 & 8620 & 8756 & 8951 & 9035 & 9341 & 9570 & 9752 & 9879 & 9971 & 9987\\
\hline
20 & 6600 & 7229 & 8113 & 8711 & 9349 & 9752 & 9969 & 9998 & 10000 & 10000\\
\hline
50 & 5921 & 7620 & 8220 & 9330 & 9964 & 10000 & 10000 & 10000 & 10000 & 10000\\
\hline
100 & 6783 & 7401 & 8794 & 9989 & 10000 & 10000 & 10000 & 10000 & 10000 & 10000\\
\hline
200 & 7209 & 7593 & 9976 & 10000 & 10000 & 10000 & 10000 & 10000 & 10000 & 10000\\
\hline
\end{tabular}
\caption{Frequency of occassions where $I_{deg} \cap I_{clos} \neq \emptyset$.}
\label{table::coinciding}
\end{table}
\begin{table}[h!]
\centering
\begin{tabular}{| c || c | c | c | c | c | c | c | c | c | c |}
\hline
\backslashbox{n}{p} & 0.05 & 0.10 & 0.15 & 0.20 & 0.25 & 0.30 & 0.35 & 0.40 & 0.45 & 0.50 \\
\hline\hline
10 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\
\hline
20 & 11 & 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\
\hline
50 & 30 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\
\hline
100 & 14 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\
\hline
200 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\
\hline
\end{tabular}
\caption{Frequency of occassions where $I_{deg} \cap I_{clos} \neq \emptyset$ and $I_{dc}^{\delta} \nsubseteq (I_{deg} \cap I_{clos})$ for some value of $\delta$.}
\label{table::coincidingandintermediate}
\end{table}
We turn our attention to the case where $I_{deg}$ and $I_{clos}$ do not intersect. In this case, we expect from theory a transition in the nodes that belong to $I_{dc}^{\delta}$ as $\delta$ increases. It turns out that most of the times the transition is immediate, meaning that any node in $I_{dc}^{\delta}$ belongs to either $I_{deg}$ or $I_{clos}$. This can become apparent in Figure~\ref{fig::frequency}, which contains the percentage frequencies with which $I_{dec}^{\delta} \subseteq I_{deg}$, $I_{dec}^{\delta} \subseteq I_{clos}$ and $I_{dc}^{\delta} \cap \left(I_{deg}\cup I_{clos}\right)=\emptyset$. For each value of $\delta$ the frequencies correspond to the fraction of the simulated networks in which each of the three conditions held true. The fact that in the left subfigure $I_{deg}$ and $I_{clos}$ can intersect means that the sum of the frequencies may exceed 100\%, which seems to be the case rather often. This is no longer possible in the right subfigure, where we include only the cases where $I_{deg}$ and $I_{clos}$ do not intersect, therefore the three conditions are mutually exclusive. Note that, observing a node that maximizes decay centrality without belonging either to $I_{deg}$ or $I_{clos}$ is never observed in more than a 2\% of the trials, with the percentage becoming much lower away from $\delta=0.5$. This suggests that for most networks there is a threshold value of $\delta$ below which $I_{dec}^{\delta} \subseteq I_{deg}$ and above which $I_{dec}^{\delta} \subseteq I_{clos}$.
\begin{figure}[htbp]
\centering
\makebox[\linewidth][c]{%
\begin{subfigure}[b]{.54\textwidth}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{onehist_clean.jpg}
}
\caption{Including cases where $I_{deg} \cap I_{clos} \neq \emptyset$}
\end{subfigure}
\begin{subfigure}[b]{.54\textwidth}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{onehist_clean_noequal.jpg}
}
\caption{Excluding cases where $I_{deg} \cap I_{clos} \neq \emptyset$}
\end{subfigure}
}\\
\caption{The dotted (dashed) line shows the frequency with which $I_{dec}^{\delta} \subseteq I_{deg}$ ($I_{dec}^{\delta} \subseteq I_{clos}$), whereas the solid line shows the frequency with which it does not belong to any of the two sets. }
\label{fig::frequency}
\end{figure}
Figures~\ref{fig::histpernet_prob05_to10} and \ref{fig::histpernet_prob15_to20} contain the three percentage frequencies of interest for all network sizes and values of $p \in \{0.05, 0.1, 0.15, 0.2\}$, focusing on networks where $I_{deg}$ and $I_{clos}$ do not intersect. The results are qualitatively similar in all configurations, presenting an inverted S-shaped curve for the frequency of $I_{dec}^{\delta} \subseteq I_{deg}$, an S-shaped curve for $I_{dec}^{\delta} \subseteq I_{clos}$ and an inverted bell curve for $I_{dc}^{\delta} \cap \left(I_{deg}\cup I_{clos}\right)=\emptyset$.
\begin{figure}[htbp]
\vspace{-1cm}
\centering
\makebox[\linewidth][c]{%
\begin{subfigure}[b]{.58\textwidth}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{histpernet_clean_noequal_degree_p05.jpg}
}
\end{subfigure}
\begin{subfigure}[b]{.58\textwidth}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{histpernet_clean_noequal_degree_p10.jpg}
}
\end{subfigure}
}\\
\vspace{-0.3cm}
\makebox[\linewidth][c]{%
\begin{subfigure}[b]{.58\textwidth}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{histpernet_clean_noequal_intermediate_p05.jpg}
}
\end{subfigure}
\begin{subfigure}[b]{.58\textwidth}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{histpernet_clean_noequal_intermediate_p10.jpg}
}
\end{subfigure}
}\\
\vspace{-0.3cm}
\makebox[\linewidth][c]{%
\begin{subfigure}[b]{.58\textwidth}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{histpernet_clean_noequal_closeness_p05.jpg}
}
\end{subfigure}
\begin{subfigure}[b]{.58\textwidth}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{histpernet_clean_noequal_closeness_p10.jpg}
}
\end{subfigure}
}
\caption{Percentage frequency of networks for which $I_{dec}^{\delta} \subseteq I_{deg}$ (first row), $I_{dc}^{\delta} \cap \left(I_{deg}\cup I_{clos}\right)=\emptyset$ (second row) and $I_{dec}^{\delta} \subseteq I_{clos}$ (third row) for each value of the decay parameter, presented separately for each network size. The two columns correspond to $p=0.05$ (left) and $p=0.1$ (right).}
\label{fig::histpernet_prob05_to10}
\end{figure}
\begin{figure}[htbp]
\vspace{-1cm}
\centering
\makebox[\linewidth][c]{%
\begin{subfigure}[b]{.58\textwidth}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{histpernet_clean_noequal_degree_p15.jpg}
}
\end{subfigure}
\begin{subfigure}[b]{.58\textwidth}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{histpernet_clean_noequal_degree_p20.jpg}
}
\end{subfigure}
}\\
\vspace{-0.3cm}
\makebox[\linewidth][c]{%
\begin{subfigure}[b]{.58\textwidth}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{histpernet_clean_noequal_intermediate_p15.jpg}
}
\end{subfigure}
\begin{subfigure}[b]{.58\textwidth}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{histpernet_clean_noequal_intermediate_p20.jpg}
}
\end{subfigure}
}\\
\vspace{-0.3cm}
\makebox[\linewidth][c]{%
\begin{subfigure}[b]{.58\textwidth}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{histpernet_clean_noequal_closeness_p15.jpg}
}
\end{subfigure}
\begin{subfigure}[b]{.58\textwidth}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{histpernet_clean_noequal_closeness_p20.jpg}
}
\end{subfigure}
}
\caption{Percentage frequency of networks for which $I_{dec}^{\delta} \subseteq I_{deg}$ (first row), $I_{dc}^{\delta} \cap \left(I_{deg}\cup I_{clos}\right)=\emptyset$ (second row) and $I_{dec}^{\delta} \subseteq I_{clos}$ (third row) for each value of the decay parameter, presented separately for each network size. The two columns correspond to $p=0.15$ (left) and $p=0.2$ (right).}
\label{fig::histpernet_prob15_to20}
\end{figure}
\noindent The latter one never exceeds 10\%, with this being the case only for $p=0.05$ and values of $\delta$ close to 0.5. We also observe the transition of maximizers of decay centrality from belonging to the set of nodes with maximum degree to that of nodes with maximum closeness occurs for lower values of $\delta$ as networks become larger, with the result being more prevalent for low values of $p$; as $p$ increases we observe a sharp transition occurring for $\delta$ around 0.5.\footnote{For $p\geq0.2$ there are very few observations where $I_{deg}$ and $I_{clos}$ do not intersect.}
At this point one might wonder what is the rank in terms of decay centrality of nodes with maximum degree and closeness, when they are not ranked first? Figures~\ref{fig::rank_degree_p05_allnets} and~\ref{fig::rank_closeness_p05_allnets} show the average rank, as well as the 5th and 95th percentiles of rank distribution in decay centrality of nodes with maximum degree and closeness respectively. It turns out that nodes belonging to $I_{deg}$ or $I_{clos}$ are highly ranked in terms of decay centrality for all values of $\delta$, even when they are not ranked first. The result is similar if we exclude the networks in which $I_{deg}$ and $I_{clos}$ intersect.
\begin{figure}[htbp]
\vspace{-1.1cm}
\centering
\begin{subfigure}[b]{0.63\textwidth}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{rank_degree_p05_n50.jpg}
}
\end{subfigure}
\centering
\begin{subfigure}[b]{0.63\textwidth}
\vspace{-0.6cm}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{rank_degree_p05_n100.jpg}
}
\end{subfigure}
\begin{subfigure}[b]{0.63\textwidth}
\vspace{-0.6cm}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{rank_degree_p05_n200.jpg}
}
\end{subfigure}
\vspace{-0.3cm}
\caption{Average rank (solid), 5th (dotted) and 95th (dashed) percentiles of decay centrality of nodes with maximum degree. Subfigures correspond to $n=50, 100, 200$ from top to bottom and $p=0.05$.}
\label{fig::rank_degree_p05_allnets}
\end{figure}
\begin{figure}[htbp]
\vspace{-1.1cm}
\centering
\begin{subfigure}[b]{0.63\textwidth}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{rank_closeness_p05_n50.jpg}
}
\end{subfigure}
\centering
\begin{subfigure}[b]{0.63\textwidth}
\vspace{-0.6cm}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{rank_closeness_p05_n100.jpg}
}
\end{subfigure}
\begin{subfigure}[b]{0.63\textwidth}
\vspace{-0.6cm}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{rank_closeness_p05_n200.jpg}
}
\end{subfigure}
\vspace{-0.3cm}
\caption{Average rank (solid), 5th (dotted) and 95th (dashed) percentiles of decay centrality of nodes with maximum closeness. Subfigures correspond to $n=50, 100, 200$ from top to bottom and $p=0.05$.}
\label{fig::rank_closeness_p05_allnets}
\end{figure}
Nevertheless, focusing only on degree or only on closeness leads to increasingly suboptimal choices as $\delta$ moves towards the extremes. Hence, it remains unclear which of the two sets would provide better candidates depending on the value of $\delta$. Ideally, we would like to have a simple rule of thumb that would facilitate this choice. A natural rule would be to choose a node from $I_{deg}$ for $\delta<0.5$ and a node from $I_{clos}$ for $\delta>0.5$. Given that the two sets usually contain few nodes, it should not be too costly to calculate the decay centrality of each of these nodes and pick the one that maximizes it. It turns out that this rule of thumb is sufficient to ensure that the chosen node will be highly ranked in terms of decay centrality. Figure~\ref{fig::rank_combined_p05_allnets} shows the same statistics as Figures \ref{fig::rank_degree_p05_allnets} and \ref{fig::rank_closeness_p05_allnets} for $p=0.05$ and all network sizes, in which it can be seen that in all networks a node chosen according to this rule will be ranked in terms of decay centrality among the top three with probability 95\%. Additional figures in the Online Appendix suggest that the result is robust to different parameter values and even for networks with 1000 nodes.
\begin{figure}[htbp]
\vspace{-1.2cm}
\centering
\begin{subfigure}[b]{0.625\textwidth}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{rank_combined_p05_n50.jpg}
}
\end{subfigure}
\centering
\begin{subfigure}[b]{0.625\textwidth}
\vspace{-0.65cm}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{rank_combined_p05_n100.jpg}
}
\end{subfigure}
\begin{subfigure}[b]{0.625\textwidth}
\vspace{-0.65cm}
\centering
\resizebox{\linewidth}{!}{
\includegraphics{rank_combined_p05_n200.jpg}
}
\end{subfigure}
\vspace{-0.4cm}
\caption{Average rank (solid), 5th (dotted) and 95th (dashed) percentiles, of decay centrality of nodes with maximum degree for $\delta<0.5$ and maximum closeness for $\delta>0.5$. Subfigures correspond to $n=50, 100, 200$ from top to bottom and $p=0.05$.}
\label{fig::rank_combined_p05_allnets}
\end{figure}
\section{Discussion}
We have established, both analytically and numerically, a relation between decay centrality, degree and closeness, showing that nodes that maximize one of the two measures are natural candidates for maximizing decay centrality as well. In fact, the majority of networks has a threshold value of $\delta$ below which maximum decay centrality coincides with maximum degree and above which it coincides with maximum closeness. We show that a simple rule of thumb that considers a common threshold at $\delta=0.5$ performs particularly well. It still remains rather unexplored whether there are particular characteristics of the network that can allow the accurate characterization of this threshold value with some accuracy. Finally, simulations are limited to networks of small to medium size. In the Online Appendix we provide some additional support in favor of our results by considering networks with 1000 nodes, however a more systematic extension of the analysis to large networks would ensure their applicability to problems where decay centrality has been shown to play an important role.
\addcontentsline{toc}{section}{References}
|
\section{Introduction}
The pioneering work on the separability of the Teukolsky master equation on the C\,--\,metric background has been done in \cite{Prestidge1998} and generalized to the rotating case in \cite{Bini2008}. In the preceding paper \cite{KofSep1} we independently tackled the same problem and provided a separation for the ``extreme'' Newman\,--\,Penrose (NP) field components on the nonrotating C\,--\,metric background. To incorporate rotation in the C\,--\,metric is a natural generalization as some of the relativistic effects are present in rotating solutions only. As compared with \cite{Prestidge1998,Bini2008} we use a more convenient of the C\--\,metric: (i) an explicit factorization of the structure function, which has been provided only recently in \cite{ht2}; (ii) asymptotically ``nonrotating'' form of the metric, as discussed in \cite{BiKofAcc} and (iii) the canonical coordinates; and therefore we obtain simpler results.
For the Kerr solution the magnetic and electric field becomes increasingly expelled from the horizon itself as the Kerr black hole becomes more extremal -- this is known as the Meissner effect, see e.g. \cite{Penna} for recent review or \cite{BicakJanis,DKL} for calculations on misaligned fields. This effect has been studied not only in the test field approximation, but through full exact solutions as well \cite{BicakLedvinka,KarasBudinova}.
Although the Meissner effect is known for a long time its origins have not been still explained satisfactorily. Suggestions as the infinite distance to the horizon (in the extremal case) cannot cause the Meissner effect. The infinite distance is universal property meanwhile only the axially symmetric field are subject to the Meissner effect.
The rotating C\,--\,metric is a special case of the boost\,--\,rotation symmetric solutions of Einstein (--Maxwell) field equations. At the same time it is a member of Pleba\'{n}ski\,--\,Demia\'{n}sky class of solutions \cite{pd}; therefore of algebraic type D. The C\,--\,metric represents two ``uniformly accelerated charged and rotating black holes''.
In type D spacetimes the equations for radiative (ingoing and outgoing radiation) NP field components of (a) massless Klein-Gordon field ($s=0$), (b) neutrino field ($s=\nicefrac{1}{2}$), (c) test Maxwell field ($s=1$), (d) Rarita\,--\,Schwinger field ($s=\nicefrac{3}{2}$) and (e) linear gravitational perturbations ($s=2$) can be decoupled \cite{Teuk,RS} and the Debye potential, i.e. a single scalar field from which all the field components can be generated, for them can be found \cite{Keg1,Keg2}.
We present these equations in Geroch\,--\,Held\,--\,Penrose (GHP) formalism \cite{GHP,RindlerPenrose1} which is not only succinct but provides a deeper geometrical insight than the NP formalism.
The background metric -- the rotating C\,--\,metric -- is presented in Section \ref{sec:CM}. For more details see, e.g., \cite{kw,ht2,gkp} which encompass comprehensive historical introduction. The null tetrad and corresponding spin coefficients are presented.
In section \ref{sec:meq} the Teukolsky master equation and the equation for the Debye potential are summarized together with the separability ansatz. The equation for separated radial and angular functions are exposed.
The asymptotic behaviour of the radial function close the outer black hole horizon is analyzed in section \ref{sec:radial}.
These general results are then used to investigate the static axially symmetric electromagnetic field in section \ref{sec:elmag}. Utilizing the Debye potential the field is easy to find. Computing the electric and magnetic flux through the horizon in the extremal case we prove that the Meissner effect works even for accelerated black holes.
\section{Charged rotating C\,--\,metric in canonical form}\label{sec:CM}
Using the signature $(-,\,+,\,+,\,+)$ the canonical form of the charged rotating C\,--\,metric is given by \cite{ht2,BiKofAcc}
\begin{widetext}
\begin{multline}
\vec{\mathrm{d}} s^2 = \frac{B^2}{A^2(x-y)^2} \; \biggl\{
\frac{\mathcal{G}(y)}{1+\left( aAxy \right)^2}\Bigl[ \left( 1+\left( aAx \right)^2 \right)\Ktau \,\vec{\mathrm{d}} \tau+aA\left( 1-x^2 \right)\Kphi \,\vec{\mathrm{d}}\varphi \Bigr]^2
-\frac{1+\left( aAxy \right)^2}{\mathcal{G}(y)}\,\vec{\mathrm{d}} y^2 \\
+\frac{1+\left( aAxy \right)^2}{\mathcal{G}(x)}\,\vec{\mathrm{d}} x^2
+\frac{\mathcal{G}(x)}{1+\left( aAxy \right)^2}\;\Bigl[ \left( 1+\left(aAy\right)^2 \right)\Kphi \,\vec{\mathrm{d}}\varphi-aA\left( 1-y^2 \right)\Ktau \,\vec{\mathrm{d}} \tau \Bigr]^2\biggr\}\,,
\label{eq:rCM}
\end{multline}
\end{widetext}
where the quartic structure function $\mathcal{G}(\xi)$ can be in physically interesting cases factorized (see \cite{ht2} for detail on this factorization) as
\begin{equation}
\mathcal{G}(\xi) = \left( 1-\xi^2 \right)\left( 1+\Rp A\xi \right)\left( 1+\Rm A\xi \right),
\label{eq:G}
\end{equation}
where
\begin{eqnarray}
\Rp &=& m+\sqrt{m^2-q^2-a^2}\,,\\
\Rm &=& m-\sqrt{m^2-q^2-a^2}\,.
\end{eqnarray}
The metric \re{rCM} is together with electromagnetic 4-potential
\begin{multline}
\fp = \frac{Bqy}{1+\left( aAxy \right)^2}\Bigl[ \left( 1+\left( aAx \right)^2 \right)\Ktau \, \vec{\mathrm{d}} \tau \\ + aA \left(1-x^2 \right)\Kphi \,\vec{\mathrm{d}}\varphi \Bigr] ,
\label{eq:r4p}
\end{multline}
a solution of the Einstein\,--\,Maxwell equations for $\tau\in \mathbb{R}$, $x\in \mathbb{R}$, $y\in\mathbb{R}$ and $\varphi \in \langle 0, \,2\pi \rangle$ in general.
We are interested in spacetime which can be interpreted as uniformly accelerated black hole with acceleration $A>0$. This can be achieved by defining the ranges of values $\Rp$ and $\Rm$, which are roots of the structure function $\mathcal{G}(\xi)$ rescaled by $A$ by
\begin{equation}
-\infty \leq -1/A\Rm \leq -1/A\Rp < -1 \,.
\end{equation}
Also the admissible values of coordinates are restricted to
\begin{align}
x & \in \langle -1,\, 1 \rangle\,, &
y & \in \langle -1/A\Rm,\, 1 \rangle\,, &
x-y & > 0\,,
\label{eq:range}
\end{align}
where $y=-1/A\Rm$ is the inner black hole horizon, $y=-1/A\Rp$ is the outer black hole horizon, $y=-1$ is acceleration horizon and $x-y=0$ is asymptotic infinity. In these coordinates the axis is divided in two parts by the black hole: $x=1$ corresponds to the part of axis connecting black hole horizon with the acceleration horizon and the part of axis where $x=-1$ connects black hole horizon with infinity.
The coordinate singularity at the acceleration horizon in \re{rCM} can be avoided in the global boost\,--\,rotation symmetric coordinates \cite{gkp}. Only in these global coordinates the second black hole appears after we proceed with the analytic continuation across the acceleration horizon.
The basic set of parameters, which enters the metric, is the acceleration $A$,the rotation $a$, and the position of outer $\Rp$, resp. inner $\Rm$, horizon. $\Rm$. But the mass and the charge parameters are encoded in the metric via relations $m=(\Rp+\Rm)/2$ and $q^2=\Rp\Rm-a^2$, as follows from \re{G}. Only three of these parameters are independent.
From the newly introduced dimensionless constants $\Kphi$, $\Ktau$ and $B$ only the $\Ktau$ explicitly demonstrates coordinate freedom. The other ones, $\Kphi$ and $B$, change the physical properties of the solution. The constant $B$ serves as a conformal factor which cannot be transformed out by a coordinate transformation and $\Kphi$ defines the conical singularities along the axis. We keep, however, even the $\Ktau$, because it is useful in demonstrating the ``visual symmetry'' of some equations.
For the sake of simplicity of our relations, let us also define function $\mathcal{H}(\xi)$, conformal factor $\Omega_0$ and constant $\Gamma$ as
\begin{align}
\mathcal{H}(\xi) &\equiv \frac{\mathcal{G}(\xi)}{1+\left( aAxy \right)^2}\,,&
\Omega_0 &= \frac{B}{A(x-y)}\,,&
\Gamma &= \sqrt{1+a^2\!A^2}\,.
\label{eq:H}
\end{align}
Moreover, we have to introduce a null tetrad $(\vec{l},\,\vec{n},\,\vec{m},\,\bar{\vec{m}})$ such that $l^an_a = -1$ and $m^a\bar{m}_a=1$ as follows
\begin{widetext}
\begin{align}
\vec{l} &= -\frac{\Omega_0}{\sqrt{2}}\left[
\left( 1+a^2\!A^2x^2 \right)\sqrt{-\varepsilon\mathcal{H}(y)}\, \Ktau \,\vec{\mathrm{d}}\tau
- \frac{1}{\sqrt{-\varepsilon\mathcal{H}(y)}}\;\vec{\mathrm{d}} y
+ aA\left( 1-x^2 \right)\sqrt{-\varepsilon\mathcal{H}(y)}\,\Kphi \,\vec{\mathrm{d}}\varphi \right], \label{eq:PND-l}\\
\vec{n} &= -\frac{\Omega_0}{\sqrt{2}}\left[
\left( 1+a^2\!A^2x^2 \right)\sqrt{-\varepsilon\mathcal{H}(y)}\, \Ktau \,\vec{\mathrm{d}}\tau
+ \frac{1}{\sqrt{-\varepsilon\mathcal{H}(y)}}\;\vec{\mathrm{d}} y
+ aA\left( 1-x^2 \right)\sqrt{-\varepsilon\mathcal{H}(y)}\,\Kphi \,\vec{\mathrm{d}}\varphi \right],
\label{eq:PND-n} \\
\vec{m} &=
\frac{\Omega_0}{\sqrt{2}}\, \left[
- iaA\left( 1-y^2 \right)\sqrt{\mathcal{H}(x)}\,\Ktau \,\vec{\mathrm{d}}\tau
- \frac{1}{\sqrt{\mathcal{H}(x)}}\;\vec{\mathrm{d}} x
+ i\left( 1+a^2\!A^2y^2 \right)\sqrt{\mathcal{H}(x)}\, \Kphi \,\vec{\mathrm{d}}\varphi \right], \label{eq:PND-m}
\end{align}
\end{widetext}
and $\vec{\bar{m}}$ is given as a complex conjugate of $\vec{m}$. This null tetrad definition is valid in the region between outer black hole horizon and acceleration horizon, where $\mathcal{G}(y)<0$ for $\varepsilon=1$. In the regions where $G(y)>0$ the parameter $\varepsilon$ has to be set to -1. Note that this does not affect the separability of the equation at all.
In this tetrad $\vec{l}$ and $\vec{n}$ are parallel to principal null directions, but the vector field itself are not tangent to affinely parameterized null congruences.
There exist two Killing vector fields
\begin{align}
\vec{\xi_{\tau}} &= \Ntau \, \frac{\vec{\partial}}{\vec{\partial \tau}}\,, &
\vec{\xi_\varphi} &= \Nphi \,\frac{\vec{\partial}}{\vec{\partial\varphi}}\,,
\label{eq:KV}
\end{align}
although the normalization of axial Killing vector field $\vec{\xi_\varphi}$ is quite obvious, this is not true for boost Killing vector field $\vec{\xi_\tau}$. Therefore, we leave both these constants in subsequent equations.
According to \cite{KrKuCKYPD}, there are also nontrivial conformal Killing\,--\,Yano tensors which we readapted to our metric and coordinates:
\begin{eqnarray}
\vec{k} &=& \phantom{-}\Omega_0\,\Big[\, aAxy\, \vec{l} \wedge \vec{n} -i \vec{m}\wedge\vec{\bar{m}}\, \Big]\,, \\
\star\vec{k} &=& -\Omega_0 \, \Big[\, \vec{l}\wedge \vec{n}+iaAxy\,\vec{m}\wedge \bar{\vec{m}}\, \Big]\,.
\label{eq:cky}
\end{eqnarray}
This directly leads to a conformal Killing tensor $\hat{Q}_{ab} = k_{ac}k^{c}_{\phantom{c}b} = \frac{1}{2}\left[ \left( aAxy \right)^2\,\vec{l}\otimes\vec{n}+\vec{m}\otimes\bar{\vec{m}} \right].$
The NP spin coefficients corresponding to the tetrad \re{PND-l}\,--\,\re{PND-m} are
\begin{align}
\pi & = \phantom{-}\frac{1}{\sqrt{2}}\,\frac{A}{B}\,\frac{\left( 1+iaAxy \right)\left( 1-iaAy^2 \right)\sqrt{\mathcal{G}(x)}}{\left( 1+(aAxy)^2 \right)^{\nicefrac{3}{2}}}\,, \\
\mu &= -\frac{1}{\sqrt{2}}\,\frac{A}{B}\,\frac{\left( 1+iaAxy \right)\left( 1-iaAx^2 \right)\sqrt{-\varepsilon\mathcal{G}(y)}}{\left( 1+(aAxy)^2 \right)^{\nicefrac{3}{2}}}\,, \\
\epsilon &= \frac{1}{4\sqrt{2}}\,\frac{A}{B}\, \frac{\left( 1-iaAxy \right)^2\left( x-y \right)^3}{\sqrt{1+(aAxy)^2}\sqrt{-\varepsilon\mathcal{G}(y)}} \nonumber\\
&\qquad\qquad\qquad\times \frac{\partial}{\partial y} \left( \frac{-\varepsilon\mathcal{G}(y)}{\left( 1-iaAxy \right)^2\left( x-y \right)^2} \right)\,, \\
\beta &= -\frac{1}{4\sqrt{2}}\,\frac{A}{B}\, \frac{\left( 1-iaAxy \right)^2\left( x-y \right)^3}{\sqrt{1+(aAxy)^2}\sqrt{\mathcal{G}(x)}} \nonumber\\
&\qquad\qquad\qquad\times \frac{\partial}{\partial x} \left( \frac{\mathcal{G}(x)}{\left( 1-iaAxy \right)^2\left( x-y \right)^2} \right)\,,
\label{eq:sc}
\end{align}
and then
\begin{align}
\pi &= -\tau\,, & \varrho &= \varepsilon\mu\,, & \epsilon &= \varepsilon\gamma\,, & \beta &= -\alpha\,, \\
\kappa &=0\,, & \sigma &= 0\,, & \nu &= 0\,, & \lambda &= 0\,.
\label{}
\end{align}
Finally, the only non-zero Weyl NP scalar is
\begin{align}
\psi_2 &= \frac{1}{12}\,\frac{A^2\left( x-y \right)^2}{B^2} \, \frac{\left( 1-iaAxy \right)^3}{\left( 1+(aAxy)^2 \right)}\times\\
&\left[ \frac{\partial^2}{\partial x^2}\left( \frac{\mathcal{G}(x)}{\left( 1-iaAxy \right)^3} \right) - \frac{\partial^2}{\partial y^2}\left( \frac{\mathcal{G}(y)}{\left( 1-iaAxy \right)^3} \right)\right]. \nonumber
\end{align}
\section{Master equation}\label{sec:meq}
Teukolsky \cite{Teuk} provided decoupled equations for NP components of gravitational perturbation $\Psi_0$ and $\Psi_4$, of test electromagnetic field $\Phi_0$ and $\Phi_2$ and neutrino field $\chi_0$ and $\chi_1$ in general type D spacetime. For the Rarita\,--\,Schwinger field this was done in \cite{RS}. In \cite{KofSep1} we summarized these equations using GHP formalism. However, we realized that it is not sufficient in general to denote the NP field components by its spin weight $s$ only, but it is necessary to take the spin $S$ of the field into account as well. Here we, therefore, rewrite the equations, using a little bit different notation. Namely, every field component is denoted as $\Phi_{[p,q]}^{(s,S)}$ where $p$ and $q$ are GHP weights, $s$ is the spin weight\footnote{Of course, $s=\nicefrac{1}{2}(p-q)$.} and $S$ is spin of the field, see Tab. \ref{tab:T1}.
\begin{table*}
\centering
\caption{Notation: spin a GHP weight of field components.}
\renewcommand{\arraystretch}{1.5}
\begin{ruledtabular}
\begin{tabular}{l|ccccccccc}
&
$\Psi_4$ & $\Sigma^{RS}_3$ &
$\Phi^{EM}_2$ & $\chi_1$ &
$\Phi^{KG}$ &
$\chi_0$ & $\Phi^{EM}_0$ &
$\Sigma^{RS}_0$ & $\Psi_0$ \\
$\Phi$ &
$\Phi^{(-2,2)}_{[-4,0]}$ & $\Phi^{(-\nicefrac{3}{2},\nicefrac{3}{2})}_{[-3,0]}$ &
$\Phi^{(-1,1)}_{[-2,0]}$ & $\Phi^{(-\nicefrac{1}{2},\nicefrac{1}{2})}_{[-1,0]}$ &
$ \Phi^{(0,0)}_{[0,0]}$ &
$\Phi^{(\nicefrac{1}{2},\nicefrac{1}{2})}_{[1,0]}$ &$\Phi^{(1,1)}_{[2,0]}$&
$\Phi^{(\nicefrac{3}{2},\nicefrac{3}{2})}_{[2,0]}$ &$\Phi^{(2,2)}_{[4,0]}$
\end{tabular}
\end{ruledtabular}
\label{tab:T1}
\end{table*}
In general the equation for $\Phi^{(S,S)}_{[2S,0]}$ field component reads
\begin{multline}
\bigl[ \left( \Th -\bar{\varrho}-2S\varrho \right)\left( \Th' -\varrho' \right) - \left( \Ed-\bar{\tau}'-2S\tau \right)\left( \Ed'-\tau' \right)\\ -(2S-1)(S-1)\psi_2 \bigr] \Phi^{(S,S)}_{[2S,0]} = 0 \,,
\label{eq:SWplus}
\end{multline}
with the separable ansatz\footnote{The function $\mathcal{X}$ should bear also indices $l,\,m,\,\omega$ as follows from the necessity to label the basis of the solutions of Sturm\,--\,Liouville problem. And so does $\mathcal{Y}$. We omit these indices -- the reader is kindly asked to imagine them anywhere $\mathcal{X}$ and $\mathcal{Y}$ is present.}
\begin{multline}
\hat{\Phi}^{(S,S)}_{[2S,0]} = e^{-i\omega\tau} e^{im\varphi} \left( x-y \right)^{1+S} \\
\times \left( 1-iaAxy \right)^{-S}
\mathcal{X}^{(S)}_{(S)}(x)\,\mathcal{Y}^{(-S)}_{(S)}(y)\,. \label{eq:sepSWp}
\end{multline}
For field component $\Phi^{(-S,S)}_{[-2S,0]}$ the equation is
\begin{multline}
\bigl[ \left( \Th'-\bar{\varrho}'-2S\varrho' \right)\left( \Th-\varrho \right)-\left( \Ed'-\bar{\tau}-2S\tau' \right)\left( \Ed-\tau \right)\\ -\left( 2S-1 \right)\left( S-1 \right)\psi_2 \bigr] \Phi^{(-S,S)}_{[-2S,0]} = 0\,,
\label{eq:SWminus}
\end{multline}
and the separable ansatz reads
\begin{multline}
\hat{\Phi}^{(-S,S)}_{[-2S,0]} = e^{-i\omega\tau} e^{im\varphi} \left( x-y \right)^{1+S}\\
\times \left( 1-iaAxy \right)^{-S}
\mathcal{X}^{(-S)}_{(S)}(x)\,\mathcal{Y}^{(S)}_{(S)}(y)\,. \label{eq:sepSWm}
\end{multline}
Observe that Eq. \re{SWminus} is just a primed version of Eq. \re{SWplus}.
In \cite{Keg1} and \cite{Keg2} Cohen and Kegeles provided Debye potentials for field of arbitrary spin (the discussion of $S=\nicefrac{3}{2}$ is done in \cite{RS2}). In a slightly modified notation ($s\rightarrow -S$ and complex conjugation), and rewritten in GHP formalism, this equation reads
\begin{multline}
\bigl[ \left( \Th'-\varrho' \right)\left( \Th+(2S-1)\bar{\varrho} \right)-\left( \Ed-\tau \right)\left( \Ed'+(2S-1)\bar{\tau} \right) \\
-\left( 2S-1 \right)\left( S-1 \right)\bar{\psi_2} \bigr] \hat{\psi}_{[0,-2S]}\equiv \square_S\hat{\psi}_{[0,-2S]} = 0\,.
\label{eq:DebP}
\end{multline}
which admits a separable ansatz
\begin{multline}
\hat{\psi}_{[0,-2S]} = e^{-i\omega\tau} e^{im\varphi} \left( x-y \right)^{-S+1} \\
\times \left( 1+iaAxy \right)^S \mathcal{Y}^{(S)}_{(S)}(y)\mathcal{X}^{(S)}_{(S)}(x) \,.\label{eq:sepDeb}
\end{multline}
The operator $\square_S$ defined in Eq. \re{DebP} is invariant under the composition of $^*$ and $'$ operation, i.e. $\square_S=\square_S^{\prime *}=\square_S^{*\prime}$. From this follows $\hat{\psi}_{[0,-2S]} = \pm (\hat{\psi}^{\prime *})_{[0,-2S]} = \pm (\hat{\psi}^{*\prime})_{[0,-2S]}$.
The equations \re{SWplus}, \re{SWminus} and \re{DebP} with ansatz \re{sepSWp}, \re{sepSWm} and \re{sepDeb} leads to separated equations for $\mathcal{X}^{(s)}_{(S)}$ and $\mathcal{Y}^{(s)}_{(S)}$ where $s=\pm S$ and the indices $l,\,m$ and $\omega$ are again omitted
\begin{widetext}
\begin{align}
\frac{\left[ \mathcal{G} \mathcal{X}^{(s)}_{(S),x} \right]_{,x}}{{\mathcal{X}^{(s)}_{(S)}}} + \frac{s^2\!+\frac{1}{2}}{3} \; \mathcal{G}_{,xx} + \Lambda^{(s)}_{(lm)} - \frac{\left( \left( 1+a^2\!A^2x^2 \right)\hat{m} -\frac{s}{2}\,\mathcal{G}_{,x} + aA\left( 1-x^2 \right)\hat{\omega} \right)^2}{\mathcal{G}} - 4aA\left( aA\hat{m}-\hat{\omega} \right)sx &= 0\,, \label{eq:gX} \\
\frac{\left[ \mathcal{G} \mathcal{Y}^{(s)}_{(lm),y} \right]_{,y}}{{\mathcal{Y}}^{(s)}_{(S)}} + \frac{s^2\!+\frac{1}{2}}{3} \; \mathcal{G}_{,yy} + \Lambda^{(s)}_{(S)} - \frac{\left(i \left( 1+a^2\!A^2y^2 \right)\hat{\omega} +\frac{s}{2}\,\mathcal{G}_{,y} -i aA\left( 1-y^2 \right)\hat{m} \right)^2}{\mathcal{G}} + 4iaA\left( aA\hat{\omega}+\hat{m} \right)sy &= 0\,, \label{eq:gY}
\end{align}
\end{widetext}
where
\begin{align}
\hat{m}&\equiv \frac{m}{K_\varphi\Gamma^2}\,,&
\hat{\omega}&=\frac{\omega}{\Ktau\Gamma^2}\,,
\end{align}
and the symbol $\mathcal{G}$ represents either $\mathcal{G}(x)$ or $\mathcal{G}(y)$; which one is clear from the context.
The equation for the Debye potential for scalar field ($S=0$) is in fact just a massless Klein\,--\,Gordon equation (complex conjugated). All the field components $\Phi^{(s,S)}_{[2s,0]}$ with $s=-S,\,-S+1,\,\dots,\,S-1,\,S$ are generated from the Debye potential $\hat{\psi}_{[0,2S]}$, see \cite{Keg1,Keg2} for neutrino field, Maxwell field and gravitational perturbations; for Rarita\,--\,Schwinger field see \cite{RS,RS2} (following equations can be proven to be equivalent with their results) and leads to
\begin{align}
\Psi_0 = \Phi^{(2,2)}_{[4,0]} &= \left( \Th-\bar{\varrho} \right)^3\left( \Th+3\bar{\varrho} \right)\hat{\psi}_{[0,-4]}\,,\nonumber\\
\Sigma^{RS}_0 = \Phi^{(\nicefrac{3}{2},\nicefrac{3}{2})}_{[3,0]} &= \left( \Th-\bar{\varrho} \right)^2\left( \Th+2\bar{\varrho} \right)\hat{\psi}_{[0,-3]}\,,\nonumber\\
\Phi_0 = \Phi^{(1,1)}_{[2,0]} &= \left( \Th -\bar{\varrho} \right)\left( \Th+\bar{\varrho} \right)\hat{\psi}_{[0,-2]}\,, \nonumber\\
\chi_0 = \Phi^{(\nicefrac{1}{2},\nicefrac{1}{2})}_{[1,0]} &= \Th \hat{\psi}_{[0,-1]}\,, \nonumber\\
\Phi_{KG} = \Phi^{(0,0)}_{[0,0]}&= \hat{\psi}_{[0,0]}\,,\nonumber\\
\chi_1 = \Phi^{(-\nicefrac{1}{2},\nicefrac{1}{2})}_{[-1,0]} &= \Ed' \, \hat{\psi}_{[0,-1]}\,, \nonumber\\
\Phi_2 = \Phi^{(-1,1)}_{[-2,0]} &= \left( \Ed' -\bar{\tau} \right)\left( \Ed' +\bar{\tau} \right)\hat{\psi}_{[0,-2]}\,, \nonumber\\
\end{align}
\begin{align}
\Sigma^{RS}_3 = \Phi^{(-\nicefrac{3}{2},\nicefrac{3}{2})}_{[-3,0]} &= \left( \Ed' -\bar{\tau} \right)^2\left( \Ed' +2\bar{\tau} \right)\hat{\psi}_{[0,-3]}\,, \nonumber\\
\Psi_4 = \Phi^{(-2,2)}_{[-4,0]} &= \left( \Ed'-\bar{\tau} \right)^3\left( \Ed'+3\bar{\tau} \right)\hat{\psi}_{[0,-4]}\,,\nonumber
\end{align}
or, in general
\begin{align}
\Phi^{(S,S)}_{[2S,0]} &= \left( \Th-\bar{\varrho} \right)^{2S-1}\left( \Th+(2S-1)\bar{\varrho} \right)\hat{\psi}_{[0,-2S]} \\
\Phi^{(-S,S)}_{[-2S,0]} &= \left( \Ed'-\bar{\tau} \right)^{2S-1}\left( \Ed'+(2S-1)\bar{\tau} \right)\hat{\psi}_{[0,-2S]}
\label{}
\end{align}
These equations can be beautifully simplified by application the following easy-to-prove identities
\begin{align}
\left( \Th-\bar{\varrho} \right)\left( \Th+w\bar{\varrho} \right) &=
\left( \Th+(w-1)\bar{\varrho} \right)\Th \,, \\
\left( \Ed'-\bar{\tau} \right)\left( \Ed'+w\bar{\tau} \right) &=
\left( \Ed'+(w-1)\bar{\tau} \right)\Ed' \,,
\label{}
\end{align}
to the form
\begin{align}
\Phi^{(S,S)}_{[2S,0]} &= \Th^{2S} \hat{\psi}_{[0,-2S]}\,,
& \Phi^{(-S,S)}_{[-2S,0]} &= \Ed^{\prime\, 2S} \hat{\psi}_{[0,-2S]} \,.
\label{eq:ExS}
\end{align}
Note that operators $\Ed$ and $\Ed'$ incorporate the derivatives with respect to $\tau$ (multiplication by $i\omega$), $\varphi$ (multiplication by $im$) and (the only nontrivial derivative) with respect to $x$, according to \re{PND-m}.
Furthermore, the following symmetry holds
\begin{equation}
\left(\Phi^{(s,S)}_{[2s,0]}\right)^{\prime *} = i^{2S} \Phi^{(-s,S)}_{[-2s,0]}\,,
\label{eq:}
\end{equation}
which expresses the results as a derivative of the Debye potential $\hat{\psi}_{[0,-2S]}$ and interchanges $\Th \leftrightarrow \Ed'$ in \re{ExS}.
\section{Radial function}\label{sec:radial}
The \emph{regular} singular points of the equation \re{gY} are
\begin{equation}
-\frac{1}{A\Rm},\,-\frac{1}{A\Rp},\,-1,\,1,\,\infty\,.
\end{equation}
The behavior of the two linearly independent solutions around the outer black hole horizon is governed by characteristic exponents of the regular singular point $y=-1/A\Rp$:
\begin{align}
e_1 &= \phantom{-}\frac{S}{2} - i\frac{a}{\Rp-\Rm}\,\left( \hat{m}-\frac{\Ntau}{\Nphi\Kphi}\,\frac{\hat{\omega}}{\Omega_H}\right)\,, \\
e_2 &= -\frac{S}{2} + i\frac{a}{\Rp-\Rm}\,\left( \hat{m}-\frac{\Ntau}{\Nphi\Kphi}\,\frac{\hat{\omega}}{\Omega_H}\right)\,,
\label{}
\end{align}
where the angular velocity of the outer horizon $\Omega_H$ is defined with respect to the Killing vector field $\vec{\xi_H} = \vec{\xi_\tau}+\Omega_H \vec{\xi_\varphi}$ and reads
\begin{equation}
\Omega_H = -\frac{a}{A}\,\frac{\Ktau \Ntau}{\Kphi\Nphi}\, \frac{a^2+\Rp^2}{1-A^2\Rp^2}\,.
\label{eq:AngVel}
\end{equation}
Then the solutions behave as
\begin{align}
\mathcal{Y}_1 & \sim \left(1+A\Rp y\right)^{e_1}\left( c_0 + c_1\left( 1+A\Rp y \right)+\dots \right),\\
\mathcal{Y}_2 & \sim \left(1+A\Rp y\right)^{e_2}\left( d_0 + d_1\left( 1+A\Rp y \right)+\dots \right) \,,
\label{}
\end{align}
if $e_1-e_2 \notin \mathbb{N}$, otherwise logarithmic terms will appear in $\mathcal{Y}_2$.
In the extremal case ($\Rp =\Rm$) the point $y=-1/A\Rp$ is in general \emph{irregular} singular point, and, thus, the asymptotic behaviour is difficult to investigate. But, in the case of static and axisymmetric configurations $y=-1/A\Rp$ becomes \emph{regular} singular point, and the exponents then depend on the eigenvalues $\Lambda^{(S)}_{(l0)}$, which we found in \cite{KofSep1} to be
\begin{equation}
\Lambda^{(S)}_{(l0)} = \left( 1-A^2\Rp^2 \right)\left[l\left( l+1 \right) + \frac{1}{3}\left( 1-S^2 \right)\right],
\end{equation}
and the calculations lead to the exponents\footnote{In \cite{KofSep1} the exponents were $(l-s,\,-l-s-1)$, but our current definition of the radial function $\mathcal{Y}$ differs from the one used in \cite{KofSep1} by the factor $\mathcal{G}^{s/2}(y)$.}
\begin{align}
e_1 & = l\,, & e_2 &= -l-1\,.
\label{eq:exponents}
\end{align}
In this extremal, static and axisymmetric case the Eq. \re{gY} can be solved exactly, one of the solution is simply polynomial, the second one is polynomial containing logarithms, see \cite{KofSep1}, where we also found the regular solution of the angular equation.
\section{Electromagnetic field --- the Meissner effect}\label{sec:elmag}
By reconstructing the self conjugated electromagnetic tensor $\vec{F}^*=\vec{F}+\frac{i}{2}\,\vec{\star F}$ from NP field components and in the null tetrad \re{PND-l}\,--\,\re{PND-m} we obtain
\begin{multline}
F^*_{x\varphi} = \Omega_0^2 \Kphi \Biggl[ i\left( 1+a^2A^2y^2 \right)\Phi_1\\ + \frac{a}{2}\,\frac{1-x^2}{\sqrt{\mathcal{G}(x)}}\,\sqrt{-\varepsilon \mathcal{G}(y)} \left( -\varepsilon^{-1} \Phi_0 + \Phi_2 \right) \Biggr]\,,
\label{eq:Flux}
\end{multline}
which represents the density of flux of ``magnetic + i $\times$ electric'' field.
The last NP component of electromagnetic field $\Phi_1$ was given in \cite{Keg1} as
\begin{equation}
2\Phi_1
= \Bigl[ \left( \Th+\varrho-\bar{\varrho} \right)\left( \Ed'+\bar{\tau} \right)
+\left( \Ed'+\tau'-\bar{\tau} \right)\left( \Th+\bar{\varrho} \right) \Bigr]\hat{\psi}_{[0,-2]}
\nonumber
\end{equation}
which can be again simplified. Thus, all the NP components of electromagnetic field in terms of the Debye potential $\hat{\psi}_{[0,-2]}$ read
\begin{align}
\Phi_0 & = \Th\Th\,\hat{\psi}_{[0,-2]} \label{eq:P0}\\
2\Phi_1
& = \left[ \left( \Th+\varrho \right)\Ed' +\left( \Ed'+\tau' \right)\Th \right]\hat{\psi}_{[0,-2]}\\
\Phi_2 & =\Ed'\Ed'\hat{\psi}_{[0,-2]} \,. \label{eq:P2}
\end{align}
Meissner effect affect static axially symmetric field configurations in the vicinity of rapidly rotating outer horizon. For the extremal case the radial equation can be solved explicitly as a rational function, see \cite{KofSep1}, but only the series expansion above the horizon is necessary\footnote{Here we denote the solution by all of the separation constant: $l$, $m=0$ and spin $S=1$ as $Y^{(1)}_{(1l0)}$.}:
\begin{equation}
Y^{(1)}_{(1l0)} \sim \left( 1+A\Rp y \right)^l\left( c_0+c_1\left( 1+A\Rp y \right)^2\dots \right)\,.
\label{eq:DebSol}
\end{equation}
Calculating the field components $\Phi_0$, $\Phi_1$ and $\Phi_2$ using \re{P0}\,--\,\re{P2} and evaluating the density of the flux \re{Flux} in the limit $y\rightarrow -1/A\Rp$ with the solution \re{DebSol} yields
\begin{equation}
\lim_{y\rightarrow -1/A\Rp} F^*_{x\varphi} = 0\,,
\label{eq:}\end{equation}
independently of the exact knowledge of the angular function $\mathcal{X}^{(1)}_{(1l0)}(x)$.
This means that there is no flux of static axially symmetric magnetic or electric field through the outer black hole horizon of extremely rotating uniformly accelerated black hole.
\section{Conclusions}
We have separated the Teukolsky master equation on a \emph{rotating C\,--\,metric} background for a test field of arbitrary spin $S$. We have separated also the equation for the Debye potential of these fields.
Utilizing the Debye potential for electromagnetic field we calculated the component $F^*_{x\varphi}$, which represents the density of flux through surfaces $y=\;$const and $\tau=\;$const, of a electromagnetic field tensor and we have showed that for extremely rotating C\,--\,metric this flux is zero in the case of axially symmetric field configurations.
We have reformulated some results on the Debye potential in GHP formalism by applying a unifying notation and by using commutation relations simplifying known results.
\begin{acknowledgements}
D.K. acknowledges the support from the Czech Science Foundation, Grant No. 14-37086G --- the Albert Einstein Centre. Moreover, D.K. would like to thank Prof. J. Bi\v{c}\'ak for introducing him to the C\,--\,metric, to Dr. M. Scholtz and Dr. Georgios Loukes-Gerakopoulos for inspiring discussions and comments on the manuscript.
We were unaware of the previous work of T. Prestidge \cite{Prestidge1998} and D. Bini, C. Cherubini and A. Geralico \cite{Bini2008} and therefore we would like to apologize for not paying the proper attention to their pioneering work in \cite{KofSep1}.
\end{acknowledgements}
\input{rs.bbl}
\end{document}
|
\section{Introduction}
This paper introduces a variation of Edelsbrunner-Harer nerves which are collections of Vorono\"{i} regions (called nucleus clusters) endowed with one or more proximity relations. Harer-Edelsbrunner nerves are introduced in~\cite[\S III.2, p. 59]{Edelsbrunner2010compTop}.
\setlength{\intextsep}{0pt}
\begin{wrapfigure}[8]{R}{0.35\textwidth}
\begin{minipage}{5.2 cm}
\centering
\includegraphics[width=20mm]{nucleusClusterN}
\caption[]{$\mbox{}$\\ Nucleus Cluster}
\label{fig:nucleusCluster}
\end{minipage}
\end{wrapfigure}
Vorono\"{i} tessellation has great utility and has many applications such as the creation of synthetic poly-crystals, computer graphics~\cite{Fukushige2007VoronoiTessellationPolygonVisibility}, geodesy~\cite{Gold2009VoronoiTessellationGeodesy}, non-parametric sampling~\cite{Villagran2016VoronoiTessellationSampling} and geometric modelling in physics, astrophysics, chemistry and biology~\cite{Qiang2010VoronoiTessellationApplications}. The form of clustering introduced in this article has proved to be important in the analysis of brain tissue~\cite{Peters2016brainTissueTessellation}, cortical activity and brain symmetries~\cite{Tozzi2016JNeuroSciSymmetries,Duyckaerts2000neuronTessellation} and capillary loss in skeletal and cardiac muscle~\cite{AlShammari2012VoronoiTessellationCapillaryLoss}. Vorono\"{i} nucleus clustering also has great utility in the study of digital images (see,{\em e.g.},~\cite[\S 1.13]{Peters2016ComputationalProximity},~\cite{Aiyeh2015TAMCSquality},~\cite{Wang2011VoronoiTessellationSegmentation}). The focus of this paper is not on the applications of MNCs, recently proved to be of great utility~\cite{Tozzi2016JNeuroSciSymmetries,Peters2016brainTissueTessellation}. Instead, the focus is on maximal nucleus clusters (MNCs) in proximity spaces and MNCs that are strongly proximal Edelsbrunner-Harer nerves. A proximity space setting for MNCs makes it possible to investigate the strong closeness of subsets in MNCs as well as the spatial and descriptive closeness of MNCs themselves.
\section{Preliminaries}
This section introduces the axioms for traditional as well as strong proximity spaces. Strong proximities were introduced in~\cite{Peters2015AMSJmanifolds}, elaborated in~\cite{Peters2016ComputationalProximity} (see, also,~\cite{Inan2015}) and are a direct result of earlier work on proximities~\cite{DiConcilio2006,DiConcilio2013mcs,Naimpally1970,Naimpally2009,Naimpally2013}.
\subsection{Spatial and Descriptive Lodato Proximity}
This section briefly introduces spatial and descriptive forms of proximity that provide a basis for two corresponding forms of strong Lodato proximity introduced in~\cite{Peters2015AMSJmanifolds} and axiomatized in~\cite{Peters2016ComputationalProximity}.
Let $X$ be a nonempty set. A \emph{Lodato proximity}~\cite{Lodato1962,Lodato1964,Lodato1966} $\delta$ is a relation on the family of sets $2^X$, which satisfies the following axioms for all subsets $A, B, C $ of $X$:\\
\begin{description}
\item[{\rm\bf (P0)}] $\emptyset \not\delta A, \forall A \subset X $.
\item[{\rm\bf (P1)}] $A\ \delta\ B \Leftrightarrow B \delta A$.
\item[{\rm\bf (P2)}] $A\ \cap\ B \neq \emptyset \Rightarrow A \near B$.
\item[{\rm\bf (P3)}] $A\ \delta\ (B \cup C) \Leftrightarrow A\ \delta\ B $ or $A\ \delta\ C$.
\item[{\rm\bf (P4)}] $A\ \delta\ B$ and $\{b\}\ \delta\ C$ for each $b \in B \ \Rightarrow A\ \delta\ C$. \qquad \textcolor{blue}{$\blacksquare$}
\end{description}
\vspace{3mm}
Further $\delta$ is \textit{separated }, if
\vspace{3mm}
\begin{description}
\item[{\rm\bf (P5)}] $\{x\}\ \delta\ \{y\} \Rightarrow x = y$. \qquad \textcolor{blue}{$\blacksquare$}
\end{description}
\noindent We can associate a topology with the space $(X, \delta)$ by considering as closed sets those sets that coincide with their own closure. For a nonempty set $A\subset X$, the closure of $A$ (denoted by $\mbox{cl} A$) is defined by,
\[
\mbox{cl} A = \{ x \in X: x\ \delta\ A\}.
\]
The descriptive proximity $\delta_{\Phi}$ was introduced in~\cite{Peters2012ams}. Let $A,B \subset X$ and let $\Phi(x)$ be a feature vector for $x\in X$, a nonempty set of non-abstract points such as picture points. $A\ \delta_{\Phi}\ B$ reads $A$ is descriptively near $B$, provided $\Phi(x) = \Phi(y)$ for at least one pair of points, $x\in A, y\in B$. From this, we obtain the description of a set and the descriptive intersection~\cite[\S 4.3, p. 84]{Naimpally2013} of $A$ and $B$ (denoted by $A\ \dcap\ B$) defined by
\begin{description}
\item[{\rm\bf ($\boldsymbol{\Phi}$)}] $\Phi(A) = \left\{\Phi(x)\in\mathbb{R}^n: x\in A\right\}$, set of feature vectors.
\item[{\rm\bf ($\boldsymbol{\dcap}$)}] $A\ \dcap\ B = \left\{x\in A\cup B: \Phi(x)\in \Phi(A)\ \mbox{and}\ \Phi(x)\in \Phi(B)\right\}$.
\qquad \textcolor{blue}{$\blacksquare$}
\end{description}
Then swapping out $\near$ with $\dnear$ in each of the Lodato axioms defines a descriptive Lodato proximity.
That is, a \textit{descriptive Lodato proximity $\dnear$} is a relation on the family of sets $2^X$, which satisfies the following axioms for all subsets $A, B, C $ of $X$.\\
\begin{description}
\item[{\rm\bf (dP0)}] $\emptyset\ \dfar\ A, \forall A \subset X $.
\item[{\rm\bf (dP1)}] $A\ \dnear\ B \Leftrightarrow B\ \dnear\ A$.
\item[{\rm\bf (dP2)}] $A\ \dcap\ B \neq \emptyset \Rightarrow\ A\ \dnear\ B$.
\item[{\rm\bf (dP3)}] $A\ \dnear\ (B \cup C) \Leftrightarrow A\ \dnear\ B $ or $A\ \dnear\ C$.
\item[{\rm\bf (dP4)}] $A\ \dnear\ B$ and $\{b\}\ \dnear\ C$ for each $b \in B \ \Rightarrow A\ \dnear\ C$. \qquad \textcolor{blue}{$\blacksquare$}
\end{description}
\vspace{3mm}
Further $\dnear$ is \textit{descriptively separated }, if
\vspace{3mm}
\begin{description}
\item[{\rm\bf (dP5)}] $\{x\}\ \dnear\ \{y\} \Rightarrow \Phi(x) = \Phi(y)$ ($x$ and $y$ have matching descriptions). \qquad \textcolor{blue}{$\blacksquare$}
\end{description}
\vspace{3mm}
\noindent The pair $\left( X,\dnear \right)$ is called a \emph{descriptive proximity space}. Unlike the Lodato Axiom (P2), the converse of the descriptive Lodato Axiom (dP2) also holds.
\begin{proposition}\label{prop:dnear}
Let $\left(X,\dnear\right)$ be a descriptive proximity space, $A,B\subset X$. Then $A\ \dnear\ B \Rightarrow A\ \dcap\ B\neq \emptyset$.
\end{proposition}
\begin{proof}
$A\ \dnear\ B \Leftrightarrow$ there is at least one $x\in A, y\in B$ such that $\Phi(x)=\Phi(y)$ (by definition of $A\ \dnear\ B$)\ Hence, $A\ \dcap\ B\neq \emptyset$.
\end{proof}
\subsection{Spatial and Descriptive Strong Proximities}
This section briefly introduces spatial strong proximity between nonempty sets and descriptive strong Lodato proximity.
Nonempty sets $A,B$ in a topological space $X$ equipped with the relation $\sn$, are \emph{strongly near} [\emph{strongly contacted}] (denoted $A\ \sn\ B$), provided the sets have at least one point in common. The strong contact relation $\sn$ was introduced in~\cite{Peters2015JangjeonMSstrongProximity} and axiomatized in~\cite{PetersGuadagni2015stronglyNear},~\cite[\S 6 Appendix]{Guadagni2015thesis}.
Let $X$ be a topological space, $A, B, C \subset X$ and $x \in X$. The relation $\sn$ on the family of subsets $2^X$ is a \emph{strong proximity}, provided it satisfies the following axioms.
\begin{description}
\item[{\rm\bf (snN0)}] $\emptyset\ \sfar\ A, \forall A \subset X $, and \ $X\ \sn\ A, \forall A \subset X$.
\item[{\rm\bf (snN1)}] $A \sn B \Leftrightarrow B \sn A$.
\item[{\rm\bf (snN2)}] $A\ \sn\ B$ implies $A\ \cap\ B\neq \emptyset$.
\item[{\rm\bf (snN3)}] If $\{B_i\}_{i \in I}$ is an arbitrary family of subsets of $X$ and $A \sn B_{i^*}$ for some $i^* \in I \ $ such that $\Int(B_{i^*})\neq \emptyset$, then $ \ A \sn (\bigcup_{i \in I} B_i)$
\item[{\rm\bf (snN4)}] $\mbox{int}A\ \cap\ \mbox{int} B \neq \emptyset \Rightarrow A\ \sn\ B$.
\qquad \textcolor{blue}{$\blacksquare$}
\end{description}
\noindent When we write $A\ \sn\ B$, we read $A$ is \emph{strongly near} $B$ ($A$ \emph{strongly contacts} $B$). The notation $A\ \notfar\ B$ reads $A$ is not strongly near $B$ ($A$ does not \emph{strongly contact} $B$). For each \emph{strong proximity} (\emph{strong contact}), we assume the following relations:
\begin{description}
\item[{\rm\bf (snN5)}] $x \in \Int (A) \Rightarrow x\ \sn\ A$
\item[{\rm\bf (snN6)}] $\{x\}\ \sn \{y\}\ \Leftrightarrow x=y$ \qquad \textcolor{blue}{$\blacksquare$}
\end{description}
For strong proximity of the nonempty intersection of interiors, we have that $A \sn B \Leftrightarrow \Int A \cap \Int B \neq \emptyset$ or either $A$ or $B$ is equal to $X$, provided $A$ and $B$ are not singletons; if $A = \{x\}$, then $x \in \Int(B)$, and if $B$ too is a singleton, then $x=y$. It turns out that if $A \subset X$ is an open set, then each point that belongs to $A$ is strongly near $A$. The bottom line is that strongly near sets always share points, which is another way of saying that sets with strong contact have nonempty intersection. Let $\near$ denote a traditional proximity relation~\cite{Naimpally1970}. \\
\begin{figure}[!ht]
\centering
\includegraphics[width=55mm]{nearSetsStrong}
\caption[]{Near Sets: $A = \left\{(x,0): 0.1\leq x\leq 1\right\},B = \left\{(x,sin(5/x)):0.1\leq x\leq 1\right\}$}
\label{fig:nearSetsStrong}
\end{figure}
$\mbox{}$\\
\vspace{2mm}
Next, consider a proximal form of a Sz\'{a}z relator~\cite{Szaz1987}. A \emph{proximal relator} $\mathscr{R}$ is a set of relations on a nonempty set $X$~\cite{Peters2016relator}. The pair $\left(X,\mathscr{R}\right)$ is a proximal relator space. The connection between $\sn$ and $\near$ is summarized in Prop.~\ref{thm:sn-implies-near}.
\begin{proposition}\label{thm:sn-implies-near}
Let $\left(X,\left\{\near,\dnear,\sn\right\}\right)$ be a proximal relator space, $A,B\subset X$. Then
\begin{compactenum}[1$^o$]
\item $A\ \sn\ B \Rightarrow A\ \near\ B$.
\item $A\ \sn\ B \Rightarrow A\ \dnear\ B$.
\end{compactenum}
\end{proposition}
\begin{proof}$\mbox{}$\\
1$^o$: From Axiom (snN2), $A\ \sn\ B$ implies $A\ \cap\ B\neq \emptyset$, which implies $A\ \near\ B$ (from Lodato Axiom (P2)).\\
2$^o$: From 1$^o$, there are $x\in A, y\in B$\ common to $A$ and $B$. Hence, $\Phi(x) = \Phi(y)$, which implies $A\ \dcap\ B\neq \emptyset$. Then, from the descriptive Lodato Axiom (dP2), $A\ \dcap\ B \neq \emptyset \Rightarrow\ A\ \dnear\ B$. This gives the desired result.
\end{proof}
\begin{example}
Let $X$ be a topological space endowed with the strong proximity $\sn$ and $A = \left\{(x,0): 0.1\leq x\leq 1\right\}$,$B = \left\{(x,sin(5/x)):0.1\leq x\leq 1\right\}$. In this case, $A,B$ represented by
Fig.~\ref{fig:nearSetsStrong} are strongly near sets with many points in common.
\qquad \textcolor{blue}{$\blacksquare$}
\end{example}
The descriptive strong proximity $\snd$ is the descriptive counterpart of $\sn$.
To obtain a \emph{descriptive strong Lodato proximity} (denoted by \emph{\bf dsn}), we swap out $\dnear$ in each of the descriptive Lodato axioms with the descriptive strong proximity $\snd$.
Let $X$ be a topological space, $A, B, C \subset X$ and $x \in X$. The relation $\snd$ on the family of subsets $2^X$ is a \emph{descriptive strong Lodato proximity}, provided it satisfies the following axioms.
\vspace{2mm}
\begin{description}
\item[{\rm\bf (dsnP0)}] $\emptyset\ {\sdfar}\ A, \forall A \subset X $, and \ $X\ \snd\ A, \forall A \subset X$.
\item[{\rm\bf (dsnP1)}] $A\ \snd\ B \Leftrightarrow B\ \snd\ A$.
\item[{\rm\bf (dsnP2)}] $A\ \snd\ B$ implies $A\ \dcap\ B\neq \emptyset$.
\item[{\rm\bf (dsnP4)}] $\mbox{int}A\ \dcap\ \mbox{int} B \neq \emptyset \Rightarrow A\ \snd\ B$.
\qquad \textcolor{blue}{$\blacksquare$}
\end{description}
\noindent When we write $A\ \snd\ B$, we read $A$ is \emph{descriptively strongly near} $B$.
For each \emph{descriptive strong proximity}, we assume the following relations:
\begin{description}
\item[{\rm\bf (dsnP5)}] $\Phi(x) \in \Phi(\Int (A)) \Rightarrow x\ \snd\ A$.
\item[{\rm\bf (dsnP6)}] $\{x\}\ \snd\ \{y\} \Leftrightarrow \Phi(x) = \Phi(y)$.
\qquad \textcolor{blue}{$\blacksquare$}
\end{description}
So, for example, if we take the strong proximity related to non-empty intersection of interiors, we have that $A\ \snd\ B \Leftrightarrow \Int A\ \dcap\ \Int B \neq \emptyset$ or either $A$ or $B$ is equal to $X$, provided $A$ and $B$ are not singletons; if $A = \{x\}$, then $\Phi(x) \in \Phi(\Int(B))$, and if $B$ is also a singleton, then $\Phi(x)=\Phi(y)$.
The connections between $\snd,\dnear$ are summarized in Prop.~\ref{thm:sn-implies-dnear}.
\begin{proposition}\label{thm:sn-implies-dnear}
Let $\left(X,\left\{\sn,\dnear,\snd\right\}\right)$ be a proximal relator space, $A,B\subset X$. Then
\begin{compactenum}[1$^o$]
\item For $A,B$ not equal to singletons, $A\ \snd\ B \Rightarrow \Int A\ \dcap\ \Int B\neq \emptyset \Rightarrow \Int A\ \dnear\ \Int B$.
\item $A\ \sn \ B \Rightarrow (\Int A\ \dcap\ \Int B)\ \snd \ B$.
\item $A\ \snd\ B \Rightarrow A\ \dnear\ B$.
\end{compactenum}
\end{proposition}
\begin{proof}$\mbox{}$\\
1$^o$:
$A\ \snd\ B$ implies that interior of $A$ is descriptively near the interior of $B$. Consequently, $\Int A\ \dcap\ \Int B\neq \emptyset$. Hence, from Axiom (dP2), $\Int A\ \dnear\ \Int B$.\\
1$^o\Rightarrow\ 2^o$.
3$^o$: Immediate from Axioms (dsnP2) and (dP2).
\end{proof}
\subsection{Vorono\"{i} regions}
Let $E$ be the Euclidean plane, $S\subset E$ (set of mesh generating points), $s\in S$. A Vorono\"{i} region (denoted by $V(s)$) is defined by
\[
V(s) = \left\{x\in E: \norm{x - s}\leq \norm{x - q}, \mbox{for all}\ q\in S\right\}\ \mbox{(Vorono\"{i} region)}.
\]
Let $X$ be a collection of Vorono\"{i} regions containing $N$, endowed with the strong proximity $\sn$. A nucleus mesh cluster (denoted by $\Cn\ N$) in a Vorono\"{i} tessellation is defined by
\[
\Cn\ N = \left\{A\in X: \cl\ A\ \sn\ N\right\}\ \mbox{(Vorono\"{i} mesh nucleus cluster)}.\\
\]
\begin{example}
A partial view of a Vorono\"{i} tessellation of a plane surface is shown in Fig.~\ref{fig:nucleusCluster}. The Vorono\"{i} region $N$ in this tessellation is the nucleus of a mesh cluster containing all of those polygons adjacent to $N$.
\qquad \textcolor{blue}{\Squaresteel}
\end{example}
A \emph{concrete} (\emph{physical}) set $A$ of points $p$ that are described by their location and physical characteristics, {\em e.g.}, gradient orientation (angle of the tangent to $p$. Let $\varphi(p)$ be the gradient orientation of $p$. For example, each point $p$ with coordinates $(x,y)$ in the concrete subset $A$ in the Euclidean plane is described by a feature vector of the form $(x,y,\varphi(p(x,y))$. Nonempty concrete sets $A$ and $B$ have descriptive strong proximity (denoted $A\ \snd\ B$), provided $A$ and $B$ have points with matching descriptions.
In a region-based, descriptive proximity extends to both abstract and concrete sets~\cite[\S 1.2]{Peters2016ComputationalProximity}. For example, every subset $A$ in the Euclidean plane has features such as area and diameter. Let $(x,y)$ be the coordinates of the centroid $m$ of $A$. Then $A$ is described by feature vector of the form $(x,y,area, diameter)$. Then regions $A,B$ have descriptive proximity (denoted $A\ \snd\ B$), provided $A$ and $B$ have matching descriptions.
The notion of strongly proximal regions extends to convex sets. A nonempty set $A$ is a \emph{convex set} (denoted $\cv A$), provided, for any pair of points $x,y\in A$, the line segment $\overline{xy}$ is also in $A$. The empty set $\emptyset$ and a one-element set $\left\{x\right\}$ are convex by definition. Let $\mathscr{F}$ be a family of convex sets. From the fact that the intersection of any two convex sets is convex~\cite[\S 2.1, Lemma A]{Edelsbrunner2014}, it follows that
\[
\mathop{\bigcap}\limits_{A\in\mathscr{F}} A\ \mbox{is a convex set}.
\]
Convex sets $\cv A, \cv B$ are strongly proximal (denote $\cv A \sn\ \cv B$), provided $\cv A, \cv B$ have points in common. Convex sets $\cv A, \cv B$ are descriptively strongly proximal (denoted $\cv A \snd\ \cv B$), provided $\cv A, \cv B$ have matching descriptions.
Let $X$ be a Vorono\"{i} tessellation of a plane surface equipped with the strong proximity $\sn$ and descriptive strong proximity $\snd$ and let $A,N\in X$ be Vorono\"{i} regions. The pair $\left(X,\left\{\sn,\snd\right\}\right)$ is an example of a proximal relator space~\cite{Peters2016relator}.
The two forms of nucleus clusters (ordinary nucleus cluster denoted by $\Cn$) and descriptive nucleus clusters are examples of mesh nerves~\cite[\S 1.10, pp. 29ff]{Peters2016ComputationalProximity}, defined by
\begin{align*}
\Cn N &= \left\{A\in X: A\ \sn\ N\right\}\ \mbox{(nucleus cluster)}.\\
\Cdn N &= \left\{A\in X: A\ \snd\ N\right\}\ \mbox{(descriptive nucleus cluster)}.
\end{align*}
A nucleus cluster is \emph{maximal} (denoted by $\Cmn N$), provided $N$ has the highest number of adjacent polygons in a tessellated surface (more than one maximal cluster in the same mesh is possible). Similarly, a descriptive nucleus cluster
is maximal (denoted by $\Cmdn N$), provided $N$ has the highest number of polygons in a tessellated surface descriptively near $N$, {\em i.e.}, the description of each $A\in \Cmdn N$ matches the description of nucleus $N$ and the number of polygons descriptively near $N$ is maximal (again, more than one $\Cmdn N$ is possible in a Vorono\"{i} tessellation).\\
\vspace{3mm}
\begin{figure}[!ht]
\centering
\includegraphics[width=65mm]{nucleusClustersN1N2N3}
\caption[]{ $\Cn\ N_1\ \sn\ \Cn\ N_2\ \mbox{and}\ \Cn\ N_1\ \snd\ \Cn\ N_2$}
\label{fig:N1N2N3maxClusters}
\end{figure}
\begin{example}$\mbox{}$\\
Let $X$ be the collection of Vorono\"{i} regions in a tessellation of a subset of the Euclidean plane shown in Fig.~\ref{fig:N1N2N3maxClusters} with nuclei $N_1,N_2,N_3\in X$. In addition, let $2^X$ be the family of all subsets of Vorono\"{i} regions in $X$ containing maximal nucleus clusters $\Cn N_1,\Cn N_2,\Cn N_3\in 2^X$ in the tessellation. Then, for example, $\Int\Cn N_2\ \cap\ \Int\Cn N_3\neq\emptyset$, since $\Cn N_2,\Cn N_3$ share Vorono\"{i} regions. Hence, $\Cn N_2\ \sn\ \Cn N_3\neq\emptyset$ (from Axiom (snN4)). Similarly,
$\Cn N_1\ \sn\ \Cn N_2$.
Let $\Phi(A)$ the description of a Vorono\"{i} equal the number of sides of $A\in X$. Since the nuclei $N_1,N_2,N_3$ have matching descriptions, $\Int\Cn N_1\ \dcap\ \Int\Cn N_2\neq \emptyset$. Consequently, $\Cn N_1\ \snd\ \Cn N_2$ (from Axiom (dsnP4)). Similarly, $\Cn N_1\ \snd\ \Cn N_3$ and $\Cn N_2\ \snd\ \Cn N_3$.
\qquad \textcolor{blue}{\Squaresteel}
\end{example}
\setlength{\intextsep}{0pt}
\begin{wrapfigure}[13]{R}{0.30\textwidth}
\begin{minipage}{3.8 cm}
\centering
\begin{pspicture}
(-0.5,-1.0)(3.5,2.5)
\providecommand{\PstPolygonNode}
\psdots[showpoints=true,linecolor=black](1;\INode)}
\PstPentagon[unit=0.75]
\psline[showpoints=true,linecolor=gray](-0.55,-0.83)(-1.10,-0.83)
\psline[showpoints=true,linecolor=gray](-0.55,-0.83)(-0.30,0.12)
\psline[showpoints=true,linecolor=gray](-1.10,-0.83)(-1.18,0.12)
\psline[showpoints=true,linecolor=gray](-0.75,1.50)(0.08,2.12)
\psline[showpoints=true,linecolor=gray](0.08,2.12)(0.38,1.82)
\psline[showpoints=true,linecolor=gray](0.38,1.82)(-0.00,1.00)
\rput(-1.05,0.80){\Large \textcolor{black}{$\boldsymbol{N}$}}
\rput(-0.20,1.50){\Large \textcolor{black}{$\boldsymbol{A_{_{1}}}$}}
\rput(-0.80,-0.30){\Large \textcolor{black}{$\boldsymbol{A_{_{2}}}$}}
\end{pspicture}
\caption[]{$\mbox{}$\\ {MNC Spokes}}
\label{fig:MNCnerve}
\end{minipage}
\end{wrapfigure}
\section{Main Results}
Homotopy types are introduced in~\cite[\S III.2]{Edelsbrunner1999} and lead to significant results for Vorono\"{i} maximal nucleus clusters.
Let $f,g:X\longrightarrow Y$ be two continuous maps. A \emph{homotopy} between $f$ and $g$ is a continuous map $H:X\times[0,1]\longrightarrow Y$ so that $H(x,0) = f(x)$ and $H(x,1) = g(x)$. The sets $X$ and $Y$ are \emph{homotopy equivalent}, provided there are continuous maps $f: X\longrightarrow Y$ and $g:Y\longrightarrow X$ such that $g\circ f \simeq \mbox{id}_X$ and $f\circ g \simeq \mbox{id}_Y$. This yields an equivalence relation $X\simeq Y$. In addition, $X$ and $Y$ have the same \emph{homotopy type}, provided $X$ and $Y$ are homotopy equivalent.
Let $\mathscr{F}$ be a finite collection of sets. An \emph{Edelsbrunner-Harer nerve} (denoted by $\mbox{Nrv}\ \mathscr{F}$) consists of all nonempty subcollections of $\mathscr{F}$ that have a nonvoid common intersection, {\em i.e.},
\[
\mbox{Nrv}\ \mathscr{F} = \left\{X\in \mathscr{F}: \bigcap X\neq \emptyset\right\}.
\]
Let $\mathscr{F}_{_{MNC}}$ be a collection of polygons in a Vorono\"{i} MNC endowed with the strong proximity $\sn$, $A$ be a Vorono\"{i} region in a MNC $\Cn N$ with nucleus $N$ and let subscollection $\mathscr{S}_A = \left\{A,N\right\}\in\Cn N$. The pair $\left(\mathscr{F}_{_{MNC}},\sn\right)$ is a proximity space. For each MNC $\Cn N$ endowed with $\sn$, the nucleus $N$ together with its adjacent polygons is a Vorono\"{i} structure (denoted by $\mbox{Nrv}\mathscr{F}_{_{MNC}}$) defined by
\[
\mbox{Nrv}\mathscr{F}_{_{MNC}} = \left\{\mathscr{S}\in \Cn N: \left(N,A\right)\in \sn\ \mbox{for}\ A\in \mathscr{S}\right\}\ \mbox{(MNC nerve)}.
\]
Each pair $\left(N,A\right)\in \sn$ in $\mbox{Nrv}\mathscr{F}_{_{MNC}}$ is called a \emph{\bf spoke} (denoted by $\mathscr{S}_A$), with a shape similar to the spoke in a wheel. A spoke contains a Vorono\"{i} region $A\in \Cn N$ that shares an edge with $N$. Hence, the $A\ \sn\ N$, {\em i.e.}, there is a strong proximity between the subsets in a spoke.
\begin{example}\label{ex:MNCnerve}
A pair of spokes $\mathscr{S}_{A_1},\mathscr{S}_{A_2}$ in a fragment of an MNC $\Cn N$ with nucleus $N$ is represented in Fig.~\ref{fig:MNCnerve}.
\qquad \textcolor{blue}{\Squaresteel}
\end{example}
Every MNC $\Cn N$ is a finite collection of closed convex sets in the Euclidean plane. Let $\Cn N$ be endowed with the strong proximity $\sn$. All non-nucleus polygons in $\Cn N$ share an edge with $N$. The collection of spokes $\mathscr{S}_A\in \Cn N$ each contain the nucleus $N$, which is common to all of the spokes, {\em i.e.}, the spokes in $\mbox{Nrv}\mathscr{F}_{_{MNC}}$ have a nonvoid common intersection. Let $\mathscr{S}_A,\mathscr{S}_{A'}$ be spokes in $\Cn N$ that share nucleus $N$. Consequently, $\Int\left(\mathscr{S}_A\right)\ \cap\ \Int\left(\mathscr{S}_{A'}\right) \neq\emptyset$ implies $\mathscr{S}_A\ \sn\ \mathscr{S}_{A'}$ (from Axiom (snN4)). Hence, $\mbox{Nrv}\mathscr{F}_{_{MNC}}$ is an Edelsbrunner-Harer nerve. From this, we obtain the result in Lemma~\ref{lem:MNCnerves}.
\begin{lemma}\label{lem:MNCnerves}
Let $\mathscr{F}_{_{MNC}}$ be a collection of polygons in a Vorono\"{i} MNC endowed with the strong proximity $\sn$.
The structure $\mbox{Nrv}\mathscr{F}_{mnc}$ is an Edelsbrunner-Harer nerve.
\end{lemma}
\begin{proof}
Let $\mathscr{S}_A,\mathscr{S}_{A'}$ be a pair of spokes in a maximal nucleus cluster MNC $\Cn N$. Since $\mathscr{S}_A\ \sn\ \mathscr{S}_{A'}$ have $N$ in common, $\mathscr{S}_A\sn \mathscr{S}_{A'}$ implies $\mathscr{S}_A\cap \mathscr{S}_{A'}\neq \emptyset$ (from Axiom (snN2)). This holds true for all spokes in $\Cn N$. Consequently, $\mathop{\bigcap}\limits_{\mathscr{S}_A\in \Cn N} \mathscr{S}_A\neq \emptyset$. Hence, the structure $\mbox{Nrv}\mathscr{F}_{_{MNC}}$ is an Edelsbrunner-Harer nerve.
\end{proof}
\begin{theorem}\label{thm:sn-nerve}
Let $\left(\mbox{Nrv}\mathscr{F}_{_{MNC}},\left\{\near,\dnear,\sn\right\}\right)$ be a proximal relator space, spokes $\mathscr{S}_A,\mathscr{S}_{A'}\in \mbox{Nrv}\mathscr{F}_{_{MNC}}$. Then
\begin{compactenum}[1$^o$]
\item $\mathscr{S}_A\ \sn\ \mathscr{S}_{A'} \Rightarrow \mathscr{S}_A\ \near\ \mathscr{S}_{A'}$.
\item $\mathscr{S}_A\ \sn\ \mathscr{S}_{A'} \Rightarrow \mathscr{S}_A\ \dnear\ \mathscr{S}_{A'}$.
\end{compactenum}
\end{theorem}
\begin{proof}$\mbox{}$\\
1$^o$: From Lemma~\ref{lem:MNCnerves}, $\mbox{Nrv}\mathscr{F}_{_{MNC}}$ is an Edelsbrunner-Harer nerve. Consequently, $\mathscr{S}_A\ \sn\ \mathscr{S}_{A'}$ for every pair of spokes $\mathscr{S}_A,\mathscr{S}_{A'}$ in the nerve.
Then, $\mathscr{S}_A\ \sn\ \mathscr{S}_{A'}$ implies $\mathscr{S}_A\ \cap\ \mathscr{S}_{A'}\neq \emptyset$, which implies $\mathscr{S}_A\ \near\ \mathscr{S}_{A'}$ (from Prop.~\ref{thm:sn-implies-near}).\\
2$^o$: Spokes $\mathscr{S}_A,\mathscr{S}_{A'}$ have nucleus $N$ in common. Hence, $\mathscr{S}_A\ \dcap\ \mathscr{S}_{A'}\neq \emptyset$. Then, from Prop.~\ref{thm:sn-implies-near}, $\mathscr{S}_A\ \dcap\ \mathscr{S}_{A'} \neq \emptyset \Rightarrow\ \mathscr{S}_A\ \dnear\ \mathscr{S}_{A'}$. This gives the desired result for each pair of spokes in the nerve.
\end{proof}
\begin{theorem}\label{EHnerve}{\rm ~\cite[\S III.2, p. 59]{Edelsbrunner1999}}
Let $\mathscr{F}$ be a finite collection of closed, convex sets in Euclidean space. Then the nerve of $\mathscr{F}$ and the union of the sets in $\mathscr{F}$ have the same homotopy type.
\end{theorem}
\begin{theorem}
Let the nucleus cluster $\Cn N$ be a finite collection of closed, convex sets in a Vorono\"{i} mesh $V$ in the Euclidean plane. The nerve $\mbox{Nrv}\mathscr{F}_{MNC}$ in $\Cn N$ and the union of the sets in $\Cn N$ have the same homotopy type.
\end{theorem}
\begin{proof}
Let $\Cn N$ be a MNC be nucleus $N$ in a Vorono\"{i} mesh. From Lemma~\ref{lem:MNCnerves}, $\mbox{Nrv}\mathscr{F}_{_{MNC}}$ is a Edelsbrunner-Harer nerve. From Theorem~\ref{EHnerve}, we have that the union of the sets in $\Cn N$ and $\mbox{Nrv}\mathscr{F}_{_{MNC}}$ have the same homotopy type.
\end{proof}
\begin{theorem}\label{lem:sndMNCnerves}
Let $X$ be a finite collection of MNC Edelsbrunner-Harer nerves $\mbox{Nrv}\mathscr{F}_{_{MNC}}$ in a Vorono\"{i} mesh with nuclei $N$ in the Euclidean plane and let $X$ be equipped with the relator $\left\{\sn,\snd\right\}$ with strongly close mesh nerves. Each nucleus $N$ has a description $\Phi(N) = \mbox{number of sides of}\ N$. Then $\mathop{\bigcap}\limits_{\Phi}\mbox{Nrv}\mathscr{F}_{_{MNC}} \neq \emptyset$.
\end{theorem}
\begin{proof}
Each $\mbox{Nrv}\mathscr{F}_{_{MNC}}$ is a collection of Vorono\"{i} regions containing a nucleus polygon $N$ with the same number of sides, since $\mbox{Nrv}\mathscr{F}_{_{MNC}}\in \Cn N$, which is maximal.
Let $\mathscr{N},\mathscr{N'}\in X$ be nerves with nuclei $N_1,N_2$ in maximal nucleus clusters. $\Phi(N_1) = \Phi(N_2)$, since $\Cn N_1,\Cn N_2$ are maximal, {\em i.e.}, $N_1,N_2$ have same number of sides. This means that all nuclei in $\mathscr{N},\mathscr{N'}$ have the same description. Consequently, $\mathscr{N}\ \snd\ \mathscr{N'}$ implies $\Int N_1\ \dcap\ \Int N_2\neq \emptyset$ (from Axiom (dsnP2)). Hence,
$N_1\ \snd\ N_2$ implies $\mathscr{N}\ \dnear\ \mathscr{N'}$ (from Prop.~\ref{thm:sn-implies-dnear}). Then $\mathscr{N}\ \dnear\ \mathscr{N'}$ implies $\mathscr{N}\ \dcap\ \mathscr{N'}\neq\emptyset$ (from Prop.~\ref{prop:dnear}). Therefore, $\mathop{\bigcap}\limits_{\Phi}\mbox{Nrv}\mathscr{F}_{_{MNC}} \neq \emptyset$.
\end{proof}
\bibliographystyle{amsplain}
|
\section{Introduction}
Near-resonance optical excitation of quantized matter underpins the field of quantum photonics. It enables the initialization, coherent manipulation, and read-out of the quantum states~\cite{Press08,Ramsay08,Kim11,Poem11,Yale13,Xia15} and, via resonance fluorescence~\cite{Muller07,Wrigge08}, the generation of indistinguishable single photons~\cite{Lettow10,He13natnanotech,Sipahigil14,Proux15} - a crucial resource for future quantum technologies~\cite{Gao15}. In the solid-state, such quantum optical demonstrations have been made with quantum dots~\cite{Press08,Ramsay08,Kim11,Poem11,Muller07,Proux15}, single molecules~\cite{Wrigge08,Lettow10} and crystal defects~\cite{Yale13,Sipahigil14,Xia15}. For fundamental investigations, resonant or near-resonant optical excitation is invaluable to probe the coherence and dephasing mechanisms in few-level quantum systems.
Rapid progress has recently been made in understanding the two-dimensional exciton (2D-$X$), spin, and valley-pseudospin properties in monolayer transition metal dichalcogenide semiconductors~\cite{Liu15,Xu14,Glazov15}. The first Brillouin zone of a monolayer transition metal dichalcogenide (TMD) such as MoS$_2$, WS$_2$, WSe$_2$, or MoSe$_2$ has a hexagonal shape that accommodates three pairs of degenerate but inequivalent edges, often denoted by $K$ and -$K$, which exhibit a direct band-gap with unique selection rules: for W-based TMDs left- (right-) handed circular polarized photons couple to interband transitions in the $K$ (-$K$) valley only ~\cite{Xu14,Glazov15,Jones13,Wang14}. Further, strong spin-orbit coupling links the spin and the valley-pseudospin, giving rise to spin-dependent optical selection rules. Also unique to these semiconductors with intrinsic two-dimensional confinement are very strong Coulomb interactions, large effective masses, and reduced dielectric screening which lead to large exciton binding energies ($\approx$0.5\,eV) and small Bohr radii ($<$1\,nm)~\cite{He14,Chernikov14}.
Recently, localized excitons that exhibit substantially reduced linewidths compared to 2D-$X$ have been discovered in two-dimensional materials \cite{Srivastava15,He15,Koperski15,Chakraborty15,Tonndorf15,Kumar15,Tran16,Branny16}. However, besides their basic magneto-optical properties, these quantum emitters have yet to be explored in detail. Fundamental open questions revolve around the precise nature of the three-dimensional confinement and its effect on emitter properties, e.g. spin-orbit coupling and valley hybridization~\cite{Liu14}, which impact the potential for a coherent spin-valley qubit that can be coherently controlled with near-resonance optical excitation~\cite{Wu16}. Here we focus on single quantum emitters in WSe$_2$, in which a range of observed magneto-optical properties are broadly categorized as follows. (i) Emitters with a fine-structure splitting (FSS) of 0.6 to 0.8 meV caused by exchange interactions that exhibit a large (7 to 10) exciton g-factor ~\cite{Srivastava15,He15,Koperski15,Chakraborty15,Kumar15}. The FSS doublet typically exhibits equal intensity and orthogonal linear polarization but not exclusively~\cite{Kumar15}. (ii) Emitters with a smaller FSS doublet ($\approx$ 0.3 meV) with approximately parallel linear polarization and very small g-factor. (iii) Spectral lines without measurable FSS that do not exhibit any or extremely small Zeeman splitting even for $B_{\text{ext}}$\,=\,9\,T~\cite{He15}. These emitters are typically linearly polarized and the degree of linear polarization is unchanged with magnetic field. In this Letter we investigate emitters in categories (ii) and (iii).
\begin{figure} \centering
\includegraphics[width =83.668mm]{fig1.pdf}
\caption{\textbf{A highly isolated quantum emitter with high purity single photon emission.}~\textbf{a,}~A low-resolution emission spectrum from location A on the WSe$_2$ monolayer showing a single emission line from a single localized emitter. Inset left: A high-resolution spectrum showing the zero-phonon line (ZPL) and a low-energy phonon sideband (PSB). Inset right: Color-coded normalized peak intensities map of emitter A and a neighboring emitter B with strong spatial localization of both emitters.\textbf{b,}~Normalized second-order correlation function $g^{(2)}(\tau)$ of the emission line of emitter A exhibiting nearly perfect antibunching. The solid red line is a 95$\%$ confidence band for fitting of the measured data, yielding a deconvolved $g^{(2)}(0)$\,=\,0.022\,$\pm$\,0.004 and a lifetime of 9.55\,$\pm$\,0.11\,ns. The thin cyan line is the calculated deconvolved $g^{(2)}(\tau)$. Inset left: Power dependence of integrated intensity and photon counting rate for emitter A. Inset right: The probability density of $g^{(2)}(0)$ calculated using the probabilistic values of the fitted parameters. The most probable value of $g^{(2)}(0)$ and its standard deviation have been estimated from this plot. The size of time bin is 128\,ps. The measurements were performed using non-resonant CW excitation at $\lambda\,=\,532$\,nm with powers of 4\,nW for (a) and 2\,nW for (b).} \label{fig1}
\end{figure}
\begin{figure*} \centering
\includegraphics[width =168.21mm]{fig2.pdf}
\caption{\textbf{Resonance fluorescence from a single quantum emitter.}~Simultaneous time traces of the fluorescence from emitter B under resonant CW excitation at $\lambda\,=\,784.69460$\,nm with a power of 1\,\textmu W as recorded on~\textbf{a,}~an APD and~\textbf{b,}~a high-resolution spectrometer with 70\,ms and 5\,s integration, respectively. The background level of ~0.4\,MHz shown by a horizontal dashed line in (a) and the line at $\sim$1580.03\,meV in (b) are due to the scattered excitation laser.~\textbf{c,}~The time trace of emitter detuning $\delta\,=\,E_{\text{laser}}\,-\,E_{\text{p1}}$ of the dominant emission line $p_{\text{1}}$ of emitter B. The gray area is the fitting errors of $\delta$.~\textbf{d,}~ $g^{(2)}(\tau)$ for a time-interval when $\delta \approx 0$ showing that the total signal is antibunched. The solid thick line is a 95$\%$ confidence band for fitting of the measured data, yielding a deconvolved $g^{(2)}(0)$\,=\,0.341\,$\pm$\,0.007 and a decay-time of 2.87\,$\pm$\,0.05\,ns. The relatively high value of $g^{(2)}(0)$ is due to a signal-to-background ratio of $\sim$4.3~\textbf{e,}~The fluorescence spectra of emitter B at two different time instances marked by black (blue) dashed lines in (a-c) corresponding to time $t$\,=\,12.8 (16.2)\,min. for $\delta$\,= 10 (190)\,\textmu eV. The black (blue) closed circles are measured data and solid lines are fits.} \label{fig2}
\end{figure*}
\begin{figure*} \centering
\includegraphics[width =168.98mm]{fig3.pdf}
\caption{\textbf{High-resolution PLE spectroscopy and observation of a weakly-fluorescent blue-shifted exciton (BS-$X$).}~\textbf{a,}~Polarization-resolved single fluorescence spectrum (top) and color-coded intensity map (bottom) of emitter B under non-resonant excitation showing three emission lines.~\textbf{b,}~A demonstration of high-resolution PLE spectroscopy. The fitted peak intensities (open circles) of the emission line $p_{\text{1}}$ in ~\ref{fig2}b as a function of $\delta$ identify $p_{\text{1}}$ beyond the resolution limit of the spectrometer and the spectral fluctuations. The gray area is the fitting errors of $\delta$.~\textbf{c,}~High-resolution PLE spectra showing the resonance for the line $p_{\text{0}}$. Open blue circles represent peak $p_{\text{1}}$ intensity as a function of $\delta$ and the closed red circles are the 5-point nearest-neighbour smoothed data. Two resonant excitation wavelengths (784.3800 and 784.4090\,nm) were used at a power of 10\,\textmu W and with a 50\,ms acquisition time.~\textbf{d,}~The PLE spectrum of emitter B shows the two bright-exciton peaks $p_{\text{0}}$ and $p_{\text{1}}$ and a BS-$X$ resonance. Closed blue circles are integrated intensities of the line $p_{\text{1}}$ obtained by scanning the laser wavelengths. The open red circles are PLE resonances for peaks $p_{\text{1}}$ and $p_{\text{0}}$. Peak $p_{\text{2}}$ is not visible in this experiment due to its lower emission energy.} \label{fig3}
\end{figure*}
\begin{figure} \centering
\includegraphics[width =89.796mm]{fig4.pdf}
\caption{\textbf{Emitter C: Observation of BS-$X$ in PLE spectroscopy and emission of high-purity single photons under resonant excitation of the BS-$X$.}~\textbf{a,}~PLE spectrum of emitter C identifying the ``BS-$X$" exciton at an energy $\sim$5\,meV higher than the ground-state exciton. The closed blue circles are data points while the solid red curve is guide to the eye composed of three Gaussian functions.~\textbf{b,}~The fluorescence spectrum of emitter C under non-resonant excitation at a power of 4\,\textmu W. The spectrum matches the energy range for which PLE is performed in (a). Insets: The full fluorescence spectrum of emitter C acquired on the high-resolution grating under non-resonant excitation. \textbf{c,}~ The fluorescence spectrum of emitter C under resonant excitation of the BS-$X$. The excitation power for (a) and (c) was 80\,nW.~Inset: Schematic of resonant excitation to the BS-$X$ and emission via the ground-state exciton.~\textbf{d,}~$g^{(2)}(\tau)$ under resonant CW excitation of the BS-$X$. High-purity single photon emission is observed with a deconvolved $g^{(2)}(0)$\,$<$\,0.002 and a decay-time of 3.50\,$\pm$\,0.05\,ns.}\label{fig4}
\end{figure}
\section{Results}
First, we show in Fig.~\ref{fig1} that monolayer WSe$_2$ is a suitable host for a pure single photon emitter. Under non-resonant excitation, a highly spectrally and spatially isolated emitter delivers single photon emission with a single photon purity $g^{\text{(2)}}(0)$\,$\approx$\,2$\%$ and a single photon count-rate $>$ 3\,MHz at saturation. A low-resolution microphotoluminescence (\textmu -PL) spectrum of emitter A, described by emitter category (iii) above, is shown in Fig.~\ref{fig1}a. In contrast to all previous observations where sharp emission lines have been accompanied by extraneous emission from other localized emitters or 2D-$X$~\cite{Srivastava15,He15,Koperski15,Chakraborty15,Tonndorf15,Kumar15,Tran16,Branny16}, here we demonstrate an emission spectrum dominated by a single quantum emitter. The 2D-$X$ emission is highly suppressed as the optically excited electron-hole pairs are efficiently captured by a single confined exciton. The left inset of Fig.~\ref{fig1}a shows the high-resolution \textmu -PL spectrum, revealing the zero-phonon line (ZPL) and a low energy phonon sideband (PSB). The intensity ratio of ZPL:PSB is $\approx$\,60:40. Figure~\ref{fig1}b presents the second-order correlation function $g^\text{(2)}(\tau)$ under non-resonant CW excitation. Using Bayesian statistics (see Supplementary Sec.~III for details) to fit (solid lines) the measured data (closed circles), we obtain a deconvolved $g^{(2)}(0)$\,=\,0.022\,$\pm$\,0.004 (see also the right inset of Fig.~\ref{fig1}b for the probability density plot). This high single photon purity is essential for future quantum photonic applications.
We now present resonance fluorescence (RF) from a single quantum emitter in monolayer WSe$_2$. Emitter B, belonging to emitter category (ii) above, was chosen due to its favourable wavelength ($\lambda\,\approx$\,784.69\,nm) for our tunable laser diode and good spatial and spectral isolation (see right-inset of Fig.~\ref{fig1}a). Non-resonant \textmu -PL was first used to identify the ZPL wavelength and then the excitation laser was tuned into resonance. The background laser scattering was highly, but not completely, suppressed using orthogonal linear polarizers in the excitation and collection arms of the microscope (see Methods). We split the RF signal into two parts: 70$\%$ was measured by an APD (see Fig.~\ref{fig2}a) and 30$\%$ by a spectrometer (see Fig.~\ref{fig2}b). By fitting each spectrum in Fig.~\ref{fig2}b, the emitter peak energy detuning ($\delta\,=\,E_{\text{laser}}\,-\,E_{\text{p1}}$, where $E_{\text{p1}}$ is the peak emission energy), is determined (see Fig.~\ref{fig2}c). Two example spectra with fits are shown in Fig.~\ref{fig2}e. The single photon count-rate dynamics are directly correlated with $\delta$. We ascribe the slow spectral fluctuations to charge noise in the emitter environment, similar to that observed in semiconductor quantum dots~\cite{Houel12,Dekker91}. A maximum count-rate (mean background level) of $\sim$1.7 ($\sim$0.4)\,MHz is observed when the emitter is in-resonance (out-of-resonance) with the excitation laser. A second-order coherence measurement during a time-interval when the emitter was in resonance with the excitation laser yielded $g^{(2)}(0)$\,=\,0.341\,$\pm$\,0.007, conclusively demonstrating that the RF signal is indeed composed of quantum light (see Fig.~\ref{fig2}d). A signal-to-background of $\approx$ 4.3 is obtained by fitting the measured antibunching data, in agreement with the the maximum signal and background count-rates shown in Fig.~\ref{fig2}a.
The polarization properties of emitter B under non-resonant excitation are shown in Fig.~\ref{fig3}a. The brightest peak $p_{\text{1}}$ is accompanied by peak $p_{\text{2}}$ ($p_{\text{0}}$) on its low- (high-) energy side, energetically separated by $\sim$330 (600)\,\textmu eV. The polarization-resolved PL map (see Fig.~\ref{fig2}a: bottom) shows that peaks $p_{\text{1}}$ and $p_{\text{2}}$ are linearly polarized along almost the same direction and peak $p_{\text{0}}$ is polarized at slightly different angle. Notably, under resonant excitation conditions (see Fig.~\ref{fig2}b and e and Supplementary Figs.~S1 and S2), emission from $p_{\text{2}}$ is highly suppressed compared to non-resonant excitation. This result provides a hint that a specific valley-index can be optically addressed, encouraging further investigations.
We take advantage of the spectral fluctuations to perform high-resolution photoluminescence excitation (PLE) spectroscopy. This method allows us to measure several $\delta$ values with an accuracy of $\pm$5\,\textmu eV at a fixed excitation laser wavelength. Figure~\ref{fig3}b plots fitted peak intensities of peak $p_{\text{1}}$ versus $\delta$ (see open circles), both extracted from Fig.~\ref{fig2}b. It shows a resonance for peak $p_{\text{1}}$ with FWHM of 42\,\textmu eV. Similarly, PLE of peak $p_{\text{0}}$ is performed (see Fig.~\ref{fig3}c). The resonances of peaks $p_{\text{1}}$, $p_{\text{0}}$ and a high-energy phonon band are clearly resolved (see Fig.~\ref{fig3}d). More importantly, an additional resonance peak, blue-shifted by $\sim$4.75\,meV from $p_{\text{1}}$, is also observed. The PL spectrum of emitter B under non-resonant excitation shows negligible emission at this energy (see Supplementary Fig.~S1).\\
\indent To investigate if the blue-shifted exciton (BS-$X$) observed in PLE is an intrinsic property of quantum emitters in monolayer WSe$_2$, we probe a third emitter. Emitter C, from the same monolayer flake and belonging to emitter category (iii), exhibits a BS-$X$ detuned from the primary exciton by 5.07\,$\pm$0.01\,meV. We compare the PLE spectrum (Figs.~\ref{fig4}a) with a PL spectrum (Fig.~\ref{fig4}b, which has a high-resolution, logarithmic intensity scale spectrum shown in the inset). Compared to the ground-state exciton, the PL emission from BS-$X$ is suppressed by a factor of 1,250.
Finally, we establish high-purity single photon emission from the ground-state exciton under resonant-excitation of the BS-$X$ state. Figure~\ref{fig4}c shows a spectrum consisting of the low energy bright exciton emission and the scattered excitation laser due to imperfect polarization cancellation. This laser peak can be filtered with high fidelity, enabling clean $g^{(2)}(\tau)$ measurements as shown in Fig.~\ref{fig4}d. A deconvolved $g^{(2)}(0)$ value of $<$\,0.002 is achieved, demonstrating a single photon source with perfect purity.
\section{Discussion}
The experiments have revealed that optical absorption into the BS-$X$ state quickly relaxes into the ground exciton state from which it can emit pure single photons. We currently do not understand the nature or origin of BS-$X$, but coupling to phonon modes can be excluded~\cite{Zhang15phonon}. One possibility is the BS-$X$ is a charged species that quickly relaxes into the ground state exciton, motivating future experiments with charge-tunable samples~\cite{Chakraborty15}. Another possibility is that the BS-$X$ state is a mostly optically inactive dark exciton. For W-based monolayer TMDs (WS$_2$, WSe$_2$), the electron spin in the lowest conduction band is antiparallel to carrier spin in the highest valence band, leading to optically inactive dark states which limits their quantum efficiency for light emission at low temperatures in comparison to Mo-based monolayer TMDs such as MoSe$_2$ ~\cite{Zhang15,Wang15}. Similar to valley hybridization, the optical activity of localized dark excitons could be linked to the symmetry of the confinement potential and underlying crystal lattice. Further investigations are required to understand the nature of the ground state excitons and BS-$X$. Tantalizingly, unlike with strict resonance fluorescence, the BS-$X$ offers future opportunities to investigate spin-valley coupling using excitation and fluorescence detection in both co-polarized or cross-polarized configurations.
In summary, we have demonstrated that monolayer WSe$_2$ is a benevolent host for a pure single photon emitter. These quantum emitters yield bright, stable, and highly-pure quantum light. The two-dimensional nature of the platform provides unique opportunities to engineer the light-matter interaction and integrate onto quantum photonic chips. We unambiguously achieve resonance fluorescence from the quantum emitters in spite of significant spectral fluctuations and background laser scattering. Strategies such as incorporating the single photon emitters into tunable electronic devices and surface passivation or encapsulation are likely to provide significant improvement. While the spectral fluctuations create challenges for quantum control and resonance fluorescence, we also demonstrate its utility for high-resolution PLE spectroscopy. PLE yields the direct observation of a three-dimensionally confined weakly-fluorescent exciton state that is energetically blue-shifted by $\sim$5\,meV. Resonant excitation of this BS-$X$ state provides an extremely robust and pure single photon source. The high-resolution characterization of the bright-exciton fine-structure and the experimental observation of the BS-$X$ are important results to better understand the specific nature of these localized excitons. The resonance fluorescence and laser spectroscopy techniques demonstrated here raise the prospect for indistinguishable single photon generation and investigations of the spin and valley coherence of strongly confined excitons in 2D-TMDs.
\section{Methods}
Using an all-dry viscoelastic stamping procedure~\cite{Castellanos-Gomez14}, we integrate a mechanically exfoliated WSe$_2$ flake onto a few layers of h-BN on top of a piezoelectric actuator so that in-plane dynamic strain could be induced in the flake by applying an out-of-plane electric field to the actuator. The actuator is made of a PMN-PT substrate. In the context of this letter, all experiments have been performed at zero external electric field to the actuator, and therefore, both top and bottom Ti/Au (5/100\,nm) electrodes of the actuator have been grounded. All measurements have been performed on a single monolayer, which has been identified using optical micrographs and spatial maps of \textmu -PL.
A confocal microscope with an objective lens with NA of 0.82, yielding a diffraction limited focus of $\sim$ 460 nm at $\lambda$ = 750 nm, was used for resonant laser spectroscopy. A CW tunable laser diode, covering a wavelengths range of 765 - 805\,nm, was used for resonant excitation. $\lambda$\,=\,532\,nm was used for non-resonant CW excitation. The fluorescence signal was separated from the excitation laser via orthogonally oriented linear polarizers in the excitation and collection arms of the microscope. This yields a 10$^7$ suppression of laser counts on smooth substrates, but the rough gold surface used here yields 10$^5$ suppression at best. The sample was placed on automated nanopositioners at T = 4\,K in a closed-cycle cryostat. All spectra were acquired with a 0.5 m focal length spectrometer and nitrogen-cooled charge-coupled device with a measured spectral resolution of $\sim$75 \textmu eV at $\lambda$\,=\,784\,nm for an 1800 l/mm grating. A separate confocal microscope is used to perform the polarization-resolved \textmu -PL measurements. A fiber based Hanbury-Brown and Twiss interferometer was used for second-order correlation measurements and photon counting was performed using Si avalanche photodiodes.
\begin{acknowledgments}
We thank B. Urbaszek for fruitful discussion, A. Rastelli for data analysis software, and A.C. Dada for assistance with the experimental setup. This work was supported by a Royal Society University Research Fellowship, the EPSRC (grant numbers EP/I023186/1, EP/K015338/1, and EP/L015110/1) and an ERC Starting Grant (number 307392), and ESP Grants. Support from the Spanish Government (Grant No. TEC2014-53727-C2-1-R), Comunidad Valenciana Government (Grant No. PROMETEOII/2014/059), and University of Valencia (UV-INV-PREDOC13-110538) is acknowledged.\\
\end{acknowledgments}
|
\section*{Holographic Smarr formula and equation of state}
The expectation value of the dual energy-momentum tensor is related to the
quasilocal stress tensor (incuding the counterterms):
\begin{equation}
\left\langle \mathcal{T}_{ab}\right\rangle =-\frac{2}{\sqrt{-\gamma}}%
\frac{\delta I}{\delta\gamma^{ab}}=\lim_{r\rightarrow\infty}\frac{r^{2}}%
{l^{2}}\mathcal{T}_{\mu\nu}^{BK}+\lim_{r\rightarrow\infty}\frac{r^{2}}{l^{2}%
}\mathcal{T}_{\mu\nu}^{ext}+\mathcal{T}_{ab}^{ct}\text{ ,} \label{Tensor}%
\end{equation}
where the first term is the Balasubramanian-Kraus part
\cite{Balasubramanian:1999re}%
\begin{equation}
\mathcal{T}_{\mu\nu}^{BK}=-\frac{1}{\kappa}\left( K_{\mu\nu}-h_{\mu\nu
}K+\frac{3}{l}h_{\mu\nu}-\frac{l}{2}\mathcal{G}_{\mu\nu}\right) \text{ ,}%
\end{equation}
with $\mathcal{G}_{\mu\nu}$ the Einstein tensor constructed with the metric
$h$. The second term and the third term are the extrinsic scalar field term
and the finite contribution contributions introduced in this paper:%
\begin{equation}
\mathcal{T}_{\mu\nu}^{ext}=\frac{1}{2}h_{\mu\nu}n^{\mu}\phi\partial_{\mu}%
\phi\text{ ,}\qquad\mathcal{T}_{ab}^{ct}=\frac{1}{l^{5}}\gamma_{ab}\left[
\frac{\alpha\left( t\right) \beta\left( t\right) }{2}-W(\alpha)\right]
\text{ .}%
\end{equation}
The relevant divergence coming from the bulk and Gibbons-Hawking contributions
is canceled out by the divergence from the counterterm and we obtain the
following regularized stress tensor of the dual field theory:
\begin{align}
\left\langle \mathcal{T}_{ab}\right\rangle & =\frac{\gamma_{ab}}{l^{3}%
}\left[ -\frac{3M_{h}\left( t\right) }{2\kappa}+\frac{k^{2}l^{2}}{8\kappa
}+\frac{2\mu\left( t\right) }{\kappa}+\frac{1}{l^{2}}\left( \alpha\left(
t\right) \beta\left( t\right) -\beta\left( t\right) ^{2}-W(\alpha
)\right) \right] \nonumber\\
& +\frac{1}{\kappa l}\delta_{a}^{0}\delta_{b}^{0}\left[ \frac{k^{2}}%
{2}+\frac{2\mu\left( t\right) }{l^{2}}\right] \text{ .} \label{BT}%
\end{align}
Taking the trace and using the field equations (\ref{grr}), we obtain
\begin{equation}
\gamma^{ab}\left\langle \mathcal{T}_{ab}\right\rangle =\frac{1}{l^{5}}\left[
\frac{1}{2}\alpha\left( t\right) ^{2}+2\alpha\left( t\right) \beta\left(
t\right) -4W(\alpha)\right] \text{ ,}%
\end{equation}
which vanishes for the AdS invariant boundary conditions. Using the normalized
timelike vector $u^{a}=\partial_{t}$, the energy density of the fluid is
\[
\rho=u^{a}u^{b}\left\langle \mathcal{T}_{ab}\right\rangle =\frac{1}{l^{3}%
}\left[ \frac{3M_{h}\left( t\right) }{2\kappa}+\frac{3k^{2}l^{2}}{8\kappa
}-\frac{1}{l^{2}}\left( \alpha\left( t\right) \beta\left( t\right)
-\beta\left( t\right) ^{2}-W(\alpha)\right) \right] \text{ }.
\]
The total mass is the energy density integrated on a spacelike section
\begin{equation}
M=\int_{\Sigma}\rho l^{3}d\Sigma=\left[ \frac{3M_{h}\left( t\right)
}{2\kappa}+\frac{3k^{2}l^{2}}{8\kappa}+\frac{1}{l^{2}}\left( -\alpha\left(
t\right) \beta\left( t\right) +W(\alpha)+\beta\left( t\right)
^{2}\right) \right] \sigma_{k}=H\text{ ,}%
\end{equation}
where the last equality is to remark that this result is in agreement with the
Hamiltonian with $H_{h}=\frac{3k^{2}l^{2}}{8\kappa}$. The counterterm
computation also provides the Casimir energy of the large N limit of
$\mathcal{N}=4$ Super Yang-Mills theory --- a cross check of our computation
is its exact agreement with the original paper of Balasubramanian-Kraus when
the scalar field vanishes \cite{Balasubramanian:1999re}.
The introduction of the scalar field yields a dual perfect fluid with energy
momentum tensor $\left\langle \mathcal{T}_{ab}\right\rangle =\left(
\rho+p\right) u_{a}u_{b}+p\gamma_{ab}$. Hence, we can identify%
\begin{equation}
p=\frac{1}{l^{3}}\left[ \frac{\mu\left( t\right) }{2\kappa}+\frac
{k^{2}l^{2}}{8\kappa}+\frac{1}{8l^{2}}\left( \alpha\left( t\right)
^{2}+4\alpha\left( t\right) \beta\left( t\right) -8W(\alpha)\right)
\right] \label{press}%
\end{equation}%
\begin{equation}
\rho=\frac{1}{l^{3}}\left[ \frac{3\mu\left( t\right) }{2\kappa}%
+\frac{3k^{2}l^{2}}{8\kappa}-\frac{1}{8l^{2}}\left( \alpha\left( t\right)
^{2}+4\alpha\left( t\right) \beta\left( t\right) -8W(\alpha)\right)
\right] \label{dens}%
\end{equation}
where we have used the relation (\ref{grr}). Note that when there is
no scalar field we get a thermal gas of massless particles $\rho=3p$
\cite{Myers:1999psa}. When $k=0$, an interesting implication of
(\ref{press}) and (\ref{dens}) is that the entropy density,~$s=$
$\frac{\mathcal{A}}{4G},$ and temperature $T$ of a perfect fluid is
defined by the relation\footnote{The
$l^{3}$ factor is due to our definition of the dual metric (\ref{dual}).}%
\begin{equation}
Ts=l^{3}\left( \rho+p\right) =\frac{2\mu\left( t\right) }{\kappa}\text{ ,}
\label{Smarr}%
\end{equation}
which exactly coincides with the generalized Smarr formula of
\cite{Liu:2015tqa}. Our calculation shows that the same formula holds when the
gravitational configuration is time dependent. When $k=0,$ the temperature of
the configuration has the form $T=\frac{8G}{\kappa l^{2}}\mu_{h}\left(
\phi_{h}\right) \mathcal{A}^{\frac{1}{3}}$. Using (\ref{Smarr}) we find
\begin{equation}
\mu\left( \phi_{h},\mathcal{A}\right) =\frac{\mu_{h}\left( \phi_{h}\right)
\mathcal{A}^{\frac{4}{3}}}{l^{2}}\text{ .} \label{mu}%
\end{equation}
Inserting (\ref{map}) in the first law of black hole thermodynamics, $\delta
H=T\delta S$, with the knowledge of (\ref{mu}) and (\ref{map 1}) shows that
the terms proportional $\delta\mathcal{A}$ cancel, provided that
\begin{equation}
\beta\left( \phi_{h},\mathcal{A}\right) =-\frac{1}{2}\alpha\ln\left(
\alpha\right) -\frac{1}{2}\alpha\beta_{h}\left( \phi_{h}\right) ~.
\label{beta}%
\end{equation}
Moreover, the terms proportional to the variation of $\delta\phi_{h}$ cancel
if and only if%
\begin{equation}
6d\mu_{h}+\kappa\alpha_{h}^{2}d\beta_{h}=0\text{ .} \label{mu0}%
\end{equation}
By the implicit function theorem we can take the independent variable to be
$\alpha_{h}$ and hence reduce the problem to that of finding only one
function, for instance $\beta_{h}$ obtaining then $\mu_{h}$ by a direct
integration of (\ref{mu0}). Finally, note that fixing the boundary condition
implies a functional relation between $\phi_{h}$ and $\mathcal{A}$. This is
equivalent to say that the black hole is characterized by a single integration constant.
A useful consequence of the general considerations made so far, it that there
are two $\mathcal{A}-$independent functions
\begin{equation}
\frac{\beta}{\alpha}+\frac{1}{2}\ln\left( \alpha\right) \text{ },\qquad
\frac{\mu}{\alpha^{2}}\text{ }.\label{Aind}%
\end{equation}
Whenever there is a hairy black hole with AdS invariant conditions at
$\phi_{h}=\phi_{\ast}$ then the functions $\left( \beta_{h},\alpha_{h}%
,\mu_{h}\right) $ admit a Taylor expansion around $\phi_{h}=\phi_{\ast
}\footnote{The numerical integration of the field equations done in
\cite{Hertog:2004dr} shows that for a scalar field belonging to a consistent
truncation of a type IIB supergravity Lagrangian it is possible to obtain black
holes with AdS invariant boundary conditions for any finite constant value of
$\beta_{h}$.}$. It follows then that one can obtain the generic form of the
surface of existence of hairy black holes around a given regular point
\begin{equation}
\frac{\beta}{\alpha}+\frac{1}{2}\ln\left( \alpha\right) =-\frac{1}{2}%
\beta_{\ast}+\frac{C}{\kappa}\left( \frac{\mu l^{2}}{\alpha^{2}}-\frac
{\mu_{\ast}}{\alpha_{\ast}^{2}}\right) +O\left( \frac{\mu^{2}}{\alpha^{4}%
}\right) \text{ ,}\label{eq1}%
\end{equation}
where it was used that exists a hairy black hole with AdS invariant boundary
conditions at $\left( \beta_h,\alpha_h,\mu_h\right) =\left( \beta
_\ast,\alpha_\ast,\mu_\ast\right) $ and $C$ is a constant that depends on the
theory. When higher order corrections are neglected, insertion of
equation (\ref{eq1}) in (\ref{mu0}) yields
\begin{equation}
\mu_{h}=\mu_{\ast}\left( \frac{\alpha_{h}}{\alpha_{\ast}}\right) ^{\frac
{2C}{-3+C}},\qquad\beta_{h}=\beta_{\ast}-\frac{2C}{\kappa}\frac{\mu_{\ast}%
}{\alpha_{\ast}^{2}}\left( \left( \frac{\alpha_{h}}{\alpha_{\ast}}\right)
^{\frac{6}{-3+C}}-1\right) \text{ .}\label{eq1a}%
\end{equation}
upon using \eqref{beta}, the integration constant being fixed by
requiring $\alpha_{h}=\alpha_{\ast
}\Longrightarrow\mu_{h}=\mu_{\ast}$.
We have now the tools to further analyze the connection between infrared
regularity and the equation of state. When the boundary conditions are fixed,
$\beta=\frac{dW(\alpha)}{d\alpha}$, the density and pressure are specified by
the choice of $W(\alpha)$. This is tantamount to defining an equation of
state. Conversely, specification of an equation of state $p=p(\rho)$
necessarily determines $W(\alpha)$ from equations (\ref{press}) and
(\ref{dens}). Indeed, we need the function $\mu(\alpha)$ to determine the
exact boundary condition associated to a given equation of state of the dual
fluid. However, we will keep the discussion general and treat a simple case.
For instance, we find that the equation of state $p=c_{s}^{2}\rho$ where
$c_{s}^{2}$ is the (constant) speed of sound squared, is equivalent to the
following one-parameter family of boundary conditions
\begin{equation}
W(\alpha)=\frac{l^{2}\left( 3c_{s}^{2}-1\right) }{\kappa\left( c_{s}%
^{2}+1\right) }\left( \alpha^{2}\omega\left( \alpha\right) -\frac
{k^{2}l^{2}}{8}\right) -\frac{\alpha^{2}}{4}\left( \ln(\alpha)+\alpha
_{1}-\frac{1}{2}\right) \text{ ,} \label{1param}%
\end{equation}
where we have parameterized $\mu=\alpha^{3}\frac{d\omega\left( \alpha\right)
}{d\alpha}$ and $\alpha_{1}$ is an integration constant. We readily see that
when $c_{s}^{2}=\frac{1}{3}$ we recover the description of the gas of massless
particles and the AdS invariant boundary conditions. Now, let us assume that
there is a black hole with AdS invariant boundary conditions at $\alpha
_{1}=\beta_{\ast}$. Then, we can use (\ref{eq1}) to find $\mu(\alpha)$ and
finally to find $\beta(\alpha):$%
\begin{equation}
\beta(\alpha)=\alpha\left[ \left( -\frac{1}{2}+\frac{3\mu_{\ast}\left(
3c_{s}^{2}-1\right) }{2\kappa\alpha_{\ast}^{2}}\right) \ln(\alpha
)-\frac{\beta_{\ast}}{2}-\frac{\mu_{\ast}\left( 3c_{s}^{2}-1\right) \left(
C-3\right) }{4\kappa\alpha_{\ast}^{2}}\right] +O\left( \left( 3c_{s}%
^{2}-1\right) ^{2}\right) \label{BC}%
\end{equation}
where we linearize around $c_{s}^{2}=\frac{1}{3}$ to be consistent with the
fact that we have neglected $O\left( \frac{\mu^{2}}{\alpha^{4}}\right) ~$in
(\ref{eq1}). The dependence of the boundary condition (\ref{BC}) on $C$ shows
that for any theory defined by the surface (\ref{eq1}) one can find the
corresponding boundary condition that yields the desired equation of state.
This is valid in a neighborhood of the AdS invariant boundary condition.
Hence, it is valid for equations of state that can be made arbitrarily close
to the gas of massless particles.
Although we have worked in five dimensions for scalar fields saturating the
Breitenlohner-Freedman bound, our results are easily generalized to any
spacetime dimension for any scalar field with masses between this bound and
the unitarity bound. This can be of particular use to the holographic
description of metals, superconductors and different kind of materials
\cite{Hartnoll:2008vx}. The holographic description of condensed matter
systems has recently been discussed in the hydrodynamic regime
\cite{Davison:2016hno}. The results bringed in here allow to actually
introduce a detailed description of the condensed matter system through its
equation of state, in the holographic picture.
The formalism introduced here has a direct application on the exact, time
dependent hairy black hole solutions in Einstein-dilaton gravity with general
moduli potential, recently constructed in \cite{Zhang:2014sta, Lu:2014eta,
Zhang:2014dfa, Fan:2015tua, Fan:2015ykb}. Indeed, all these collapsing black
holes are dual to some process in fluid/gravity with a very precise equation
of state that can now be unveiled.
We have seen that contrary to the common belief, there are many dual fluid
equations of state associated to one theory. This is particularly relevant for
string theory. The construction of the map (\ref{map}) along the lines
described in this letter for type IIB supergravity will provide an holographic
description of the fluid dynamics associated to the deformations of
$\mathcal{N}=4$ super Yang-Mills well behaved in the infrared.
\section*{Acknowledgments}
Research of AA is supported in part by Fondecyt Grant 1141073 and
Newton-Picarte Grants DPI20140053 and DPI20140115. The work of DA is supported
by the Fondecyt Grant 1161418 and Newton-Picarte Grant DPI20140115. This work
was supported in part by the Natural Sciences and Engineering Research Council
of Canada.
|
\section{Introduction}
\label{sec:intro}
Studies of two-particle correlations play an important role in understanding the underlying mechanism of particle production in high-energy nuclear collisions ~\cite{Adams:2005dq,Adcox:2004mh,Back:2004je}. Typically, these correlations are studied in a two-dimensional $\dphi$-$\deta$ space, where $\dphi$ and $\deta$ are the differences in the azimuthal angle $\phi$ and the pseudorapidity $\eta$ of the two particles.
A notable feature in the two-particle correlations is the so-called ``ridge,'' which is an extended correlation structure in
relative pseudorapidity $\deta$ concentrated at small relative azimuthal angle $|\dphi| \approx 0$. The ridge, first observed
in nucleus-nucleus (AA) collisions~\cite{Adams:2005ph,Alver:2009id,Abelev:2009af}, has been studied both at RHIC and LHC over a
wide range of collision energies and system sizes~\cite{Adams:2005ph,Alver:2009id,Abelev:2009af,Abelev:2009jv,Chatrchyan:2012wg,Khachatryan:2010gv,CMS:2012qk,Abelev:2012ola,Aad:2012gla,phenix2015,star2015Eta,Adare:2015ctn}. In AA collisions, such long-range two-particle correlations have been associated with
the development of collective hydrodynamic flow, which transfers the azimuthal anisotropy in the initial energy density distribution to the final state momentum anisotropy
through strong rescatterings in the medium produced in such collisions~\cite{Ollitrault:1992bk,Andrade:2006yh,Alver:2010gr,Gale:2013da,Heinz:2013th}.
A recent study suggests that anisotropic escape probabilities may already produce large final-state anisotropies without the need for significant rescattering~\cite{He:2015hfa}.
Another possible mechanism proposed to account for the initial-state correlations is the color glass condensate (CGC), where the two-gluon density
is enhanced at small $\dphi$ over a wide $\deta$ range~\cite{Dumitru:2010iy,Dusling:2013oia}. However, to reproduce the magnitude of the ridge in
AA collisions, the CGC-based models also require a late-stage collective flow boost to produce the observed stronger angular collimation
effect~\cite{Gavin:2008ev,Dusling:2012iga}. As a purely initial-state effect, the CGC correlations are expected to be independent of the formation of a
thermally equilibrated quark-gluon plasma, while the collective hydrodynamic flow requires a medium that is locally thermalized. The latter condition might
not be achieved in small systems.
Measurements at the LHC led to the discovery of a long-range ridge structure in small systems. The ridge has been observed in high-multiplicity
proton-proton (\pp)~\cite{PhysRevLett.116.172302,Khachatryan:2010gv,ppAtlas} and proton-lead (\pPb) collisions~\cite{CMS:2012qk,Abelev:2012ola,Aad:2012gla, Adam2016126}. A similar long-range structure was
also found in the most central deuteron-gold (dAu) and $^{3}$He-gold collisions at RHIC~\cite{phenix2015,star2015Eta,Adare:2015ctn}. To investigate
whether collective flow is responsible for the ridge in pPb collisions, multiparticle correlations were studied at the LHC~\cite{Chatrchyan:2013nka,atlasVn,cmsVn} in events with different multiplicities. The second harmonic anisotropy parameter, $v_2$, of the particle azimuthal distributions measured using four-, six-, eight-, or all-particle correlations were found to have the same value~\cite{cmsVn}, as expected in a system with global collective flow ~\cite{globalflow}.
In addition, the $v_2$ parameters of identified hadrons were measured as a function of transverse momentum (\pt) in \pPb ~\cite{cmsPID,alicePID} and in dAu collisions~\cite{phenix2015}. The $v_2(\pt)$ distributions were found to be ordered by the particle mass, i.e., the distributions for the heavier particles are boosted to higher \pt, as expected from hydrodynamics, where the particles move with a common flow velocity.
The similarities between the correlations observed in the small systems and in heavy ion collisions suggest a common hydrodynamic origin~\cite{Bozek:2010pb,Bozek:2012gr,Chatrchyan:2013nka}. However it is still under investigation whether hydrodynamics can be applied reliably to \pp or \pA systems.
As predicted by hydrodynamics and CGC~\cite{Bozek:2014plb,Duraes:2015qoa}, as well as phenomenological models like EPOS~\cite{Pierog:2013ria}, the average transverse momentum, $\langle\pt\rangle$, of the produced particles should depend on
pseudorapidity. This pseudorapidity dependence of $\langle\pt\rangle$ could translate into a pseudorapidity dependence of the long-range correlations which also depend on $\pt$~\cite{Bozek:2015plb}.
While hydrodynamics predicts that the pseudorapidity dependence of $\langle\pt\rangle$ follows that of the charged particle
pseudorapidity density $dN/d\eta$ which increases at negative pseudorapidity, in the CGC both a rising or a falling trend
of $\langle\pt\rangle$ with pseudorapidity may be possible~\cite{Duraes:2015qoa}. Thus, a measurement of the pseudorapidity dependence of the ridge may provide further insights into its origin.
The pseudorapidity dependence of the Fourier coefficients extracted using the long-range two-particle correlations could also be influenced by
event-by-event fluctuations of the initial energy density~\cite{PhysRevC.83.034911,PhysRevC.87.011901,PhysRevC.91.044904}. The pressure gradients that
drive the hydrodynamic expansion may differ in different pseudorapidity regions, causing a pseudorapidity-dependent phase shift in the event-plane orientation
determined from the direction of maximum particle emission. Evidence for such event-plane decorrelation has been found in pPb collisions~\cite{14012}.
Additional studies of the pseudorapidity dependence of the ridge may contribute to elucidating the longitudinal dynamics of the produced system.
The two-particle correlation measurement is performed using ``trigger'' and ``associated'' particles as described in Ref.~\cite{Chatrchyan:2011eka}. The trigger particles are defined as charged particles detected within a given $\pttrg$ range. The particle pairs are formed by associating each trigger particle with the remaining charged particles from a certain $\ptass$ range.
Typically, both particles are selected from a wide identical range of pseudorapidity, and therefore by construction the $\Delta\eta$ distribution is symmetric about $\deta = 0$~\cite{Chatrchyan:2013nka}.
Any $\Delta\eta$ dependence in the ridge correlation signal would be averaged out by the integration over the trigger and associated particle pseudorapidity distributions~\cite{Xu:2013sua}.
To gain further insights about the long-range ridge correlation in the \pPb system, in this paper we perform a $\deta$-dependent analysis by restricting the trigger particle to a narrow pseudorapidity range.
With this method, the combinatorial background resembles the single-particle density. Therefore, the correlation function in \pPb collisions is nonuniform in $\deta$.
The ridge correlation is often characterized by the Fourier coefficients $V_n$. The $V_n$ values are determined from a Fourier decomposition
of long-range two-particle \dphi\ correlation functions, given by:
\begin{linenomath}
\begin{equation}
\label{eq:Vn}
\frac{1}{N_\text{trig}}\frac{\rd N^\text{pair}}{\rd\Delta\phi} = \frac{N_\text{assoc}}{2\pi} \left[1+\sum\limits_{n} 2V_{n} \cos (n\Delta\phi)\right],
\end{equation}
\end{linenomath}
as described in Refs.~\cite{Chatrchyan:2011eka,Chatrchyan:2012wg}, where
$N^\text{pair}$ is the total number of correlated hadron pairs.
$N_\text{assoc}$ represents the total number of associated particles per trigger particle
for a given $(\pttrg, \ptass)$ bin.
To remove short-range correlations from jets and other sources, a pseudorapidity separation may be applied between the trigger and associated particle; alternatively, the correlations in low multiplicity events may be measured and subtracted from those in high multiplicity events after appropriate scaling, to remove the short-range correlations, which are likely to have similar $\deta$-$\dphi$ shapes in high- and low-multiplicity collisions.
Both methods are used in this analysis.
The single-particle anisotropy parameters $v_n$ are extracted from the particle-pair Fourier coefficients $V_n$, assuming that they factorize~\cite{Aamodt:2011by}.
The $v_n$ values are then normalized by their lab frame mid-rapidity values
and are studied as a function of $\etacm$. These distributions are compared to
the normalized pseudorapidity distributions of the mean transverse momentum.
\section{CMS detector}
\label{sec:cmsdetector}
A detailed description of the CMS detector, together with a definition of the coordinate system used and the relevant kinematic variables, can be found in Ref.~\cite{Chatrchyan:2008zzk}.
The main results in this paper are based on data from the silicon tracker.
This detector consists of 1440 silicon pixel and 15\,148 silicon
strip detector modules, and is located in the 3.8\unit{T} magnetic field of the superconducting solenoid.
It measures the trajectories of the charged particles emitted within the pseudorapidity range $\abs{\etalab}< 2.5$,
and provides an impact parameter resolution of ${\sim}15\mum$ and a transverse
momentum resolution of about 1\% for particles with $\pt= 2\GeVc$, and 1.5\% for particles at $\pt= 100$\GeVc.
The electromagnetic calorimeter (ECAL) and the hadron calorimeter (HCAL) are also
located inside the solenoid. The ECAL consists of 75\,848
lead-tungstate crystals, arranged in a quasi-projective geometry and distributed in a
barrel region ($\abs{\etalab} < 1.48$) and two endcaps that extend up to $\abs{\etalab} = 3.0$.
The HCAL barrel and endcaps are sampling calorimeters composed of brass and
scintillator plates, covering $\abs{\etalab} < 3.0$. Iron/quartz fiber Cherenkov
Hadron Forward (HF) calorimeters cover the range $2.9 < \abs{\etalab} < 5.2$ on either
side of the interaction region.
The detailed MC simulation of the CMS detector response is based
on \GEANTfour~\cite{GEANT4}.
\section{Data samples and event selection}
\label{sec:data}
The data used are from \pPb collisions recorded by the CMS detector in 2013, corresponding to an integrated luminosity of about 35\nbinv ~\cite{lumiRef}.
The beam energies were 4\TeV for protons and 1.58\TeV per nucleon
for lead nuclei, resulting in a center-of-mass energy per nucleon pair of $\rootsNN = 5.02\TeV$.
The direction of the higher-energy proton beam was initially set up to be clockwise, and then reversed.
Massless particles emitted at $\etacm = 0$ were detected at $\etalab = -0.465$ (clockwise proton beam) or at $\etalab = 0.465$ (counterclockwise proton beam) in the laboratory frame.
Both datasets were used in this paper. The data in which the proton beam traveled clockwise were reflected about $\etalab = 0$ and combined with the rest of the data, so that the proton beam direction is always associated with the positive $\etalab$ direction.
The online triggering, and the offline reconstruction and selection follow the same procedure as described in Ref.~\cite{Chatrchyan:2013nka}.
Minimum-bias events were selected by requiring that at least one track with $\pt>0.4\GeVc$ was found in the pixel tracker for a \pPb bunch crossing.
Because of hardware limits on the data acquisition rate, only a small fraction ($10^{-3}$) of all minimum bias triggered events were recorded (i.e., the trigger was ``prescaled'').
The high-multiplicity triggers were implemented using the Level-1 (L1) trigger and High Level Trigger (HLT) to enhance high multiplicity events that are of interest for the particle correlation studies.
At L1, two event streams were triggered by requiring the total transverse energy summed over ECAL and HCAL to be greater than 20 or 40\GeVc.
Charged tracks were then reconstructed online at the HLT using the three layers of pixel detectors, and requiring a track origin within a cylindrical region of 30 cm length along the beam and 0.2\unit{cm} radius perpendicular to the beam~\cite{Stenson:2010xx}.
In the offline analysis, hadronic collisions were selected by requiring at least 3\GeVc of total energy in at least one HF
calorimeter tower on each side of the interaction region (positive and negative $\etalab$).
Events were also required to contain at least one reconstructed primary
vertex within 15\unit{cm} of the nominal interaction point along the beam axis ($z_\text{vtx}$)
and within 0.15\unit{cm} distance transverse to the beam trajectory.
The \pPb instantaneous luminosity provided by the LHC in the 2013 \pPb run resulted in approximately a 3\% probability that at least one additional interaction occurs in the same bunch crossing, i.e. pileup events. A pileup rejection procedure~\cite{Chatrchyan:2013nka} was applied to select clean, single-vertex \pPb events. The residual fraction of pileup events was estimated to be no more than 0.2\% for the highest multiplicity \pPb interactions studied in this paper~\cite{Chatrchyan:2013nka}.
Based on simulations using the \HIJING~\cite{Gyulassy:1994ew}
and the \textsc{epos}~\cite{Porteboeuf:2010um} event generators, these event selections
have an acceptance of 94--97\% for \pPb interactions
that have at least one primary particle with $E>3$\GeV in both $\etalab$ ranges of
$-5<\etalab<-3$ and $3<\etalab<5$.
The charged-particle information was recorded in the silicon tracker and the tracks were reconstructed within the pseudorapidity range $\abs{\etalab}< 2.5$.
A reconstructed track was considered as a primary track candidate if the impact parameter
significance $d_{xy}/\sigma(d_{xy})$ and the significance of $z$ separation between the
track and the best reconstructed primary vertex (the one associated with the largest
number of tracks, or best $\chi^{2}$ probability if the same number of tracks was found)
$d_z/\sigma(d_z)$ are both less than 3.
In order to remove tracks with poor momentum estimates,
the relative uncertainty in the momentum measurement $\sigma(\pt)/\pt$ was required
to be less than 10\%.
To ensure high tracking efficiency and to reduce the rate of misreconstructed tracks, primary tracks with $\abs{\etalab}<2.4$ and $\pt>0.3\GeVc$ were used in the analysis.
The events are classified by \noff, the measured number of primary tracks within $\abs{\etalab}<2.4$ and $\pt>0.4\GeVc$ (a $\pt$ cutoff of 0.4\GeVc was used in the multiplicity determination to match the HLT requirement), in a method similar to the approach used in Refs.~\cite{CMS:2012qk,Khachatryan:2010gv}.
The high- and low-multiplicity events in this paper are defined by $220\leq\noff<260$ and $2\leq\noff<20$, respectively.
The high-multiplicity selection corresponds to an event fraction of $3.4\times 10^{-6}$ of the events.
Data from the minimum bias trigger are used for low-multiplicity event selection, while the high-multiplicity triggers with online multiplicity thresholds of 100, 130, 160, and 190 are used for high multiplicity events~\cite{Chatrchyan:2013nka}.
\section{Analysis procedure}\label{sec:}
The dihadron correlation is quantified by azimuthal angle $\phi$ and pseudorapidity differences between the two particles.
\begin{linenomath}
\begin{equation*}
\label{detadphi}
\Delta\phi=\phi_\text{assoc}-\phi_\text{trig},\quad \Delta\eta=\etaassoc-\etatrg,
\end{equation*}
\end{linenomath}
where $\phi_\text{assoc}$ and $\etaassoc$ are the associated particle coordinates and $\phi_\text{trig}$ and $\etatrg$ are the trigger particle coordinates, both measured in the laboratory frame.
The per-trigger normalized associated particle yield is defined by:
\begin{linenomath}
\begin{equation*}
\label{eq:signal}
S(\Delta\eta,\ \Delta\phi) = \frac{1}{N_\text{trig}}\frac{\rd^{2}N}{\rd\Delta\eta\, \rd\Delta\phi}.
\end{equation*}
\end{linenomath}
Unlike in previous studies ~\cite{Adams:2005ph,Alver:2009id,Abelev:2009af,Khachatryan:2010gv,CMS:2012qk,Abelev:2012ola,Aad:2012gla}, the trigger particles in this analysis are restricted to two narrow $\etalab$ windows: $-2.4<\etatrg<-2.0$ (Pb-side) and $2.0<\etatrg<2.4$ (p-side). The associated particles are from the entire measured $\etalab$ range of $-2.4<\etaassoc<2.4$.
The associated particles are weighted by the inverse of the
efficiency factor, $\varepsilon_\text{trk}(\etalab,\pt)$,
as a function of the track's pseudorapidity and $\pt$~\cite{Chatrchyan:2011eka}.
The efficiency factor accounts for the detector
acceptance $A(\etalab,\pt)$, the reconstruction efficiency $E(\etalab,\pt)$,
and the fraction of misidentified tracks, $F(\etalab,\pt)$,
\begin{linenomath}
\begin{equation*}
\varepsilon_\text{trk}(\etalab,\pt) = \frac{A E}{1-F}\,.
\end{equation*}
\end{linenomath}
The corresponding correction function is obtained from a {\PYTHIA6 }~(tune Z2)~\cite{PYTHIA} plus \GEANTfour \cite{GEANT4} simulation.
\subsection{Quantifying the jet contributions}
\begin{figure*}[thb]
\centering
\includegraphics[width=0.48\textwidth]{Figure_001-a.pdf}
\includegraphics[width=0.48\textwidth]{Figure_001-b.pdf}
\includegraphics[width=0.48\textwidth]{Figure_001-c.pdf}
\includegraphics[width=0.48\textwidth]{Figure_001-d.pdf}
\caption{ \label{fig:pPb_corrYield_2D}(Color online) Efficiency-corrected 2D associated yields with Pb-side trigger particle ($-2.4<\etatrg<-2.0$, left panels) and p-side trigger particle ($2.0<\etatrg<2.4$, right panels) in low-multiplicity ($2\leq\noff<20$, upper panels) and high-multiplicity ($220\leq\noff<260$, lower panels) are shown for \pPb collisions at $\rootsNN = 5.02\TeV$. The associated and trigger particle $\pt$ ranges are both $0.3<\pt<3\GeVc$.}
\end{figure*}
Figure~\ref{fig:pPb_corrYield_2D} shows the two-dimensional (2D) correlated yield for the two trigger particle pseudorapidity windows in low and high multiplicity events. The same $\pt$ range of $0.3<\pt<3.0\GeVc$ is used for trigger and associated particles.
The peak at $(0,0)$ is the near-side jet-like structure.
In the high multiplicity events, one can notice a ridge-like structure in $\abs{\deta}$ at $\dphi=0$ atop the high combinatorial background. A similar extensive structure can also be seen on the away side $\dphi=\pi$, which contains the away-side jet.
Unlike correlation functions from previous studies, the correlated yield is asymmetric in $\Delta\eta$; it reflects the asymmetric single particle $\rd N/\rd\eta$ distribution in the \pPb system.
\begin{figure*}[thb]
\centering
\includegraphics[width=\textwidth]{Figure_002.pdf}
\caption{\label{fig:dphi} (Color online) Examples of the distribution of the associated yields after ZYAM subtraction for both low-multiplicity ($2\leq\noff<20$, blue triangles) and high-multiplicity ($220\leq\noff<260$, red circles) are shown for \pPb collisions at $\rootsNN = 5.02\TeV$. The results for Pb-side (left panels) and p-side (right panels) trigger particles are both shown; small $\deta$ in the upper panels and large $\abs{\deta}$ in the lower panels. The trigger and associated particle \pt ranges are both $0.3<\pt<3\GeVc$.}
\end{figure*}
The $\dphi$ distribution of the associated yield is projected within each $\Delta\eta$ bin (with a bin width of 0.2).
Before quantifying jet contributions, the zero-yield-at-minimum (ZYAM) technique~\cite{Ajitanand:2005jj} is used to subtract a uniform background in $\Delta\phi$.
To obtain the ZYAM background normalization, the associated yield distribution is first projected into the range of $0<\Delta\phi<\pi$, and then scanned to find the minimum yield within a $\dphi$ window of $\pi/12$ radians. This minimum yield is treated as the ZYAM background. The ZYAM background shape as a function of $\deta$ is similar to the shape of the single particle density.
After ZYAM subtraction, the signal will be zero at the minimum.
For example, the $\dphi$ distributions in high- and low-multiplicity collisions are depicted in Fig.~\ref{fig:dphi} for two, short-($0<\abs{\deta}<0.2$) and long-range($2.8<\abs{\deta}<3.0$), $\deta$ bins.
They are composed of two characteristic peaks:
one at $\Delta\phi=0$ (near-side) and the other at $\Delta\phi=\pi$ (away-side), with a minimum valley between the two peaks.
For low-multiplicity collisions at large $\deta$, no near-side peak is observed.
First, the $\deta$ dependence of the correlated yield is analyzed. In each $\deta$ bin, the correlated yield is averaged within the near side ($|\Delta\phi|<\pi/3$. The correlated yield reaches a minimum at around $\pi/3$). The near-side averaged correlated yield per radian, $(1/N_\text{trig}) (\rd N)/(\rd\deta)$, is shown as a function of $\deta$ in Fig.~\ref{fig:near_fit_pPb}.
In low-multiplicity collisions, the near-side $\deta$ correlated yield is consistent with zero at large $\deta$. This indicates that the near side in low-multiplicity \pPb collisions is composed of only a jet component after ZYAM subtraction. In high-multiplicity collisions, an excess of the near-side correlated yield is seen at large $\deta$ and it is due to the previously observed ridge~\cite{CMS:2012qk}.
\begin{figure*}[thb]
\centering
\includegraphics[width=\textwidth]{Figure_003.pdf}
\caption{\label{fig:near_fit_pPb}(Color online) The near-side ($|\dphi|<\pi/3$) correlated yield after ZYAM subtraction in low-multiplicity $2\leq\noff<20$ (upper panels) and high-multiplicity $220\leq\noff<260$ (lower panels) are shown for \pPb collisions at $\rootsNN = 5.02\TeV$. The trigger and associated particle $\pt$ ranges are both $0.3<\pt<3\GeVc$. The trigger particles are restricted to the Pb-side ($-2.4<\etatrg<-2.0$, left panels) and the p-side ($2.0<\etatrg<2.4$, right panels), respectively. Fit results using Eq.~(\ref{eq:nearfit}) (black solid curves) are superimposed; the red dashed curve and the blue open points are the two fit components, jet and ridge, respectively.
}
\end{figure*}
In order to quantify the near-side jet contribution, the near-side correlation function is fitted with a two-component functional form:
\begin{linenomath}
\ifthenelse{\boolean{cms@external}}{
\begin{multline}
\label{eq:nearfit}
\frac{1}{N_\text{trig}}\,\frac{\rd N_\text{near}(\deta)}{\rd\deta}=\frac{Y\beta}{\sqrt{2}\sigma\Gamma(1/2\beta)}\exp\left[-\left(\frac{\Delta\eta^{2}}{2\sigma^{2}}\right)^{\beta}\right]\\
+(C+k\Delta\eta)\mathrm{ ZYAM}(\Delta\eta).
\end{multline}
}{
\begin{equation}
\label{eq:nearfit}
\frac{1}{N_\text{trig}}\,\frac{\rd N_\text{near}(\deta)}{\rd\deta}=\frac{Y\beta}{\sqrt{2}\sigma\Gamma(1/2\beta)}\exp\left[-\left(\frac{\Delta\eta^{2}}{2\sigma^{2}}\right)^{\beta}\right]+(C+k\Delta\eta)\mathrm{ ZYAM}(\Delta\eta).
\end{equation}
}
\end{linenomath}
The first term represents the near-side jet; $Y$ is the correlated yield, and $\sigma$ and $\beta$ describe the correlation shape. Neither a simple Gaussian nor an exponential function describes the jet-like peak adequately. However, a generalized Gaussian form as in Eq.~(\ref{eq:nearfit}) is found to describe the data well.
The second term on the right-hand side of Eq.~(\ref{eq:nearfit}) represents the ridge structure. Since the ridge is wide in $\Delta\eta$ and may be related to the bulk medium, its shape is modeled as dominated by the underlying event magnitude, $\mathrm{ZYAM}(\deta)$. However, the background shape multiplied by a constant is not adequate to describe the ridge in high multiplicity events. Instead, the background shape multiplied by a linear function in $\Delta\eta$, as in Eq.~(\ref{eq:nearfit}), can fit the data well, with reasonable $\chi^2/\mathrm{ndf}$ (where ndf is the number of degree of freedom) (see Table~\ref{tab:fit_para}). Here $C$ quantifies the overall strength of the ridge yield relative to the underlying event, and $k$ indicates the $\deta$ dependence of the ridge in addition to that of the underlying event.
\begin{table}
\topcaption{\label{tab:fit_para} Summary of fit parameters for low- and high-\noff\ ranges in \pPb collisions. }
\centering
\begin{scotch}{ l x x }
\multicolumn{3}{c}{\rule[-0.4\ruleht]{0pt}{1.25\ruleht}$\noff <20$} \\
\hline
Parameter & \multicolumn{1}{c}{Pb-side trigger} & \multicolumn{1}{c}{p-side trigger}\\[1.2ex]
Y &0.130,0.003 & 0.156,0.003\\
$\sigma$ &0.445,0.011 &0.446,0.010\\
$\beta$ &0.943,0.057 & 0.870,0.043\\
$C$ & 0.0045,0.0009 & 0.0045,0.0010\\
$k$ & \multicolumn{1}{l}{0 (Fixed)} & \multicolumn{1}{l}{0 (Fixed)}\\
$\chi^2/\mathrm{ndf}$ & \multicolumn{1}{l}{0.279} & \multicolumn{1}{l}{0.459}\\
\hline
\multicolumn{3}{c}{\rule[-0.4\ruleht]{0pt}{1.25\ruleht}$220 \leq \noff <260$}\\
\hline
Parameter & \multicolumn{1}{c}{Pb-side trigger} & \multicolumn{1}{c}{p-side trigger}\\[1.2ex]
Y & 0.401,0.011&0.489,0.011\\
$\sigma$ &0.457,0.008 &0.492,0.007 \\
$\beta$ & 0.757,0.003 &0.782,0.025 \\
$C$ & 0.0137,0.0004 & 0.0098,0.0004 \\
$k$ & - 0.0011,0.0001& 0.0002,0.0001 \\
$\chi^2/\mathrm{ndf}$ & \multicolumn{1}{l}{1.074} & \multicolumn{1}{l}{0.463}\\
\end{scotch}
\end{table}
The fits using Eq.~(\ref{eq:nearfit}) are superimposed in Fig.~\ref{fig:near_fit_pPb} and the fit parameters are shown in Table~\ref{tab:fit_para}. For low-multiplicity collisions, the $k$ parameter is consistent with zero and, in the fit shown, it is set to zero.
For high-multiplicity collisions, the $C$ parameter is positive, reflecting the finite ridge correlation, and the $k$ parameter is nonzero, indicating that the ridge does not have the same $\deta$ shape as the underlying event.
As already shown in Fig.~\ref{fig:near_fit_pPb}, the ridge (correlated yield at large $\deta$) is not constant but $\deta$-dependent.
The fitted $Y$ parameter shows that the jet-like correlated yield in high-multiplicity collisions ($Y_{ 220 \leq \noff <260}$) is larger than that in low-multiplicity collisions ($Y_{ \noff <20}$). The ratio is
\begin{linenomath}
\ifthenelse{\boolean{cms@external}}
{\begin{multline}
\alpha=Y_{ 220 \leq \noff <260}/Y_{ \noff <20} \\=
\begin{cases}
3.08\pm 0.11^{+0.96}_{-0.31} & \quad \text{for Pb-side triggers}; \\
3.13\pm 0.09^{+0.28}_{-0.28} & \quad \text{for p-side triggers}, \\
\end{cases}
\label{eq:alphaEq222}
\end{multline}
}
{\begin{equation}
\alpha=Y_{ 220 \leq \noff <260}/Y_{ \noff <20} =
\begin{cases}
3.08\pm 0.11^{+0.96}_{-0.31} & \quad \text{for Pb-side triggers}; \\
3.13\pm 0.09^{+0.28}_{-0.28} & \quad \text{for p-side triggers}, \\
\end{cases}
\label{eq:alphaEq222}
\end{equation}
}
\end{linenomath}
where the $\pm$ sign is followed by the statistical uncertainty from the fit. The upper ``$+$'' and lower ``$-$'' are followed by the systematic uncertainty, which is obtained by fitting different functional forms, such as Gaussian and exponential functions, and by varying the $\deta$ range to calculate the ZYAM value.
The $\alpha$ values are used as a scaling factor when correlations from low-multiplicity collisions are removed in determining the Fourier coefficients in high-multiplicity events.
\subsection{Fourier coefficients}
For each $\deta$ bin,
the azimuthal anisotropy harmonics, $V_n$, can be calculated from the two-particle correlation $\dphi$ distribution,
\begin{linenomath}
\begin{equation*}
V_n=\langle\cos n \Delta\phi\rangle.
\label{eq:2pVn}
\end{equation*}
\end{linenomath}
The $\langle\rangle$ denotes the averaging over all particles and all events.
At large $\deta$, the near-side jet contribution is negligible, but the away-side jet still contributes. The jet contributions may be significantly reduced or eliminated by subtracting the low-multiplicity collision data, via a prescription described in Ref.~\cite{Chatrchyan:2013nka},
\begin{linenomath}
\begin{equation}
\label{eq:VnSub}
V_n^\text{sub} = V_n^\mathrm{HM} - V_n^\mathrm{LM} \frac{N_\text{assoc}^\mathrm{LM}}{N_\text{assoc}^\mathrm{HM}}\alpha.
\end{equation}
\end{linenomath}
Here LM and HM stand for low-multiplicity and high-multiplicity, respectively. $N_\text{assoc}^\mathrm{HM}$ and $N_\text{assoc}^\mathrm{LM}$ are the associated particle multiplicities in a given pseudorapidity bin, and $V_n^\mathrm{HM}$ and $V_n^\mathrm{LM}$ are the Fourier coefficients in high- and low-multiplicity collisions, respectively. The $\alpha$ value is obtained from Eq.~(\ref{eq:alphaEq222}).
This procedure to extract $V_{n}$ is tested by studying the \pPb collisions generated by the \HIJING 1.383 model ~\cite{Gyulassy:1994ew}.
The basic \HIJING model has no flow, so a flow-like signal is added~\cite{afterburnerflow} by superimposing an azimuthal modulation on the distributions of the produced particles.
The measured $V_2$ using Eq.~(\ref{eq:VnSub}) is consistent with the input flow value within a relative 5\% difference.
To quantify the anisotropy dependence as a function of $\etalab$, assuming factorization, \breakhere$V_n(\etatrg, \etaassoc) = v_n(\etatrg)v_n (\etaassoc)$, a self-normalized anisotropy is calculated from the Fourier coefficient $V_{n}$.
\begin{linenomath}
\begin{equation}
\label{eq:vnNormalize}
\frac{v_n(\etaassoc)}{v_n(\etaassoc=0)}=\frac{V_n(\etaassoc)}{V_n(\etaassoc=0)}.
\end{equation}
\end{linenomath}
Here the $\etaassoc$ is directly calculated from $\deta$, assuming the trigger particle is at a fixed $\etalab$ direction
\begin{linenomath}
\begin{equation}
\label{eq:etaShift}
\etaassoc=\deta+\etatrg,
\end{equation}
\end{linenomath}
in which $\etatrg=-2.2\,(2.2)$ for the Pb-side (p-side) trigger. Hereafter, we write only $\etalab$, eliminating the superscript `assoc' from $\etaassoc$.
To avoid short-range correlations that remain even after the subtraction of the low-multiplicity events, only correlations with large $\abs{\deta}$ are selected to construct the $v_n$ pseudorapidity distributions.
\section{Systematic uncertainties}
\label{sec:systematic}
The systematic uncertainties in the Fourier coefficient $V_n$ are estimated from the following sources:
the track quality requirements by comparing loose and tight selections; bias in the event selection from the HLT trigger, by using different high-multiplicity event selection criteria; pileup effect, by requiring a single vertex per event; and the event vertex position, by selecting events from different z-vertex ranges.
In the low multiplicity $V_n$ subtraction, the jet ratio parameter $\alpha$ is applied.
The systematic uncertainties in $\alpha$ are assessed by using fit functions different from Eq.~(\ref{eq:nearfit}), as well as by varying the $\deta$ range when obtaining the ZYAM value. This systematic effect is included in the final uncertainties for the multiplicity-subtracted $V_n$.
In addition, the effect of reversing the beam direction is studied. This is subject to the same systematic uncertainties already described above; thus it is not counted in the total systematic uncertainties, but is used as a cross-check.
The estimated uncertainties from the above sources are shown in Table~\ref{tab:syst-table-new}. Combined together, they give a total uncertainty of 3.9\% and 10\% for $V_2$ and $V_3$ coefficients, respectively, as determined without the subtraction of signals from low-multiplicity events. For low-multiplicity-subtracted results, the systematic uncertainties rise to 5.8\% and 15\%, respectively.
The systematic uncertainties from the track-quality and jet-ratio selection are correlated among the pseudorapidity bins,
so they cancel in the self-normalized anisotropy parameter, \breakhere$v_n(\etalab)/v_n(\etalab=\nobreak 0)$.
The systematic uncertainties in other sources are treated as completely independent of pseudorapidity and are propagated in $v_n(\etalab)/v_n(\etalab\nobreak =0)$.
The estimated systematic uncertainties in $v_2(\etalab)/v_2(\etalab=0)$ and $v_3(\etalab)/v_3(\etalab=0)$ without low multiplicity subtraction are estimated to be 3.6\% and 10\%, respectively. For low-multiplicity-subtracted results, the systematic uncertainties rise to 5.7\% and 14\%, respectively .
\begin{table*}[htb]
\topcaption{\label{tab:syst-table-new} Summary of systematic uncertainties in the second and third Fourier harmonics in \pPb collisions. The label ``low-mult sub'' indicates the low-multiplicity subtracted results, while "no sub" indicates the results without subtraction.}
\centering
\begin{scotch}{ccccc}
\multicolumn{5}{c}{\rule[-0.4\ruleht]{0pt}{1.25\ruleht} $220\leq\noff < 260$ } \\
\hline
Source & $V_2$ (no sub) & $V_2$ (low-mult sub) & $V_3$ (no sub) & $V_3$ (low-mult sub)\\[1.2ex]
Track quality requirement & 3.0\% & 3.0\% & 7.0\% & 11.0\%\\
HLT trigger bias & 2.0\% & 2.5\% & 2.0\% & 2.5\% \\
Effect from pileups & 1.5\% & 3.0\% & 3.5\% & 3.5\% \\
Vertex dependence & 0.5\% & 1.0\% & 6.0\% & 9.0\% \\
Jet ratio &\NA & 3.0\% &\NA & 3.0\% \\[1.2ex]
Total & 3.9\% & 5.8\% & 10\% & 15\% \\
\end{scotch}
\end{table*}
\section{Results}
The $V_2$ and $V_3$ values in high-multiplicity collisions for Pb-side and p-side trigger particles are shown in Fig.~\ref{fig:V2_pPb}. The strong peak is caused by near-side short-range jet contributions.
The Fourier coefficients, $V_2^\text{sub}$ and $V_3^\text{sub}$, after the low-multiplicity data are subtracted, are also shown. The short-range jet-like peak is largely reduced, but may not be completely eliminated due to different near-side jet-correlation shapes for high- and low-multiplicity collisions. The long-range results are not affected by the near-side jet, but the away-side jet may still contribute if its shape is different in high- and low-multiplicity collisions or if its magnitude does not scale according to $\alpha$.
\begin{figure*}[thb]
\centering
\includegraphics[width=\textwidth]{Figure_004.pdf}
\caption{(Color online) Fourier coefficients, $V_2$ (upper) and $V_3$ (lower), of two-particle azimuthal correlations in high-multiplicity collisions ($220\leq\noff<260$) with (circles) and without (triangles) subtraction of low-multiplicity data, as a function of $\etalab$. Left panel shows data for Pb-side trigger particles and the right panel for the p-side. Statistical uncertainties are mostly smaller than point size; systematic uncertainties are $3.9\%$ and $10\%$ for $V_2$ and $V_3$ without low-multiplicity subtraction, $5.8\%$ and $15\%$ for $V_2$ and $V_3$ with low-multiplicity subtraction, respectively. The systematic uncertainties are shown by the shaded bands.}
\label{fig:V2_pPb}
\end{figure*}
By self-normalization via Eq.~(\ref{eq:vnNormalize}), the Fourier coefficient from both trigger sides can be merged into a single distribution by combining the negative and positive $\etalab$ range. The lab frame central value $\etalab=0$ is used so that the separation of the central value to both $\etatrg$ is the same. In this way, possible contamination from jets is kept at the same level as a function of $\etalab$. This is more important for the Fourier coefficients determined without the subtraction of the low-multiplicity data.
Figure~\ref{fig:v2_pPb} shows the $v_2(\etalab)/v_2(\etalab=0)$ and $v_3(\etalab)/v_3(\etalab=0)$ results obtained from the corresponding $V_2$ and $V_3$ data in Fig.~\ref{fig:V2_pPb}. The curves show the $v_n(\etalab)/v_n(\etalab=0)$ obtained from the high-multiplicity data alone, $V_n^\mathrm{HM}$, without subtraction of the low-multiplicity data. The data points are obtained from the low-multiplicity-subtracted $V_n^\text{sub}$; closed circles are from the Pb-side trigger particle data and open circles from the p-side. To avoid large contamination from short-range correlations, only the large $\abs{\deta}$ range is shown, but still with enough overlap in mid-rapidity $\etalab$ between the two trigger selections; good agreement is observed.
Significant pseudorapidity dependence is observed for the anisotropy parameter; it decreases by about $(24 \pm 4)\%$ (statistical uncertainty only) from $\etalab=0$ to $\etalab=2$ in the p-direction.
The behavior of the normalized $v_2(\etalab)/v_2(\etalab=0)$ is different in the Pb-side, with the maximum difference being smaller.
The $v_2$ appears to be asymmetric about $\etacm = 0$, which corresponds to $\etalab=0.465$. A non-zero $v_3$ is observed, however the uncertainties are too large to draw a definite conclusion regarding its pseudorapidity dependence.
When using long-range two-particle correlations to obtain anisotropic flow, the large pseudorapidity separation between the particles, while reducing nonflow effects, may lead to underestimation of the anisotropic flow because of event plane decorrelation stemming from the fluctuating initial conditions~\cite{PhysRevC.87.011901,PhysRevC.91.044904}. This effect was studied in pPb and PbPb collisions~\cite{14012}. The observed decrease in $v_2$ with increasing absolute value of pseudorapidity could be partially due to such decorrelation.
\begin{figure}[thb]
\centering
\includegraphics[width=0.48\textwidth]{Figure_005-a.pdf}
\includegraphics[width=0.48\textwidth]{Figure_005-b.pdf}
\caption{(Color online) Self-normalized anisotropy parameters, $v_2(\etalab)/v_2(\etalab=0)$ (\cmsLeft panel) and $v_3(\etalab)/v_3(\etalab=0)$ (\cmsRight panel), as a function of $\etalab$. Data points (curves) are results with (without) low-multiplicity data subtraction; filled circles and solid lines are from the Pb-side trigger. Open circles and dashed lines are from the p-side trigger. The bands show systematic uncertainties of $\pm5.7\%$ and $\pm14\%$ for $v_2(\etalab)/v_2(\etalab=0)$ and $v_3(\etalab)/v_3(\etalab=0)$, respectively. The systematic uncertainties in $v_n(\etalab)/v_n(\etalab=0)$ without subtraction are similar. Error bars indicate statistical uncertainties only. }
\label{fig:v2_pPb}
\end{figure}
The asymmetry of the azimuthal anisotropy is studied by taking the ratio of the $v_{n}$ value at positive $\etacm$ to the value at $-\etacm$ in the center-of-mass frame, as shown in Fig.~\ref{fig:v2_asym_ratio}.
The ratio shows a decreasing trend with increasing $\etacm$.
\begin{figure}[thb]
\centering
\includegraphics[width=0.48\textwidth]{Figure_006.pdf}
\caption{(Color online) $v_2(\etacm)/v_2(-\etacm)$, as a function of $\etacm$ in the center-of-mass frame. The data points are results from $V_n^\text{sub}$ with low-multiplicity data subtracted. The bands show the systematic uncertainty of $\pm$5.7\%.
Error bars indicate statistical uncertainties only.}
\label{fig:v2_asym_ratio}
\end{figure}
In pPb collisions, the average $\pt$ of charged hadrons depends on pseudorapidity. As stated in Ref.~\cite{Bozek:2014plb}, the pseudorapidity dependence of $\langle\pt\rangle$ could influence the pseudorapidity dependence of $v_2$. This may have relevance to the shape of the normalized $v_2$ distribution as observed in Fig.~\ref{fig:v2_pPb}.
To compare $v_2$ and the $\langle\pt\rangle$ distribution,
the $\pt$ spectra for different $\etacm$ ranges are obtained from Ref.~\cite{RpA12017}. The charged particle $\pt$ spectra in minimum-bias events are then fitted with a Tsallis function, as done in Ref.~\cite{10002}.
The inclusive-particle $\pt$ is averaged within $0< \pt<6 \GeVc$. In addition, the average momentum for the particles used in this analysis, $0.3< \pt<3 \GeVc$ and $220\leq\noff<260$, is calculated and plotted in Fig.~\ref{fig:v2_pt_compare}.
The $\langle\pt\rangle$ as a function of $\etacm$ does not change for different multiplicity ranges within 1\%. Thus, the minimum bias $\langle\pt\rangle$ distribution is compared directly to the high-multiplicity anisotropy $v_2$ result.
The $\langle\pt\rangle$ distribution is normalized by its value at $\etacm$ = -0.465.
Self-normalized $\langle\pt\rangle(\etacm)/\langle\pt\rangle(\etacm= -0.465)$ is plotted on Fig.~\ref{fig:v2_pt_compare}, compared to the self-normalized $v_2(\etacm)/v_2(\etacm=-0.465)$ distribution in the center-of-mass frame. The systematic uncertainty band for $\langle\pt\rangle(\etacm)/\langle\pt\rangle(\etacm= -0.465)$ is obtained by averaging the upper and lower limits of the systematic uncertainty band from the underlying $\pt$ spectra.
The hydrodynamic theoretical prediction for $\langle\pt\rangle(\etacm)/\langle\pt\rangle(\etacm= -0.465)$ is also plotted.
\begin{figure}[thb]
\centering
\includegraphics[width=0.48\textwidth]{Figure_007.pdf}
\caption{(Color online) Self-normalized $v_2(\etacm)/v_2(\etacm=-0.465)$ distribution with low-multiplicity subtraction from Pb-side (filled circles) and p-side (open circles) triggers, and $\langle\pt\rangle(\etacm)/\langle\pt\rangle(\etacm = - 0.465)$ of $0< \pt<6 \GeVc$ range from minimum-bias events (solid line) and $0.3< \pt < 3\GeVc$ range from high-multiplicity ($220\leq\noff<260$) events (dotted line) as functions of $\etacm$. Dashed curve is the hydrodynamic prediction for $\langle\pt\rangle(\etacm)/\langle\pt\rangle(\etacm= -0.465)$ distribution.
\label{fig:v2_pt_compare}}
\end{figure}
As shown in Fig.~\ref{fig:v2_pt_compare}, the hydrodynamic calculation~\cite{Bozek:2014plb} for $\langle\pt\rangle$ falls more rapidly than the $\langle\pt\rangle$ for data (solid and dotted lines) towards positive $\etacm$. The distribution is asymmetric for both data and theory.
The comparison of the $\langle\pt\rangle$ and the $v_2$ distributions shows that both observables have a decreasing trend towards large $|\etacm|$, but the decrease in $\langle\pt\rangle$ at forward pseudorapidity is smaller. The decrease of $v_2$ with $\etacm$ does not appear to be entirely from a change in $\langle\pt\rangle$; other physics is likely at play. The value of $v_2$ decreases by $(20 \pm 4)\%$ (statistical uncertainty only) from $\etacm = 0$ to $\etacm\approx 1.5$.
\section{Summary}
Two-particle correlations as functions of $\dphi$ and $\deta$ are reported in \pPb\ collisions at $\rootsNN = 5.02$\TeV by the CMS experiment. The trigger particle is restricted to narrow pseudorapidity windows.
The combinatorial background is assumed to be uniform in $\dphi$ and normalized by the ZYAM procedure as a function of $\deta$. The near-side jet correlated yield is fitted and found to be greater in high-multiplicity than in low-multiplicity collisions.
The ridge yield is studied as a function of $\dphi$ and $\deta$ and it is found to depend on pseudorapidity and the underlying background shape ZYAM($\deta$). The pseudorapidity dependence differs for trigger particles selected on the proton and the Pb sides.
The Fourier coefficients of the two-particle correlations in high-multiplicity collisions are reported, with and without subtraction of the scaled low-multiplicity data. The pseudorapidity dependence of the single-particle anisotropy parameters, $v_2$ and $v_3$, is inferred.
Significant pseudorapidity dependence of $v_2$ is found. The distribution is asymmetric about $\etacm = 0$ with an approximate $(20 \pm 4)\%$ decrease from $\etacm = 0$ to $\etacm\approx 1.5$, and a smaller decrease towards the Pb-beam direction. Finite $v_3$ is observed, but the uncertainties are presently too large to draw conclusions regarding the pseudorapidity dependence.
{\tolerance=1200
The self-normalized $v_2(\etacm)/v_2(\etacm=-0.465)$ distribution is compared to the \breakhere$\langle\pt\rangle(\etacm)/\langle\pt\rangle(\etacm= -0.465)$ distribution as well as from hydrodynamic calculations.
The $\langle\pt\rangle(\etacm)/\langle\pt\rangle(\etacm= -0.465)$ distribution shows a decreasing trend towards positive $\etacm$. The $v_2(\etacm)/v_2(\etacm = -0.465)$ distribution also shows a decreasing trend towards positive $\etacm$, but the decrease is more significant in the case of the $v_2$ measurement. This indicates that physics mechanisms other than the change in the underlying particle
spectra, such as event plane decorrelation over pseudorapidity, may influence the anisotropic flow.
\par}
\begin{acknowledgments}
\hyphenation{Bundes-ministerium Forschungs-gemeinschaft Forschungs-zentren} We congratulate our colleagues in the CERN accelerator departments for the excellent performance of the LHC and thank the technical and administrative staffs at CERN and at other CMS institutes for their contributions to the success of the CMS effort. In addition, we gratefully acknowledge the computing centers and personnel of the Worldwide LHC Computing Grid for delivering so effectively the computing infrastructure essential to our analyses. Finally, we acknowledge the enduring support for the construction and operation of the LHC and the CMS detector provided by the following funding agencies: the Austrian Federal Ministry of Science, Research and Economy and the Austrian Science Fund; the Belgian Fonds de la Recherche Scientifique, and Fonds voor Wetenschappelijk Onderzoek; the Brazilian Funding Agencies (CNPq, CAPES, FAPERJ, and FAPESP); the Bulgarian Ministry of Education and Science; CERN; the Chinese Academy of Sciences, Ministry of Science and Technology, and National Natural Science Foundation of China; the Colombian Funding Agency (COLCIENCIAS); the Croatian Ministry of Science, Education and Sport, and the Croatian Science Foundation; the Research Promotion Foundation, Cyprus; the Ministry of Education and Research, Estonian Research Council via IUT23-4 and IUT23-6 and European Regional Development Fund, Estonia; the Academy of Finland, Finnish Ministry of Education and Culture, and Helsinki Institute of Physics; the Institut National de Physique Nucl\'eaire et de Physique des Particules~/~CNRS, and Commissariat \`a l'\'Energie Atomique et aux \'Energies Alternatives~/~CEA, France; the Bundesministerium f\"ur Bildung und Forschung, Deutsche Forschungsgemeinschaft, and Helmholtz-Gemeinschaft Deutscher Forschungszentren, Germany; the General Secretariat for Research and Technology, Greece; the National Scientific Research Foundation, and National Innovation Office, Hungary; the Department of Atomic Energy and the Department of Science and Technology, India; the Institute for Studies in Theoretical Physics and Mathematics, Iran; the Science Foundation, Ireland; the Istituto Nazionale di Fisica Nucleare, Italy; the Ministry of Science, ICT and Future Planning, and National Research Foundation (NRF), Republic of Korea; the Lithuanian Academy of Sciences; the Ministry of Education, and University of Malaya (Malaysia); the Mexican Funding Agencies (BUAP, CINVESTAV, CONACYT, LNS, SEP, and UASLP-FAI); the Ministry of Business, Innovation and Employment, New Zealand; the Pakistan Atomic Energy Commission; the Ministry of Science and Higher Education and the National Science Centre, Poland; the Funda\c{c}\~ao para a Ci\^encia e a Tecnologia, Portugal; JINR, Dubna; the Ministry of Education and Science of the Russian Federation, the Federal Agency of Atomic Energy of the Russian Federation, Russian Academy of Sciences, and the Russian Foundation for Basic Research; the Ministry of Education, Science and Technological Development of Serbia; the Secretar\'{\i}a de Estado de Investigaci\'on, Desarrollo e Innovaci\'on and Programa Consolider-Ingenio 2010, Spain; the Swiss Funding Agencies (ETH Board, ETH Zurich, PSI, SNF, UniZH, Canton Zurich, and SER); the Ministry of Science and Technology, Taipei; the Thailand Center of Excellence in Physics, the Institute for the Promotion of Teaching Science and Technology of Thailand, Special Task Force for Activating Research and the National Science and Technology Development Agency of Thailand; the Scientific and Technical Research Council of Turkey, and Turkish Atomic Energy Authority; the National Academy of Sciences of Ukraine, and State Fund for Fundamental Researches, Ukraine; the Science and Technology Facilities Council, UK; the US Department of Energy, and the US National Science Foundation.
Individuals have received support from the Marie-Curie program and the European Research Council and EPLANET (European Union); the Leventis Foundation; the A. P. Sloan Foundation; the Alexander von Humboldt Foundation; the Belgian Federal Science Policy Office; the Fonds pour la Formation \`a la Recherche dans l'Industrie et dans l'Agriculture (FRIA-Belgium); the Agentschap voor Innovatie door Wetenschap en Technologie (IWT-Belgium); the Ministry of Education, Youth and Sports (MEYS) of the Czech Republic; the Council of Science and Industrial Research, India; the HOMING PLUS program of the Foundation for Polish Science, cofinanced from European Union, Regional Development Fund; the Mobility Plus program of the Ministry of Science and Higher Education (Poland); the OPUS program of the National Science Center (Poland); MIUR project 20108T4XTM (Italy); the Thalis and Aristeia programs cofinanced by EU-ESF and the Greek NSRF; the National Priorities Research Program by Qatar National Research Fund; the Rachadapisek Sompot Fund for Postdoctoral Fellowship, Chulalongkorn University (Thailand); the Chulalongkorn Academic into Its 2nd Century Project Advancement Project (Thailand); and the Welch Foundation, contract C-1845.
\end{acknowledgments}
|
\section{Introduction}
\IEEEPARstart{S}{imultaneous} wireless information and power transfer (SWIPT) is a new communication paradigm, where wireless devices extract information and energy from the received radio-frequency (RF) signals \cite{RUI}. Due to practical limitations, SWIPT cannot be performed from the same signal without losses and practical implementations divide the received signal in two parts; one part is used for information transfer and another part is used for power transfer \cite{KRI}. In this work, we focus on the power splitting (PS) technique, where the signal is split into two separate streams of different power \cite{RUI,KRI}. The time-splitting technique, where SWIPT is performed in the time domain, is a special case of PS with a binary power splitting parameter and therefore is excluded in our analysis.
Due to the importance of path-loss degradation on SWIPT systems, recent works study SWIPT from a large-scale point of view by taking into account the spatial randomness of the network deployment \cite{DIN}. In these works, the location of the devices and/or base stations is mainly modeled as a homogeneous Poisson point process (HPPP) in the 2-D space. To further boost the information/power transfer, directivity gains through antenna sectorization is introduced as a promising technology. Sectorized antennas not only efficiently handle multi-user interference through spatial separation \cite{HUN}, but also facilitate wireless power transfer in environments with poor propagation characteristics \cite{KHA}. However, most of the work on SWIPT with/without sectorized antennas is limited to 2-D networks and does not consider recent 3-D deployments.
A first effort to analyze the performance of a 3-D network is presented in \cite{PAN}, where the authors study the coverage probability for a 3-D large-scale cellular network with omnidirectional antennas. On the other hand, recent standardization activities (e.g, 3GPPP) focus on full-dimension multiple-input multiple-output technology, where antenna elements are placed into a 2-D planar array in order to achieve spatial separation to both the elevation and the traditional azimuth domains \cite{KIM,SIM}. The employment of 3-D antenna sectorization on 3-D networks with geometric randomness is a new research area without any previous work.
In this letter, we study antenna sectorization for 3-D bipolar ad hoc networks with SWIPT-PS capabilities. Conventional 2-D sectorized models \cite{HUN} are extended to capture spatial separation in the horizontal domain as well as the vertical domain. The impact of 3-D sectorized antennas on the achieved information and power transfer is evaluated by using stochastic geometry tools. Closed-form expressions as well as asymptotic simplifications are derived for the success probability and the average harvested energy. We reveal that 3-D antenna sectorization is a useful tool for 3-D networks and the associated additional degree of freedom (i.e., vertical spatial separation) achieves a higher information/power transfer than conventional antenna configurations. The main contributions of this letter are the development of a new 3-D antenna sectorization model for 3-D networks and its study in the context of SWIPT.
\vspace{-0.1cm}
\section{System model}\label{sec2}
We consider a 3-D Poisson bipolar ad hoc network consisting of a random number of transmitter-receiver pairs \cite{HUN}. The transmitters form a HPPP $\Phi=\{x_k \!\in\! \mathbb{R}^3\}$ with $k \geq 1$ of density $\lambda$ in the 3-D space \cite[Sec. 2.4]{HAN}, where $x_k$ denotes the coordinates of the node $k$. Each transmitter has a dedicated receiver (not a part of $\Phi$) at an Euclidean distance $d_0$ in some random direction. The time is considered to be slotted and in each time slot all the sources are active without any coordination or scheduling process. To analyze the performance of the network, consider a typical receiver located at the origin \cite{HAN}. This 3-D set-up is inline with ultra-dense urban networks, where a vast number of devices such as smartphones, tablets, sensors, connected objects will be connected in the same frequency band. These devices can be randomly spread in both the horizontal/vertical spatial domains in a chaotic manner.
We assume that wireless links suffer from both small-scale block fading and large-scale path-loss effects. The fading is Rayleigh distributed\footnote{More sophisticated 3-D channel models that capture the spatial correlation between the horizontal and the vertical domain are beyond the scope of this letter \cite{SIM}.} so the power of the channel fading is an exponential random variable with unit variance \cite{PAN}. We denote by $h_x$ the power of the channel fading for the link between the interfering transmitter $x$ and the typical receiver; $h_0$ denotes the power of the channel for the typical link. The path-loss model assumes that the received power is proportional to $1/(1+d^{\alpha})$ where $d$ is the Euclidean distance between the transmitter and the receiver, $\alpha>3$ denotes the path-loss exponent\footnote{The condition $\alpha>3$ ensures the validity of the derived closed-form expressions and holds for practical urban/suburban wireless environments. The considered path-loss model ensures that the path-loss is always larger than one for any distance \cite{HAN} and is appropriate for SWIPT systems.}. In addition, all wireless links exhibit additive white Gaussian noise (AWGN) with variance $\sigma^2$.
\begin{figure}[t]
\centering
\includegraphics[width=0.87\linewidth]{figure1.eps}\\
\vspace{-0.9cm}
\caption{a) 2-D sectorized planar antenna array with $M_1\times M_2=M$ elements; b) Sector size for the horizontal domain ($2\pi/M_1$) and the vertical domain ($\pi/M_2$).}\label{model1}
\end{figure}
\vspace{-0.2cm}
\subsection{Azimuth and vertical sectorization (AVS)}
The transmitters are equipped with sectorized planar antenna arrays to steer their beams to specific directions in both the azimuth and the vertical domains \cite{KIM,SIM}. Each planar antenna array consists of $M_1\times M_2=M$ equally spaced antenna elements arranged in a regular rectangular array in the 2-D plane; Fig. \ref{model1} schematically shows the planar antenna array and the associated azimuth/vertical sectorization. Each antenna element covers an angle $\frac{2\pi}{M_1}$ in the horizontal domain with directivity gain $M_1$, and an angle $\frac{\pi}{M_2}$ in the vertical domain with associated directivity gain $M_2$. We also assume a constant sidelobe level $\gamma$ for each antenna dimension with $0<\gamma< 1$ without loss of generality, where $\gamma$ denotes the ratio of the sidelobe level to the main lobe for the out-of-sector transmitted power for the horizontal and the vertical dimensions \cite{HUN}. The receivers are equipped with a single isotropic antenna and intercept RF electromagnetic fields equally in all possible directions in 3-D space.
Due to the 3-D sectorization, three interference terms appear as follows: $\Phi_1$ represents the set of interferers which are aligned with the location of the typical receiver in the horizontal/vertical domains; $\Phi_1$ is a HPPP in the 3-D space with density $\lambda_1=\lambda\frac{1}{M_1}\frac{1}{M_2}=\frac{\lambda}{M}$ (thinning operation \cite{HAN}). $\Phi_2$ is the set of interferers which are aligned with the typical receiver only in the horizontal domain; $\Phi_2$ is a 3-D HPPP with density $\lambda_2=\lambda\frac{1}{M_1}\frac{M_2-1}{M_2}=\lambda\frac{M_2-1}{M}$. $\Phi_3$ refers to the set of interferers which transmit in different horizontal and vertical directions from the location of the typical receiver; it is also a 3-D HPPP with density $\lambda_3=\lambda\frac{M_1-1}{M_1}$. We note that the vertical domain is taken into account only when the interference link is aligned with the horizontal location of the typical receiver; if an interferer occurs in a misaligned horizontal sector then its vertical sector is also misaligned.
By using similar arguments with \cite{HUN}, Table \ref{table1} summarizes the interference characteristics for each of these processes, where $\lambda_i$ is the node density of the process $\Phi_i$, $\theta_i$ and $\phi_i$ denote the sector size over which the process occurs for the horizontal and the vertical domain, respectively, and $\psi_i$ is the total antenna directivity gain. As for the direct links, we assume that the transmitters are perfectly aligned with their associated receivers i.e., the antenna directivity gain equals to $\psi_1$.
\vspace{-0.2cm}
\subsection{SWIPT at the receivers}
The transmitters have their own power supply (e.g., battery) and transmit with full power $P_t$. The receivers employ a SWIPT-PS technique and thus split the received signal into two flows of different power for information decoding and energy harvesting, respectively. Let $0<\nu\leq 1$ denote the PS parameter for each receiver \cite{KRI}. The energy harvesting circuit is ideal and the corresponding activation threshold is equal to zero \cite{RUI,DIN}. As for the information decoding process, an extra noise is introduced by the conversion operation of the RF band to baseband signal, which is modeled as AWGN with zero mean and variance $\sigma_C^2$. The signal-to-interference-plus-noise ratio (SINR)
at the typical receiver can be written as
\begin{align}
{\sf SINR}=\frac{\frac{\nu P_t \psi_1 h_0}{\eta}}{\nu[\sigma^2+P_t(I_1+I_2+I_3)]+\sigma_C^2},
\end{align}
where $\eta\triangleq 1+d_0^\alpha$ and $I_i\triangleq\psi_i\sum_{x \in \Phi_i} \frac{h_x}{1+d_x^\alpha}$ denotes the (normalized) aggregate interference generated by the transmitters in $\Phi_i$.
\begin{table}[t]
\caption{Characteristics of the interference processes}
\centering
\resizebox{\columnwidth}{!}{
\begin{tabular}{|c||c||c||c||c|}
\hline
- & $\lambda_i$ & $\theta_i$ & $\phi_i$ & $\psi_i$ \\
\hline
$I_1$ & $\lambda \frac{1}{M}$ & $2\pi\frac{1}{M_1}$ & $\pi \frac{1}{M_2}$ & $\frac{M}{(1+\gamma(M_1-1))(1+\gamma(M_2-1))}$ \\
\hline
$I_2$ & $\lambda \frac{M_2-1}{M}$ & $2\pi\frac{1}{M_1}$ & $\pi \frac{M_2-1}{M_2}$ & $\frac{\gamma M}{(1+\gamma(M_1-1))(1+\gamma(M_2-1))}$ \\
\hline
$I_3$ & $\lambda \frac{(M_1-1)}{M_1}$ & $2\pi\frac{M_1-1}{M_1}$ & $\pi$ & $\frac{\gamma^2 M}{(1+\gamma(M_1-1))(1+\gamma(M_2-1))}$\\
\hline
\end{tabular}}\label{table1}
\end{table}
\vspace{-0.1cm}
\section{Performance analysis}
In this section, we study the information and power transfer performance of the 3-D bipolar ad hoc network.
\vspace{-0.3cm}
\subsection{Information transfer- success probability}
The information transfer performance of the network is analyzed in terms
of success probability i.e., the probability that the receiver
can support a target SINR threshold $\beta$. The probability of successful transmission for the typical receiver is
\vspace{-0.1cm}
\begin{align}
\Pi_{\textrm{AVS}}&=\mathbb{P}\{{\sf SINR}\geq \beta \} \nonumber \\
&\;=\mathbb{P}\left\{h_0\geq \frac{\beta \eta [\nu(\sigma^2+P_t\sum_{i=1}^3 I_i)+\sigma_C^2]}{\nu \psi_1 P_t} \right\} \nonumber \\
&\;=\exp\left(-\frac{\beta \eta (\nu\sigma^2+\sigma_C^2)}{\nu \psi_1 P_t}\right)\mathbb{E}_{\Phi} \exp\left(-\frac{\beta \eta \sum_{i=1}^3 I_i}{\psi_1} \right) \nonumber \\
&\;=\exp\left(-\frac{\beta \eta (\nu\sigma^2+\sigma_C^2)}{\nu \psi_1 P_t}\right)\prod_{i=1}^3\mathbb{E}_{\Phi_i}\! \exp\left(-\beta \eta I_i/\psi_1 \right) \nonumber \\
&\;=\exp\!\left(-\frac{\beta \eta (\nu\sigma^2+\sigma_C^2)}{\nu \psi_1 P_t}\right) \prod_{i=1}^3\mathcal{L}_I \left(\frac{\psi_i\beta \eta}{\psi_1},\lambda_i \right), \label{res1} \\
\Pi_{\textrm{AVS}}^\infty &\;\rightarrow \mathcal{L}_I \left(\beta \gamma^2 \eta,\lambda \right)\;\;\;\text{for $P_t,M\rightarrow \infty$}, \label{ass1}
\end{align}
where $L_I(\cdot)$ denotes the Laplace transform of the interference and is given in the Appendix.
The performance of conventional antenna configurations yields from \eqref{res1} with appropriate parameterization. Specifically, the case where the transmitters are equipped with single omnidirectional antennas (case ``OM'') \cite{PAN} refers to $M_1=M_2=1$ and is equal to
\begin{align}
\Pi_{\text{OM}}&=\exp\left(-\frac{\beta \eta(\nu\sigma^2+\sigma_C^2)}{\nu P_t}\right)\mathcal{L}_I(\beta\eta,\lambda), \\
\Pi_{\text{OM}}^\infty &\rightarrow \mathcal{L}_I(\beta\eta,\lambda)\;\;\;\text{for $P_t\rightarrow \infty$}. \label{ass2}
\end{align}
The case where the transmitters are equipped with a 1-D linear array of $M$ sectorized antennas for spatial separation on the traditional azimuth domain \cite{HUN} (case ``AS'') refers to $M_1=M$, $M_2=1$ and corresponds to
\begin{align}
\Pi_{\text{AS}}&=\exp\left(-\frac{\beta \eta (\nu\sigma^2+\sigma_C^2)(1+\gamma(M-1))}{\nu M P_t}\right) \nonumber \\
&\;\;\;\;\times \mathcal{L}_I(\beta \eta,\lambda/M)\mathcal{L}_I(\beta\gamma \eta,\lambda(M-1)/M), \\
\Pi_{\text{AS}}^\infty &\rightarrow \mathcal{L}_I(\beta\gamma \eta,\lambda)\;\;\;\text{for $P_t,M\rightarrow \infty$}. \label{ass3}
\end{align}
By comparing the asymptotic expressions in \eqref{ass1}, \eqref{ass2}, and \eqref{ass3}, we have $\Pi_{\text{OM}}^\infty<\Pi_{\text{AS}}^\infty<\Pi_{\text{AVS}}^\infty$.
\vspace{-0.2cm}
\subsection{Power transfer- average harvested energy}
On the other hand, RF energy harvesting is a long term operation and is expressed in terms of average harvested energy \cite{RUI}. If $100(1-\nu)\%$ of the received energy is used for rectification, the average energy harvesting at the typical receiver is expressed as
\begin{align}
{\sf E}_{\text{AVS}}&=\zeta \mathbb{E}_{\Phi,h} (1-\nu)\left[\frac{P_t \psi_1 h_0}{\eta}+P_t(I_1+I_2+I_3) \right] \nonumber \\
&=\zeta (1-\nu)P_t\left(\frac{\psi_1}{\eta}+\sum_{i=1}^3 \psi_i\mathbb{E}_{\Phi_i} \sum_{x \in \Phi_i}\frac{1}{1+d_x^\alpha} \right) \nonumber \\
&=\zeta (1-\nu)P_t\left(\frac{\psi_1}{\eta}+\sum_{i=1}^3\psi_i \mathcal{E}_I(\lambda_i) \right), \\
{\sf E}_{\text{AVS}}^\infty \!&\rightarrow\!{\zeta (1\!-\!\nu)P_t}\bigg(\!\frac{1}{\eta \gamma^2}\!+\!\frac{4\pi^2 \lambda}{\alpha\sin(\frac{3\pi}{\alpha})} \!\bigg)\;\text{for $M_1,M_2\!\rightarrow\!\! \infty$}, \label{asym}
\end{align}
\noindent where $\zeta \in (0 , 1]$ denotes the conversion efficiency from RF signal to DC voltage \cite{RUI,DIN}, and $\mathcal{E}_I(\cdot)$ is given in the Appendix. It is worth noting that the RF energy harvesting from the AWGN noise is considered to be negligible. The expression in \eqref{asym} shows that the average harvested energy asymptotically depends on the ratio $\gamma$, while the ambient harvested energy (i.e., harvesting from the multi-user interference) is equal to the one of the OM and AS cases (see \eqref{exp1}, \eqref{exp11}).
For the special cases with $M_1=M_2=1$ (single omnidirectional antenna) and $M_1=M$, $M_2=1$ (sectorized linear antenna array), the average harvested energy is given by
\begin{align}
{\sf E}_{\text{OM}}&=\zeta (1-\nu)P_t\left[\frac{1}{\eta}+\frac{4\pi^2\lambda}{\alpha \sin(\frac{3\pi}{\alpha})} \right], \label{exp1}\\
{\sf E}_{\text{AS}}&={\zeta (1-\nu)P_t}\bigg(\frac{M}{\eta[1+\gamma(M-1)]}+\frac{4\pi^2 \lambda}{\alpha\sin(\frac{3\pi}{\alpha})} \bigg), \label{exp11} \\
{\sf E}_{\text{AS}}^\infty &\rightarrow {\zeta (1-\nu)P_t}\bigg(\frac{1}{\eta \gamma}\!+\!\frac{4\pi^2 \lambda}{\alpha\sin(\frac{3\pi}{\alpha})} \bigg)\;\text{with $M\rightarrow \infty$}. \label{asym3}
\end{align}
By comparing the asymptotic expressions in \eqref{asym}, \eqref{exp1} and \eqref{asym3}, we have
${\sf E}_{\text{OM}}<{\sf E}_{\text{AS}}^\infty<{\sf E}_{\text{AVS}}^\infty$.
\section{Numerical results}
Computer simulations are carried-out in order to evaluate the performance of the proposed schemes. The simulation set-up follows the description in Section \ref{sec2} with parameters (unless otherwise defined) $P_t=-10$ dBm, $\alpha=4$, $\lambda=10^{-4}$, $d_0=3$ m, $M=16$, $M_1=M_2=\sqrt{M}$, $\beta=20$ dBm, $\sigma^2=-130$ dBm, $\sigma_C^2=-30$ dBm, $\nu=0.5$, $\zeta=1$, and $\gamma=0.3$ \cite{LU}. Conventional antenna configurations such as omnidirectional antennas (OM) and azimuth sectorization (AS) are used as benchmarks. In addition, existing 2-D models are presented for the OM and AS configurations \cite{PAN}.
Fig. \ref{fig1} plots the success probability and the average harvested energy versus the transmit power $P_t$ and the SINR threshold $\beta$. As it can be seen, antenna sectorization significantly facilitates the information and the power transfer in environments with low ambient interference. 3-D sectorization achieves a more accurate beam orientation by exploiting the horizontal as well as the vertical spatial separation and thus achieves higher success probability and more efficient energy transfer by boosting the direct links. The comparison with the conventional 2-D models shows the compatibility of the 3-D models and confirms the superiority of the 3-D antenna sectorization. Theoretical curves perfectly match with the numerical results and validate our analysis.
\begin{figure}[t]
\centering
\includegraphics[width=0.89\linewidth]{figure2.eps}\\
\vspace{-0.3cm}
\caption{a) Success probability versus $P_t$; b) Average harvested energy versus $P_t$; c) Success probability versus $\beta$; d) Average harvested energy versus $\beta$.}\label{fig1}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{figure3.eps}\\
\vspace{-0.4cm}
\caption{a) Success probability versus $\lambda$; b) Average harvested energy versus $\lambda$; c) Success probability versus $\nu$; d) Average harvested energy versus $\nu$.}\label{fig2}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{figure4.eps}\\
\vspace{-0.4cm}
\caption{a) Success probability versus $d_0$; b) Average harvested energy versus $d_0$; c) Success probability versus $M$; d) Average harvested energy versus $M$.}\label{fig3}
\end{figure}
The impact of the network density and the PS parameter on the achieved information/power transfer is depicted in Fig. \ref{fig2}. Fig. \ref{fig2}.(a) shows that sectorization is more beneficial for moderate network densities, where multi-user interference can be efficiently controlled through spatial separation. From the harvesting standpoint, Fig. \ref{fig2}.(b) illustrates that as $\lambda$ increases, multi-user interference becomes the dominant component of the harvested energy and thus antenna sectorization has limited interest. On the other hand, Fig.'s \ref{fig2}.(c), \ref{fig2}.(d) capture the fundamental trade-off between information and energy transfer; as $\nu$ increases the success probability is improved while the harvesting process becomes less efficient, since most of the received energy is used for information decoding.
Finally, Fig. \ref{fig3} illustrates the impact of the Euclidean distance $d_0$ as well as the number of antenna elements $M$ on the system performance. As it can be seen from Fig.'s \ref{fig3} (a), \ref{fig3} (b), as the distance $d_0$ increases, the success probability decreases since the direct links become more weak. In addition, harvesting gains from antenna sectorization become less significant as $d_0$ increases, since the ambient multi-user interference dominates the harvesting process. Fig.'s \ref{fig3} (c), \ref{fig3} (d) indicate that as the number of antenna elements increases, antenna steering becomes more accurate which is beneficial for information/power transfer through direct links; the convergence floor of the curves also validates the proposed asymptotic expressions.
\vspace{-0.2cm}
\section{Conclusion}
In this letter, a new tractable 3-D model for bipolar ad hoc
networks with antenna sectorization has been presented and studied from a SWIPT point of view. Simple closed-form expressions for the success probability and the average harvested energy have been derived for the SWIPT-PS technique. Our results have shown that the 3-D antenna sectorization fully exploits the three spatial dimensions of the network and outperforms conventional antenna configurations, mainly at moderate network densities and transmitter-receiver distances.
\vspace{-0.4cm}
|
\section{Introduction}
Spin liquids are strongly-interacting spin systems which exhibit long-range entanglement in the ground state and an exotic excitation spectrum which is not smoothly connected to the non-interacting limit. See Reference \onlinecite{lucile} for an entanglement-based review of spin liquids. Theoretically, spin liquids correspond to the deconfined phases of compact gauge theories, and we can mostly use the terms spin liquid and gauge theory interchangeably.\cite{foot1} In the gapped case, an equivalent concept is topological order.\cite{xgtop} We use the term spin liquid in a more general sense, referring to stable phases with long-range entanglement, either gapped or gapless, all described quite naturally as deconfined gauge theories. To date, the study of spin liquids has mostly focused on rank 1 gauge theories, by which we mean the gauge variable $A_i$ (``vector potential") is a vector, $i.e.$ a rank 1 tensor. This gauge variable can take discrete or continuous values, or even be matrix valued in the case of non-abelian gauge theories.
In the present work however, we will investigate a class of spin liquids going beyond this traditional paradigm. While the vector potential theory is most familiar, there is in general no reason why the emergent gauge variable should be a vector, as opposed to a more complicated geometric object, such as a higher rank tensor. (The meaning of ``tensor" will be discussed further in the Appendix.) We therefore wish to explore the consequences of a tensor gauge structure, to see whether or not we get any exotic new physics. We will briefly discuss later the prospects of realizing such phases in experiments, but ultimately the motivation here will not come from a specific physical system. Rather, we are seeking to expand the boundaries of what sorts of emergent behavior can occur in highly entangled phases of matter. We will find that the phases considered here represent a drastic departure from the properties of rank 1 spin liquids and represent an exciting new field of study.
As a first thought on going beyond the vector framework, one might consider working in the language of differential forms. For example, while a vector potential $A_i$ naturally describes a phase of condensed strings ending on point-like excitations, the antisymmetric rank 2 tensor $A_{ij}$ (``two-form") naturally describes a phase of condensed surfaces terminating on string-like excitations. However, antisymmetric tensor theories do not lead to new stable spin liquids in 3+1 dimensions. In the discrete case, the two-form simply leads to a dual description of the conventional one-form, swapping the role of the electric particles and magnetic loops. In the $U(1)$ case, the two-form (called a Kalb-Ramond field in the string theory literature) actually is distinct from the one-form, but unfortunately these models are unstable to gapped confined phases in 3+1 (or fewer) dimensions \cite{analytic,numerical,soojong,gerbe}, so they do not correspond to realistic spin liquids in our world.\cite{foot2} We also expect non-abelian generalizations to share this instability to confinement, so pure antisymmetric tensor gauge theories seem to contain no new physics in 3+1 dimensions.
On the other hand, the case of symmetric tensor $U(1)$ gauge theories\cite{foot3} has recently been analyzed by Rasmussen, You, and Xu\cite{alex}, who showed that there are multiple theories of a given rank, and all such theories are stable gapless phases in 3+1 dimensions (but unstable in 2+1). Furthermore, there is reason to be optimistic that these phases can be found in realistic spin systems, as we shall comment on further in the conclusion. We shall therefore focus on (3+1)-dimensional symmetric tensor theories in this work, though a brief mention will be made of mixed-symmetry tensors, which may or may not be stable. (Note that a general tensor gauge theory does not fall into the category of ``higher form" gauge theory, which refers specifically to antisymmetric tensor objects. A generic tensor does not correspond to a differential form.) The symmetric tensor gauge theories represent a new class of stable gapless spin liquids, where the gaplessness is protected against any microscopic perturbation to the Hamiltonian, without regards to symmetry. However, important questions remain. How do we know that these theories are really distinct from previously understood spin liquids, and are not simply dual formulations of some rank 1 theory, or perhaps several rank 1 theories stapled together? Is there some property of these theories which fundamentally distinguishes them from the rank 1 case? We shall here answer this question in the affirmative for the $U(1)$ case by examining the particle structure of higher rank $U(1)$ spin liquids. It will be found that, though the gapless gauge modes of these systems are qualitatively similar to the rank 1 case, the particle structure is dramatically different.
The particles in these theories have severe restrictions placed on their motion, in a manifestation of a phenomenon encountered earlier in the condensed matter literature. This sort of restricted mobility was first seen in work due to Chamon\cite{chamon} and in several other models\cite{bravyi,cast,yoshida}, including the famous Haah's code\cite{haah,haah2}. These principles were later more systematically developed in the ``fracton" theories of Vijay, Haah, and Fu \cite{fracton1,fracton2}. (We also refer the reader to other recent literature on the fracton phenomenon.\cite{williamson,sagarlayer,hanlayer,abhinav}) Depending on the gauge constraint, particles in tensor gauge theories can be restricted to motion along lower-dimensional subspaces, or they can be prevented from moving at all without the creation of other excitations. Following the conventions of Reference \onlinecite{fracton1}, the former will be referred to as $d$-dimensional particles, where $d$ is the dimension of the subspace, and the latter will be called fractons. We will use the phrase ``subdimensional particles" as a blanket term for both, since fractons can naturally be viewed as $0$-dimensional particles.
The restriction of charge motion in all of these models has a very natural interpretation in terms of higher-moment conservation laws. For example, whereas the rank 1 spin liquids only had a charge neutrality constraint, we will see that particle configurations in a rank 2 spin liquid with scalar charge must have both charge neutrality and vanishing dipole moment. Other higher rank theories can be understood in terms of similar conservation laws. These extra conservation laws lead immediately to subdimensional behavior in these models.
While all of these models are stable against the usual Polyakov mechanism for confinement\cite{polyakov}, where monopoles proliferate and gap the gauge field, we find that a subset of these models exhibit a new, milder form of ``confinement," which we term electrostatic confinement, arising from the large energy stored in the electric field of a static point charge. This mechanism involves a large energy cost for widely separated particles, but it does not gap the gauge field. And in many cases, surprisingly, it still leaves the particles as well-defined excitations, so most of the usual properties of deconfined phases are still present. In other cases, however, electrostatic confinement does destroy the charge sector of the theory, but subdimensional neutral bound states are still possible and the gapless sector remains untouched, so these theories are still far from trivial. This unconventional confinement mechanism is closely connected with the restricted motion of fractons.
The models here represent gapless generalizations of the original discrete fracton models. It seems clear that, by considering discrete symmetric tensor gauge theories, such as by breaking the $U(1)$ theories down to a discrete subgroup, we could generate a large class of gapped models with subdimensional particles. It seems likely that the ``generalized gauge theory" of Reference \onlinecite{fracton2} can be naturally phrased in terms of a higher rank gauge theory (not necessarily symmetric). It will be an interesting future project to see if all of the fracton models can be described in terms of higher rank gauge theories, and to investigate what the precise relation is between higher rank gauge theory and the techniques from algebraic geometry used in the original fracton constructions.
We also note that the higher rank gauge theories described here might have been called ``higher spin" gauge fields by our high energy colleagues. (We have avoided this terminology, since we have abused the term spin enough already.) Fields of spin higher than 2 receive much less attention in the high energy literature than their lower spin counterparts, since it is hard to construct consistent interacting theories. For many years, the only consistently interacting higher spin framework has been Vasiliev theory\cite{vasiliev} and its variants, which involve an infinite tower of all possible higher spin gauge fields. In light of the present work, it would seem that individual higher spin gauge fields actually can be consistently coupled to matter, as long as the matter particles have subdimensional behavior. The difficulties which have plagued the development of higher spin gauge fields in the high energy literature seem to be a manifestation of the fact that such fields cannot be consistently coupled to fully mobile matter fields.
\section{Review of the Rank 1 $U(1)$ Spin Liquid}
The models that we will consider in this paper are natural extensions of the conventional rank 1 $U(1)$ spin liquid in three spatial dimensions. It will therefore be useful to first review the basic structure of this more familiar theory. In such a theory, we can take our basic degrees of freedom to be those of a vector field $A_i$. We let the theory be compact, so that the components of $A_i$ are $U(1)$-valued. In other words, we identify $A_i\sim A_i + 2\pi$, which effectively makes each component a quantum rotor variable. We will also work with the canonical conjugate of the gauge field, which is the electric field $E_i$, as is familiar from classical electromagnetism. $E_i$ effectively serves as the angular momentum of the rotor variable and as such is quantized to have integer eigenvalues. In a concrete lattice model, $A_i$ and $E_i$ are most conveniently taken to live on links of the lattice. The choice of lattice will not be important for the essential physics, though oftentimes the presence of a lattice makes issues like compactness easier to deal with.
We now seek a low-energy theory which has the properties of the familiar Maxwell theory of a $U(1)$ gauge field. Motivated by this, we often characterize a $U(1)$ spin liquid by demanding that the low-energy theory be invariant under the gauge transformation $A_i(x)\rightarrow A_i(x) + \partial_i \alpha(x)$, for a function $\alpha(x)$ with arbitrary spatial dependence. For our purposes, however, it is actually more convenient to work in terms of the canonically conjugate variable $E_i$. For a state to be invariant under $A_i\rightarrow A_i + \partial_i \alpha$, it is easy to show that the state must obey the constraint $\partial_i E^i = 0$, which is simply the source-free Gauss's law. It is also easy to show that the low-energy theory consistent with this gauge constraint/transformation takes the form:
\begin{equation}
H =\frac{1}{2}g\sum_{links} E^2 - \sum_{plaquettes}\cos B
\label{lowen}
\end{equation}
where $B_i = \epsilon_{ijk}\partial^jA^k$ is the magnetic field (conventionally defined on plaquettes of the lattice), and $g$ is a tuning parameter. When $g$ is small, the fluctuations of the cosine around its minimum will be small, and we can write:
\begin{equation}
H\rightarrow\int d^3x\frac{1}{2}(gE^2 + B^2)
\end{equation}
which should look familiar from electromagnetism. This Hamiltonian leads to a gapless photon mode with linear dispersion and two polarizations, as we expect. At large $g$, however, the electric term dominates, effectively disordering the magnetic cosine term, and the system picks up a gap of order $g$. All of the properties of Maxwell theory are destroyed, and the system will no longer be in a spin liquid phase. This is the phenomenon of confinement.
At low energies, we have characterized the $U(1)$ spin liquid by a gapless photon mode whose field configurations obey the constraint $\partial_i E^i = 0$. However, there is another important set of excitations to consider. While the low-energy sector obeyed a constraint, there can be states higher up in energy which violate this constraint, $\partial_i E^i = \rho\neq 0$. These violations of the low-energy constraint correspond precisely to the charges of the emergent electric field, which must be present on very general grounds. If we demand that our spin liquid exist within a tensor product Hilbert space (as any real spin liquid does), then the existence of emergent charges is guaranteed \cite{me}. Furthermore, this charge represents a conserved quantity, as no local operator can create a net charge in the system. If we work on a closed manifold, then we are always guaranteed to have no net charge in the system:
\begin{equation}
\int \rho = \int \partial_iE^i = 0
\end{equation}
since we are integrating a total derivative. In general, these charges will exist as gapped excitations of the system and will have a corresponding energy penalty in the Hamiltonian. While Equation \ref{lowen} was valid for the gapless sector, we will more generally have:
\begin{equation}
H =\frac{1}{2}g\sum_{links} E^2 - \sum_{plaq}\cos B + U\sum_{vert}(\partial_iE^i)^2 + \cdot\cdot\cdot
\label{genham}
\end{equation}
where the $U$ term is defined on each vertex of the lattice. The ``$\cdot\cdot\cdot$" represents the terms which do not commute with the gauge constraint, which are irrelevant to the low-energy physics. These ``$\cdot\cdot\cdot$" terms are important, however, as they will dictate the dynamics of the charges.
There is also one last class of excitations to consider: magnetic monopoles. The magnetic field $B_i = \epsilon_{ijk}\partial^j A^k$ seems to automatically obey $\partial_iB^i = 0$. However, we must remember that $A_i$ is only defined modulo $2\pi$, which allows us to more generally have $\partial_iB^i = 2\pi n$, for integer $n$. Such a defect corresponds to a magnetic charge of strength $n$. The properties of these excitations closely mirror those of the analogous electric particles. This is a manifestation of electromagnetic duality. By making use of the constraint $\partial_i E^i = 0$, we can write $E^i = \epsilon^{ijk}\partial_j \tilde{A}_k$ within the low-energy sector, for a dual gauge potential $\tilde{A}$. In terms of this variable, the electric and magnetic fields have simply switched places, giving us a useful dual formulation of the theory.
The most important fact about the magnetic monopoles in this three-dimensional system is that they are excitations with a well-defined energy. We can therefore imagine a phase in which the monopoles are gapped and do not effect the low-energy physics. This allows the $U(1)$ spin liquid to exist as a stable phase of matter. This is in sharp contrast to the two-dimensional case, in which magnetic monopoles are ``instantons" (spacetime defects) corresponding to phase slip events, not excitations. There is no meaningful sense in which instantons can be gapped, so there is no guarantee that they will not effect the low-energy physics. Indeed, it was shown by Polyakov\cite{polyakov} that the two-dimensional $U(1)$ gauge theory is totally destroyed by these instantons. There do exist schemes for stabilizing a two-dimensional $U(1)$ spin liquid against instantons, such as by coupling to a large number of gapless charges\cite{stability}, but this is a more complicated problem. In this work, we will only consider theories which are instanton-free and are therefore stable.
\section{Rank 2 Theories}
We now proceed to the higher rank theories. We will first go through the rank 2 case to illustrate the principle, which is readily generalized. We take our degrees of freedom to be those of a compact $U(1)$-valued symmetric tensor $A_{ij}$, with canonically conjugate variable $E_{ij}$, representing a generalized electric field. Each component $A_{ij}$ is essentially a quantum rotor, and the momentum variable $E_{ij}$ is quantized to have integer values. In the simplest lattice models\cite{cenke2}, off-diagonal elements ($e.g.$ $E_{xy}$) naturally live on faces of a lattice, while diagonal elements ($e.g.$ $E_{xx}$) naturally live on vertices, but the precise choice of lattice system will not be important for the discussion here. As discussed in Reference \onlinecite{alex}, there are three different sorts of gauge transformations of $A_{ij}$ which can be considered at rank 2, leading to three different analogues of Gauss's law which one can write down: $\partial_i E^{ij} = 0$, $\partial_i\partial_j E^{ij} = 0$, and $E^i_{\,\,i} = 0$, some of which can be applied on top of each other. Furthermore, each valid combination of Gauss's laws will represent a stable phase, as we will discuss.
To construct Hamiltonians for these theories, the authors of Reference \onlinecite{alex} considered the natural generalization of the rank 1 compact U(1) Hamiltonian, Equation \ref{genham}. The generalized $E$ term and gauge constraint term can be written down immediately, whereas the generalized $B$ term requires a bit more cleverness. The structure of the $B$ term depends on the gauge constraint, and there can be different numbers of spatial derivatives in $B$ depending on the theory, leading to different dispersions for the gauge mode. We will delay discussion of the magnetic tensor until a later section, since most of our analysis will not need to make any use of the specifics of these Hamiltonians, except for the $U$ term enforcing the generalized Gauss's law. Almost all of the important physics follows directly from the Gauss's law. The other terms in the Hamiltonian only serve to define the dynamics of the gapless gauge mode and the structure of the magnetic defects. We will, of course, need to check later that these magnetic defects are not instantons, so that the theory is stable. This will indeed be the case, so that these phases will all have a stable deconfined phase at small $g$ (and obviously a trivial confined phase at large $g$). As first shown in Reference \onlinecite{alex}, many of the models considered in this paper will have an electric-magnetic duality, so the behavior of the magnetic particles will often be the same as that of the electric particles, which we focus on first. All we will need for the present discussion is that the gauge field is not confining\cite{foot4} at small $g$, so that particles exist as well-defined excitations in this phase.
It is also worth noting that we expect such rank 2 symmetric tensor theories to have some relationship with the theory of gravity, which is also described by a symmetric tensor gauge field. There is actually a deep connection between the models considered here and emergent gravity, but this relationship will not be apparent at the level of the analysis we will conduct here. The emergent gravitational behavior of these phases is a topic of its own and is being treated in a separate work.\cite{mach}
\subsection{Scalar Charge Theory}
Let us first take the example of imposing only the constraint $\partial_i\partial_j E^{ij} = 0$, corresponding to the gauge transformation $A_{ij} \rightarrow A_{ij} + \partial_i\partial_j \phi$ for arbitrary scalar function $\phi$. Of course, the source-free gauge constraint applies only to the low-energy subspace, achieved for example via a term in the Hamiltonian of the form $U(\partial_i\partial_j E^{ij})^2$ for large $U$. States which violate the source-free Gauss's law must appear higher up in energy as particle states of the theory in order to have a tensor product Hilbert space, as is the situation in any condensed matter problem (see Reference \onlinecite{me} for further discussion of this issue). For a general state, we can therefore write the generalized Gauss's law as $\partial_i\partial_j E^{ij} = \rho$, defining $\rho$ as the scalar charge density.
So what conservation laws do we have in this system? Obviously we have charge neutrality, just as in the rank 1 case:
\begin{equation}
\int \rho = \int \partial_i\partial_j E^{ij} = 0
\end{equation}
where the integrals are over three-dimensional space, and we have integrated a total derivative term. (We choose to work on a closed manifold for simplicity, so that the integral of the total derivative vanishes. Everything works similarly on an open manifold.) This conservation law leads to the usual constraint that the emergent charges cannot be created or destroyed unless it is accompanied by the creation/destruction of other charges in order to preserve neutrality. Naively, one such allowed neutrality-preserving operation is a local hop: a particle is destroyed on one site and created on a neighboring site, in accordance with our usual intuition of particle mobility.
However, interestingly, this rank 2 theory has an additional dipolar conservation law:
\begin{equation}
\int \vec{x} \rho = \int x^k \partial_i\partial_j E^{ij} = -\int \partial_j E^{kj} = 0
\end{equation}
where we have integrated by parts in the middle step.\cite{foot5} (The choice of origin for $\vec{x}$ is arbitrary, since the system is neutral.) In this theory, therefore, any creation/annihilation operation must not only respect the neutrality of the system, but also the vanishing of its dipole moment. As a concrete example, take the lattice model discussed in Reference \onlinecite{cenke2}, where the diagonal components ($E_{xx}$, $E_{yy}$, and $E_{zz}$) live on the vertices of a cubic lattice, and the off-diagonal components ($E_{xy}$, $E_{xz}$, and $E_{yz}$) live on the faces, with all components taking integer values. The basic creation and annihilation operators can be found by examining the effect of changing one component of $E$ at a single location by 1 unit. Doing so leads to two distinct types of creation and annihilation operators, as shown in Figures \ref{fig:quadrupole1} and \ref{fig:quadrupole2}.
\begin{figure}[t!]
\includegraphics[scale=0.35]{quadrupole1.jpg}
\caption{Increasing an off-diagonal component of $E_{ij}$ by 1 creates excitations at the four corners of a plaquette, in a quadrupolar configuration.}
\label{fig:quadrupole1}
\end{figure}
\begin{figure}[t!]
\includegraphics[scale=0.4]{quadrupole2.jpg}
\caption{Increasing a diagonal component of $E_{ij}$ by 1 creates three excitations in a line, again with vanishing charge and dipole moment.}
\label{fig:quadrupole2}
\end{figure}
The uniting feature of all such operators is that they correspond to quadrupolar configurations of charge, obeying both charge neutrality and vanishing dipole moment. In fact, it would seem that this quadrupolar principle is the fundamental feature of this model which would allow it to be generalized to other types of lattices besides cubic. Putting rotors on the vertices and faces of a cubic lattice allowed for the simplest lattice regularization, since there were effectively six degrees of freedom at each location, corresponding to the six degrees of freedom of a 3$\times$3 symmetric tensor. Similarly, the simplest lattice regularization of a rank 1 $U(1)$ gauge theory would be on the links of a cubic lattice, giving us three degrees of freedom per site. Nevertheless, the rank 1 theory can be defined on any lattice, with the fundamental feature being that the electric strings live on links, and these electric strings can essentially be thought of as tiny dipoles. From this perspective, the rank 2 generalization is quite natural. Whereas in the rank 1 case microscopic bosonic degrees of freedom were mapped onto dipole creation operators, in a rank 2 theory they map onto quadrupolar creation operators. We will see that a similar story holds for other higher rank gauge theories.
The consequences of this extra conservation law for the motion of individual charges are quite dramatic. First of all, let us examine an isolated point charge, well-separated from any other charges in the system. (This situation can indeed be created, for example by a ``membrane" operator, formed by applying the operator of Figure \ref{fig:quadrupole1} on every plaquette of a two-dimensional surface. Isolated charges will naturally occur at the corners of this membrane.) We then wish to move such an isolated charge to an adjacent site by means of a local operator. Upon examining our toolbox, it is obvious that there is no local operator that can hop the particle without simultaneously creating other excitations. This is a simple consequence of the conservation of dipole moment. Moving the particle by itself changes the global dipole moment of the system and must be compensated by the creation/destruction/motion of other particles. Such a particle excitation, which cannot move without the creation of extra excitations, was dubbed a ``fracton" in Reference \onlinecite{fracton1}, and we will stick with that convention. The difference between the present case and the previously studied fracton models is that the models here also possess gapless gauge modes (which propagate normally).
It should be noted that not all excitations in the system have this fracton property. Whereas isolated point charges are fractons, a dipolar bound state of two opposite charges will be freely hopping. The operators seen in Figures \ref{fig:quadrupole1} and \ref{fig:quadrupole2} represent the transverse and longitudinal hops of these dipoles, so dipolar excitations can freely propagate in any direction in space.
\subsection{Vector Charge Theory}
There is another important class of rank 2 symmetric $U(1)$ spin liquids, which has related properties, but with important differences. In this theory, we take the Gauss's law constraint on the ground state to be $\partial_i E^{ij} = 0$, corresponding to the gauge transformation $A_{ij}\rightarrow A_{ij} + \partial_i \lambda_j + \partial_j \lambda_i$ for arbitrary vector $\lambda_i$. As in the previous case, we regard the violations of the ground state Gauss's law as the particle states of the theory, defining a vector-valued charge as:
\begin{equation}
\rho^j = \partial_i E^{ij}
\end{equation}
Once again, we first identify the various conservation laws of the system. We obviously have the natural analogue of charge neutrality in the system:
\begin{equation}
\int \vec{\rho} = \int \partial_i E^{ij} = 0
\end{equation}
We can also consider the moment of this vector charge around an arbitrary origin:
\begin{equation}
\int \vec{x}\times \vec{\rho} = \int \epsilon_{ijk} x^j \partial_n E^{nk} = -\int \epsilon_{ijk}E^{jk} = 0
\end{equation}
where we have once again integrated by parts, and the last step follows since $E^{ij}$ is a symmetric tensor. If we heuristically regard $\vec{\rho}$ as a ``momentum density", then the two conservation laws here represent the conservation of linear and angular momentum. And just as any particle creation/annihilation operator in the scalar case needed to respect the neutrality and dipolar conservation laws, any creation/annihilation operator in this theory must respect both linear and angular momentum conservation.
\begin{figure}[b!]
\includegraphics[scale=0.45]{vector1.jpg}
\caption{Increasing a diagonal component of $E_{ij}$ by 1 creates two oppositely directed vector charges on adjacent collinear links.}
\label{fig:vector1}
\end{figure}
\begin{figure}[b!]
\includegraphics[scale=0.35]{vector2.jpg}
\caption{Increasing an off-diagonal component of $E_{ij}$ by 1 creates a ``loop" of vector charge around a plaquette, in such a way that both linear and angular momentum are conserved.}
\label{fig:vector2}
\end{figure}
As a concrete example, let us consider the operators present in the rank 2 vector charge model considered in Reference \onlinecite{alex}. Once again, diagonal components live on vertices of a cubic lattice and off-diagonal components live on the faces. The vector charges live naturally on the links of the lattice. There are two sorts of local creation/annihilation operators, as depicted in Figures \ref{fig:vector1} and \ref{fig:vector2}. One creates opposite vector charges on adjacent collinear links. The other creates vector charges in a sort of loop around a plaquette. However, it is important to note that the ``loop" changes orientation at every vertex, in order to respect the conservation laws. (A unidirectional loop would violate angular momentum conservation.) The first of these operators, acting on collinear links, can be thought of as a longitudinal hopping operator for the vector charge, allowing each particle to hop freely along the direction of its charge. However, it is easy to see that an analogous transverse hopping operator is disallowed, since it would violate conservation of angular momentum.\cite{foot6} We can regard the plaquette operator as generating a transverse hop, but only at the cost of introducing two new excitations on perpendicular links to compensate for the change in angular momentum. Therefore, the vector charges are ``one-dimensional particles." They are restricted to motion only along their charge vector, being effectively ``confined" in the transverse direction due to the need to create extra particles. Note that one-dimensional particles are not unique to the $x$ direction or $y$ direction in Figures \ref{fig:vector1} and \ref{fig:vector2}. A bound state of an $x$ charge and a $y$ charge will be freely hopping in the $(1,1,0)$ direction, with operators like Figure \ref{fig:vector2} generating the hops of bound states.
Once again, the principles identified here provide for a natural way to generalize the model to any lattice, beyond the cubic models considered in previous work \cite{alex,cenke2}. The principle is to engineer a Gauss's law which maps the local bosonic degrees of freedom onto creation/annihilation operators on links which are consistent with conservation of linear and angular momentum. Since the vector charges naturally live on links, the gauge variable will naturally live either on faces, sites, or both. The vector charges will be able to freely hop only along one-dimensional subspaces.
\subsection{Traceless Rank 2 Theories}
We have so far discussed two different forms of gauge constraints, $\partial_i\partial_j E^{ij} = 0$ and $\partial_i E^{ij} = 0$. If we had the second constraint, then the first is automatic, so there is no sense in combining these two gauge constraints. The only additional thing we can add on top is a tracelessness constraint, $E^i_{\,\,i} = 0$. There is actually a subtle distinction regarding whether we treat this constraint as being breakable or not. In all previous Gauss's law constraints, derivatives were present, and allowing for gauge charges was necessary to preserve the tensor product structure of the Hilbert space (see Reference \onlinecite{me} for further discussion). However, for a local constraint such as the trace condition, there are two options. We can allow the constraint to be broken, in which case there are trace charges in the theory, $E^i_{\,\,i} = n$, for integer $n$. Alternatively, we are free to stipulate that our microscopic variables were traceless and say that $E^i_{\,\,i} = 0$ identically in the Hilbert space. We shall focus on this latter case, since it leads to more novel physics. Trace charges would only add in extra local scalar degrees of freedom which could modify the energetics of our system, but otherwise not do much harm.
The tracelessness condition looks fairly benign at first glance. Its only effect is to even further constrain the particles with an additional conservation law. For example, consider the scalar charge theory, in which case the extra conservation law is:
\begin{equation}
\int x^2\rho = \int x^kx_k\partial_i\partial_j E^{ij} = 2\int E^i_{\,\,i} = 0
\end{equation}
It can readily be checked that this conservation law rules out linear quadrupole operators such as that in Figure \ref{fig:quadrupole2}, only allowing square quadrupoles such as that in Figure \ref{fig:quadrupole1} and various bound states of the linear quadrupoles. The fractons remain fractonic, so not much changes in the charge sector. However, the neutral dipolar bound state, which was formerly freely hopping, now becomes a two-dimensional particle, as its longitudinal hopping operator has been removed from the theory.
However, the tracelessness condition has a much more dramatic effect on the vector charge theory. Without tracelessness, the two conservation laws of this theory were conservation of ``linear momentum" and ``angular momentum." In the presence of the tracelessness constraint, there is yet another conservation law:
\begin{equation}
\int \vec{x}\cdot \vec{\rho} = \int x^j\partial_i E^{ij} = -\int E^i_{\,\,i} = 0
\end{equation}
In the presence of this extra conservation law, the vector charges become fully fractonic in nature. Longitudinal hops, formerly allowed in the traceful theory, are now in violation of the new conservation law. The formerly one-dimensional particles have been turned into fractons via this extra gauge condition. As a general principle, adding in additional gauge constraints can only lead to an increase in the number of conservation laws of the system, thereby putting extra constraints on particle motion.
\subsection{``Mixed" Gauge Constraints}
Strictly speaking, there is still one more layer of complication which can be thrown into these theories, which we shall mention only in passing. Until now, we have always taken the gauge transformation to be on $A$, while the constraint is on $E$, as seems usual. One could also think up theories in which $A$ and $E$ switch roles, which is probably not dramatically different, resulting in a dual description of the same phases. However, we could also consider a ``mixed" gauge theory, where constraints are simultaneously imposed on both $A$ and $E$, so long as care is taken to ensure that these constraints commute. For example, the gauge transformations and constraints of the emergent gravity model in Reference \onlinecite{gu} are effectively:
\begin{equation}
A_{ij}\rightarrow A_{ij} + \partial_i\lambda_j + \partial_j \lambda_i\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\partial_i E^{ij} = 0
\label{gauge1}
\end{equation}
\begin{equation}
E_{ij} \rightarrow E_{ij} + (\delta_{ij}\partial^2 -\partial_i\partial_j)\phi \,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\, (\delta_{ij}\partial^2 - \partial_i\partial_j)A^{ij} = 0
\label{gauge2}
\end{equation}
for arbitrary gauge parameters $\lambda_i$ and $\phi$. It is easy to verify that the two constraints commute, simply by examining the effect of the opposite gauge transformation on each constraint. We therefore see that mixed gauge constraints are in principle possible, though it is unclear when such theories are physically reasonable. For example, in the rank 1 case, we could have tried to simultaneously impose both the normal Gauss's law, $\partial_iE^i = 0$, and also a curl constraint on $A$ as $\nabla \times\vec{A} = \vec{B} = 0$. However, imposing this extra magnetic flux constraint takes us to a ``sick" limit of the $U(1)$ gauge theory, where the speed of light has gone to infinity, so there is no longer a sensible gapless gauge mode. As shown in Reference \onlinecite{gu}, it is possible in certain circumstances for such mixed gauge constraints to lead to reasonable theories. It is unclear under what conditions these mixed gauge constraints will leave us with a reasonable theory, as opposed to some ``sick" theory with infinite velocities.
For these theories, we have both an ``electric" gauge constraint and a ``magnetic" gauge constraint, both of which have corresponding particles. It is important to note that these particles are not the magnetic defects of the theory. On top of these magnetic gauge constraint particles, the theory can still have magnetic defects arising from the compact nature of the gauge variable. These mixed gauge theories therefore have a particularly complicated particle spectrum.
\subsection{Magnetic Defects and Stability}
Up to this point, we have technically only been speaking in terms of the ``electric" particles of the theory, $i.e.$ those coming from violations of the Gauss's law constraint. However, we must also consider the magnetic defects of the theory. We argued earlier that, in the rank 1 $U(1)$ gauge theory, the magnetic defects have basically the same structure as the electric charges, as determined by an electric-magnetic duality. Similar stories hold in many of the higher rank theories. For example, take the rank 2 vector charge theory, with low-energy gauge constraint $\partial_i E^{ij} = 0$. Writing down the gauge invariant quantity with the fewest number of derivatives, the magnetic field in this case was shown\cite{alex} to also be given by a rank 2 symmetric tensor:
\begin{equation}
B_{ij} = \epsilon_{iab}\epsilon_{jcd}\partial_a\partial_c A_{bd}
\end{equation}
which appears to obey $\partial_i B^{ij} = 0$. However, once again, due to the compactness of $A$, we can have $\partial_i B^{ij} = 2\pi n^j$, for integer valued vector $n^j$. By the definition of $n^j$, we can apply the same arguments used on the vector electric charge to show that these vector defects obey conservation of both ``linear momentum" and ``angular momentum," so they are also one-dimensional particles. As in the conventional $U(1)$ spin liquid, this could have been determined by electric-magnetic duality. As shown in Reference \onlinecite{alex}, one can use the low-energy gauge constraint to write $E^{ij} = \epsilon^{iab}\epsilon^{jcd}\partial_a\partial_c \tilde{A}_{bd}$ for a dual gauge field $\tilde{A}_{ij}$, which effectively swaps the role of $E$ and $B$ in the theory, so the magnetic charges must have the same structure as the electric charges. For any theory with such duality, it is automatically the case that the magnetic charges have the same subdimensional behavior as the electric charges, which is a nice result. In particular, this immediately tells us that the magnetic defects are excitations, not instantons. This allows us to consider a regime in which the magnetic excitations are gapped and are irrelevant to the low-energy physics, providing us with a stable phase of matter.
There are also some phases which are not self-dual. For example, consider the scalar charge theory, without any trace conditions. One can show that the lowest order magnetic field tensor which can be written down has the non-symmetric form\cite{foot9} $B_{ij} = \epsilon_{iab}\partial^a A^b_{\,\,j}$. Despite the lack of self-duality, the theory still has magnetic monopole excitations of the form $\partial_iB^{ij} = 2\pi n^j$, which one can show are 2-dimensional particles, moving transversely to their charge vector. The presence of pointlike magnetic monopole excitations seems to be quite universal to these symmetric tensor phases, which indicates that they are all stable phases of matter.
We can now say that these phases are stable against confinement by instanton proliferation, but one might worry that one deconfined phase could be unstable to a different deconfined phase. For example, suppose we take the gauge constraint to be $\partial_i\partial_j E^{ij} = 0$. Our ideal Hamiltonian will have the schematic form:
\begin{equation}
H = E^2 + B^2 + U(\partial_i\partial_j E^{ij})^2
\end{equation}
where $B$ is the appropriate magnetic tensor. However, one might complain that we could throw in a term corresponding to the lower order gauge constraint of the vector charge theory:
\begin{equation}
H = E^2 + B^2 + U(\partial_i\partial_j E^{ij})^2 + U'(\partial_i E^{ij})^2
\label{thisone}
\end{equation}
The $U'$ term naively looks more relevant, since it has fewer derivatives. If this were indeed true, then the scalar charge rank 2 theory would flow towards the vector charge rank 2 theory. However, this is not the case. We cannot view the last two terms of Equation \ref{thisone} on equal footing, since the particular $B$ here commutes only with the first gauge constraint, not the second. The $B$ term and the $U'$ term are in competition. At small $U'$, the magnetic term will win out and the $U'$ term is irrelevant. Another way to see this is to note that the addition of the $U'$ term only modifies the dispersion relation from $\omega^2-k^4 = 0$ to $\omega^2 + U'\omega^2k^2 - k^4 = 0$, which is an irrelevant change at low energies. Of course, at large enough $U'$ (far away from the fixed point), the vector gauge constraint will take over. But for small perturbations, the scalar charge fixed point is a locally stable one against the vector perturbation. Similar arguments hold for the stability of any of the gauge constraint combinations against any other. The $B$ term for each theory is constructed specifically for that gauge constraint and will destroy any other gauge constraint added in perturbatively.
\section{Electrostatic Confinement}
(Special thanks are due in this section to Senthil Todadri, who first pointed out that electrostatic energy may be an issue in these models.)
Before moving ahead with these models, we need to take a close look at the energy needed to create well-separated fractons. Besides the energy $U$ representing the mass of the particles, we must also consider the energy stored in the ``electric field" $E_{ij}$ of a static point charge. (More correctly, we are looking at the expectation value $\langle E_{ij}\rangle$, but we shall simply write $E_{ij}$ to avoid clutter.) Suppose we have a static delta function source for our Gauss's law, representing an isolated point charge. In the vector charge theory, which has one derivative in Gauss's law, we have:
\begin{equation}
\partial_i E^{ij} = v^j \delta(r)
\end{equation}
for some constant vector $v^j$. Fourier transforming gives us:
\begin{equation}
q_i \tilde{E}^{ij} = v^j
\end{equation}
The scaling of our electric field then behaves as:
\begin{equation}
\tilde{E}^{ij}(q) \sim \frac{1}{q}
\end{equation}
\begin{equation}
E^{ij}(r) \sim \int d^3q \,\tilde{E}^{ij}(q) e^{i\vec{q}\cdot\vec{r}} \sim \frac{1}{r^2}
\end{equation}
like a normal Coulomb field. The electrostatic energy associated with these single-derivative theories is much the same as in a rank 1 $U(1)$ spin liquid, and no further comments are necessary.
On the other hand, consider a delta function source for the scalar charge theory:
\begin{equation}
\partial_i\partial_j E^{ij} = \delta(r)
\end{equation}
The scaling of the electric field now behaves as:
\begin{equation}
\tilde{E}^{ij}(q) \sim \frac{1}{q^2}
\end{equation}
\begin{equation}
E^{ij}(r) = \int d^3q\, \tilde{E}^{ij}(q) e^{i\vec{q}\cdot\vec{r}} \sim \frac{1}{r}
\end{equation}
so the electric field in this case only falls off as $1/r$, as opposed to the $1/r^2$ behavior of the usual three-dimensional Coulomb field. This is quite a significant difference, since the energy stored in the field is given by:
\begin{equation}
\int d^3r E^2 \sim \int dr\, r^2 \frac{1}{r^2} \sim R
\end{equation}
where $R$ is a large-distance cut-off. In the usual Coulomb field, $1/r^2$, the integral converges at large distances, indicating a finite amount of energy stored in the electric field of a point particle. (Short-distance divergences are of course cut off at a finite value in any condensed matter context.) For the rank 2 scalar charge theory, however, the electrostatic energy of a truly isolated point charge is actually infinite (cut off by the system size). And if a group of particles with zero net charge and dipole moment is separated from each other by a distance of order $R$, then the natural energy scale of this system is of order $R$. We have therefore found that a configuration with particles separated by distance $R$ requires an energy linear in the separation. This is the standard metric by which one normally measures confinement.
However, it is important to note that this is a different sort of ``confinement" which is much milder than the usual notion. Usually, one speaks in terms of the interaction between particles. If a linear potential, $V(R)\sim R$, exists between two particles, then there is an attractive force which is independent of distance, always seeking to recombine the particles into the vacuum, indicating that the particles never truly become independent and are thus confined. However, in the fracton case, the electric field of a point particle dies off as $1/r$, so at large separation, the particle's do not feel each other's fields. The linear energy associated with separating particles does not arise from one particle working against the field of the other, but rather from the large energy associated with building up each individual charge's electrostatic field.
Normally, these two notions are equivalent to each other. For example, a standard undergraduate electromagnetism problem\cite{griffiths} is to calculate the electrostatic energy of a uniformly charged ball. One way to do the problem is to examine the energy stored in the final electric field configuration. Another way to do it is to imagine building up the ball incrementally and calculate the work done in bringing each charge into the ball against the field of the charges already present. In the present problem, however, the two descriptions are inequivalent, as we just found. The resolution to this seeming paradox is that there simply is no notion of ``force" (or even equations of motion) when dealing with fractons. Since fractons cannot hop, there is no sense in which one can do work to move a particle against the field. Any treatment involving incremental energy changes would need to account for the creation of extra particles along the way, which seems like a complicated task. It is much simpler to just look at the final electrostatic energy, which unambiguously yields the answer.
The fractons in this theory are somewhat ``confined," in the sense that it requires a macroscopically large energy to acquire well-separated fractons, but many aspects of traditional Polyakov confinement do not carry over. First of all, the gauge field is not gapped out (none of the present issues affect the stability arguments of Reference \onlinecite{alex}), so this ``confinement" does not destroy the Coulomb phase. Furthermore, once the large energy cost has been paid to create well-separated fractons, there will be no ``restoring force" on an isolated fracton which seeks to recombine it with its neutralizing partners into the vacuum. A fracton cannot move towards its neutralizing partners without creating extra particles, the cost of which will energetically outweigh any marginal decrease in the electrostatic energy.
In this sense, it is still perfectly reasonable to speak of these fractons as well-defined excitations of the system. The isolated fracton state is stable, just energetically costly, much like vortices in a two-dimensional superfluid. A configuration with well-separated fractons is too energetically costly to be created via thermal fluctuations, so they will have little effect on low-energy thermal properties of the system. But if one prepared a system with well-separated fractons, then these fractons would continue to exist as stable excitations of the system. We will give this new milder version of confinement the name ``electrostatic confinement." For this rank 2 theory, electrostatic confinement actually has more in common with a conventional deconfined phase (well-defined particle excitations and a gapless gauge field) and the main effect of the confinement is just to inhibit the fractons from being generated by thermal fluctuations.
Despite its benign behavior in rank 2 theories, we will later face examples at rank 3 and higher where electrostatic confinement is more severe. In the traditional Gauss's law with one derivative, we had $E\sim 1/r^2$. In the case with two derivatives, we have $E\sim 1/r$. In a theory with three derivatives in Gauss's law, such as $\partial_i\partial_j\partial_k E^{ijk} = 0$, we would have $E\sim \log(r)$, which actually increases with distance. For higher derivatives, the growth is even faster. In such a theory, we face the additional problem that, not only is the total electrostatic energy divergent, but the energy \emph{density} at large distances diverges as well. At some critical distance, the energy density would be large enough to start nucleating new particles out of the vacuum. This would then place a fundamental limitation on the isolatability of an individual fracton. Thus, in theories with three derivatives or more in Gauss's law, isolated fractons do not appear as stable excitations at long-distances. Such theories will still have protected gaplessness, but all particles will only occur in neutral bound states. The neutral bound states can actually still be subdimensional in certain cases, so these theories are far from trivial, but the charge sector will be totally destroyed.
\section{Generalization to Higher Rank}
We can very straightforwardly generalize the principles of the rank 2 theories to higher rank cases. At higher ranks, there are a wider variety of available Gauss's laws, as discussed in Reference \onlinecite{alex}. All of these theories can be analyzed from the same perspective as the rank 2 case: identify all charge conservation laws, then consider all local creation/annihilation operators respecting those conservation laws. For one- or two-derivative Gauss's laws, we can then immediately identify all subdimensional behavior in the theory. As noted earlier, if there are three or more derivatives present in the Gauss's law, then the fundamental charge sector will be destroyed by electrostatic confinement, even though the gapless gauge mode is still protected. Neutral bound states will still exist as well-defined excitations, and may still be subdimensional. Such theories are therefore not uninteresting, but are slightly less interesting than the one- and two-derivative cases, which have well-defined charge sectors. We will focus on these two cases for the sake of brevity. Any other specific case can be studied via the same principles.
In the first case, we take our Gauss's law to have two derivatives, $\partial_{i_1}\partial_{i_2} E^{i_1...i_n} = \rho^{i_3...i_n}$. (We can also allow the other indices to be contracted with each other in various ways. Contracting internal indices of $E$ just lowers its effective rank, which would allow us to study it by previous methods.) Since there are two derivatives, it is easy to see that the first two charge moments vanish:
\begin{equation}
\int \rho^{i_3...i_n} = 0
\end{equation}
\begin{equation}
\int x^j \rho^{i_3...i_n} = 0
\end{equation}
after performing an integration by parts in the second equation. When coupled together, these equations imply that isolated charges are fully immobile. Such a tensor charge cannot hop to an adjacent site while preserving both charge and first moment. Thus, two-derivative models are all fracton theories. Just as in the rank 2 case, two-derivative theories at any rank will have a $1/r$ electric field for isolated charges. Electrostatic confinement will then prevent the particles from being generated via thermal fluctuations, but they will still exist as stable states. (Higher derivative theories would also have all charged particles be fractonic, but these charges would be destabilized by electrostatic confinement.)
The other case occurs when there is one derivative in Gauss's law, $\partial_{i_1} E^{i_1...i_n} = \rho^{i_2...i_n}$. We obviously have charge conservation:
\begin{equation}
\int \rho^{i_2...i_n} = 0
\end{equation}
But we no longer have the vanishing of a generic first moment:
\begin{equation}
\int x^j \rho^{i_2...i_n} = \int x^j \partial_{i_1} E^{i_1...i_n} = -\int E^{ji_2...i_n} \neq 0
\end{equation}
However, by taking advantage of the symmetry of $E$, a fully symmetric tensor, we can easily construct a contracted first moment which does obey a conservation law:
\begin{equation}
\int \epsilon^{nji_2} x_j \rho_{i_2...i_n} = -\int \epsilon^{nji_2} E_{ji_2...i_n} = 0
\end{equation}
by the symmetry of $E$. In the rank 2 case, the analogous conservation law was conservation of angular moment, which still allowed the charges to freely hop along the direction of their charge vector. However, for rank higher than 2, the charges have at least two indices. In this case, the conservation law is more constraining. For a charge $\rho_{i_2...i_n}$ to be freely hopping in direction $v_j$, we require $\epsilon^{nji_2}v_j\rho_{i_2...i_n} = 0$ for all $n$, so that the moment does not change with hops in the $v_j$ direction. For this to be true for arbitrary $n$ (and considering the index symmetry of $\rho$), $\rho$ cannot have components in any index direction except the $v_j$ direction. In other words, a charge can only propagate in direction $v_i$ if the charge tensor has the specific form $\rho_{i_2...i_n} = v_{i_2}v_{i_3}...v_{i_n}$. A generic tensor will have no directions of free propagation and will therefore be a fracton. This case therefore has a mix of both one-dimensional particles and fractons (with most charges being fractons).
\section{Conclusion}
We have shown that spin liquids described by higher rank $U(1)$ gauge theories \cite{alex} lead naturally to subdimensional particle excitations, as in fracton models \cite{fracton1,fracton2}. The theories with scalar charge are fractonic, with particles unable to hop in any direction without generating additional excitations. The vector charge theories can be either fractonic or have one-dimensional excitations. These models provide a natural gapless generalization of the original discrete fracton models. Furthermore, it seems like tensor gauge theories may provide a more general framework for understanding and classifying phases with subdimensional particles. There is much work to be done on this front, and exciting new theoretical questions are waiting to be answered. Do all of the previously discovered fracton models fit into the tensor gauge theory framework? How do these formulations relate to previous work relying on algebraic geometry? How does one obtain the fractal structure of Haah's code? Also, recent work has indicated that discrete fracton models can be obtained from a layer construction method, where layers of two-dimensional topologically ordered systems are strongly coupled together\cite{sagarlayer,hanlayer}. It will be interesting to see if the $U(1)$ fracton models can be obtained from similar constructions.
There are plenty of even more exotic (and perhaps less well-defined) questions. Can we get anything new by considering non-abelian generalizations of these tensor gauge theories? What is the nature of transitions to the confined phase? Can subdimensional particles condense? Can subdimensional particles exist on curved submanifolds? For example, could we construct a model in which two-dimensional particles are restricted to a sphere instead of a plane? There are many new avenues to explore.
On a more practical level, how and where can we observe such subdimensional particles in the lab? As to how, we note that the conservation laws of these theories will have dramatic consequences for thermalization properties of these phases\cite{abhinav}. Any energy injected into the particle modes of these systems tends to be localized to a particular subspace (or to one particular point, in the case of fractons). Thus, these systems have severe obstructions to thermalization and will have unusual thermal transport properties. Furthermore, this localization exists in the clean limit, without the disorder that is often needed in the context of many-body localization, which may serve as a way to distinguish subdimensional particles from more conventional localization effects.
The specific lattice models of Reference \onlinecite{alex} are theoretically well-established to be in the higher rank $U(1)$ spin liquid phase. These models apply most directly to a system of quantum rotors, which is essentially equivalent to a large $S$ limit of spins. Such a quantum rotor system could possibly be engineered directly in ultra-cold atom systems, which philosophically is enough to establish the reality of the phases described here. But there is every reason to believe that these phases are realizable in Mott-insulating solid-state systems as well. A finite $S$ system, even a simple spin-$1/2$ system, can flow towards a rotor description under the renormalization group\cite{balents}. The precise choice of microscopics is unimportant. The key ingredient for these models is the Gauss's law, which corresponds to some strong frustrating interaction of the underlying spins. The precise form of the geometric frustration will necessarily be more complicated in these models than in the rank 1 case, since spin flips should create more than just two particles. But such geometries are definitely accessible, such as the face-centered cubic geometry of the models in Reference \onlinecite{alex}. It remains to be seen if any fcc material will have the precise interaction structure necessary to produce higher rank $U(1)$ spin liquids, or if a more complicated geometry is needed. Either way, there is every reason to be optimistic that higher rank $U(1)$ spin liquids are experimentally accessible.
\section*{Acknowledgements}
Numerous thanks are due to my advisor, Senthil Todadri, first and foremost for calling my attention to the large electrostatic energy of these models. He is also due thanks for many insightful discussions, for financial support throughout, and for allowing me sufficient freedom to follow my own interests. I would also particularly like to thank Liujun Zou for numerous helpful discussions during the early stages of this project. Thanks are due to Sagar Vijay, Jeongwan Haah, and Liang Fu for sharing their expertise on fracton models. I would also like to thank Yahui Zhang, Lucile Savary, and Samuel Lederer for useful discussions. This work was supported by NSF DMR-1305741.
\section*{Appendix: What is a Tensor?}
\begin{figure}[t!]
\includegraphics[scale=0.35]{squarescalar.jpg}
\caption{We can regard the spins of this system as living on the sites of a square lattice, in a scalar representation.}
\label{fig:square}
\end{figure}
\begin{figure}[t!]
\includegraphics[scale=0.35]{squarevector.jpg}
\caption{We could regard the same system of spins as living on the links of a tilted square lattice, in a vector representation.}
\label{fig:square2}
\end{figure}
In the main text, we have considered models in which the fundamental bosonic degrees of freedom map onto a ``tensor" structure. This is essentially just a rewriting of our Hilbert space. For example, consider spins on the square lattice in Figure \ref{fig:square}. Under a $\pi/2$ rotation, the lattice returns to itself, and every spin ends up at an equivalent location. The fundamental variables are not endowed with any extra sense of directionality (not counting any internal spin direction), so this is effectively a ``scalar" representation of the spin model, as would be appropriate to describe a ferromagnet, for example. On the other hand, let us now take the exact same spin system, but represent the spins as living on the links of a tilted square lattice, as seen in Figure \ref{fig:square2}. An excitation of the spin can then be represented by a string on this link, which can be directed or undirected, depending on the microscopic details. When we perform a $\pi/2$ rotation, the system still returns to itself. However, spins which formerly lived on links in the $(1,1)$ direction now live on links in the $(-1,1)$ direction, so the spin variables transform as a vector in this representation, which is the correct representation for conventional vector gauge theories. Similarly, in a three-dimensional system, one could view the spins as living on plaquettes, leading naturally to an antisymmetric tensor representation. One can also get a symmetric tensor representation by allowing some spins to live on vertices and others to live on plaquettes, as discussed in the main text. We therefore see that we can always map any given spin system back and forth between various representations of the lattice symmetry group. However, obviously the system is not well-described by all of these tensors at the same time. The question is which representation is convenient for describing the spectrum of the theory. Depending on the Hamiltonian of the system in question, one representation may be more useful than any other. Or sometimes there are multiple useful tensor representations, such as a three-dimensional $\mathbb{Z}_2$ spin liquid, which can be described either as a vector or antisymmetric tensor theory. Typically, such a tensor rewriting of the theory is only useful when there is some local energetic constraint which forces loops to be closed, or surfaces to be closed, or some other convenient geometric constraint. This will give a ``Gauss's law" for the system, which means the system can be described by a gauge theory.
Tensors are usually defined in terms of their behavior under spatial rotations. A generic tensor can be decomposed in terms of certain simpler tensors. For example, a rank 2 tensor can be written as a sum of an antisymmetric tensor, a traceless symmetric tensor, and a scalar trace. Studying these irreducible tensor theories should then allow us to describe arbitrary tensor gauge theories. But throughout the present work, and in any typical condensed matter system, we are always working on a lattice, so there is no full rotational symmetry. However, on any crystalline lattice there still exists some subgroup of the full rotation group which is a symmetry of the system. We should then technically switch to speaking in terms of irreducible representations of the discrete rotational group, instead of the full rotational group, but the situation should not be dramatically different, just a bit more tedious to classify. Plus, it is often the case that full rotational symmetry is restored in the low-energy theory, since anisotropy terms are often irrelevant under the renormalization group.
On this basis, we have focused on tensor representations of the full rotation group in this paper. However, we would like to emphasize, as was noted in Reference \onlinecite{alex}, that this rotational symmetry is not needed to protect the gaplessness of the gauge mode. This gaplessness is enforced by the gauge invariance ($i.e.$ by the source-free Gauss's law of the ground state), which is enforced by the gapped nature of the particles. Essentially, the particle gap protects the gauge mode gaplessness. In turn, the gauge constraint requires a certain structure for the particle sector of the theory, which in the models considered here turns out to be subdimensional. In this sense, these phases are stable against microscopic perturbations, without having to worry about symmetry protection. We also note that our rewriting of the fundamental bosonic degrees of freedom in tensor representations did not assume rotational symmetry of the phase. It merely took advantage of the symmetry of the underlying Hilbert space. These tensor representations could be used just as well to describe a phase with broken rotational symmetry.
|
\section{Introduction}
\label{sec:Intro}
A novel class of functional surfaces has recently been introduced, known as Liquid Infused Surfaces\cite{Epstein2012} (LIS),
Lubricant Impregnated surfaces\cite{Smith2013} (LIS), or Slippery Liquid Infused Porous Surfaces\cite{Wong2011} (SLIPS).
They are typically constructed by infusing rough or porous materials with lyophilic oils.
Thus, as shown in Fig. \ref{fgr:sketch_LIS}, a typical liquid infused system involves three fluids:
the oil phase, and typically a (water) droplet in a gas environment.
They have been shown to exhibit many advantageous wetting properties, including low contact angle
hysteresis, self-cleaning, drag reduction, anti-icing and anti-fouling
\cite{Wong2011,Lafuma2011,Kim2012,Epstein2012,Smith2013,Schellenberger2015,Wexler2015}.
Furthermore, compared to competing technologies such as superhydrophobic surfaces,
LIS are robust against pressure-induced instabilities and failure
\cite{Kusumaatmaja2008a,Hejazi2011,Butt2013,Tuteja2008a}, which makes them
favourable for applications to a wide-range of problems, ranging from marine fouling and product
packaging to heat exchanger and medical devices\cite{Leslie2014,Xiao2013,Epstein2012}.
\begin{figure}[tbh]
\centering
\includegraphics[width=1.0\columnwidth]{Fig1.eps}
\caption{(a) The geometry of an axisymmetric water droplet placed on a liquid infused surface, indicating
the three Neumann angles $\theta_{\rm o}$, $\theta_{\rm w}$ and $\theta_{\rm g}$,
and the wetting angles relative to the oil phase $\theta_{\rm ow}$ and $\theta_{\rm og}$.
Depending on the oil-water contact angle, the water droplet can touch the solid substrate (b)
or be separated by a thin oil layer (c). Similarly, the oil-gas contact angle determines whether
the infusing oil merely hemi-wicks the rough surface (d) or coat the surface corrugations (e).
Based on the combinations of (b-c) and (d-e), four wetting states of interest on liquid infused
surfaces can be distinguished.
}
\label{fgr:sketch_LIS}
\end{figure}
For LIS, oil penetration in between the surface corrugations is key. As such, the material contact
angles characterizing the infusing oil relative to the air and water phases have to satisfy two wicking criteria:
$\theta_{\rm ow}^{\rm Y}<\theta^\ast$ for the oil--water and
$\theta_{\rm og}^{\rm Y}<\theta^\ast$ for the oil--gas interfaces.
In general the critical angle $\theta^\ast$ will depend on the details of the surface corrugations\cite{Semprebon2014a},
but in many cases of interest a simple thermodynamic criterion can be expressed in terms of global statistical quantities.
These are the roughness factor $r$, corresponding to the ratio of total area of the textured surface to its projected area,
and the fraction $f$ of the projected area that is occupied by a solid\cite{Bico2002,Quere2008}
\begin{equation} \label{eq:wicking}
\cos \theta^{\ast} = \frac{(1-f)}{(r-f)}.
\end{equation}
It is worth noting that, even with these thermodynamic criteria satisfied, impalement of the water and gas phases
may still occur at high pressures\cite{Kusumaatmaja2008a,Hejazi2011,Butt2013}, but we will exclude such cases here.
For vanishing contact angles $\theta^{\rm Y}_{\rm ow} = 0^\circ$ or $\theta^{\rm Y}_{\rm og} = 0^\circ$,
it is possible that a thin oil layer will completely cover the surface roughness.
Provided the wicking criteria are satisfied, we can identify four relevant thermodynamic
wetting states\cite{Smith2013}. As illustrated in Fig. \ref{fgr:sketch_LIS}(b-e)
these states depend on the presence and absence of a thin oil film between the water and/or gas phases and the solid surface.
The presence of a thin oil film prevents a direct contact between the water/gas phase
and the solid. It has been proposed to explain the smooth displacement of contact lines on
liquid infused surfaces\cite{Guan2015}, as opposed to stick-slip commonly observed on other surfaces\cite{Varagnolo2013}.
While thermodynamics arguments are sufficient to predict the presence of different wetting states on LIS,
to date there is no theory for computing the corresponding values of the contact angle
and contact angle hysteresis, despite their relevance as key design parameters for any
application. For example, low contact angle and low contact angle hysteresis are preferred
for efficient heat transfer \cite{Xiao2013,Rykaczewski2014}, while high contact angle and low contact
angle hysteresis are desirable for high droplet mobility\cite{Mahadevan1999}.
For standard wetting scenarios, the Young's equation determines how the contact angle depends
on the three independent (solid-liquid, solid-gas and liquid-gas) surface tensions. In contrast,
there are six independent surface tensions for LIS, and an equivalent relation for the contact
angle as a function of these surface tensions is, to date, not yet available. We will derive
such a relation in this paper. Furthermore, an oil ridge is drawn to the
oil-water-gas three-phase contact line in LIS system due to capillary action, as depicted in Fig. \ref{fgr:sketch_LIS}(a).
It is unclear how the shape and size of this ridge can be controlled, and correspondingly
what the consequences might be. Indeed, a distinguishing feature of LIS we will show here is that
both the apparent contact angle and contact angle hysteresis are not uniquely defined by material parameters;
instead, they also have a strong dependence on the size of the oil ridge relative to the droplet,
which in return can be manipulated by tuning the oil pressure.
This paper is organised as follows. In section 2, we will list the essential
physical assumptions for the theoretical model and describe the computational method we employ
to calculate drop morphologies on liquid infused surfaces.
We will derive a closed form expression for the contact angle in the limit
of vanishing oil ridge in section 3. In section 4, supported by numerical
calculations, we will address the influence of oil pressure on the apparent contact angle.
In section 5, we will discuss how the theoretical results in sections 3 and 4
can be extended to predict contact angle hysteresis generated by contact line pinning
at the edges of the surface corrugations.
Finally, we will conclude and describe future works in section 6.
\section{Physical model and numerical method}
\label{sec:Model}
\subsection{Physical model}
For concreteness, let us consider a typical LIS system consisting of a water droplet (w)
deposited on a porous/rough solid substrate (s) infused by an oil liquid (o), and immersed
in a surrounding gas phase (g) as shown in Fig. \ref{fgr:sketch_LIS}. Our theory
is valid, without any loss of generality, if other fluids are used instead of water, oil and gas.
Let us also define $\gamma_{\rm wg}$, $\gamma_{\rm ow}$ and $\gamma_{\rm og}$ as the surface
tensions between the water--gas, oil--water and oil--gas components respectively.
We further assume that the typical length scales in the problem (the droplet size, the oil
ridge, and the surface corrugation) are smaller than the capillary length, such that
we can neglect gravity. For water and most oils, the capillary length is of the order of a few millimetres.
The total energy $E_{\rm b}$ has several different contributions. First, this energy contains
two terms that depend on the volumes of the water droplet $V_{\rm w}$ and infusing oil $V_{\rm o}$,
and on the pressure differences $\Delta P_{\rm wg}$ between the water droplet and
the surrounding gas, and $\Delta P_{\rm og}$ between the oil and the surrounding gas.
It is also convenient to define $\Delta P_{\rm ow}$ as the
pressure difference between oil and water.
Second, each fluid-fluid interface contributes with a term proportional to
$\gamma_{\alpha\beta}\,A_{\alpha\beta}$.
The subscripts $\alpha,\beta$ correspond to the water, oil and gas phases.
Third, if any of the phases is in contact with a portion of the solid substrate, it also
contributes with a term $\gamma_{\alpha \rm s}\,A_{\alpha \rm s}$, where ${\rm s}$ indicates
the solid surface. Thus, the total energy is given by:
\begin{equation}
E_{\rm b} = \Delta P_{\rm wg} V_{\rm w} + \Delta P_{\rm og} V_{\rm o} + \sum_{\alpha\neq \beta} \gamma_{\alpha\beta} A_{\alpha\beta} + \sum_{\alpha} \gamma_{\alpha \rm s} A_{\alpha \rm s}.
\label{eq:energy}
\end{equation}
Let us now discuss the suitable ensemble for the water and oil phases.
Usually the volume of the water droplet is fixed in experiments.
As such, in our calculations the pressure difference $\Delta P_{\rm wg}$ acts as a Lagrange multiplier
to the droplet volume.
For the oil phase, instead it is appropriate to assume the pressure ensemble due to
the presence of a large amount of oil infused in between the surface
corrugations, which to a good approximation can be considered as an infinite reservoir.
In this context, the definition of $V_{\rm o}$ in Eq. \eqref{eq:energy} corresponds to the
amount of oil drawn from the reservoir into the ridge. The oil which fills the surface
roughness is not included in the computation of $V_{\rm o}$.
In equilibrium, the Laplace pressures of the fluid-fluid interfaces determine their mean curvatures
$\kappa$ through the Laplace law
\begin{equation}
\Delta P_{\rm \alpha\beta}=2\kappa_{\rm \alpha\beta}\gamma_{\rm \alpha\beta}.
\label{eq:Laplace}
\end{equation}
As before, the subscripts $\alpha,\beta$ correspond to the water, oil and gas phases.
At the triple point junction, where the three fluid interfaces meet, the stresses are balanced
\begin{equation}
\vec{\gamma}_{\rm ow}+\vec{\gamma}_{\rm og}+\vec{\gamma}_{\rm wg} = 0.
\label{eq:balance}
\end{equation}
As shown in Fig. \ref{fgr:sketch_LIS}(a), Eq. \eqref{eq:balance}
leads to the Neumann angles\cite{Rowlinson}, $\theta_{\rm o}$, $\theta_{\rm w}$ and $\theta_{\rm g}$, where
\begin{equation}
\frac{\gamma_{\rm ow}}{\sin \theta_{\rm g}} = \frac{\gamma_{\rm wg}}{\sin \theta_{\rm o}}
= \frac{\gamma_{\rm og}}{\sin \theta_{\rm w}},
\label{eq:Neumann}
\end{equation}
and $\theta_o+\theta_w+\theta_g=2\pi$.
It is worth noting that, for $\gamma_{\rm wg} > \gamma_{\rm ow} + \gamma_{\rm og}$, the water
droplet is encapsulated by a thin layer of oil, and Eq. \eqref{eq:Neumann} is ill-defined.
In this work, we will exclude such a case, and assume that the Neumann angles can be computed
according to Eq. \eqref{eq:Neumann} for a given set of water--gas, oil--gas, and water--oil surface tensions.
The interaction between the ternary fluids (water--oil--gas) with the (smooth) solid surface
can be characterised by three material contact angles, $\theta^{\rm Y}_{\rm wg}$, $\theta^{\rm Y}_{\rm ow}$
and $\theta^{\rm Y}_{\rm og}$, given by the Young's relation
\begin{equation}
\cos{\theta^{\rm Y}_{\rm \alpha\beta}} = \frac{\gamma_{\rm \beta s}-\gamma_{\rm \alpha s}}{\gamma_{\rm \alpha\beta}},
\end{equation}
where once again the subscripts $\alpha,\beta$ correspond to the water, oil and gas phases.
For liquid infused surfaces, the solid surface is not smooth. In fact, the surface roughness is
key for maintaining the infusing oil. As shown in Fig. \ref{fgr:sketch_LIS}, a typical
substrate can be modelled as a composite between solid and oil.
If pinning effects and the related energy barriers are negligible, the contact
angles can be described by weighted averages as proposed
by Cassie and Baxter \cite{Cassie1944},
\begin{equation}
\cos{\theta^{\rm CB}_{\alpha\beta}} = f \cos{\theta^{\rm Y}_{\rm \alpha\beta}} +(1-f),
\label{eq:CB}
\end{equation}
where $\alpha$ represents the oil phase, $\beta$ is either the water or gas phase,
and $f$ is the fraction of the projected solid area exposed to the water or gas phase.
We will explicitly consider pinning phenomena at the sharp edges of the surface roughness
in section \ref{sec:CAH}, where we address the emergence of contact angle hysteresis on LIS.
Formally the weighted average will enter the total energy $E_{\rm b}$ by redefining the
surface energy of the composite substrate
$\gamma_{\alpha \rm s} \rightarrow f \gamma_{\alpha \rm s} +(1-f)$,
where $\alpha$ represents the water or gas phase.
For the Cassie-Baxter equation to be valid, the water droplet needs to cover
a sufficiently large number (e.g. several tens) of posts.
As it is not the aim of this work to resolve the liquid morphology down to the molecular scale,
we will not specifically model a thin oil film on top of the roughness, because the effect of
a thin microscopic film on the equilibrium shape of a macroscopic droplet is negligible.
Its presence, however, will affect the choice of $\theta_{\rm ow}^{\rm Y}$ and
$\theta_{\rm og}^{\rm Y}$. Typically it is convenient to assume the value of
$0^\circ$ when a thin wetting oil film is present, though in general a thin oil film does not
necessarily lead to vanishing contact angles, depending on the effective interfacial potentials
describing the intermolecular forces acting between the substrate and the water and oil
molecules \cite{Dufour2016,Kuchin2014}.
\subsection{Numerical method}
In our calculations we assume the distortion induced by the underlying pattern geometry to be
negligible and the drop to retain an axial symmetry. We will numerically compute droplet
configurations by employing a finite element approach based on the free software SURFACE EVOLVER \cite{Brakke1996}.
Eq. \eqref{eq:energy} will be minimized according to standard minimisation algorithms.
Without losing generality, we will set the reference pressure of the gas phase to zero,
while the pressures of the water and oil phases will correspond to the Laplace pressures
of the water--gas and oil--gas interfaces. The oil phase will be controlled by its
pressure $\Delta P_{\rm og}$, by including the corresponding term in the total energy in
Eq. \eqref{eq:energy}, while the volume $V_{\rm w}$ of the water droplet will be imposed by a
global constraint. The Laplace Pressure $\Delta P_{\rm wg}$ is the Lagrange multiplier to
the constant droplet volume constraint.
\section{The limit of vanishing oil ridge}
\label{sec:2Dsolution}
In the most general case, the shapes of fluid-fluid interfaces with constant mean curvature
and an axial symmetry must belong to the family of Delaunay surfaces\cite{Delaunay1841}.
In our problem, the water-gas interface will be a portion of sphere, while the oil-water and oil-gas interfaces can
be described either by nodoids or unduloids, depending on the boundary conditions (wetting contact angles).
In this section we are interested in the case where the size of the oil ridge is infinitely small compared to the water
drop. For most liquid infused surfaces, this limit is equivalent to the condition of large and negative oil
pressure in comparison to the Laplace pressure in the water droplet, $-\Delta P_{\rm wg}/\Delta P_{\rm og}\rightarrow 0$,
since $\Delta P_{\rm og} < 0$ is the regime of physical interest.
In section 4, we will comment on the implications related to the case of $\Delta P_{\rm og} > 0$.
In the $-\Delta P_{\rm wg}/\Delta P_{\rm og}\rightarrow 0$ limit,
the geometry near the oil ridge can be simplified. The water-gas interface is
effectively flat. The curvature in the $x-y$ plane for the oil-water and oil-gas interfaces can be neglected.
Their profiles in the $x-z$ plane are circular arcs to an excellent approximation.
\begin{figure}[tb]
\centering
\includegraphics[width=0.85\columnwidth]{Fig2.eps}
\caption{Sketch illustrating the derivation of the closed form expression for $\theta^S_{\rm app}$. In the limit of small
oil ridge the water-gas interface is flat in proximity of the Neumann triangle, while the oil-water and oil-gas interfaces
can be approximated by circular arcs. The surface corrugation is not explicitly depicted in this sketch. However, its effect
is accounted by using the Cassie-Baxter's rather than the Young's angles for the oil-water and oi-gas interfaces.
}
\label{fgr:sketch2D}
\end{figure}
Referring to the sketch shown in Fig. \ref{fgr:sketch2D}, we introduce two auxiliary angles $\varphi$ and $\psi$
for the water-oil and oil-gas interfaces. These interfaces respectively have radii of curvature $r_{\rm ow}$ and $r_{\rm og}$.
The oil-water and oil-gas interfaces approach the substrate with contact angles $\theta^{\rm CB}_{\rm ow}$ and
$\theta^{\rm CB}_{\rm og}$. Since $-\Delta P_{\rm wg}/\Delta P_{\rm og}\rightarrow 0$, we can deduce that
\begin{equation}
-\Delta P_{\rm ow} = \frac{\gamma_{\rm ow}}{r_{\rm ow}} = -\Delta P_{\rm og} = \frac{\gamma_{\rm og}}{r_{\rm og}},
\label{eq:radiusratio0}
\end{equation}
which, combined with Eq. \eqref{eq:Neumann}, leads to
\begin{equation}
\frac{r_{\rm ow}}{r_{\rm og}} = \frac{\gamma_{\rm ow}}{\gamma_{\rm og}}=\frac{\sin\theta_{\rm g}}{\sin\theta_{\rm w}}.
\label{eq:radiusratio}
\end{equation}
In Fig. \ref{fgr:sketch2D}, we can identify a triangle with interior angles given by
$\varphi+\theta^{\rm CB}_{\rm ow}$, $\psi+\theta^{\rm CB}_{\rm og}$ and $\theta_{\rm o}$.
Imposing their sum to be $\pi$, we have a trigonometrical relation
\begin{equation}
\label{eq:phipsirelation}
\varphi+\theta^{\rm CB}_{\rm ow}+\psi+\theta^{\rm CB}_{\rm og}+\theta_{\rm o}=\pi.
\end{equation}
Similarly we can derive geometrical relations for the apparent contact angle $\theta^S_{\rm app}$:
\begin{equation}
\label{eq:relationrecedingpressure}
\theta^S_{\rm app}=\theta_{\rm w}-\theta^{\rm CB}_{\rm ow}-\varphi=\pi-\theta_{\rm g}+\theta^{\rm CB}_{\rm og}+\psi.
\end{equation}
Here we have used the superscript $S$ to denote the limiting case of vanishing oil volume.
To complete the set of equations and express $\theta^S_{\rm app}$ in terms of the remaining material parameters,
we can deduce one more equation by comparing the expressions for the height $h$ given by both menisci,
\begin{eqnarray}
h &= & r_{\rm og}\left[\cos\theta^{\rm CB}_{\rm og}-\cos(\psi+\theta^{\rm CB}_{\rm og})\right] \nonumber \\
&= & r_{\rm ow}\left[\cos\theta^{\rm CB}_{\rm ow}-\cos(\varphi+\theta^{\rm CB}_{\rm ow})\right].
\label{eq:hconditionrec}
\end{eqnarray}
Substituting Eqs. \eqref{eq:radiusratio}, \eqref{eq:phipsirelation} and \eqref{eq:relationrecedingpressure}
into Eq. \eqref{eq:hconditionrec}, we obtain
\begin{eqnarray}
\label{eq:anglepressureeq}
&\sin\theta_{\rm w} \left[\cos\theta^{\rm CB}_{\rm og}+\cos(\theta^S_{\rm app}+\theta_{\rm g})\right]= \nonumber \\
&\sin\theta_{\rm g} \left[\cos\theta^{\rm CB}_{\rm ow}-\cos(\theta_{\rm w}-\theta^S_{\rm app})\right].
\end{eqnarray}
Eq. \eqref{eq:anglepressureeq} can be inverted to express $\theta^S_{\rm app}$ only in terms of the
Neumann, oil-water and oil-gas contact angles:
\begin{equation}
\label{eq:anglepressure}
\cos\theta^S_{\rm app} =\left(\frac{\cos\theta^{\rm CB}_{\rm ow}\sin\theta_{\rm g} - \cos\theta^{\rm CB}_{\rm og} \sin\theta_{\rm w}}
{\cos\theta_{\rm g}\sin\theta_{\rm w} + \cos\theta_{\rm w} \sin\theta_{\rm g}}\right).
\end{equation}
Note that the specific value of the Laplace pressures for the oil-water and oil-gas interface
do not appear in Eq. \eqref{eq:anglepressure}.
It is convenient to express the apparent contact angle in Eq. \eqref{eq:anglepressure} in terms of the
fluid-fluid surface tensions and the oil-water and oil-gas contact angles. This is the form
in which the material properties are most commonly reported and tabulated.
To do this, we observe that the denominator in Eq. \eqref{eq:anglepressure} can be simplified as
$\sin (\theta_{\rm g} + \theta_{\rm w}) = \sin (2\pi - \theta_{\rm o}) = -\sin (\theta_{\rm o})$.
Taking further advantage of Eq. \eqref{eq:Neumann}, we obtain
\begin{equation}
\cos\theta^S_{\rm app} = - \cos\theta^{\rm CB}_{\rm ow} \frac{\gamma_{\rm ow}}{\gamma_{\rm wg}} + \cos\theta^{\rm CB}_{\rm og} \frac{\gamma_{\rm og}}{\gamma_{\rm wg}},
\label{eq:angle2}
\end{equation}
or alternatively
\begin{equation}
\cos\theta^S_{\rm app} = \cos\theta^{\rm CB}_{\rm wo} \frac{\gamma_{\rm ow}}{\gamma_{\rm wg}} + \cos\theta^{\rm CB}_{\rm og} \frac{\gamma_{\rm og}}{\gamma_{\rm wg}},
\label{eq:angle3}
\end{equation}
where we have used $\cos\theta^{\rm CB}_{\rm wo} =-\cos\theta^{\rm CB}_{\rm ow} $.
In general Eq. \eqref{eq:angle2} can be interpreted as a `weighted sum' between the oil-water and oil-gas
contact angles, where the weighting coefficients are given by ratios of the fluid-fluid surface tensions.
Several limiting cases are worth being pointed out:
(i) When the oil-gas surface tension is very small, such that $\gamma_{\rm og} \rightarrow 0$ and
$\gamma_{\rm ow} \rightarrow \gamma_{\rm wg}$, we recover
$\cos\theta^S_{\rm app} = - \cos\theta^{\rm CB}_{\rm ow} = \cos\theta^{\rm CB}_{\rm wo}$.
This is the Cassie-Baxter contact angle for a water droplet on a composite solid-oil substrate;
(ii) Similarly, in the limit of $\gamma_{\rm ow} \rightarrow 0$ and $\gamma_{\rm og} \rightarrow \gamma_{\rm wg}$,
we recover the Cassie-Baxter angle, $\cos\theta^S_{\rm app} = \cos\theta^{\rm CB}_{\rm og}$;
(iii) For $\theta^{\rm CB}_{\rm ow}\rightarrow 0$ and $\theta^{\rm CB}_{\rm og}\rightarrow 0$,
we recover a condition equivalent to the Young's equation,
$\cos\theta^S_{\rm app} = (\gamma_{\rm og} - \gamma_{\rm ow})/\gamma_{\rm wg}$, corresponding to a water
droplet spreading on a flat substrate made of oil;
(iv) The apparent contact angle approaches $\theta^S_{\rm app} \sim 90^\circ$ if the oil-gas and oil-water
interfaces have `symmetric' properties,
$\gamma_{\rm og} \sim \gamma_{\rm ow}$ and $\theta^{\rm CB}_{\rm ow} \sim \theta^{\rm CB}_{\rm og}$,
irrespective of their actual values.
\section{Role of the Laplace pressures}
\label{sec:Angles}
\begin{figure}[tb]
\centering
\includegraphics[width=0.6\columnwidth]{Fig3.eps}
\caption{Sketch of a drop with finite oil ridge displaying two possible definitions
for the apparent angle: $\theta_{\rm app}$ at the triple junction of
the fluid phases and $\theta'_{\rm app}$ at the virtual contact line where the
interpolated water-gas interface meets the solid substrate.}
\label{fgr:definition_angles}
\end{figure}
In the previous section we derived an analytical expression for the apparent contact angle $\theta_{\rm app}^{\rm S}$
in the limit of small oil ridge, assuming a vanishing Laplace pressure for the water-gas interface.
In this section we will extend our analysis to the general case, explicitly accounting for the role of finite Laplace pressures
for the three interfaces. In particular we will show that $\theta_{\rm app}$ is not uniquely determined by the material parameters,
but it is affected by the relative size of the oil ridge relative to the size of the water drop.
To illustrate the role of the Laplace pressures, we have numerically
computed ternary drop morphologies for representative systems typical of an oil with low
surface tension, choosing $\theta_{\rm o}=30^\circ$ and symmetrically
$\theta_{\rm w}=\theta_{\rm g}=165^\circ$ (see Fig. \ref{fgr:pressure}).
We consider two different combinations of wetting angles for the oil phase, given by
(i) $\theta^{\rm CB}_{\rm ow}=0^\circ$, $\theta^{\rm CB}_{\rm og}=15^\circ$ and
(ii) $\theta^{\rm CB}_{\rm ow}=30^\circ$, $\theta^{\rm CB}_{\rm og}=0^\circ$.
We then vary the oil pressure while keeping the water volume constant.
In our calculations the specific value of the volume is not relevant, as it
simply sets the length scale of the system.
In the presence of finite Laplace pressures it is necessary to adapt the definition
of $\theta_{\rm app}$ as the water-gas interface is no longer represented by a straight line.
Two meaningful geometric choices are possible as shown in Fig. \ref{fgr:definition_angles},
either (i) as the slope of the water-gas interface at the triple junction $\theta_{\rm app}$, or (ii)
as the slope of the virtual water-gas interface as it is extrapolated down to the
solid substrate $\theta'_{\rm app}$. The water-gas interface assumes a spherical
cap geometry and the extrapolation procedure is unique.
We find $\theta'_{\rm app}>\theta_{\rm app}$.
In the limit of vanishing oil ridge, both definitions converge to the same value
corresponding to Eq. \ref{eq:angle3}, which describes the energy balance at the contact line.
The deviation grows larger with increasing size of the ridge.
In this work we favour the apparent contact angle definition at the triple junction,
$\theta_{\rm app}$, for two main reasons:
(a) the angle at the Neumann triangle can be easily identified from the kink in the drop profile,
and it can be directly measured both in simulations and experiments, and
(b) it represents a direct measure of the rigid rotation of the Neumann triangle with respect to
the solid surface. In contrast, $\theta'_{\rm app}$ describes the slope of a portion of an interface
that is only virtual, and cannot be measured directly.
The sequence of morphologies reported in Fig. \ref{fgr:pressure}(b-d) shows
the impact of increasing $-\Delta P_{\rm wg}/\Delta P_{\rm og}$ to the
growth of the oil ridge. As depicted in Fig. \ref{fgr:pressure}(a), for the chosen
sets of parameters, this is accompanied with a decrease in $\theta_{\rm app}$,
as consequence of the rigid rotation of the Neumann triangle (see the insets).
For the two specific examples we have shown here, the variation in the apparent contact angle
between $-\Delta P_{\rm wg}/\Delta P_{\rm og}\rightarrow 0$ (small ridge) and
$-\Delta P_{\rm wg}/\Delta P_{\rm og}\rightarrow \infty$ (large ridge) is above $30^\circ$.
Similarly $\theta'_{\rm app}$ also decreases with increasing $-\Delta P_{\rm wg}/\Delta P_{\rm og}$
but its variation is considerably smaller, limited to a few degrees.
It is worth noting that our definition for the apparent contact angle is intended to characterise not just the water droplet shape,
but instead the combined water droplet-oil ridge configuration, which spreads out as $-\Delta P_{\rm wg}/\Delta P_{\rm og}$ increases. In this context, $\theta_{\rm app}$ captures this behaviour better than $\theta'_{\rm app}$.
From here on, we will only focus on $\theta_{\rm app}$.
\begin{figure}[tb]
\centering
\includegraphics[width=0.8\columnwidth]{Fig4.eps}
\caption{Numerical results for a water droplet placed on liquid infused surfaces.
The Neumann angles for the oil, water and gas phases are respectively
$\theta_{\rm o}=30^\circ$ and $\theta_{\rm w}=\theta_{\rm g}=165^\circ$.
Two sets of wetting angles are considered:
(i) $\theta^{\rm CB}_{\rm ow}=0^\circ$ and $\theta^{\rm CB}_{\rm og}=15^\circ$, and
(ii) $\theta^{\rm CB}_{\rm ow}=30^\circ$ and $\theta^{\rm CB}_{\rm og}=0^\circ$.
a) Variation of the apparent contact angle as a function of $-\Delta P_{\rm wg}/\Delta P_{\rm og}$.
The results for both definitions of apparent contact angles are shown,
$\theta_{\rm app}$ (points) and $\theta'_{\rm app}$ (straight lines).
The insets illustrate how the Neumann triangles rotate as the oil pressure is varied.
b-d) Snapshots of numerically evaluated ternary drop configurations.
}
\label{fgr:pressure}
\end{figure}
One aspect differentiating the two combinations of angles reported in Fig. \ref{fgr:pressure} is worth
discussing further. In both cases the
limit $-\Delta P_{\rm wg}/\Delta P_{\rm og}\rightarrow \infty$ (i.e. $\Delta P_{\rm og}\rightarrow 0$)
implies that the oil-gas ridge approaches
a catenoid shape. However, while for finite contact angle $\theta^{\rm CB}_{\rm og}$
the oil ridge has a finite size, in the case of $\theta^{\rm CB}_{\rm og}=0^\circ$ the
radius of the oil ridge is diverging. The latter corresponds to a liquid lens configuration,
where the Neumann triangle is oriented such that the oil-gas interface is flat and lies parallel to the
solid substrate.
For small but finite $-\Delta P_{\rm wg}/\Delta P_{\rm og}$, we are also able to derive a closed form
expression for the apparent contact angle $\theta_{\rm app}$.
To proceed, we note that the profiles of the oil-gas and
oil-water interfaces can be still be assumed to be circular arcs in the $x-z$ plane, as shown in
Fig. \ref{fgr:sketch2D}. Following the convention displayed in Fig. 2, we have $r_{\rm og} > 0$ and
$r_{\rm ow} > 0$ when $\Delta P_{\rm og} < 0$ and $\Delta P_{\rm ow} < 0$.
The curvature in the $x-y$ plane can also be neglected for the oil-gas and oil-water interfaces.
Taking into account the Laplace pressure difference (or the curvature) of the water-gas interface,
we can write
\begin{equation}
\Delta P_{\rm wg}=\frac{\gamma_{\rm ow}}{r_{\rm ow}}-\frac{\gamma_{\rm og}}{r_{\rm og}}.
\label{eq:Laplace_balance}
\end{equation}
A straightforward manipulation invoking Eq. \eqref{eq:Laplace} leads to the following equation,
\begin{equation}
\frac{r_{\rm og}}{r_{\rm ow}}=\frac{\gamma_{\rm og}}{\gamma_{\rm ow}} \left(1- \frac{\Delta P_{\rm wg}}{\Delta P_{\rm og}} \right).
\label{eq:pressure_parameter}
\end{equation}
\vspace{0.0cm}
Following the same route leading to Eq. \eqref{eq:anglepressureeq} in the previous section,
we can write down an equivalent relation with a correction term due to finite $-\Delta P_{\rm wg}/\Delta P_{\rm og}$,
\begin{equation}
\frac{\sin\theta_{\rm g} \left[\cos\theta^{\rm CB}_{\rm ow}-\cos(\theta_{\rm w}-\theta_{\rm app})\right]}
{\sin\theta_{\rm w} \left[\cos\theta^{\rm CB}_{\rm og}+\cos(\theta_{\rm app}+\theta_{\rm g})\right]}
=\left(1- \frac{\Delta P_{\rm wg}}{\Delta P_{\rm og}} \right).
\label{eq:anglepressureeq2}
\end{equation}
This relation can be inverted for the apparent contact angle, and it is given by
\begin{equation}
\cos\theta_{\rm app}=\frac{A C + B\sqrt{A^2 + B^2 - C^2}}{A^2 + B^2},
\label{eq:anglepresssol0}
\end{equation}
where
\begin{eqnarray}
&A=\sin\theta_{\rm g} \cos\theta_{\rm w} + \left(1- \frac{\Delta P_{\rm wg}}{\Delta P_{\rm og}} \right) \sin\theta_{\rm w} \cos\theta_{\rm g}, \label{eq:anglepresssol1} \\
&B=\sin\theta_{\rm g} \sin\theta_{\rm w} \left(\frac{\Delta P_{\rm wg}}{\Delta P_{\rm og}} \right), \label{eq:anglepresssol2} \\
&C=\sin\theta_{\rm g} \cos\theta^{\rm CB}_{\rm ow} - \left(1- \frac{\Delta P_{\rm wg}}{\Delta P_{\rm og}} \right) \sin\theta_{\rm w} \cos\theta^{\rm CB}_{\rm og} . \label{eq:anglepresssol3}
\end{eqnarray}
As we can observe in Fig. \ref{fgr:compare_evolver_model}, the analytical expression compares well with the full numerical results
for $ -\Delta P_{\rm wg}/\Delta P_{\rm og}<1$. For larger $ -\Delta P_{\rm wg}/\Delta P_{\rm og}$, the model departs from the numerical solution,
as the circular arc approximation for the oil-water and oil-gas interfaces breaks down.
Furthermore, it is useful to extract a linear correction to the apparent contact angle due to the parameter $-\Delta P_{\rm wg}/\Delta P_{\rm og}$,
given by
\begin{widetext}
\begin{equation}
\cos\theta_{\rm app}=\cos\theta_{\rm app}^{\rm S} + \Lambda \left( \frac{\Delta P_{\rm wg}} {\Delta P_{\rm og}} \right)
+ \mathcal{O}\left( \frac{\Delta P_{\rm wg}} {\Delta P_{\rm og}} \right)^2,
\label{eq:anglepressexpansion}
\end{equation}
with
\begin{equation}
\Lambda= \frac{\sin\theta_{\rm g}\sin\theta_{\rm w} \left(\cos\theta_{\rm g} \cos\theta^{\rm CB}_{\rm ow} + \cos\theta^{\rm CB}_{\rm og} \cos\theta_{\rm w} +
\sqrt{\sin(\theta_{\rm g} + \theta_{\rm w})^2 -(\cos\theta^{\rm CB}_{\rm ow} \sin\theta_{\rm g} - \cos\theta^{\rm CB}_{\rm og} \sin\theta_{\rm w})^2 }\right)}
{\sin(\theta_{\rm g}+\theta_{\rm w})^2}.
\label{eq:chipar}
\end{equation}
\end{widetext}
In experiments, the pressure difference between the oil and gas phases is usually kept constant. The Laplace pressure
of the water droplet is given by $\Delta P_{\rm wg} = 2 \gamma_{\rm wg}/r_{\rm wg} \simeq 2\gamma_{\rm wg}\sin\theta^S_{\rm app}/R$,
where $R/r_{\rm wg} \simeq \sin\theta^S_{\rm app}$ is the effective contact radius, taken at the Neumann's triple junction between the
water, oil and gas phases. Since we are only interested in the first order correction here, we have also approximated $\theta_{\rm app}\simeq\theta^S_{\rm app}$.
As such, Eq. \eqref{eq:anglepressexpansion} can be written as
\begin{equation}
\cos\theta_{\rm app}=\cos\theta_{\rm app}^{\rm S} + \frac{2\Lambda\gamma_{\rm wg}\sin\theta^S_{\rm app}}{\Delta P_{\rm og}} \times \frac{1}{R},
\label{eq:linetension}
\end{equation}
This equation has a familiar interpretation in the literature of wetting phenomena:
it is reminiscent to the correction term in Young's angle due to line tension \cite{Gaydos1987,Marmur1997},
with an effective line tension given by
\begin{equation}
\tau = -\frac{2\Lambda\gamma^2_{\rm wg}\sin\theta^{\rm S}_{\rm app}}{\Delta P_{\rm og}}.
\label{eq:linetension2}
\end{equation}
The linear approximation in Eq. \eqref{eq:linetension} is also shown in Fig. \ref{fgr:compare_evolver_model}. It is in good agreement with the
full numerical results for $ -\Delta P_{\rm wg}/\Delta P_{\rm og}<0.5$.
\begin{figure}[tb]
\centering
\includegraphics[width=0.8\columnwidth]{Fig5.eps}
\caption{Comparison between numerical results (dots), Eq. \eqref{eq:anglepresssol0}
(full lines), and Eq. \eqref{eq:linetension} (dashed lines). The analytical expressions
are valid for small $-\Delta P_{\rm wg}/\Delta P_{\rm og}$.
The system parameters are the same as those reported in Fig. \ref{fgr:pressure}.
}
\label{fgr:compare_evolver_model}
\end{figure}
The coefficient $\Lambda$ in Eq. \eqref{eq:chipar} can in principle assume both positive and negative values.
It can be shown that $\Lambda<0$ if $\theta_{\rm w}+\theta_{\rm g}>\pi +\theta_{\rm ow} +\theta_{\rm og}$.
Using the fact that $\theta_{\rm w}+\theta_{\rm g} +\theta_{\rm o}=2\pi$,
we obtain $\pi > \theta_{\rm o} +\theta_{\rm ow} +\theta_{\rm og}$, which has a simple geometrical interpretation.
From Fig. \ref{fgr:press_pos} it is clear that when $\theta_{\rm o} +\theta_{\rm ow} +\theta_{\rm og}=\pi$
the three angles of the oil ridge form a triangle.
When $\Lambda<0$ the sum of these angles is smaller than $\pi$, and the oil ridge is stable only if
$\Delta P_{\rm og}<0$ and $\Delta P_{\rm ow}<0$.
In contrast, if $\Lambda>0$, the sum of the three angles is larger than $\pi$, and the ridge is stable only if
$\Delta P_{\rm og}>0$ and $\Delta P_{\rm ow}>0$.
$\Lambda<0$ and $\Delta P_{\rm og}<0$ represent the most relevant physical regime for liquid infused surfaces.
The case of $\Lambda>0$ implies larger oil-water and oil-gas wetting angles, which are often in conflict with the
wicking criterion in Eq. \eqref{eq:wicking}. Additionally, for $\Delta P_{\rm og}>0$, the fluid configuration
could be unstable against non-axisymmetric perturbations\cite{Schafle2010}.
Taking advantage of Eqs. \eqref{eq:angle2} and \eqref{eq:linetension2}, we have computed the apparent
contact angles and effective line tensions for several LIS systems reported in the literature\cite{Wong2011,Anand2015,Smith2013}
in Table \ref{tbl:example}. In Fig. \ref{fgr:exp_angle}, we have also shown one experimental drop morphology from Smith et al. \cite{Smith2013}. Here the oil ridge is small in comparison to the droplet size, and as such, we expect Eq. \ref{eq:angle2} to provide an excellent approximation. Indeed the measured contact angle is in agreement with the theoretical prediction, assuming $f=0.44$.
To compute the effective line tension in Table \ref{tbl:example}, we have assumed a typical Laplace pressure
$|\Delta P_{\rm og}|=10^3$ Pa for the oil--gas ridge, corresponding to a radius of curvature of
$r_{\rm og} \sim 100 \, \rm{\mu m}$. The computed effective line tension values are comparable to
those measured for gas-liquid-solid contact line tensions\cite{Amirfazli2004}. Noticeably,
liquid infused surfaces always have negative effective line tensions since the signs of $\Lambda$ and
$\Delta P_{\rm og}$ are always the same for the system to be stable.
Thus, the apparent contact angle of a water droplet on a liquid infused surface increases with increasing droplet volume.
Our analytical expressions are readily applicable to other solid surfaces and fluids
(both for the droplet and the lubricant) for cases where $\gamma_{\rm wg} < \gamma_{\rm ow} + \gamma_{\rm og}$.
\begin{figure}[tb]
\centering
\includegraphics[width=0.8\columnwidth]{Fig6.eps}
\caption{Sketch illustrating the stable shapes of the oil ridge depending on the sign of $\Lambda$:
a) the common case, with $\Lambda<0$ and $\Delta P_{\rm og}<0$; and
b) with $\Lambda>0$ and $\Delta P_{\rm og}>0$
}
\label{fgr:press_pos}
\end{figure}
\begin{table*}
\small
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}ccccccccccccccc}
\hline
Source & Solid & droplet (w) & lubricant (o) & $\gamma_{\rm wg}$ & $\gamma_{\rm og}$ & $\gamma_{\rm wo}$
& $\theta_{\rm w}$ & $\theta_{\rm o}$ & $\theta_{\rm g}$ & $\theta_{\rm ow}$ & $\theta_{\rm og} $
& $\theta_{\rm app}^{\rm S}$ & $\Lambda$ & $\tau$ \\
\hline
Ref. \cite{Schellenberger2015} & inv. opal & H$_2$O& decanol & $30$ & $28.5$ & $8.6$ & $108.3$ & $88.4$ & $163.3$ & $0$ & $0$ & $48.4$ & $-0.14$ & $-1.9\times 10^{-7}$ \\
Ref. \cite{Smith2013} & OTS & H$_2$O& BMIm & $42$ & $34$ & $13$ & $135.4$ & $60.2$ & $164.4$ & $37$ & $64$ & $70.9$ & $-0.15$ & $-4.9\times 10^{-7}$ \\
Ref. \cite{Wong2011} & S.Epoxy & H$_2$O & FC-70 & $72.4$ & $17.1$ & $56.0$ & $175.6$ & $18.8$ & $165.6$ & $36.5$ & $14.1$ & $118.2$ & $-0.29$ & $-2.7\times 10^{-6}$ \\
Ref. \cite{Wong2011} & Epoxy & H$_2$O & FC-70 & $72.4$ & $17.1$ & $56.0$ & $175.6$ & $18.8$ & $165.6$ & $71.7$ & $33.5$ & $108.7$ & $-0.24$ & $-2.3\times 10^{-6}$ \\
Ref. \cite{Wong2011} & Silicon & C16 H$_34$ & H$_2$O & $27.2$ & $72.4$ & $51.1$ & $47.2$ & $164.0$ & $48.8$ & $5.6$ & $13.1$ & $28.8$ & $0.028$ & $-2.0\times 10^{-8}$ \\
\hline
\end{tabular*}
\caption{Theoretical prediction for the apparent contact angle $\theta_{\rm app}^{\rm S}$ for
several LIS systems reported in the literature\cite{Wong2011,Anand2015,Smith2013},
as given by Eq. \eqref{eq:angle2}. For the Cassie-Baxter contact angles, we have assumed rough surfaces with projected solid area fraction $f=0.44$.
The surface tensions are expressed in the unit of mN/m, the line tension $\tau$ is in Newton (N), and the angles are in degrees ($^\circ$).
$\Lambda$ is the dimensionless parameter needed for computing the line tension as defined in Eq. \eqref{eq:chipar}.
For the computation of the line tension in Eq. \eqref{eq:linetension2}, we have assumed
a typical Laplace pressure $\Delta P_{\rm og}=10^3$ Pa for the oil--gas ridge, corresponding to a radius of curvature, $r_{\rm og} \sim 100 \, \rm{\mu m}$.
}
\label{tbl:example}
\end{table*}
\section{Contact Angle Hysteresis}
\label{sec:CAH}
\begin{figure}[tb]
\centering
\includegraphics[width=0.95\columnwidth]{Fig7.eps}
\caption{a) Experimental image of a water droplet on an OTS surface infused by BMIm as the lubricant, taken from Smith et al. \cite{Smith2013}.
The observed apparent contact angle $\theta_{\rm app}\simeq 70^\circ \pm 2^\circ$ is in excellent agreement with the prediction
of Eq. \ref{eq:angle2} $\theta_{\rm app}^{\rm S}= 70.9^\circ$. b) The surface pattern used in the experiment, again taken from Smith et al. \cite{Smith2013}.
From panel b), we estimate that the projected solid fraction exposed to the water and gas phases is $f=0.44$.
}
\label{fgr:exp_angle}
\end{figure}
In the previous sections we computed the apparent contact angles
in thermodynamic equilibrium, by introducing the Cassie-Baxter contact angles on the composite substrate.
In this section we will address how pinning of the oil-water and oil-gas contact lines give rise to contact angle
hysteresis on liquid infused surfaces. In general contact line pinning can be generated either by chemical
heterogeneities or surface topographies\cite{Kusumaatmaja2007,Semprebon2012,Lenz1998,Oliver1977,Johnson1964}.
Here we will focus on the latter. Following the Gibbs condition \cite{Gibbs}, a pinned contact line does not exhibit a unique contact angle,
instead it can take a range of values.
\begin{figure}[tb]
\centering
\includegraphics[width=0.99\columnwidth]{Fig8.eps}
\caption{Contact angle hysteresis on liquid infused surfaces arises due to the pinning of oil-water
and oil-gas contact lines by the surface corrugations.
Sketches illustrating the configurations of the oil ridge for (a) an advancing and (b) a receding water droplet.
}
\label{fgr:sketch_hysteresis}
\end{figure}
There are four wetting states on liquid infused surfaces:
(i) When $\theta^{\rm CB}_{\rm ow} = \theta^{\rm CB}_{\rm og} = 0^\circ$, we expect the contact angle hysteresis to be negligible;
(ii) In contrast, for $\theta^{\rm CB}_{\rm ow} > 0^\circ$ and $\theta^{\rm CB}_{\rm og} > 0^\circ$,
the oil-water and oil-gas contact lines can both be pinned. As illustrated in Fig. \ref{fgr:sketch_hysteresis},
for a droplet to advance on a liquid infused surface, the oil-water
contact line has to recede and the oil-gas contact line has to advance. Similarly, a receding droplet requires the oil-water contact
line to advance and the oil-gas contact line to recede; (iii) For $\theta^{\rm CB}_{\rm ow} > 0^\circ$ and $\theta^{\rm CB}_{\rm og} = 0^\circ$,
contact line pinning only occurs at the oil-water contact line, while (iv) for $\theta^{\rm CB}_{\rm ow} = 0^\circ$ and $\theta^{\rm CB}_{\rm og} > 0^\circ$,
pinning only takes place for the oil-gas contact line.
In our model the contact angles are defined with respect to the oil phase, and are required
to be small to guarantee the validity of the hemi-wicking criterion, Eq. \eqref{eq:wicking}. The complementary angles,
defined with respect to the water and gas phases are therefore large, and the analogy to superhydrohobic materials
is appropriate.
A large body of work on contact angle hysteresis on superhydrophobic materials leads to the
surprisingly simple result that the liquid (e.g. water) advancing contact angle occurs for
$\theta^{\rm A}_{\rm wg}=180^\circ$ \cite{Kusumaatmaja2007,Schellenberger2016},
where deviations reported in literature are most likely due to experimental difficulties
of measuring very large angles. The estimate for the receding angle is more
debated, and several models have been proposed in the literature. These include
(i) the sparse defect model proposed by Joanny and DeGennes \cite{Joanny1984}, and experimentally tested by
Reyssat and Quere \cite{Reyssat2009}, which suggests the receding contact angle
has a logarithmic dependence with respect to the pillar spacing;
(ii) thermodynamic approaches based on a linear average of the contact angle along the contact line\cite{Raj2012}, and
(iii) the Cassie-Baxter model based on the area average\cite{McHale2004a}.
It is not in the scope of this work to assess the accuracy of such models in general, but we remark that
the thermodynamic Cassie-Baxter approach is more aligned with the approximations assumed here.
The sparse defect model is not consistent with the requirement of dense patterns,
while the linear averaging model implies a strong effect of the orientation of the contact line on the global
shape of a drop, which appears negligible in the currently available experimental data \cite{Guan2015,Schellenberger2016}.
\begin{figure}[tb]
\centering
\includegraphics[width=0.7\columnwidth]{Fig9.eps}
\caption{Contact angle hysteresis ($\Delta \theta$) of a water droplet on a liquid infused surface with
$\theta_{\rm o}=30^\circ$, $\theta_{\rm w}=\theta_{\rm g}=165^\circ$,
$\theta^{\rm CB}_{\rm ow}=30^\circ$ and $\theta^{\rm CB}_{\rm og}=15^\circ$.
For this set of parameters, the two sets of data shown in Fig. \ref{fgr:pressure}(a) in fact correspond to the
advancing ($\theta^{\rm R}_{\rm ow}=0^\circ$, $\theta^{\rm A}_{\rm og}=15^\circ$)
and receding ($\theta^{\rm A}_{\rm ow}=30^\circ$, $\theta^{\rm R}_{\rm og}=0^\circ$) contact angles.
We only focus on the definition of the apparent contact angle as defined at the triple junction, $\theta_{\rm app}$.
}
\label{fgr:hysteresis}
\end{figure}
\begin{figure*}[tb]
\centering
\includegraphics[width=1.5\columnwidth]{Fig10.eps}
\caption{A typical contact angle hysteresis loop for a water droplet on a liquid infused surface.
Here $\theta_{\rm o}=30^\circ$, $\theta_{\rm w}=\theta_{\rm g}=165^\circ$,
$\theta^{\rm CB}_{\rm ow}=30^\circ$ and $\theta^{\rm CB}_{\rm og}=15^\circ$.
Panel (a) shows the apparent contact angle of the droplet as a function of its volume,
while panel (b) shows the radii of the oil-water ($R_{\rm ow}$) and oil-gas ($R_{\rm og}$) contact lines.
The oil pressure, $\Delta P_{\rm og} = -2\gamma_{\rm og}/r_{\rm og}$, is kept constant in these calculations,
where $r_{\rm og}$ is the radius of curvature of the oil-gas interface. We use $r_{\rm og}$
to normalise the droplet volume and the contact line radii. (c-h) Drop morphologies as indicated in panels (a) and (b).
}
\label{fgr:hysteresis_loop}
\end{figure*}
Contact angle hysteresis is usually evaluated employing two alternative experimental approaches.
The first one relies on applying a body force to the droplet\cite{Semprebon2014}, and the
advancing and receding angles are measured at the front and back of the droplet
just before it starts to move. To aid the discussion, let us now consider a specific example
where the oil-water and oil-gas (Cassie-Baxter) contact angles are respectively $\theta^{\rm CB}_{\rm ow}=30^\circ$
and $\theta^{\rm CB}_{\rm og}=15^\circ$. The Neumann angles are chosen to be
$\theta_{\rm o}=30^\circ$ and $\theta_{\rm w}=\theta_{\rm g}=165^\circ$.
Based on our discussion in the previous paragraph, the conditions for an advancing contact line are
$\theta^{\rm R}_{\rm ow}=0^\circ$ and $\theta^{\rm A}_{\rm og}=\theta^{\rm CB}_{\rm og}=15^\circ$, while for a receding
contact line we have $\theta^{\rm A}_{\rm ow}=\theta^{\rm CB}_{\rm ow}=30^\circ$ and $\theta^{\rm R}_{\rm og}=0^\circ$.
As such, the curves in Fig. \ref{fgr:pressure} represent the advancing and receding apparent contact angles for
a water drop on a liquid infused substrate with the aforementioned Cassie-Baxter contact angles, parametrized by the pressure
ratio $-\Delta P_{\rm wg}/ \Delta P_{\rm og}$. The contact angle hysteresis is defined as the
difference between the advancing and receding contact angles,
$\Delta \theta_{\rm app} = \theta^{\rm A}_{\rm app} - \theta^{\rm R}_{\rm app}$, and it is shown in Fig. \ref{fgr:hysteresis}
as a function of $-\Delta P_{\rm wg}/ \Delta P_{\rm og}$. The contact angle hysteresis shows a strong dependence on the pressure ratio
(or equivalently the size of the oil ridge relative to the water droplet). It increases logarithmically in the
limit of $ -\Delta P_{\rm wg}/ \Delta P_{\rm og} \rightarrow \infty$, and it approaches a constant value
as $-\Delta P_{\rm wg}/ \Delta P_{\rm og} \rightarrow 0$. Interestingly, the curve is also non
monotonic, and exhibits a shallow minimum close to $-\Delta P_{\rm wg}/ \Delta P_{\rm og} =0.2$.
This is in contrast with binary systems (e.g. water-gas on a solid surface), where the advancing and receding
angles (correspondingly, contact angle hysteresis) can be regarded as constant material parameters.
The analytical expressions in Eqs. \eqref{eq:angle2}, \eqref{eq:anglepresssol0} and \eqref{eq:linetension}
can be modified to predict advancing and receding contact angles in the limit of small oil ridge,
with the following replacement:
$\theta^{\rm R}_{\rm ow}=0^\circ$,
$\theta^{\rm A}_{\rm ow}=\theta^{\rm CB}_{\rm ow}$,
$\theta^{\rm A}_{\rm og}=\theta^{\rm CB}_{\rm og}$,
$\theta^{\rm R}_{\rm og}=0^\circ$.
In the limit of $-\Delta P_{\rm wg}/ \Delta P_{\rm og} \rightarrow 0$, we obtain
\begin{equation}
\label{eq:angleAdv}
\cos\theta^{S,A}_{\rm app} = - \frac{\gamma_{\rm ow}}{\gamma_{\rm wg}} + \cos\theta^{\rm CB}_{\rm og} \frac{\gamma_{\rm og}}{\gamma_{\rm wg}},
\end{equation}
and
\begin{equation}
\label{eq:angleRec}
\cos\theta^{S,R}_{\rm app} = - \cos\theta^{\rm CB}_{\rm ow} \frac{\gamma_{\rm ow}}{\gamma_{\rm wg}} + \frac{\gamma_{\rm og}}{\gamma_{\rm wg}}.
\end{equation}
Furthermore, the resisting force due to contact angle hysteresis is given by
\begin{equation}
\label{eq:force}
F = 2 R \gamma_{\rm wg} \Delta \cos\theta,
\end{equation}
where $R$ is the contact radius and $\Delta \cos\theta = \cos\theta^{\rm R} - \cos\theta^{\rm A}$.
We can straightforwardly obtain the expression for
$\Delta \cos\theta^{S}_{\rm app}=\cos\theta^{S,A}_{\rm app}-\cos\theta^{S,R}_{\rm app}$
by combining Eqs. \eqref{eq:angleAdv} and \eqref{eq:angleRec}. Here we assume the action of the
body force does not significantly deform the droplet.
The resulting closed form expression once again
can be interpreted as a 'weighted sum' of the contact angle hysteresis for the oil--water and oil--gas contact lines.
Similar expressions for $\Delta \cos\theta_{\rm app}$ can also be obtained for small but finite $-\Delta P_{\rm wg}/ \Delta P_{\rm og}$
by exploiting Eqs. \eqref{eq:anglepresssol0} or \eqref{eq:linetension}.
The second approach to measure contact angle hysteresis is by varying the volume of the water droplet \cite{Ruiz-Cabello2011}.
It is important to keep in mind that this protocol, unlike the previous one, involves measurements at
different pressure ratio $-\Delta P_{\rm wg}/ \Delta P_{\rm og}$. To elucidate the relevance of $-\Delta P_{\rm wg}/ \Delta P_{\rm og}$,
we report a typical hysteresis loop in Fig. \ref{fgr:hysteresis_loop}. As before, we consider
$\theta^{\rm CB}_{\rm ow}=30^\circ$, $\theta^{\rm CB}_{\rm og}=15^\circ$,
$\theta_{\rm o}=30^\circ$ and $\theta_{\rm w}=\theta_{\rm g}=165^\circ$
such that the data shown in Fig.\ref{fgr:pressure}(a) are the
advancing and receding apparent contact angles as function of the pressure ratio
for this set of parameters.
Let us begin with the drop configuration shown in panel (c) of Fig. \ref{fgr:hysteresis_loop}.
When the drop volume is increased, the apparent contact angle also increases. Here both the oil-water
and oil-gas contact lines are pinned. At (d), the oil-water contact angle locally reaches
$0^\circ$, and as a result, its contact line depins. With increasing the droplet
volume, the oil-water contact line slides while the oil-gas one remains pinned.
From configuration (e), both contact lines reach their corresponding
advancing and receding contact angles, and become free to move.
Correspondingly we observe a clear change of slope in
the volume-contact angle relation.
Once reaching (f), we reverse the process and decrease the droplet volume.
Similar to the advancing scenario, initially both contact lines are pinned. As before, the depinning of the
oil--water and oil--gas contact lines are not simultaneous,
and occur at (g) for the oil--water contact line and at (h) for the
oil--gas contact line. Both contact lines move freely from (h) to (c), which
is our starting configuration.
\section{Discussion}
\label{sec:Conclusions}
In this work we have theoretically investigated the apparent contact angle and contact angle hysteresis
of a droplet on liquid infused surfaces (LIS). We derived a closed form expression for the apparent contact angle
in the limit of vanishing oil ridge that captures the energy balance of the three fluid phases in contact with the solid substrate.
Moreover, we computed the first order correction to the contact angle accounting for the influence of a small but finite oil
wetting ridge surrounding the droplet, and showed that the correction term can be interpreted as a negative line tension.
We also employed numerical calculations to explore the full range of negative oil-gas Laplace pressures,
showing that the apparent contact angles indeed vary as a function of pressure.
Unlike usual wetting scenarios involving two fluids (e.g. water--gas), the apparent contact angle for LIS cannot be regarded
as a constant material property. We further note that our analytical expressions are in excellent agreement with the numerical results.
By introducing appropriate models for pinning and depinning
of the oil--water and oil--gas contact lines, we showed how the analytical expression for the apparent contact angles
can be readily manipulated to predict contact angle hysteresis on liquid infused surfaces.
We presented a typical contact angle hysteresis loop, where we demonstrated that
the depinning of the oil--water and oil--gas contact lines are in general not simultaneous.
Contact angle hysteresis on LIS also depends on the oil pressure, or alternatively
the relative size of the oil wetting ridge to the water droplet.
Numerical calculations indicate that the contact angle hysteresis is smaller
for large and negative oil pressure (small ridge), compared to small and negative oil pressure
(large ridge). This finding provides a useful design principle for LIS,
suggesting that the contact angle hysteresis can be tuned by the oil pressure, which
can be achieved for example by under-filling/over-filling the substrate with oil.
Our results so far are limited to equilibrium morphologies. A full characterisation of
wetting dynamics on liquid infused surfaces is an important and open problem.
To this end, we recently developed a ternary free energy Lattice Boltzmann approach\cite{Semprebon2016},
well suited for handling the fluid dynamics of the water droplet and infusing oil, and
for taking into account the Neumann angles and wetting contact angles involved in the problem.
Another important direction for future work is to investigate the possible presence of thin oil film
coating the surface corrugations and/or the water droplet \cite{Smith2013}, including the molecular mechanism
that determines the film thickness and its influence to the shape of the water droplet when
their length scales are comparable. When the infusing oil cloaks the droplet, the water-gas surface tension
is not the appropriate variable to use in Eqs. \eqref{eq:balance}, \eqref{eq:angle2} and \eqref{eq:linetension}.
Instead, it should be replaced by a composite water-oil and oil water interfaces,
$\gamma_{\rm wg} \rightarrow \gamma_{\rm ow}+\gamma_{\rm og} - \Delta(e)$, where the binding potential
$\Delta(e)$ is a function of the oil film thickness, and its form depends on the intermolecular interactions
of the fluids. We will explore this case in more details in future works.
\section{Acknowledgements}
\label{sec:acknowledgements}
C.S. and H.K. are grateful to Dr. M. Brinkmann for inspiring comments on this manuscript,
and thank Dr. M. Wagner, Dr. Y. Gizaw, Dr. P. Koenig, Prof. G. Mistura and M. S. Sadullah
for fruitful discussions. We acknowledge Procter and Gamble (P\&G) for funding.
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
The large-scale structure of the Universe has been extensively studied since the publication of the first galaxy surveys. Some analysis with high-redshift samples shows the existence of fractal correlations \citep{celerier2005fractal}. The large network of matter filaments that connect galaxies is characterised by fractal structures \citep{martinez2010statistics}. In a cosmological context, fractals were introduced into a model by Mandelbrot as a hypothesis to solve the Olbers paradox \citep{martinez1990universe}. According to the fractal hypothesis, if the set of stars forms a structure with self-similarity properties, such as a Cantor dust, the paradox is solved, for even in a mathematically infinite universe, the sky can have dark regions. We currently know that at small scales, the galaxy distribution exhibits fractal behaviour \citep{peebles1980large,mandelbrot1982the}. Some observational estimates suggest that the Universe behaves as a multifractal object with a Hausdorff dimension $D_H=2.1\pm0.1$ and correlation dimension $D_2=1.3\pm0.1$ \citep{durrer1998fractal}.\\
The introduction of the fractal concept allows interpretation of the hierarchical clustering in terms of self-similarity properties or an invariant scale of the galaxy distribution \citep{chaconstudy}. It seems clear that for $r<10^{-1}~Mpc$, galaxy clustering exhibits fractal behaviour \citep{peebles1993principles}; at larger scales, the structure is more complex and can be characterised as a multifractal object, requiring a generalisation of the fractal dimension in order to explain this distribution on the same scale range \citep{conde2015fractal}.\\
Fractal analysis is a useful mathematical tool that quantifies galactic clustering using data from galaxy surveys by calculating quantities such as the fractal dimension, making it possible to establish relationships between these values and other statistical descriptors. The possible cosmological implications of fractal analysis of the galaxy distribution are discussed in detail in \citep{baryshev1998fractal, martinez1991fractal}. Because cosmic clustering developed under the influence of gravity alone, there is no physical motivation to consider smaller scales related to the processes of galaxy formation. In this context, the multifractal measures are described as a natural scaling result of the matter interactions in the Universe. A formal and general proof of the fractal nature of the matter distribution on large scales requires additional observations and some solutions of the Einstein field equations that take into account the conditional cosmological principle \citep{martinez1990universe}. An additional motivation to study the matter distribution on large scales from a multifractal viewpoint is the transition to a homogeneity scale, which can be defined according to \citep{yadav2010fractal} as the value of $r$ above which the fractal dimension $D_q$ of the distribution is equal to the dimension of the physical space in which the points are distributed, i.e., $D=3$; this implies that if the distribution is finite in size and weakly clustered, such as that in the first galaxy catalogues, it is difficult to achieve this equality. For a non-integer dimension, the galaxy distribution is in a state of transition to homogeneity \citep{grujic2009fractal}.\\
The clustering description under the fractal hypothesis has encouraged the development of several theoretical models. One of the first models was proposed by \citet{mandelbrot1982the} using L\'evy flights to simulate the distribution of galaxy clusters regardless of the underlying physical phenomena. Similarly, \citet{peebles1989fractal} developed the basis for a statistical description of structures in the Universe; using correlation functions of up to four points for the first galaxy catalogues, he showed that the fractal dimension depends on the size of the distribution. Hence, this quantity changes from a pure fractal with dimension $D=1.26$ for $r<15.33~Mpc$ to structures consistent with the standard cosmological principle on larger scales \citep{peebles1989fractal}. Some observations have shown that the galaxy distribution grows in proportion to $r$ raised to an exponent related to the correlation dimension. For the galaxy distribution on a significant range of scales, this dimension is approximately 2 \citep{martinez2010statistics}, a very different value from 3, which is what would be observed for a homogeneous distribution because the number of galaxies grows in proportion to the volume of a sphere, as expected according to the standard cosmological model given the homogeneity and isotropy conditions.\\
Recent studies have also reported fractal and multifractal behaviour of the spatial distribution of galaxies and dark matter halos using galaxy surveys, in which large-scale fractal structures are evident \citep{pino1995evidence}; in particular, see \citep{zheng1988fractal} for \emph{CfA survey data}, \citep{xia1992fractal} for the \emph{IRAS survey}, and \citep{seshadri1999multi} for \emph{the Campanas redshift survey}.
\citet{chacon2012millennium} performed a multifractal analysis of dark matter halos from the \emph{Millennium} simulation \citep{springel2005simulations}; they found a transition to homogeneity between $100$ and $120~Mpc/h$, in strong agreement with the LCDM model. In contrast, an analysis with volume-limited samples from SDSS-DR7 \citep{chaconstudy, chacon2016multi, munoz2012galaxy} reported fractal behaviour at large scales until $165~Mpc/h$, where $D_q$ is always below the homogeneity limit $D=3$ for all values of the structure parameter. This result is consistent with \citep{joyce1999fractal, labini2009absence}, who found hierarchical patterns with self-similarity properties and a fractal dimension smaller than the dimension of physical space at scales greater than $100~Mpc/h$. For fractal analysis with WiggleZ, \citep{scrimgeour2012wigglez} reports a transition to homogeneity at $r_H=71\pm8~Mpc/h$ with $z\leq0.2$; this indicates that the galaxy distribution does not behave as a fractal object. This result is also consistent with those of \citep{hogg2005cosmic, yadav2005testing, sarkar2009scale}, who report a transition to homogeneity at $\sim70~Mpc/h$. \citet{wu1999large} and \citet{yadav2005testing} are in agreement with this value for the homogeneity scale; however, their studies show that at smaller scales the galactic cluster has fractal properties with dimension $D\approx1.2-2.2$.\\
In all cases, the effects of the geometry of the surveys must be taken into account according to \citep{yadav2010fractal}; in particular, the fractal calculations may be affected by the presence of holes and borders in the catalogues that are inherent to the process of observation using astronomical instruments. In this paper, we studied the multifractal behaviour of a spatial distribution of points using homogeneous samples built from the Sloan Digital Sky Survey-Baryon Oscillation Spectroscopic Survey (SDSS-BOSS) footprint \texttt{boss survey.ply}, including a random distribution of observational holes in right ascension and declination. We determined the fractal dimension in the range $-6<q<6$, the lacunarity spectrum, and the homogeneity scale using the sliding window technique for each sample.\\
In section 2, we define the main concepts of fractal formalism, such as the fractal dimension, multifractal description of galaxy clustering, lacunarity, and generalised fractal dimension. In section 3, we review the origin of observational holes, borders, and footprints of SDSS masks. Moreover, we present the construction of synthetic homogeneous samples limited in redshift, including the distribution of observational holes. Then, in section 4, we present the results of the application of our method to the constructed samples with the BOSS footprint. In section 5, the results for the multifractal dimension and lacunarity are discussed. Finally, our conclusions are given in section 6.
\section{Fractal and multifractal formalism
\subsection{Concepts of fractal dimension}\label{sec:fractalconcept}
The concept of dimension can be associated with the number of degrees of freedom or the minimum number of coordinates to specify any point within a distribution of points in a metric space. Topologically, the dimension indicates how much space a set occupies near each of its points \citep{falconer2004fractal}. The most intuitive definition of dimension is the \emph{topological dimension} $D_T$, introduced by Poincare and generalised by Lebesgue. In this definition, given a set of topological space $\mathbb{X}\in\mathbb{R}^n$, the dimension is the minimum value of $n$ for which every open cover admits a locally finite open refinement the order which does not exceed $n+1$. If there is not a minimum value of $n$, we say that the set is infinite-dimensional.
~In Euclidean space, $\mathbb{R}^n$, $D_T$ takes integer values in $(0,~n)$ \citep{mandelbrot1982the}.\\%, for instance an empty set has $D_T=-1$, for a point $D_T=0$, for a line segment $D_T=1$, for a surface $D_T=2$, and for a solid $D_T=3$.\\
When the sets describe irregular shapes, the concept of dimension in terms of the number of coordinates is insufficient to describe them. This fact has motivated the introduction of new concepts beyond the classical geometry \citep{mandelbrot1982the}; therefore, fractal geometry was developed. Thus, it is possible to give a different concept of \emph{dimension}; for instance, the \emph{self-similar dimension} can be explained by further fragmentation of an object or set, and the ratio of the number of identical parts, where each part is scaled down by the ratio $r$. For any set $X~\in~R^n$ that supports division into a finite number of subsets $N(k)$, where all of them are consistent with each other by translations and rotations, and it is a reduced copy of the initial set by a factor $r=1/k$, the self-similar dimension of $X$ is defined as the unique value $D$ satisfying the equation $N(k)=k^D$ \citep{mandelbrot1982the}, i.e.,
\begin{equation}\label{eq:DimensionFractal1}
D=\frac{\log N}{\log (k)}.
\end{equation}
Here $D$ is not necessarily an integer number; in some cases, it may be an integer and match the topological dimension. When the object cannot be subdivided into exact copies of itself, we can use the \emph{box-counting dimension}, in this case, a set $\mathscr{A}$ covered by a grid or regular boxes, all equal to each other, with side $\delta>0$; then, the number of boxes $N(\delta)$ needed to cover the figure is determined. This process is very natural for a computer, and it is not necessary that the figure be self-similar.
~Hausdorff and Besicovitch \citep{mandelbrot1982the} proposed a more general definition of \emph{dimension} that considers fractional values and can be defined for any set of points. Mandelbrot defined a fractal object as a set with Hausdorff dimension $D_H$ strictly exceeding its topological dimension; thus, sets with non-integer Hausdorff dimension are fractals.
In practice, the Hausdorff dimension is not always easy to calculate \citep{falconer2004fractal}. In this case, following the idea of the \emph{self-similar fractal dimension}, the \emph {mass-radius fractal dimension} $D_m$ is defined by a power law; this dimension is the measure of the total mass contained in a sphere of radius $R$ whose center is a point of the set, and the mass contained as a function of the radial size is determined as
\begin{equation}\label{eq:DimensionMasa1}
M(R)=FR^{D_m},
\end{equation}
where the factor $F$ is a function that may be different for fractals with identical dimension, and the density number of galaxies decreases as $n(R)\approx R^{D_m-d}$ for a set in $\mathbb{R}^d$ \citep{martinez2010statistics}.
\subsection{Multifractals, lacunarity, and generalised fractal dimension}
For point distributions from galaxy catalogues, the analysis is more complex than that for modelled distributions because fluctuations associated with the intrinsic characteristics of the mass distribution appear, so we need to use more general geometric estimators \citep{william2000distribution}. The fractal dimension is not sufficient to characterise a set unambiguously; consequently, the concept of multifractals and lacunarity have been introduced \citep{seshadri1999multi}.\\
Multifractals provide the most detailed description possible of the fractal properties of a distribution of points. Given a distribution in which each region exhibits a fractal behaviour, but the dimension changes from one place to another, it is possible to establish correlations between these dimensions and to do a complete analysis from a higher level \citep{blumenfeld1997levy}. The lacunarity is an estimator associated with the size of the holes and borders within the distribution; furthermore, it describes the ``texture'' of a fractal. This concept can be extended by other definitions of the fractal dimension as a correlation dimension \citep{martinez2010statistics}. Generally, if a fractal has large empty regions within the clustering, i.e., ``lacunae'' or holes, it will have a high lacunarity value, which increases the heterogeneity of the structure, and if a fractal is translation-invariant, the lacunarity will be low, and the structure will be very close to homogeneous. Different fractals can have the same dimension but very different lacunarity values.\\
For physical properties that depend on the scale, the fractal behaviour at small scales reported by \citet{peebles1989fractal}, \citet{bagla2008fractal}, and \citet{martinez1990universe} can be extended to a matter distribution at large scales using the multifractal formalism. To determine the lacunarity, first we must know the $F$ factor in equation~(\ref{eq:DimensionMasa1}), which is related to the average distance between neighbours. For each center of the point distribution, the number of particles $n_i(r)$ contained within a sphere of radius $r$ measured from the position of the $i$th particle is given by
\begin{equation}\label{eq:ni}
n_i(r)=\sum_{j=1}^N \Theta(r-|\mathbf{x_i}-\mathbf{x_j}|),
\end{equation}
where the sum is over all particles in the sample, and $N$ is the total number of particles. The coordinates of each particle in the three-dimensional space are denoted as $\mathbf{x_j}$, $j\neq i$, and $\Theta$ is the Heaviside function. The number of particles $n_i(r)$ around each galaxy taken as the center, with coordinates $\mathbf{x_i}$, is determined by counting the particles around the center that are located within a comoving sphere of radius $r$ \citep{celerier2005fractal}.\\
The correlation dimension is defined similarly to the mass-radius fractal dimension \citep{seshadri2005fractal}. To characterise the distribution, we must have all the information about the statistical moments in order to define the generalised dimension. The generalised correlation integral $C_q(r)$ is defined as
\begin{equation}\label{eq:Cq}
C_q(r)=\frac{1}{NM}\sum_{i=1}^M[n_i(r)]^{q-1},
\end{equation}
where $q$ is called the structure parameter and corresponds to an arbitrary real number, and $M$ is the number of particles used as centers. According to \citep{murante1997density}, from the correlation integral it is possible to perform a power series expansion of $\log{(r)}$ [equation~(\ref{eq:ExpanLOG})] and thus to calculate directly the multifractal dimension $D_q$ and the lacunarity spectrum $\Phi_q$. It is sufficient to keep the first two terms on the right side of equation~(\ref{eq:ExpanLOG}), which is simplified so that a simple relation between the generalised correlation integral and generalised fractal dimension is obtained as in equation~(\ref{eq:Cqrq}) \citep{chaconstudy}.
\begin{equation}\label{eq:ExpanLOG}
\log[C_q^{1/(q-1)}]=D_q\log(r)+\log(F_q)+\mathcal{O}\left(\frac{1}{\log(r)}\right).
\end{equation}
\begin{equation}\label{eq:Cqrq}
C_q(r)^{1/(q-1)}=F_qr^{D_q}.
\end{equation}
Thus, the generalised fractal dimension and generalised lacunarity can be defined in the same way as the mass-radius fractal dimension. The generalised dimension $D_q$ is given by
\begin{equation}\label{eq:DqDEF}
D_q=\frac{1}{q-1}\frac{d\log{C_q(r)}}{d\log{r}},
\end{equation}
and the generalised lacunarity, using the previously determined factor $F$ for each structure parameter $q$, will be
\begin{equation}\label{eq:Lacunarity}
\Phi_q=\frac{\langle(F_q-\langle F_q\rangle)^2\rangle}{\langle F_q\rangle^2}=\frac{\langle F_q^2\rangle}{\langle F_q\rangle^2}-1.
\end{equation}
For some values of $q$ such that $q_i\neq q_j$ which satisfy $D_{q_i}=D_{q_j}$, i.e., $D_q$ is independent of $q$ and $r$, the distribution is called monofractal because its dimension is constant. In addition, if $D_q$ is equal to the Euclidean dimension, the distribution is homogeneous. For $q\geq1$, $D_q$ explores the scaling behaviour in high-density environments within the distribution, which are associated mainly with clusters and superclusters, whereas for $q<1$, $D_q$ explores the behaviour in low-density environments, i.e., those associated with \emph{voids} \citep{sarkar2009scale}. A full spectrum of the generalised fractal dimension provides detailed information about the entire distribution, whether in regions of high density or low density. This allows us to connect the concept of fractal dimension with statistical measures used to quantify the distribution of matter on large scales. If the distribution of galaxies undergoes a transition to homogeneity, all values of the fractal dimension tend to the dimension of the physical space, that is, $D_q\simeq D=3$ for any value of $q$; at small scales, we expect to see a spectrum of values of the fractal dimension all different from $3$, as strongly structures defined before of a homogeneity transition; in this case, the lacunarity spectrum must approach zero at the same radial scale $r$.
\section{Data samples}
\subsection{Observational holes, borders, and footprint}
Galaxy surveys that use an optical fiber multi-object spectrograph require automated methods to controlling each fiber and effectively determining the regions of interest to record information from the target. In SDSS-I/SDSS-II, the fibers capture light from 640 objects simultaneously; in SDSS-III, this value increased to 1000. In both cases, the objects were observed in a circular field of radius $1.49º$ called ``\emph{tile}'' \citep{blanton2003efficient}. Although automated mechanisms were used, there are effects that cause erroneous or unmapped regions, which result in holes in the catalogue. Specifically, these effects are associated with the position of the optical fibers; the most significant of these effects is fiber collisions, which occur when two objectives are close enough together in the same observation that the fibers cross each other, exceeding the allowed angle. In SDSS-I/II, the collision radius was $55''$; in SDSS-III BOSS, it was $62''$.
There are additional effects inherent to the process of observation that cannot be ignored and that also produce sky regions that should be eliminated from the samples for various reasons, including effects due to seeing, bright stars that saturate the detectors, trails caused by meteors and satellites or nearby objects, and failures in the observation plates. These effects are identified and quantified in detail using masks.\\
The accumulation of these effects results in a distribution of observational holes in the galaxy samples. The effective area covered by the catalogue depends on the masks' quality and the sectors to be excluded from observation. The various types of masks depend exclusively on the catalogue. SDSS-I/II has the following masks: \emph{Bleeding}, \emph{Bright\_star}, \emph{Trail}, \emph{Hole}, and \emph{Seeing}. SDSS-III uses \emph{Bright Star Mask}, \emph{Centerpost Mask}, \emph{Bad Field Mask}, and \emph{Collision Priority Mask} [for details on each mask, see \citep{blanton2001efficient, smee2012multi, dawson2013baryon}]. All of these masks are exclusion masks and indicate the regions in which the data quality is unacceptable; i.e., if a random point is within the mask, it will be excluded from the catalogue. According to \citep{blanton2001efficient}, these areas are relevant to and useful for large-scale structure studies. For fractal analysis in particular, \citet{yadav2010fractal} refers to some effects that may be caused by the geometry of the catalogues, as holes not only produce incompleteness regions, but also modify the geometry of the masks by adding edges or borders.\\
Although masks include observational holes, there is a footprint, i.e., an observational template with the SDSS geometry, that represents the ideal galaxy mapping without holes, and it covers the largest area possible. In this paper, we used the BOSS footprint \texttt{boss\_survey.ply}. The masks and SDSS footprint are designed in convex polygons that are independently generated for each of the five filters in the catalogue. The geometry of the masks is given in terms of spherical polygons and arrays; these files are manipulated using the \emph{Mangle} code \citep{hamilton2004scheme}. In this case, a mask is an arbitrary union of weighted angular regions bounded by a number of edges.
\subsection{Synthetic samples}
The synthetic samples have a random point distribution that represents galaxies observed by the SDSS; they were built on the basis of the BOSS \emph{footprint} to cover the largest possible observation area, corresponding to $14555~deg^2$. The BOSS footprint is made up of 19 polygons for the two hemispheres of the catalogue. In this paper, we used only the northern hemisphere from the spectroscopic catalogue to cover $\sim9399$ deg${}^2$ (see Fig.~\ref{fig:HemisNorte}). The southern hemisphere polygons were removed from \texttt{boss\_survey.ply}. Then absolute equatorial coordinates $(\alpha,~\delta)\equiv(RA,~DEC)$ were assigned to each point within the \emph{footprint}. First, we selected a random polygon from the mask with a probability proportional to the product of the statistical weight and area of the polygon; this was then limited, and a random point was generated within the circle. We checked whether the point was located within the polygon. If it was, the point was kept; otherwise, the algorithm was repeated. The number density of points generated in the footprint corresponds to $3,273,548$, the same value as in the random catalogue \texttt{random0\_DR11v1\_LOWZ\_North.dat} available in \emph{SAS} \url{http://data.sdss3.org/sas/} (see Table~\ref{tab:NumDatCatalogos}). The \emph{random} catalogues have the same structure observed in SDSS-DR11 \texttt{galaxy\_DR11v1\_LOWZ\_North.dat}, with $2,401,952$ galaxies in the redshift range $0.00014<z<0.98723$.
\begin{figure}
\includegraphics[width=\columnwidth]{Footprint.pdf}
\caption{North Galactic projection in absolute equatorial coordinates from BOSS \emph{footprint} \texttt{boss\_survey.ply} spectroscopic catalogue generated using \emph{Mangle}. Colour bar indicates the redshift distribution limited to $0.002<z<0.2$.}
\label{fig:HemisNorte}
\end{figure}
\begin{table*}
\footnotesize
\centering
\caption{Number density in the SDSS galaxy survey, random and synthetic samples.}
\label{tab:NumDatCatalogos}
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
\multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}Density\\Number \end{tabular}}} & \multicolumn{2}{c|}{\textbf{DR7}} & \multicolumn{2}{c|}{\textbf{DR10-LOWz}} & \multicolumn{2}{c|}{\textbf{DR11-LOWz}} & \multicolumn{1}{c|}{\textbf{BOSS Footprint}} \\ \cline{2-8}
& \textbf{Random} & \textbf{Galaxy} & \textbf{Random} & \textbf{Galaxy} & \textbf{Random} & \textbf{Galaxy} & \textbf{Shuffler} \\ \hline
\textbf{\begin{tabular}[c]{@{}c@{}}Unlimited sample\\$0.00014<z<0.98723$\end{tabular}} & 1925442 & 105832 & 10162336 & 203614 & 14035192 & 281103 & \multirow{2}{*}{3273548} \\ \cline{1-7}
\textbf{\begin{tabular}[c]{@{}c@{}}Limited sample \\$0.002<z<0.2$\end{tabular}} & 1765923 & 97064 & 2329271 & 49420 & 3273548 & 66314 & \\ \hline
\end{tabular}
\end{table*}
Synthetic catalogues based on the BOSS footprint were limited in redshift to $0.002<z<0.2$, so the time interval between two events at these points is sufficiently small, and there are not significant changes in the large-scale structure patterns \citep{laporte2014evolution}. To model the redshift distribution of galaxies, it is necessary to have a complete theoretical model that takes into account the behaviour of the population of galaxies. Although a complete model is not available\citep{ross2012clustering}, the effects of the galaxy distribution in redshift can be known from information provided by the observed catalogues, some cosmological simulations, and random catalogues, considering that they always follow a radial selection function $n(z)$ \citep{ross2012clustering, kazin2010baryonic}. Naturally, the samples will be affected by the effects of the radial selection function, which result from the instrumental inability to detect faint galaxies at great distances. To determine $n(z)$, we used the shuffle method suggested by \citet{ross2012clustering} and \citet{kazin2010baryonic}, which consists of assigning randomly to each point of the sample a redshift value selected from the galaxy catalogue.\\
To build homogeneous synthetic samples, first we filtered the redshift values of the DR11 galaxy catalogue on $0.002<z<0.2$. Next, we used the shuffle method to assign the radial selection function and distribute all the points in the BOSS footprint to complete the synthetic sample.\\
For each point in the synthetic samples, the coordinates $(RA,~DEC,~z)\equiv(\alpha,~\delta,~z)$ are known. The area of observational holes in the SDSS-DR10 and DR11 masks has a maximum size of approximately $36.727~deg^2$; these holes can be modelled as ellipses with a semi-major axis $a_{max}=3.893^\circ$ and semi-minor axis $b_{max}=3.003^\circ$. The footprint holes are generated randomly by assigning a random point with coordinates ($\alpha_{rnd},~\delta_{rnd}$); an ellipse centered at this point is drawn with semi-axes $a\in(0,3.893]^\circ$ and $b\in(0,3.003]^\circ$, and then the ellipse is removed from the footprint. This process is done for different percentages of holes based on the total number of points in the mask. Thus, we obtain the samples \emph{footprint-10}, \emph{footprint-20}, \emph{footprint-30}, \emph{footprint-40}, \emph{footprint-50}, \emph{footprint-60}, and \emph{footprint-70} for the corresponding percentages. All the samples with holes are correlated because they are built cumulatively, i.e., \emph{footprint-10} contains \emph{footprint-20}, which contains \emph{footprint-30}, and so on. In Fig.~\ref{fig:PorcentajeHuecos}, the projections in $\alpha-\delta$ are illustrated for the samples used in the multifractal analysis described in the next section.
\begin{figure}
\includegraphics[width=1.0\linewidth,clip]{CatalogoHuecosSerie1.pdf}
\caption{$\alpha-\delta$ projections of the samples with holes limited to $0.002<z<0.2$. The percentages of holes were calculated on the basis of the total number of data points from the DR11 random sample.}
\label{fig:PorcentajeHuecos}
\end{figure}
\section{Multifractal analysis and lacunarity spectrum}
In the previous section, we described how we built the limited synthetic samples with $0.002<z<0.2$. The homogeneous footprint without holes is the densest sample in terms of the number of points, with the positions of $3,273,548$ objects in absolute equatorial coordinates (see Table~\ref{tab:SumarySamples}). The observational hole distribution in the samples covers values from $10\%$ to $70\%$ in the northern hemisphere BOSS footprint and contains an ideal sampling of all points registered in the catalogue with a total area of $\sim9399$ $deg^2$ (see Fig.~\ref{fig:PorcentajeHuecos}). The multifractal analysis was done in comoving Cartesian coordinates in order to compare the results with the standard cosmology.
Moreover, to perform the multifractal analysis, we made a sampling within the limited redshift samples without any assumptions about the shape of the group of points; consequently, it is necessary to determine the radius of a larger sphere than these can contain. According to \citet{gabrielli2006statistical}, the effective depth $R_s$ for all samples according to their limits in $\alpha-\delta$ can be determined by equation~(\ref{eq:RsFractal}).
\begin{equation}\label{eq:RsFractal}
R_s=\frac{R_d\sin(\delta\theta/2)}{1+\sin(\delta\theta/2)},
\end{equation}
where $R_d$ corresponds to the maximum radial distance from the sample, and $\delta\theta=\text{min}(\alpha_2-\alpha_1,~\delta_2-\delta_1)$, where $\alpha_2,~\alpha_1$ are the limits in right ascension, and $\delta_2,~\delta_1$ are the limits in declination.
The distortions in redshift space due to peculiar movements may be important for low-redshift galaxy samples. For synthetic catalogues, this effect can be neglected given the nature of the radial selection function, which was built from a statistically homogeneous and isotropic random field in the BOSS footprint. In Fig.~\ref{fig:MuestraEnComovil}, a projection in comoving Cartesian coordinates is shown for different samples used in our analysis.
\begin{table}
\small
\caption{Summary of redshift-limited samples created from the BOSS footprint with different percentages of observational holes used in the multifractal analysis.}
\label{tab:SumarySamples}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline\hline
\multicolumn{1}{|c|}{\textbf{Sample}} & \begin{tabular}[|c|]{@{}c@{}}\textbf{Area}\\$\mathbf{\Omega~(\textbf{deg}^2)}$\end{tabular} & \begin{tabular}[|c|]{@{}c@{}}\textbf{Area}\\$\mathbf{\Omega_{\text{holes}}~(\textbf{deg}^2)}$\end{tabular} & \multicolumn{1}{l|}{$\mathbf{R_s~(Mpc/h)}$} \\ \hline
0\% & 9399 & 0 & \multirow{8}{*}{396.68} \\ \cline{1-3}
10\% & 8459 & 940 & \multicolumn{ 1}{c|}{} \\ \cline{1-3}
20\% & 7519 & 1880 & \multicolumn{ 1}{c|}{} \\ \cline{1-3}
30\% & 6579 & 2820 & \multicolumn{ 1}{c|}{} \\ \cline{1-3}
40\% & 5639 & 3759 & \multicolumn{ 1}{c|}{} \\ \cline{1-3}
50\% & 4699 & 4699 & \multicolumn{ 1}{c|}{} \\ \cline{1-3}
60\% & 3759 & 5639 & \multicolumn{ 1}{c|}{} \\ \cline{1-3}
70\% & 2820 & 6579 & \multicolumn{ 1}{c|}{} \\ \hline\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth,clip]{MuestraEnComovil.pdf}
\caption{Redshift-limited samples in comoving Cartesian coordinates for different hole percentage estimates on the BOSS footprint.}
\label{fig:MuestraEnComovil}
\end{figure}
For each synthetic sample, we calculated the generalised correlation integral $C_q$, the fractal dimension $D_q$, and the lacunarity spectrum $\Phi_q$. The statistical uncertainty of the $C_q$ values was determined using the equation~(\ref{eq:Cq}) for the error propagation.
\begin{equation}\label{eq:PropErrorCq}
\Delta C_q(r)=\frac{1}{MN}\sum_{i=1}^M(q-1)\left[n_i(r)\right]^{q-2}\Delta n_i(r).
\end{equation}
The $\Delta n_i(r)$ uncertainty is determined numerically by counting the number of points that could have been included in $n_i(r)$ but were omitted because of the uncertainty in the position, that is, those points that statistically are outside the spherical shell, despite being very close to its borders \citep{chacon2012millennium}.\\
The distance step for the $C_q$ calculation is $\Delta r=r_{i}-r_{i-1}=1.0~Mpc/h$; this value corresponds to the average distance between two galaxies in clusters \citep{narlikar2002introduction, sharan2009spacetime, martinez2010statistics}. The value of $ C_q $ was determined for each sample using 13 structure parameters in the range $-6\leq q\leq6$ in order to compare the results with those of \citep{chaconstudy, scrimgeour2012wigglez}. Fig.~\ref{fig:Cq_BOSS_Huecos} shows the behaviour of $C_q$ with respect to the comoving distance; $q\geq0$ indicates high-density regions within the distribution, whereas $q<0$ indicates low-density regions in the same distribution. To avoid divergence in equation~(\ref{eq:Cqrq}) when $q=1$, $C_q$ is calculated using the numerical limit, in which case it converges to the average of the left- and right-hand limits of $q$.\\
We determine the fractal dimension $D_q$ using equation~(\ref{eq:DqDEF}). The sliding window technique presented by \citep{martinez2003statistics, rodrigues2004self} was applied to the logarithm of the correlation integral as a function of the logarithm of the comoving distance $r$, because in this case we have a linear relationship between these variables \citep{pietronero2004statistical, chacon2012millennium}, i.e., $\log(C_q)\propto\log(r)$, where the constant of proportionality is related to the fractal dimension; see equation~(\ref{eq:ExpanLOG}). Thus, the slope $\frac{1}{q-1}D_q$ and the intersection $\frac{1}{q-1}\log(F_q)$ approached a segment of a curve. Fig.~\ref{fig:HUECOS-Dim} shows the fractal dimensions of low- and high-density regions with the structure parameter in the range $-6<q<6$. From these plots, it is possible to obtain the behaviour of $D_q$ in terms of the structure parameter, as shown in Fig.~\ref{fig:Dq_VS_q-CadaR-Huecos}.\\
Our analysis is complemented by the lacunarity spectrum, which explain the hole distribution and the effect of border holes, and indicates how the set of points fills the space of the masks. Fig.~\ref{fig:Phi-HUECOS} presents the results for these quantities. In addition, we determined the homogeneity scale for the distribution of points with a margin of error of 1\% (see Fig.~\ref{fig:PorcenHuecos_VS_Escala_Homoge}). The dependence of the homogeneity scale $r_H$ on the density of holes in each sample is shown in Table~\ref{tab:ResultsRH}.
\begin{table}
\tiny
\caption{Homogeneity scale $r_H$ for $-6\leq q\leq6$ in each sample.}
\label{tab:ResultsRH}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline\hline
\multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}Hole\\percentage\end{tabular}}} & \multicolumn{6}{c|}{\textbf{q [($r_H\pm0.03$) Mpc/h]}}\\ \cline{2-7}
\textbf{} & \textbf{-6} & \textbf{-5} & \textbf{-4} & \textbf{-3} & \textbf{-2} & \textbf{-1}\\ \hline
\textbf{0} & 23.64 & 20.77 & 19.11 & 17.23 & 17.06 & 18.95\\ \hline
\textbf{10} & 154.58 & 155.28 & 150.11 & 148.65 & 151.16 & 144.86 \\ \hline
\textbf{20} & 73.10 & 78.85 & 72.39 & 71.75 & 56.49 & 69.28 \\ \hline
\textbf{30} & - & - & 64.38 & 47.69 & 37.47 & 67.55 \\ \hline
\textbf{40} & 87.33 & 85.49 & 84.85 & 79.27 & 76.69 & 53.33\\ \hline
\textbf{50} & - & 52.19 & 51.88 & 54.36 & 62.16 & 67.04 \\ \hline
\textbf{60} & - & - & - & - & 159.25 & 149.48 \\ \hline
\textbf{70} & - & - & - & - & - & - \\ \hline
\end{tabular}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
\multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}Hole\\percentage\end{tabular}}} & \multicolumn{6}{c|}{\textbf{q [($r_H\pm0.03$) Mpc/h]}}\\ \cline{2-7}
\textbf{} & \textbf{1} & \textbf{2} & \textbf{3} & \textbf{4} & \textbf{5} & \textbf{6}\\ \hline
\textbf{0} & 12.40 & 13.70 & 18.60 & 16.67 & 26.15 & 32.34 \\ \hline
\textbf{10} & 63.72 & 68.08 & 65.49 & 71.89 & 72.54 & 79.36 \\ \hline
\textbf{20} & 70.61 & 70.05 & 68.44 & 68.07 & 69.46 & 67.04 \\ \hline
\textbf{30} & 73.53 & 79.85 & 81.81 & 81.28 & 83.25 & 86.46 \\ \hline
\textbf{40} & - & - & - & - & - & - \\ \hline
\textbf{50} & 75.21 & 77.42 & 81.14 & 83.13 & 83.74 & 88.48 \\ \hline
\textbf{60} & 93.67 & 98.87 & 112.15 & 116.65 & 125.60 & - \\ \hline
\textbf{70} & 88.14 & 91.98 & 90.02 & 93.58 & 92.34 & 95.31 \\ \hline\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure*}
\centering
\includegraphics[width=1.0\linewidth,clip]{019Cq_BOSS_Huecos.pdf}
\caption{Generalised correlation integral $C_q(r)$ for all the values of $q$ studied in this paper for the footprint samples: left, high densities ($q \geqslant 1 $); right, low densities ($q < 1$).}
\label{fig:Cq_BOSS_Huecos}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.38\linewidth,clip]{HUECOS-Serie-1.pdf}
\includegraphics[width=0.38\linewidth,clip]{HUECOS-Serie-2.pdf}\\
\includegraphics[width=0.38\linewidth,clip]{HUECOS-Serie-3.pdf}
\includegraphics[width=0.38\linewidth,clip]{HUECOS-Serie-4.pdf}\\
\includegraphics[width=0.38\linewidth,clip]{HUECOS-Serie-5.pdf}
\includegraphics[width=0.38\linewidth,clip]{HUECOS-Serie-6.pdf}\\
\includegraphics[width=0.38\linewidth,clip]{HUECOS-Serie-8.pdf}
\includegraphics[width=0.38\linewidth,clip]{HUECOS-Serie-9.pdf}\\
\includegraphics[width=0.38\linewidth,clip]{HUECOS-Serie-10.pdf}
\includegraphics[width=0.38\linewidth,clip]{HUECOS-Serie-11.pdf}\\
\includegraphics[width=0.38\linewidth,clip]{HUECOS-Serie-12.pdf}
\includegraphics[width=0.38\linewidth,clip]{HUECOS-Serie-13.pdf}
\caption{Multifractal spectrum $D_q(r)$ as a function of the comoving radial distance $r$. For low-density environments, $-6\leq q\leq0$; for high-density environments, $q\geq0$. Error bars have an uncertainty of $\pm1\sigma$ and coincide with the width line of the curves.}
\label{fig:HUECOS-Dim}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=1.0\linewidth,clip]{015Dq_VS_q-CadaR-Huecos.pdf}
\caption{Multifractal spectrum $D_q(r)$ for each sample with structure parameter $q\in[-6,~6]$. The distance $r$ measured from the centers is shown in each subplot. The fractal dimension is bounded by $2.9\leq D_q\leq 3.2$ and converges to the physical dimension as $q$ increases. The sample with 40\% holes exhibits weak fractal behaviour with dimension $2.9<D_q<3.0$ at $r>90~Mpc/h$ around the centers in this sample. Solid line shows the Bezier interpolation of the data.}
\label{fig:Dq_VS_q-CadaR-Huecos}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.38\linewidth,clip]{Phi-Qpos-Serie-1.pdf}
\includegraphics[width=0.38\linewidth,clip]{Phi-Qpos-Serie-2.pdf}\\
\includegraphics[width=0.38\linewidth,clip]{Phi-Qpos-Serie-3.pdf}
\includegraphics[width=0.38\linewidth,clip]{Phi-Qpos-Serie-4.pdf}\\
\includegraphics[width=0.38\linewidth,clip]{Phi-Qpos-Serie-5.pdf}
\includegraphics[width=0.38\linewidth,clip]{Phi-Qpos-Serie-6.pdf}\\
\includegraphics[width=0.38\linewidth,clip]{Phi-Qpos-Serie-8.pdf}
\includegraphics[width=0.38\linewidth,clip]{Phi-Qpos-Serie-9.pdf}\\
\includegraphics[width=0.38\linewidth,clip]{Phi-Qpos-Serie-10.pdf}
\includegraphics[width=0.38\linewidth,clip]{Phi-Qpos-Serie-11.pdf}\\
\includegraphics[width=0.38\linewidth,clip]{Phi-Qpos-Serie-12.pdf}
\includegraphics[width=0.38\linewidth,clip]{Phi-Qpos-Serie-13.pdf}
\caption{Lacunarity spectrum $\Phi_q(r)$ in logarithmic scale as a function of comoving radial distance $r$, in low-density environments ($-6\leq q\leq0$) and high-density environments ($0\leq q\leq6$) for each redshift-limited sample.}
\label{fig:Phi-HUECOS}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=1.0\linewidth,clip]{021PorcenHuecos_VS_Escala_Homoge.pdf}
\caption{Comparison of homogeneity scale $r_H$ as a function of the hole density: left, low-density environments; right, high-density environments.}
\label{fig:PorcenHuecos_VS_Escala_Homoge}
\end{figure*}
\section{Discussion}
According to our results, the spectrum of the fractal dimension provides sufficient information to show a transition or rupture of the homogeneity scale at a specific scale. This justifies the use of statistically homogeneous samples following the observed radial selection function and geometric features of the galaxy masks. Our calculations take into account the maximum radius of the comoving circumscribing sphere of the edges for each sample; therefore, the fractal study can be performed by small and successive iterations within this sphere \citep{joyce2005basic}.\\
The method used to determine the fractal quantities was validated using a main sample without holes, which represents an ideal sample based on the BOSS footprint with 3,273,548 points distributed uniformly across the mask limited to $0.002<z<0.2$. In this sample, fractal behaviour does not occur; for all values of the parameter structure, the fractal dimension $D_q$ converges rapidly to the dimension of physical space, $D=3$, and the transition to homogeneity occurs at $r_H\sim19.436~Mpc/h$.\\%For different samples with holes and borders distribution our results are comparable to those of \citep{chacon2012millennium, chaconstudy, chacon2016multi, scrimgeour2012wigglez}.\\
The fractal dimension $D_q(r)$ was calculated from the generalised correlation integral $C_q(r)$ using the sliding window technique for radial depths that reach as large as $150~Mpc/h$. Fig.~\ref{fig:Cq_BOSS_Huecos} is consistent with the results of \citep{chaconstudy, sarkar2009scale}; this set of integrals shows the change in the average number of neighbours around the centers, so they increase for $q>1$ and decrease for $q<1$. For some $q$ values, these functions indicate the probability that a pair of points are at a distance less than or equal to $r$. All the samples studied exhibit the same behaviour for $C_q$ independent of the hole density; thus, the fractal dimension will be convergent and will be similar for different values of $q$. For distances $r>10~Mpc/h$, $C_q\rightarrow0$, which means that at small scales, defined structures appear that fill the space in a granular manner because the number of neighbours is less than $q\geq0$.\\
The effect of borders is evident in the lacunarity spectrum; this quantity reveals how the space is filled in terms of the hole density and sample morphology, which is given only by the irregular borders of the holes \citep{gefen1983geometric, allain1991characterizing}. The relative maxima and minima are observed in Fig.~\ref{fig:Phi-HUECOS}; these points indicate the distance at which there is a strong dependence on the density of holes. In addition, the distance between peaks is related to the size of the holes. $\Phi_q$ is proportional to the density of holes; for all values of the structure parameter, the lacunarity spectrum presents oscillatory behaviour consistent with the result of \citep{chaconstudy}. The minimum values of $\Phi_q$ are between $40$ and $120~Mpc/h$; that is, there are fewer holes at these distances, and the matter is distributed more evenly than on other scales.\\
The behaviour of the spectrum of the fractal dimension $D_q$ depends on the radial distance. In low-density regions, $D_q$ fluctuates around the dimension of physical space. For $D_{q<0}$, the dimension tends to $D=3$ as $q$ increases, indicating that the distribution is grouped into small spatial regions with self-similarity properties. When $q<0$, the amplitude of the fluctuations is large; this causes the growth rate of the fractal dimension to converge to the homogeneity scale $r_H$. In high-density regions, there is a strong tendency to homogeneity because the values of the fractal dimension are very close of the physical space dimension: $2.97\leq D|_{r_H}\leq3.03$. $D_{q\geq0}$ increases for large $r$ values to reach homogeneity; this means that on average, the space is completely filled at greater depths than $r_H$. Finally, for $r>80~Mpc/h$, the fractal behaviour disappears.\\
Our results show a relationship between the hole density from the masks and $r_H$. In particular, observational holes cause shifts in the homogeneity scale. When the hole density is lower but comparable to the density of objects in the sample, at least for a hole percentage of $40\%$, a fractal behaviour occurs with $2.70<D_{q\geq0}<3.20$ before the homogeneity scale is reached. For $q>0$ and hole percentages near $40\%$, $r_H$ shifts to scales on the order of $120~Mpc/h$. For percentages between $10\%$ and $30\% $, $r_H\sim70-90~Mpc/h$, and for percentages below $10\%$, $r_H\sim20~Mpc/h$, which is very close to the $r_H$ value for the BOSS footprint without holes. This value depends exclusively on the density of points in the samples and the radial selection function. For an infinitely dense sample, the homogeneity scale is $\sim1~Mpc/h$, which is the accuracy of the distance at which sampling is performed.
\section{Conclusions}
In this paper, we analysed the multifractal behaviour of synthetic samples built using the BOSS footprint and including different distributions of observational holes and borders. We conclude that the lacunarity increases if the borders describing complex curves that depend on the hole distribution in the masks. The relative minimum of the lacunarity spectrum in high-density environments ($q\geq1$) makes it possible to detect the spatial regions where there is a strong dependence on the density of holes. The minimum values of $\Phi_{q>0}$ are between $40$ and $120~Mpc/h$; in these regions, the matter is distributed more evenly than on other scales. In particular, for all the samples with holes and for each $q>0$, the values of the scale at the homogeneity transition are around $80~Mpc/h$.\\
Synthetic samples without observational holes exhibit a transition to homogeneity around $r_H\sim19.5~Mpc/h$ for all structure parameters. The analysis of samples with holes revealed that for low scales with $q<0$, there is a fractal behaviour, where fluctuations around the dimension of the physical space are smoothed to achieve homogeneity at $D_q=3.00\pm0.03$ for $r>130~Mpc/h$. This behaviour depends exclusively on the holes and borders in the sample. For $q>0$, the fractal dimension $D_q$ converges rapidly to $r_H$, with low lacunarity values and a tendency to fill the physical space on larger scales than $80~Mpc/h$.\\
For the samples studied, we found that $r_H$ depends on the density of observational holes. The holes cause two effects in the calculation of the fractal dimension: shifts in the homogeneity scale proportional to the number density of holes and weakly fractal behaviour with $2.9<D_q<3.0$ for a sample with about $40\%$ holes (see Fig. \ref{fig:Dq_VS_q-CadaR-Huecos}), where the hole distribution begins to be more important than the point distribution, so the homogeneity scale disappears slowly for all $q$ values
~In this case, the area covered by holes is smaller but is comparable to the area covered by objects. For high-density regions, with $q>0$ and a percentage of holes near $40\%$, $r_H$ is shifted to scales on the order of $120~Mpc/h$. For areas with $10\%$ and $30\%$ holes, the homogeneity scales are around $70$ and $90~Mpc/h$, respectively. Similarly, for lower percentages of holes, the homogeneity scale tends to $\sim19.5~Mpc/h$, which corresponds to the $r_H$ value for the BOSS footprint. Finally, the observational hole percentage in the masks must be taken into account in analyses of the homogeneity scale measured from galaxy samples.
\section*{Acknowledgements}
JEGF acknowledges financial support from the Facultad de Ciencias at Universidad Nacional de Colombia - Sede Bogot\'a to perform master studies. The authors thank Dr. C\'esar Alexander Chac\'on for providing the code for calculating fractal quantities and Dr. Florian Beutler for suggestions about BOSS footprint.\\
Funding for SDSS-III has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, and the U.S. Department of Energy Office of Science. The SDSS-III web site is http://www.sdss3.org/.\\
SDSS-III is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS-III Collaboration including the University of Arizona, the Brazilian Participation Group, Brookhaven National Laboratory, Carnegie Mellon University, University of Florida, the French Participation Group, the German Participation Group, Harvard University, the Instituto de Astrofisica de Canarias, the Michigan State/Notre Dame/JINA Participation Group, Johns Hopkins University, Lawrence Berkeley National Laboratory, Max Planck Institute for Astrophysics, Max Planck Institute for Extraterrestrial Physics, New Mexico State University, New York University, Ohio State University, Pennsylvania State University, University of Portsmouth, Princeton University, the Spanish Participation Group, University of Tokyo, University of Utah, Vanderbilt University, University of Virginia, University of Washington, and Yale University.
\bibliographystyle{mnras}
|
\section{Introduction}
Each generation of wireless mobile technology has been
driven by the need to meet new requirements
that could not be completely achieved by its predecessor.
Following this trend, fifth generation cellular
(5G) systems are now expected
to meet unprecedented speeds, near-wireline latencies and ubiquitous connectivity with uniform user Quality of Experience (QoE) \cite{samsung,osseiran2014scenarios}.
While current microwave bands below 3~GHz
have become nearly fully occupied, the millimeter wave
(mmWave) frequencies, roughly above $10$ GHz, have
enormous amounts of unused available spectrum.
These bands are widely expected to become a key means of addressing the challenge of higher required data rates \cite{mmW,rappaportmillimeter,RanRapE:14}.
However, one of the key challenges for cellular
systems in the mmWave bands is the rapid channel dynamics.
In addition to the high Doppler shift,
mmWave signals are completely blocked by many common
building materials such as brick and
mortar~\cite{rappaport2014millimeter}, and even the human body can cause
up to 35~dB of attenuation~\cite{lu2012modeling}.
As a result, the movement of obstacles and reflectors,
or even changes in the orientation of a handset relative to a body or a hand, can cause the channel to rapidly appear or disappear.
This high level of channel variability has widespread implications
for virtually every aspect of cellular design. This paper
focuses on one particular important design issue
which is the tracking of the
downlink channel quality and signal-to-noise ratio (SNR) at the mobile
user equipment (UE).
Measuring the SNR and reporting the value in periodic \emph{channel quality indicator} (CQI) reports is an essential component of any modern
cellular system -- see, for example \cite{LTE_book,schwarz2010calculation}
for a detailed description of the methods in 3GPP LTE.
Most importantly, CQI reports are the basis for rate prediction
and adaptive modulation and coding.
While CQI errors can be mitigated somewhat
via Hybrid ARQ (HARQ), HARQ requires retransmissions that may
result in excess delay. One of the goals of 5G systems is to achieve
very low ($< 1$ ms) air link latencies.
CQI and related signals measurements are
also necessary for proper handover determination and radio link failure detection, which are likely to become more common in mmWave
due to the small cell topology and the intermittency of the channel.
While CQI estimation is relatively straightforward in current
cellular systems, there are at least three
potentially complicating issues for mmWave:
(i) the rapid dynamics due to blockage events that strongly affect the link quality;
(ii) the need to track the CQI in multiple spatial directions with very
narrow beams; and
(iii) the limited number of available measurements
since the
\emph{cell reference signal} (CRS) used in current 3GPP LTE
systems may not be available for mmWave (see Section~\ref{sec:cqilte}).
To address these challenges, this paper presents two key contributions.
First, we propose a novel method for estimating the channel quality
using synchronization signals and directional scanning.
This signaling mechanism was also considered for initial access in
\cite{Barati,barati2015initial}.
We derive an unbiased estimate for the instantaneous
wideband SNR in a particular pointing direction.
The estimate can then be filtered over time to trade off noise
reduction and tracking speed.
Secondly, we evaluate the SNR tracking through real measurements
using a novel high-speed measurement system.
There are currently a large number of
measurements of mmWave outdoor channels and detailed statistical models
\cite{Mustafa,Samimi2015Prob,MacCartney2015Wideband,ns3_nokia}.
However, these measurements have been largely
performed in static locations with minimal local blockage.
The dynamics of the channel are not fully understood
-- see some initial work in \cite{eliasi2015stochastic,ferrante2015mm}.
In this work, we experimentally measure the dynamics of the channel
in various common blockage scenarios using a high-speed channel
sounder at 60~GHz. We then
combine the measured channel traces with the
statistical models to evaluate the SNR tracking algorithms.
\section{System Model}
\label{sec:system_model}
\subsection{CQI Estimation in 3GPP LTE} \label{sec:cqilte}
CQI estimation of the downlink channel is relatively straightforward
in 3GPP LTE \cite{LTE_book,schwarz2010calculation}:
the downlink channel quality is measured from what is
called the cell reference signal (CRS). This is a wideband signal
transmitted essentially continuously with one signal being sent
from each BS cell transmit antenna port.
Each UE in connected mode
monitors these signals to create a wideband channel estimate that can be used
both for demodulating
downlink transmissions and for estimating the channel quality.
However, in addition to the rapid variations of the channel,
there are two issues for CQI estimation in mmWave.
First, a CRS will likely not be available since downlink
transmissions at mmWave frequencies
will be directional and specific to the UE. Demodulation
reference signals
will thus likely follow the format of
LTE's UE-specific reference signals, which are
transmitted in-band with the data.
Thus, there will likely be no reference signals that are broadcast to all UEs.
Secondly, mmWave UEs are likely to use analog beamforming, meaning that
the UE can only measure the channel quality in one direction at a time
\cite{sun2014mimo,heath2015overview}.
\begin{figure}[t!]
\centering
\includegraphics[trim= 0cm 0cm 0cm 0cm , clip=true, width=0.8 \columnwidth]{slot_scheme_LTE.pdf}
\caption{Periodic transmission of narrowband synchronization signals from the BS. This structure is similar to the LTE PSS.}
\label{fig:slot}
\end{figure}
\subsection{Synchronization Signal Transmission Format}
In the absence of CRS, each UE
must find an alternate signal to measure the downlink channel quality.
For this work, we propose that
the UE estimates the channel quality from
periodic synchronization signals similar to the LTE primary or secondary
synchronization signals (PSS or SSS) used for initial access
and cell search.
These signals are transmitted at a much lower duty cycle and
the estimation of the channel
from these limited measurements
is one of the key challenges addressed in this paper.
For the structure of the synchronization signals, we assume the format
described in \cite{Barati} and reported in Figure \ref{fig:slot}. Similar to the LTE PSS, we assume that
each BS cell transmits a synchronization signal once every $T_{\rm per}$ seconds
for a duration of $T_{\rm sig}$ seconds.
These signals will be transmitted
omni-directionally or in a fixed pattern covering the cell area.
Each transmission consists of $N_{\rm sig}$ sub-signals where each
sub-signal is transmitted over a narrow band of $W_{\rm sig}$ Hz.
The use of multiple transmissions is for frequency diversity.
At the UE side, we assume that the UE
receiver attempts to estimate the received SNR of the
synchronization signals in $N_{\rm dir}$ different angular
directions.
As discussed above, we assume the UE performs analog beamforming
and hence can measure the synchronization signal in only
one direction at a time.
We thus assume that in each synchronization signal
period, the UE measures the received signal strength in
one of the $N_{\rm dir}$ angular directions. Hence, it can make
a received signal measurement in a particular angular direction once every
$N_{\rm dir}T_{\rm per}$ seconds.
The specific parameter values will be discussed in
Section \ref{sec:sis_par}.
\subsection{Channel Model and SNR Tracking}
\label{sec:SNR_tracking}
Let $p_{ik}(t)$ be the $k$-th transmitted sub-signal in the $i$-the
synchronization period. Let $t_i$ denote the time of the synchronization period
and $f_k$ the frequency location of the sub-signal within that period.
We assume that the sub-signal is received at the receiver as
\[
r_{ik}(t) = \mathbf{w}_i^{rx*}\mathbf{H}(t_i,f_k)\mathbf{w}_i^{tx} p_{ik}(t) + n_{ik}(t),
\]
where $\mathbf{w}_i^{rx}$ is the RX beamforming vector at the UE,
$\mathbf{w}_i^{tx}$ is the TX beamforming vector at the BS cell,
$\mathbf{H}(t_i,f_k)$ is the narrowband channel response for the synchronization signal,
and $n_{ik}(t)$ is AWGN. Note that, as described above, we assume that
each sub-signal is transmitted in a sufficiently narrow band that we can assume flat fading across the transmission. We let $N_0$ denote the noise
power spectral density.
We assume a standard multi-path channel model \cite{Mustafa}
where the time-varying channel response is given by
\begin{equation} \label{eq:Hmp}
\mathbf{H}(t,f) = \frac{1}{\sqrt{L}} \sum_{\ell =1}^L \sqrt{g_\ell(t)}
e^{2\pi j (f_{d,\ell} t - \tau_\ell f)}
\mathbf{u}^{rx}_\ell \mathbf{u}^{tx *}_\ell,
\end{equation}
where $L$ is the number of paths and, for each path $\ell$,
$g_\ell(t)$ is the time-varying channel power,
$f_{d,\ell}$ is the path Doppler shift, and $\mathbf{u}^{rx}_\ell$ and $\mathbf{u}^{tx}_\ell$
are the RX and TX spatial signatures of the path that depend on the
angles of arrival and departure of the path from the antenna arrays.
In this work, we are interested in tracking the SNR in a single
TX and RX pointing direction.
As described in the previous
subsection, the BS cell will use a fixed transmit direction
and the UE receiver will scan $N_{\rm dir}$ beamforming
directions and estimate the SNR separately in each direction.
For the remainder of this paper, we focus on a subset of the
transmission times $i$ where
the TX and the RX are pointed in a particular direction $\mathbf{w}_i^{tx} = \mathbf{w}^{tx}$
and $\mathbf{w}_i^{rx} = \mathbf{w}^{rx}$, for some fixed $\mathbf{w}^{tx}$ and $\mathbf{w}^{rx}$.
Given TX and RX directions $\mathbf{w}^{tx}$ and $\mathbf{w}^{rx}$,
define the wideband average channel gain as
\[
G(t) = \frac{1}{W_{\rm tot}}
\int_{f_c-W_{\rm tot}/2}^{f_c+W_{\rm tot}/2}
|\mathbf{w}^{rx*}\mathbf{H}(t,f) \mathbf{w}^{tx}|^2df,
\]
where the integral is over the total
system bandwidth of $W_{\rm tot}$ at center frequency $f_c$.
If the base station transmits at a power $P_{tx}$, then the average wideband SNR would be
\begin{equation} \label{eq:gamwide}
\gamma(t) := \frac{G(t)P_{tx}}{N_0W_{\rm tot}},
\end{equation}
where $W_{\rm tot}$ is the total system bandwidth.
We call $\gamma(t)$ the \emph{true wideband} SNR.
As stated before, since mmWave cells will not transmit a CRS,
we wish to estimate the wideband SNR $\gamma(t)$
from the synchronization signal.
The wideband SNR can be estimated as follows:
let $E_s = \int |p_{ik}(t)|^2dt$ denote the transmitted signal energy per sub-signal.
We assume this does not vary with $i$ or $k$.
If the transmit power is $P_{tx}$, the signal duration is $T_{\rm sig}$ and there are
$N_{\rm sig}$ signals,
\begin{equation} \label{eq:Es}
E_s = \frac{P_{tx}T_{\rm sig}}{N_{\rm sig}}.
\end{equation}
Now suppose that the receiver applies
a matched filter for each sub-signal to obtain the statistic
\begin{align}
\MoveEqLeft z_{ik} = \frac{1}{\sqrt{E_s}}\int p_{ik}^*(t)r_{ik}(t)dt \label{eq:zik} \\
&= \sqrt{E_s}\mathbf{w}^{rx*}\mathbf{H}(t_i,f_k)\mathbf{w}^{tx} + v_{ik},
\quad v_{ik} \sim {\mathcal CN}(0,N_0). \nonumber
\end{align}
It is easy to verify that if the frequency $f_k$ is uniformly randomly distributed
over the system bandwidth, then
\[
\mathbb{E}\left[|z_{ik}|^2\right] = \frac{G(t)P_{tx}}{N_{\rm sig}} + N_0.
\]
Hence, we can form an unbiased estimate of $\gamma(t)$ in \eqref{eq:gamwide} by
\begin{equation} \label{eq:gamhati}
\hat{\gamma}_i = \frac{1}{N_0T_{\rm sig}W_{\rm tot}}
\sum_{k=1}^{N_{\rm sig}} \left[ |z_{ik}|^2 - N_0 \right],
\end{equation}
which sums the received power on the $N_{\rm sig}$ sub-signals and subtracts the noise.
\subsection{Filtering Algorithms}
\label{sec:filtering_algh}
\begin{table}[!t]
\centering
\renewcommand{\arraystretch}{1.2
\begin{tabular}{|l|l|}
\toprule
\textbf{Symbol} & \textbf{Description} \\
\hline
$\gamma_i$ & \begin{tabular}{@{}l @{}} Wideband true SNR \end{tabular} \\
\hline
$\hat{\gamma}_i$ & \begin{tabular}{@{}l @{}} Raw SNR estimate of $\gamma_i$ \\ from the synchronization signals \end{tabular} \\
\hline
$\bar{\gamma}_i$ & \begin{tabular}{@{}l @{}} Time-filtered SNR estimate of $\gamma_i$ from $\hat{\gamma}_i$ \end{tabular} \\
\bottomrule
\end{tabular}
\caption{Symbols for the SNR and its estimates.}
\label{tab:SNR_traces}
\end{table}
Since $\hat{\gamma}_i$ in \eqref{eq:gamhati} is an estimate of the wideband SNR that has been obtained starting from the synchronization signals, it may deviate from the true SNR due to noise.
We call the measurement $\hat{\gamma}_i$ the \emph{raw} SNR.
To reduce the noise, we can filter the raw SNR producing a time-averaged value
that we will denote by $\bar{\gamma}_i$. We consider three possible filtering schemes \cite{Dcom}:
\begin{itemize}
\item \emph{No filtering:} In this case, we simply take $\bar{\gamma}_i = \hat{\gamma}_i$.
\item \emph{First-order filtering:} This uses a simple low-pass filter:
\begin{equation} \label{eq:gamFO}
\bar{\gamma}_i = (1-\alpha) \bar{\gamma}_{i-1} + \alpha \hat{\gamma}_i,
\end{equation}
for some constant $\alpha \in (0,1)$.
\item \emph{Moving average filtering:} In this algorithm, we simply average the last $M$
values,
\begin{equation} \label{eq:gamMA}
\bar{\gamma}_i = \frac{1}{M} \sum_{j=1}^M \hat{\gamma}_{i-j+1}.
\end{equation}
\end{itemize}
Therefore $\bar{\gamma}_i $ is a \emph{filtered SNR estimate} of $\gamma_i$, obtained starting from the noisy raw SNR $\hat{\gamma}_i$. Our goal is to find the optimum scheme to minimize the \emph{average estimation error} ${e_i =\mathbb{E}\left[|\bar{\gamma}_i -\gamma_i|\right]}$, in order to derive an SNR stream that can be used to reliably estimate the channel quality.
\section{Experimental Evaluation}
\label{sec:testbed}
\subsection{Channel Modeling Overview}
\begin{figure*}[t!]
\centering
\includegraphics[trim= 0cm 0cm 0cm 0cm , clip=true, width= \textwidth]{horn-antenna_small}
\caption{Our mmWave testbed. We introduce an obstacle (person walking, hand, metal plate) in front of the receiver to observe the received power drop.}
\label{fig:TX}
\end{figure*}
While there has been considerable progress in understanding the mmWave channel
for long-range outdoor cellular links, most of the studies have been performed
in stationary locations with minimal local blockage. For example,
in the New York City studies in \cite{Mustafa,Samimi2015Prob,MacCartney2015Wideband,ns3_nokia},
the RX was placed in a fixed location on a cart. In addition, there were
no obstacles in the immediate vicinity of the RX, such as a hand or a person,
whose movement would cause signal variations due to blockage.
Unfortunately, measuring a wideband spatial channel model with dynamics is not possible
with our current experimental equipment. Such a measurement would require that the
TX and RX directions be swept rapidly during the local blockage event. Since our
platform relies on horn antennas mounted on mechanically rotating gimbals, such
rapid sweeping is not possible.
In this work, we thus propose the following alternate approximate method to generate
realistic dynamic models for link evaluation:
\begin{enumerate}
\item We first randomly generate the number of paths, relative power, delay and
angles of arrival and departure based on the wideband channel models in \cite{Mustafa} and \cite{ns3_nokia}.
These models are based on extensive measurements in New York City in links similar to a
likely urban micro-cellular deployment, and would reflect the characteristics of a
stationary ground-level mobile with no motion nor local obstacles.
\item Combining the angles of arrivals and departure with the
antenna array patterns at the BS and UE, we can then determine the spatial signatures
$\mathbf{u}_\ell^{rx}$ and $\mathbf{u}_\ell^{tx}$ in \eqref{eq:Hmp}. The randomly generated parameters
from Step 1 will also provide the delay and power of each path,
that we will denote by $\tau_\ell$ and $P_\ell$, respectively.
\item We assume a random direction of motion of the UE receiver. Based on the UE velocity
and angle of motion relative to the angle of arrivals of the path, we can compute the Doppler
shifts $f_{d,\ell}$ in \eqref{eq:Hmp} by $f_{d,\ell} = f_{d,max}\cos\theta_\ell$, where $f_{d,max}$ is the maximum
Doppler shift and $\theta_\ell$ is the angle between the path
angle of arrival and direction of motion.
\item Finally, if there were no local blockage,
then the path powers $g_\ell(t)$ in \eqref{eq:Hmp} could be fixed as
$g_\ell(t) = P_\ell$, where the values
$P_\ell$ are the path powers generated in the static model in Step 1.
To simulate local blockage, we assume that these powers will be modulated as
\begin{equation} \label{eq:gell}
g_\ell(t) = \beta P_\ell h(t),
\end{equation}
where $h(t)$ is a time-varying scaling factor accounting for the blockage and $\beta$ is a scaling factor.
Since there are no statistical models for the blockage dynamics, we measure
traces of $h(t)$ experimentally in various blockage scenarios.
The factor $\beta$ can then be adjusted to set a desired test SNR, according to the envisioned target rate a mmWave user is expected to reach. We refer to Section \ref{sec:sis_par} for further details on the choice of this parameter.
\end{enumerate}
This four step procedure thus provides a semi-statistical model, in which
(i) the spatial characteristics of the channel are determined from
static statistical models derived from outdoor measurements and
(ii) local blockage events are measured experimentally and modulated on top of the
static parameters.
An important simplification in \eqref{eq:gell} is that we assume that
the local blockage $h(t)$ equally attenuates all paths, which may not be realistic.
For example, a hand may block only paths in a limited range of directions.
However, this work considers the SNR tracking in only one direction at a time.
In any fixed direction, most of the power is contributed only from paths within
a relatively narrow beamwidth and thus the approximation that the
paths are attenuated together may be reasonable.
\subsection{Channel Sounding System}
We will call the scaling term $h(t)$ in \eqref{eq:gell} the \emph{local blockage factor}\footnote{Note that the absolute value of $h(t)$ is immaterial, since the total channel power will be scaled by the factor $\beta$ in \eqref{eq:gell} to target a particular SNR.}. The key challenge in measuring the dynamics of local blockage is that we need
relatively fast measurements. To perform these fast measurements, we used the
experimental channel sounding system in Figure \ref{fig:TX}: a high-bandwidth baseband processor, built on a PXI (a rugged PC-based platform for measurement and automation systems) from National Instruments, which engineers a real-world mmWave link. The transmitter and receiver operate in two separate boxes, each of which have the parts listed below:
\begin{itemize}
\item[(i)] an 8-slot chassis, capable of holding a variety of expansion cards.
\item[(ii)] a $1.73$ GHz quad-core PXIe controller that runs a realtime
operating system (RTOS) called PharLap, and communicates with the computer used to run the experiments through an Ethernet connection to coordinate the operation of each peripheral card in the chassis.
\item[(iii)] two FPGA cards for the baseband signal processing.
\item[(iv)] a FlexRIO adapter module (FAM) card and a converter between
the baseband signal and an IF signal, which are connected to the antenna.
\item[(v)] mmW Converters, to convert the IF signal to mmWave in the range of $57-63$ GHz. The IF signal is mixed with the output of a local oscillator (LO), filtered, amplified, and sent over a waveguide output. We use 23~dBi directional
horn antennas (manufactured by Sage millimeter) to interface with the waveguide. This converter works in tandem with a power supply and a controller card. An identical converter at the receiver performs the down-conversion from mmWave frequencies to IF.
\end{itemize}
To sound the channel, we used a standard frequency-domain
method: the \emph{transmitter} sent a continuous
repeating pattern created from an IFFT of a 128 point pseudo random
QPSK sequence. We will call each group of 128 samples a symbol.
The sample rate is 130~MHz corresponding
to a symbol period of approximately 1~$\mu$s. Note that this symbol period
is larger than the maximum delay spread.
The \emph{receiver} segments the received time domain sequence
into symbols, takes the FFT of each symbol and derotates it by the frequency-domain representation of the transmitted sequence.
Since the transmitted signal
is periodic, the derotated signal at the receiver will provide
an estimate of the frequency-domain response of the channel.
To reduce the effect of the noise, the sequence is averaged over 32 symbols,
providing one averaged response every $32 \times 1 = 32$ ~$ \mu$s.
The averaged response is then converted to time-domain, to produce
the power delay profile (PDP) of the channel.
The phase noise at the receiver can be large (the manufacturer specification
is up to $-80$ dBc). This is a common problem in many mmWave RF units.
A characterization of this receiver
in \cite{aditya} found the maximum frequency deviation to be up 50 kHz,
which would be too large to leave uncompensated.
To compensate for the phase noise, in each
32 symbol measurement period the receiver derotated the signal by
9 frequency hypotheses spaced uniformly from $-50$ to 50 kHz,
and a potential PDP was generated from each of the 9 different
hypotheses. The PDP with the maximum peak was then selected amongst the
9 hypotheses. After this phase compensation, the received
symbols are sufficiently coherent over the 32 $\mu$s period needed for a new averaged response to be provided.
\subsection{Measurement of Local Blockage}
\label{sec:meas_local_block}
\begin{figure}[t!]
\centering
\includegraphics[trim= 0cm 0cm 0cm 0cm , clip=true, width= \columnwidth]{blockage_40.pdf}
\caption{ SNR trace perceived when receiving the synchronization signals. The solid line is obtained by simulating the statistical channel described in \cite{Mustafa} and \cite{ns3_nokia} (without local obstacles). In the dashed line, the experimentally measured local blockage dynamics are modulated on top of the statistical trace. The blockage is referred to a person walking multiple times between the transmitter and the receiver.}
\label{fig:blockage}
\end{figure}
Using the above system,
the blockage experiments were conducted by
placing the transmitter and the receiver on a one-meter high pedestal,
facing each other, at a distance of $4$ meters.
A laser pointer was used to improve the alignment between the two devices.
After this set-up, the system is then run to continuously
collect PDPs during a blockage event.
Blockage events are simulated by
placing moving obstacles between the transmitter and the receiver. In
this work, we considered three common blockage events: (i) a person walking (or running) between TX and RX; (ii) a wood (or metal) plate held between the two communication edges; (iii) a hand holding a cellular phone.
The system was run during each of these blockage events
for a total time of 10 seconds.
During this time, PDPs were measured
at a rate of one PDP per 32~$\mu$s.
We found that the dynamics of the channel varied considerably
slower than this rate, so we decimated the results
by a factor of four, recording one PDP per 128 $\mu$s.
Since each experiment was run for $10$ seconds,
each experiment resulted in $10^7/128 = 78125$ PDP recordings.
To determine the local blockage function, we are only interested
in the line-of-sight path. The power on this path was determined
from the maximum peak in the PDP. Reflected paths would appear in other
samples and thus be rejected. This received power then provides
the trace for the local blockage function $h(t)$ in \eqref{eq:gell}.
As described above, this local blockage function is then
used to modulate the time-varying channel response obtained
from the statistical channel model.
As an example, Figure \ref{fig:blockage} shows an SNR trace in which the blockage event is referred to a person walking multiple times between the transmitter and the receiver.
\subsection{Evaluation Results}
Once the SNR trace $\gamma(t)$ from the synchronization signals has been obtained, combining the simulated statistical channel described in \cite{Mustafa} and \cite{ns3_nokia} with the local blockage dynamics measured experimentally, the raw SNR $\hat{\gamma}(t)$ can be estimated from the synchronization signals following Section \ref{sec:SNR_tracking}. In Section \ref{sec:sis_par} we describe the system parameters that we use in our simulations, while in Section \ref{sec:res} we analyze and compare the performance of the presented linear filters, applied to the raw SNR trace, to obtain the estimate $\bar{\gamma}(t)$.
\section{Simulation parameters}
\label{sec:sis_par}
\begin{figure*}[t!]
\centering
\includegraphics[trim= 0cm 0cm 0cm 0cm , clip=true, width=1 \textwidth]{Trace_new_40.pdf}
\caption{ SNR trace $\gamma(t)$ together with its raw version $\hat{\gamma}(t)$ and its estimation $\bar{\gamma}(t)$ after different filtering schemes are applied. The upper lines refer to the SNR of a $50^{\text{th}}$ percentile typical user, while the lower lines refer to the SNR of a $5^{\text{th}}$ percentile edge user.}
\label{fig:trace_filters}
\end{figure*}
In this section, we derive some parameters that we will use in the SNR tracking simulations: (i) the scaling factor $\beta$ of Equation \eqref{eq:gell}, to set a desired test SNR, and (ii) the SNR trace downsampling factor.
\subsection{SNR scaling factor $\beta$}
\begin{table}[!t]
\centering
\renewcommand{\arraystretch}{1.2
\begin{tabularx}{1\columnwidth}{@{\extracolsep{\fill}}ccc}
\toprule
\textbf{Description}& \textbf{$50^{\text{th}}$ percentile} & \textbf{$5^{\text{th}}$ percentile} \\
\toprule
\begin{tabular}{@{}c @{}} LTE spectral efficiency \\ $\rho$ (bit/s/Hz/$W_{\rm tot}$) (from \cite{3GPP}) \end{tabular} & $\textbf{3.28}$ & $\textbf{0.154}$ \\
\midrule
\begin{tabular}{@{}c @{}} LTE rate (Mbps)\\ ($R_\mu = \rho \cdot W_{\rm tot} $) \end{tabular} & $3.28 \cdot 50 = \textbf{164}$ & $0.154 \cdot 50 = \textbf{7.7} $ \\
\midrule
\begin{tabular}{@{}c @{}} mmWave rate (Mbps)\\ (from \cite{Boccardi}, $R_{mmW} \simeq 9 R_\mu) $ \end{tabular} & $\textbf{1480}$ & $\textbf{70}$ \\
\bottomrule
\end{tabularx}
\caption{Cell user 4G-LTE and expected 5G rate, for average-cell-position users ( $50^{\text{th}}$ percentile) and cell-edge users ($5^{\text{th}}$ percentile). For the LTE case, we refer to a DL SU-MIMO $4 \times 4$ TDD baseline for a microwave system using $50$ MHz of bandwidth. For the mmWave case, we refer to a system with $500$ MHz of bandwidth and a single user.}
\label{tab:rate}
\end{table}
As previously asserted, the
scaling factor $\beta$ in \eqref{eq:gell}
is selected to bring the average SNR to some desired test level.
We consider two test cases:
\begin{enumerate}
\item the user belongs to the $50^{\text{th}}$ percentile, so it presents average propagation conditions;
\item the user belongs to the $5^{\text{th}}$ percentile, to simulate the $5\%$ worst user rate at cell edges.
\end{enumerate}
For each case, the target test SNR is obtained by: (i) determining a reliable 4G-LTE target rate for the user, according to \cite{3GPP}; (ii) determining the corresponding mmWave target rate, according to \cite{Boccardi}; and (iii) finding the corresponding target test SNR through a Shannon capacity evaluation.
In order to find a reliable data rate for the two user cases we are considering, we refer to the actual LTE 3GPP user data rate. We consider the performance of the DL SU-MIMO $4 \times 4$ TDD baseline in \cite{3GPP}, for a microwave system using $50$ MHz of bandwidth. In Table \ref{tab:rate}, we show the cell user spectral efficiency $\rho$ that we use to compute the LTE data rate $R_\rho$ (actually $R_\rho=\rho \cdot W_{\rm tot}$).
The corresponding DL data rate for a mmWave system with $500$ MHz of bandwidth can be determined referring to \cite{Boccardi}, where it is reported that the data rate of a mmWave user is expected to be around $9$ times higher than the rate of current LTE systems. According to this result, Table \ref{tab:rate} reports the corresponding rates that can realistically be achieved by mmWave users.
Based on the mmWave user data rate, we can estimate the corresponding target SNR $\gamma_t$ using the Shannon capacity:
\begin{equation}
R = \delta \cdot W_{\rm tot} \cdot \log_2(1+\gamma_t) \Longrightarrow \gamma_t = 2^{\tfrac{R}{\delta \cdot W_{\rm tot} }}-1
\label{eq:shannon}
\end{equation}
where $R$ is the data rate, $W_{\rm tot}$ is the system
bandwidth ($500$ MHz for the mmWave system we are considering), and $\gamma_t$ is the target data SNR. $\delta = 0.8$ is a parameter that accounts for a $20\%$ control overhead \cite{Mustafa}.
Solving \eqref{eq:shannon}, we obtain the target SNR $\gamma_t$.
The value $\gamma_t$ is the wideband SNR we would expect on a data channel.
The data channel would be received with the BS and UE performing beamforming.
However, the synchronization signals would be transmitted omni-directionally
and thus would be received at a lower SNR. In the experiments below,
following \cite{Mustafa},
we will assume that the BS cell has $N_{tx}=64$ antennas, allowing
up to a 18 dB beamforming gain. This gain would not be available for
the synchronization and thus the synchronization signals would be received
at a much lower SNR -- this is one of the main challenges in SNR tracking.
Thus, we assume that the wideband SNR (in linear scale) should be
$\gamma_t/N_{tx}$.
To set this SNR, we first
generate a random trace of using the statistical model. Then, we set the
factor $\beta$ in \eqref{eq:gell} to scale the average value of the
wideband SNR $\gamma(t)$ in the experiment to the desired target level $\gamma_t/N_{tx}$.
This generates the sequence for the wideband SNR $\gamma(t)$.
The raw estimate of the SNR $\hat{\gamma}(t)$
is then computed according to Section \ref{sec:system_model}.
\subsection{SNR Trace Downsampling Factor}
As we stated in Section \ref{sec:meas_local_block}, the measured SNR trace is composed of $78125$ samples, one every $128 \: \, \mu s$. According to Section \ref{sec:system_model} and the results in \cite{CISS}, we assume that each synchronization signal is transmitted periodically once every $T_{\rm per}=1$ ms for a duration of $T_{\rm sig} = 10 \:\mu$s, to maintain an overhead of $1\%$. Moreover, we also assume that the user directionally receives such signals by performing an exhaustive search of the angular space through $N_{\rm slot}=16$ directions. Therefore, the transmitter and the receiver will be perfectly aligned just once every $N_{\rm slot} T_{\rm per} = 16 $ ms. For this reason, the original SNR trace has been downsampled, keeping just one sample every $16$ ms.
\section{Performance Evaluation}
\label{sec:res}
In Section \ref{sec:res_filter}, we analyze and compare the performance of the filtering algorithms described in Section \ref{sec:filtering_algh}. The goal is to determine the filter that minimizes the estimation error ${e(t) =\mathbb{E}\left[|\bar{\gamma}(t) -\gamma(t)|\right]}$, for different user propagation characteristics ($50^{\text{th}}$ and $5^{\text{th}}$ percentiles). Furthermore, Section \ref{sec:res_error} shows how the estimation error changes when considering different SNR regimes.
\subsection{Filters performance comparison}
\label{sec:res_filter}
In Figure \ref{fig:trace_filters}, we plot the SNR trace $\gamma(t)$, whose blockage events refer to a person walking multiple times between the transmitter and the receiver. The upper line refers to a $50^{\text{th}}$ percentile typical user and the lower line to a $5^{\text{th}}$ percentile edge user. The figure also shows the noisy version $\hat{\gamma}(t)$ of the SNR trace, together with its estimate $\bar{\gamma}(t)$ after the presented linear filters have been applied. Two different scaling factors $\beta$ have been applied, when computing the SNR trace from the synchronization signals, according to the two user propagation regimes.
We see that, for low SNR regimes, the raw SNR trace $\hat{\gamma}(t)$ shows a very noisy trend, which is considerably different from its original version. The reason is that, when the user receives the synchronization signals with very low power (e.g., when located at the cell edge), the noise component dominates the SNR unbiased estimate in \eqref{eq:gamhati}, and therefore the raw SNR substantially differs from $\gamma(t)$. A filtering algorithm is thus required, to recover a stream $\bar{\gamma}(t)$ that can be used to more accurately estimate the channel.
The main concern is that it is hard to discern between the downspikes which refer to an actual blockage and those which accidentally manifested due to the additive noise. In such a way, the detection of real radio link failure situations might be distorted and might lead to false alarm or missed blockage detection events.
However, a simple first-order filter can produce an estimated SNR trace $\bar{\gamma}(t)$ that appears very similar to the measured one. Therefore, even without designing much more complex and expensive nonlinear adaptive filters, we can properly restore the desired SNR stream and perform reliable link failure detection and channel estimation. It should finally be noted that this filter requires a transient phase before reaching its normal operation, as can be seen in the first milliseconds of the traces in Figure \ref{fig:trace_filters}.
It is interesting to note that, when considering a good SNR regime (upper lines of Figure \ref{fig:trace_filters}, simulating a $50^{\text{th}}$ percentile user), even the raw SNR trace $\hat{\gamma}(t)$ (when no filters are applied) almost overlaps with its measured original version. Therefore, when an average position mmWave user receives uncorrupted synchronization signals, it estimates an SNR trace $\hat{\gamma}(t)$ that sufficiently resembles the measured one, and can hence perform an adequately reliable channel estimation without any further signal processing.
\begin{figure}[t!]
\centering
\includegraphics[trim= 0cm 0cm 0cm 0cm , clip=true, width=1 \columnwidth]{CDF_new.pdf}
\caption{CDF of the estimation error ${e(t) =|\bar{\gamma}(t) -\gamma(t)|}$ for a $5^{\text{th}}$ percentile edge user, when different linear filters are applied to the noisy SNR trace $\hat{\gamma}(t)$.}
\label{fig:CDF}
\end{figure}
In Figure \ref{fig:CDF} we plot the CDFs of the estimation error for a $5^{\text{th}}$ percentile edge user, when different linear filters are applied to the noisy SNR trace $\hat{\gamma}(t)$. The trace after a first-order filter is used shows much better performance with respect to the moving-average filtered trace which, besides its poor efficiency, is also affected by a non-negligible delay.
Therefore, among the options we considered, a first-order filter is the best choice to reduce the estimation error and properly track the SNR trace.
\subsection{Analysis of the estimation error }
\label{sec:res_error}
In Figure \ref{fig:error_vs_gamma}, we show the average estimation error ${e(t) =\mathbb{E}\left[|\bar{\gamma}(t) -\gamma(t)|\right]}$ versus different target SNR values $\gamma_t$, obtained by adjusting the scaling factor $\beta$ in Equation \eqref{eq:gell}. Multiple linear filter algorithms are applied to the raw SNR trace.
For low SNR regimes, we recognize again the better performance of the first-order filter, with respect to the capabilities of the moving-average filter. However, it is interesting to note that, after a certain threshold ($\gamma_t \geq 24$ dB), the moving-average is an even worse estimate than the noisy SNR trace $\hat{\gamma}(t)$ where no filters or further digital signal processing have been applied.
Moreover, in the same high-SNR range, the trend of $\hat{\gamma}(t)$ almost overlaps with the performance of the first-order filter in Figure \ref{fig:error_vs_gamma}. This agrees with the result of Subsection \ref{sec:res_filter} where we stated that, when simulating a $50^{\text{th}}$ percentile user, the AWGN noise does not significantly affect the estimated raw SNR trace $\hat{\gamma}(t)$, which therefore faithfully tracks the actual SNR evolution.
\begin{figure}[t!]
\centering
\includegraphics[trim= 0cm 0cm 0cm 0cm , clip=true, width=1 \columnwidth]{beta.pdf}
\caption{Average estimation error ${e(t) =\mathbb{E}\left[|\bar{\gamma}(t) -\gamma(t)|\right]}$ vs. target SNR, for different linear filter configurations.}
\label{fig:error_vs_gamma}
\end{figure}
\section{Conclusions and Future Work}
\label{sec:concl}
A key concern for the feasibility of mmWave system is the
rapid channel dynamics. Two broad questions need understanding:
how fast do channels actually change and how can systems be designed
to deal with these variations. This paper has attempted to develop some fundamental understanding in
the context of one particularly important problem -- namely
the tracking of SNR. We have considered a simplified
procedure to estimate the SNR that can be readily implemented
in next generation systems using synchronization signals.
These signals will be necessary for initial access and thus will not
introduce further overhead. Simple estimates for the SNR for these
were derived. The methods were then evaluated in a novel
semi-statistical model, where the spatial characteristics were derived from an
existing statistical model based on
outdoor measurements and the local blockage was derived from new
experimental measurements.
Our high level finding is that the SNR can be mostly tracked within
a few dB of error, even when the measurements
are in very low SNR. Nevertheless, using very simple filtering mechanisms,
the SNR tracking does incur some delay, particularly during periods of
very rapid changes.
Further work is still needed. Most directly, it is useful to test nonlinear
and/or adaptive mechanisms that could track this SNR more effectively.
Also, this SNR tracking can then be used to assess the effects on other
higher layer functions including rate prediction, handover and
radio link failure detection.
\bibliographystyle{IEEEtran}
|
\section*{Introduction }
\medskip
Milnor's fibration theorem for holomorphic maps is a key-stone in singularity theory; it says that for a given holomorphic map-germ
$f : (\mathbb{C}^n,0)\rightarrow (\mathbb{C},0)$ with an isolated critical point at the origin, the map $\frac{f}{||f||}$ is the projection map of a locally trivial bundle
$$\frac{f}{||f||}: \mathbb{S}^{2n-1}_{\epsilon} \setminus f^{-1}(0) \rightarrow \mathbb{S}^1$$
for sufficiently small sphere $\mathbb{S}^{2n-1}_{\epsilon}$ centered at the origin.
Extensions of this result in different directions became an important theme of study. In particular the extension to the case of real analytic maps has been considered by Milnor himself in \cite{JMilnor-hypersurfaces}. He proved that if $f : (\mathbb{R}^n,0)\rightarrow (\mathbb{R}^p,0) $ with $n\geq p\geq2$ is real analytic with an isolated critical point at the origin, then
$$\frac{f}{||f||}: \mathbb{S}^{n-1}_{\epsilon} \setminus N(f^{-1}(0)) \rightarrow \mathbb{S}^1$$
is a fibration on the complement of a tubular neighborhood of $f^{-1}(0)$ in sufficiently small spheres.
Furthermore, he proved that this fibration can be extended to $\mathbb{S}^{n-1}_{\epsilon} \setminus f^{-1}(0)$. However this extension may not be given by the natural map $\frac{f}{||f||}$. He actually gave an example where this natural map does not induce a fibration on the sphere.
Many authors started investigating conditions to ensure that a real analytic map induces a fibration on small spheres.
In this work we will be concerned with the $(c)$ and $(m)$-regularity conditions introduced by K. Bekka in \cite{KBekka-cregularite}.
M. Ruas and R.A. dos Santos, proved in \cite{MASRuas-DSantos}, that if a real analytic map $f: (\mathbb{R}^n, 0) \rightarrow (\mathbb{R}^2 , 0)$ with an isolated critical point is $(c)$-regular (in a certain sense they define), then $\frac{f}{||f||}: \mathbb{S}^{n-1}_{\epsilon} \setminus f^{-1}(0) \rightarrow \mathbb{S}^1$ is a fibration. However they noticed that $(c)$ condition is too strong.
We prove here that for a real analytic map $f: (R^n, 0) \rightarrow (\mathbb{R}^p, 0)$ with an isolated critical point and $p\leq n$, the map $\frac{f}{||f||}$ is a fibration on small spheres if and only if the map $f$ satisfies the weaker regularity condition $(m)$.
We also prove that, for surjective real analytic map $f: (\mathbb{R}^n, 0) \rightarrow (\mathbb{R}^p, 0)$ with an isolated critical value, $(m)$-regularity condition (in a certain sense we precise here) ensures not only that $\frac{f}{||f||}$ is a fibration on small spheres but also that $f$ induces a fibration on the tube
$$f: \mathbb{B}_{\epsilon}\cap f^{-1}(\partial \mathbb{D}_{\delta}\setminus \{0\}) \rightarrow \partial \mathbb{D}_{\delta}\setminus \{0\}$$
where $\delta$ is sufficiently small with respect to a sufficiently small $\epsilon$; $\mathbb{B}_{\epsilon}$ and $\mathbb{D}_{\delta}$ being respectively balls in $\mathbb{R}^n$ and $\mathbb{R}^p$. In this case these two fibrations will be equivalent.
Actually what we prove here is that $(m)$-regularity condition with control function {\it distance at the origin} is equivalent to $(d)$-regularity, and $(m)$-regularity with control function {\it distance to the special fiber} is equivalent to the transversality of fibers, near the special fiber, with sufficiently small spheres. In other words, we show that the arguments behind $(m)$-regularity that ensure Milnor fibrations are the ones introduced in \cite{Cisneros-Seade-Snoussi-IJM}, namely $(d)$-regularity and transversality of fibers near the special one with small spheres.
\medskip
\section{A view of some classical regularity conditions}
Many of the regularity criteria we use in this work, include some stratification condition.
Let us state, very briefly, the main stratification concept we use.
Consider an analytic or semi-analytic subspace $X\subset U \subset \mathbb{R}^n$, where $U$ is an open set of $\mathbb{R}^n$. A semi-analytic stratification of $X$ (resp. $U$), is a locally finite partition of $X$ (resp. $U$) into locally closed, connected, non singular, semi-analytic spaces $X_\alpha$ (resp. $U_\alpha$). Such a stratification of $U$ is adapted to $X$ if $X$ is a union of strata.
A stratification $X = \cup X_\alpha$ satisfies the frontier condition when for each indices $\alpha$ and $\beta$ we have ${\overline \mathcal X_\alpha} \cap X_\beta \neq \emptyset \Rightarrow X_\beta \subset {\overline X_\alpha}$.
\vglue .5cm
{\bf Whitney's conditons $a$ and $b$:}
For a locally finite stratification satisfying the frontier condition, H. Whitney introduced the following two conditions:
\begin{definition}
Whitney $(a)$-condition: A pair $(X_\alpha,X_\beta)$ with $X_\beta \subset {\overline X_\alpha}$ is $(a)$-regular at $y \in X_\beta$ if, whenever $\{x_i\} \in X_{\alpha}$ is a sequence converging to $y$ such that the sequence of tangent spaces
$\{T_{x_i} X_{\alpha}\} $ has a limit $T$, then $T_y S_{\beta} \subset T.$\\
\medskip
Whitney $(b)$-condition: The pair $(X_\alpha,X_\beta)$, with $X_\beta \subset {\overline X_\alpha}$, is $(b)$-regular at $y\in X_\beta$, if whenever we have sequences of points $\{x_i\} \in X_{\alpha}$ and $\{y_i\} \in X_{\beta}$ both converging to $y$ and such that $\{T_{x_i} X_{\alpha}\} $ has a limit $T$, and the sequence of lines $\overline{x_i y_i}$ has a limit $\ell$, then
$\ell \subset T.$
\end{definition}
It is well known that condition (b) implies condition (a), see for example \cite{DTrotman-ENS}.
A stratification is called a {\it Whitney stratification}, or Whitney regular, if it satisfies condition (b)
for all pairs of strata at every point of the small stratum.
Given an arbitrary $(U,X)$ as above, there exist Whitney stratifications of $U$ adapted to $X$. This was proved by Whiney in \cite {HWhitney-tangents} in the complex analytic setting, and by Hironaka in \cite{HHironaka-subanalytic-sets} in general.
An important consequence of Whitney conditions is the topological triviality. In fact, If $(X_\beta ,X_\alpha)$ are two regular strata with $X_\alpha \subset {\overline X_\beta}$, then $X_\beta$ is homeomorphic to a product of $X_\alpha$ with a section of $X_\beta$ of suitable dimension.
We now let $f : X \rightarrow Y$ be an analytic map
between analytic spaces.
The map $f $ is said to be stratified if there exist Whitney stratifications $\Sigma =\{S_{\alpha}\}$
and $\sigma = \{s_{\beta}\}$ of $X$ and $Y$ respectively, such that $f$ induces an analytic submersion on each stratum of $\Sigma$ into a stratum of $\sigma$.
Thom's first isotopy lemma says that a proper smooth stratified map is the projection of a smooth fiber bundle (see for instance \cite{JMather-notes}). This generalizes to the case of singular varieties, the classical fibration theorem of Ehresman.
\medskip
\vglue .5cm
{\bf Thom's} $a_f$-{\bf condition:}
Assume now that $f : (U, 0)\subset (\mathbb{R}^n , 0) \rightarrow (\mathbb{R}^p,0)$, $ n>p$, is a real analytic map. A point $x \in U$ is critical for $f$ when the derivative of $f$ at $x$ has rank less than $p$. A point $y\in \mathbb{R}^p$ is a critical value of $f$ when it is the image of a critical point.
\begin{definition}
Let $f: X \rightarrow Y$ be a stratified real analytic map between real analytic Whitney stratified spaces $X = \cup X_\alpha$ and $Y = \cup Y_\beta$.
Let $x\in X_{\alpha_1} \subset {\overline X_{\alpha_2}}$.
The map $f$ satisfies {\it Thom $a_f$-condition} at $x$, if for every sequence of points $x_n \in X_{\alpha_2}$ converging to $x$, for which the sequence of tangent spaces
${T_{x_n}(f^{-1}(f(x_n))\cap X_{\alpha_2})}$ converges to a space $T$, we have $T_x(f^{-1}(f(x))\cap X_{\alpha_1}) \subset T$.
The map $f$ is said to satisfy {\it Thom} $a_f$-condition, if it satisfies it at each point $x\in X_\alpha$ for any pair of strata $X_\alpha \subset {\overline X_\beta}$.
We say that a real analytic map $f: X \rightarrow Y$ has {\it Thom property} when there exist Whitney stratifications of $X$ and $Y$, for which $f$ satisfies Thom's $a_f$-condition.
\end{definition}
In particular, let $f: (U, 0) \subset (\mathbb{R}^n, 0) \rightarrow (\mathbb{R}^p, 0)$ be a real analytic map with an isolated critical value at the origin. Let $V=f^{-1}(0)$ and consider a Whitney stratification $V = \cup V_\alpha$.
The stratification $(U\setminus V) {\displaystyle \bigcup _\alpha} V_\alpha$ is a Whitnay regular stratification of $U$ adapted to $V$. In order to check Thom's $a_f$-property for $f$ with respect to that stratification, it is enough to check it for the pairs of strata of the form $(U\setminus V, V_\alpha)$.
We know by the conical structure theorem (\cite[lemma 3.2]{Burghalea}) that for a sufficiently small positive integer $\epsilon$, the sphere $\mathbb{S}_{\epsilon}$ centered at the origin with radius $\epsilon$ is transversal to every stratum of a Whitney regular stratification of $V$. So Thom's $a_f$ condition ensures the following property:
\begin{proposition}\label{thom-transversal}
Let $f: (U,0) \subset (\mathbb{R}^n, 0) \rightarrow (\mathbb{R}^p, 0)$ be an analytic map-germ as above with Thom property.
There exists $\epsilon_0 >0$ such that for every $0<\epsilon \leq \epsilon_0$, there exists $\delta >0$ such that for every $t\in \mathbb{R}^p$ with $\Vert t \Vert \leq \delta$, the fiber $f^{-1}(t)$ intersects the sphere $\mathbb{S}_\epsilon$ transversally.
\end{proposition}
We summarize this property by saying that in a sufficiently small ball, every sphere is transversal to fibers sufficiently close to the special fiber.
\vglue .5cm
{\bf Bekka's $c$-regularity condition:}
Consider a smooth analytic space $M$ and a pair of strata $(X,Y)$ with $Y \subset \overline X$.
Let $\rho : M \rightarrow \mathbb{R}_+$ be a non-negative smooth function with $\rho^{-1}(0) = Y$. We say that $\rho$ is a control function for the pair in $M$ with respect to $Y$.
\begin{definition}[(c)-regularity]
The pair $(X,Y)$ is $(c)$-regular at $y \in Y$ with respect to the control
function $\rho$ if, for any sequence of points in $X,$ $\{x_i\}\rightarrow y$ such that the sequence of planes $\{ ker (d\rho(x_i))\bigcap T_{{x_i}}X\}$
converges to a plane $T$ in the Grassmann space of $(dimX-1)$-planes, then $T_{y} Y \subset T.$
The pair $(X,Y)$ is $(c)$-regular with respect to the control function
$\rho$ if it is $(c)$-regular for every point $y \in Y$ with respect to the function $\rho$.
\end{definition}
The $c$-regularity condition for a pair of strata $(X,Y)$ with respect to a control function $\rho$ is equivalent to Thom's $a_{\rho}$ condition for these strata.
Bekka proved in \cite[theorem 1]{KBekka-cregularite}, that $(c)$-regularity condition ensures topological triviality. This makes it somehow a weaker condition than $(b)$-condition for getting topological triviality. This is the main reason for using it in \cite{MASRuas-DSantos}.
\vglue .5cm
{\bf The $m$-regularity condition:}
Assume that the manifold $M$ has a Riemannian metric. Consider a pair of strata $(X,Y)$ with $Y\subset {\overline X}$.
Let $(T_Y,\pi,\rho)$ be a tubular neighborhood of $Y$ in $M$ together with a projection $\pi : T_Y \rightarrow Y $, associated to a smooth non-negative control function $\rho$ such that $\rho^{-1} (0)=Y$ and $\triangledown \rho (x) \in ker (d\pi(x))$.
\begin{definition}[(m)-regularity]\label{defmreg}
The pair of strata $(X,Y)$ satisfies \textbf{condition $(m)$ }if there exists a positive real number $\epsilon > 0$ such that
$$(\pi,\rho)\mid_{X\cap T_Y^{\epsilon}} : X\cap T_Y^{\epsilon} \rightarrow Y\times [0,\epsilon) $$
$$x\longmapsto (\pi(x),\rho(x))$$ is a submersion, where $T_Y^{\epsilon}:= \{x \in T_Y, \rho (x) \textless \epsilon\}.$
When this happens we say that $(X,Y)$ is (m)-regular with respect to the control function $\rho$.
\end{definition}
The following can be found in \cite{KBekka-cregularite}.
\begin{proposition}
If a pair $(X,Y)$ is (c)-regular with respect to a control function $\rho$, then this pair is Whitney (a) regular and (m)-regular with respect to $\rho$.
\end{proposition}
If $(X,Y)$ is (a) regular and also (m)-regular with respect to a control function $\rho$, it is not yet known whether this implies $(X,Y)$ is (c)-regular.
\vglue .5cm
{\bf The (d)-regularity for map-germs:}
Consider an analytic map $f: (\mathbb{R}^n,0) \rightarrow (\mathbb{R}^p,0)$.
Following \cite{JSeade-BSMM, MRuas-JSeade-AVerjovsky, MASRuas-DSantos, Cisneros-Seade-Snoussi-IJM},
we associate to $f$ a family of varieties: For each line $\mathcal{L}$
through the origin in $\mathbb{R}^p$ we define:
$$X_{\mathcal{L}}:= \{x\in \mathbb{R}^n \mid f(x)\in\mathcal{L} \}.$$
\begin{definition}
The family of analytic spaces $X_{\mathcal{L}}$, as $\mathcal{L}$ varies in $\mathbb{R} P^{p-1}$, is called the {\it canonical pencil} associated to $f.$
\end{definition}
This pencil was first introduced in \cite{JSeade-BSMM} for a certain family of maps into $\mathbb{R}^2$, to show that these have Milnor open-book fibrations. This was later used in \cite{MASRuas-DSantos} in relation with (c)-regularity (see Theorem \ref{Ruas-Santos} below).
The union of all $X_{\mathcal{L}}$ is the whole $\mathbb{R}^n$ and their intersection is $V$. It is in this sense that this family forms a pencil with axis $V$.
In particular,
when $f$ has an isolated critical value, these analytic varieties are all smooth
away from $V$.
The following concept was introduced in \cite[definition 5.4]{Cisneros-Seade-Snoussi-Advances} and \cite[definition 2.4]{Cisneros-Seade-Snoussi-IJM}.
\begin{definition}
Let $f: U \rightarrow \mathbb{R}^p$ be a locally surjective real analytic map with isolated critical value at $0.$ We say that $f$ is \textbf{$d$-regular} if there
exists $\epsilon_0 > 0$ such that for any $\epsilon\leq \epsilon_0$ and for any line $0 \in\mathcal{L}\subset \mathbb{R}^p $ the sphere $S_{\epsilon}^{n-1}$
and the space $X_{\mathcal{L}}$ are transverse.
\end{definition}
In other words, $d-$regularity means transversality of sufficiently small spheres with every element of the canonical pencil.
The following important characterization of d-regularity comes from \cite[proposition 3.2]{Cisneros-Seade-Snoussi-IJM}:
\begin{proposition} The following statements are equivalent:
\begin{itemize}
\item The map-germ $f$ is d-regular.
\item The map $\phi =
\frac{f}{\norm{f}}\colon \mathbb{S}_{\eta}^{n-1} \setminus K_{\eta}
\longrightarrow\mathbb{S}^{p-1}$
is a submersion for every sufficiently small sphere
$\mathbb{S}_{\eta}^{n-1}$.\label{it:phi-sub}
\end{itemize}
\end{proposition}
\section{Milnor fibrations}
Consider a real analytic map $f: \mathbb{R}^n \rightarrow \mathbb{R}^p$, $n\geq p$, with an isolated critical value. The critical points of $f$ are then concentrated in the fiber $f^{-1}(0)$. a natural question is to ask to what extend does $f$ induce a fibration outside $f^{-1}(0)$.
Following J. Milnor's pioneer work in \cite{JMilnor-hypersurfaces}, one looks for two posible types of fibrations:
{\bf Fibration in the tube:} Intuitively, we refer to the situation where the map $f$ induces a fibration in the space made of nearby fibers to $f^{-1}(0)$ still in a small neighborhood of the origin. More precisely:
\begin{definition}
We say that $f: (\mathbb{R}^n, 0) \rightarrow (\mathbb{R}^p, 0)$ induces a fibration in the tube if there exists $\epsilon_0 >0$ such that for every $0<\epsilon \leq \epsilon_0$ there exists $0<\delta$ such that the restriction map:
$$f: \mathbb{B}_{\epsilon} \cap f^{-1}(\mathbb{B}^{p}_{\delta}\setminus \{0\}) \rightarrow \mathbb{B}^{p}_{\delta} \setminus \{0\}$$ is a fibration; where $\mathbb{B}_{\epsilon}$ (resp. $\mathbb{B}^{p}_{\delta}$) is the open ball of $\mathbb{R}^n$ (resp. of $\mathbb{R}^p$) centered at the origin with radius $\epsilon$ (resp. $\delta$).
\end{definition}
When $f$ has an isolated critical point, Milnor showed in \cite{JMilnor-hypersurfaces}, that one always has a fibration in a tube. An other classical case is when $p=2$, $n$ is even and $f$ is holomorphic. In this situation, we have an isolated critical value and D.T. L\^e proved in \cite{DTLe-remarks}, that such an $f$ induces a fibration in the tube.
In the general setting, J.L. Cisneros-Molina, together with the first and third authors, proved in \cite[Proposition 5.1 and Remark 5.7]{Cisneros-Seade-Snoussi-IJM}, the following:
\begin{theorem}\label{tube}
Let $f: (\mathbb{R}^n, 0) \rightarrow (\mathbb{R}^p, 0)$, $n \geq p$, be a locally surjective real analytic map with isolated critical value.
If there exists $\epsilon_0>0$ such that for every $0<\epsilon \leq \epsilon_0$ there exists $\delta >0$ such that for any $y\in \mathbb{R}^p$ with $0<\Arrowvert y\Arrowvert <\delta$ one has $f^{-1}(y)$ is transversal to $\mathbb{S}_{\epsilon}$, then the map $f$ induces a Milnor fibration in a tube; where $\mathbb{S}_{\epsilon}\subset \mathbb{R}^n$ is the sphere centered at the origin with radius $\epsilon$.
\end{theorem}
In particular, one obtains from proposition \ref{thom-transversal}, that a locally surjective map germ with isolated critical value that satisfies Thom property induces a fibration in a tube.
{\bf Fibration in the sphere:} This comes from the holomorphic case of maps from $\mathbb{C}^n \rightarrow \mathbb{C}$, and it refers to the fact that the argument of $f$ restricted to a sufficiently small sphere of $\mathbb{R}^{2n}$ is a fibration. More precisely and in a more general setting we mean:
\begin{definition}
Let $f: (\mathbb{R}^n, 0) \rightarrow (\mathbb{R}^p,0)$ be a locally surjective real analytic map with an isolated critical value, with $n\geq p$. We say that $f$ induces a fibration in the sphere if there exists $\epsilon_0 >0$ such that for any $0<\epsilon \leq \epsilon_0$ the map
$$\frac{f}{\Arrowvert f\Arrowvert}: \mathbb{S}_{\epsilon} \setminus f^{-1}(0) \rightarrow \mathbb{S}^{p-1}$$
is a fibration.
\end{definition}
For holomorphic functions with isolated critical points, J. Milnor proved that they always induce a fibration in a sphere. This was extended to Holomorphic functions on a singular variety $X \rightarrow \mathbb{C}$ with isolated critical value in \cite{Cisneros-Seade-Snoussi-Advances}.
Many authors got interested in the general case of real analytic maps, see for example \cite{Jaquemard-Milnor, Durfee-Milnor, JSeade-BSMM, Tibar-etAl1, Tibar-etAl2}.
In particular, Ruas and Dos Santos, in \cite{MASRuas-DSantos}, related $(c)$-regularity condition to the fibration in the sphere in a particular case; this will be discussed in next section.
In \cite[Theorem 5.3 and Remark 5.7]{Cisneros-Seade-Snoussi-IJM}, the following result is achieved:
\begin{theorem}\label{CSS-int-j}
Let $f: (\mathbb{R}^n,0) \rightarrow (\mathbb{R}^p, 0)$ be a locally surjective analytic map with an isolated critical value with $n\geq p$. Assume $f$ is $d$-regular and there exists $\epsilon_0>0$ such that for every $0<\epsilon \leq \epsilon_0$ there exists $\delta>0$ such that for any $y\in \mathbb{R}^p$ with $\Arrowvert y \Arrowvert < \delta$ the fiber $f^{-1}(y)$ is transversal to the sphere $\mathbb{S}_{\epsilon}$.
Then the map $f$ induces a fibration in the sphere and a fibration in the tube, and both are equivalent.
\end{theorem}
In the particular case of a map with an isolated critical point, since the transversality of the nearby fibers with the spheres is always achieved, theorem \ref{CSS-int-j} can be reduced to the following:
\begin{corollary}
Let $f: (\mathbb{R}^n,0) \rightarrow (\mathbb{R}^p,0)$ be an analytic map with an isolated critical point, $n\geq p$. It induces a fibration in the sphere if and only if it is $d$-regular.
\end{corollary}
\section{ Milnor fibrations and c-regularity}
In this section we explain briefly the main result in \cite{MASRuas-DSantos}
relating $c$-regularity to Milnor fibrations for real
analytic map-germs. The statement and proof in \cite{MASRuas-DSantos} use something similar to what we call the canonical pencil.
The situation considered in \cite{MASRuas-DSantos} is the one of a germ of real analytic map from $\mathbb{R}^n$ to $\mathbb{R}^2$ with an isolated critical point. Since our goal is to make their statement more precise and to extend it to a more general situation, we will start in the following setting.
Consider an analytic map
$$\begin{array}{rcl}
f: (\mathbb{R}^n, 0) & \rightarrow & (\mathbb{R}^p,0) \\
x=(x_1, \cdots, x_n) & \mapsto & (f_1(x), \cdots, f_n(x))
\end{array}$$
with an isolated critical value, $n\geq p$.
Call $V:= f^{-1}(0)$ the special fiber of $f$. The critical points of $f$ are then all contained in $V$.
Consider the map
$$\begin{array}{rcl}
\phi : \mathbb{R}^n \setminus V & \rightarrow & \mathbb{R}\mathbb{P}^{p-1}\\
x & \mapsto & (f_1(x) : \cdots : f_p(x))
\end{array}$$
Note that the fibers of the map $\phi$ together with the subspace $V$ are the elements of the canonical pencil associated to $f$.
We define the blow-up of $V$ in $\mathbb{R}^n$, or more precisely, of the ideal $(f_1, \cdots , f_p)$ as follows:
Consider the subspace $X \subset \mathbb{R}^n \times \mathbb{R}\mathbb{P}^{p-1}$ defined by the equations
$$f_it_j - f_jt_i =0 , 0<i<j\leq p$$
where $(t_1: \cdots : t_p)$ is a system of homogeneous coordinates in $\mathbb{R}\mathbb{P}^{p-1}$.
The restriction of the first projection to $X$ induces an analytic map
$$e: X \rightarrow \mathbb{R}^n$$
that we will call the blow-up of $V$ in $\mathbb{R}^n$.
The map $e$ induces an isomorphism $X \setminus e^{-1}(V) \cong \mathbb{R}^n\setminus V$.
The inverse image of $V$ by $e$ is $e^{-1}(V) = V \times \mathbb{R}\mathbb{P}^{p-1}$.
The second projection of $\mathbb{R}^n \times \mathbb{R}\mathbb{P}^{p-1}$ induces a map
$$\psi : X \rightarrow \mathbb{R}\mathbb{P}^{p-1}$$
that extends $\phi$ in the sense that $\phi \circ e = \psi$.
We want to consider a stratification of $X$ compatible with $Y:= V\times \mathbb{R}\mathbb{P}^{p-1}$.
Consider a decomposition $V = {\displaystyle \cup _{\alpha}} V_{\alpha}$ into semi-analytic smooth connected varieties all adherent to $0$. The corresponding stratification of $Y$ is given by $Y = {\displaystyle \cup _ {\alpha}}Y_{\alpha}$, where $Y_{\alpha} = V_{\alpha} \times \mathbb{R}\mathbb{P}^{p-1}.$
Similarly, consider a decomposition $X\setminus Y = {\displaystyle \cup _{\beta}} X_{\beta}$ into semi-analytic smooth and connected varieties all adherent to $0$. Note that $X\setminus Y$ being smooth, this decomposition can be taken as the decomposition of $X \setminus Y$ into connected components.
We may choose these decompositions minimal with that property.
Notice that some strata $V_{\alpha}$ may not be contained in the closure of $X$. The reason is that the blow-up of $V$ may not coincide with the closure of the graph of the map $\phi$ in $\mathbb{R}^n \times \mathbb{R}\mathbb{P}^{p-1}$.
\begin{example}
Consider the map
$$\begin{array}{rcl}
f: \mathbb{R}^3 & \rightarrow & \mathbb{R}^2\\
(x, y, z) & \mapsto & (xy, xz)
\end{array}$$
$V = (x=0) \cup (y=z=0)$. The component defined by the hyperplane $(x=0) \times \mathbb{R}\mathbb{P}^1$ in not in the closure of $X$ which has the same dimension.
\end{example}
By abuse of language, and in order to simplify the statements, in what follows we will call a pair of strata $(X, Y)$ any pair of strata $(X_{\beta}, Y_{\alpha})$ with the frontier condition: $Y_{\alpha} \subset {\overline X_{\beta}}$.
Let us now go toward $(c)$-regularity.
define the control function
$$\begin{array}{rcl}
\rho_0 : X & \rightarrow & \mathbb{R}_+\\
(x_1, \cdots , x_n, t_1 : \cdots : t_p) & \mapsto & {\displaystyle \sum_{i=1}^{n}} x_i^2
\end{array}$$
The inverse image $\rho^{-1}(0) = \{0\} \times \mathbb{R}\mathbb{P}^{p-1}$. We can admit it as a stratum in $Y$. Let us call it $Y_0$. We have $Y_0 \subset {\overline X_\beta}$, for any $\beta$. We can then consider them as admissible strata.
\begin{theorem}\label{Ruas-Santos}\cite[Theorem 3]{MASRuas-DSantos}
Assume $f$ has an isolated critical point at $0$ and suppose $p=2$. If the above pair $(X_{\alpha}\setminus Y_0,Y_0)$ is $(c)$-regular with respect to the control function $\rho_0$ (distance to the origin), then the map $f$ induces a Milnor fibration on the sphere.
\end{theorem}
In the same work, M. Ruas and R.A. dos Santos gave an example showing that $(c)$-regularity in this context is strictly stronger than Milnor fibration on the sphere. More precisely, their example consists in a map for which the considered strata do not satisfy Whitney's condition $(a)$ and however the map induces a fibration in the sphere. In the following section we will clarify this situation.
\section{Milnor fibrations and m-regularity}
Let $f= (f_1, \cdots , f_p) : \mathbb{R}^n \rightarrow \mathbb{R}^p$ be a real analytic map with isolated critical value. Define $V$, $X$, $Y$, $Y_0$ and $\rho_0$ as in the previous section.
\begin{definition}
We say that the map $f$ is $m_0$-regular if the strata $(X_{\alpha}\setminus Y_0, Y_0)$ are $m$-regular with respect to the control function $\rho_0$.
\end{definition}
In this case the contraction $\pi$ is the natural map $\mathbb{R}^n\times \mathbb{R}\mathbb{P}^{p-1} \rightarrow \{0\} \times \mathbb{R}\mathbb{P}^{p-1}$.
\begin{lemma}\label{m0ssid}
The map $f$ is $m_0$-regular if and only if it is $(d)$-regular.
\end{lemma}
\begin{proof}
Recall that $Y_0:= \{0\}\times \mathbb{R}\mathbb{P}^{p-1}$. A tubular neighborhood of $Y_0$ in the ambient space $\mathbb{R}^n \times \mathbb{R}\mathbb{P}^{p-1}$ is of the form $T^{\epsilon}_{Y_0} := B^n_{(0,\epsilon)}\times \mathbb{R}\mathbb{P}^{p-1}$ where $B^n_{(0,\epsilon)}$ is the open ball of $\mathbb{R}^n$ centered at the origin with radius $\epsilon$.
Consider the map
$$\begin{array}{rcl}
(\pi_0, \rho_0): T^{\epsilon}_{Y_0} & \rightarrow & Y_0 \times [0, \epsilon)\\
(x_1, \cdots , x_n, t_1: \cdots : t_p) & \mapsto & ((0, t_1 : \cdots : t_p), \rho_0(x_1, \cdots , x_n))
\end{array}$$
Following definition \ref{defmreg}, $m_0$-regularity means that the restriction of $(\pi_0, \rho _0)$ to $X\cap T_{Y_0}^{\epsilon}$ is a submersion.
Since the map $(\pi_0, \rho_0)$ is a submersion, $m_0$-regularity is equivalent to transversality of the stratum $X_{\alpha}$ with the fibers of $(\pi_0, \rho _0)$.
Consider the projection map:
$$\psi : X \rightarrow \mathbb{R}\mathbb{P}^{p-1}$$
induced by the second factor projection on $\mathbb{R}^n\times \mathbb{R}\mathbb{P}^{p-1}$. The fibers of the map $\psi$ are precisely the elements $X_{\mathcal L}$ of the canonical pencil associated to $f$, expanded along the projective space $\mathbb{R}\mathbb{P}^{p-1}$.
Let $(0, \xi, \delta) \in \{0\}\times \mathbb{R}\mathbb{P}^{p-1}\times [0, \epsilon)$. The fiber $(\pi_0, \rho_0)^{-1}(0, \xi, \delta)$ is $\mathbb{S}_{\delta} \times \{\xi\}$. This fiber is transversal to $X_{\alpha}$ at a point of the form $(x,\xi)\in\mathbb{R}^n\times \mathbb{R}\mathbb{P}^{p-1}$ if and only if the fiber $\psi^{-1}(\xi)$ is transversal to $\mathbb{S}_{\delta}\times \{\xi\}$; Equivalently, the sphere $\mathbb{S}_{\delta}$ is transversal to the elements of the canonical pencil. This last statement is $d$-regularity condition.
\end{proof}
In \cite[Theorem 3]{MASRuas-DSantos}, it is proved that, for maps with isolated critical points, $(c)$-regularity implies Milnor Fibration in the sphere. We know that $(c)$-regularity implies $(a)$ and $(m)$-regularity. We also know from \cite{Cisneros-Seade-Snoussi-IJM} that for such maps, Milnor fibration in the sphere is equivalent to $d$-regularity. Therefore, Lemma \ref{m0ssid} makes Ruas and dos Santos's result more precise:
\begin{corollary}
Let $f: (\mathbb{R}^n,0) \rightarrow (\mathbb{R}^p,0)$ be analytic with an isolated critical point. The map $f$ induces Milnor Fibration in the sphere if and only if it is $m_0$-regular.
\end{corollary}
Let us now explore the isolated critical value situation.
Let $f$ be as in the beginning of this section. Recall that $V= f^{-1}(0)$ and $Y= V\times \mathbb{R}\mathbb{P}^{p-1}$.
$V$ may have many irreducible components, each of them can be expressed as a union of strata. since all the strata are adherent to the origin, any tubular neighborhood of any of these strata will intersect all the other strata. In the definition of $m$-regularity, we need a control function whose zero set inside a tubular neighborhood of a stratum $Y_{\alpha}$ is precisely $Y_{\alpha}$. If we keep this definition we will not be able to use it in the case where $V$ has more than one irreducible component. That is why we need to modify slightly the definition of $m$-regularity, making it point-wise.
\begin{definition}\label{mlocal}
Let $(X_\beta, Y_\alpha)$ be a pair of strata with $Y_\alpha \subset Y$, $X_\beta \subset X\setminus Y$ and $Y_\alpha \subset {\bar X}_\beta$. Let $T_\alpha$ be a tubular neighborhood of $Y_\alpha$ and $\pi: T_\alpha \rightarrow Y_\alpha $ a ${\mathcal C}^1$ retraction. Consider a non-negative function $\rho:
T_\alpha \rightarrow \mathbb{R}_+$.
Let $y_0 \in Y\alpha$.
We say that the pair $(X_\beta, Y_\alpha)$ is $m$-regular at $y_0$ if there exist a neighborhood $U$ of $y_0$ in the ambient space, and positive number $\delta_0$ such that for any $\delta \leq \delta_0$ we have:
- $\rho^{-1}(0)\cap U = Y_\alpha\cap U$
- and the restriction of $(\pi , \rho) : T_\alpha^{\delta} \rightarrow Y_\alpha \times [0, \delta)$ to $X_\beta \cap T_{\alpha}^{\delta}$ is a submersion
where $T_\alpha^{\delta} = T_\alpha \cap \rho^{-1}([0, \delta))$.
We will say that the pair $(X_\beta, Y_\alpha)$ is $m$-regular if it is $m$-regular at every point of $Y_\alpha$.
\end{definition}
Note that if a pair of strata is $m$-regular in the sense of definition \ref{defmreg}, then it is still $m$-regular in the sense of definition \ref{mlocal}. When the small stratum is just a point, they obviously coincide.
The difference between these two definitions appears clearly in the case of different strata adherent to the same one, and all of them in $\rho^{-1}(0)$.
From now on, when we refer to $m$-regularity, we mean definition \ref{mlocal}
Let us consider a stratum $V_{\alpha}$ of $V$ and call $Y_{\alpha} = V_{\alpha}\times \mathbb{R}\mathbb{P}^{p-1}$. We will say that $Y_{\alpha}$ is a maximal stratum of $Y$ if it is not contained in the closure of any other stratum $V_{\beta}\times \mathbb{R}\mathbb{P}^{p-1}$.
Let $T_{\alpha}$ be a tubular neighborhood of $Y_{\alpha}$ in $\mathbb{R}^n\times \mathbb{R}\mathbb{P}^{p-1}$. Let $\pi : T_{\alpha} \rightarrow Y_{\alpha}$ be a ${\mathcal C}^1$ retraction and
$$\rho: T_{\alpha} \rightarrow \mathbb{R}_+$$
$$\rho(x_1, \cdots , x_n, t_1: \cdots : t_p) = {\displaystyle \sum_{i=1}^p} f_i^2(x_1, \cdots , x_n)$$
a control function such that $\nabla \rho_p \in T_p{\rm Ker}\pi$ at any point $p$. Recall that $X$ is the real blow-up of $V$ in $\mathbb{R}^n$ and $(X_{\alpha})$ is a stratification of $X\setminus Y$.
\begin{theorem}\label{mreg-to-tube}
Let $f: (\mathbb{R}^n,0) \rightarrow (\mathbb{R}^p,0)$ be a surjective analytic map, with isolated critical value and $n\geq p$.
Let $Y_{\alpha} = V_{\alpha}\times \mathbb{R}\mathbb{P}^{p-1}$ be any maximal stratum of $Y$ and $X_\beta$ a stratum of $X\setminus Y$ adherent to $Y_\alpha$. If the pair $(X_\beta, Y_{\alpha})$ is $m$-regular for any ${\mathcal C}^1$ retraction $\pi: T_{\alpha} \rightarrow Y_{\alpha}$ and for the previously defined control function $\rho$, then
the map $f$ induces a Milnor fibration on the tube.
\end{theorem}
\begin{remark}
In case of non maximal strata, it does not make sense to talk about $m$-regularity because, a tubular neighborhood of such a stratum will necessarily intersect an other stratum of $V\times \mathbb{R}\mathbb{P}^{p-1}$.
\end{remark}
\begin{proof}
Following theorem \ref{tube}, we are going to prove that for sufficiently small spheres, a fiber over a point with sufficiently small norm is transversal to the given sphere. We claim that the $m$-regularity statement, for different retraction maps, implies this property.
In order to prove that, let us explore the meaning of $m$-regularity on strata $X_\beta$ and $Y_{\alpha}$ of our situation.
First of all, since $Y_{\alpha}= V_{\alpha}\times \mathbb{R}\mathbb{P}^{p-1}$, a tubular neighborhood $T_{\alpha}$ in $\mathbb{R}^n \times \mathbb{R}\mathbb{P}^{p-1}$ is of the form ${\mathcal T}_{\alpha} \times \mathbb{R}\mathbb{P}^{p-1}$, where ${\mathcal T}_{\alpha}$ is a tubular neighborhood of $V_{\alpha}$ in $\mathbb{R}^n$. A retraction $\pi : T_{\alpha} \rightarrow Y_{\alpha}$ decomposes into $(\tau , Id) : {\mathcal T}_{\alpha} \times \mathbb{R}\mathbb{P}^{p-1} \rightarrow V_{\alpha} \times \mathbb{R}\mathbb{P}^{p-1}$.
Consider a positive real number $\delta_0$ such that
$$(\pi, \rho): X\cap T_{\alpha}^{\delta_0} \rightarrow Y_{\alpha} \times [0, \delta_0)$$
is a submersion near a point $y \in Y_{\alpha}$. This is equivalent to saying that the fibers of $(\pi, \rho)$, before intersecting with $X$, are transversal to $X$.
So let $y= (v,t) \in Y = V \times \mathbb{R}\mathbb{P}^{p-1}$.
Let $\delta \in [0, \delta_0)$. Recall that $\rho = \sum f_i^2$. The fiber $(\pi, \rho)^{-1}(y,\delta)$ is $(f^{-1}(\partial \mathbb{D}_{\delta^{1/2}}) \cap \tau^{-1}(v))\times \{t\}$.
This fiber is transversal to $X$ if and only if, for every $z\in \mathbb{R}^p$ with $\Arrowvert z\Arrowvert = \delta$, the fiber $f^{-1}(z)$ is transversal to $\tau^{-1}(v)$ in $\mathbb{R}^n$.
Consider a sufficiently small real number $0< \epsilon$ and $\mathbb{S}_{\epsilon}\subset \mathbb{R}^n$ the sphere centered at the origin with radius $\epsilon$. For every point in $\mathbb{S}_{\epsilon}\cap V_{\alpha}$ and for every stratum $V_{\alpha}$, we have a positive integer $\delta_0$ given by the $m$-regularity statement. Since the intersection $V\cap \mathbb{S}_{\epsilon}$ is compact and the number of strata of $V$ is finite, we can choose a value ${\delta_0}$ valid for all the intersection points.
Consider $0<\delta <\delta_0$ and $z\in \mathbb{R}^p$ with $\Arrowvert z \Arrowvert = \delta$.
Let $x\in f^{-1}(z) \cap \mathbb{S}_{\epsilon}$. We are going to prove that this last intersection is transversal.
We can choose a ${\mathcal C}^1$ retraction for which the tangent space $T_x\pi^{-1}(\pi(x))$ is contained in the tangent space $T_x\mathbb{S}_{\epsilon}$ to the sphere. The transversality property between $f^{-1}(z)$ and $\pi^{-1}(\pi(x))$, consequence of the $m$-regularity, ensures transversality between the fiber $f^{-1}(z)$ and the sphere $\mathbb{S}_{\epsilon}$ at the point $x$.
This is precisely the condition required in theorem \ref{tube}, to have a fibration in the tube.
\end{proof}
Combining Theorem \ref{mreg-to-tube}, Lemma \ref{m0ssid} with Theorem \ref{CSS-int-j}, we obtain:
\begin{theorem}
Let $f: \mathbb{R}^n \rightarrow \mathbb{R}^p$ be a locally surjective real analytic map with an isolated critical value.
If $f$ is $m_0$-regular and the pair $(X_\beta, Y_{\alpha})$ is $m$-regular for any ${\mathcal C}^1$ retraction $\pi: T_{\alpha} \rightarrow Y_{\alpha}$, for any maximal stratum $Y_{\alpha}$ of $V= f^{-1}(0)$, any adherent stratum $X_\beta$ of $X\setminus Y$, and for the control function $\rho = {\displaystyle \sum_i} f_i^2$, then the map $f$ induces a Milnor fibration on the tube and on the sphere; and both fibrations are equivalent.
\end{theorem}
\medskip
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]{%
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
|
\section{Introduction}
de Sitter space plays a central role in modern theoretical physics. It
is not only relevant for the description of the late time
cosmology. It is also widely believed that the universe has
underwent a period of inflationary expansion described by a quasi-de
Sitter metric. As such it is of utmost importance to understand the quantum
dynamics of de Sitter space. Despite intensive research,
e.g. \cite{witten-ds, strominger, KKLT} for some of the approaches, the
problem of quantum gravity in de Sitter space remains open.
In fact even the much less ambitious problem of a quantum field theory dynamics
in de Sitter space is already quite a nontrivial problem. Although the
problem of the definition of Hamiltonian and particles for a
quantum field theory (QFT) in a generic curved spacetime
\cite{BD}
can be dealt with in perturbation theory for QFT in de Sitter space,
it is not necessary so in the strongly coupled regime, in which case
we also expect new phenomena may arise. Besides, the understanding of the
long time secular effects in de Sitter space is another important
problem that called for a treatment beyond the usual perturbation
scheme \cite{starobinsky,dRG,schwinger-dyson}.
It has been speculated that
\cite{poly,ford}
the infrared (IR) quantum effects in dS space could provide a screening
effect on the cosmological constant and offers a solution to the cosmological
constant problem. One of the motivations of this work is the desire
for a better understanding of the dynamics of quantum field theories
in de Sitter spacetime beyond perturbation theory.
A powerful tool in this regard is the AdS/CFT correspondence
\cite{mal}.
The holographic nature of quantum gravity was originally
suggested by the discovery of the thermodynamic
nature of the black hole mechanics, most notably the Bekenstein-Hawking
formula for black hole entropy \cite{ber,hawking}.
The AdS/CFT correspondence
provides the first explicit example of how holography
can be realized in string theory
\cite{mal,witten,GKP}. Over the years, it has also been used as an
important tool to learn about the physical properties of various
quantum field theories in the strongly coupled
nonperturbative regime. See, for example, the reviews
\cite{r1,r2,r3}. In some cases,
when sufficient amount of supersymmetry or integrability is
present, even exact results are possible \cite{polchinski,r4}. In the case of
de Sitter space, gauge/gravity dual has been studied in
\cite{d1}-\cite{d24}.
In particular,
evidence of dynamical phase transition for confining gauge theory on
de Sitter space was found in the strongly coupled regime
in \cite{d15}; entanglement of entropy for strongly coupled field theories
on de Sitter space with a gravity dual was computed
in \cite{d17} and their results
suggested that the FRW cosmologies is contained in the field theory
description. One notes that for the kind of questions addressed
in these previous works, the gauge/gravity
correspondence was only needed to be considered in the generic sense without
having to spell out in details the involved
string theory and the boundary field theory.
Nevertheless it is certainly interesting to have a concrete
duality so that one can ask precise questions of other kinds.
Another motivation of this work is to construct such a more precise
gauge/gravity correspondence in de Sitter spacetime.
The construction of global supersymmetric field theory in four dimensional
de Sitter spacetime is however impossible \cite{nogo1,nogo2}. First of all,
there does not exist Majorana Killing spinor on de Sitter spacetime,
which are necessary for the construction of supersymmetry. Moreover,
the usual de Sitter superalgebra has no unitary representation. These
no-go theorems are likely to be the reason why a holographic duality
involving a
supersymmetric field theory on de Sitter spacetime has not been
constructed. However it has been realized quite recently that the
no-go theorem can be bypassed if global superconformal symmetry is
considered instead of global supersymmetry.
In particular
the $\cN=4$ superconformal non-abelian Yang-Mills theory on dS${}_4$
has been constructed \cite{dzf}.
The existence of this theory is also anticipated from the works of
\cite{Hristov:2013spa,Cassani:2012ri}.
This superconformal Yang-Mills theory is
expected to enjoy exact $SU(2,2|4)$ supersymmetry \cite{dzf}. One
might also expect that
this superconformal theory may share some of the remarkable
properties such as integrability and S-duality
like its cousin, the $\cN=4$
supersymmetric Yang-Mills theory on Minkowski spacetime, which would
then allow
exact results to be obtained. Hence
it's gauge/gravity correspondence would be a perfect laboratory to
see how some of the nontrivial results of the AdS/CFT correspondence would
extend in the presence of a time dependent background spacetime.
The desire to study the properties of the $\cN=4$ superconformal
Yang-Mills theory on de Sitter spacetime is another motivation of this work.
As it turns out, de Sitter spacetime can be obtained as the boundary
of the AdS spacetime. In the Poincare patch of the AdS space, a
boundary of Minkowski spacetime with a conformal structure was
created \cite{witten}. However one may also choose a dS-sliced
coordinates \eq{ds1} of the AdS${}_5$ spacetime and
obtain dS${}_4$ as the boundary manifold \cite{pope}.
Based on this
observation
and the results of \cite{dzf}, we propose in this paper the AdS/dS CFT
correspondence:
Type IIB string theory on AdS${}_5 \times S^5$ with boundary
condition imposed on the boundary dS${}_4$ is dual
to the $\cN=4$ superconformal $SU(N)$ Yang-Mills theory on
dS${}_4$.
The plan of the paper is as follows.
In section 2 we review the global
definition of dS and AdS spacetime, and the
coordination of AdS which give rises to Minkowski or de Sitter
spacetime as boundaries.
We also describe the $\cN=4$ superconformal Yang-Mills theory and some
general properties of conformal field theory on de Sitter spacetime.
In section 3, we introduce the bulk to
boundary propagator formulation of the AdS/CFT correspondence \cite{witten,GKP},
generalized for a general bulk metric. In sections 4 and 5, we apply
this formulation to compute
the boundary correlators for
conformal operators of scalar and spin 1/2 types.
We show that the results obtained agree with the
de Sitter conformal field theory.
The paper is ended with further discussion in section 6.
\section{AdS/dS CFT Correspondence}
Our proposal is that
Type IIB string theory on AdS${}_5 \times S^5$ with boundary
condition specified on the boundary dS${}_4$ is dual
to the $\cN=4$ superconformal $SU(N)$ Yang-Mills theory on
dS${}_4$. Below let us spell out some of the basic elements of the
duality.
\subsection{dS Embedding in AdS}
The $(d+1)$-dimensional Anti-de Sitter space AdS${}_{d+1}$ is a
maximally symmetric space with negative cosmological constant.
It can be most easily defined by an embedding
\begin{equation}\label{ads1}
-X_0^2+X_1^2+\ldots+X_{d}^2-X_{d+1}^2=-L^2~,
\end{equation}
in the $(d+2)$-dimensional flat space ${\bf R}^{d+2}$ with the metric
$\eta_{\unM \unN} = \rm{diag} (-1, \mathbb{1}_d, -1)$. Here $L$ is the radius of
the AdS space and the cosmological constant is given by
\begin{equation}\label{lc}
\L_0=-\frac{ d(d-1)}{2 L^2}~.
\end{equation}
AdS${}_{d+1}$ is invariant under the group $SO(2,d)$ as both the
embedding metric and the embedding equations are invariant under this
transformation.
Similarly the de Sitter space dS is a maximally symmetric
space with positive cosmological constant. For $d$-dimensional dS
space, it is given by the hyperboloid
\begin{equation}
-Y_0^2+Y_1^2+\ldots+Y_{d}^2= L^2~,\lab{yconstr}
\end{equation}
in the flat space ${\bf R}^{d+1}$ with the metric
$\eta_{\unM \unN} = \rm{diag} (-1, \mathbb{1}_d)$. The cosmological constant
is
\begin{equation}\label{lc1}
\L_1=\frac{ (d-1)(d-2)}{2 L^2}~,
\end{equation}
and dS${}_d$ has the symmetry group $SO(1,d)$.
In the standard application of AdS/CFT correspondence,
one uses the Poincare
coordinates
\begin{align} \label{poin}
&& X_0= \frac{r}{2}\left[ 1+ \frac{x^2- t^2 + L^2}{r^2}\right],
\qquad
X_i = \frac{ L x_i}{r}, \quad i =1, \cdots d-1, {\nonumber} \\
&& X_d = \frac{r}{2}\left[ 1+ \frac{x^2- t^2 -L^2}{r^2}\right],
\qquad
X_{d+1} = \frac{L t}{r},
\end{align}
in which case the AdS metric takes the form
\begin{equation}\label{adsp}
ds^2=\frac{L^2}{r^2}(d r^2 - dt^2 +d x_i^2), \quad r \geq 0.
\end{equation}
It is clear that each constant $r$-slice describes
a copy of Minkowski space. The reason of the constraint
$r \geq 0$ is because of the singularity at $r =0$ and so the
metric \eq{adsp} can cover only half of the AdS space.
Hence the Poincare patch
describes a patch of the AdS
space with boundary consisting of a copy of the Minkowski space
$M_d$ at $r =0$,
together with a single point $P$ at $r = \infty$. This fact has
been of crucial
importance in the prescription of \cite{witten,GKP}
for the realization of the holography of gravity in AdS
space.
It is interesting to note that the $(d+1)$-dimensional Anti-de Sitter
space AdS${}_{d+1}$ also admit a
coordinate patch with $d$-dimensional de Sitter space
dS${}_d$ slicing. This fact has been used in the construction
of braneworld with cosmological constant \cite{pope}.
The embedding of dS${}_d$
can be realized with the following change of coordinates of
the AdS${}_{d+1}$:
\begin{align}\lab{cc1}
X_{d+1} = L \cosh\frac{z}{L}~, \qquad\quad
X_\mu} \def\n{\nu= Y_\mu} \def\n{\nu\sinh\frac{z}{L}, \quad
\mu} \def\n{\nu=0,1,\ldots,d,
\end{align}
with $Y_\mu} \def\n{\nu$ satisfying \eq{yconstr} and describes de Sitter space dS${}_d$.
In this coordinate patch, the AdS metric takes the form
\begin{equation}\lab{ds1}
ds^2= dz^2+\sinh^2(\frac{z}{L}) \; ds^2_{dS}~, \quad z \geq 0.
\end{equation}
The metric \eq{ds1} describes a portion of the AdS
space with boundary consisting of a copy of the de Sitter space
dS${}_d$ at $z =\infty$ , together with a single point at $z=0$.
An explicit solution of the constraint \eq{yconstr} is given by
\begin{equation}\lab{cc2}
Y_0=\frac{\sinh H t}{H}-\frac{1}{2}H x_i^2 e^{-H t},\qquad
Y_i=x_i e^{-H t}, \qquad
Y_{d}=\frac{\cosh H t}{H}-\frac{1}{2}H x_i^2 e^{-H t},
\end{equation}
with $H= 1/L$ and
$i = 1, \cdots, d-1$. This gives the dS${}_d$ metric in terms of the planar
coordinates $(t,x^i)$:
\begin{equation}\lab{ds2}
ds^2_{dS} =-dt^2+e^{-2 H t} d x_i^2~ .
\end{equation}
The normalized distance $P$
between two points $X$ and $X'$ in the ambient space
\begin{equation}\label{inv1}
P(X,X'):=\frac{\eta_{\unM \unN} X^\unM X'^\unN}{L^2}~,
\end{equation}
is a convenient quantity that can be used to express
the geodesic distance ${\cal D}} \def\cE{{\cal E}} \def\cF{{\cal F}$ between any two points in the AdS or dS.
For time like separated points in AdS, $P$ is related to
the geodesic distance ${\cal D}} \def\cE{{\cal E}} \def\cF{{\cal F}$ between two points $X, X'$ by
$ P = - \cos ({\cal D}} \def\cE{{\cal E}} \def\cF{{\cal F}/L)$.
For points in the same causal
diamond in dS, the relation is $P = \cos
({\cal D}} \def\cE{{\cal E}} \def\cF{{\cal F}/L)$.
For AdS${}_{d+1}$, $P$ is
given by
\begin{equation}
P_{\rm AdS}(X,X')=-\frac{1}{\xi}~,
\end{equation}
where
\begin{equation}\lab{geod-poin}
\xi^{-1} =\frac{r^2+r'^2+(x_i-x_i')^2-(t-t')^2}{2 r r'}
\end{equation}
in the Poincare coordinates \eq{poin};
and
\begin{equation}\label{geod-planar}
\xi^{-1} = \cosh H z \, \cosh H z' - \sinh Hz \, \sinh Hz'
\times P_{dS} (x^\mu, x'{}^\mu)
\end{equation}
in the dS planar coordinates \eq{cc1}, \eq{cc2}.
Here
\begin{equation}
P_{\rm dS} := \cosh H(t-t') - \frac{e^{-H(t+t')}}{2} \; H^2
(x_i-x_i')^2
\end{equation}
is the normalized distance in dS${}_d$. Note that $P_{\rm dS}=1$ for coincident
points, therefore it is more convenient to consider the following
quantity
\begin{equation} \label{sigma}
\sigma} \def\S{\Sigma^2(x,x'):= e^{-H(t+t')} \prt{x_i-x_i'}^2
-\frac{\cosh H \prt{t-t'}-1}{H^2/2} ~
\end{equation}
as a measurement of distance between any two points $(t,x_i), (t', x'_i)$ in
dS${}_d$.
In general it is
\begin{equation}
\frac{P_{\rm dS} -1}{H^2} = -\frac{1}{2}\sigma} \def\S{\Sigma^2.
\end{equation}
The quantity $\sigma} \def\S{\Sigma^2$ has the property that it coincides with the proper
distance in Minkowski spacetime in the flat space limit $H \to 0$,
\begin{equation}
\sigma} \def\S{\Sigma^2 \to |x-x'|^{2} := -(t-t')^2 + (x-x')^2 .
\end{equation}
In terms of the conformal time $x^0
= H^{-1}\exp (Ht)$, the dS metric \eq{ds2} can be written as
\begin{equation} \label{ds3}
ds^2 = \frac{1}{H^2 x_0^2} \prt{-dx_0^2+ dx_i^2 }
\end{equation}
and it is
\begin{equation} \label{sigma1}
\sigma} \def\S{\Sigma^2 = \frac{(x_\mu} \def\n{\nu -x'_\mu} \def\n{\nu)^2}{H^2 x_0 x'_0}~,
\end{equation}
where indices are raised and lowered by the Minkowski metric $\eta_{\mu} \def\n{\nu\n}$.
\subsection{ $\cN=4$ superconformal Yang-Mills theory}
Let us review the $\cN=4$ superconformal Yang-Mills theory on dS${}_4$
constructed in \cite{dzf}. There the metric is taken to be of the
form \eq{ds3}. Crucial to the construction of \cite{dzf} is the
existence of conformal Killing spinor on dS${}_4$.
Unlike the Killing spinor equation,
the conformal Killing spinor defined by the equation
\begin{equation}
\prt{ D_\mu - \frac{1}{4} \gamma} \def\G{\Gamma} \def\dc{{\dot\gamma}_\mu} \def\n{\nu {D \hspace{-7.5pt} \slash}\;} \epsilon = 0
\end{equation}
is compatible with the Majorana condition on spinor. This can be
solved and one obtain the conformal Killing spinors on dS${}_4$
\begin{equation}
\epsilon(x) = \frac{1}{\sqrt{H x^0}}(\eta_0 + x^\mu} \def\n{\nu \gamma} \def\G{\Gamma} \def\dc{{\dot\gamma}_\mu} \def\n{\nu \eta_1 )~,
\end{equation}
where $\eta_0, \eta_1$ are arbitrary Majorana spinors.
This gives $\cN=1$ superconformal symmetry in dS${}_4$ and
corresponds to a basis of 8 real supercharges.
The $\cN=4$ maximal superconformal Yang-Mills theory on dS${}_4$ contains the
gauge potentials $A_\mu} \def\n{\nu^a$, four Majorana gauginos $\lambda} \def\L{\Lambda^a_\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}$ and six
real scalars $X^{a}_i$, where the indices $a$ is in the adjoint of the gauge group
$SU(N)$. The Lagrangian is $\cL = \cL_2+ \cL_3 + \cL_4$ where
\begin{align}
\cL_2 & = & - \Big[
\frac{1}{4} F_{\mu\nu}^a F^{\mu} \def\n{\nu\n a} + \bar{\l}^{a\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}} \gamma} \def\G{\Gamma} \def\dc{{\dot\gamma}^\mu} \def\n{\nu D_\mu} \def\n{\nu P_L
\lambda} \def\L{\Lambda^a_\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}+
\frac{1}{2} D_\mu X_i^a D^\mu X^a_i + H^2 X^a_i X^a_i
\Big]~, \\
\cL_3 &=& - \frac{1}{2} f^{abc} X_i^a \Big[
C_i^{\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}\beta} \def\db{{\dot\beta}} \bar{\l}^b_\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A} P_L\lambda} \def\L{\Lambda^c_\beta} \def\db{{\dot\beta} +C_{i\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}\beta} \def\db{{\dot\beta}} \bar{\l}^{b\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}}P_R \lambda} \def\L{\Lambda^{c\beta} \def\db{{\dot\beta}}
\Big]~, \\
\cL_4 &=& -\frac{1}{4}f^{abc} f^{a'b'c'} X^b_i X^c_j X^{b'}_i X^{c'}_j~.
\end{align}
Here $P_L, P_R$ are chiral projectors, $C_i$ are the six `t
Hooft instanton matrices:
\begin{align}
C_1 &=& \begin{pmatrix} 0 & \sigma_1\\ -\sigma} \def\S{\Sigma_1 & 0 \end{pmatrix}~, \quad
C_2 = \begin{pmatrix} 0 & -\sigma} \def\S{\Sigma_3\\ \sigma} \def\S{\Sigma_3 & 0 \end{pmatrix}~, \quad
C_3 = \begin{pmatrix} i \sigma} \def\S{\Sigma_2 & 0\\ 0& i\sigma} \def\S{\Sigma_2 \end{pmatrix}~,\\
C_4 &=& -i \begin{pmatrix} 0 & i\sigma} \def\S{\Sigma_2\\ i\sigma} \def\S{\Sigma_2 & 0 \end{pmatrix}~, \quad
C_5 = -i \begin{pmatrix} 0 & 1\\ -1 & 0 \end{pmatrix}~, \quad
C_6 = -i \begin{pmatrix} -i \sigma} \def\S{\Sigma_2 & 0\\ 0& i\sigma} \def\S{\Sigma_2 \end{pmatrix}
\end{align}
and $\sigma} \def\S{\Sigma_i$ are the Pauli matrices. Note that $C_1, C_2, C_3$ are
real, $C_4, C_5, C_6$ are imaginary.
The action admits an $SU(4)$ R-symmetry
and the superconformal symmetry:
\begin{align}
\delta}\def\D{\Delta}\def\ddt{\dot\delta A_\mu} \def\n{\nu^a &=& - \bar{\epsilon}^\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A} \gamma} \def\G{\Gamma} \def\dc{{\dot\gamma}_\mu P_L \lambda} \def\L{\Lambda^a_\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A} - \bar{\epsilon}_\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A} \gamma} \def\G{\Gamma} \def\dc{{\dot\gamma}_\mu} \def\n{\nu P_R \lambda} \def\L{\Lambda^{a\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}}~, \\
\delta}\def\D{\Delta}\def\ddt{\dot\delta X^a_i &=& -\bar{\epsilon}_\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A} P_L C_i^{\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}\beta} \def\db{{\dot\beta}} \lambda} \def\L{\Lambda_\beta} \def\db{{\dot\beta} - \epsilon^\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A} P_R C_{i\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}\beta} \def\db{{\dot\beta}}
\lambda} \def\L{\Lambda^{a\beta} \def\db{{\dot\beta}}~, \\
\delta}\def\D{\Delta}\def\ddt{\dot\delta \lambda} \def\L{\Lambda^a_\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A} &=& \frac{1}{2} \gamma} \def\G{\Gamma} \def\dc{{\dot\gamma}^{\mu\nu} F_{\mu} \def\n{\nu\n}^a \epsilon_\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A} -\gamma} \def\G{\Gamma} \def\dc{{\dot\gamma}^\mu} \def\n{\nu D_\mu} \def\n{\nu X_i^a
(P_L C_i^{\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}\beta} \def\db{{\dot\beta}} \epsilon_\beta} \def\db{{\dot\beta} +P_R C_{i \alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}\beta} \def\db{{\dot\beta}} \epsilon^\beta} \def\db{{\dot\beta})
-\frac{1}{2} X_i^a (P_R C_i^{\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}\beta} \def\db{{\dot\beta}} {D \hspace{-7.5pt} \slash}\; \epsilon_\beta} \def\db{{\dot\beta} +P_L C_{i \alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}\beta} \def\db{{\dot\beta}} {D \hspace{-7.5pt} \slash}\;
\epsilon_\beta} \def\db{{\dot\beta}) {\nonumber}\\
&& - \frac{1}{2}f^{abc} X_i^b X_j^c [(C_i C_j)^\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}{}_\beta} \def\db{{\dot\beta} P_R \epsilon^\beta} \def\db{{\dot\beta}
+ (C_i C_j)_\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}{}^\beta} \def\db{{\dot\beta} P_L \epsilon_\beta} \def\db{{\dot\beta}]~,
\end{align}
where $P_L\epsilon_\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}, P_R \epsilon^\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}$ are an $SU(4)$ quartet of Majorana
conformal Killing spinors. Here $(C_i C_j)^\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}{}_\beta} \def\db{{\dot\beta} := C_i^{\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}\gamma} \def\G{\Gamma} \def\dc{{\dot\gamma}} C_{j
\gamma} \def\G{\Gamma} \def\dc{{\dot\gamma}\beta} \def\db{{\dot\beta}}$ and $ (C_i C_j)_\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}{}^\beta} \def\db{{\dot\beta} := C_{i\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}\gamma} \def\G{\Gamma} \def\dc{{\dot\gamma}} C^{\gamma} \def\G{\Gamma} \def\dc{{\dot\gamma}\beta} \def\db{{\dot\beta}}_j$.
Due to its large amount of superconformal symmetry, the theory is
expected to be UV finite. Since there is no massless minimal coupled
field, it is also expected that there is no IR divergence.
Hence the theory is expected to enjoy exact $SU(2,2|4)$ supersymmetry.
Adding a $\theta} \def\Th{\Theta} \def\vth{\vartheta$-term and restore the Yang-Mills coupling $g$, one
expect the theory also enjoy exact $SL(2,Z)$ strong-weak duality, just
as the type IIB superstring theory does.
The moduli space of the theory can be easily worked out. With fermions
and the gauge fields set to zero. The
equation of motion for the scalar fields read
\begin{equation}
-D_\mu} \def\n{\nu D^\mu} \def\n{\nu X_i + 2 H^2 X_i - [X_j,[X_i,X_j]] =0~.
\end{equation}
One class of solution (static) is given by product of fuzzy spaces described by
$SU(2)$. A more interesting solution is
\begin{equation}
X_i = e^{Ht} Z_i~,
\end{equation}
where $Z_i$ are diagonal. This describes an expanding $R^6$
and is
expected from the form \eq{ds2} of the de Sitter metric used.
\subsection{Conformal field theory in dS space}
The $SO(1,d)$ isometries of
de Sitter space dS${}_d$ is generated by the generators:
\begin{equation}
L_{AB} = Y^A \frac{\del}{\del Y^B} - Y^B \frac{\del}{\del Y^A}, \quad
A, B = 0, 1, \cdots, d,
\end{equation}
which acts linearly on the de Sitter hyperboloid.
In terms of the dS space conformal coordinates \eq{ds3},
the de Sitter isometries are generated by the
spatial rotations $J_{ij}$, dilation $D$,
spatial translations $P_i$ and special conformal
transformation $K_i$:
\begin{align}
J_{ij} &=& -i L_{ij}= -i( x_i \del_j - x_j \del_i) ,
\quad \mbox{where $i =1,\cdots, d-1$},
\\
D &=& -i L_{0d} = -i x^\mu} \def\n{\nu \del_\mu} \def\n{\nu, \\
P_i &=& -i (L_{id} +L_{0 i}) = -i H^{-1}\del_i , \\
K_i &=& i (L_{id} - L_{0 i}) = -2iHx_i \, x^\mu} \def\n{\nu \del_\mu} \def\n{\nu - iH
x^2 \del_i .
\end{align}
The corresponding finite transformations are:
\begin{align}
x_i'& =& \L_i^j x_j, \qquad \mbox{$\L \in SO(d-1)$ rotations}, \\
x'{}^\mu} \def\n{\nu &=& \lambda} \def\L{\Lambda x^\mu} \def\n{\nu,\\
x_i' &=& x_i +a_i, \\
x'{}^\mu} \def\n{\nu &=& \frac{x^\mu} \def\n{\nu + b^\mu} \def\n{\nu x^2}{1+2b_\mu} \def\n{\nu x^\mu} \def\n{\nu + b^2 x^2}, \quad b^\mu} \def\n{\nu
= (0, b^i).
\end{align}
As the de Sitter metric is related to the flat space Minkowski
metric by a Weyl factor, the conformal symmetry of dS${}_d$ is the same
$SO(2,d)$ as that of the Minkowski spacetime, and is obtained by
adding to the dS isometries the generators:
the Lorentz boosts
$J_{0i}$, time translation $P^0$ and the special conformal transformation $K^0$.
The corresponding finite transformations are:
\begin{align}
x'{}^\mu} \def\n{\nu& =& \L^\mu} \def\n{\nu_\n x^\n, \qquad \mbox{nonvanishing $\L^0_i, \L^i_0$
= Lorentz boost}, \\
x'{}^0 &=& x^0 +a^0, \\
x'{}^\mu} \def\n{\nu &=& \frac{x^\mu} \def\n{\nu + b^\mu} \def\n{\nu x^2}{1+2b_\mu} \def\n{\nu x^\mu} \def\n{\nu + b^2 x^2}, \quad b^\mu} \def\n{\nu
= (b^0, 0)
\end{align}
and the metric transforms as
\begin{equation}
ds^2 \to ds'{}^2 = \L^2(x) ds^2,
\end{equation}
where, respectively,
\begin{equation}
\L(x)^2 = \prt{\frac{x_0}{\L^0_\mu} \def\n{\nu x^\mu} \def\n{\nu}}^2, \qquad
\prt{\frac{x_0}{x_0+a_0}}^2, \qquad 1 .
\end{equation}
As usual, the special
conformal transformations can be constructed out of translations and
the inversion
\begin{equation} \label{inv}
x^\mu} \def\n{\nu \to x'{}^\mu} \def\n{\nu = \frac{x^\mu} \def\n{\nu}{x^2} := \prt{\frac{1}{x} }^\mu} \def\n{\nu.
\end{equation}
Inversion is an isometry and induces an $SO(1,d-1)$ rotation on vector
\begin{equation}
\frac{\del}{\del x'{}^\mu} \def\n{\nu} = x^2 M_\mu} \def\n{\nu^\n(x) \frac{\del}{\del x^\n} ,
\end{equation}
where
\begin{equation}
M_\mu} \def\n{\nu^\n(x) := \delta}\def\D{\Delta}\def\ddt{\dot\delta_\mu} \def\n{\nu{}^\n - 2\frac{x_\mu} \def\n{\nu x^\n}{x^2}
\end{equation}
and satisfies $M_\mu} \def\n{\nu^\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A} M_\n^\beta} \def\db{{\dot\beta} \eta_{\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}\beta} \def\db{{\dot\beta}} =
\eta_{\mu} \def\n{\nu\n}$. For a spinor in fundamental representation, inversion induces the
transformation
\begin{equation}
\psi'(x) =
S_\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}^\beta} \def\db{{\dot\beta}(x) \psi_\beta} \def\db{{\dot\beta}\prt{\frac{1}{x}},
\end{equation}
where the matrix $ S^\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}_\beta} \def\db{{\dot\beta}(x)$ satisfies
\begin{equation}
S^\dagger(x) \G^\mu} \def\n{\nu S(x) = M^\mu} \def\n{\nu_\n \G^\n
\end{equation}
and is given by
\begin{equation} \lab{smatrix1}
S(x) = \frac{\G_\mu} \def\n{\nu x^\mu} \def\n{\nu}{|x|}.
\end{equation}
In a conformal field theory,
scalar (with respect to the rotation group)
operator $\cO$ of conformal dimension $\D$ satisfies
under conformal transformation $x\to x'$ as
\begin{equation}
\quad\cO'(x') = \frac{1}{|\L(x)|^{\D}} \cO(x).
\end{equation}
For two such operators,
dS invariance implies that their
$2$ point function must be function of the geodesic distance
$\sigma} \def\S{\Sigma(x,y)^2$. Furthermore, invariance under the conformal
transformations implies that
\begin{equation}
\langle \cO_1 (x) \cO_2(y) \rangle = \left\{
\begin{array}{ll}
\dfrac{C_{12}}{\sigma} \def\S{\Sigma(x,y)^{2\D}} & \mbox{if $\D_1 = \D_2 = \D$}, \\
0 & \mbox{if $\D_1 \neq \D_2$}.
\end{array}\right.
\end{equation}
As for operator
$\cO^\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}(x)$
of dimension $\D$ in spin 1/2 representation, it
satisfies under translation in $x^0$:
\begin{equation}
\cO'{}^\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}(x') = \frac{1}{|\L(x)|^{\D}} S^\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}_\beta} \def\db{{\dot\beta}(x) \cO^\beta} \def\db{{\dot\beta}(x^\mu} \def\n{\nu).
\end{equation}
Together with it's property under inversion
\begin{equation}
\cO'{}^\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}(x') = S^\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}_\beta} \def\db{{\dot\beta}(x) \cO^\beta} \def\db{{\dot\beta}(x^\mu} \def\n{\nu),
\end{equation}
one can easily show that the two point function
for any two such operators is
\begin{equation} \label{OOD}
\langle \cO^\dagger_{1\,\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}}(x) \cO_2^\beta} \def\db{{\dot\beta}(y) \rangle =
\frac{D_\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}{}^\beta} \def\db{{\dot\beta}(x,y)}{\sigma} \def\S{\Sigma(x,y)^{2\D}} \delta}\def\D{\Delta}\def\ddt{\dot\delta_{\D_1, \D_2},
\end{equation}
where $\D= \D_1$. Here $D$ satisfies the relation
\begin{equation}
D(x,y) = S^\dagger(x) D\prt{\frac{1}{x},\frac{1}{y}} S(y)
\end{equation}
and is given by
\begin{equation}
D_\alpha} \def\da{{\dot\alpha}} \def\dA{{\dot A}{}^\beta} \def\db{{\dot\beta}(x,y) = \frac{(x-y)_\mu} \def\n{\nu \G^\mu} \def\n{\nu}{|x-y|}.
\end{equation}
Constraints on 3 and higher point functions of dS CFT
can be similarly worked out.
\section{Bulk Propagators and
Boundary Correlators: General Setup} \label{b2b-general}
One of the very basic tools for the study of gauge/gravity
correspondence is the bulk to boundary formalism
\cite{witten, GKP}. In this section
we develop this formalism for a general bulk metric. Most of what we do in
section \ref{b2b-scalar} is a simple generalization of
known results in the literature. The general
result obtained in section \ref{b2b-fermion} is new.
Consider a $(d+1)$-dimensional manifold ${\cal M}} \def\cN{{\cal N}} \def\cO{{\cal O}$ with boundary and the metric
\begin{equation}
ds^2=g_{MN} dy^M dy^N~.
\end{equation}
Without loss of generality, we assume that
the metric has an expansion of the form near the boundary,
\begin{equation} \label{bdy-form}
ds^2 =dz^2+\gamma} \def\G{\Gamma} \def\dc{{\dot\gamma}_{\mu} \def\n{\nu \n}(z,x) dx^\mu} \def\n{\nu dx^\n~,\qquad \gamma} \def\G{\Gamma} \def\dc{{\dot\gamma}_{\mu} \def\n{\nu\n}(z,x) = p^2(z)h_{\mu} \def\n{\nu\n}(x)~,
\end{equation}
for some function $p (z)$ and
$h_{\mu} \def\n{\nu\n} (\mu} \def\n{\nu,\n = 0\cdots, d-1)$ is the boundary metric.
The boundary is supposed to be at some location $z = a$ in this
coordinate system.
\subsection{Scalar case} \label{b2b-scalar}
First let us consider the case of a real scalar field
with the action
\begin{equation}
I(\phi)=-\frac{1}{2} \int d^{d+1}y \sqrt{g} (g^{MN}
\partial_M\varphi \partial_N \varphi+m^2 \varphi^2),
\end{equation}
where $g$ is the absolute value
of the determinant of the matrix $g_{MN}$.
Performing an integration by parts leads to
\begin{equation}\label{act1}
I(\phi)=-\frac{1}{2} \int_{\del M} d^{d}x\sqrt{\gamma} \def\G{\Gamma} \def\dc{{\dot\gamma}}\varphi n^M\partial_M\varphi
+\frac{1}{2} \int d^{d+1}y\sqrt{g} \varphi(\square-m^2)\varphi,
\end{equation}
where $n^M:=\partial x^M/\partial z$
is the normal vector to the boundary surface. For our metric
\eq{bdy-form}, the only nonvanishing component is $n^z =1$.
The first term of \eq{act1} is a boundary term and the second term gives
the equations of motion
\begin{equation}\label{kg1}
(\square-m^2) \varphi=0~,
\quad \square :=\frac{1}{\sqrt{g}}\partial_M(\sqrt{g} g^{MN}\partial_N \; ),
\end{equation}
where $\square$ is the $(d+1)$-dimensional d'Alembertian operator of ${\cal M}} \def\cN{{\cal N}} \def\cO{{\cal O}$.
When evaluated on shell, the bulk term in \eq{act1} vanishes and
only the boundary term contributes. It is clear that the boundary contribution
depends on the boundary behavior of the solution of the equation of
motion. Without loss of generality, consider a solution of $\varphi$ with the
following leading asymptotic behavior near the boundary,
\begin{equation}
\varphi \sim f(z) \varphi_0(x) , \quad z \sim a,
\end{equation}
for some function $f$ of $z$. It is then convenient to introduce the
bulk to boundary propagator $K$ defined by the following differential
equation
\begin{equation} \label{K-def}
\square K(z,x,x')=0~,
\end{equation}
and has the boundary behavior
\begin{equation} \label{K-def2
\lim_{z\rightarrow a} K(z,x,x')\sim f(z) \frac{\delta^{(d)}(x-x')}{\sqrt{h}}~,
\end{equation}
where $h$ is the absolute value of the determinant of the boundary
metric $h_{\mu} \def\n{\nu\n}$
and $\sim$ denotes the leading order contributing term in the sense of
distribution.
For convenience, we have chosen to include a volume factor
$\sqrt{h}$ in our definition of $K$, making it scalar.
The introduction of the bulk to boundary propagator
allows us to write $\varphi(z,x)$ as
\begin{equation}\label{phi1}
\varphi(z,x)=\int d^d x'\sqrt{h(x')} K(z,x,x')\varphi_0(x')~,
\end{equation}
in agreement with the definition \eq{K-def2}. If we substitute \eq{phi1} into \eq{act1}, we get
\begin{equation}
I(\phi)=-\frac{1}{2} \int d^{d}x\sqrt{h(x)} d^{d}x' \sqrt{h(x')} \;
\varphi_0(x) {\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}(x,x') \varphi_0(x') ~,
\end{equation}
with
${\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}(x,x')$ defined as
\begin{equation}\label{cg1}
{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}(x,x'):=\lim_{z\rightarrow a}
\prt{f(z) p^d(z) \partial_z K(z,x,x')} ~.
\end{equation}
According to the prescription of AdS/CFT correspondence \cite{witten,GKP},
the 2-point function of the dual field theory is given by ${\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I} (x,x')$.
The equation \eq{cg1} expresses it in terms of the normal derivative
of the bulk-to-boundary propagator $K$ at the boundary.
The key goal now is to obtain the bulk-to-boundary propagator. In the
original works \cite{witten,GKP}, this is obtained by solving the
differential equation \eq{K-def} directly. There is however a more
effective way.
Let us introduce the Green function for the bulk
\begin{equation} \label{green}
(-\square+m^2)G(z,x;z',x')=\frac{1}{\sqrt{g}}\delta^{(d)} (x-x')\delta(z-z')~.
\end{equation}
Using the Green's identity, one can easily obtain
the solution of the scalar Klein-Gordon equation \eq{kg1}
in terms of the Green function as
\begin{equation}
\varphi(z,x)=\int d^{d} x'\varphi_0(x') \sqrt{\gamma} \def\G{\Gamma} \def\dc{{\dot\gamma}(z',x')} \Big(
G \del_{z'} f(z') -f(z') \partial_{z'} G
\Big)_{z'\rightarrow a}~.
\end{equation}
Comparing with \eq{phi1}, the bulk-to-boundary propagator $K$ can be
written in terms of the Green function as
\begin{equation}\label{ktog}
K(z,x,x')= \lim_{z' \rightarrow a} p^d(z')
\Big(
G(z,x;z',x') \del_{z'} f(z') -f(z') \partial_{z'} G(z,x;z',x')
\Big)
\end{equation}
and subsequently the two point function ${\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}(x,x')$ can be obtained
in terms of $K$ using \eq{cg1}.
This formula displays clearly how the bulk physics, as encoded in the
bulk Green function, is translated to the physics on the
holographic field theory
through the boundary data: the asymptotic behavior $f$ of
the field and of the
metric volume factor $p^d(z)$ near the boundary.
The relation of the propagator with the Green
function and its derivative, turns out to be
always
scaled in such a way
that the limit at the boundary is finite.
Higher point functions can be obtained by the Witten diagrams \cite{witten}.
\subsection{Spin 1/2 case} \label{b2b-fermion}
The action for a massive free spin 1/2 fermion
on our $(d+1)$-dimensional space ${\cal M}} \def\cN{{\cal N}} \def\cO{{\cal O}$ reads
\begin{equation}\lab{actionf}
I_0=\int d^{d+1} x \sqrt{g} \bar{\psi}\prt{{D \hspace{-7.5pt} \slash}\;-\mu} \def\n{\nu}\psi,
\end{equation}
where $g_{MN} $ is the metric on ${\cal M}} \def\cN{{\cal N}} \def\cO{{\cal O}$. Without loss of generality, we
assume $\mu} \def\n{\nu>0$. The Dirac equation of motion reads
\begin{equation}\lab{dir1}
\prt{{D \hspace{-7.5pt} \slash}\;-\mu} \def\n{\nu}\psi=0~, \quad \bar{\psi} (- \overleftarrow{{D \hspace{-7.5pt} \slash}\;} -\mu} \def\n{\nu) =0.
\end{equation}
The action \eq{actionf} vanishes on-shell. For the application of
AdS/CFT, one needs to supplement it with the boundary action \cite{sfetsos}
\begin{equation}
\frac{1}{2}\int_{\del {\cal M}} \def\cN{{\cal N}} \def\cO{{\cal O}} d^d x\sqrt{\gamma} \def\G{\Gamma} \def\dc{{\dot\gamma}} \bar{\psi} \psi ,
\end{equation}
where $\del M$ is the boundary of ${\cal M}} \def\cN{{\cal N}} \def\cO{{\cal O}$ and the metric takes the form
\eq{bdy-form} near the boundary. In practical calculation,
as various quantities in $I_b$ are divergent near the boundary,
one needs to consider it as the limit
\begin{equation} \lab{i1}
I_b := \lim_{z \to a} \frac{1}{2}\int_{\del {\cal M}} \def\cN{{\cal N}} \def\cO{{\cal O}} d^d x\sqrt{\gamma} \def\G{\Gamma} \def\dc{{\dot\gamma}} \bar{\psi} \psi .
\end{equation}
The necessity of the boundary term was justified in \cite{henneaux}
which demonstrated that only then the variational principle for the
fermionic action is well defined:
it ensures that by decomposing the spinor $\psi$ in terms of the
eigenvalues
of say, the Gamma matrix $\G^z$, the on-shell action is not a function
of both components, since the regularity of the solution on ${\cal M}} \def\cN{{\cal N}} \def\cO{{\cal O}$
restrict that only half the components of the spinor $\psi$ can be
prescribed on the boundary.
Now let us consider a solution $\psi$ with the leading asymptotic behavior near
the boundary
\begin{equation}
\psi \sim f(z) \psi_0(x), \quad z\sim a,
\end{equation}
where $f(z)$ is a function and $\psi_0$ is a spinor living on the boundary.
It is easy to see that for positive $\mu} \def\n{\nu$, the non-normalizable mode
is obtain from $\psi_0$ of negative chirality:
\begin{equation}
\G^z \psi_0 =- \psi_0.
\end{equation}
The fermionic bulk to boundary propagator $S(z,x,x')$ from a point
$(z,x^\mu} \def\n{\nu)$ in the interior to a point $x'{}^\mu} \def\n{\nu$ on the boundary
is defined by the
differential equation
\begin{equation} \label{S-def1}
({D \hspace{-7.5pt} \slash}\; -\mu} \def\n{\nu) S =0
\end{equation}
and the boundary behavior
\begin{equation} \label{S-def2
\lim_{z \to a} S(z,x,x') \sim f(z) \frac{\delta}\def\D{\Delta}\def\ddt{\dot\delta^{(d)}(x-x')}{\sqrt{h}} {\bf 1}.
\end{equation}
It allows us to write the on-shell configuration of $\psi$ as
\begin{equation}\lab{sola1}
\psi(z,x)=\int d^d x' \sqrt{h(x')}S\prt{z,x,x'} \psi_0\prt{x'}~,
\end{equation}
in agreement with the definition \eq{S-def2}.
Substituting \eq{sola1} into $I_b$, we obtain
\begin{equation}\lab{i2}
I_b=\int d^d x'\sqrt{h(x')} d^d x''\sqrt{h(x'')}\;
\bar{\psi}_0(x'){\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}(x',x'') \psi_0(x'')~,
\end{equation}
where
\begin{equation} \label{G2-fermion}
{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}(x',x'') := \lim_{z\to a} \frac{1}{2}
\int d^d x \sqrt{\gamma} \def\G{\Gamma} \def\dc{{\dot\gamma}(z,x)} S^\dagger(z,x,x') S(z,x,x'')~.
\end{equation}
As $S$ behaves like a delta function near the boundary, the integral
picks up its contribution from the two regions: $x \sim x'$ and $ x\sim
x''$ and we obtain
\begin{equation} \label{cg2}
{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}(x',x'') = \lim_{z\to a} \frac{1}{2} f(z) p^d(z) \prt{ S^\dagger(z,x'',x')
+ S(z,x',x'') }.
\end{equation}
This formula is analogous to \eq{cg1} and gives the fermionic
two point function
in terms of the fermionic bulk to boundary propagator $S$. To find
$S$, one can try to solve for it directly from the defining equations
\eq{S-def1} and \eq{S-def2}. This has been carried out in
\cite{sfetsos,mueck,henneaux} for the original AdS/CFT correspondence
with Minkowski CFT living on the boundary.
However this is not necessary as we only
need $S$ near the boundary in \eq{cg2}.
In the next section, we will show that $S$ in \eq{cg2}
can be obtained in terms of the
scalar bulk to boundary propagator $K$:
\begin{equation} \label{cg3}
S = {D \hspace{-7.5pt} \slash}\; K.
\end{equation}
As noted in \cite{pope}, apart from having Minkowski spacetime M${}_4$ and de
Sitter spacetime dS${}_4$ as boundaries, it is also possible to consider another
coordinate patch of the AdS${}_{5}$ and have AdS${}_4$ as
boundary. In this case boundary dual theory is a superconformal
Yang-Mills theory living on AdS${}_4$. The general
results obtained here can be applied to all these cases.
\section{Scalar 2-Point Function in dS Dual Field Theory}
In the above we have shown how the bulk to boundary propagator and the
two point correlators in the boundary field theory can be derived from
a knowledge of the bulk Green function.
In this section, we consider different boundaries of the AdS
space and will use these formulas to show how the same AdS bulk physics
could manifest itself differently on the boundary field theories.
In the standard AdS/CFT correspondence which employ the Poincare
coordinate
patch \eq{adsp} of the AdS space, one can put the metric in the
form \eq{bdy-form} by writing $ r = \exp(-z/L)$. It is
\begin{equation}
\sqrt{-\gamma} \def\G{\Gamma} \def\dc{{\dot\gamma} } = \frac{1}{r^d}= e^{dz/L}.
\end{equation}
Near the boundary $r' \to 0$ or $z'\to \infty$,
from \eq{geod-poin} we have
\begin{equation}
\xi \approx \frac{2 r r'}{r^2 + |x-x'|^{2} } \to 0.
\end{equation}
To compute the bulk to boundary propagator $K$ from \eq{ktog},
we need to know the Green function near the
boundary. For AdS${}_{d+1}$ in the Poincare coordinates, we have, see
for example \cite{r2},
\begin{equation}\lab{greenads}
L^{d-2}G(X,X')=\frac{C_{\D}}{2\n}\prt{\frac{\xi}{2}}^{\D}
F\prt{\frac{\D}{2},\frac{\D+1}{2},\n+1;\xi^2}~,
\end{equation}
where
\begin{equation}
C_{\D}:=\frac{\G(\D)}{\pi^{\frac{d}{2}}\G\prt{\D-\frac{d}{2}}}
\end{equation}
and $\D$ can be either $\D_+$ or $\D_-$
\begin{equation}
\D_\pm := \frac{d}{2}\pm \n~,\quad \n:=\sqrt{\frac{d^2}{4}+m^2 L^2}~ .
\end{equation}
Using the properties
\begin{equation}
\lim_{x\to 0} F(a,b,c;x) =1~, \quad
\frac{d}{dx}F(a,b,c;x) =\frac{ab}{c} F(a+1,b+1,c+1;x)~,
\end{equation}
of the hypergeometric functions, we get
\begin{align}{\nonumber}
\partial} \def\del{\partial_{z'} \prt{\xi^{\D} F}= -\frac{\D}{L} \xi^{\D} +\cO\prt{\xi^{\D+1}}
\end{align}
near the boundary.
Now it is the non-normalizable mode with the asymptotic
behavior $f = r^{\D_-}$ that defines a field at the boundary.
It is easy to see that in order to
get the desired boundary behavior \eq{K-def2} for $K$, we need to
adopt the root $\D =\D_+$ in the Green function \eq{greenads} and
we obtain the bulk to boundary propagator (up to a constant)
\begin{equation}
K = \Big(\frac{r}{r^2+ |x-x'|^{2} } \Big)^\D.
\end{equation}
It follows immediately from \eq{cg1} that
\begin{equation}
{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}(x,x')= \frac{1}{|x-x'|^{2\D}},
\end{equation}
which is the expected form of the two point function for operator of
dimension $\D$ in CFT living on Minkowski spacetime.
Next we consider the dS slicing \eq{ds1} of AdS.
To compute the 2-point function in the corresponding boundary
conformal
field
theory, we use the fact that the Green function
is a scalar and therefore invariant under
coordinate transformations. The AdS${}_{d+1}$ metric $g$ we consider is
\begin{equation}
ds^2=dz^2+\sinh^2 (Hz)\, ds^2_{\rm dS}~,
\end{equation}
and has a dS boundary with metric $h$, where
$\sqrt{g}=\sinh^d(Hz)\sqrt{h}$.
Near the boundary $z' \rightarrow \infty$, \eq{geod-planar} gives
\begin{equation}\lab{sdef}
\xi=\frac{e^{-H\prt{z+z'}}}{H^2 \rho^2/8} \rightarrow 0 ,
\end{equation}
where
\begin{equation}
\rho^2 = \sigma} \def\S{\Sigma^2 + e^{-2Hz}\cdot \frac{1-H^2\sigma} \def\S{\Sigma^2/4}{H^2/4}
\end{equation}
and $\sigma} \def\S{\Sigma^2$ is given by \eq{sigma1}.
Working similarly as above, we need the asymptotics of the
non-normalizable mode near
the boundary $z \to \infty$ in the dS slicing. In our coordinates,
the Klein-Gordon operator $\square$ is given by
\begin{equation}
\square=\partial} \def\del{\partial_z^2+ \frac{d H}{\tanh H z}\partial} \def\del{\partial_z+\frac{1}{\sinh^2 H
z}\square_{\rm dS}, \quad \square_{\rm dS} =
\frac{1}{\sqrt{-h}}\partial} \def\del{\partial_\mu} \def\n{\nu\prt{\sqrt{-h} h^{\mu} \def\n{\nu\n}\partial} \def\del{\partial_\n}.
\end{equation}
Consider a solution of the form $\varphi\prt{z,x}=f(z) \varphi_0 (x)$,
near the boundary we obtain the second order ordinary differential equation
\begin{equation}
f''+ d Hf'-m^2 f=0~,
\end{equation}
with the non-normalizable solution
\begin{equation}\lab{solkg}
f=e^{-\D_- H z}~,\qquad \D_\pm:=
\frac{d}{2}\pm\sqrt{\frac{d^2}{4}+\frac{m^2}{H^2}}~.
\end{equation}
The bulk to boundary propagator is given by (up to a constant),
\begin{equation}
K = \Big(\frac{e^{-Hz}}{\rho^2}\Big)^\D, \quad
\mbox{with $\D = \D_+$}
\end{equation}
and we obtain from \eq{cg1} the two point function
\begin{equation} \label{res1}
{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I} (x,x') = \frac{1}{\sigma} \def\S{\Sigma(x,x')^{2\D}}.
\end{equation}
This is the expected form of the two point function for operators of
dimension $\D$ of a conformal field theory in dS spacetime.
\section{Fermion 2-Point Function in dS Dual Field Theory}
\subsection{Flat space conformal field theory}
Let us first consider the canonical case \eq{adsp} of AdS in Poincare
coordinates, with Minkowski spacetime located at the boundary at
$r=0$. The vielbein reads
\begin{equation}
e^A_M=\frac{1}{r}\delta}\def\D{\Delta}\def\ddt{\dot\delta^A_M~,
\end{equation}
with the curved indices $M=(r,\mu} \def\n{\nu)$ and flat indices $A=(r,a)$
having range $a,\mu} \def\n{\nu=0,\ldots,d-1 $. For convenience,
we will assume $L=1$ in this
subsection. $L$ can be restored easily by dimensional analysis.
The Dirac operator is given by
\begin{equation}
{D \hspace{-7.5pt} \slash}\;= e^M_A \G^A(\del_M + \frac{1}{2}\omega_M{}^{BC} \S_{BC}) =
r \G^r \del_r+r\G^\mu} \def\n{\nu \del_\mu} \def\n{\nu-\frac{d}{2}\G^r~,
\end{equation}
where the matrices $\G^A = (\G^r,\G^\mu} \def\n{\nu)$ are the flat space gamma matrices:
\begin{equation}
\{\G^A,\G^B\}=2\eta^{AB}~, \quad \eta_{AB} = {\rm diag} (1, \eta_{d}).
\end{equation}
The Dirac equation \eq{dir1} in AdS${}_{d+1}$ reads
\begin{equation}\lab{dir2}
\prt{\G^r\prt{r \del_r-\frac{d}{2}} +r \G^\mu} \def\n{\nu\del_\mu} \def\n{\nu-\mu} \def\n{\nu}\psi=0 ~.
\end{equation}
The asymptotic behavior of the on-shell mode is govern by the
behavior of the Dirac operator near the boundary, we obtain
\begin{equation}
\psi \sim r^{\tilde{\D}_-} \psi_0(x), \quad r\to 0,
\end{equation}
where
\begin{equation}
{\tilde{\D}}_\pm := \frac{d}{2}\mp \mu} \def\n{\nu \G^r.
\end{equation}
In order for this mode to be non-normalizable, it is necessary to take the
boundary spinor $\psi_0$ to be of
negative chirality
\begin{equation}
\G^r \psi_0 = - \psi_0
\end{equation} and $\mu} \def\n{\nu > d/2$
such that $\tilde{\D}_- <0$.
To construct $S$, it is useful to note that the Dirac operator
${D \hspace{-7.5pt} \slash}\;$ satisfies the following relation
\begin{equation} \label{Dss2-flat}
{D \hspace{-7.5pt} \slash}\;^2=\prt{\square+\frac{d^2}{4}} {\bf 1}-r\G^\mu} \def\n{\nu\G^r\del_\mu} \def\n{\nu~,
\end{equation}
where
\begin{equation}
\square=r^2\del_r^2+r(1-d)\del_r+r^2\del_\mu} \def\n{\nu^2~
\end{equation}
is the d'Alembertian on AdS${}_d$ space in the metric \eq{adsp}.
\eq{Dss2-flat} can be written in the more convenient form
\begin{equation}
{D \hspace{-7.5pt} \slash}\;^2+\prt{{D \hspace{-7.5pt} \slash}\;-\mu} \def\n{\nu}\G^r=\prt{\square+\frac{d^2}{4}}+\prt{
r\del_r-\frac{d}{2}-\mu} \def\n{\nu \G^r}~.
\end{equation}
We propose to construct the fermion bulk
to boundary propagator $S$
as follow:
\begin{equation} \label{pr1}
S=\prt{{D \hspace{-7.5pt} \slash}\;+\mu} \def\n{\nu +\G^r}K+\delta}\def\D{\Delta}\def\ddt{\dot\delta~,
\end{equation}
where
\begin{equation}
K = \Big(\frac{r}{\rho^2} \Big)^{\D_+}
\quad \mbox{with} \quad
\rho^2 := r^2 + (x-x')^2 ~,
\end{equation}
is the bulk
to boundary propagator for an auxiliary scalar field of mass $m$;
\begin{equation} \label{pr2}
\delta}\def\D{\Delta}\def\ddt{\dot\delta:=- \prt{{D \hspace{-7.5pt} \slash}\;-\mu} \def\n{\nu}^{-1}\prt{r\del_r-\tilde{\D}_-}K~.
\end{equation}
We remark that
our notation for $\tilde{\D}_\pm$ takes into account that our boundary spinor $\psi_0$
lives in the $\G^r=-1$ sector.
It is easy to see that in order for $S$ to satisfy the equation of
motion \eq{S-def1}, $m$ has to be given by
\begin{equation} \label{pr4}
m^2= \mu} \def\n{\nu^2 -\frac{d^2}{4}.
\end{equation}
This relation is also
precisely
what is needed to guarantee that $S$ satisfies the
desired boundary condition. In fact, since $K$ has the
boundary behavior
\beta} \def\db{{\dot\beta}
\lim_{r\to 0} K \sim r^{\D_-} \delta}\def\D{\Delta}\def\ddt{\dot\delta^{(d)}(x-x').
\end{equation}
As a result, $\delta}\def\D{\Delta}\def\ddt{\dot\delta$ vanishes at the boundary
since $\D_- = \tilde{\D}_-$ (due to \eq{pr4} and
$\G^r =-1$ when acting on $\psi_0$). Now, it is easy to check that
\begin{equation} \label{DssK}
{D \hspace{-7.5pt} \slash}\; K = \mu} \def\n{\nu K - 2 \tilde{\D}_- \frac{r}{\rho} U K~,
\end{equation}
where
\begin{equation}
U := \frac{r \G^r + (x-x')_\mu} \def\n{\nu\G^\mu} \def\n{\nu}{\rho}~.
\end{equation}
Note that $U^2 = {\bf 1}$.
As one approaches the boundary, $K$ imposes $x=x'$ implying
\begin{equation}
\frac{r}{\rho} \to 1, \quad U \to \G^r,
\end{equation}
and so $S$
satisfies the same (up to proportional constant) boundary condition as $K$:
\beta} \def\db{{\dot\beta}
\lim_{r \to 0} S \sim r^{\tilde{\D}_-} \delta}\def\D{\Delta}\def\ddt{\dot\delta^{(d)}(x-x').
\end{equation}
In principle one can use \eq{pr1} to compute the full expression of
$S$. However for our purpose of computing the two point function using
\eq{cg2}, this is not necessary as we only need to know $S$
near the boundary.
Note that the $\mu$ and $\G^r$ term of \eq{pr1}
do not contribute to the boundary action \eq{i2} since
\begin{equation}\lab{cons}
\bar{\psi}_0 \psi_0 =0~,\qquad \bar{\psi}_0 \G^r \psi_0= 0,
\end{equation}
where we have used the fact that $\psi_0$ and $\bar{\psi}_0$ has opposite
chirality. Also, since $\delta}\def\D{\Delta}\def\ddt{\dot\delta =0$ at the boundary.
As a result, the form of $S$ to be used in \eq{cg2} is given by
\begin{equation}
S= {D \hspace{-7.5pt} \slash}\; K.
\end{equation}
Since only the second term in \eq{DssK} contributes, so
we obtain the fermionic two point function
in the conformal field theory living on the boundary flat space
\begin{equation}\lab{fermiong11}
{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}(x,x')= \frac{(x-x')_\mu} \def\n{\nu \Gamma^\mu} \def\n{\nu}{|x-x'|}
\frac{1}{|x-x'|^{2 \tilde{\D}_+}}~.
\end{equation}
The result \eq{fermiong11} was first obtained by \cite{sfetsos}
using the AdS/CFT correspondence.
\subsection{dS conformal field theory}
Next let us consider the AdS${}_{d+1}$ space with a $dS$ boundary with
the metric
\begin{equation}\lab{metric1}
ds^2=dz^2 +\sinh^2 H z\prt{-dt^2+e^{-2 H t}dx_i^2}.
\end{equation}
The vielbein $e^M_A$ is
\begin{equation}
e_{z}^M=\delta}\def\D{\Delta}\def\ddt{\dot\delta^M_z~,\quad e_t^M=\frac{1}{\sinh Hz} \delta}\def\D{\Delta}\def\ddt{\dot\delta_t^M ~,\quad
e^M_i=\frac{e^{H t}}{\sinh H z} \delta}\def\D{\Delta}\def\ddt{\dot\delta_i^M~.
\end{equation}
The non-zero spin connection elements $\omega_M{}^{AB}$ are
\begin{equation}
\omega_M{}^{t~ z}=H \cosh H z \delta_M^t~,\qquad
\omega_M{}^{i~ z}=H \cosh H z e^{-H t} \delta_M^i~,\qquad
\omega_M{}^{i~ t}=H e^{-H t} \delta_M^i~
\end{equation}
and the Dirac operator reads
\begin{equation}
{D \hspace{-7.5pt} \slash}\;=\G^z \partial_z +\frac{1}{\sinh H z}\prt{\G^t\del_t+e^{Ht} \G^i
\del_i}
- \frac{H\prt{d-1}}{2 \sinh H z} \G^t+\frac{d H}{2 \tanh H z} \G^z.
\end{equation}
We note in passing that
\begin{equation}
{D \hspace{-7.5pt} \slash}\;=\G^z \partial_z +\frac{1}{\sinh H z} {D \hspace{-7.5pt} \slash}\;_{\rm dS} +
\frac{d H}{2 \tanh H z} \G^z.
\end{equation}
The asymptotic behavior of the solution to the Dirac equation can be
easily worked out. It is
\begin{equation}
\psi \sim e^{-\tilde{\D}_- H z} \psi_0(x), \quad z \to \infty~,
\end{equation}
where here
\begin{equation}
{\tilde{\D}}_\pm := \frac{d}{2}\pm \mu} \def\n{\nu \G^z~.
\end{equation}
In order for this mode to be non-normalizable, we need
\begin{equation}
\G^z \psi_0 = \psi_0
\end{equation}
and $\mu} \def\n{\nu>d/2$.
The general solution of the Dirac equation can be written in the form
\eq{sola1} with the boundary spinor $\psi_0$ satisfying $\G^z \psi_0=\psi_0$.
As before we note the useful relation:
\begin{align} \label{DD-ds}
&&{D \hspace{-7.5pt} \slash}\;^2-({D \hspace{-7.5pt} \slash}\; -\mu) H \coth H z ~\G^z
= \prt{\square+\frac{d^2 H^2}{4 \tanh^2 H z}
-\frac{\prt{d-3}\prt{d+1} H^2}{4 \sinh^2 H z }} {\nonumber} \\
&&\qquad\qquad\qquad\qquad
-H \coth H z \prt{\del_z - \mu \G^z +
\frac{d H}{2 \tanh H z}}+\frac{H e^{Ht}}{\sinh^2 H z}\G^t \G^i \del_i~,
\end{align}
where
\begin{equation}
\square=\del_z^2 -\frac{1}{\sinh^2 H z}\del_t^2+
\frac{e^{2 H t}}{ \sinh^2 H z}\del_i^2+\frac{d H}{\tanh H z}\del_z
+\frac{(d-1)H}{\sinh^2 H z}\del_t ~
\end{equation}
is the d'Alembertian on AdS${}$ space with the metric \eq{metric1}.
This allows us to construct the fermion bulk to
boundary propagator $S$ as
\begin{equation} \label{S-ds}
S =\prt{{D \hspace{-7.5pt} \slash}\;+\mu} \def\n{\nu-H \coth H z~ \Gamma^z}K +\delta}\def\D{\Delta}\def\ddt{\dot\delta~,
\end{equation}
where
\begin{equation}
K = \prt{\frac{e^{-Hz}}{\rho^2}}^{\D_+}
\end{equation}
is the bulk to boundary propagator
for an auxiliary scalar field of mass $m$;
\begin{equation}
\delta}\def\D{\Delta}\def\ddt{\dot\delta:=-\prt{{D \hspace{-7.5pt} \slash}\;-\mu} \def\n{\nu}^{-1}\prt{q(z)-H \coth Hz
\prt{\del_z + H \tilde{\Delta}_-} +\frac{H e^{Ht}}{\sinh^2 Hz}\G^t
\G^i \del_i }K~,
\end{equation}
and
\begin{equation}
q(z):=\frac{d H^2}{2}\frac{\tanh H z-1}{\tanh^2 H z}-
\frac{\prt{2d+3} H^2}{4}\frac{1}{\sinh^2 H z }~.
\end{equation}
It is easy to see that one needs again \eq{pr4} in order for $S$ to
satisfy its defining differential equation. One also easily see that $\delta}\def\D{\Delta}\def\ddt{\dot\delta$
vanishes at the boundary and $S$ satisfies the desired boundary
condition.
To calculate the boundary two point function using \eq{cg2}, we
only need to know $S$ near the boundary, which is
$S = \prt{{D \hspace{-7.5pt} \slash}\; + \mu} \def\n{\nu - H \coth H z~ \Gamma^z} K $. As terms
that are proportional to the unit matrix or $\G^z$ do not
contribute to \eq{i2}, we can drop them in $S$ and we obtain again the relevant
expression \eq{cg3} for $S$. In terms of conformal coordinates,
it is
\begin{equation}\lab{sk}
S =\frac{1}{2}
e^{- H z}\, x_0 \G^\mu} \def\n{\nu \del_\mu} \def\n{\nu K -\frac{H(d-1)}{2}
\G^0 e^{- H z} K \quad\mbox{for large $z$}.
\end{equation}
Now the first term of \eq{sk} is equal to
$-\D_+ b K D$,
where $b$ is the function
$ b:= e^{-H z} / \rho$
and $D$ is the matrix
$ D := x^0 x^\mu} \def\n{\nu \del_\mu} \def\n{\nu \sigma} \def\S{\Sigma^2/(2 \sigma} \def\S{\Sigma) $.
Near the boundary, $K$ imposes $\sigma} \def\S{\Sigma =0$ and so $b=2/H$
and
\begin{equation}
D = \frac{(x-x')_\mu} \def\n{\nu \G^\mu} \def\n{\nu}{|x-x'|}.
\end{equation}
It is then
clear that the second term in \eq{sk} is sub-leading compared to the
first term. As a result, the boundary two-point function is given by
(up to a constant)
\begin{equation}
{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}\prt{x,x'}=
\frac{D}{\sigma} \def\S{\Sigma^{2\tilde{\D}_+}}~.
\end{equation}
This agrees with the result
\eq{OOD}
for a conformal field theory in dS spacetime.
\section{Discussion}
In this paper, we have proposed an AdS/CFT
correspondence of the Type IIB string theory on $AdS_5 \times S^5$ in
terms of a
superconformal field theory on the dS${}_4$ boundary. As a first step, we have
provided evidence of it
by showing that the boundary correlators
of the dS conformal field theory
can be
reproduced from the AdS bulk dynamics.
It would be interesting to compute the higher point functions and see
if there is any nonrenormalization theorem for chiral operators as in
\cite{seiberg}.
It would also be interesting to look
at minimal surfaces
with different boundary conditions
and compute the interacting potential between quarks
anti-quarks and
study properties of the entanglement entropy in a quantum field
theory in a time dependent background.
According to the holographic principle
\cite{hooft,susskind} (see also \cite{bousso} for a review),
the full description of quantum gravity in a
region requires only a quantum field theory living at the
boundary.
Gauge/gravity correspondence is a nice illustration of the
holographic principle. In the original works of \cite{witten, GKP}, a
boundary flat $R^4$ was created when the Poincare patch of the AdS
spacetime
was considered.
In this paper we considered a
different coordinate patch with a different boundary, and showed
that the same bulk dynamics (e.g. the same
IIB supergravity equations in the classical supergravity limit),
with different boundary conditions
would results in different
holograms.
This is of course
in consistent with the statement
of the holographic principle. Nevertheless this is an aspect of the
holographic principle that
has not been emphasized much so far in the literature.
In the full non-nonperturbative formulation of string theory on
$AdS_5 \times S^5$, all the
different boundaries and boundary conditions should be contained in the
moduli space of the theory itself. This implies that the different dual quantum
field theories may also be considered to be contained in some bigger
theory of quantum field theories.
On the $AdS_5 \times S^5$ side, the existence of the Lax pair and an
infinite set of
classically conserved nonlocal charges are properties of the
Green-Schwarz
string sigma model
\cite{polchinski}. Note that these properties were established
using the global AdS space without referencing to the Poincare
coordinate system. As such, the construction \cite{polchinski}
of charges also apply in
our case. One may speculate that the $\cN=4$
superconformal Yang-Mills on dS${}_4$ may also be integrable in some of
its sectors. This is an
interesting aspect that deserves further studies.
\vskip7mm
\section*{Acknowledgements}
It is our pleasure to thank Jean-Pierre Derendinger for many
useful discussion and
collaboration in the early stage of this work.
CSC also thank Koji Hashimoto and Piljin Yi for helpful
discussion and suggestion.
This work is
supported in part by the National Center of Theoretical Science
(NCTS) and the grant
104-2112-M-007-001 -MY3 of the Ministry of Science and
Technology of Taiwan.
|
\section{Introduction}\label{sec:intro}
Let $A$ be a commutative ring and $\mathcal{P}(A)$ the category of finitely-generated projective $A$-modules. Let $\End\mathcal{P}(A)$ be the category whose objects are endomorphisms $P\to P$ where $P\in\mathcal{P}(A)$, and whose morphisms $\alpha:(P\to P)\to (Q\to Q)$ are morphisms $\alpha:P\to Q$ in $\mathcal{P}(A)$ such that the diagram
\begin{equation*}
\xymatrix{
P\ar[d]\ar[r]^\alpha & Q\ar[d]\\
P\ar[r]^\alpha & Q
}
\end{equation*}
commutes. The category $\End\mathcal{P}(A)$ is naturally equivalent to the category of $A[t]$-modules that are finitely generated and projective as $A$-modules via the inclusion map $A\to A[t]$. The category $\End\mathcal{P}(A)$ is an exact category and one would like to calculate its Quillen $K$-theory groups. This type of $K$-theory calculation falls under the more general problem of calculating for a ring homomorphism $R\to S$, the $K$-theory of the category of $S$-modules that are finitely generated and projective as $R$-modules.
The calculation of $K_i(\End\mathcal{P}(A))$ was given by Kelley and Spanier \cite{KelleySpanier1968} when $A$ is a field and $i=0$. Almkvist \cite{Almkvist1978} did the calculation when $A$ is an arbitrary commutative ring and $i=0$, and when $A$ is a field and $i$ is arbitrary. To describe these computations, define the abelian group whose underlying set is
\begin{align*}
\tilde{A}_0 = \left\{ \frac{1 + a_1t + \cdots + a_nt^n}{1 + b_1t + \cdots + b_mt^m} : a_i,b_j\in A\right \}
\end{align*}
and whose binary operation is given by the usual multiplication of rational functions. If $f:M\to M$ is an endormorphism of a finitely-generated projective $A$-module, then the characteristic polynomial $\lambda_t(f)$ may be defined by extending $f$ to an endomorphism of a free module, and then defining $\lambda_t(f) = \det(1 + tf)$; see \cite{Almkvist1978} for an alternative definition. We can use this map to describe the results of Kelley and Spanier and Almkvist:
\begin{thm}There is an isomorphism
\begin{align*}
K_0(\End\mathcal{P}(A))&\longrightarrow K_0(A) \times \tilde{A}_0\\
\end{align*}
given on generators by
\begin{align*}
[M\xrightarrow{f} M] &\longmapsto ([M], \lambda_t(f)).
\end{align*}
\end{thm}
In this paper, we generalise this result to the case of finitely many commuting endomorphisms and where $A=k$ is a field:
\begin{thm}
The algebraic $K$-groups of the exact category $\End_n\mathcal{P}(k)$ are given by
\begin{align}
K_i(\End_n\mathcal{P}(k)) \cong \bigoplus_{M} K_i(k[t_1,\dots,t_n]/M)
\end{align}
where $M$ ranges over all the maximal ideals of the polynomial ring $k[t_1,\dots,t_n]$.
\end{thm}
We remark that our our proof is different than the proofs in \cite{Almkvist1978} and in \cite{KelleySpanier1968}. One feature is that our result for $i=0$ is not phrased in terms of a characteristic polynomial map, which is an advantage in the sense that it is more explicit in terms of the structure of the $K$-groups, but a disadvantage in the sense that working with products is more difficult. The result for $n=1$ and higher $K$-theory is \cite[Theorem 5.2]{Almkvist1978}, but again, our proof is entirely different.
\section{The Category $\End_n\mathcal{P}(k)$}
Let $\End_n\mathcal{P}(k)$ denote the exact category of $k[T]:=k[t_1,\dots,t_n]$-modules that are finitely generated as $k$-modules. In this section we calculate $K_i(\End_n\mathcal{P}(k))$. A finite-dimensional $k$-vector space $V$ with any $n$ commuting endomorphisms $f_1,f_2,\dots,f_n$ of $V$ may also be considered as a $k[t_1,\dots,t_n]/I$ module where $I$ is the kernel of the map $k[t_1,\dots,t_n]\to k[f_1,\dots,f_n]$ given by $t_i\mapsto f_i$; hence:
\begin{prop}
The category $\End_n\mathcal{P}(k)$ is naturally equivalent to the filtered direct limit
\begin{align}\label{eqn:dirlimit}
\varinjlim_I \left(\fgmod{k[T]/I}\right)
\end{align}
of exact categories, where $I$ runs over the set of ideals of $k[T]$ such that the quotient $k[T]/I$ is finite-dimensional over $k$, and if $I\subseteq J$ are two such ideals, then the functor $\fgmod{k[T]/J}\to\fgmod{k[T]/I}$ is the forgetful functor induced by the quotient map $k[T]/I\to k[T]/J$.
\end{prop}
\begin{proof}
The limit is indeed filtered, since $k[T]/(I\cap J)$ is finite-dimensional whenever $k[T]/I$ and $k[T]/J$ are finite-dimensional. Define a functor
\begin{align*}
F:\varinjlim_I \left(\fgmod{k[T]/I}\right)\to\End_n\mathcal{P}(A)
\end{align*}
as follows. Any element in $\varinjlim_I \left(\fgmod{k[T]/I}\right)$ is represented by $V\in\fgmod{k[T]/I}$ for some $I$. We let $F(V)$ to be the $k[T]$ module given by the forgetful functor induced by the map $k[T]\to k[T]/I$. This functor is well-defined because $k[T]/I$ is finite-dimensional, and it is easy to see that $F$ gives the required natural equivalence.
\end{proof}
Using this observation and the result that taking $K$-groups of exact categories commutes with filtered direct limits \cite[\S2]{Quillen1972}, we obtain the following corollary.
\begin{cor}
The $K$-theory of the category $\End_n\mathcal{P}(A)$ may be calculated as the direct limit
\begin{align}
K_i(\End_n\mathcal{P}(A)) \cong \varinjlim_I K_i\left(\fgmod{k[T]/I}\right)
\end{align}
\end{cor}
For any ring $R$, let $\Jac(R)$ denote the Jacobson radical of $R$. Any surjective ring homomorphism $R\to S$ induces a surjective ring homomorphism $R/\Jac(R)\to S/\Jac(S)$, and a commutative diagram of rings
\begin{equation*}
\xymatrix{
R\ar[r]\ar[d] & S\ar[d]\\
R/\Jac(R)\ar[r] & S/\Jac(S)
}
\end{equation*}
In turn, whenever $S$ is finitely generated as an $R$-module, we have a commutative diagram of forgetful functors
\begin{equation}\label{eqn:forgetful}
\xymatrix{
\fgmod{R} & \fgmod{S}\ar[l]\\
\fgmod{R/\Jac(R)}\ar[u] & \fgmod{S/\Jac(S)}\ar[l]\ar[u]
}
\end{equation}
In particular, this applies to the ring homomorphisms $k[T]/I\to k[T]/J$ for $I\subseteq J$ appearing in the direct limit of~\eqref{eqn:dirlimit}.
\begin{prop}\label{thm:lastreduction}
The natural transformation of the direct limits of categories induced by the forgetful functors as in \eqref{eqn:forgetful} for $R = k[T]/I$ induces an isomorphism algebraic $K$-groups
\begin{align*}
\varinjlim_I K_i\left( \fgmod{\frac{k[T]/I}{\Jac(k[T]/I)}}\right)\overset{\sim}{\longrightarrow} \varinjlim_I K_i\left(\fgmod{k[T]/I}\right).
\end{align*}
\end{prop}
\begin{proof}
For any Artinian ring $R$, the Jacobson radical $\Jac(R)$ is nilpotent (e.g. \cite[Theorem 4.12]{Lam1991}), and so devissage \cite[\S5, Theorem 4]{Quillen1972} shows that the inclusion $\fgmod{R/\Jac(R)}\to \fgmod{R}$ induces an isomorphism
\begin{align*}
K_i(\fgmod{R/\Jac(R)})\to K_i(\fgmod{R})
\end{align*}
(see \cite[Page 439]{Weibel2013} for more details). In particular, this applies to $R = k[T]/I$ where $k[T]/I$ is finite-dimensional over $k$.
\end{proof}
\begin{thm}\label{thm:maintheorem}
The algebraic $K$-groups of the exact category $\End_n\mathcal{P}(k)$ are given by
\begin{align}
K_i(\End_n\mathcal{P}(k)) \cong \bigoplus_{M} K_i(k[t_1,\dots,t_n]/M)
\end{align}
where $M$ ranges over all the maximal ideals of the polynomial ring $k[t_1,\dots,t_n]$.
\end{thm}
\begin{proof}
We must calculate the limit
\begin{align}\label{eqn:finaldirectlimit}
\varinjlim_I K_i\left( \fgmod{\frac{k[T]/I}{\Jac(k[T]/I)}}\right)
\end{align}
given in Proposition~\ref{thm:lastreduction}. So, fix an ideal $I$ such that $k[T]/I$ is finite-dimensional. Then there are finitely many maximal ideals $M_1,\dots,M_k$ of $k[T]$ that contain $I$ and
\begin{align*}
k[T]/I \cong \bigoplus_{j=1}^k (k[T]/I)_{M_j}
\end{align*}
via the obvious map (e.g. \cite[Theorem 5.20]{Goertz2010}). Here, by $(k[T]/I)_{M_i}$, we abuse notation and mean the localization of $k[T]/I$ away from the maximal ideal $M_i(k[T]/I)$. If $J$ is an ideal containing $I$, then there is a subset $\{N_1,\dots,N_\ell\}$ of the maximal ideals $\{M_1,\dots,M_k\}$ that contain $J$ and the quotient homomorphism $k[T]/I\to k[T]/J$ induces the map
\begin{align*}
\bigoplus_{j=1}^k (k[T]/I)_{M_j}\longrightarrow\bigoplus_{j=1}^\ell (k[T]/J)_{N_j}
\end{align*}
which must be the projection map. The induced map on $K$-groups that fits into the direct limit in~\eqref{eqn:finaldirectlimit}, by is then the map
\begin{align*}
\bigoplus_{j=1}^\ell K_i(k[T]/N_j)\longrightarrow\bigoplus_{j=1}^k K_i(k[T]/M_j).
\end{align*}
given by the sum of the inclusion maps. Taking the direct limit gives the stated result, once we note that $k[T]/M$ is finite-dimensional over $k$ for any maximal ideal $M$.
\end{proof}
In particular, we obtain the following amusing, possibly well-known result.
\begin{cor}
For a field $k$, there is an isomorphism of abelian groups
\begin{align*}
\tilde{k}_0 := \left\{ \frac{1 + a_1t + \cdots + a_nt^n}{1 + b_1t + \cdots + b_mt^m} : a_i,b_j\in k\right \} \cong \bigoplus_{|\mathrm{Spec}(k[T])| - 1}\mathbb{Z}.
\end{align*}
\end{cor}
\bibliographystyle{alpha}
|
\section{Introduction}\label{sec:INTRO}
Over the past decades, evidence has accumulated for the existence of massive black holes (MBHs) in the centre of most, possibly all, galaxies \citep[e.g.][]{KormendyRichstone1995, RichstoneEtAl1998, MagorrianEtAl1998}. These objects are thought to be the engine behind the extremely luminous quasars observed in the early universe. To explain these observations, the MBHs must have accreted material at a rate of $\sim 1\,\mathrm{M}_{\odot}\,\mathrm{yr}^{-1}$ in order to reach their inferred masses of up to $\sim 10^{10}\,\mathrm{M}_{\odot}$. It is generally believed that gas infall from large scales provides the necessary feeding for those MBHs to outshine their host galaxies and reach present day masses \citep[e.g.][]{Soltan1982, HaehneltRees1993, YuTremaine2002}, other models \citep{MK2005} argue that the necessary fuel is provided by stars captured by a gas disc as they plunge through it, and are then destroyed and accreted by the MBH.
Both scenarios require the presence of a gas disc. The idea that emission from a thin accretion disc (AD) is the source of quasar luminosity is consistent with the observed quasar spectral energy distribution in the optical and UV \citep{Shields1978, SunMalkan1989, Laor1990}. The disc forms as gas infalls into the central region of the galaxy and initially loses energy and settles into a rotating flattened structure. Angular momentum is gradually transported outwards by various processes, allowing matter to progressively move inwards, towards the centre of gravity. While falling deeper into the gravitational well, viscous processes heat up the now ionized gas and cause yet more energy and angular momentum to be thermally radiated out; this process also cools the disc leading to a steady state \citep{AbramowiczStraub2014}. When a gas parcel eventually reaches a region in the disc where a circular orbit is no longer possible due to General Relativistic effects, it will then plunge into the MBH. A model for such a stationary AD was developed by \citet{SS1973}, and this model has been the basis of many other models. The stationary AD model was extended by \citet{NT1973} to include General Relativistic effects, which are critically important close to the inner boundary of the AD. It is important to note that such an AD may extend to the parsec scale or larger, where the environment is dominated by stars, orders of magnitude more than the MBH's gravitational (Schwarzschild) radius.
When a gas disc is not present, a MBH does not normally emit any detectable electromagnetic radiation. However, when a star occasionally falls in from the surrounding cluster close enough to be disrupted, the material then heats up by dissipative processes and radiates some of its binding energy, further falling further into the MBH. This process is only possible when the tidal disruption radius (which depends on the stellar radius and the MBH mass) is larger than the Schwarzschild radius, which is the case for a Sun-like star when $M_\mathrm{bh} \lesssim 10^8\,\mathrm{M}_{\odot}$. Because of the great Keplerian speed so deep in the MBH's potential well, the gaseous debris is hot enough to produce X-rays. Since these tidal disruption events (TDEs) \citep{Hills1975, Hills1976, FR1976} happen close to the MBH event horizon, they help constrain some of the MBH's parameters such as mass and spin. Tidal disruption of stars has famously been used as a boundary condition on the Boltzmann equation to determine the stellar mass distribution around MBHs \citep{BahcallWolf1976, BahcallWolf1977, DokuchaevOzernoi1977a, DokuchaevOzernoi1977b}, following an equivalent idea in plasma physics \citep{Gurevich1964}.
In the past decade, some TDE candidate events have been published \citep{Komossa2002, KomossaMerritt2008, WangEtAl2012}, with potentially many more hidden in data archives. There is only a handful of observations simply because such events are rare and often very distant \citep{Komossa2002}. \citet{MagorrianTremaine1999} calculated stellar TDE rates in simple galaxy models using loss-cone theory and predicted a rate of no more than about one event in $10^4$ years per galaxy (largely in agreement with \citealt{SyerUlmer1999}, who used similar methodology); this value is increased by 1--2 orders of magnitude when using a more recent MBH mass distribution and is roughly consistent with the inferred rate of soft X-ray outbursts from normally quiescent galactic nuclei \citep{DonleyEtAl2002}.
The loss-cone (also referred to as loss-wedge or loss-cylinder) is defined as the region in angular momentum and energy (or radius) phase space where stars have orbits that bring them inside their tidal disruption radius at pericentre. This region is quickly emptied on a time-scale of the orbital crossing time as stars directly plunge onto the MBH. The analysis of \citet{MagorrianTremaine1999} assumed the loss-cone is repopulated by 2-body relaxation and large-scale torques; but other processes can keep the loss-cone full for a longer time (1) accelerating angular momentum diffusion (reducing the effective relaxation time) and (2) enlarging the effective size of the loss-cone. For example, \citet{PeretsEtAl2007} used an extension of loss-cone theory and found that massive perturbers were far more efficient than 2-body relaxation in bringing stars within the zone of influence of the MBH. The presence of a second MBH in the galactic nucleus can potentially increase the rate dramatically \citep{ChenEtAl2011} or under some circumstances decrease it \citep{MerrittWang2005}. Likewise, rotation of the star cluster around the MBH \citep{FiestasSpurzem2010, FiestasEtAl2012} as well as the recoil of a post-coalescence MBH from a galactic nucleus \citep{GualandrisMerritt2009, LiEtAl2012} can have a dramatic influence on the TDE rate as compared to the simple loss-cone estimation. Finally, \citet[hereafter Paper I]{JustEtAl2012} explored the influence of both the gaseous disc and star cluster centred on the MBH on the TDE rates.
Research into the interaction of stars with an AD goes back to \citet{VilkoviskijBekbasarov1981} and \citet{Vilkoviskij1983}, who realised that the gas density at the inner regions was potentially high enough to induce substantial drag on stars crossing it, changing their orbits. More detailed studies that followed in the 1990s aimed to find an equilibrium state between the transport of stars into the central regions and their removal from the AD \citep{SyerEtAl1991, ArtymowiczEtAl1993, VokrouhlickyKaras1998, SubrKaras1999}. \citet{VilkoviskijCzerny2002} introduced a semi-analytic model based on the balance between relaxation and dissipation processes, noting in particular that 2-body relaxation both supplies the AD with stars and also has the opposite effect of elevating back previously trapped stars. \citet{SubrEtAl2004} used similar methodology with a more realistic disc model and found that the presence of an AD causes significant anisotropies such as depletion of counter-rotating stars and flattening of the stellar system. More progress was made by \citet{Rauch1995, Rauch1999} who for the first time applied numerical simulations to the problem. Notably, General Relativistic effects were included in his work. He showed that the orbits of stars quickly align with the AD's plane, as well as that a constant density core forms due to collisional effects.
These studies all made simplifying assumptions about the AD, adopting infinitely-thin static disc models (no feedback from stars), and about the star cluster, which was often assumed to be spherically symmetric and to start in equilibrium (neglecting relevant time-scales). {\citetalias{JustEtAl2012}} have relaxed some of these assumptions by conducting, for the first time, self-consistent direct \textit{N}-body simulations to investigate the interaction of a central star cluster surrounding a MBH with an AD. They used a stationary Keplerian rotating disc model (which was vertically extended rather than infinitely-thin) and resolved the dissipation of kinetic energy of stars passing through it (ram pressure effects), as well as star accretion by the MBH and 2-body relaxation. They showed that the rate of star capture by the AD and subsequent accretion onto the MBH was enhanced by a significant factor compared to the case where no disc was included. They also showed that, as predicted by \citet{VilkoviskijCzerny2002}, there is a stationary flow of stars within the disc toward the central MBH, which is determined by an equilibrium between diffusion by 2-body encounters and energy loss by the dissipative force from the AD.
In this work, we extend the models of \citetalias{JustEtAl2012}. Our new models are significantly more detailed, with larger number of particles (up to an order of magnitude), higher resolution in the central region (artificial accretion radius decreased by up to two orders of magnitude) and a more physical disc model (with disc thickness varying with radius rather than constant). This new simulation suite allows us to explore a large range of the free parameters, examine the statistical properties of the orbits of plunging stars and the relevant time-scales, as well as for the first time make an accurate scaling analysis and extrapolate our results to real galactic centers.
This paper is organised as follows: in Section~\ref{sec:method} we present the method for modelling star--star and star--gas interactions as well as the initial conditions for the star cluster and AD; in Section~\ref{sec:plunge} we closely examine the types of stellar orbits that result in accretion or tidal disruption (based on the distribution of final orbital eccentricity of accreted stars and mass growth rate of the MBH); in Section~\ref{sec:RES} we investigate the dependence on these on the simulation particle number resolution and accretion radius spatial resolution; in Section~\ref{sec:hrdisc} the effect of the structure of the gas disc is investigated by changing the disc height, while keeping the surface density profile fixed; in Section~\ref{sec:realgc} we discuss the scaling analysis of our simulation results and applicability to real galactic centres; finally, in Section~\ref{sec:CON} we present a discussion of our key results and conclusions.
\section{Methods}
\label{sec:method}
In this paper we present a suite of 39 new simulations, spanning a factor of 16 in particle number $N$ and up to $N=1.28\times10^5$ particles; all models are listed in Table~\ref{ModelParas}. This number is still quite small compared to the number of stars in a real galaxy centre, our \textit{star particles} are thus ``superparticles'' representing a group of stars. The computation times varied greatly between models, with one of the longest runs (\texttt{128k03r}) taking 180 days to complete two relaxation times ($\sim 1800$ H\'enon time units\footnote{\citet{Heggie2014} suggested naming the standard \textit{N}-body unit system (previously known just as \textit{N}-body units) after M. H\'enon to commemorate his work; we use this throughout this study. In this unit set, the total stellar mass $M_\mathrm{cl}$ and the gravitational constant $G$ are set to unity, and the total energy of the isolated stellar system is set to $-0.25$. For convenience, we will drop the unit name when unambiguous.}) using the $\phi$\textsc{grape} code (see Section~\ref{sec:Stellar_component}) on 8 nodes, each equipped with one NVIDIA Tesla M2050 GPU.
\begin{table}
\caption{Model parameters for all simulation runs}
\label{ModelParas}
\begin{centering}
\begin{tabular}{lrrll}
\hline
\hline
Model name & $N$ & $r_\mathrm{acc}^*$ & $\epsilon_\mathrm{bh}$ & AD type\\
\hline
\texttt{008k01ng} & 8k & 1 & $10^{-6}$ & no gas\\
\texttt{008k03ng} & 8k & 3 & $10^{-6}$ & no gas\\
\texttt{008k03ns} & 8k & 3 & $0$ & $h_z$\\
\texttt{008k03} & 8k & 3 & $10^{-6}$ & $h_z$\\
\texttt{008k03rns} & 8k & 3 & $0$ & $h(R)$\\
\texttt{008k03r} & 8k & 3 & $10^{-6}$ & $h(R)$\\
\texttt{008k10ng} & 8k & 10 & $10^{-6}$ & no gas\\
\texttt{008k10} & 8k & 10 & $10^{-6}$ & $h_z$\\
\texttt{008k50ng} & 8k & 50 & $10^{-6}$ & no gas\\
\texttt{008k50} & 8k & 50 & $10^{-6}$ & $h_z$\\
\texttt{016k01ng} & 16k & 1 & $10^{-6}$ & no gas\\
\texttt{016k03ng} & 16k & 3 & $10^{-6}$ & no gas\\
\texttt{016k03} & 16k & 3 & $10^{-6}$ & $h_z$\\
\texttt{016k10ng} & 16k & 10 & $10^{-6}$ & no gas\\
\texttt{016k10} & 16k & 10 & $10^{-6}$ & $h_z$\\
\texttt{016k50ng} & 16k & 50 & $10^{-6}$ & no gas\\
\texttt{016k50} & 16k & 50 & $10^{-6}$ & $h_z$\\
\texttt{032k01ng} & 32k & 1 & $10^{-6}$ & no gas\\
\texttt{032k03ng} & 32k & 3 & $10^{-6}$ & no gas\\
\texttt{032k03ns} & 32k & 3 & $0$ & $h_z$\\
\texttt{032k03} & 32k & 3 & $10^{-6}$ & $h_z$\\
\texttt{032k03rns} & 32k & 3 & $0$ & $h(R)$\\
\texttt{032k03r} & 32k & 3 & $10^{-6}$ & $h(R)$\\
\texttt{032k10ng} & 32k & 10 & $10^{-6}$ & no gas\\
\texttt{032k10} & 32k & 10 & $10^{-6}$ & $h_z$\\
\texttt{032k50ng} & 32k & 50 & $10^{-6}$ & no gas\\
\texttt{032k50} & 32k & 50 & $10^{-6}$ & $h_z$\\
\texttt{064k01ng} & 64k & 1 & $10^{-6}$ & no gas\\
\texttt{064k03ng} & 64k & 3 & $10^{-6}$ & no gas\\
\texttt{064k10ng} & 64k & 10 & $10^{-6}$ & no gas\\
\texttt{064k50ng} & 64k & 50 & $10^{-6}$ & no gas\\
\texttt{128k01ng} & 128k & 1 & $10^{-6}$ & no gas\\
\texttt{128k03ng} & 128k & 3 & $10^{-6}$ & no gas\\
\texttt{128k03} & 128k & 3 & $10^{-6}$ & $h_z$\\
\texttt{128k03r} & 128k & 3 & $10^{-6}$ & $h(R)$\\
\texttt{128k10ng} & 128k & 10 & $10^{-6}$ & no gas\\
\texttt{128k10} & 128k & 10 & $10^{-6}$ & $h_z$\\
\texttt{128k50ng} & 128k & 50 & $10^{-6}$ & no gas\\
\texttt{128k50} & 128k & 50 & $10^{-6}$ & $h_z$\\
\hline
\end{tabular}
\end{centering}
\par\medskip
\textbf{Notes.} All models completed at least two relaxation times, with many completing nine. Column 2 gives the number of particles $N$ used in each model where k stands for $10^3$ particles; Column 3 gives the spatial resolution at the centre, where $r_\mathrm{acc}^*$ is a shorthand for convenience and represents the numerical accretion radius in units of $10^{-4} R_\mathrm{d}$ (where $R_\mathrm{d}$ is the scale of the disc); Column 4 gives the softening length in H\'enon units of the star--MBH force or zero when there was no softening (\texttt{ns}), the star--star softening was $10^{-4}$ in all models; Column 5 gives the particular type of AD used for the mode, where $h_z$ means constant thickness, $h(R)$ means varying thickness as described in the text, and `no gas' are pure stellar dynamical models with no AD.
\end{table}
The spatial resolution is characterised by the \textit{numerical} accretion radius $r_\mathrm{acc}$, which is the distance from the MBH at which particles are said to be accreted by the MBH. When this happens in the simulation, the star particle disappears and its mass is added to the MBH. The value of $r_\mathrm{acc}$ is chosen as a compromise between physical realism and computational cost. While we improved the spatial resolution significantly from {\citetalias{JustEtAl2012}} by up to two orders of magnitude (the lowest resolution in this study was the highest in theirs), the smallest value of $r_\mathrm{acc}$ used here, when scaled to real galaxies, is still 2--3 orders of magnitudes larger than the tidal radius for a Sun-like star (the motivation here is to ensure that the $\phi$\textsc{grape} code can treat star accretion reliably). In Section~\ref{sec:realgc} we discuss the implications of this choice. For convenience we define the quantity $r_\mathrm{acc}^* \equiv r_\mathrm{acc}/(10^{-4}R_\mathrm{d})$, where $R_\mathrm{d}$ is the scale of the disc as explained below.
In all models listed in Table~\ref{ModelParas}, the star cluster mass $M_\mathrm{cl}$ is set to unity. The MBH is fixed at the centre or the coordinate system and its mass is $M_\mathrm{bh}=0.1$.
In the following subsections, we give a brief summary of the methods used in {\citetalias{JustEtAl2012}}, focusing on the changes made in the $\phi$\textsc{grape} code since then in terms of the disc model, stellar component and the interaction between the stars and the disc.
\subsection{Accretion disc component}\label{sec:discmodel}
We have two models for the AD, denoted $h_z$ (constant thickness) and $h(R)$ (varying thickness); both correspond to an axisymmetric thin disc inspired by \citet{SS1973} and \citet{NT1973}. The gas density is given by
\begin{eqnarray}
\rho_\mathrm{g}(R,z) & = & \frac{2-p}{2\pi \sqrt{2 \pi}} \frac{M_\mathrm{d}}{h R_\mathrm{d}^2} \left( \frac{R}{R_\mathrm{d}} \right)^{-p}\nonumber\\
& & \exp \left[ -\beta_s \left( \frac{R}{R_\mathrm{d}} \right)^s \right] \exp \left( \frac{-z^2}{2 h^2} \right),
\label{gasdensity}
\end{eqnarray}
where $p=3/4$ is the structure power-law corresponding to the outer region of standard thin disc model, $R$ is the radial distance from the MBH, $z$ is the vertical distance from the disc plane, $R_\mathrm{d} = 0.22$ is the radial extent of the disc; its value is determined such that the enclosed stellar mass within $R_\mathrm{d}$ equals the MBH mass at the beginning of the simulation (after letting the stellar system relax for $\sim 5$ crossing times in the MBH potential in order to get a dynamical equilibrium). The parameters $s=4$ and $\beta_s=0.7$ are associated with the smoothness of the outer cutoff of the disc (introduced for numerical reasons) and $h$ is the scale height (note that {\citetalias{JustEtAl2012}} defined $h$ differently, as a dimensionless ratio). The total disc mass is fixed to be $M_\mathrm{d} = 0.01$, with the numerical minimum and maximum radii given by $r_\mathrm{acc}$ and 1, respectively (but the density drops sharply around $R=R_\mathrm{d}$). The gas in the disc is taken to have a Keplerian rotation profile.
The two disc models differ by the vertical disc structure. The first model, denoted by $h_z$, has a constant scale height $h_z = 10^{-3} R_\mathrm{d}$. This corresponds to a self-gravitating isothermal profile, which is sufficiently accurate for large radii, and was thus adopted by {\citetalias{JustEtAl2012}} as they could not resolve the innermost part of the disc.
The second disc model modifies the former by using a simple linear relation for the scale height in the inner region. This is based on the fact that an AD in hydrostatic equilibrium with in the MBH gravitational field curves like $R^{9/8}$, which is reasonably similar to constant slope. The disc height is thus proportional to $R$ up to a radius where the disc is vertically self-gravitating at $R = R_\mathrm{sg}$, beyond which it is given by $h=h_z$, i.e.
\begin{equation}
h(R)=\begin{cases}
\frac{R}{R_\mathrm{sg}}h_z & R \leq R_\mathrm{sg}\\
h_z & R > R_\mathrm{sg}.
\end{cases}\label{eq:hR}
\end{equation}
The transition between the two regions can be estimated by equating the vertical component of the spherically symmetric force from the MBH at $z=h_z$ with the vertical self-gravitation of a thin disc above the AD:
\begin{equation}
\frac{GM_\mathrm{bh}h_z}{R_\mathrm{sg}^3} = 2\pi G \Sigma(R_\mathrm{sg}),\label{eq:find_Rsg}
\end{equation}
where $\Sigma(R)$ is the surface density of the disc which is the integral $\mathrm{d}z$ of equation~(\ref{gasdensity}); as only the last exponent is a function of $z$, the result is a simple power-law in $R$ with the same index $p$ and an exponential cutoff.
We note that both sides of equation~(\ref{eq:find_Rsg}) are approximate; we are only interested in an order of magnitude estimate of $R_\mathrm{sg}$ since the model for the AD structure is very rough anyway. On the left hand side, we only keep the first order term in a Taylor expansion of the vertical force component, which is linear in $z$ (the exact result is only $\sim 15$ per cent smaller). On the right hand side, the expression is the standard thin disc approximation solving the vertical Poisson equation while neglecting the radial derivatives \citep[for more information see][]{BinneyTremaine2008}. The solution is
\begin{equation}
R_\mathrm{sg}^{3-p} = \frac{1}{2-p} \frac{M_\mathrm{bh}}{M_\mathrm{d}} h_z R_\mathrm{d}^{2-p} .\label{rsg}
\end{equation}
For our choice of parameters, we get $R_\mathrm{sg} \approx 0.026$ which gives half-opening angle of $0.5^\circ$. The effects of changing the disc height profile on the results are examined in Section~\ref{sec:hrdisc}.
\subsection{Stellar component}
\label{sec:Stellar_component}
The stellar cluster initially had \citet{Plummer1911} density profile; it was allowed to dynamically evolve in the presence of the MBH for $\sim 5$ crossing times before $t=0$ and the AD is added. The numerical integration is done with a version of the $\phi$\textsc{grape} code \citep{HarfstEtAl2007} which has been modified to include the effect of an AD and a central MBH. It is a direct \textit{N}-body code based on the fourth-order Hermite integration scheme. It is a parallel code that also uses accelerator technology such as GPUs to calculate mutual gravitational interactions between star particles. The $\phi$\textsc{grape} code has been used in {\citetalias{JustEtAl2012}} as well as many other publications about galactic centres \citep[e.g.][and references thereafter]{BerczikEtAl2005}.
The $\phi$\textsc{grape} code suits this investigation due to its computational speed and simplicity, which allows it to be easily modified. The version used here does not include stellar evolution, star formation, nor a mass function. There is no special treatment of two (or more) body encounters; i.e. no regularisation for close encounters. Neglecting these factors is justified here since the particle resolution is not high enough to model individual stars.
Indeed, a star particle does not actually represent a single star, but rather a group of stars. The number of stars in such a so-called \textit{superparticle} varies wildly between models and the galaxies they are scaled to; in the best case (models with largest $N$) ranges between a few hundreds (small galaxies) to hundreds of thousands (giant galaxies). This means that softening of the gravitational force is needed between the star particles, and we used $\epsilon = 10^{-4}$ in all models.
An additional softening of the same form is used in the interaction between a star particle and the MBH; as seen in Table~\ref{ModelParas}, most of the models use $\epsilon_\mathrm{bh} = 10^{-6}$. This value is not expected to significantly affect the results as the softening is smaller than all $r_\mathrm{acc}$ values. Using softening for the MBH interaction does complicate the orbital elements for the stars at the point of accretion. By correcting the MBH-star distance by $|\boldsymbol{r}|^2 \rightarrow |\boldsymbol{r}|^2+\epsilon_\mathrm{bh}^2$ in the angular momentum and binding energy, the orbital elements can be accurately computed. For models with $\epsilon_\mathrm{bh}>0$ the corrected values for eccentricity and semi-major axis are used in all presented results. The effect of softening on the final orbits of accreted stars and the growth of the MBH was tested by setting $\epsilon_\mathrm{bh}=0$ in four of the models in Table~\ref{ModelParas}. We examined the effects of softening length and saw very little difference between the softened and unsoftened models. The eccentricity distribution is slightly different, with stars being accreted on more eccentric orbits for the unsoftened models.
\subsection{Star--disc interaction}
\label{sec:SDI}
As a star particle passes through the accretion disc, a dissipative force is applied against its motion. Once its orbit lies within the disc (meaning that the apocentre $< R_\mathrm{d}$ and the highest point of the orbit above the plane of the disk $< h_z$), the particle is considered to be captured into the disc. If the distance of a star particle to the MBH comes within $r_\mathrm{acc}$, it is considered accreted by the MBH and is removed from the simulation.
The dissipative force that an individual star experiences while crossing the disc comes from ram pressure and originates at the bow shock in front of the star. It is proportional to the geometric cross section and local gas density, and is quadratic in velocity. It is given by the drag equation:
\begin{equation}
\boldsymbol{F}_\mathrm{drag} = - Q_\mathrm{d} \pi r_\star^2 \rho_\mathrm{g}(R,z) \left| \boldsymbol{V}_\mathrm{rel} \right| \boldsymbol{V}_\mathrm{rel},\label{eq:fdrag_star}
\end{equation}
where $Q_\mathrm{d}$ is the drag coefficient, $\rho_\mathrm{g}$ is the local gas density (equation \ref{gasdensity}), $r_\star^2$ is the stellar radius and $\boldsymbol{V}_\mathrm{rel}$ is the relative velocity between the star and the gas. The drag coefficient $Q_\mathrm{d}$ incorporates much of the uncertainty and can be estimated from the shock conditions \citep{CourantFriedrichs1948}; we use $Q_\mathrm{d}=5$. This prescription is only appropriate when the star is supersonic with respect to the disc and ignores gravitational focusing; see section 2.2 in {\citetalias{JustEtAl2012}} for more details.
Since a star particle does not represent a single Sun-like star but rather a group of stars, equation~(\ref{eq:fdrag_star}) has to be properly scaled in order to be used in a simulation. We do this by introducing an effective dissipative parameter,
\begin{equation}
Q_\mathrm{tot}(N) \equiv Q_\mathrm{d} N \left( \frac{r_\star}{R_\mathrm{d}}\right)^2.\label{Qtotdef}
\end{equation}
Substituting into equation~(\ref{eq:fdrag_star}), we get
\begin{equation}
\boldsymbol{a}_\mathrm{drag} = - Q_\mathrm{tot} \frac{\pi R_\mathrm{d}^2 \rho_\mathrm{g}(R,z)}{M_\mathrm{cl}} \left| \boldsymbol{V}_\mathrm{rel} \right| \boldsymbol{V}_\mathrm{rel}.\label{fdrag}
\end{equation}
Note that equation~(\ref{fdrag}) contains only \textit{global} quantities related to the entire cluster, and thus easily expressed in H\'enon units. Despite that, $Q_\mathrm{tot}$ is still very much a `real world' quantity rather than a simulation one (i.e. the $N$ in equation~\ref{Qtotdef} is $N_\mathrm{real}$, the number of stars in a galactic centre, not the number $N_\mathrm{sim}$ of particles in the model). Before we can use equation~(\ref{fdrag}) in the simulation, $Q_\mathrm{tot}$ has to be treated carefully because it introduces a new physical time-scale.
The main concern when modelling a system with superparticles rather than particles representing individual stars is that the relaxation time (which depends on the number of particles) is much shorter, while the dynamical time (only depends on the mass) is the same. The following expression is the relaxation time at the half mass radius \citep{Spitzer1987}:
\begin{eqnarray}
t_\mathrm{rx}(N) & = & \frac{0.14 N}{\ln \left( 0.4 N \right)} t_\mathrm{dyn}\label{trelax}\\
t_\mathrm{dyn} & = & \left( \frac{r_\mathrm{h}^3}{G M_\mathrm{cl}} \right)^{1/2}.\label{tdyn}
\end{eqnarray}
As the half-mass radius at the beginning of the simulation (after letting the system evolve with no AD for 5 crossing times) is measured to be $r_\mathrm{h} = 0.66$, the dynamical time of the system is $\sim 0.54$ H\'enon time units, and the relaxation time ranges between 74 ($N=8$k) to 886 ($N=128$k) time units.
The way to deal with the fact that $t_\mathrm{rx}$ depends on the number of particles is to present the results in terms of relaxation times, and then scale to physical time units when comparing with a real galactic centre. However, this cannot work if there are more physical processes associated with their own time-scales. The accretion rate of stars to the MBH is a secular process and is driven by two physical mechanisms: relaxation and dissipation. Two-body relaxation scatters stars into the empty part of the loss cone, which are then accreted quickly on high eccentricity orbits, its characteristic timescale increases with particle number $N$ (equation \ref{trelax}). The second process is energy loss by the drag force with the gas disk, which depends on the strength of the drag force at each crossing and the inverse of the orbital time (orbital frequency). It scales with the dissipation timescale as described in {\citetalias{JustEtAl2012}} and is independent of the particle number.
In order to correctly reproduce the accretion process, we need to remove the different $N$-dependences of the two mechanisms, i.e. we need to hold the ratio of dissipation and relaxation time constant. We achieve this by rescaling the
parameter $Q_\textrm{tot}$ as function of $N_\textrm{sim}$ particles by
\begin{equation}
Q_\mathrm{tot}(N_\mathrm{sim}) = \frac{t_\mathrm{rx}(N_\mathrm{real})}{t_\mathrm{rx}(N_\mathrm{sim})} Q_\mathrm{tot}(N_\mathrm{real}),
\label{QtotNrel}
\end{equation}
where $N_\mathrm{real}$ is the number of stars in a galactic centre with and an effective dissipative parameter $Q_\mathrm{tot}(N_\mathrm{real})$. In order to calibrate all our models using equation~(\ref{QtotNrel}), some real world values, i.e. $N_\mathrm{real}$ and $Q_\mathrm{tot}(N_\mathrm{real})$, have to be chosen. If we consider an AD of $R_\mathrm{d}=10\,\mathrm{pc}$ surrounded by approximately $2 \times 10^9$ Sun-like stars, equation~(\ref{Qtotdef}) gives $Q_\mathrm{tot} \approx 5\times10^{-8}$. Using a combination of equations~(\ref{QtotNrel}) and (\ref{trelax}) to scale this to arbitrary $N$, we set
\begin{equation}
Q_\mathrm{tot}(N) \approx 5.42 \ln(0.4N)/N.
\end{equation}
This method of treating the star--disc interaction does not include: feedback onto the disc from stellar crossings; gravity force from the disc mass, acting on either the stars or the gas; and finally the contribution of stellar winds to the gas disc via mass or energy. These effects will be explored in future work.
\section{Star plunging}
\label{sec:plunge}
\subsection{Plunge types}
\label{starplunge}
The drag force due to the gas disc (equation~\ref{fdrag}) will act to dissipate the orbital binding energy of a star particle. This will shrink the orbit and slowly align it to the plane of the disc. When the orbit lies completely within the disc it can be said to be captured into the disc. However, this process can be interrupted at any time if the star comes within $r_\mathrm{acc}$ of the MBH.
This then leads to three paths for a star to be accreted onto the MBH: (1) \textit{disc capture} of particles with low eccentricity and low relative inclination between the orbital plane and the accretion disc, (2) \textit{gas assisted accretion} of particles with moderate eccentricity, characterised by a uniform inclination distribution, and (3) \textit{direct accretion} onto the MBH of particles on nearly radial orbits (also characterised by a uniform inclination distribution). The three paths correspond to $e_\mathrm{acc}\sim0$, $<1$, and $\sim 1$ respectively, where $e_\mathrm{acc}$ is the particle's orbital eccentricity at the time of accretion. These three broad \textit{plunge types} are thus categorised by the eccentricity at the time of accretion.
There are five basic phases that a star can go through (in each of the above categories): (1) scattering, (2) slow decay, (3) fast decay, (4) disc migration, and (5) radial infall. Note that not every accreted star will go through all of these phases. For example, before a gas assisted accretion event, a star will go through the scattering and possibly some of the slow migration phase before it is almost radially accreted onto the MBH, while the radial fall phase is unique to the radially accreted type. The scattering phase is when the star is far from the MBH, almost unaware of the disc, and any change in binding energy and angular momentum is caused by interactions between stars. Once the star is scattered close enough into the disc that disc crossings can significantly alter the binding energy, then it is in the slow decay phase. If the star is aligned with the disc when the radius is within $\sim 10^{-3}$ length units, then its orbit will rapidly decay in $\lesssim 10$ time units. If the star becomes completely aligned with the disc on a nearly circular orbit, it enters the final phase where it migrates with the gas disc until eventually falling inside $r_\mathrm{acc}$, also in $\lesssim 10$ time units. We use the word \textit{plunge} to refer to all phases an accreted star goes through after the scattering phase, and thus the plunge type as defined above refers to path to accretion a star has taken.
Fig.~\ref{FigPlunges} shows the semi-major axis, eccentricity and inclination as a function of time \textit{before} the star is accreted (i.e. time increases to the right of the horizontal axis) for selected stars in the above three categories. The orbital elements $a$ and $e$ are calculated from the particle's orbital energy and angular momentum under the assumption of a Keplerian orbit, they are thus effective quantities rather than geometric ones. The orbits are not perfectly Keplerian due to the self-gravity of the stellar system, however the relevant orbits are deep enough in the MBH's potential well to justify this approximation (as noted earlier, $r_\mathrm{inf}=R_\mathrm{d}=0.22$). The ``wiggling'' seen especially in panels (c) and (d) is due to the fact that these orbits are on the outskirts of the MBH's sphere of influence, where this approximation is poor. The black arrows show an estimate of the beginning of the decay phase. This time is characterised by a consistent decrease in the binding energy of the orbit. The method for determining this point is involved and the details are discussed in the next Section.
The first category of disc capture can be further divided into two subcategories based on the orbit's orientation relative to the disc's angular momentum. An example of a star captured into the disc and later accreted onto the MBH from a prograde orbit is shown in Fig.~\ref{FigPlunges} (a), while an example of an accretion from a retrograde orbit is shown in Fig.~\ref{FigPlunges} (b). For both prograde and retrograde disc capture stars, the star becomes aligned with the disc and the orbit rapidly circularises once it lies within the disc. Note that the semi-major axis of the particle on the retrograde orbit decays much more rapidly than the prograde one due to head wind effect.
For a particle orbiting within the disc (either prograde or retrograde), by the time it reaches $r_\mathrm{acc}$, the orbital eccentricity has decayed to $\sim 0$. The details of the motion of the particle once it is captured into the disc are complicated and not accurately modelled here. This disc migration phase is similar to a Type I migration in planetary dynamics, i.e. large mass ratio between the captured star and the gas disc, so no gap opening is expected \citep[cf.][]{BaruteauEtAl2011, KocsisEtAl2011, McKernanEtAl2011, BellovaryEtAl2015}. The final migrating phase is slow in our simulations since the relative velocity between the star and the gas goes to zero, and so does the force in equation~(\ref{fdrag}). In nature, this decay will be much faster since the gas will have an inward radial velocity component. This level of realism is not included in our models since we cannot resolve a single star, nor do we model all of the microphysics.
The second type (gas assisted accretion) is shown in Fig.~\ref{FigPlunges} (c). This example starts with a large relative inclination with respect to the disc plane (and retrograde orientation) such that the star particle is accreted onto the MBH before there is sufficient time for the eccentricity to decay. This lack of time to circularise is the only difference between this type and the first type where stars are accreted after spending time within the gas disc itself. In this case the particle quickly begins to plunge through the disc (black arrow) and enters the slow decay phase where it steadily loses binding energy until it comes within the accretion radius. We found many similar examples in the simulation results that showed a long scattering phase (see below) before the slow decay. These can potentially last multiple relaxation times until the star crosses the disc close enough for it to enter into the slow decay phase. The statistics of how deep (i.e. close to the MBH) a star must cross the disc to begin the decay phase is discussed in Section~\ref{plungestats}.
The third type is a direct accretion of the star particle on a nearly radial orbit onto the MBH. These events are caused by an interaction between two distant stars, one of which is sling shot into the loss-cone of the MBH. The example particle shown in Fig.~\ref{FigPlunges} (d) comes from a radial orbit ($e_\mathrm{acc} \sim 1$) at a large distance from the MBH. This type has little to no interaction with the gas disc, and so has no arrow as there is no decay phase. The accretion happens on an orbital time which is not visible in the plot. Note that the gaps in the semi-major axis versus time are due to the particle being on hyperbolic (unbound) orbits and thus negative semi-major axis. This process is well studied and for an estimation of the mass growth of the MBH based on loss-cone filling see Section~\ref{sec:losscone}. The scattering time-scale before the orbital decay is largely variable and can be of the order of a relaxation time, whereas the time-scale for the final phase for direct accretion is just the free-fall time to the MBH.
\begin{figure*}
\includegraphics[width=1\textwidth]{PlungeExamples.eps}
\caption{Semi-major axis, eccentricity and inclination of a few sample orbits for the three types described in Section \ref{starplunge}. The semi-major axis is plotted in red when the orbits is prograde ($i<90^\circ$) and green when retrograde. Panel (a) shows a disc captured star with $e_\mathrm{acc} \sim 0$ and low inclination prograde orbit, (b) shows a star captured into the disc on a retrograde orbit, (c) shows a gas assisted accretion $e_\mathrm{acc} < 1$ where the inclination distribution for those orbits is uniform, (d) shows a direct accretion onto the MBH with $e_\mathrm{acc} \sim 1$ and also from a uniform inclination distribution. Black arrows indicate where the decay phase begins (see text). Panel (d) shows a direct capture, and thus has no decay phase; the gaps in the semi-major axis occur when the orbit is instantaneously unbound.}
\label{FigPlunges}
\end{figure*}
\subsection{Plunge statistics}
\label{plungestats}
The aim of this Section is twofold: (1) to estimate the plunge time-scales and (2) to estimate the distance from the MBH where the slow decay phase begins. This distance is used to estimate an effective radius for the disc. Since not all stars coming within this radius will instantly be accreted, this radius cannot be converted into an effective cross section, like the tidal radius can for loss-cone calculations in stellar systems (see Section~\ref{sec:losscone}).
Before proceeding, let us note that the beginning of the plunge is the end of the scattering phase, and could mean either the beginning of the decay (types 1 and 2) or beginning of the radial infall (type 3). The latter is not so relevant in the context of the star-disc interaction as type 3 plunge is a 2-body effect that has nothing to do with the disc.
To define the time of beginning of the plunge for types 1 and 2, one first needs to find the peak of the fast orbital decay phase, i.e. the most recent minimum in $\mathrm{d}a/\mathrm{d}t$ while the star particle is bound to the MBH ($a>0$) and before the star is accreted. Then, the time derivative $\mathrm{d}a/\mathrm{d}t$ is followed back in time to find the next local maximum in semi-major axis, this indicates the previous increase in semi-major axis. Since the semi-major axis can only be increased by an interaction with another star, this marks the end of the scattering phase, and is defined as the beginning of the plunge. The time derivative is calculated from $a(t)$ after being smoothed using a moving average with a window of 1 time unit to remove spurious low amplitude oscillations during the plunge caused by scattering events and deviations from the assumed Keplerian potential. Examples of the beginning of plunges for those two types are shown as the black arrows in Fig.~\ref{FigPlunges} (a--c).
This procedure was applied to all accretion events from the most realistic gas disc model from Table~\ref{ModelParas}, namely \texttt{128k03r}. After two relaxation times, the total number of accreted stars was 14657, of which 19.3 per cent had $e_\mathrm{acc} > 0.8$ (most of which direct accretion onto the MBH and so had no orbital decay phase), 18.6 per cent had eccentricities in the range $0.1 < e_\mathrm{acc} < 0.8$ and 62.1 per cent had $e_\mathrm{acc} < 0.1$, typically very close to zero. For more details on the orbital distribution of accreted stars from this model see Section~\ref{sec:hrdisc}.
The distribution of semi-major axes at the \textit{beginning} of each plunge is shown in Fig.~\ref{Plunge:semi} for all accreted stars broken down to three groups according to their \textit{final} eccentricities $e_\mathrm{acc}$. For plunge types 1 and 2, the plotted quantity is the semi-major axis at the time of the beginning of the plunge as defined above; for type 3, the value shown is the semi-major axis at the time of accretion for bound (elliptical) orbits only, which has the same value as in the beginning of the plunge because in this case the plunge (radial infall) is very quick. One difference between the three distributions is that stars with $e_\mathrm{acc}<0.1$ often begin their plunge in a much denser regions of the disc, as seen by the semi-major axis spread closer to the MBH.
Another difference is in the plunge time-scales, as shown in Fig.~\ref{Plunge:time} where it is clear that while stars with low eccentricities at accretion begin the plunge from smaller semi-major axes, it takes longer for the semi-major axis to decay. Two examples of these plunges with $e_\mathrm{acc} \sim 0$ were given in Fig.~\ref{FigPlunges}, in both cases the particle has time to circularise before the semi-major axis decays and brings the star into the MBH. This process also aligns them to the disc and orbit in either a prograde or retrograde sense, this is examined further in Section~\ref{sec:hrdisc}. It is worth noting that the plunge time-scale can take up to $\sim 0.1$ times the relaxation time for stars that will become trapped in the disc (low $e_\mathrm{acc}$).
\begin{figure}
\begin{centering}
\includegraphics[width=\columnwidth]{semihist.eps}
\par\end{centering}
\caption{Distribution of semi-major axis at the \textit{beginning} of the plunge in model \texttt{128k03r} where 14657 particles plunged during about two relaxation times, divided into three groups according to \textit{final} eccentricity. Low $e_\mathrm{acc}$ tend to begin their plunge from lower semi-major axis than higher eccentricity particles. Particles are removed from the simulation at $r_\mathrm{acc} = 6.6 \times 10^{-5}$ H\'enon length units. $a_\mathrm{disc}$ is the distribution of radii where the particle had an orbit completely inside the disc. This shows that stars go into the disc before they are accreted at the accretion radius.}
\label{Plunge:semi}
\end{figure}
\begin{figure}
\begin{centering}
\includegraphics[width=\columnwidth]{tplungehist.eps}
\par\end{centering}
\caption{Distribution of plunge time-scale (same model and eccentricity grouping as in Fig.~\ref{Plunge:semi}). Among the particles accreted through the disc (low and moderate $e_\mathrm{acc}$), it is interesting to note that that the particles with low eccentricity take longer to be accreted. This is due to the final stage taking longer than it would in reality, since the particles end up moving with the gas disc velocity, which in our model has no radial component. For comparison, the half-mass dynamical time is $\sim 0.54$ H\'enon time units while the relaxation time for $N=128\mathrm{k}$ is $\sim 886$ time units.}
\label{Plunge:time}
\end{figure}
To get a sense of the effective radius of the disc, the semi-major axis (the distribution of which is shown in Fig.~\ref{Plunge:semi}) is not the most important quantity to look at. Instead, we examine the \textit{projected disc crossing radius}, which is the radius where the orbit of the particle at the time it began its plunge crosses the disc; in this case only plunge types 1 and 2 are considered, since type 3 has no interaction with the disc. This value will be between the pericentre and the semi-minor axis. An approximation to the disc crossing distance is given by the distance from the MBH when the true anomaly is $\pm \pi/2$, i.e. $a(1-e^2)$, the distribution of which is shown in Fig.~\ref{Plunge:rdisc}.
\begin{figure}
\begin{centering}
\includegraphics[width=\columnwidth]{rdiskhist.eps}
\par\end{centering}
\caption{Distribution of the projected disc crossing radius, where the orbit crosses the disc just as the particle begins its plunge through the disc and onto the MBH (same model as Fig.~\ref{Plunge:semi}). The red horizontal bar indicates the median and standard deviation of the distribution, giving a characteristic location in the disc where most stars begin the plunge of $R_\mathrm{eff} = 0.032$.}
\label{Plunge:rdisc}
\end{figure}
The median value of the distribution given in Fig.~\ref{Plunge:rdisc} is used as an effective radius of the disc $R_\mathrm{eff} = 0.032$, which corresponds to an enclosed disc mass of $\sim 42$ per cent. This value is an estimate due to the ambiguity of defining the beginning of the plunge itself. Furthermore, the enclosed mass definition is not likely to be generally true for any kind of an AD, and investigation of a larger range of disc models and mass ratios is needed. This radius cannot be used to calculate a cross section as not every particle that crosses the disc within this distance will be eventually accreted.
A better way of thinking about the effective radius is as a divider between potential plungers ($R<R_\mathrm{eff}$) and stars that will be sped up (in the prograde direction) without being accreted ($R>R_\mathrm{eff}$). The latter category of stars will cross the disc and have their velocities changed by the gas velocity, which will induce a rotation in the star cluster over time. This effect of the accretion disc on the star cluster is examined in more detail in a future work.
\begin{figure}
\includegraphics[width=1\columnwidth]{InitialScatter.eps}
\caption{Distribution of the \textit{initial} semi-major axis and eccentricities of the accreted particles grouped according to the \textit{final eccentricity} (plunge type). Cf. Fig. \ref{Plunge:semi} that show the semi-major axis at the beginning of the plunge. The blue points represent the initial conditions of particles that went through type 3 plunge ($e_\mathrm{acc}>0.8$) while the red points are particles that go through type 1 plunge ($e_\mathrm{acc}<0.1$); only one in seven points is shown in order not to overcrowd the figure, and type 2 has an intermediate distribution which we do not show for the same reason. The background color represents the initial conditions of all particles.
\label{fig:InitialScatter}}
\end{figure}
In contrast with the above figures that present quantities measured at the beginning of the plunge, Fig. \ref{fig:InitialScatter} shows the distribution of the \textit{initial} semi-major axis and eccentricities of the accreted particles grouped according to the \textit{final eccentricity} (plunge type).
The blue points represent the initial conditions of particles that went through type 3 plunge ($e_\mathrm{acc}>0.8$) while the red points are particles that go through type 1 plunge ($e_\mathrm{acc}<0.1$); only one in seven points is shown in order not to overcrowd the figure, and type 2 has an intermediate distribution which we do not show for the same reason. The background color represents the initial conditions of all particles (rather than just accreted ones). When we discussed the orbital elements previously, it was in the context of the plunge, which occurs within the sphere of influence of the MBH (otherwise the gas is just too sparse to initiate orbital decay), so the Keplerian approximation to calculate the orbital element was valid. This is not the case here, and we calculated the geometric semi-major axis and eccentricity from the peri- and apocentres by integrating each orbit in the smoothed potential using the ETICS-SCF code \citep{MeironEtAl2014}
\begin{eqnarray}
a_\mathrm{geo} & = & (r_\mathrm{apo}+r_\mathrm{peri})/2\\
e_\mathrm{geo} & = & (r_\mathrm{apo}-r_\mathrm{peri})/2a.
\end{eqnarray}
The two histograms in the bottom panel show the initial semi-major axis distribution of these types, and are roughly log-normal distributions of similar width (cf. the distribution of semi-major axes at the beginning of the plunge in Fig. \ref{Plunge:semi}); the red histogram contains more than three times the number of particles in the blue one, which represents the ratio between type 1 and type 3 plunges (as noted above) this could be understood by looking at the disc as a ``bigger target'' (however as noted before, $R_\mathrm{eff}$ is not equivalent to a geometric cross section).
Particles which started at a semi-major axis of $\gtrsim 1$ are reasonably ``safe'' from being accreted onto the MBH: a little less than 0.5 per cent of particles with $a_\mathrm{init}>1$ (and any eccentricity) were accreted after two relaxation times, compared to 15 per cent for the complementary set. This is due to the fact that 2-body relaxation at the outskirts is really much slower than at the half-mass radius, where the relaxation time is measured. Thus, particles with very large initial $a$, and low to moderate initial $e$, usually continue along these orbits uninterrupted for a long time.
For a similar reason, we see that the red histogram peaks to the left of the blue one. It may seem surprising, since relaxation is the dominant process and we may expect complete mixing in phase space, how come then there seems to be some memory? Stars which were accreted from small $a$ (circular orbits) seem to have a tendency to be at smaller $a$ also at $t=0$. This is not a surprise, because relaxation is much less efficient in energy (semi-major axis) space than in angular momentum (eccentricity) space. In eccentricity, we can expect fast mixing within one relaxation time. In fact, the classical definition of $t_\mathrm{rx}$ in standard textbooks relates exactly to this point, that the velocity vector changes by $90^\circ$ or so, it does not relate to energy changes, which only occur in real systems rather than the idealized one discussed by \citet{Chandrasekhar1943}.
\section{Mass growth and eccentricities}
\label{sec:RES}
The bulk of the results from the models with constant disc height ($h=h_z$) listed in Table~\ref{ModelParas} are presented here with the aim of showing the effect of increasing the particle and spatial resolution. The effect of resolution is most evident with regards to the growth of the MBH and the eccentricity distribution of the accreted stars. Note that all models have a fixed gas disc mass, so there is no contribution to the growth of the MBH from gas, only from the accretion of stars.
In addition to the MBH mass growth, the eccentricity distribution of accreted stars is a key focus rather than any other orbital element. This is due to the fact that all accreted stars have (by definition) a pericentre distance $\leq r_\mathrm{acc}$; thus, the semi-major axis is constrained by the eccentricity according to $a \leq r_\mathrm{acc}/(1-e)$. In Section~\ref{sec:hrdisc}, the orbital inclination of the accreted stars is also investigated in the context of different gas disc profiles.
The effect of particle resolution on the growth of the MBH is shown in Fig.~\ref{Res:FixedRVaryN} (a). This figure shows the MBH mass versus time in units of the relaxation time, which depends on the total particle number $N$ via equation~(\ref{trelax}). To better illustrate the effect of the particle resolution, the best spatial resolution ($r_\mathrm{acc}^* = 3$) is used throughout. The colours indicate each particle number as stated in the figure caption. Thick lines indicate models with a gas disc and thin lines indicate models with no gas disc (for clarity, this convention is adopted throughout the paper. Note that the mass growth versus time is effectively constant once $N \geq 32$k, i.e. the 32k (blue lines) and 128k (green lines) almost completely overlap in Fig.~\ref{Res:FixedRVaryN} (a).
\begin{figure}
\begin{centering}$\begin{array}{c}
\multicolumn{1}{l}{\mbox{(a) MBH growth}}\\[-0.1cm]
\includegraphics[width=\columnwidth]{mass-old2.eps}\\
\multicolumn{1}{c}{\mbox{}}\\
\multicolumn{1}{l}{\mbox{(b) Cumulative eccentricity distribution}}\\[-0.1cm]
\includegraphics[width=\columnwidth]{ecc-old2.eps}
\end{array}$
\par\end{centering}
\caption{The top panel shows the MBH mass growth with time (in units of the relaxation time) for $r_\mathrm{acc}^* = 3$ with different particle numbers $N$ (denoted by different colours as indicated by the legend). Constant disc thickness models and models without gas are shown as thick and thin lines, respectively; note that the lines for $N=32$k and 128k (with gas disc) overlap. The bottom panel shows the cumulative eccentricity distribution for the same models after two relaxation times (note that the thin lines are difficult to make out, they hug the bottom-right corner of axes, as in this case all accreted particles have very high eccentricities).}
\label{Res:FixedRVaryN}
\end{figure}
Fig.~\ref{Res:FixedRVaryN} (b) shows the cumulative eccentricity distribution of accreted stars after two relaxation times. In this figure, $N_\mathrm{accr}$ is the total number of accreted stars after two relaxation times while $N_\mathrm{cum}$ is the number of accreted stars in each with eccentricity $\leq e$. The cumulative eccentricity distribution is then given by $N_\mathrm{cum}/N_\mathrm{accr}$ and is calculated from $e_\mathrm{acc}$, the final eccentricity of the particle when it was removed from the simulation. Note that there is no convergence of the results with increasing particle number. In fact, as the particle number increases, the fraction of stars accreted with low eccentricity falls, i.e. more stars are being accreted with relatively high eccentricity.
Similar plots for the mass growth and cumulative eccentricity distribution after two relaxation times are shown in Fig.~\ref{Res:FixedNVaryR} for a fixed particle number of $N=128$k and varying the accretion radius to be $r_\mathrm{acc}^* =3,\ 10$ and 50. Fig.~\ref{Res:FixedNVaryR} (a) shows that the mass growth of the MBH does converge as $r_\mathrm{acc}$ decreases for the cases with gas (thick lines), but not for the cases without gas (thin lines). Note that the green line exactly reproduces the result in \citetalias{JustEtAl2012}.
\begin{figure}
\begin{centering}$\begin{array}{c}
\multicolumn{1}{l}{\mbox{(a) MBH growth}}\\[-0.1cm]
\includegraphics[width=\columnwidth]{mass-old2-r.eps}\\
\multicolumn{1}{c}{\mbox{}}\\
\multicolumn{1}{l}{\mbox{(b) Cumulative eccentricity distribution}}\\[-0.1cm]
\includegraphics[width=\columnwidth]{ecc-old2-r.eps}
\end{array}$
\par\end{centering}
\caption{Similar to Fig. \ref{Res:FixedRVaryN} but the particle number is fixed $N=128$k while we vary the spatial resolution, as represented by the numerical accretion radius $r_\mathrm{acc}^* \equiv r_\mathrm{acc}/(10^{-4} R_\mathrm{d})$. Note that the lines for $r_\mathrm{acc}^*=10$ and 3 (with gas disc) overlap in panel (a).}
\label{Res:FixedNVaryR}
\end{figure}
Another way to compare the different models is to show their evolutionary track in a phase space consisting of MBH mass and the fraction $f(e_\mathrm{acc}<e_\mathrm{crit})$ of accreted stars with eccentricity smaller than some value $e_\mathrm{crit}$. We choose $e_\mathrm{crit} = 0.8$ for this critical value as it clearly distinguishes between scenarios where mostly radial orbits are accreted and those where the gas disc dominates. When there is no disc, all accreted stars come from nearly radial orbits, so this fraction $f$ would be zero, otherwise its value carries information about the eccentricity distribution. For the bulk of the simulations with gas, the full time evolution is shown in Fig.~\ref{Res:modelevoln} where all models begin at the bottom and evolve upwards. The filled circles indicate integer relaxation times. Note that the MBH growth slows down (filled circles get closer) for all models as time progresses because of the depletion of stars particles replenishing the disc and MBH loss-cones.
\begin{figure}
\includegraphics[width=\columnwidth]{evoln-old9.eps}
\caption{The time evolution of all models with constant height gas discs for all 12 combinations of $N$ and $r_\mathrm{acc}^*$. Time increases upwards and filled circles indicate integer relaxation times. Colours indicate the particle resolution with $N=8$k (black), 16k (red), 32k (green) and $N=128$k (blue); line styles indicate the spatial resolution with $r_\mathrm{acc}^*=3$ (thick solid lines), 10 (thin solid lines) and 50 (dashed lines). Note that the $N=128$k runs go only up to 2 relaxation times due to computational cost.}
\label{Res:modelevoln}
\end{figure}
The main results from Fig.~\ref{Res:FixedRVaryN} are that more stars are directly accreted as $N$ increases but increasing the particle resolution beyond $N=32$k does not alter the growth rate of the MBH through star accretion. Increasing the number of stars means that the probability for a star to be scattered into the MBH loss-cone increases and that the loss-cone is kept full for a longer time. This effect must have a steeper-than-linear dependence on $N$, otherwise it would be cancelled out by the linear scaling with $N$ of the 2-body relaxation time. The saturation of the growth rate with $N$ means that the most important physics for capturing stars into the accretion disc and subsequently onto the MBH is adequately captured by only $N=32$k particles. Since the effective loss-cone for capturing stars into the accretion disc is $R_\mathrm{eff} = 0.032$ (see Section~\ref{plungestats}), the diffusion of stars into this region is modelled correctly with relatively small particle numbers as compared to real galactic centres.
From Fig.~\ref{Res:FixedNVaryR} we learn that stars are more easily accreted onto the MBH for larger $r_\mathrm{acc}$ values. This is also clear from Fig.~\ref{Res:modelevoln} and can be understood by considering that models with larger $r_\mathrm{acc}$ have a larger target to hit and so it is easier for the MBH to directly accrete stars. This also explains why the presence of the gas disc is less important for larger values of $r_\mathrm{acc}$, as seen in Fig.~\ref{Res:FixedNVaryR} (b).
This has ramifications for applying these results to real galactic centres. As $r_\mathrm{acc}$ will tend to the stellar tidal disruption radius, from Fig.~\ref{Res:FixedNVaryR} (b) we would expect all orbits to have $e \sim 0$, i.e. $f(e<e_\mathrm{crit}) \sim 1$. However, real galactic centres also have larger particle numbers with $N\sim 10^{9}$ and as noted before, we expect more stars with radial orbits to be accreted with increased particle number. The result of this competition between two effects for real systems is discussed in more detail in Section~\ref{sec:realgc}.
\section{Increasing model realism}
\label{sec:hrdisc}
In this Section we test the effects of two improvements in the code as compared to {\citetalias{JustEtAl2012}}: smoothing length for the MBH was removed ($\epsilon_\mathrm{bh}=0$) for the models ending with \texttt{ns} in Table~\ref{ModelParas}, and the disc height was changed from a constant $h_z$ to a two-part function $h(R)$ of equation~(\ref{eq:hR}) to represent the transition to a vertically self gravitating disc.
The scale height profile, which was kept constant in {\citetalias{JustEtAl2012}} and Section \ref{sec:RES}, can give incorrect orbits for captured stars when the tidal radius is smaller than the disc scale height. Constant scale height means that there is a cylinder of gas vertically above the MBH, rotating with the planar Kepler velocity; this is unphysical. In this Section we examine the effect of changing the scale height such that it linearly increases with radius, up to the radius $R_\mathrm{sg}$ where the disc becomes vertically self-gravitating given by equation~(\ref{rsg}). For $R > R_\mathrm{sg}$ the scale height is kept constant as $h=h_z=10^{-3}R_\mathrm{d}$ as described in Section~\ref{sec:discmodel}. Note that changing the scale height does not change the total mass of the disc or the surface density at each radius, by consideration of equation~(\ref{gasdensity}).
Fig.~\ref{Res:DiscHeight} (a) shows that the MBH growth rate is unaffected by the different disc model; the differences we see can be attributed to stochastic variations between numeric models with the same physics. This is expected, since the only difference between the two disc models occurs inside of $R_\mathrm{sg}$, which is inside the effective disc radius $R_\mathrm{eff}$ (determined in Section~\ref{plungestats}). This means that any star close enough to notice the difference between the disc models is already plunging through the accretion disc and onto the MBH.
\begin{figure}
\begin{centering}$\begin{array}{c}
\multicolumn{1}{l}{\mbox{(a) Black hole growth}}\\[-0.1cm]
\includegraphics[width=\columnwidth]{mass-new2.eps}\\
\multicolumn{1}{c}{\mbox{}}\\
\multicolumn{1}{l}{\mbox{(b) Cumulative eccentricity distribution}}\\[-0.1cm]
\includegraphics[width=\columnwidth]{ecc-new2.eps}
\end{array}$
\par\end{centering}
\caption{The top panel shows the MBH mass growth for $r_\mathrm{acc}^* = 3$ models for 2 relaxation times for different disc height models. Three particle resolutions were chosen to examine this effect, namely $N=8$k, 32k and 128k.}
\label{Res:DiscHeight}
\end{figure}
The only difference in results between the two disc models is in the orbits of the captured stars, as seen in Fig.~\ref{Res:DiscHeight} (b). This Figure shows that the new $h(R)$ disc model leads to a sharper cut off between radial and low eccentricity orbits, i.e. there is a smaller fraction of stars in the eccentricity mid-range. This is true for all particle resolutions. Also note that when $N$ increases, the fraction of stars on radial orbits is also increased, which was seen previously for the constant height disc model in Fig.~\ref{Res:FixedRVaryN} (b).
To show the effect of the disc model on the orbits of the captured stars in greater detail, a high spatial and particle resolution simulation was performed, namely \texttt{128k03r} in Table~\ref{ModelParas}. This model was previously used in Section~\ref{plungestats} to produce statistics for the radius at which stars begin to plunge into the disc and the time taken for this process. Here we focus on the differences between accreted star orbits for $h_z$ and $h(R)$ models, having already established that mass growth is unaffected. Fig.~\ref{Res:AccrHR} shows the semi-major axis and orbital inclination for all accreted stars at the moment of accretion for models \texttt{128k03} (panel a) and \texttt{128k03r} (panel b). Note that the horizontal axis is composed of two log scales, both ending at $90^\circ$, with one beginning near zero (left) and the other beginning at $180^\circ$ (right). This non-standard plotting method highlights the differences between stars captured on prograde ($i < 90^\circ$) and retrograde ($i > 90^\circ$) orbits.
\begin{figure}
\begin{centering}$\begin{array}{c}
\multicolumn{1}{l}{\mbox{(a) $h=h_z$}}\\
\includegraphics[width=\columnwidth]{inc-new-hz.eps}\\
\multicolumn{1}{c}{\mbox{}}\\
\multicolumn{1}{l}{\mbox{(b) $h=h(R)$}}\\
\includegraphics[width=\columnwidth]{inc-new-hr.eps}
\end{array}$
\par\end{centering}
\caption{Semi-major axis and orbital inclination for all accreted stars at the moment of accretion for models with constant height disc (panel a; \texttt{128k03}) and varying disc heigh (panel b; \texttt{128k03r}). Particles are divided into three (final) eccentricity groups and are coloured accordingly: $e_\mathrm{acc}>0.8$ (red), $0.1<e_\mathrm{acc}<0.8$ (green) and $e_\mathrm{acc}<0.1$ (blue). Note the non-standard horizontal axis which is used to highlight the differences between particles captured on prograde and retrograde orbits.}
\label{Res:AccrHR}
\end{figure}
The stars accreted from radial orbits show very similar characteristics in both disc models in Fig.~\ref{Res:AccrHR} (red dots), in particular the semi-major axis distribution. This is expected, since most radial accretions should originate at the maximum of the stellar density profile, which is the same independent of disc model. There are two key differences between the accreted star populations in the different disc models. First, the population of disc-captured low-eccentricity and prograde-planar stars ($e\sim0$, $i<1^\circ$) is more well-defined for the $h(R)$ model; second, the $h_z$ model has very few stars accreted from inclined ($i>30^\circ$), prograde orbits with semi-major axes close to $r_\mathrm{acc}$. The second point can be understood by removing the unphysical gas above the MBH and inside of $R_\mathrm{sg}$, causing these orbits to rapidly decay in a violent way before they are properly aligned.
Violent or chaotic accretion is seen for retrograde orbits in both disc models where the backwind effect (described in Section~\ref{sec:plunge}) causes the orbit to decay faster than it is aligned with the disc. This behaviour was seen for the plunges shown in Fig.~\ref{FigPlunges} where the star particle either has time to circularise before accretion, panel (b), or is accreted before circularisation, panel (c). For the $h(R)$ disc far more stars are circularised and aligned in the disc (Fig.~\ref{FigPlunges} a) than for the $h_z$ disc. The $h(R)$ model also cause accretion of stars which are aligned in a retrograde sense but have not circularised (green points near $i=179^\circ$), which the constant height model does not.
We note here that we use the word `circularised' for all particles with low $e_\mathrm{acc}$ since all accreted particles, i.e. having pericentre distance $\lesssim r_\mathrm{acc}$, are \textit{initially} on nearly radial orbits in the star cluster. So if they end up with low eccentricity just before accretion, it is because they have been circularised by interactions with the disc. Star--star interactions can refill the loss-cone, but only with particles nearly radial orbits as is evident from the no-gas results.
In both models the number of accreted stars from radial orbits is similar: 22 per cent for $h_z$ and 19 percent for $h(R)$. Far more stars are circularised in the $h(R)$ model compared to the $h_z$ model: 63 percent and 10 percent, respectively. Overall, the more physical $h(R)$ model better separates the radially accreted population and the disc captured population with only 19 percent having $0.1<e_\mathrm{acc}<0.8$ compared to 67 per cent for $h_z$. A more detailed breakdown of the $h(R)$ model, examining the statistics of the plunge and comparing the types of plunges was presented in Section~\ref{plungestats}.
\section{Application to real galactic centres}
\label{sec:realgc}
The number of particles required to accurately model a galactic centre using a direct \textit{N}-body simulation is well beyond current computing capabilities. To apply our results to real galactic centres, we first compare the accretion rate expected from loss-cone theory with the results of the no-gas simulations (see Table \ref{RealEst}), thus ensuring that our models are broadly correct. Then, we extrapolate the results of our models which do have a gas disc to estimate MBH mass growth due to star accretion in real galactic centres.
\subsection{Without gas}
\label{sec:losscone}
In this Section we use loss-cone theory \citep{FR1976} to derive
a formula for the accretion or tidal disruption rate as a function of free parameters,
and in particular we are interested in the scaling with $r_{\mathrm{acc}}$
and $N$. We will then show that our runs without a gas disc scale
roughly in agreement with the formula. Let us assume the MBH to be
a stationary target of radius $r_{\mathrm{acc}}$ and that $N$ particles
form a \citet{BahcallWolf1976} cusp around it, where the
stellar density runs with $r^{-7/4}$, and that the only process
that can cause particles to walk in energy-angular momentum space
is 2-body relaxation. The definition of the loss-cone then is: the
region in angular momentum and energy phase space where stars on
unperturbed orbits in the given gravitational potential have
an orbital pericentre distance to the MBH less than the accretion radius
$r_{\mathrm{acc}}$. The loss cone ansatz has been used many times to
estimate black hole growth rates
\citep[cf.][]{LightmanShapiro1977,PauEtAl2004,BaumgardtEtAl2004,ZhongEtAl2014}.
Its main assumption is that the bulk of the tidally accreted stars originates
from a critical radius $r_{\mathrm{crit}}$, where draining of the loss
cone through tidal accretion and refilling due to local 2-body relaxation
balance each other. The MBH accretion rate derived from loss cone
theory is denoted here with $\dot{M}_{\mathrm{bh}}^{\prime(\mathrm{stars})} $
to distinguish it from the value measured in our simulations; the superscript indicates that the change of
mass is due to accretion of stars (rather than gas). It is dominated by the slowest
process and can be estimated through
\begin{equation}
\dot{M}_{\mathrm{bh}}^{\prime(\mathrm{stars})} = \frac{\rho(r_{\mathrm{crit}})r_{\mathrm{crit}}^{3}}{t_{\mathrm{rx}}(r_{\mathrm{crit}})}\label{eqn:mdot}
\end{equation}
Note that close 2-body encounters in the cusp, which kick stars away, may differ
only by a logarithmic factor from the flux of stars towards the black hole
\citep{LinTremaine1980}. However, since usually good agreement between loss cone
models and $N$-body simulations is found, we conclude that this process may be
neglected at least with regard to the MBH growth. Here we follow a slightly
different ansatz of \citet{PauEtAl2004} where no specific model is assumed
for the density distribution of the central star cluster:
\begin{equation}
\frac{r_{\mathrm{crit}}}{r_{\mathrm{acc}}} = \frac{4t_{\mathrm{rx}}(r_{\mathrm{crit}})}{t_{\mathrm{dyn}}(r_{\mathrm{crit}})}.\label{eqn:rc}
\end{equation}
Here, we have used the conditions at the critical radius as given in equations (11) to (14) of
\citet{PauEtAl2004}.
We now rescale the dynamical and relaxation time scales to the influence radius of the
MBH $r_{\mathrm{inf}} = G M_{BH}/\sigma^2$, assuming $r_{\mathrm{crit}} < r_{\mathrm{inf}}$ and using
$\sigma \propto r^{-1/2}$ for the velocity dispersion in the stellar system dominated
by the MBH gravity (i.e. $r < r_{\mathrm{inf}}$). The stellar density inside the influence radius is taken as the Bahcall Wolf cusp scaled to the density at the influence radius, given by
\begin{equation}
\rho(r) = \rho(r_{\mathrm{inf}}) \left( \frac{r_{\mathrm{inf}}}{r} \right)^{7/4}.
\end{equation}
Substituting these into the following expressions for the relaxation time and the dynamical time at $r=r_{\mathrm{crit}}$
\begin{eqnarray}
t_{\mathrm{rx}}(r_{\mathrm{inf}}) & \propto & \sigma(r_{\mathrm{inf}})^{3}/[\rho(r_{\mathrm{inf}}) m_\star \ln\Lambda ]\label{eq:trx-rc}\\
t_{\mathrm{dyn}}(r_{\mathrm{inf}}) & \propto & r_{\mathrm{inf}}/\sigma(r_{\mathrm{inf}}).\label{eq:tdyn-rc}
\end{eqnarray}
with $m_\star = 1/N$ and noting that $r_{\mathrm{inf}} \propto M_{BH}$ we get
\begin{equation}
r_{\mathrm{crit}}^{9/4} = \zeta r_{\mathrm{acc}} \frac{N}{\ln\Lambda} M_{\mathrm{bh}}^{1/4}
\end{equation}
Here we have collected in the constant $\zeta$ all other quantities, which do {\em not} depend
on $r_{\mathrm{acc}}$, $N$, or $M_{\mathrm{bh}}$, which are the quantities relevant for our
discussion. $\zeta$ will vary depending on the stellar system quantities at $r_{\mathrm{inf}}$ and
some physical constants; it is not dimensionless. With this we get the relation
\begin{equation}
r_{\mathrm{crit}} \propto \left[ r_{\mathrm{acc}} \frac{N}{\ln\Lambda} \right]^{4/9} M_{\mathrm{bh}}^{1/9}
\end{equation}
The experimental quantity we measure from the simulations in our numerical experiments (and would like to compare to theory) is the mass growth per fixed time interval (here two relaxation times at the half-mass radius), which we denote as $\Delta M_\mathrm{bh} = 2 t_\mathrm{rx}(r_\mathrm{h}) \dot{M}_{\mathrm{bh}}$; so we get a final scaling of
\begin{equation}
\label{eqn:delMth}
\frac{\Delta M_\mathrm{bh}}{M_\mathrm{bh}} = k \left[ r_{\mathrm{acc}} \frac{N}{\ln(0.4 N)} \right]^{4/9} M_{\mathrm{bh}}^{10/9}.
\end{equation}
In the equation above we have identified the Coulomb logarithm at the critical radius $\ln\Lambda$ with the Coulomb logarithm $\ln (0.4N)$ occurring in the half-mass relaxation time, as in {\citetalias{JustEtAl2012}}. We compare equation (\ref{eqn:delMth}) directly with our numerical results from the
models without gas as shown in Fig.~\ref{fig:DeltaMng}. Since we
expect models with smaller $r_{\mathrm{acc}}$ to give a better approximation
to the loss-cone theory, we ran an extra set of models (without a
gas disc) with $r_{\mathrm{acc}}^{*}=1$ for each $N$ (filled black
circles). The models with larger $r_{\mathrm{acc}}^{*}$ indeed show
a large deviation from the theoretical expectations. Notice that small
deviations from the theoretical curve may depend either on $M_{\mathrm{bh}}$
or on the quantities from outside the MBH influence radius absorbed in the
constant $k$ above.
\begin{figure}
\begin{centering}
\includegraphics[width=1\columnwidth]{MassGrowth}
\par\end{centering}
\protect\caption{Mass growth of the MBH relative to its initial mass after two relaxation times for all
models without a gas disc (i.e. the evolution is purely stellar dynamics);
the horizontal axis is a combination of the numerical accretion radius
$r_{\mathrm{acc}}$ and particle number $N$. The solid line shows
the best-fitting power-law with index of $4/9$ to the models with
$r_{\mathrm{acc}}^{*}=1$ only (filled black circles), as expected
from loss-cone theory (equation \ref{eqn:delMth}). Other coloured
circles represent different values of $r_{\mathrm{acc}}^{*}$ as indicated
by the legend.}
\label{fig:DeltaMng}
\end{figure}
A small accretion radius for the MBH and no gas disc should have the
same power-law relationship as equation~(\ref{eqn:delMth}). We put
it to the test by fitting a power-law to the five data-points with
$r_{\mathrm{acc}}^{*}=1$. The best fitting power-law index was $\approx0.456$,
which is within 3 per cent from the theoretical value of $4/9$. To
extrapolate our results to real galactic centres in this case, we
force the index to be $4/9$ and find the best-fitting constant, we
get $k\approx0.273$. This fit is shown as the solid black line in Fig.~\ref{fig:DeltaMng}.
\subsection{With gas}
\label{sec:realgcs}
The previous section established that simulation results without gas are consistent with theoretical expectations, giving us confidence in the numerical code. Here, the results from models with a gas disc are extrapolated into the regime of real galactic centres. The highest resolution and most realistic model with a gas disc, \texttt{128k03r}, had $N=128$k, $r_{\mathrm{acc}}^{*}=3$ and $h(R)$ disc height profile. The MBH mass growth as a function of time for this model is shown as the blue curve in Fig.~\ref{Res:DiscHeight} (a) for two relaxation times. We found that we could fit this curve using a simple power-law given by
\begin{equation}
g(t)=a\left(\frac{t}{t_{\mathrm{rx}}}\right)^{b}\label{mgrowthfit}
\end{equation}
where $a\approx0.0643$, $b\approx0.843$ and $t_{\mathrm{rx}}$ is the relaxation time for $N=128$k. The residuals were less than 0.01 mass units, i.e. better than a 10 per cent error relative to the initial MBH mass.
\begin{table*}
\caption{Estimation of mass growth via accretion of stars and gas for a sample of galactic nuclei.}
\label{RealEst}
\begin{tabular}{lllllll}
\hline
Object & $M_\mathrm{bh}$ & $r_\mathrm{inf}$ & $\dot{M}_\mathrm{bh}^{\prime(\mathrm{stars})}$ & $\dot{M}_\mathrm{bh}^{(\mathrm{stars})}$ & $\dot{M}_\mathrm{bh}^{(\mathrm{gas})}$ & $\dot{M}_\mathrm{Edd}$\\
& ($\mathrm{M}_{\odot}$) & (pc) & ($\mathrm{M}_{\odot}\,\mathrm{yr}^{-1}$) & ($\mathrm{M}_{\odot}\,\mathrm{yr}^{-1}$) & ($\mathrm{M}_{\odot}\,\mathrm{yr}^{-1}$) & ($\mathrm{M}_{\odot}\,\mathrm{yr}^{-1}$)\\
\hline
M 87 & $6.6 \times 10^{9}$ & 291 & $4 \times 10^{-5}$ & $9 \times 10^{-5}$ & $1 \times 10^{-2}$ & $1 \times 10^{2}$\\
NGC 3115 & $9.6 \times 10^{8}$ & 78 & $3 \times 10^{-5}$ & $1 \times 10^{-4}$ & $2 \times 10^{-2}$ & $2 \times 10^{1}$\\
NGC 4291 & $3.2 \times 10^{8}$ & 24 & $7 \times 10^{-5}$ & $3 \times 10^{-4}$ & $5 \times 10^{-2}$ & $7$\\
M 31 & $1.5 \times 10^{8}$ & 25 & $2 \times 10^{-5}$ & $2 \times 10^{-4}$ & $2 \times 10^{-2}$ & $3$\\
NGC 4486A & $1.3 \times 10^{7}$ & 4.5 & $4 \times 10^{-5}$ & $3 \times 10^{-4}$ & $3 \times 10^{-2}$ & $3 \times 10^{-1}$\\
MW & $4.0 \times 10^{6}$ & 1.4 & $1 \times 10^{-4}$ & $7 \times 10^{-4}$ & $7 \times 10^{-2}$ & $9 \times 10^{-2}$\\
M 32 & $3.0 \times 10^{6}$ & 2.3 & $3 \times 10^{-5}$ & $3 \times 10^{-4}$ & $2 \times 10^{-2}$ & $7 \times 10^{-2}$\\
\hline
\end{tabular}
\par\medskip
\begin{flushleft}\textbf{Notes.} We extrapolate results to this sample of galactic nuclei (adopted from {\citetalias{JustEtAl2012}}). Columns 1--3 are the object's name, MBH mass and radius of influence (calculated from the stellar velocity dispersion), respectively. Column 4 gives the expected accretion rate of stars given pure stellar dynamical evolution (no gas disc); Column 5 gives the expected accretion rate of stars assuming a gas disc of the form assumed in this work is present (averaged over 100\,Myr); Column 6 gives the expected accretion rate of gas from said disc model; Column 7 is the accretion rate corresponding to the Eddington luminosity for each MBH, assuming 10 per cent accretion efficiency.\end{flushleft}
\end{table*}
In order to extrapolate from the simulation results, a number of assumptions are made: (1) mass ratios between disc, MBH and nuclear stellar cluster are roughly the constant adopted in Section \ref{sec:method}, (2) the time evolution of the mass ratios does not significantly affect the MBH mass growth, (3) we only need to extraplolate to a larger $N$, but the value $r_{\mathrm{acc}}^{*}=3$ used in our models is already small enough as justified in Section \ref{sec:RES}, and (4) the disc can be described by a thin $\alpha$-disc with gas on Keplerian orbits. These assumptions are clearly flawed to some extent, but they do allow us to perform an order of magnitude calculation. A more realistic model of the galactic centre is our long term goal.
Table~\ref{RealEst} shows a sample of galactic nuclei, also used for scaling in {\citetalias{JustEtAl2012}} along with their MBH masses and influence radii; these are the only two quantities used for scaling. The stellar mass of the cluster $M_\mathrm{cl}$ is assumed to be $\times 10$ that of the MBH, and the number of stars is assumed $N=M_\mathrm{cl}/\mathrm{M}_{\odot}$. The disc radius $R_\mathrm{d}$ is assumed $=r_\mathrm{inf}$ and the cluster half-mass radius $r_\mathrm{h}$ is assumed $=3R_\mathrm{d}$. We need to use a physical accretion radius when scaling equation~(\ref{eqn:delMth}) to real galaxies. For the four lowest MBH masses in the Table, we approximate it with the rigid-body Roche limit for a Sun-like star
\begin{equation}
r_\mathrm{acc} = \mathrm{R}_{\odot} \left( 2 M_\mathrm{bh}/\mathrm{M}_{\odot} \right)^{1/3},
\end{equation}
while for the three highest masses this turns out to be smaller than the Schwarzschild radius, so accretion will take place without tidal disruption, and for the scaling of equation~(\ref{eqn:delMth}) we use
\begin{equation}
r_\mathrm{acc} = 2 G M_\mathrm{bh}/c^2.
\end{equation}
Note that it is only used to calculate $\dot{M}_\mathrm{bh}^{\prime(\mathrm{stars})}$, the other quantities in the Table are agnostic to the real accretion radius. The extrapolation of $\dot{M}_\mathrm{bh}^{(\mathrm{stars})}$ is straightforward using equation~(\ref{mgrowthfit}), one just needs to find the relaxation time at the half-mass radius for the given nuclei using equation~(\ref{trelax}) and convert to physical mass units by multiplying $g(t)$ by $M_\mathrm{cl}$. For details on the origin of these values see {\citetalias{JustEtAl2012}} and references therein. We use 100\,Myr as the lifetime of the disc \citep[see][]{MillerEtAl2003} and divide the mass growth by this time-scale to get the average accretion rate.
Another contribution to the growth of the MBH is gas from the accretion disc itself. The \citet{SS1973} model assumes a constant gas accretion rate to construct the density structure of the thin disc. Since we assume an ad-hoc density structure, we can work back to find the corresponding gas accretion rate. Using equation~(2.19) of \citet{SS1973} and our equation~(\ref{gasdensity}) we find by algebraic manipulation that
\begin{align}
\dot{M}_\mathrm{bh}^{(\mathrm{gas})} &= 1.8\times10^{-6} \, {\mathrm{M}_{\odot}\,\mathrm{yr}^{-1}} \alpha^{8/7} \left(\frac{M_{\mathrm{d}}}{\mathrm{M}_{\odot}}\right)^{10/7}\nonumber\\
&\phantom{=} \left(\frac{M_\mathrm{bh}}{\mathrm{M}_{\odot}}\right)^{-5/14}\left(\frac{R_{\mathrm{d}}}{\mathrm{pc}}\right)^{-25/14},
\end{align}
where $\alpha$ is the viscosity parameter which is set to $0.2$ for the purpose of calculating the values of $\dot{M}_\mathrm{bh}^{(\mathrm{gas})}$ in Table \ref{RealEst}.
Overall, the contribution of stars to the growth of the MBH is typically $\sim 100$ times weaker than the growth due to accretion of gas from the disc itself. However, this very simple calculation depends strongly on the thin-disc model while accretion in reality is much more complex, so these results should be treated as approximate at best.
\section{Summary and discussion}
\label{sec:CON}
Using a modified version of the $\phi$\textsc{grape} code, we performed a suite of 39 high-accuracy direct $N$-body simulations of a system containing a star cluster, an MBH and a static gas disc. The mass ratios were $M_\mathrm{bh}/M_\mathrm{cl} = M_\mathrm{d}/M_\mathrm{bh} = 0.1$, the rotation profile of the disc was Keplerian and its density profile derived from the Shakura--Sunyaev thin disc model. We tested two disc models (constant height and variable height) and a large range of particle numbers (from 8k to 128k) and accretion radii (from $\sim 10^{-5}$ to $\sim 10^{-3}$ H\'enon length units); we also looked at the effect of the softening length of the MBH--star interaction. All simulations finished at least two relaxation times.
The disc used in our model represents the inner Keplerian part of the gaseous material surrounding the central AGN engine. It is very difficult to resolve it observationally and there is no standard value known for the mass of such disks in general. From modelling and theory, one can argue about the stability of such disks against self-gravity (Toomre $Q$ analysis) or non-axisymmetric perturbations. According to \citet{ArtymowiczEtAl1993}, $\mu_\mathrm{d} = 0.01$ (where $M_\mathrm{d} = \mu_\mathrm{d} M_\mathrm{bh}$) results in a clearly stable disc, while $\mu_\mathrm{d} = 0.2$ is prone to instability \citep{RoedigEtAl2011}. In {\citetalias{JustEtAl2012}} we chose $\mu_\mathrm{d} = 0.1$, and argued that such a disc is already marginally unstable (for $R > 0.26 R_\mathrm{d}$). Such marginally stable disc could still survive the life cycle of quasars (of order $10^8$ years), and gives the maximal dissipative effect on stars \citep{Rauch1995}. The gas reservoir after galaxy mergers outside $R_\mathrm{d}$ (which is also about equal to the gravitational influence radius of the MBH) is, according to many simulations \citep[see e.g.][]{CallegariEtAl2009, ChaponEtAl2013}, much larger, even larger than the black hole mass. Therefore it seems plausible to assume that our AGN disc assumes the maximum stable mass. We choose the value $\mu_\mathrm{d} = 0.1$ for simplicity, because the exact stability boundary can only be determined by much more detailed disc modelling \citep[see e.g.][]{SyerEtAl1991} and depends on parameters which are unknown or poorly defined in our work (e.g. disc viscosity, equation of state).
For the models with a gas disc, the MBH growth due to star accretion converges for sufficiently large particle number ($N \sim 32$k) and small accretion radius ($r_\mathrm{acc}^* \sim 3$). This convergence in MBH growth is not seen for models without gas, precisely as expected from loss-cone theory as explained in Section~\ref{sec:losscone}. There is no convergence in the eccentricity distribution of accreted stars for models with a gas disc due to two competing effects. On the one hand, as the particle number $N$ increases, the fraction of accreted stars on low eccentricity orbits falls (characterised by the fraction $f$ of particles accreted with eccentricity smaller than 0.8). On the other hand, as the spatial resolution increases (i.e. $r_\mathrm{acc}^*$ decreases towards the stellar tidal disruption radius), the fraction $f$ increases as more stars are circularised by the disc before being accreted. It is thus difficult to extrapolate the eccentricity distribution to real galaxies, but from inspection of Figs. \ref{Res:FixedRVaryN} (b) and \ref{Res:DiscHeight} (b) it seems that a large fraction of stars will be disrupted when their orbit has been fully circularised (i.e. $e_\mathrm{acc}\sim 0$). This has ramifications for the search of electromagnetic counterparts of gravitational wave emitters.
The improved variable height disc model resulted in a significantly higher fraction of stars that had been circularised before being accreted, compared to the model with constant disc height. There was no difference in the growth of the MBH, however. Data from our most realistic simulation using the improved disc model ($N=128$k, $r_\mathrm{acc}^* = 3$) were used to identify the three main paths whereby stars accreted onto the MBH: \textit{disc capture} ($e_\mathrm{acc}\sim 0$), \textit{gas assisted accretion} ($e_\mathrm{acc}<1$) and \textit{direct accretion} ($e_\mathrm{acc}\sim 1$); these were discussed in detail in Section~\ref{plungestats} along with the statistics of the different types. Using these data we also calculated the effective radius in the disc where stars began to plunge, which was $R_\mathrm{eff} = 0.032$ ($\sim 0.15$ per cent of the disc radius $R_\mathrm{d}$) and the time-scale for this to occur was between 1 and 10 per cent of the 2-body relaxation time at the half-mass radius.
When extrapolating our results to real galactic nuclei, we found that typically the stellar accretion rate is increased by a factor of $\sim 10$ in the presence of the disc, and is typically a few times $10^{-4}\,\mathrm{M}_{\odot}\,\mathrm{yr}^{-1}$. This assumes that 2-body relaxation is the main process that replenishes the loss-cone, and that our most realistic model with the disc is convergent in terms of both $N$ and $r_\mathrm{acc}$. The former assumption is not justified when the system is strongly non-spherical or massive perturbers are present, both cause large scale torques that can redistribute angular momentum in the stellar system. The latter assumption should be made carefully, since while it is clear that our results for the accretion rate converge between $N=32$k and 128k and between $r_\mathrm{acc}^*=10$ and 3, a realistic system is still a few orders of magnitude away in both variables.
We estimated the accretion rate of gas directly from the disc using the assumptions of the Shakura--Sunyaev model, and found that it is typically $\sim 100$ times larger than the accretion rate of stars, which means that star accretion is not a major contributor to MBH mass growth \citep[cf.][]{MK2005,KennedyEtAl2011}. But while the star accretion rate seems to be robust given the density profile and mass of the disc, the estimated gas rate is potentially much more model-dependent and could be very different, e.g. due to winds. Moreover, our assumption that all the material from the tidally disrupted star is added to the MBH mass is clearly off by some factor.
Since many stars spend some time captured within the disc as they migrate inwards under the effect of friction, if the gas were to suddenly disappear, it would leave behind a disc of stars. The Galactic centre is known to have one \citep{LevinBeloborodov2003} or two \citep{PaumardEtAl2006} such stellar discs, but whether or not this observation is consistent with the subsystem of trapped stars we find in our models will be examined in future work. We note that other mechanism can form a disc-shaped stellar subsystem such as in-situ star formation in the gas disc and vector resonant relaxation \citep{KocsisTremaine2015}. Different mechanisms will have different signatures in terms of mass and age segregation, which cannot be distinguished in our models where all stars are identical and no stellar evolution is included. Multiple stellar populations as well as other realistic physics such as massive perturbers and dynamic disc model will be included in future work.
\citet{ArcaviEtAl2014} find that in their archival search of the Palomar Transient Factory, all three candidate TDEs are in E+A galaxies, that have spectra typical of the host galaxies of type 2 Seyferts \citep{DresslerGunn1983}. Since these galaxies are rare in the local universe, comprising just 0.1 per cent \citep{Goto2007,SnyderEtAl2011}, there seems to be an enhancement of the TDE rate in these galaxies. \citet{ArcaviEtAl2014} speculate that if E+A galaxies are indeed the result of a merger \citep{ZabludoffEtAl1996}, then the TDE rate could be enhanced by the binary MBH mechanism \citep{ChenEtAl2011}. However, because of the similarity of the spectra to AGN hosts, we postulate that a gaseous disc may be present and is responsible to TDE rate enhancement as described in this work. \citet{LiEtAl2015} show in their simulation how the TDE rate increases due to mergers, however, the duration of the event is short; an estimate how this affects TDE rates in a cosmological background is not yet done.
\section*{Acknowledgements}
We thank Maxim Makukov, Chingis Omarov, Emmanuil Y. Vilkoviskij and Ari Laor for valuable comments, discussion, support and supervision. We also thank the anonymous referee for useful comments and suggestions. This work has benefited very much from funding of exchange and collaboration between Germany and Kazakhstan by Volkswagen Foundation under the project ``STARDISK -- Simulating Dense Star-Gas Systems in Galactic Nuclei using special hardware'' (\textit{I/81 396}). GRAPE and GPU hardware at Fesenkov Astrophysical Institute in Kazakhstan used for this work have been supported by the STARDISK project, too. We acknowledge support from the Strategic Priority Research Program ``The Emergence of Cosmological Structure'' of the Chinese Academy of Sciences (No. \textit{XDB09000000}) (Pilot B programme).
We acknowledge support by Chinese Academy of Sciences through the Silk Road Project at NAOC, through the Chinese Academy of Sciences Visiting Professorship for Senior International Scientists, Grant Number \textit{2009S1-5} (RS), and through the ``Qianren'' special foreign experts program of China. RS has also been partially supported by NSFC (National Natural Science Foundation of China), grant Nr. \textit{11073025}. GK acknowledges a Chinese Academy of Sciences Young International Postdoctoral Fellowship Grant No. \textit{2011Y2JB09}. YM acknowledges support from the China Postdoctoral Science Foundation through grant No. \textit{2015T80011}, from the excellence initiative at the Univ. of Heidelberg through mobility measures for international research collaborations project nr. \textit{7.1.47}, and the European Research Council under the European Union's Horizon 2020 Programme, ERC-2014-STG grant GalNUC 638435. PB acknowledges the special support by the NASU under the Main Astronomical Observatory GRID/GPU computing cluster project.
The special GPU accelerated supercomputer {\tt laohu} at the Center of Information and Computing at National Astronomical Observatories, Chinese Academy of Sciences, funded by Ministry of Finance of People's Republic of China under the grant \textit{ZDYZ2008-2}, has been used for some of the largest simulations. We also used smaller GPU clusters {\tt titan}, {\tt hydra} and {\tt kepler}, funded under the grants \textit{I/80041-043} and \textit{I/84678/84680} of the Volkswagen Foundation and grants \textit{823.219-439/30} and \textit{/36} of the Ministry of Science, Research and the Arts of Baden-W\"urttemberg, Germany.
|
\section{Introduction}
The present paper is devoted to several pointwise inequalities involving several nonlocal operators. We focus on two types of pointwise inequalities: the C\'ordoba-C\'ordoba inequality and the Kato inequality. In order to keep the presentation simple, we state the inequalities in question in the case of the fractional laplacian, i.e. $(-\Delta)^s$, in ${\mathbb R}^n$. Actually, in subsequent sections, we will generalize these inequalities to a lot of different contexts. Furthermore, we will present a unified proof for both inequalities based on some extension properties of some nonlocal operators. Our proofs are elementary and simplify the original arguments.
The fractional
Laplacian can be defined in various ways, which we review now.
It can be defined using Fourier transform by
$$
\mathcal F((-\Delta)^s v) = \left| \xi \right|^{2s}\mathcal F(v),
$$
for $v\in H^s({\mathbb R}^n)$. It can also
be defined through the kernel representation (see the book by Landkof \cite{landkof})
\begin{equation}\label{defPV}
(-\Delta)^{s} v(x)=C_{n,s}\ \textrm{P.V.}\int_{\mathbb{R}^n} \frac{v(x)-v(\overline x)}
{|x-\overline x|^{n+2s}}\,d\overline x,
\end{equation}
for instance for $v\in\mathcal{S}({\mathbb R}^n)$, the Schwartz space of rapidly decaying functions. Here we will only consider $s \in (0,1). $
The inequalities considered in the present paper are the following
\begin{thm}[C\'ordoba-C\'ordoba inequality]\label{CC}
Let $\varphi$ be a $C^2({\mathbb R}^n)$ convex function. Assume that $u$ and $\varphi(u)$ are such that $(-\Delta)^s u $ and $(-\Delta)^s \varphi(u) $ exist. Then the following holds
\begin{equation}\label{CCeq}
(-\Delta)^s\varphi(u) \leq \varphi'(u) (-\Delta)^s\,u .
\end{equation}
\end{thm}
The next theorem is the Kato inequality.
\begin{thm}[Kato inequality]\label{kato}
The following inequality holds in the distributional sense
\begin{equation}\label{k}
(-\Delta)^s |u| \leq \text{sgn}(u) (-\Delta)^s\,u .
\end{equation}
\end{thm}
The previous two theorems are already known: Theorem \ref{CC} is due to C\'ordoba and C\'ordoba (see \cite{CCCMP, CCPNAS}). Theorem \ref{kato} is due to Chen and V\'eron (see \cite{CV}). Both original proofs are based on the representation formula given in \eqref{defPV}. This formula holds only when the fractional laplacian is defined on ${\mathbb R}^n$. The C\'ordoba-C\'ordoba inequality is a very useful result in the study of the quasi-geostrophic equation (see \cite{CCCMP}). This inequality has been generalized in several contexts in \cite{CordobaAdv} for instance or \cite{constantin}. In this line of research we propose a unified way of proving these inequalities based on some extension properties for nonlocal operators.
\section{Some new inequalities}\label{new}
In this section, we derive by a very simple argument several inequalities at the nonlocal level, i.e. without using any extensions, which are not available in these frameworks.
\subsection{A pointwise inequality for nonlocal operators in non-divergence form}
Nonlocal operators in non-divergence form are defined by
$$
\mathcal I u(x)=-\int_{{\mathbb R}^n} (u(x+y)+u(x-y)-2u(x))K(y)\,dy
$$
for a kernel $K \geq 0$. Denote
$$
\delta_y u (x)= -\Big ( u(x+y)+u(x-y)-2u(x) \Big ).
$$
Then, considering a $C^2$ convex function $\varphi$, one has by the fact that a convex function is above its tangent line
$$
\delta_y \varphi(u)(x)=-\Big ( \varphi(u(x+y))+\varphi(u(x-y))-2\varphi(u(x))\Big )=$$
$$-\Big (\varphi(u(x+y))-\varphi(u(x))+\varphi(u(x-y))-\varphi(u(x))\Big )
$$
$$
\leq \varphi '(u(x)) \delta_y u(x).
$$
Hence for the operator $\mathcal I$ one has also an analogue of the original C\'ordoba-C\'ordoba estimate.
\subsection{The case of translation invariant kernels}
Consider the operator
$$
\mathcal L u(x)=\int_{{\mathbb R}^n} (u(x)-u(y))K(x-y)\,dy
$$
where $K$ is symmetric. Hence one can write
$$
\mathcal L u(x)=\int_{{\mathbb R}^n} (u(x)-u(x-h))K(h)\,dh
$$
or in other words, by a standard change of variables
$$
\mathcal L u(x)=\frac12 \int_{{\mathbb R}^n} \delta_h u (x)K(h)\,dh
$$
We start with the following lemma, which is a direct consequence of the symmetry of the kernel
\begin{lemma}
$$
\int_{{\mathbb R}^n} \mathcal L u(x)=0.
$$
\end{lemma}
The following lemma is consequence of straightforward computations
\begin{lemma}
$$
\delta_h uv(x)=u\delta_h v+v\delta_h u +
$$
$$
(v(x+h)-v(x))(u(x+h)-u(x))+(v(x-h)-v(x))(u(x-h)-u(x)).
$$
\end{lemma}
Hence by the two previous lemma one has the useful identity
$$
0=\int_{{\mathbb R}^n} \mathcal L u^2=2\int_{{\mathbb R}^n} u\mathcal L u + 2\int_{{\mathbb R}^n}\int_{{\mathbb R}^n} (u(x)-u(y))^2K(x-y)\,dxdy.
$$
\subsection{Some integral operators on geometric spaces}
In this section, we describe new operators involving curvature terms. These operators appear naturally in harmonic analysis, as described below. They are of the form
$$
\mathcal L u(x)=\int (u(x)-u(y))K(x,y)\,dy
$$
where the non-negative kernel $K$ is symmetric and has some geometric meaning. The integral sign runs either over a Lie group or over a Riemannian manifold. By exactly the same argument as in the previous section, one deduces trivially C\'ordoba-C\'ordoba estimates for these operators. We now describe these new operators.
\subsubsection*{The case of Lie groups}
Let $G$ be a unimodular connected Lie group endowed with the Haar measure $dx$. By ``unimodular'', we mean that the Haar measure is left and right-invariant. If we denote by $\mathcal G$ the Lie algebra of $G$, we consider a family
$$\mathbb X= \left \{ X_1,...,X_k \right \}$$
of left-invariant vector fields on $G$ satisfying the H\"ormander condition, i.e. $\mathcal G$ is the Lie algebra generated by the $X_i's$. A standard metric on $G$ , called the Carnot-Caratheodory metric, is naturally associated with $\mathbb X$ and is defined as follows: let $\ell : [0,1] \to G$ be an absolutely continuous path. We say that $\ell$ is admissible if there exist measurable functions $a_1,...,a_k : [0,1] \to \mathbb C$ such that, for almost every $t \in [0,1]$, one has
$$\ell'(t)=\sum_{i=1}^k a_i(t) X_i(\ell(t)).$$
If $\ell$ is admissible, its length is defined by
$$|\ell |= \int_0^1\left(\sum_{i=1}^k |a_i(t)|^2 \,dt \right)^{ \frac 12 }.$$
For all $x,y \in G $, define $d(x,y)$ as the infimum of the lengths of all admissible paths joining $x$ to $y$ (such a curve exists by the H\"ormander condition). This distance is left-invariant. For short, we denote by $|x|$ the distance between $e$, the neutral element of the group and $x$, so that the distance from $x$ to $y$ is equal to $|y^{-1}x|$.
For all $r>0$, denote by $B(x,r)$ the open ball in $G$ with respect to the Carnot-Caratheodory distance and by $V(r)$ the Haar measure of any ball. There exists $d\in \mathbb{N}^{\ast}$ (called the local dimension of $(G,\mathbb X)$) and $0<c<C$ such that, for all $r\in (0,1)$,
$$
cr^d\leq V(r)\leq Cr^d,
$$
see \cite{nsw}. When $r>1$, two situations may occur (see \cite{guivarch}):
\begin{itemize}
\item Either there exist $c,C,D >0$ such that, for all $r>1$,
$$c r^D \leq V(r) \leq C r^D$$
where $D$ is called the dimension at infinity of the group (note that, contrary to $d$, $D$ does not depend on $\mathbb X$). The group is said to have polynomial volume growth.
\item Or there exist $c_1,c_2,C_1,C_2 >0$ such that, for all $r>1$,
$$c_1 e^{c_2r} \leq V(r) \leq C_1 e^{C_2r}$$
and the group is said to have exponential volume growth.
\end{itemize}
When $G$ has polynomial volume growth, it is plain to see that there exists $C>0$ such that, for all $r>0$,
\begin{equation} \label{homog}
V(2r)\leq CV(r),
\end{equation}
which implies that there exist $C>0$ and $\kappa>0$ such that, for all $r>0$ and all $\theta>1$,
\begin{equation} \label{homogiter}
V(\theta r)\leq C\theta^{\kappa}V(r).
\end{equation}
On a Lie group as previously described, one introduces the Kohn sub-laplacian
$$
\Delta_G=\sum_{i=1}^k X_i^2.
$$
On any Lie group $G$, it is natural by functional calculus to define the fractional powers $(-\Delta_G)^s$, $s \in (0,1)$ of the Kohn sub-laplacian $-\Delta_G$. It has been proved in \cite{MRS, RS} (see also \cite{SW}) that for Lie groups with polynomial volume
$$
\|(-\Delta_G)^{s/2}u\|^2_{L^2(G)} \leq C \int_{G \times G} \frac{|u(x)-u(y)|^2}{V(|y^{-1}x|)|y^{-1}x|^{2s}}\,dx \,dy.
$$
It is therefore natural to consider the operator which is the Euler-Lagrange operator of the Dirichlet form in the R.H.S. of the previous equation given by
$$
\mathcal Lu(x)=\int_G \frac{u(x)-u(y)}{V(|y^{-1}x|)|y^{-1}x|^{2s}}\,dy.
$$
It defines a new Gagliardo-type norm, suitably designed for Lie groups (of any volume growth). By the structure itself of this norm, one can prove as before a C\'ordoba-C\'ordoba inequality.
\subsubsection*{The case of manifolds}
Let $M$ be a complete riemannian manifold of dimension $n$. Denote $d(x,y)$ the geodesic distance from $x$ to $y$. Similarly to the previous case it is natural to introduce the new operators, Euler-Lagrange of suitable Gagliardo norms, given by
$$
\mathcal Lu(x)=\int_{M} \frac{u(x)-u(y)}{d(x,y)^{n+2s}}\,dy
$$
These new operators also satisfy C\'ordoba-C\'ordoba estimates (see \cite{RS} for an account in harmonic analysis where these quantities pop up).
\section{A review of the extension property}
\subsection{The extension property in ${\mathbb R}^n$}\label{extRn}
We first introduce the spaces
$$H^s({\mathbb R}^n)=\left \{ v \in L^2({\mathbb R}^n)\,\,:\,\,|\xi|^{s} (\mathcal F v)(\xi) \in L^2({\mathbb R}^n)
\right \},$$
where $s \in (0,1)$ and $\mathcal F$ denotes Fourier transform.
For $\Omega \subset {\mathbb R}^{n+1}_+$ a Lipschitz domain (bounded or unbounded) and $a \in (-1,1)$,
we denote
$$H^1(\Omega,y^a)=\left \{ u \in L^2
(\Omega,y^a \,dx\,dy)\,\,:\,\,|\nabla u| \in L^2(\Omega,y^a \,dx\,dy) \right \}.$$
Let $a=1-2s$. It is well known that the space $H^s({\mathbb R}^n)$
coincides with the trace on $\partial{\mathbb R}^{n+1}_{+}$ of $H^1({\mathbb R}^{n+1}_+,y^a)$.
In particular, every $v\in H^s({\mathbb R}^n)$ is the trace of a function $u\in L^2_{\rm loc}({\mathbb R}^{n+1}_+,y^a)$
such that $\nabla u \in L^2({\mathbb R}^{n+1}_+,y^a)$.
In addition, the function $u$ which minimizes
\begin{equation} \label{argmin}
{\rm min}\left\{ \int_{{\mathbb R}^{n+1}_{+}}y^a \left| \nabla u \right|^2\;dx dy \; : \;
u|_{\partial{\mathbb R}^{n+1}_{+}}=v\right\}
\end{equation}
solves the Dirichlet problem
\begin{equation}\label{bdyFrac2}
\left \{
\begin{aligned}
L_a u:= \textrm{div\,} (y^a \nabla u)&=0 \qquad
{\mbox{ in ${\mathbb R}^{n+1}_+$}}
\\
u&= v
\qquad{\mbox{ on $\partial{\mathbb R}^{n+1}_+$.}}\end{aligned}\right .
\end{equation}
By standard elliptic regularity, $u$ is smooth in ${\mathbb R}^{n+1}_{+}$. It turns out that
$-y^a u_{y} (\cdot,y)$ converges in $H^{-s}({\mathbb R}^n)$ to a distribution
$h\in H^{-s}({\mathbb R}^n)$ as $y\downarrow 0$. That is, $u$ weakly solves
\begin{equation}\label{bdyFrac3}
\left \{
\begin{aligned}
\textrm{div\,} (y^a \nabla u)&=0 \qquad
{\mbox{ in ${\mathbb R}^{n+1}_+ $}}
\\
-y^a \partial_y u &= h
\qquad{\mbox{ on $\partial{\mathbb R}^{n+1}_+$.}}\end{aligned}\right .
\end{equation}
Consider the Dirichlet to Neumann operator
$$
\begin{aligned}
&\Gamma_a: H^s({\mathbb R}^n)\to H^{-s}({\mathbb R}^n)\\
&\qquad \quad v\mapsto \Gamma_{a}(v)= h:=
- \displaystyle{\lim_{y \rightarrow 0^+}} y^{a} \partial_y u=\frac{\partial u}{\partial \nu^a},
\end{aligned}
$$
where $u$ is the solution of \eqref{bdyFrac2}.
Then, we have:
\begin{theorem}[\cite{CS}]\label{realization}
For every $v\in H^s({\mathbb R}^n)$,
\begin{equation*}
(-\Delta)^s v= d_s\Gamma_{a}(v)=
- d_s \displaystyle{\lim_{y \rightarrow 0^+}} y^{a} \partial_y u,
\end{equation*}
where $a=1-2s$, $d_s$ is a positive constant depending only on~$s$,
and the equality holds in the distributional sense.
\end{theorem}
\subsection{The extension property in bounded domains}
We consider now the case of bounded domains. In this case, two different operators can be defined.
\noindent $\bullet$ {\sl The spectral Laplacian: } If one considers the classical Dirichlet Laplacian $\Delta_{\Omega}$ on the domain $\Omega$\,, then the {spectral definition} of the fractional power of $\Delta_{\Omega}$ relies on the following formulas:
\begin{equation}\label{sLapl.Omega.Spectral}
\displaystyle(-\Delta_{\Omega})^{s}
g(x)=\sum_{j=1}^{\infty}\lambda_j^s\, \hat{g}_j\, \phi_j(x)
=\frac1{\Gamma(-s)}\int_0^\infty
\left(e^{t\Delta_{\Omega}}g(x)-g(x)\right)\frac{dt}{t^{1+s}}.
\end{equation}
Here $\lambda_j>0$, $j=1,2,\ldots$ are the eigenvalues of the Dirichlet Laplacian on $\Omega$ with zero boundary conditions\,, written in increasing order and repeated according to their multiplicity and $\phi_j$ are the corresponding normalized eigenfunctions, namely
\[
\hat{g}_j=\int_\Omega g(x)\phi_j(x)\, dx\,,\qquad\mbox{with}\qquad \|\phi_j\|_{L^2(\Omega)}=1\,.
\]
The first part of the formula is therefore an interpolation definition. The second part gives an equivalent definition in terms of the semigroup associated to the Laplacian.
We will denote the operator defined in such a way as $\mathcal A_{1,s}=(-\Delta_{\Omega})^s$\,, and call it the \textit{spectral fractional Laplacian}.
\medskip
\noindent $\bullet$~{\sl The restricted fractional laplacian:} On the other hand, one can define a fractional Laplacian operator by using the integral representation in terms of hypersingular kernels already mentioned
\begin{equation}\label{sLapl.Rd.Kernel}
(-\Delta_{{\mathbb R}^d})^{s} g(x)= C_{d,s}\mbox{
P.V.}\int_{\mathbb{R}^n} \frac{g(x)-g(z)}{|x-z|^{n+2s}}\,dz,
\end{equation}
In this case we materialize the zero Dirichlet condition by restricting the operator to act only on functions that are zero outside $\Omega$. We will call the operator defined in such a way the \textit{restricted fractional Laplacian} and use the specific notation $\mathcal A_{2,s}=(-\Delta_{|\Omega})^s$ when needed. As defined, $\mathcal A_{2,s}$ is a self-adjoint operator on $L^2(\Omega)$\,, with a discrete spectrum: we will denote by $\lambda_{s, j}>0$, $j=1,2,\ldots$ its eigenvalues written in increasing order and repeated according to their multiplicity and we will denote by $\{\phi_{s, j}\}_j$ the corresponding set of eigenfunctions, normalized in $L^2(\Omega)$.
\medskip
\noindent $\bullet$~{\sl Common notation.}
In the sequel we use $\mathcal A$ to refer to any of the two types of operators $\mathcal A_{1,s}$ or $\mathcal A_{2,s}$, $0<s<1$. Each one is defined on a Hilbert space
\begin{equation} \label{defH}
H(\Omega)=\{u=\sum_{k=1}^\infty u_k \phi_{s,k} \in L^2(\Omega)\; : \; \| u \|^2_{H} = \sum_{k=1}^\infty \lambda_{s,k}\vert u_k\vert^2 <+\infty\}\subset L^2(\Omega)
\end{equation}
with values in its dual $H^*$. The notation in the formula copies the one just used for the second operator. When applied to the first one we put here $\phi_{s,k}=\phi_k$, and $\lambda_{s,k}=\lambda_{k}^s$.
Note that $H(\Omega)$ depends in principle on the type of operator and on the exponent $s$. Moreover, the operator $\mathcal A$ is an isomorphism between $H$ and $H^*$, given by its action on the eigen-functions.
It has been proved in \cite{BSV} (see also \cite{CDDS}) that
$$
H(\Omega)=
\left\{
\begin{aligned}
H^s(\Omega)&\qquad\text{if $s\in (0,1/2)$},\\
H^{1/2}_{00}(\Omega)&\qquad\text{if $s=1/2$},\\
H^s_{0}(\Omega)&\qquad\text{if $s\in (1/2,1)$},
\end{aligned}
\right.
$$
We now introdruce the Caffarelli-Silvestre extension for these operators. In the case of the restricted fractional laplacian, the extension is precisely the one described in Section \ref{extRn}. We now concentrate on the case of the spectral fractional laplacian. Let us define
\begin{align*}
{\mathcal C} &= \Omega\times(0,+\infty),\\
\partial_L {\mathcal C}&=\partial \Omega \times [0,+\infty ).
\end{align*}
We write points in the cylinder using the notation $(x,y)\in {\mathcal C}=\Omega \times (0,+\infty)$. Given $s\in(0,1)$, it has been proved in \cite{CDDS} (see also \cite{CabreTan}) that the following holds.
\begin{lemma}\label{extCylindrical}
Consider a weak solution of
\begin{equation}
\left\{
\begin{array}{lll}
\mbox{\rm div}(y^{1-2s} \nabla w)=0 &\mbox{in }\,\,\mathcal C= \Omega \times (0,+\infty),\\
w=0\,,\; &\mbox{on }\,\,\partial \Omega \times (0,+\infty) \\
\end{array}
\right.
\end{equation}
Then $-\lim_{y \to 0} y^{1-2s}\partial_y w=\mathcal A w(\cdot,0).$
where $\mathcal A$ is the spectral fractional laplacian.
\end{lemma}
\subsection{The extension property in general frameworks}\label{extST}
To generalize the inequalities under consideration, one has to invoke a rather general version of the Caffarelli-Silvestre extension proved by Stinga and Torrea \cite{ST}. Their approach, based on semi-group theory, allows to prove the previous results in quite general ambient spaces, like Riemannian manifolds or Lie groups.
In the following theorem, we will consider three cases later for the object $\mathcal M$:
\begin{enumerate}
\item The case of complete Riemannian manifolds and the Laplace-Beltrami operator
\item The case of Lie groups and the Kohn laplacian
\item The case of the Wiener space and the Ornstein-Uhlenbeck operator
\end{enumerate}
Let $\mathcal L$ be a positive and self-adjoint operator in
$L^2(\mathcal M)$. One can define its fractional powers
by means of the standard formula in spectral theory
\[
\displaystyle \mathcal L^{s}=\frac1{\Gamma(-s)}\int_0^\infty
\left(e^{t\mathcal L}-\mbox{Id}\right)\frac{dt}{t^{1+s}},
\]
where $s\in (0,1)$
and $e^{t\mathcal L}$ denotes the heat semi-group on $\mathcal M$. Then one has
\begin{theorem}\label{extension}
Let $u \in \text{dom}(\mathcal L^s)$. A solution of the extension problem
\[
\left \{ \begin{array}{ll}
\displaystyle{\mathcal L v+ \frac{1-2s}{y}\partial_yv + \partial^2_yv=0}\qquad &\mbox{on } \, \mathcal M\times \mathbb{R}^+
\\ \\
v(x,0)=u &\mbox{on } \, \mathcal M, \\
\end{array} \right .
\]
is given by
\[
v(x,y)=\frac{1}{\Gamma(s)}\int_0^\infty e^{t\mathcal L} (\mathcal L^su)(x)e^{-y^2/4t}\frac{dt}{t^{1-s}}
\]
and furthermore, one has at least in the distributional sense
\begin{equation}\label{eqcs}
-\lim_{y\to 0^+} y^{1-2s}\partial_yv(x,y)=
\frac{2s\Gamma(-s)}{4^s\Gamma(s)}\mathcal L^s u(x).
\end{equation}
\end{theorem}
\section{Proofs of Theorem \ref{CC} and \ref{kato}}
\subsection{Proof of Theorem \ref{CC}}
We now come to the proof of Theorem \ref{CC}. We introduce the function
$$
\tilde w = \varphi(w)-v
$$
where $w$ is the Caffarelli-Silvestre extension of $u$ and $v$ the Caffarelli-Silvestre extension of $\varphi(u)$. Then $\tilde w$ satisfies
\begin{equation*}
\left \{
\begin{array}{c}
L_a \tilde w=y^a \varphi '' (w) |\nabla w|^2 \geq 0, \,\,\,\,\,\mbox{in}\,\,\,{\mathbb R}^{n+1}_+\\
\tilde w=0\,\,\,\,\,\mbox{on}\,\,\,\partial {\mathbb R}^{n+1}_+
\end{array} \right .
\end{equation*}
since $\varphi $ is convex. Hence by the Hopf lemma in \cite{CS1} (see also the Appendix) ( notice $\tilde w \geq 0$ by the weak maximum principle) , one has $\frac{\partial \tilde w}{\partial \nu^a} > 0$, hence the result.
\subsection{Proof of Theorem \ref{kato}}
We now turn to the proof of the Kato inequality in Theorem \ref{kato}. This is a consequence of the Cordoba-Cordoba inequality. Indeed consider the convex function
$$
\varphi_\epsilon(x)=\sqrt{x^2+\epsilon^2}.
$$
Then the result follows by Theorem \ref{CC} and a standard approximation argument.
\subsection{The results in bounded domains}
In the case of the spectral laplacian, the C\'ordoba-C\'ordoba estimate has been proved by Constantin and Ignatova \cite{constantin} by a rather involved use of semi-group theory. Our proof has the same flavour as the one of Theorem \ref{CC}. Furthermore, in our framework, one can also prove the C\'ordoba-C\'ordoba estimate in the case of the restricted laplacian, which is not covered by \cite{constantin}.
\begin{thm}\label{CCbdd}
Let $\varphi$ be a $C^2({\mathbb R}^n)$ convex function. Assume that $u$ and $\varphi(u)$ are such that $\mathcal A u $ and $\mathcal A \varphi(u) $ exist where $\mathcal A$ is either the restricted or spectral fractional laplacian. Then the following holds
\begin{equation}\label{CCeq}
\mathcal A \varphi(u) \leq \varphi'(u) \mathcal A \,u
\end{equation}
\end{thm}
\begin{proof}
The case of the retricted laplacian is fully covered by the proof of Theorem \ref{CC} verbatim. In the case of the spectral fractional laplacian, one considers as before
$$
\tilde w = \varphi(w)-v
$$
where $w$ is the Caffarelli-Silvestre extension of $u$ and $v$ the Caffarelli-Silvestre extension of $\varphi(u)$ where the Caffarelli-Silvestre extension is the one described in Section \ref{extCylindrical}. Then $\tilde w$ satisfies
\begin{equation*}
\left \{
\begin{array}{c}
L_a \tilde w=y^a \varphi '' (w) |\nabla w|^2 \geq 0, \,\,\,\,\,\mbox{in}\,\,\,{\mathcal C}\\
\tilde w=0\,\,\,\,\,\mbox{on}\,\,\,\partial_L {\mathcal C} \\
\tilde w=0\,\,\,\,\,\mbox{on}\,\,\,\left \{ y=0 \right \}
\end{array} \right .
\end{equation*}
By the weak maximum principle, one has $\tilde w\geq 0$ in ${\mathcal C}$ and one concludes with the Hopf lemma in the appendix.
\end{proof}
\begin{remark}
Our proof of the estimate is the same as the one in C\'ordoba and Mart\'inez in \cite{CordobaAdv} for the Dirichlet-to-Neumann operator. However, their proof covers only the case $1/2$ and for power-like convex functions. The argument can be actually generalized as we mentioned. Furthermore, it unifies all the possible proofs of the C\'ordoba-C\'ordoba estimates.
\end{remark}
\section{Geometric ambiebent spaces}
\subsection{The case of manifolds}
The case of compact manifolds, through a parabolic argument, has been proved by Cordoba and Mart\'inez \cite{CordobaAdv}. Our proof once again completely unifies the several approaches. Consider a complete Riemannian manifold $\mathcal M$ and its Laplace-Beltrami operator
$$
\mathcal L=-\Delta_g
$$
Invoking now the extension of Stinga and Torrea described in Section \ref{extST}, one proves
\begin{thm}\label{CCA}
Let $\varphi$ be a $C^2({\mathbb R}^n)$ convex function. Assume that $u$ and $\varphi(u)$ are such that $\mathcal L u $ and $\mathcal L \varphi(u) $ exist. Then the following holds
\begin{equation}\label{CCeq}
\mathcal L \varphi(u) \leq \varphi'(u) \mathcal L \,u
\end{equation}
\end{thm}
We then recover the case of compact manifolds in \cite{CordobaAdv} and even generalize it to complete non-compact manifolds. The proof of the previous theorem is identical, once the extension is well defined as described above (see \cite{ST}), to the proof of Theorem \ref{CC}.
\subsection{The case of Lie groups}
Consider a Lie group $G$ with its Kohn Laplacian
$$
\mathcal L=-\Delta_G
$$
Invoking now the extension of Stinga and Torrea described in Section \ref{extST}, one proves
\begin{thm}\label{CCB}
Let $\varphi$ be a $C^2({\mathbb R}^n)$ convex function. Assume that $u$ and $\varphi(u)$ are such that $\mathcal L u $ and $\mathcal L \varphi(u) $ exist. Then the following holds
\begin{equation}\label{CCeq}
\mathcal L \varphi(u) \leq \varphi'(u) \mathcal L \,u
\end{equation}
\end{thm}
\subsection{The case of the Wiener space}
We start by recalling the basic notions about the Wiener space and its associated operators.
An abstract Wiener space is defined as a triple $(X,\gamma,H)$ where
$X$ is a separable Banach space, endowed with the norm $\|\cdot\|_X$,
$\gamma$ is a nondegenerate centred Gaussian measure,
and $H$ is the Cameron--Martin space associated with the measure $\gamma$, that is,
$H$ is a separable Hilbert space densely embedded in $X$, endowed with the inner product
$[\cdot, \cdot ]_H$ and with the norm $|\cdot |_H$. The
requirement that $\gamma$ is a centred Gaussian measure means that
for any $x^*\in X^*$, the measure $x^*_\#\gamma$ is a centred Gaussian
measure on the real line ${\mathbb R}$, that is, the Fourier transform
of $\gamma$ is given by
\[
\hat \gamma(x^*) = \int_X
e^{-i\scal{x}{x^*}}\, d\gamma (x)=\exp\left(
-\frac{\scal{Qx^*}{x^*}}{2}
\right),\qquad \forall x^*\in X^*;
\]
here the operator $Q\in {\mathcal L}(X^*,X)$ is the covariance operator and it is
uniquely determined by the formula
\[
\scal{Qx^*}{y^*}=\int_X \scal{x}{x^*}\scal{x}{y^*}d\gamma(x),\qquad \forall x^*,y^*\in X^*.
\]
The nondegeneracy of $\gamma$ implies that $Q$ is positive definite: the boundedness
of $Q$ follows by Fernique's Theorem,
asserting that there exists a positive number $\beta>0$
such that
\[
\int_X e^{\beta\|x\|^2}d\gamma(x)<+\infty.
\]
This implies also that the maps $x\mapsto \scal{x}{x^*}$ belong to
$L^p_\gamma(X)$ for any $x^*\in X^*$ and $p\in [1,+\infty)$, where $L^p_\gamma(X)$
denotes the space of all $\gamma$-measurable functions $f:X\to \mathbb{R}$
such that
\[
\int_X |f(x)|^p d\gamma(x)<+\infty.
\]
In particular, any element $x^*\in X^*$ can be seen as a map $x^*\in L^2_\gamma(X)$, and
we denote by $R^*: X^*\to {\mathcal H}$ the identification map $R^*x^*(x):=\scal{x}{x^*}$.
The space ${\mathcal H}$ given by the closure of $R^*X^*$ in $L^2_\gamma(X)$
is usually called reproducing kernel.
By considering the map $R: {\mathcal H}\to X$
defined as
\[
R\hat{h} := \int_X \hat{h}(x)x\, d\gamma(x),
\]
we obtain that $R$ is an injective $\gamma$--Radonifying operator, which is
Hilbert--Schmidt when $X$ is Hilbert. We also have
$Q=RR^*:X^*\to X$. The space $H:=R{\mathcal H}$, equipped with the inner product $[\cdot,\cdot]_H$
and norm $|\cdot|_H$ induced by ${\mathcal H}$ via $R$, is the Cameron-Martin space
and is a dense subspace of $X$. The continuity of $R$ implies
that the embedding of $H$ in $X$ is continuous, that is, there exists $c>0$
such that
\[
\|h\|_X \leq c|h|_H,\qquad \forall h\in H.
\]
We have also that the measure $\gamma$ is absolutely continuous with respect
to translation along Cameron--Martin directions; in fact, for $h\in H$, $h=Qx^*$,
the measure
$\gamma_h(B)=\gamma(B-h)$ is absolutely continuous with respect to $\gamma$ with
density given by
\[
d\gamma_h(x)=\exp\left(
\scal{x}{x^*}-\frac{1}{2}|h|_H^2
\right)d\gamma(x) .
\]
For $j\in \mathbb{N}$ we choose $x^*_j\in X^*$ in such a way that $\hat h_j:= R^*x_j^*$, or equivalently
$h_j:=R\hat h_j=Qx^*_j$, form an orthonormal basis of $H$. We order the vectors $x^*_j$ in such
a way that the numbers $\lambda_j:=\|x^*_j\|_{X^*}^{-2}$ form a non-increasing sequence.
Given $m\in\mathbb N$, we also let $H_m:=\langle h_1,\ldots, h_m\rangle\subseteq H$,
and $\Pi_m: X\to H_m$ be the closure of the orthogonal projection from $H$ to $H_m$
\[
\Pi_m(x) := \sum_{j=1}^m \scal{x}{x^*_j}\, h_j \qquad x\in X.
\]
The map $\Pi_m$ induces the decomposition $X\simeq H_m\oplus X_m^\perp$, with $X_m^\perp:= {\rm ker}(\Pi_m)$,
and $\gamma=\gamma_m\otimes\gamma_m^\perp$,
with $\gamma_m$ and $\gamma_m^\perp$ Gaussian measures on $H_m$ and $X_m^\perp$ respectively,
having $H_m$ and $H_m^\perp$ as Cameron--Martin spaces.
When no confusion is possible we identify $H_m$ with ${\mathbb R}^m$;
with this identification the measure
$\gamma_m={\Pi_m}_\#\gamma$ is the standard Gaussian measure on ${\mathbb R}^m$ (see \cite{B}).
Given $x\in X$,
we denote by $\underline x_m\in H_m$ the projection $\Pi_m(x)$,
and by $\overline x_m\in X_m^\perp$ the infinite dimensional component of $x$, so that
$x=\underline x_m+\overline x_m$.
When we identify $H_m$ with ${\mathbb R}^m$ we rather write
$x=(\underline x_m,\overline x_m)\in {\mathbb R}^m\times X_m^\perp$.
We say that $u:X\to {\mathbb R}$ is a {\em cylindrical function} if $u(x)=v(\Pi_m (x))$ for some $m\in\mathbb N$ and
$v:{\mathbb R}^m\to {\mathbb R}$.
We denote by ${\mathcal F C}_b^k(X)$, $k\in\mathbb N$,
the space of all $C^k_b$ cylindrical functions, that is, functions of the form $v(\Pi_m (x))$
with $v\in C^k({\mathbb R}^n)$, with continuous and bounded derivatives up to the order $k$.
We denote by ${\mathcal F C}_b^k(X,H)$ the space generated by all functions of the form
$u h$, with $u\in {\mathcal F C}_b^k(X)$ and $h\in H$.
Given $u\in L^2_\gamma(X)$, we consider the canonical cylindrical approximation $\mathbb{E}_m$ given by
\begin{equation}\label{cancylapprox}
\mathbb{E}_m u (x)=\int_{X_m^\perp} u(\Pi_m(x),y) \,d\gamma_m^\perp(y).
\end{equation}
Notice that $\mathbb{E}_m u$ depends only on the first $m$ variables
and $\mathbb{E}_m u$ converges to $u$ in $L^p_\gamma(X)$ for all $1\leq p<\infty$.
We let
\[
\begin{array}{ll}
\displaystyle{\nabla_\gamma u := \sum_{j\in\mathbb N}\partial_j u\, h_j} & {\rm for\ } u\in {\mathcal F C}_b^1(X)
\\
\\
\displaystyle{\diver_\gamma \varphi := \sum_{j\geq 1}\partial^*_j [\varphi,h_j]_H} & {\rm for\ }\varphi\in {\mathcal F C}_b^1(X,H)
\\
\\
\displaystyle{{\Delta_\gamma} u := \diver_\gamma\nabla_\gamma u} & {\rm for\ } u\in {\mathcal F C}_b^2(X)
\end{array}
\]
where $\partial_j := \partial_{h_j}$ and
$\partial_j^* := \partial_j - \hat h_j$ is the adjoint operator of $\partial_j$.
With this notation, the following integration by parts formula holds:
\begin{equation}\label{inp}
\int_X u\, \diver_\gamma \varphi\,d\gamma = -\int_X [\nabla_\gamma u,\varphi]_H\, d\gamma
\qquad \forall \varphi\in {\mathcal F C}_b^1(X,H).
\end{equation}
In particular, thanks to \eqref{inp}, the operator $\nabla_\gamma$ is closable in $L^p_\gamma(X)$,
and we denote by $W^{1,p}_\gamma(X)$ the domain of its closure. The Sobolev spaces
$W^{k,p}_\gamma(X)$, with $k\in\mathbb N$ and $p\in [1,+\infty]$, can be defined analogously \cite{B},
and ${\mathcal F C}_b^k(X)$ is dense in $W^{j,p}_\gamma(X)$, for all $p<+\infty$ and $k,j\in\mathbb N$ with $k\ge j$.
Given a vector field $\varphi \in L^{p}_\gamma(X;H)$, $p\in (1,\infty]$, using \eqref{inp} we can define
$\mathrm{div}_\gamma \, \varphi$ in the distributional sense,
taking test functions $u$ in $W^{1,q}_\gamma(X)$ with
$\frac{1}{p}+\frac{1}{q} = 1$. We say that
$\mathrm{div}_\gamma\, \varphi \in L^p_\gamma(X)$ if this linear functional can be extended to all test
functions $u\in L^{q}_\gamma(X)$. This is true in particular if $\varphi\in W^{1,p}_\gamma(X;H)$.
Let $u\in W^{2,2}_\gamma(X)$, $\psi\in {\mathcal F C}_b^1(X)$ and $i,j\in \mathbb N$.
{}From \eqref{inp}, with $u=\partial_j u$ and $\varphi=\psi h_i$, we get
\begin{equation}\label{parts}
\int_X \partial_j u\,\partial_{i}\psi \,d\gamma =
\int_X -\partial_i(\partial_{j}u)\,\psi+ \partial_ju\,\psi\langle x,x^*_i\rangle d\gamma
\end{equation}
Let now $\varphi\in {\mathcal F C}_b^1(X,H)$. If we apply \eqref{parts} with $\psi=[\varphi,h_j]_H=:\varphi^j$, we obtain
\[
\int_X \partial_j u\,\partial_{i}\varphi^j \,d\gamma =
\int_X -\partial_j(\partial_{i}u)\,\varphi^j
+ \partial_ju\,\varphi^j\langle x,x^*_i\rangle d\gamma
\]
which, summing up in $j$, gives
\[
\int_X [\nabla_\gamma u,\partial_i \varphi]_H\,d\gamma = \int_X -[\nabla_\gamma (\partial _i u), \varphi]_H
+ [\nabla_\gamma u,\varphi]_H \langle x,x^*_i\rangle d\gamma
\]
for all $\varphi\in {\mathcal F C}_b^1(X,H)$.
The operator ${\Delta_\gamma}:W^{2,p}_\gamma(X)\to L^p_\gamma(X)$ is usually called the
Ornstein-Uhlenbeck operator on $X$.
Notice that, if $u$ is a cylindrical function, that is $u(x)=v(y)$ with
$y=\Pi_m(x)\in\mathbb{R}^m$ and $m\in\mathbb N$,
then
\[
{\Delta_\gamma} u = \sum_{j=1}^m \partial_{jj}u-\langle x,x_j^*\rangle\partial_{j}u =
\Delta v - \langle \nabla v,y \rangle_{\mathbb{R}^m}\,.
\]
We write $u\in C(X)$ if $u: X\to \mathbb{R}$ is continuous and $u\in C^1(X)$ if
both $u: X\to \mathbb{R}$ and $\nabla_\gamma u:X\to H$ are continuous.
{F}or simplicity of notation, from now on we omit the explicit dependence on $\gamma$
of operators and spaces.
We also indicate by $[\cdot, \cdot ]$ and $|\cdot |$ respectively the inner product and the norm in $H$.
By means of Section \ref{extST}, one can prove an extension property for the operator $(-\Delta_\gamma)^s$ and one proves in this case also a C\'ordoba-C\'ordoba estimate.
\section{Appendix}
In this appendix, we provide the Hopf lemma, which is crucial in the proof of the estimates. We state the theorem in the case of ${\mathbb R}^n$ as stated in \cite{CS}. However, an inspection of the proof shows that it is extendable to cylinders $\mathcal M \times (0,+\infty)$ where $\mathcal M$ is one of the cases covered in the present note and the associated operators. Indeed, the geometry is always the same and the Hopf lemma just depends on the structure of the equation.
We start with some notations. We introduce
\begin{align*}
& B_R^+=\{ (x,y)\in\mathbb{R}^{n+1} : y>0, |(x,y)|<R\}, \\
& \Gamma_R^0=\{ (x,0)\in\partial\mathbb{R}^{n+1}_+ : |x|<R\}, \\
& \Gamma_R^+=\{ (x,y)\in\mathbb{R}^{n+1} : y\ge 0, |(x,y)|=R\}.
\end{align*}
\begin{lemma}\label{hopf}
Consider the cylinder
$\mathcal C_{R,1}= \Gamma_R^0 \times (0,1) \subset {\mathbb R}^{n+1}_+$ where $\Gamma_R^0$
is the ball of center $0$ and radius $R$ in ${\mathbb R}^n$. Let $u \in C(\overline{\mathcal C_{R,1}})
\cap H^1(\mathcal C_{R,1},y^a)$ satisfy
\begin{equation*}
\begin{cases}
L_a u \leq 0&\text{ in } \mathcal C_{R,1} \\
u> 0&\text{ in } \mathcal C_{R,1} \\
u(0,0)=0.&
\end{cases}
\end{equation*}
Then,
$$\limsup_{y \rightarrow 0^+ }-y^a \frac{u(0,y)}{y}<0.$$
In addition, if $y^a u_y \in C(\overline{\mathcal C_{R,1}})$, then
$$\partial_{\nu^a} u (0,0) <0. $$
\end{lemma}
\section*{Acknowledgements}
Luis Caffarelli is supported by NSF grant DMS-1160802.
\bibliographystyle{alpha}
|
\section{}
\section{Microwave Amplitude Extraction}
To account for the $x$ axis variation in B$_{\mathrm{dc}}$ that the laser integrates over as it passes through the vapor cell, we take data in both the time and frequency domains, and perform a simultaneous fit to both data sets. Broadly speaking, the frequency domain data constrains the B$_{\mathrm{dc}}$ variation, and the time domain data gives us the microwave Rabi frequency. In future high resolution setups, the vapor cell will be sufficiently thin that B$_{\mathrm{dc}}$ variation through the cell will be negligible, rendering this frequency domain scan and simultaneous fitting process unnecessary. For our current proof-of-principle setup, we fit the data using
\begin{align}\label{eq:fitting}
\mathrm{OD}_{\mathrm{mw}} = \int dx \Big[A\frac{\Omega_R^2}{\Omega_R^2 + \Delta(x)^2} \sin^2\Big(\tfrac{1}{2}\sqrt{\Omega_R^2 + \Delta(x)^2} \, dt_{\mathrm{mw}}\Big) \\
\times\exp(-dt_{\mathrm{mw}}/\tau_2)\nonumber + B \frac{\Omega_R^2}{\Omega_R^2 + \Delta(x)^2} \Big(1-\exp(-dt_{\mathrm{mw}}/\tau_1)\Big) \Big],
\end{align}
where $A$, $B$, $\Omega_R$, $\Delta(x)$, $\tau_1$, and $\tau_2$ are fit parameters. The first term describes Rabi oscillations with a coherence time $\tau_2$. The second term is phenomenological, accounting for the diffusion of atoms from neighbouring regions of the cell, with a time constant $\tau_1$. We neglect variation in $\Omega_R$ along the $x$ axis, as there was little variation in distance from the microwave source for a given $y$-$z$ position through the cell. We include the B$_{\mathrm{dc}}$ inhomogeneity along the $x$ axis by using $\Delta(x) = \delta \omega_0 - \delta \omega_2 (x-x_0)^2 - \delta \omega_4 (x-x_0)^4$, where $x_0$ is the $x$ axis centre of the B$_{\mathrm{dc}}$ field. This form is derived from the measured B$_{\mathrm{dc}}$ variation along the $y$ axis (Fig.~4(b) in the main text), justified by the cylindrical symmetry of the solenoid.
For the pixel at ($z=0$~mm, $y=0$~mm), shown in Figs.~4(c+d) of the main text, the fit parameters are $A=2.1$, $B=0.88$, $\tau_1=15\,\mu$s, $\tau_2=18\,\mu$s, $x_0= 0.55\,\mathrm{mm}$, $\delta \omega_0=2\pi\,4.4\,$kHz, $\delta \omega_2 = 2\pi\,260\,\mathrm{kHz}/\mathrm{mm}^2$, $\delta \omega_4= 2\pi\,250\,\mathrm{kHz}/\mathrm{mm}^4$, and $\Omega_{\mathrm{R},\Delta(x)}=2\pi\,45\,$kHz. The coupling constant between the Rabi frequency and microwave amplitude is essentially a constant value of $\alpha_{18}=0.492$ over the cell. The extracted Rabi frequency of $\Omega_{\mathrm{R},\Delta(x)}=2\pi\,45\,$kHz thus corresponds to B$_{\mathrm{mw}}=1.6\,\mu$T. The coherence time is around $10\%$ of the optically pumped population lifetime (T$_1$), with the dominant dephasing mechanism caused by fluctuations in Zeeman shift due to atomic motion through the B$_{\mathrm{dc}}$ inhomogeneities.
\section{Further B$_{\mathrm{mw}}$ Detection Details}
Due to the large dc magnetic fields and spatial constraints imposed by the solenoid, we avoided using resistive heaters to control the cell temperature. Instead, the cell was placed between two 2~mm thick pieces of RG9 glass (strongly absorptive at 1500~nm, with better than $90\%$ transmission at 780~nm), and we heated the glass using a 2~W laser at the wavelength of 1500~nm. The direct heating of the cell windows ensures that there is minimal build-up of Rb on the windows, and the localised cell heating is advantageous for detecting microwave fields produced by temperature-sensitive devices. We were able to operate at relatively high temperatures with minimal reduction in T$_1$ lifetime, due to the suppression of Rb-Rb spin exchange relaxation in large dc magnetic fields~\cite{Kadlecek2001}.
For the data presented in Fig.~3 of the main text, the 780~nm laser intensity was $5\,\mathrm{mW}/\mathrm{cm}^2$ and the beam diameter was 0.6~mm. The data in Fig.~3(b) was taken at a cell temperature of $99^{\circ}$C.
For the data presented in Fig.~4 of the main text, the 780~nm laser beam was expanded to cover the entire vapor cell, with an intensity of $120\,\mathrm{mW}/\mathrm{cm}^2$ averaged over the central 3~mm. The cell temperature was $115^{\circ}$C. We repeated the frequency and time domain scans 3 and 5 times, respectively. Fitting was performed on the averaged data, however before further analysis, we binned the CCD pixels into $3\times3$ blocks to create $35.5\times35.5\,\mu\mathrm{m}^2$ image pixels. This reduced the computational intensity of the data analysis, with minimal reduction in image quality, as the image pixel size was still significantly smaller than the $0.11\,\mathrm{mm}$ diffusion distance of atoms during their coherence lifetime (for $P_{\mathrm{fill}}=63\,\mathrm{mbar}$ N$_2$, $T_{\mathrm{fill}}=80^{\circ}$C, $T_{\mathrm{cell}}=115^{\circ}$C, we have $D=3.4\,\mathrm{cm}^2/\mathrm{s}$. Using $\Delta x=\sqrt{2 D\,dt}$ and $dt=18\,\mu$s, $\Delta x=0.11\,\mathrm{mm}$). The frequency domain data was taken less than 5 minutes after the time domain data, allowing us to minimise the effects of drifts in the B$_{\mathrm{dc}}$ field, which occurred with a timescale of several minutes to tens of minutes. These drifts also complicated the practical implementation of shimming fields to homogenise the solenoid field. However, we note that no attempt was made to actively stabilise the solenoid, and that solenoid stability and homogeneity have likely advanced in the decades since the solenoid was purchased in 1974.
The metallic poles of the magnet will impose boundary conditions on the microwaves. This may be a significant perturbation for the detection of far-field microwaves, as the 26~mm separation of our solenoid poles is comparable to microwave wavelengths. However, we do not expect the poles to be a problem in the detection of microwave near-fields hundreds of micrometers to a few millimeters above a source, which would be the operation mode for microwave device characterisation as in Ref.~\cite{Horsley2015}.
Regarding the variation of the B$_{\mathrm{dc}}$ field along the $x$ axis, we can define three regions: a `good' central region, where the B$_{\mathrm{dc}}$ field is relatively flat and atoms are near resonance with the microwave; two `bad' middle regions, in front of and behind the `good' region, where the detuning is small enough for the atoms to still undergo Rabi oscillations, but at a significantly different oscillation frequency to the `good' region, thus acting to wash out the `good' signal; and two `neutral' outer regions, where atoms are so far detuned from the microwave that they do not interact. We minimised the `bad' effects by placing the vapor cell mount on a translation stage, and adjusting the cell position to minimise the double-resonance linewidth. In doing so, we aligned the front of the cell with the edge of the `good' region, meaning that the atoms only saw one `bad' region, and an extended `neutral' region. The sensing region for B$_{\mathrm{mw}}$ was primarily in the `good' region, thinner than the full 2~mm thickness of the cell.
\section{Hyperfine Transitions in an Arbitrary DC Magnetic Field}\label{sec:hyperfine_trasitions_arbitrary_Bdc}
\begin{figure}[t!]
\centering
\includegraphics[width=8.7cm]{Rb87_TransitionFrequencies.pdf}%
\caption{Hyperfine transition frequencies as a function of applied dc magnetic field. The legend lists the transitions in order of decreasing frequency. The $\sigma_+$ transitions are shown in red, $\pi$ transitions in green, and $\sigma_-$ transitions are shown in blue. Dashed lines are used for clarity.}
\label{fig:Rb87_TransitionFrequencies}
\end{figure}
The Hamiltonian for the hyperfine splitting of an atom in an external dc magnetic field \textbf{B}$_{\mathrm{dc}} = B_z \uv{z}$ is
\begin{equation}\label{eq:freqtun_hamiltonian}
H = H_{\mathrm{hfs}} + H_Z = A_{\mathrm{hfs}} \textbf{I} \cdot \textbf{J} + \mu_B(g_I I_z + g_J J_z) B_z,
\end{equation}
where $H_{\mathrm{hfs}}$ is the hyperfine coupling Hamiltonian, $H_Z$ is the Zeeman Hamiltonian, $\textbf{I}$ and $\textbf{J}$ are the atomic nuclear and electronic spin, respectively, $g_I$ and $g_J$ are the corresponding $g$-factors, and $A_{\mathrm{hfs}}$ is the hyperfine coupling constant~\cite{Steck87}.
The energies of each hyperfine level can be obtained numerically from the eigenvalues of $H$, or analytically by using the Breit-Rabi formula. This is shown as a function of dc magnetic field for the $^{87}$Rb 5S$_{1/2}$ hyperfine levels in Fig.~1 of the main text. The levels are labelled A$_1 \rightarrow \mathrm{A}_8$, in order of increasing energy ($\mathrm{E}_1<\mathrm{E}_2 ... <\mathrm{E}_8$). The resulting hyperfine transition frequencies are shown in Fig.~\ref{fig:Rb87_TransitionFrequencies}. Transitions between levels A$_i$ and A$_f$ are labelled T$_{if}$, with energies E$_{if}=\mathrm{E}_f-\mathrm{E}_i$. The $\sigma_+$ transitions are shown in red, $\pi$ transitions in green, and $\sigma_-$ transitions are shown in blue.
The $H_Z$ term in Eq.~(\ref{eq:freqtun_hamiltonian}) means that as the magnetic field is scanned, the eigenfunctions of the Hamiltonian must change. It is therefore best to describe the hyperfine levels in some field-independent basis, for which the $\ket{I,m_I,J,m_J}$ basis is a convenient choice. As $I=3/2,J=1/2$ for all of the levels, we abbreviate our notation to $\ket{m_I, m_J}$. The composition of each of the hyperfine levels in this basis is given Table~\ref{tbl:level_notation}. In general, each hyperfine level is a superposition of two $\ket{m_I, m_J}$ states, and the coefficients $a$ and $b$ can be determined by numerically diagonalising $H$. The two stretched states, $\ket{F=2,m_F=\pm2} \leftrightarrow \ket{m_I=\pm 3/2, m_J=\pm 1/2}$, are comprised of only a single $\ket{m_I, m_J}$ for all fields. In the absence of dc magnetic field, $a$ and $b$ are given by the Clebsch-Gordon coefficients, listed in Table~\ref{tbl:ab_coefficients_noBfield}. As the dc field strength increases, $a\rightarrow 1$ and $b\rightarrow 0$.
\begin{table}[t!]
\caption{Notation used for the $^{87}$Rb 5S$_{1/2}$ hyperfine levels. The levels A$_1 \rightarrow \mathrm{A}_8$ are in order of increasing energy.}
\centering
\begin{tabular}{c|llll}
& Weak Field & General & Strong Field \\
& ($\ket{F,m_F}$)& ($\ket{m_I,m_J}$) & ($\ket{m_I,m_J}$) \\
\hline
A$_1$ & $\ket{1,1}$ & $a_1 \ket{3/2,-1/2} + b_1 \ket{1/2,1/2}$ & $\ket{3/2,-1/2}$ \\
A$_2$ & $\ket{1,0}$ & $a_2 \ket{1/2,-1/2} + b_2 \ket{-1/2,1/2}$ & $\ket{1/2,-1/2}$ \\
A$_3$ & $\ket{1,-1}$ & $a_3 \ket{-1/2,-1/2} + b_3 \ket{-3/2,1/2}$ & $\ket{-1/2,-1/2}$ \\
A$_4$ & $\ket{2,-2}$ & $ \ket{-3/2,-1/2}$ & $\ket{-3/2,-1/2}$ \\
A$_5$ & $\ket{2,-1}$ & $a_5 \ket{-3/2,1/2} + b_5 \ket{-1/2,-1/2}$ & $\ket{-3/2,1/2}$ \\
A$_6$ & $\ket{2,0}$ & $a_6 \ket{-1/2,1/2} + b_6 \ket{1/2,-1/2}$ & $\ket{-1/2,1/2}$ \\
A$_7$ & $\ket{2,1}$ &$a_7 \ket{1/2,1/2} + b_7 \ket{3/2,-1/2}$ & $\ket{1/2,1/2}$ \\
A$_8$ & $\ket{2,2}$ & $ \ket{3/2,1/2}$ & $\ket{3/2,1/2}$ \\
\end{tabular}
\label{tbl:level_notation}
\end{table}
\begin{table}[t]
\caption{Coefficients $a$ and $b$ in the $\ket{m_I,m_J}$ basis for the $^{87}$Rb 5S$_{1/2}$ hyperfine levels in the absence of external static magnetic field. The coefficients are the relevant Clebsch-Gordon coefficients for each state. }
\centering
\begin{tabular}{c|lll}
& $\ket{F,m_F}$& $a$ & $b$ \\
\hline
A$_1$ & $\ket{1,1}$ & $\sqrt{3}/2$ & $-1/2$ \\
A$_2$ & $\ket{1,0}$ & $1/\sqrt{2}$ & $-1/\sqrt{2}$ \\
A$_3$ & $\ket{1,-1}$ & $1/2$ & $-\sqrt{3}/2$ \\
A$_4$ & $\ket{2,-2}$ & & \\
A$_5$ & $\ket{2,-1}$ & $1/2$ & $\sqrt{3}/2$ \\
A$_6$ & $\ket{2,0}$ & $1/\sqrt{2}$ & $1/\sqrt{2}$ \\
A$_7$ & $\ket{2,1}$ & $\sqrt{3}/2$ & $1/2$ \\
A$_8$ & $\ket{2,2}$ & & \\
\end{tabular}
\label{tbl:ab_coefficients_noBfield}
\end{table}
\begin{figure}[t!]
\centering
\includegraphics[width=0.5\textwidth]{Rb87_transitions_for_microwave_sensing.pdf}%
\caption{Strengths of the $\sigma_+$, $\pi$, and $\sigma_-$ hyperfine transitions within the $^{87}$Rb $5^2S_{1/2}$ ground state, as a function of the microwave transition frequency. The black vertical line is at 6.835~GHz. Data for $^{87}$Rb was taken from Ref.~\cite{Steck87}.}
\label{fig:transitions_for_microwave_sensing}
\end{figure}
The strengths of the hyperfine transitions also change with the dc magnetic field, and are proportional to the $\bra{f}J_{\gamma}\ket{i}$ matrix element, where $\gamma=-,\pi,+$ is the transition polarisation. We find that $\bra{f}J_{+}\ket{i}=a_ia_f$, $\bra{f}J_{-}\ket{i}=b_ib_f$, and $\bra{f}J_{\pi}\ket{i}=-\tfrac{1}{2}a_ib_f+\tfrac{1}{2}a_fb_i$, meaning that as the magnetic field strength is increased, $\sigma_+$ transitions strengths approach unity, whilst $\sigma_-$ transitions (rapidly) and $\pi$ transitions (slowly) become weak. The $\sigma_+$ transition strengths between states with the same $m_J$ value also go to zero at high dc fields. As a practical comparison, for $\bra{f}J_{\gamma}\ket{i}=1$, a B$_{\mathrm{mw}}=1\,\mu$T microwave field will drive $\Omega_R=2\pi\,28\,\mathrm{kHz}$ Rabi oscillations, whilst for the same microwave field, $\bra{f}J_{\gamma}\ket{i}=0.1$ results in $\Omega_R=2\pi\,2.8\,\mathrm{kHz}$.
\subsection{Hyperfine Transitions for Microwave Sensing}\label{sec:freqtun_sensing_transitions}
\begin{table*}[t]
\caption{Summary of the transitions available in various alkali species for sensing the $\sigma_+$, $\pi$, and $\sigma_-$ components of a microwave field. $I$ is the nuclear spin. The maximum $\sigma_+$ frequency is defined as the highest transition frequency available at our maximum solenoid field, B$_{\mathrm{dc}}=0.8\,\mathrm{T}$.}
\centering
\begin{tabular}{ccccccc}
Isotope & Abundance & I & E$_{\mathrm{hfs}}/h$ (GHz) & & Min. (GHz) & Max. (GHz) \\
\hline\hline
\multirow{3}{*}{$^{23}$Na}& \multirow{3}{*}{1} & \multirow{3}{*}{$3/2$} & \multirow{3}{*}{1.772} & $\sigma_+$ & 0 & 23.8 \\
& && & $\pi$ & 1.53 & 5.90 \\
& & & & $\sigma_-$ & 0.56 & 2.35 \\
\hline
\multirow{3}{*}{$^{39}$K}& \multirow{3}{*}{0.9326} & \multirow{3}{*}{$3/2$} & \multirow{3}{*}{0.462} & $\sigma_+$ & 0 & 22.8 \\
& && & $\pi$ & 0.40 & 1.54 \\
& && & $\sigma_-$ & 0.15 & 0.61 \\
\hline
\multirow{3}{*}{$^{85}$Rb}& \multirow{3}{*}{0.7217} & \multirow{3}{*}{$5/2$} & \multirow{3}{*}{3.036} & $\sigma_+$ & 0 & 25.0 \\
& && & $\pi$ & 2.26 & 10.1 \\
& && & $\sigma_-$ & 0.74 & 4.15 \\
\hline
\multirow{3}{*}{$^{87}$Rb}& \multirow{3}{*}{0.2783} & \multirow{3}{*}{$3/2$} & \multirow{3}{*}{6.835} & $\sigma_+$ & 0 & 27.9 \\
& & && $\pi$ & 5.92 & 22.8 \\
& && & $\sigma_-$ & 2.17 & 9.06 \\
\hline
\multirow{3}{*}{$^{133}$Cs}& \multirow{3}{*}{1} & \multirow{3}{*}{$7/2$} & \multirow{3}{*}{9.193} & $\sigma_+$ & 0 & 30.8 \\
& & && $\pi$ & 6.08 & 30.6 \\
& & & & $\sigma_-$ & 1.62 & 12.7 \\
\hline
\end{tabular}
\label{tbl:alkali_microwave_transitions}
\end{table*}
The important considerations when choosing a hyperfine transition for microwave sensing are: the microwave frequency of interest, the hyperfine transition strength, the optical resolution of the hyperfine transition states (i.e. the degree to which absorption due to each state can be distinguished), the microwave polarisation of interest, and the dc magnetic field required to tune a hyperfine transition to frequency of interest. Supplementary figure~\ref{fig:transitions_for_microwave_sensing} provides a useful analysis tool, showing the $\sigma_+$, $\pi$, and $\sigma_-$ transition strengths as a function of microwave transition frequency.
T$_{45}$ is the most versatile $\sigma_+$ transition, covering all microwave frequencies above dc. Above 0.6~GHz, it is also the strongest $\sigma_+$ transition for a given microwave frequency. The optical resolution of the neighbouring A$_4$ and A$_5$ states can be poor, however, particularly at low B$_{\mathrm{dc}}$ (corresponding to low microwave frequencies). For microwave sensing of frequencies above 6.835~GHz, the best $\sigma_+$ transition is therefore generally T$_{18}$. The A$_1$ and A$_8$ levels are maximally spectrally resolved from one another, and optical transitions from these levels enjoy minimal background absorption due to $^{85}$Rb. The T$_{18}$ transition is almost as strong as T$_{45}$ for a given microwave frequency, and requires much smaller B$_{\mathrm{dc}}$ to be tuned to a given frequency. For example, to achieve an 18~GHz microwave transition requires B$_{\mathrm{dc}}=0.44\,\mathrm{T}$ on the T$_{18}$ transition, but B$_{\mathrm{dc}}=0.81\,\mathrm{T}$ on the T$_{45}$ transition.
The selection of $\pi$ microwave transition is less clear-cut. T$_{26}$ is the strongest $\pi$ transition, but the difference with T$_{17}$ and T$_{35}$ is not dramatic. T$_{17}$ has the best optical resolution, due to the low $^{85}$Rb absorption background for optical transitions from A$_1$ and the large spectral separation of the A$_1$ and A$_7$ levels. The T$_{35}$ transition is first-order insensitive to dc magnetic fields around $B_{dc}=0.12\,\mathrm{T}$, corresponding to a microwave frequency of 5.92~GHz, and is thus the optimal transition around this point. Supplementary figure~\ref{fig:transitions_for_microwave_sensing} indicates that the $\pi$ transitions can be used for sensing microwaves even above 20~GHz, with the T$_{26}$ transition strength dropping to $\bra{6}J_{\pi}\ket{2}=0.15$ at 22.8~GHz.
The $\sigma_-$ transition strengths quickly drop away for microwave frequencies above 6.835~GHz, with the T$_{16}$ and T$_{25}$ transition strengths dropping to $\bra{f}J_-\ket{i}=0.15$ at 9.06~GHz. However, the T$_{34}$ transition can be used to detect microwaves below 6.835~GHz. The T$_{34}$ transition strength drops to $\bra{4}J_-\ket{3}=0.15$ at 2.17~GHz.
We can perform a similar analysis for other alkali species. The ranges of detectable frequencies for $\sigma_+$, $\pi$, and $\sigma_-$ polarised microwaves using $^{23}$Na, $^{39}$K, $^{85}$Rb, and $^{133}$Cs are summarised in Table~\ref{tbl:alkali_microwave_transitions}. The frequency range was defined as that for which there is a transition with a strength above $\bra{f}J_{\gamma}\ket{i}=0.15$, neglecting transitions between states with the same $m_J$ value. Strong $\sigma_+$ polarised transitions are available at all microwave frequencies and dc magnetic field strengths, and in order to compare the different alkali species, we took the maximum $\sigma_+$ frequency as the highest transition frequency available at our maximum solenoid field, B$_{\mathrm{dc}}=0.8\,\mathrm{T}$. Vapor cells filled with multiple species can be used to span larger frequency ranges. For example, a natural Rb cell provides $\pi$ transitions over the range $2.26-22.8\,\mathrm{GHz}$.
\section{Reconstruction of Microwave Fields of Arbitrary Frequency}
In this section, we provide a framework for reconstructing microwave magnetic fields of arbitrary frequency, using $^{87}$Rb atoms in an applied static magnetic field (\textbf{B}$_{\mathrm{dc}}$) of any strength. This builds on the framework given for the weak dc field regime in Ref.~\cite{Boehi2010a}. The framework is not restricted to $^{87}$Rb, and is valid for microwave transitions in a general system.
In the fixed lab-frame cartesian coordinate system, $(x,y,z)$, a microwave magnetic field is defined by
\[
\textbf{B} \equiv
\begin{pmatrix}
B_{x} e^{-i\phi_{x}}\\
B_{y} e^{-i\phi_{y}} \\
B_{z} e^{-i\phi_{z}} \\
\end{pmatrix}.
\]
In order to reconstruct this field, we need to find these six real values, $B_{x, y, z}$, $\phi_{x, y, z} \in \Re_{\geq 0}$, which are each an implicit function of spatial position. We do this by measuring Rabi oscillations driven by the microwave field on atomic hyperfine transitions. The quantisation axis of our measurements is defined by the applied dc magnetic field, \textbf{B}$_{\mathrm{dc}}$. Following Ref.~\cite{Boehi2010a}, we define a primed cartesian coordinate system, $(x',y',z')$, with the $z'$ axis pointing along the direction of \textbf{B}$_{\mathrm{dc}}$. The $\pi$ and $\sigma_{\pm}$ components of the microwave field in the primed frame are
\begin{equation}
B_- e^{-i\phi_-} \equiv \frac{1}{2} \Big[ B_{x'} e^{-i\phi_{x'}} + i B_{y'} e^{-i\phi_{y'}} \Big], \label{eq:B_-} \\
\end{equation}
\begin{equation}
B_{\pi} e^{-i\phi_{\pi}} \equiv B_{z'} e^{-i\phi_{z'}},\label{eq:B_pi} \\
\end{equation}
\begin{equation}
B_+ e^{-i\phi_+} \equiv \frac{1}{2} \Big[ B_{x'} e^{-i\phi_{x'}} - i B_{y'} e^{-i\phi_{y'}} \Big],\label{eq:B_+}
\end{equation}
with $B_{-, \pi, +}$, $\phi_{-, \pi, +} \in \Re_{\geq 0}$. For transitions from an initial state $\ket{1}$ to a final state $\ket{2}$, the Rabi frequencies are
\begin{align}
\Omega_- &\equiv \frac{2\mu_B}{\hbar} \bra{2}J_-\ket{1} B_- e^{-i\phi_-}, \label{eq:omega_-} \\
\Omega_{\pi} &\equiv \frac{2\mu_B}{\hbar} \bra{2}J_z\ket{1} B_{\pi} e^{-i\phi_{\pi}}, \label{eq:omega_pi} \\
\Omega_+ &\equiv \frac{2\mu_B}{\hbar} \bra{2}J_{+}\ket{1} B_{+} e^{-i\phi_{+}}, \label{eq:omega_+}
\end{align}
Where $J_z$, $J_+= J_x + iJ_y$, and $J_-= J_x - iJ_y$ are the spin $z$, raising, and lowering operators, respectively.
Note that in the definitions of $B_-$ and $B_+$, a factor of $1/\sqrt{2}$ instead of $1/2$ can also be found in some of the literature. This goes along with a change in the definitions of $J_+$ and $J_-$, which then read $J_{\pm}= \tfrac{1}{\sqrt{2}}(J_x \pm iJ_y)$. If this alternative definition is used, the coefficients $\alpha_+$ and $\alpha_-$ are larger by a factor of $\sqrt{2}$.
\subsection{Microwave Amplitude}
\label{sec:FreqTun_amplitude}
From Eq.~(\ref{eq:omega_pi}), it is straightforward to determine the amplitudes of the microwave magnetic field ($B_{x,y,z}$) when strong $\pi$ transitions are present. The matrix element $\bra{f}J_z\ket{i}$ can be calculated numerically for a any static magnetic field, and so we can obtain the amplitudes along each axis by measuring $\abs{\Omega_\pi}$ with the quantisation axis along $x$, $y$ and $z$ respectively. The $\pi$ (and $\sigma_{-}$) transitions become weak in strong dc fields, however, and in the general case, we need to determine the microwave field amplitudes using only $\sigma_{+}$ transitions.
In the following discussion, the superscript index represents the quantisation axis in the lab frame, i.e. the direction of the applied static magnetic field. Thus for example, $\Omega_+^{+ y}$ ($B_+^{+ y}$) means $\Omega_+$ ($B_+$) for \textbf{B}$_{\mathrm{dc}}$ pointing along the $y$ axis in the positive direction, whilst $\Omega_+^{- y}$ ($B_+^{- y}$) is for \textbf{B}$_{\mathrm{dc}}$ pointing along the $y$ axis in the negative direction.
We begin by finding the sum of $B_-^2$ and $B_+^2$. From Eq.~(\ref{eq:omega_+}) we see that for $\sigma_+$ transitions, we only obtain $B_+$. However, for measurements along a given axis, $B_+$ measured antiparallel to that axis is equivalent to $B_-$ measured parallel to the axis. That is, $B_+^{-} = B_-^{+}$. By measuring $\abs{\Omega_+}$ with the static field both parallel and antiparallel to our axis of measurement, we can thus obtain both $B_+$ and $B_-$. This gives us
\begin{equation}
\label{eq:b-+sum}
B_-^2 + B_+^2 = \frac{\hbar^2}{4\mu_B^2} \Big[ \frac{\abs{\Omega_+^{-}}^2}{\abs{\bra{f}J_+\ket{i}}^2} + \frac{\abs{\Omega_+^{+}}^2}{\abs{\bra{f}J_+\ket{i}}^2} \Big].
\end{equation}
We can also find the $B_-^2 + B_+^2$ sum using Eqs.~(\ref{eq:B_-}) and (\ref{eq:B_+}). Equating this with Eq.~(\ref{eq:b-+sum}) gives
\begin{equation}
\frac{1}{2} (B_{x'}^2 + B_{y'}^2) = \frac{\hbar^2}{4\mu_B^2} \Big[ \frac{\abs{\Omega_+^{-}}^2}{\abs{\bra{f}J_+\ket{i}}^2} + \frac{\abs{\Omega_+^{+}}^2}{\abs{\bra{f}J_+\ket{i}}^2} \Big].
\end{equation}
Defining
\begin{equation}
K_+ \equiv \frac{\abs{\Omega_+^{-}}^2}{\abs{\bra{f}J_+\ket{i}}^2} + \frac{\abs{\Omega_+^{+}}^2}{\abs{\bra{f}J_+\ket{i}}^2},
\end{equation}
we can write
\begin{equation}
\label{eq:B_sum}
B_{x'}^2 + B_{y'}^2 = \frac{\hbar^2}{2\mu_B^2} K_+.
\end{equation}
Next, we apply this formula with the quantisation axis defined along each of the $x$, $y$ and $z$ axes. Starting with \textbf{B}$_{\mathrm{dc}}$ (and thus $z'$) along the $x$ axis, we transform from the primed coordinate system back into the unprimed lab frame, according to the following coordinate transformation:
\begin{eqnarray}
x' &=& -z, \nonumber\\
y' &=& y, \nonumber\\
z' &=& x.
\end{eqnarray}
This transforms the microwave magnetic field phasor to
\[
\textbf{B} \equiv
\begin{pmatrix}
B_{x'} e^{-i\phi_{x'}}\\
B_{y'} e^{-i\phi_{y'}} \\
B_{z'} e^{-i\phi_{z'}} \\
\end{pmatrix} =
\begin{pmatrix}
-B_{z} e^{-i\phi_{z}}\\
B_{y} e^{-i\phi_{y}} \\
B_{x} e^{-i\phi_{x}} \\
\end{pmatrix},
\]
with
\begin{alignat}{2}
B_{x'} &= B_z &\quad \phi_{x'} &= \phi_z + \pi, \nonumber\\
B_{y'} &= B_y &\quad \phi_{y'} &= \phi_y, \nonumber\\
B_{z'} &= B_x &\quad \phi_{z'} &= \phi_x.
\end{alignat}
Applying this coordinate transformation to Eq.~(\ref{eq:B_sum}) then gives us
\begin{equation}
\label{eq:B_zy}
B_z^2+B_y^2 = \frac{\hbar^2}{4\mu_B^2} K_+^x,
\end{equation}
with $K_+^{x}$ defined in Eq.~(\ref{eq:K+}). We can follow a similar process for \textbf{B}$_{\mathrm{dc}}$ along the $y$ and $z$ axes to get
\begin{equation}
\label{eq:B_zx}
B_z^2+B_x^2 = \frac{\hbar^2}{4\mu_B^2} K_+^y,
\end{equation}
\begin{equation}
\label{eq:B_zy}
B_x^2+B_y^2 = \frac{\hbar^2}{4\mu_B^2} K_+^z.
\end{equation}
Solving these equations simultaneously gives the magnitude of the magnetic field along the (lab frame) $x$, $y$, and $z$ directions,
\begin{align}
B_x^2 &= \frac{\hbar^2}{8\mu_B^2} \Big[ -K_+^x + K_+^y + K_+^z\Big],\label{eq:B_x}\\
B_y^2 &= \frac{\hbar^2}{8\mu_B^2} \Big[ K_+^x - K_+^y + K_+^z\Big],\label{eq:B_y}\\
B_z^2 &= \frac{\hbar^2}{8\mu_B^2} \Big[ K_+^x + K_+^y - K_+^z\Big],\label{eq:B_z}
\end{align}
where $K_+^{\gamma}$ is defined as
\begin{equation}
\label{eq:K+}
K_+^{\gamma} \equiv \frac{\abs{\Omega_+^{-\gamma}}^2}{\abs{\bra{f}J_+\ket{i}}^2} + \frac{\abs{\Omega_+^{+\gamma}}^2}{\abs{\bra{f}J_+\ket{i}}^2}.
\end{equation}
$\abs{\Omega_+^{-\gamma}}$ and $\abs{\Omega_+^{+\gamma}}$ are experimentally determined quantities. The matrix element $\bra{f}J_+\ket{i}$ can be calculated numerically for a general applied static magnetic field, \textbf{B}$_{\mathrm{dc}}$.
\subsection{Microwave Phase}\label{sec:FreqTun_phase}
To reconstruct the field phases, $\phi_x$, $\phi_y$ and $\phi_z$, we begin with the difference of $B_-^2$ and $B_+^2$. Again, this can be found using Eq.~(\ref{eq:omega_+}), measuring $\abs{\Omega_+}$ with the static field both parallel and antiparallel to our axis of measurement, and also using Eqs.~(\ref{eq:B_-}) and (\ref{eq:B_+}):
\begin{align}
B_-^2 - B_+^2 &= \frac{\hbar^2}{4\mu_B^2} \Big[ \frac{\abs{\Omega_+^{-}}^2}{\abs{\bra{f}J_+\ket{i}}^2} - \frac{\abs{\Omega_+^{+}}^2}{\abs{\bra{f}J_+\ket{i}}^2} \Big] \nonumber\\
&= B_{x'} B_{y'} \sin(\phi_{y'}-\phi_{x'}).
\end{align}
This time we define
\begin{equation}
\label{eq:K-}
K_- \equiv \frac{\abs{\Omega_+^-}^2}{\abs{\bra{f}J_+\ket{i}}^2} - \frac{\abs{\Omega_+^+}^2}{\abs{\bra{f}J_+\ket{i}}^2},
\end{equation}
and so we have
\begin{equation}
\label{eq:B_phase}
\sin(\phi_{y'}-\phi_{x'}) = \frac{\hbar^2}{4\mu_B^2 B_{x'} B_{y'}} K_-.
\end{equation}
We measure $\abs{\Omega_+}^2$ parallel and antiparallel to the $x$, $y$, and $z$ axes, and use Eq.~(\ref{eq:B_phase}) with the same coordinate transformations as in Section~\ref{sec:FreqTun_amplitude}. Inserting the field magnitudes obtained with Eqs.~(\ref{eq:B_x}-\ref{eq:B_z}), we get
\begin{align}
\sin(\phi_{z}-\phi_{y}) &= \label{eq:phi_zy}\\
2 \Big[ (K_+^x - &K_+^y + K_+^z) (K_+^x + K_+^y - K_+^z) \Big]^{-1/2} K_-^x, \nonumber\\
\sin(\phi_{x}-\phi_{z}) &=\label{eq:phi_xz}\\
2 \Big[ ( -K_+^x + &K_+^y + K_+^z) (K_+^x + K_+^y - K_+^z) \Big]^{-1/2} K_-^y,\nonumber\\
\sin(\phi_{y}-\phi_{x}) &= \label{eq:phi_yx}\\
2 \Big[ ( -K_+^x + &K_+^y + K_+^z) (K_+^x - K_+^y + K_+^z) \Big]^{-1/2} K_-^z,\nonumber
\end{align}
with $K_+^{\gamma}$ as defined in equation~\ref{eq:K+} and $K_-^{\gamma}$ defined as $K_-$ for a static magnetic field along the direction $\gamma$:
\begin{equation}
\label{eq:K-_gamma}
K_-^{\gamma} \equiv \frac{\abs{\Omega_+^{-\gamma}}^2}{\abs{\bra{f}J_+\ket{i}}^2} - \frac{\abs{\Omega_+^{+\gamma}}^2}{\abs{\bra{f}J_+\ket{i}}^2}.
\end{equation}
$\abs{\Omega_+^{-\gamma}}$ and $\abs{\Omega_+^{+\gamma}}$ are experimentally determined quantities. The matrix elements $\bra{f}J_-\ket{i}$ and $\bra{f}J_+\ket{i}$ can be calculated numerically for a general applied static magnetic field, \textbf{B}$_{\mathrm{dc}}$. From Eqs.~(\ref{eq:phi_zy}-\ref{eq:phi_yx}), we can obtain the relative phases of the B$_{\mathrm{mw}}$ components.
|
\section{Introduction.}
The environmental impact of fossil fuels and the high variability in their prices have led to rising adoption of Electric Vehicles (EVs), which is supported by ambitious government policies for promoting EVs~\cite{EVUS,EVEU}. Despite the technological advances in vehicle efficiency and battery capacity, a key hurdle in EV adoption is that, barring a few pockets of densely populated areas, the distribution of EV charging stations is sparse in most regions.\footnote{Recent government mandates~\cite{EVEU-Coverage,EVUS-DRIVE} focus specifically on the expansion of refueling infrastructure to cover entire geographical regions.} Because of this, EV owners and potential buyers frequently worry about whether the vehicle will have sufficient charge to travel to their trip destinations or an intermediate charging station. On the other hand, given the high cost of building a charging station and currently low\footnote{Nevertheless, this number is rapidly growing, e.g., in the U. S., EV market share has risen~\cite{EVUS-Growth} from 0.14\% in 2011 to 0.72\% in 2014.} number of EVs, charging station operators would only want to place stations where there is sufficient demand for charging. This results in a situation where the charging station density is concentrated near the city centre and rapidly decrease as we get farther, e.g., Figure~\ref{fig:neuk-map} illustrates this for North East England.
\begin{figure}[t]
\centering
\includegraphics[width=0.33\textwidth]{motivating-map-intro}
\caption{Charging stations in North East England.\label{fig:neuk-map}}
\end{figure}
This work is an attempt to break this deadlock by improving the distribution of new charging stations so as to mitigate the range anxiety of EV users. To be effective, such a solution must simultaneously address the concerns of both EV users and charging station operators. Charging station operators are concerned about (a)~\textit{demand}: charging stations must be distributed to serve the maximum total demand for charging, and (b)~\textit{budget}: there is a limited budget available for setting up charging stations. EV users are concerned about (c)~\textit{reachability}: there must be a charging station that is reachable within a short driving distance from most locations, and (d)~\textit{waiting time}: the waiting time to begin charging at charging stations should not be prohibitively large, given that the charging event itself is time-consuming.\footnote{There may be several other concerns for charging station operators and EV users, including alternate interpretations of \textit{reachability}, e.g.,~\cite{Funke15}, but in this work, we restrict our attention to modeling just these four.}
This paper models the above placement problem within a framework that is inspired from the facility location literature \cite{Drezner95}. Given a set of candidate sites for charging stations, e.g., parking lots in a city, the objective is to select an optimal subset of these locations that maximizes the total demand satisfied, subject to the total cost not exceeding the available budget (``packing'' constraint), and given a set of locations of interest, there is at least one charging station within a specified driving range of every location in this set (``covering'' constraint). This ``mixed-packing-and-covering'' formulation (Section~\ref{ssec:prelims-placement-optimization}) covers concerns (a)-(c). In order to take care of (d), a queueing model is used to translate the constraint on waiting time and the charging demand at a location to a constraint on the minimum number of charging slots needed at a charging station at that location, which is then factored into the setup costs in a pre-processing step.
Before tackling the optimization problem itself, it is necessary to predict the demand for charging at candidate sites. Although demand prediction has been extensively studied for planning transportation infrastructure, the EV charging station demand prediction problem is difficult because of sparse deployment of such stations, and lack of sufficient and reliable historic data. Thus, in addition to the limited historical demand data for existing charging stations in other locations, two other types of features of the location which may have a strong impact on the usage are considered. First, nearby Points of Interest (PoI), e.g., shopping malls, institutions, restaurants, hospitals etc., indicate the frequency of visits that could be made by travelers to the given location, and hence, are useful for estimating the demand. Second, traffic density data at nearby road junctions are also useful as features in demand prediction. Using these features, in Section~\ref{sec:demand-prediction} proposes a supervised multi-view learning framework using Canonical Correlation Analysis (CCA) for EV charging station demand prediction.
Once the demands are estimated using the demand prediction algorithm, the next step is to solve the mixed packing and covering problem. We identify a family of heuristics in Section~\ref{mixedpackandcover} that seeks to iteratively find the optimal allocation of the available budget between satisfying the packing and the covering constraints, by alternatingly invoking algorithms that solve reduced knapsack and set-cover subproblems. For example, choosing the well known greedy algorithms for the knapsack and set-cover problems~\cite{Vazirani01} yields one instance in this family.\footnote{Since the primary mixed packing and covering problem, and the associated knapsack and set-cover problems are NP-Hard, it is unlikely that any polynomial-time solution, including this instance, would be optimal.} In Section~\ref{ssec:expt-optimal-placement}, we present results from an experimental evaluation of this instance for the EV charging station placement problem using charging data from the UK. In most cases, and especially when budget is scarce, our heuristic achieves an improvement of 10-20\% over a naive heuristic, both in terms of finding feasible solutions and maximizing demand.
The modeling and heuristic frameworks above can be easily extended to the following alternate scenarios: (a)~incremental placement of charging stations when the budget is progressively released over a period of time, and (b)~when there are multiple charging station operators, each with their respective budgets, and a government agency (subject to its own budget) seeks to optimally allocate grants to incentivize these providers to set up charging stations at selected locations (e.g., where the demand is low).
\subsection{Related work.}
With the recent increase in the adoption of EVs, the charging station placement problem has received significant attention. One of the prerequisites of an effective charging station placement is the availability of estimated charging demand at candidate sites. In this context, prior literature has studied the impact of external information on charging station demand. For example, parking demand is combined with facility location problem in~\cite{Chen13}, and the effect of demographic features is studied using regression models in~\cite{Wagner14}. However, these approaches lack a principled multi-view learning framework to integrate heterogeneous data sets to predict EV charging demand. In the machine learning literature, CCA-based models have been used for multi-variate regression~\cite{Piyush09,Sun11,Klami13}; in this paper, we propose a novel variant of CCA-based multivariate regression.
A useful charging station placement solution needs to take into account the charging demand, budget, coverage, and waiting time at the stations. Existing work has not considered all of these factors simultaneously. For example, placement of charging stations based on predicted demand has been studied in~\cite{Dai13,Klabjan12,Wagner14,Chen13,Lam13,chung2015multi}, but they do not simultaneously consider constraints on budget, coverage, and waiting times. Coverage requirements in charging station placement has been modeled as path cover in~\cite{Funke15}, and reachability cover in~\cite{Funke13} over the city road network, but they do not consider the demand satisfied by the deployment and the waiting time faced by the users.
The optimization problem considered in this work contains both packing and covering constraints, with integer decision variables. Mixed packing and covering problems have been studied for fractional decision variables~\cite{Young01,Azar13}, as well as for integer decision variables~\cite{Chakaravarthy13,Mukherjee15}. However, this work introduces an optimization problem with a knapsack packing constraint and a set covering constraint, which has not been studied earlier.
\section{Demand prediction at candidate sites.}\label{sec:demand-prediction}
Understanding the EV charging demand is key to optimal placement of charging stations. Since the distribution of charging stations is typically sparse, there is insufficient data to efficiently model the demand. Thus, it becomes important to model the relationship of EV charging demand with auxiliary data such as Point of Interest (PoI) and traffic density. Such relationships help in better understanding not only the charging demand, but also the factors behind it. Additionally, the expected demand at a candidate site where no historical demand is known, can be predicted based on external factors. A multi-view learning framework based on Canonical Correlation Analysis (CCA) is used to jointly model the charging demand data and other external data sources, and to predict the demand at new candidate sites given the external factors. Next, we briefly describe CCA and then present the CCA-based regression framework.
\subsubsection{Canonical correlation analysis.}
Canonical correlation analysis is a classical method to find a linear relationship between two sets of random variables. Let $X$ and $Y$ be data matrices of size $n \times d_x$ and $n \times d_y$ respectively, where each row corresponds to a realization of the random variables. CCA finds linear projections $u_x \in \mathbb{R}^{d_x}$ and $u_y \in \mathbb{R}^{d_y}$ such that their correlation, $corr(Xu_x,Yu_y)$, is maximized. It can be shown that maximizing correlation is equivalent to minimizing $\| Xu_x - Yu_y\|^2$ subject to $\| Xu_x\|^2 =1$ and $\| Yu_y\|^2 = 1$. A set of $k \leq min(rank(X),rank(Y))$ linear projections can be computed by solving a generalized eigenvalue problem;~\cite{Hardoon04} provides a detailed review of classical CCA.
\subsection{Multi-view regression using CCA.}
In this section, a multi-view learning based regression framework for the charging demand prediction at candidate sites is presented. The regression framework uses CCA in a supervised setting as a multivariate regression tool. CCA is, in fact, closely related to multivariate linear regression;~\cite{Borga01,Sun11} present least-square formulations. CCA-based approaches model the statistical dependence between two or more data sets and assume a single latent factor to describe information shared between all datasets. In a supervised setting such as multivariate regression task, the output variable is considered as one of the random variables and CCA models the statistical dependence between the output variable (dependent) and a covariate (independent variable). However, when there are more than two datasets (output variable and multiple covariates), modeling shared information may not always be a good choice, since it would discard dataset-specific information which may be crucial to predicting the output variable.
To address this deficiency, a multi-view learning regression framework is proposed when there are $n>2$ covariates, all of which can be used to predict the output variable. The idea is to model the statistical dependence of the output variable with each of the covariates separately and then learn a weighted combination for each such model that maximizes the prediction performance. We call this model Multiple Dependent Regression (MDR), which can be formulated as the ensemble regression function $\mathbf{f} = \sum_i w_i \mathbf{f_i}$, where $\sum_i w_i=1$. Here, $\mathbf{f_i}$ is a regression function learned from covariate $X_i$ to the output variable $Y$, and $w_i$ are the weights denoting the importance of covariate $X_i$ in the regression task. The regression function $\mathbf{f_i}$ for covariate $X_i$ can be learnt according to a CCA-based regression, given by $\argmin_{U_x^i,U_y^i} \|X_iU_x^i - YU_y^i\|^2$, where $U_x^i$ and $U_y^i$ are $k$-dimensional linear projections for each (covariate, output variable) pair $(X_i,Y)$. Hence, MDR is a two step framework of multiple regression:
\begin{enumerate}[(1)]
\item For each covariate $X_i$, apply CCA for $X_i$ and $Y$, and compute the prediction error $e_i$.
\item For each $i$, compute $w_i = \frac{1}{n-1}(1 - \frac{e_i}{\sum_j e_j})$.
\end{enumerate}
The values of parameters $U_x, U_y, w_i$ are learnt from the training data as described in the steps above. The model thus trained can be used to predict the output variable given the input covariates. In our case, the task is to predict the demand of EV charging at a new candidate location given the covariates as explained earlier.
MDR is analogous to ensemble learning approaches where multiple learning models are trained to solve a given problem. In our case, we train multiple CCA-based regression models in the first step and learn a weighted combination of models based on training error in the second step. The weights $w_i$ explain the importance of each covariate in the prediction task. This procedure is flexible--any CCA-based solution can be used in Step~(1). In this work, we use a Bayesian solution with group-wise sparsity proposed in~\cite{Klami13}.
By learning a separate CCA model for each covariate, MDR ensures that dataset-specific information of any covariate (with respect to other covariates) is not ignored if it is shared with the output variable. MDR provides an intuitive, flexible, and yet, a simple approach to multi-view regression. It can be easily extended to non-linear projections using non-linear solutions to CCA.
\subsection{Experimental results for demand prediction.}\label{ssec:expt-demand-prediction}
MDR is evaluated on EV charging data obtained from $252$ public charging points in North East England through UK's Plugged-In-Places program~\cite{PiP}. The location of charging points was obtained from UK's National Charge Point Registry data~\cite{NCPR}. PoI information is extracted from OpenStreetMap~\cite{Haklay08} API for $11$ categories: sustenance, education, transportation, financial, healthcare, entertainment, sports, gardens, places of worship, shops and public buildings. Finally, we use traffic data are for each junction-to-junction link on major road networks, provided by \cite{TrafficData}.
The following data matrices are created: (i)~$Y \in \mathbb{R}^{252 \times 24}$ represents hour-wise charging demand for each of $252$ charging points, where $Y[i,j]$ is the average energy consumed at charging point $i$ in hour $j$; (ii)~$X_1 \in \mathbb{R}^{252 \times 11}$ represents the PoIs, where $X_1[i,j]$ is the frequency of PoI category $j$ within a radius of $500$ meters around charging point $i$; (iii)~$X_2 \in \mathbb{R}^{252 \times 5}$ represents the traffic densities at the $5$ nearest traffic junctions to each charging point; and (iv)~$X_3 \in \mathbb{R}^{252 \times 5}$ represents the charging demands at the $5$ nearest charging points to each charging point. Since charging data is sparse during early mornings and late nights, only data from 07:00 hrs. till 23:00 hrs. is used.
MDR is evaluated in a leave-one-out cross-validation manner by training on all but one instance (charging point) and testing on the left-out instance. Prediction performance is measured as the average Root Mean Square Error (RMSE) over all test instances. MDR is compared against (i)~Linear Regression(LR) - all covariates are concatenated feature-wise similar to~\cite{Wagner14}, (ii)~Multiple Linear Regression (MLR) - weighted sum of multiple LR analogous to MDR, and (iii)~Bayesian CCA (bCCA) - group factor analysis~\cite{Virtanen12}. Figure~\ref{fig:MDR} shows the average error for each hour of the day over all charging stations. It can be observed that MDR clearly outperforms the other three baselines giving an overall reduction in error by 27\% compared to simple LR, 21\% compared to multiple LR and 18\% compared to bCCA.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{prediction}
\caption{Comparison of MDR against baselines.\label{fig:MDR}}
\end{figure}
\section{Optimal selection from candidate sites for placing charging stations.}\label{sec:placement-optimization}
In this section, we explain the mathematical model for the charging station placement problem, and detail the methods to solve the resulting mixed packing and covering optimization problem.
\subsection{Preliminaries.}\label{ssec:prelims-placement-optimization}
Let $\mathcal{L}=\{1,\ldots,|\mathcal{L}|\}$ denote the set of all candidate sites for placing a charging station. Let $r\in\mathbb{R}_+$ denote the desired ``reachability radius,'' that is, the maximum distance to be travelled in order to reach a charging station. Let $\mathcal{I}$ denote the set of all locations of interest that are desired to be ``covered,'' that is, lie within the reachability radius from at least one charging station. Let $B$ denote the total budget available.
For each candidate site $i\in\mathcal{L}$: $x_i\in\{0,1\}$, $d_i\in\mathbb{R}_+$, and $c_i\in\mathbb{R}_+$ denote, respectively, whether a charging station is placed, the demand for service, and the cost of setting up a charging station at $i$. $S_i^r\subseteq\mathcal{I}$ is the \textit{cover set} of $i$, that is, the set of locations of interest that would be ``covered'' (within a driving distance of $r$) if a charging station is placed at $i$.
\subsubsection{Optimizing for demand.}
For a given reachability radius $r$, set of candidate sites $\mathcal{L}$, and set of locations of interest $\mathcal{I}$, the cover sets $\left\{S_i^r\right\}_{i\in\mathcal{L}}$ can be precomputed. Given predicted demand and charging duration data, a queueing model~\cite{Kleinrock75} can estimate the number of charging slots necessary at each candidate site to meet a specified constraint on waiting times, which is then used to precompute the costs $\left\{c_i\right\}_{i\in\mathcal{L}}$. (See Appendix \ref{appendix-cost-queueing} for details.) Then, the optimization problem to be solved is the Mixed Pack \& Cover (MPC) problem shown in Figure \ref{fig:eqn-MPC}.\footnote{The total demand satisfied by installing charging stations at a subset of candidate sites may be less than the sum of the predicted demands at those sites due to overlapping reachability regions. We ignore this effect in order to keep the optimization problem simple.}
\subsubsection{Approximation algorithm for a transformed problem.}
MPC (Figure~\ref{fig:eqn-MPC}) is NP-Hard as it contains as a special case, the set-cover problem~\cite{Vazirani01}. For a closely related pure covering problem called Demand \& Set-Cover (DSC, Figure~\ref{fig:eqn-DSC}), an approximation algorithm is described next.
DSC has as subproblems, two well-known NP-Hard problems, Set-Cover (SC)~\cite{Vazirani01} and Minimization Knapsack (MinKP)~\cite{Csirik91}, which are given in Figures~\ref{fig:eqn-SC} and~\ref{fig:eqn-MinKP}, respectively. Consider, an $f_{sc}$ factor approximation algorithm $A_{sc}$ for set cover and an $f_{kp}$ factor approximation algorithm $A_{kp}$ for minimization knapsack. Let $\mathcal{L}_{sc}$ and $\mathcal{L}_{kp}$ (subsets of $\mathcal{L}$) be the solutions returned by $A_{sc}$ and $A_{kp}$, respectively. Consider algorithm $A_{dsc}$ for DSC that uses $A_{sc}$ and $A_{kp}$ as subroutines, and returns the union of the solutions returned by the two subroutines, that is, $\mathcal{L}_{sc} \cup \mathcal{L}_{kp}$, as a solution for DSC. The following lemma (proved in Appendix \ref{appendix-approx-algo}) establishes that $A_{dsc}$ is an $(f_{sc}+f_{kp})$ factor approximation algorithm for DSC.
\begin{lemma}\label{lemma:appendix-approx-algo}
$A_{dsc}$ returns a feasible solution for DSC whose cost is at most $(f_{sc} + f_{kp})$ times the optimal cost.
\end{lemma}
The heuristic framework for MPC presented next is inspired by $A_{dsc}$ and the observation that, similar to DSC, a solution for MPC can be obtained by combining set cover and knapsack algorithms.
\begin{figure*}[!t]
\centering
\subfigure[\scriptsize{Mixed Pack \& Cover (MPC)}]{\label{fig:eqn-MPC}\includegraphics[width=0.32\textwidth]{equation-mpc.png}}
\hfill
\subfigure[\scriptsize{Knapsack (KP)}]{\label{fig:eqn-KP}\includegraphics[width=0.32\textwidth]{equation-knapsack.png}}
\hfill
\subfigure[\scriptsize{Set-Cover (SC)}]{\label{fig:eqn-SC}\includegraphics[width=0.32\textwidth]{equation-set-cover.png}}
\\
\subfigure[\scriptsize{Min-Knapsack (MinKP)}]{\label{fig:eqn-MinKP}\includegraphics[width=0.32\textwidth]{equation-min-knapsack.png}}
\hspace{0.1in}
\subfigure[\scriptsize{Demand \& Set-Cover (DSC)}]{\label{fig:eqn-DSC}\includegraphics[width=0.32\textwidth]{equation-dsc.png}}
\caption{Important optimization problems.\label{fig:eqn-opt-probs}}
\end{figure*}
\subsection{Solving the mixed-pack-and-cover problem.}\label{mixedpackandcover}
Our goal is to introduce a general methodology for algorithms to solve MPC (Figure \ref{fig:eqn-MPC}), which is NP-Hard, by breaking it down into the following two subproblems:
\begin{enumerate}
\item \textit{The Knapsack Problem (KP):} This is the packing problem in Figure \ref{fig:eqn-KP}. Let $\mbox{\texttt{KP}}\left(\mathcal{L},\{d_i\}_{i\in\mathcal{L}},\{c_i\}_{i\in\mathcal{L}},B\right)$ denote any algorithm that solves this problem.
\item \textit{The Set-Cover Problem (SC):} This is the covering problem in Figure \ref{fig:eqn-SC}. Let $\mbox{\texttt{SC}}\left(\mathcal{L},\mathcal{I},\{c_i\}_{i\in\mathcal{L}},\{S_i\}_{i\in\mathcal{L}}\right)$ denote any algorithm that solves this problem.
\end{enumerate}
The algorithm \texttt{KP()} (respectively, \texttt{SC()}) need not be optimal, but we require that it be nontrivial in the sense that its solution must be \textit{maximal} (respectively, \textit{minimal}). That is, it should not be possible to add (remove) an item and still satisfy the packing (covering) constraint(s).
\subsubsection{The iterative-pack-and-cover (IPAC) framework.}\label{iter-pack-cover}
The available budget could be used very differently if the problem had been a pure packing problem (where maximizing the demand is the only concern) or a pure covering problem (where satisfying the covering constraints is the only concern). Thus, a good solution to the mixed packing and covering problem should achieve an appropriate balance by dividing the available budget between these two concerns. The core idea behind the Iterative Pack And Cover (IPAC) framework is to iteratively search for such an optimal split.
In each iteration, the budget $B$ can be thought of as comprising of three parts: $B^{p}$, $B^{c}$, and some excess (due to the integrality constraint). Here, $B^{c}$ is the portion of the budget used by a solution to a covering problem, and $B^{p}$ is the portion of the budget used by a solution to a packing problem when constrained by a reduced budget of $B-B^{c}$. A check is performed using \texttt{SC()} to determine whether the remaining budget $B-B^{p}$ is sufficient to satisfy the covering constraints left unsatisfied by the solution to the packing problem. Starting with $B^{c}=0$ (pure packing) and the corresponding solution from \texttt{KP()}, during each iteration, $B^{c}$ is increased until the covering check passes, at which point the solutions of the packing and covering problems obtained in the last iteration are merged.\footnote{As an extension, note that $B^{c}$ need not necessarily increase in each iteration, as long as it can be shown that the procedure will terminate in finite time.} The resulting solution is guaranteed to be a feasible solution to the MPC problem (Figure \ref{fig:eqn-MPC}). Optionally, \texttt{KP()} can be invoked one final time to use up any remaining portion of the budget. In the worst case, the iterations continue until $B^{c}=B$, at which point it becomes a pure covering problem, and if the covering check still fails, then IPAC fails to find a feasible solution to MPC. (But if this happens, it cannot necessarily be concluded that MPC is infeasible, unless \texttt{SC()} is optimal, which is unlikely in polynomial time since the set-cover problem is NP-Hard.)
The above framework is quite generic and encompasses a large class of algorithms. Next, we describe a particular instantiation of the framework which runs in polynomial time. The corresponding pseudocode is detailed in Algorithm \ref{alg:iter-pack-cover}.
\begin{enumerate}
\item Set $B^{c}=0$ and invoke \texttt{KP()} to solve the pure packing problem. Let $B^{p}$ be the portion of the budget used by the chosen items. $B^{free}=B-B^{p}$ is the remainder.
\item Update $B^{c}$ to be the minimum budget required to satisfy the unsatisfied covering constraints, computed by invoking \texttt{SC()} to solve the residual covering problem.
\item Repeat the following two steps until either $B^{c}\leq B^{free}$ or $B^{c} > B$:
\begin{enumerate}
\item \textit{Packing:} In response to the reduced budget $B-B^{c}$, remove one or more chosen items. In order to do so, use a ranking method, \texttt{RANK()}, that ranks the currently chosen items according to some measure of their importance. Keep removing items that are least important until the remaining items satisfy the reduced budget. Let $B^{p}$ be the portion of the budget used. $B^{free}=B-B^{p}$ is the new remainder.
\item \textit{Covering:} Since the removal of some items in the previous step might have resulted in more unsatisfied covering constraints, invoke \texttt{SC()} on the new residual covering problem and update $B^{c}$ accordingly.
\end{enumerate}
\item If $B^{c} > B$, then a feasible solution cannot be found. Otherwise, $B^{c}\leq B^{free}$, and so, add the items from the solution to the last instance of \texttt{SC()} to the remaining chosen items to obtain a feasible solution.
\item Invoke \texttt{KP()} to fill any unused portion of the budget using unallocated items.
\end{enumerate}
\begin{algorithm}[!ht]
\caption{\texttt{IPAC}\label{alg:iter-pack-cover}}
\begin{algorithmic}[1]
\scriptsize
\State \textbf{Input:} $\mathcal{L}$, $\mathcal{I}$, $\{d_i\}_{i\in\mathcal{L}}$, $\{c_i\}_{i\in\mathcal{L}}$, $B$, $\{S_i^r\}_{i\in\mathcal{L}}, \mathtt{KP}(), \mathtt{SC}(), \mathtt{RANK}()$
\State \textbf{Output:} $\{x_i\}_{i\in\mathcal{L}}$
\item[]
\State $\{x_i\}_{i\in\mathcal{L}} \leftarrow \mathtt{KP}(\mathcal{L}, \{d_i\}_{i\in\mathcal{L}}, \{c_i\}_{i\in\mathcal{L}},B)$;
\State $\mathcal{L}^{p} \leftarrow \{i\in\mathcal{L}\ :\ x_i = 1\}$; $B^{p} \leftarrow \sum_{i\in\mathcal{L}}c_ix_i$;
\State $\mathcal{I}^{p} \leftarrow \bigcup_{i\in\mathcal{L}^{p}}S_i^r$;
\LineComment{$\ \mathcal{I}^{p}$ $\leftarrow$ elements of $\mathcal{I}$ covered by $\mathcal{L}^{p}$.}
\State $\{x^{c}_i\}_{i\in\mathcal{L}\backslash\mathcal{L}^{p}} \leftarrow \mathtt{SC}(\mathcal{L}\backslash\mathcal{L}^{p}$, $\mathcal{I}\backslash\mathcal{I}^{p}$, $\{c_i\}_{i\in\mathcal{L}\backslash\mathcal{L}^{p}}$, $\{S_i^r\}_{i\in\mathcal{L}\backslash\mathcal{L}^{p}})$;
\State $B^{c} \leftarrow \sum_{i\in\mathcal{L}\backslash\mathcal{L}^{p}}c_ix^{c}_i$; $\ \ B^{free} \leftarrow B - B^{p}$;
\While{($B^{c} > B^{free}$ and $B^{c} \leq B$)}
\LineComment{$\ \rho$ is $\mathcal{L}^{p}$ in increasing order of importance.}
\State $\rho \leftarrow \mathtt{RANK}(\mathcal{L}^{p}$, $\{d_i\}_{i\in\mathcal{L}^{p}}$, $\{c_i\}_{i\in\mathcal{L}^{p}}$, $\{S_i^r\}_{i\in\mathcal{L}^{p}})$;
\LineComment{ Remove least important items.}
\State $j \leftarrow 0$;
\While{$B^{free} < B^{c}$}
\State $j \leftarrow j + 1$; $x_{\rho[j]} \leftarrow 0$; $B^{free} \leftarrow B^{free} + c_{\rho[j]}$;
\EndWhile
\LineComment{ Recompute quantities.}
\State $\mathcal{L}^{p} \leftarrow \{i\in\mathcal{L}\ :\ x_i = 1\}$; $\mathcal{I}^{p} \leftarrow \bigcup_{i\in\mathcal{L}^{p}}S_i^r$;
\State $\{x^{c}_i\}_{i\in\mathcal{L}\backslash\mathcal{L}^{p}} \leftarrow \mathtt{SC}(\mathcal{L}\backslash\mathcal{L}^{p}$, $\mathcal{I}\backslash\mathcal{I}^{p}$, $\{c_i\}_{i\in\mathcal{L}\backslash\mathcal{L}^{p}}$, $\{S_i^r\}_{i\in\mathcal{L}\backslash\mathcal{L}^{p}})$;
\State $B^{c} \leftarrow \sum_{i\in\mathcal{L}\backslash\mathcal{L}^{p}}c_ix^{c}_i$;
\EndWhile
\If{$B^{c} > B$} EXIT; \EndIf
\LineComment{ Add required items to knapsack.}
\For{$i$ in $\mathcal{L}\backslash\mathcal{L}^{p}$}
$x_i \leftarrow x_i + x^{c}_i$;
\EndFor
\LineComment{ Add more items if possible.}
\State $\mathcal{L}^{p} \leftarrow \{i\in\mathcal{L}\ :\ x_i = 1\}$; $B^{p} \leftarrow \sum_{i\in\mathcal{L}}c_ix_i$;
\State $\{x_i\}_{i\in\mathcal{L}\backslash\mathcal{L}^{p}} \leftarrow \mathtt{KP}(\mathcal{L}\backslash\mathcal{L}^{p}$, $\{d_i\}_{i\in\mathcal{L}\backslash\mathcal{L}^{p}}$, $\{c_i\}_{i\in\mathcal{L}\backslash\mathcal{L}^{p}}$, $B-B^{p})$;
\end{algorithmic}
\end{algorithm}
\subsubsection{The \texttt{RANK()} function.}\label{rank}
The effectiveness of IPAC depends on the choice of methods for \texttt{KP()}, \texttt{SC()}, and \texttt{RANK()}. There are several choices in the literature for the first two \cite{Vazirani01}, so we briefly discuss one possible choice for the third. A general observation is that an item $i\in\mathcal{L}$ is more important or desirable if its demand $d_i$ is high, or its cost $c_i$ is low, or if the number of elements it covers $|S_i|$ is high. Based on this, a viable candidate for $\mbox{\texttt{RANK}}\left(\mathcal{L},\{d_i\}_{i\in\mathcal{L}},\{c_i\}_{i\in\mathcal{L}},\{S_i\}_{i\in\mathcal{L}}\right)$ would be a method that ranks items in increasing order according to the value $v_i=\left(\frac{d_i}{\sum_{j\in \mathcal{L}} d_j} + \frac{|S_i|}{|\mathcal{I}|}\right)/c_i$, where $\mathcal{I}=\bigcup_{i\in\mathcal{L}}S_i$ denotes the set of all elements covered by items in $\mathcal{L}$.
\subsubsection{Computing charging station costs.}\label{appendix-cost-queueing}
In this section, we briefly address estimating the setup costs $c_i$, which depend on (i)~the number of charging slots necessary at $i$ to satisfy any desired SLA on waiting times, and (ii)~per-unit setup cost, which can include infrastructure as well as land/licensing costs. Since data on the per-unit costs are generally available, we focus on (i), for which we model candidate charging stations as a multi-server queue.
One possible way to model each candidate charging station at location $i$ in order to estimate its size (and hence cost) is using a multi-server queue that follows an $M$/$M$/$N_i$ discipline, where $N_i\in\mathbb{Z}_+$ is the number of charging slots to be set up. Customers arrive into the queue according to a Poisson process with rate $\lambda_i$ per time unit. The service time for each customer is exponentially distributed with mean $1/\mu_i$ time units. $\lambda_i$ can be derived from the demand $d_i$, whereas $\mu_i$ can be derived from existing data on the average customer service time in nearby existing charging stations. The average time a customer waits before service is then given by $\mathbb{E}[W] = \frac{\mbox{\texttt{ErlC}}\left(N_i,\frac{\lambda_i}{\mu_i}\right)}{N_i\mu_i-\lambda_i}$, where $\mbox{\texttt{ErlC}}(N,\rho) = \left(\frac{\rho^N}{N!}\frac{N}{N-\rho}\right) / \left(\sum_{j=0}^{N-1}\frac{\rho^j}{j!}+\frac{\rho^N}{N!}\frac{N}{N-\rho}\right)$~\cite{Kleinrock75}. Suppose the SLA to be met is given by $\mathbb{E}[W] \leq t_i$. Then, because it is well-known~\cite{Kleinrock75} that $\mathbb{E}[W]$ is a decreasing function of $N_i$, we simply choose the smallest $N_i$ (e.g., using binary search) such that the SLA is satisfied.
\subsubsection{Optimizing for reachability.}\label{ssec:reachability}
In addition to maximizing the demand for a given reachability radius, it may also be desirable to minimize the reachability radius itself. Without loss of generality, we assume that $r\in[R^{\min},R^{\max}]$, where $R^{\min} = \max_{\ell\in\mathcal{I}}\ \min_{i\in\mathcal{L}}\ \mbox{\texttt{dist}}(\ell,i)$, and $R^{\max} = \max_{\ell\in\mathcal{I}}\ \max_{i\in\mathcal{L}}\ \mbox{\texttt{dist}}(\ell,i)$. Here, $\mbox{\texttt{dist}}(\ell,i)$ denotes the distance between locations $\ell$ and $i$ according to an underlying transportation network.\footnote{The lower bound for $r$ stems from the observation that when $r<R^{\min}$, even if charging stations are placed at all candidate sites, there would be at least one uncovered location of interest; so, no feasible solution exists for such $r$. The upper bound for $r$ follows from the fact that even if there is only a single selected site where a charging station is placed, and the solution is feasible, $R^{\max}$ is, by definition, the maximum distance to be travelled in order to reach that charging station.} Let $D^*(r)$ denote the maximum demand covered (obtained from a solution to the MPC problem (Figure \ref{fig:eqn-MPC}, perhaps by using Algorithm~\ref{alg:iter-pack-cover}) for a given reachability radius $r$. (For convenience, we define $D^*(r)=0$ for those $r$ for which MPC is infeasible.) If $\alpha\in[0,1]$ is a given trade-off parameter, then the objective can be defined as:
\begin{equation}\label{reachability}
\max_{r\in[R^{\min},R^{\max}]} \alpha\frac{D^*(r)}{\sum_{i\in \mathcal{L}} d_i} + (1-\alpha)\frac{R^{\max}-r}{R^{\max}-R^{\min}}.
\end{equation}
Let $\mathcal{R}=\{\mbox{\texttt{dist}}(\ell,i)|\ell\in\mathcal{I},i\in\mathcal{L},\mbox{\texttt{dist}}(\ell,i)\geq R^{\min}\}$ denote the set of all distances from a location of interest to a candidate site that is at least $R^{\min}$, and suppose we represent this set as $\mathcal{R}=\{r_1,r_2,\ldots,r_{|\mathcal{R}|}\}$, where $R^{\min}=r_1<r_2<\ldots<r_{|\mathcal{R}|}=R^{\max}$. The following lemma presents two observations concerning $D^*(r)$:
\begin{lemma}
$D^*(r)$ is nondecreasing,\footnote{This is because, for any two radii $r_1$, $r_2$ with $r_1\leq r_2$, the feasible set of MPC for $r_1$ is a subset of that for $r_2$.} and for all $1\leq m < |\mathcal{R}|$, for all $r_m\leq r < r_{m+1}$, $D^*(r)=D^*(r_m)$.
\end{lemma}
In other words, $D^*(r)$ is a nondecreasing step function and remains unchanged between values of $r$ that do not correspond to $\mbox{\texttt{dist}}(\ell,i)$ for some $(\ell,i)$. Thus, it can be seen that the value(s) of $r$ at which the objective function in \eqref{reachability} attains its maximum must be among the elements in $\mathcal{R}$. Therefore, solving \eqref{reachability} is equivalent to solving:
\begin{equation}\label{reachability-integral}
\max_{r\in\mathcal{R}} \alpha\frac{D^*(r)}{\sum_{i\in \mathcal{L}} d_i} + (1-\alpha)\frac{R^{\max}-r}{R^{\max}-R^{\min}}.
\end{equation}
\subsection{Experimental results.}\label{ssec:expt-optimal-placement}
We conducted two sets of experiments to evaluate the performance of IPAC using real world data -- one for North East England (presented in this section), and the other for the western region of East Anglia and southern region of East Midland (presented in Appendix~\ref{appendix-second-expt}).
\subsubsection{Setup.}
We use 1305 parking locations in North East England for both the candidate sites for placing charging stations, as well as the locations of interest that need to be covered. This is equivalent to a requirement that each candidate site must either have a charging station on-site, or one that is a short driving distance away. Accordingly, the cover set $S_i^r$ for each candidate site $i$ consists of the candidate sites that are within a distance of $r$ from $i$. The charging demands at candidate sites are predicted using the MDR model trained in Section \ref{ssec:expt-demand-prediction}. The costs are computed as $c_i = N_i(L_i+F_i)$; we estimate each of these components as follows:
\begin{enumerate}
\item \textit{$L_i$, the per-unit land cost}: Points of interest around candidate sites play vital roles in determining their land values. In addition, different types of facilities affect the land value differently. Thus, we assume $L_i = p + \sum_{j\in P_i^\delta}\frac{\mbox{\texttt{Score}}_{ij}}{\mbox{\texttt{dist}}(i,j)}$, where $p=\$4000$ is the minimum per-unit land cost across all the locations in North East England (computed using data obtained from~\cite{NEUK-Land-Value}), $P_i^\delta$ is the set of points of interest that are within a radius $\delta=1$~km from candidate site $i$, and $\mbox{\texttt{Score}}_{ij}$ is the score assigned to the point of interest $j$ according to its type, as follows. Airports and railway stations have the highest score of $800$, whereas schools, restaurants and hospitals have a lower score of $300$. $\mbox{\texttt{Score}}_{ij}$ is then normalized by $\mbox{\texttt{dist}}(i,j)$.
\item \textit{$F_i$, the per-unit infrastructure cost}: For Level 2 charging, we set $F_i=\$1852$ from Table~6 of~\cite{EVCS-Cost-Luskin}.
\item \textit{$N_i$, the number of charging spots}: We assumed a Level~2 charging rate of $6.4\mbox{kW}$ and set, for each candidate site, $N_i$ to be the minimum number of Level~2 charging spots necessary (using the queueing model described in Section~\ref{appendix-cost-queueing}) to ensure that the average ``peak-demand'' waiting time (taken as the estimated maximum hourly demand at the candidate site over two years) is less than $5$ minutes.
\end{enumerate}
The predicted demands and estimated costs of the candidate sites are then used to find the optimal locations using Algorithm~\ref{alg:iter-pack-cover} (IPAC), where, for both \texttt{KP()} and \texttt{SC()}, we choose the well-known greedy approximation algorithms introduced in~\cite{Vazirani01}, and for \texttt{RANK()}, we use the function proposed in Section~\ref{iter-pack-cover}.
\begin{figure*}[!ht]
\centering
\subfigure[\scriptsize{Demand satisfied by LP-relaxation, Optimal ILP-solution, IPAC-heuristic, and Naive-heuristic.}]{\label{fig:expt1a}\includegraphics[height=1.5in, width=0.32\textwidth]{exp-demand.png}}
\hfill
\subfigure[\scriptsize{Fraction of ILP-solution demand satisfied by IPAC-heuristic and Naive-heuristic.}]{\label{fig:expt1b}\includegraphics[height=1.5in, width=0.32\textwidth]{exp-ratio.png}}
\hfill
\subfigure[\scriptsize{Minimum required budget for feasibility by IPAC-heuristic and LP-relaxation.}]{\label{fig:expt2a}\includegraphics[height=1.5in, width=0.32\textwidth]{exp1c.png}}
\\
\subfigure[\scriptsize{Increase in demand satisfied by IPAC-heuristic over Naive-heuristic.}]{\label{fig:expt1e}\includegraphics[height=1.5in, width=0.49\textwidth]{exp1e.png}}
\hfill
\subfigure[\scriptsize{Feasibility gap between IPAC-heuristic and Naive-heuristic.}]{\label{fig:expt1d}\includegraphics[height=1.5in, width=0.49\textwidth]{exp1d.png}}
\\
\subfigure[\scriptsize{Running time of IPAC as compared to directly solving the ILP using CPLEX.}]{\label{fig:expt-runningtime}\includegraphics[width=0.45\textwidth]{TimeTaken}}
\caption{Experimental results for North East England.\label{fig:expt1}}
\end{figure*}
\subsubsection{Performance.}
To evaluate the performance of IPAC when it can locate feasible solutions, we compare it with (i)~a naive heuristic that first solves the covering problem using \texttt{SC()}, and then invokes \texttt{KP()} on the remaining budget to add unselected candidate sites, (ii)~the optimal ILP solution of the MPC problem (Figure~\ref{fig:eqn-MPC}), and (iii)~a solution to the LP relaxation of MPC problem (Figure~\ref{fig:eqn-MPC}), which gives an upper bound on the actual (integer LP) optimization problem. We use a budget of $B=\$6\mbox{M}$. The number of pairwise distances between candidate sites is prohibitively large for our experiment; so, we used values of $r$ (in units of km), from the set $\{5.1,6,7,8,9,10,12,15,20,25,30,40\}$. The results are plotted in Figures~\ref{fig:expt1a}-\ref{fig:expt1b}.
As we increase the reachability radius $r$, the set of allocations that satisfy the covering constraints steadily gets larger, until, for a large enough radius, any allocation would satisfy the covering constraints, reducing the optimization to a pure packing problem. Since the feasibility set only gets larger with $r$, the demand covered should also increase accordingly. The graphs validate this expected behavior. In addition, we observe that the demand captured by IPAC is far closer to that of the LP relaxation than that captured by the naive heuristic. In particular, from Figure~\ref{fig:expt1b}, we can see that when $r=9\mbox{km}$, the IPAC heuristic already captures almost $90\%$ of the LP relaxation demand and its performance is always better than the naive heuristic. Also, the time taken to obtain the ILP solution for a given radius using CPLEX~\cite{cplex2009},\footnote{\scriptsize{Intel(R) Xeon (R) CPU @ 2.2 GHz (16 cores), 32GB RAM, 64-bit Windows.}} ranges from $23-30$ hours for most instances, whereas IPAC finished in less than 0.5 seconds; see Figure~\ref{fig:expt-runningtime}.
\subsubsection{Feasibility.}
Since IPAC is a heuristic, it may falsely deem an instance of the MPC problem (Figure~\ref{fig:eqn-MPC}) infeasible, when in fact, feasible solutions exist. Thus, it is important to first analyze the extent of this limitation. It is easy to observe that an instance of MPC is feasible if and only if the available budget $B$ is greater than or equal to the minimum required budget determined by solving the corresponding instance of the set-cover subproblem (Figure~\ref{fig:eqn-SC}). In the case of IPAC (Algorithm~\ref{iter-pack-cover}), it reduces to whether the available budget $B$ is greater than or equal to the minimum required budget as determined by \texttt{SC()}.
Therefore, given an instance of MPC, we look at the corresponding set-cover subproblem and compare its solutions (the minimum required budgets) as obtained by (i)~using the greedy approximation algorithm in~\cite{Vazirani01}, and (ii)~directly solving its relaxed LP.\footnote{\scriptsize{Computing ILP solutions for feasibility analysis exactly is prohibitively time-consuming, since it must be computed for various $r$ and $B$ values.}} We use different values of $r$ (in units of km) to generate instances. The results are plotted in Figure~\ref{fig:expt2a}. (Since the LP relaxation allows for fractional allocations, the corresponding solution is only a lower bound on the minimum budget required for feasibility.)
\subsubsection{Accounting for noisy costs.}
Since our model for estimating the costs may not be accurate, we perform experiments for two additional scenarios, where we add zero-mean gaussian noise to the costs --- one with a standard deviation of \$1000 and another with a standard deviation of \$3000. In each case, we compare IPAC and the naive heuristic:
\begin{itemize}
\item \textit{Performance:} Using a budget of \$6M, we calculate the difference in the demands satisfied by the solution obtained using IPAC and the naive heuristic. Figure~\ref{fig:expt1e} shows this difference as a percentage, for several values of the reachability radius. It can be seen, once again, that IPAC's advantage is significant (8-18\% in most cases).
\item \textit{Feasibility:} We calculate the minimum reachability radius for which IPAC and the naive heuristic are able to find a feasible solution to MPC, and find that IPAC is always able to find feasible solutions for smaller radii. Figure~\ref{fig:expt1d} shows this difference as a percentage, for several values of the available budget. It can be seen that IPAC's advantage over the naive heuristic is particularly remarkable (10-20\% in most cases) when the budget is scarce.
\end{itemize}
Thus, experimental results show that IPAC's feasibility gap as compared to the LP-Relaxation is not significant, and IPAC's performance quickly approaches that of the LP-Relaxation demand. Further, in terms of both feasibility and performance, it can be seen that the advantage of IPAC over the naive heuristic is significant (10-20\% in most cases) and this advantage is fairly robust to noise in the estimated costs.
\section{Extensions and future work.}
In this section, we explain how the optimization framework we introduced in Section \ref{sec:placement-optimization} can be extended to a situation where the deployment of charging stations is incremental, and conclude with some open questions.\footnote{Even though the work in this paper is motivated by EV charging stations, the problem formulation, modeling and solution frameworks presented here can be extended to other facility location problems which have demand, budget and coverage requirements, such as placement of bus-stop shelters, parking lots, and healthcare kiosks.} To highlight another important alternate scenario, consider a set of private charging station providers that would like to set up charging stations, but would need government subsidies to incentivize them to do so, especially in regions of low demand. The government agency would then run a grant program~\cite{Grants-CA,Grants-UK}, where the providers specify the subsidies they need to set up charging stations at candidate sites they are interested in. The government has a budget from which to allocate grants; moreover, the providers come with their own budget constraints. In Appendix \ref{ssec:government-grant}, we expand the placement optimization problem (MPC) to accommodate these additional packing constraints and explain how to extend the IPAC framework accordingly by solving a multi-dimensional knapsack subproblem.
\subsection{Incremental (multi-period) placement.}\label{ssec:incremental}
In practice, the budget for the deployment of charging stations may be released over time, and the demands at candidate sites and charging station installation costs may change over time. We now extend the problem formulation and the heuristic framework presented in Section \ref{sec:placement-optimization} for this case. Let there be $T$ time periods, $t = 1, \ldots, T$. We first extend earlier notation to include time periods, by superscripting them with $t$. Then, the single-period formulation in Figure~\ref{fig:eqn-MPC} and Equation~\eqref{reachability-integral} can be combined and generalized for the multi-period placement optimization problem as follows:
\begin{equation}\label{mpct}
\begin{split}
\max_{\substack{r^t\in[R^{\min},R^{\max}]\\ x^{t}_{i}\in\{0,1\}}} &\sum_{t=1}^{T} \left( \alpha\frac{\sum_{i\in\mathcal{L}^t}d^{t}_{i} x^{t}_{i}}{\sum_{i\in \mathcal{L}^t} d^{t}_{i}} + (1-\alpha)\frac{R^{\max}-r^t}{R^{\max}-R^{\min}} \right) \\
\mbox{subject to } &x^{t}_{i} \geq x^{t-1}_{i} \qquad \forall\ i \in\mathcal{L}^{t-1}; 1 \leq t \leq T \\
&\sum_{t=1}^{\tau} \sum_{i\in\mathcal{L}^t}c^{t}_{i} (x^{t}_{i} - x^{t-1}_{i}) \leq \sum_{t=1}^{\tau} B^t \quad \forall\ 1 \leq \tau \leq T\\
&\sum_{i\in\mathcal{L}:\ell\in S_{i}^{r^{t}}}x^{t}_{i} \geq 1 \qquad \forall\ \ell\in\mathcal{I}^t; 1 \leq t \leq T\\
\end{split}
\end{equation}
In the above formulation, initially, $x^{0}_{i}=0$ at all candidate sites $i\in\mathcal{L}^0$. If a charging station is placed at location $i$ at time $t>0$, then $x^{t}_{i}$ is set to $1$, and the first constraint ensures that the charging station remains installed for all subsequent periods. The second constraint ensures that the cost of all (new) installations done at time $\tau$ is within the budget released at $\tau$ plus any leftover budget from previous time periods, and the third set of constraints is the covering constraints for each time period $t$, as a function of the reachability radius $r^t$. The objective is to maximize the fraction of demand satisfied while minimizing the reachability radius, summed over all time periods. A straightforward ``greedy'' heuristic for this problem is to predict demands and estimate costs at the remaining candidate sites at the beginning of each time period $t$, and then using Algorithm \ref{alg:iter-pack-cover} (IPAC) to locate additional charging station installations.
A common and important scenario that any solution for incremental placement must handle is future increase in demand at previously deployed charging stations. The greedy heuristic above only indirectly addresses this concern by ensuring that the increased demand at such charging stations would be taken into account by the demand prediction algorithm in subsequent time periods. Still, a simple modification can allow for the option of expanding existing charging stations by adding more charging slots in response to increasing demand---let the set of candidate sites at time period $t+1$ be a superset of the set of candidate sites at time period $t$, i.e., $\mathcal{L}^{t+1}\supseteq\mathcal{L}^t$, but the costs $c^{t+1}_i$ at previously selected candidate sites $i$ would now be the cost of adding additional charging slots to maintain the waiting time constraints. This would let IPAC decide whether the right thing to do in response to increasing demand is to expand an existing charging station or set up a new charging station nearby. While our greedy heuristic is a starting point, more effective algorithms likely exist, especially when future demands and installation costs can be predicted with reasonable accuracy. We leave such directions to future work.
\subsection{Extending the IPAC framework.}
Further analysis of the class of algorithms defined by the IPAC framework would be of interest. In particular, it is worth exploring provable performance guarantees of an IPAC algorithm, e.g., Algorithm \ref{alg:iter-pack-cover}, in terms of those of its constituent knapsack and set-cover algorithms. Another interesting direction would be to investigate if the demand prediction model can be integrated into the optimization framework for the placement of charging stations, since the placement of a charging station at one candidate site would likely affect the demand at nearby candidate sites.
\bibliographystyle{ormsv080}
|
\section{Experimental Study}
\label{sec-exp}
In this section, we use the Microsoft Academic Graph~\cite{Sinha15:MAG} to evaluate the effectiveness of our ranking model \ewpr in terms of four aspects: (1) Time-Weighted PageRank {\em vs.} PageRank, (2) single ensembles {\em vs.} multiple ensembles, (3) ensemble models {\em vs.} mixed models and (4) ensemble models with affiliations.
\subsection{Experimental Settings}
We first present our settings.
\eat{
\begin{table}[t!]
\label{tab-statistics}
\begin{center}
\begin{scriptsize}
\vspace{1ex}
\begin{tabular}{|c|c|c|}
\hline
{\bf Entity / Relation} & {\bf Quantity in Phase~1} & {\bf Quantity in Phase~2} \\
\hline\hline
Paper & $122,675,085$ & $120,887,833$ \\ \hline
Author & $123,017,488$ & $119,892,201$ \\ \hline
Venue & $24,841$ & $24,843$ \\ \hline
Affiliation & $2,716,493$ & $19,849$ \\ \hline
Fields of study & $53,834$ & $53,830$ \\ \hline
Reference & $757,462,733$ & $952,364,264$ \\ \hline
P-A & $324,948,062$ & $312,034,259$ \\ \hline
P-V & $45,783,880$ & $45,290,168$ \\ \hline
\end{tabular}
\vspace{-5ex}
\end{scriptsize}
\end{center}
\caption{Statistics of MAG}
\vspace{-3ex}
\end{table}
}
\stitle{(1) Dataset}.
The Microsoft Academic Graph (MAG) dataset is a heterogeneous graph containing different types of literature entities and relationships. Please refer to~\cite{Sinha15:MAG} for more details about MAG.
\stitle{(2) Metric}.
We adopt {\em pairwise accuracy}~\cite{Richardson06:BPR} to evaluate the ranking quality, which is the fraction of times that a ranking agrees with the correct importance orders of scholarly article pairs:
\vspace{-1ex}
\begin{small}
\begin{equation}
\label{eq-metric}
\PairAcc=\frac{\#\mbox{ of agreed pairs}}{\# \mbox{ of all pairs}}.
\end{equation}
\end{small}
\vspace{-1ex}
The ground truth is generated by human experts, who are asked to give the orders of importance of article pairs ({\footnotesize https://wsdmcupchall-enge.azurewebsites.net/Home/Rules}).
\stitle{(3) Baselines}.
We compare our method \ewpr, with four baseline methods: \blpr, \blwpr, \ewprall, a variant of \ewpr that further uses the affiliation ensemble, and \blmulrank~\cite{Jiang12-MRank}.
\blpr simply runs PageRank on the citation network, and uses the authority scores as the importance of articles,
\blwpr is only based on the results of the citation ensemble,
\blmulrank uses the mixed model, where entities are exploited simultaneously, and
\ewprall\ further uses the affiliation ensemble on the basis of \ewpr.
\stitle{(4) Implementation}. For all algorithms: (1) the number of iterations is set to $30$, and (2) the damping parameter $d$ and decaying factor $t$ are fixed to $0.15$ and $2.5$, respectively.
\blmulrank uses the default parameters recommended in~\cite{Jiang12-MRank}.
We further fix $\alpha=1.2$ and $\beta=0.3$ for \ewpr and \ewprall.
And the weight of affiliation ensemble for \ewprall is also set to $0.3$.
All experiments were run on a PC with 2 Intel Xeon E5--2630 2.4GHz CPUs and 64 GB of memory.
\subsection{Experimental Results}
\label{subsec-expres}
We next present the experimental results, which were mainly tested using the training data of WSDM Cup Phase~1. And all experiments were tested without using the external data, which is for Phase~2. The results are reported in Table~1.
\stitle{(1) Time-Weighted PageRank {\em vs.} PageRank}. We first compare the effectiveness of Time-Weighted PageRank with PageRank. The pairwise accuracy of \blpr and \blwpr is $0.687$ and $0.701$, respectively. And our model \blwpr outperforms \blpr by $1.4\%$, which is achieved by introducing a time decaying factor. The results show that Time-Weighted PageRank combining both time and structural information is a better choice for ranking scholarly articles.
\stitle{(2) Single ensembles {\em vs.} multiple ensembles}. We then compare the effectiveness of using single ensemble, \ie the citation ensemble, with multiple diverse ensembles. The pairwise accuracy of \blwpr and \ewpr is $0.701$ and $0.733$, respectively. \ewpr outperforms \blwpr by $3.2\%$ since it combines the citation, venue and author ensembles to evaluate importance of articles. Hence, multiple ensembles are typically more effective than single ensembles.
\stitle{(3) Ensemble models {\em vs.} mixed models}. We also compare the effectiveness of the ensemble model with the mixed model. The pairwise accuracy of \blmulrank and \ewpr is $0.699$ and $0.733$, respectively.
And \ewpr is better than \blmulrank by $3.4\%$. Moreover, \blmulrank is even worse than the single ensemble method \blwpr. We believe that the effectiveness of the mixed model is impaired by error propagation in a noisy dataset, whereas
the ensemble model controls the impacts of error propagation to some extent. Combining with the previous one, we claim that in noisy heterogeneous graphs, models assembling multiple ensembles are the best choice.
\stitle{(4) Ensemble models with affiliations}. In the last set of tests, we compare \ewpr with \ewprall to evaluate the effectiveness of ensemble models with affiliations. The pairwise accuracy of \ewpr and \ewprall is $0.733$ and $0.711$, respectively. And the incorporation of the affiliation ensemble decreases the pairwise accuracy by $2.2\%$. The results verify our early claim that the correlation between the importance of articles and authorities of affiliations is not as strong as the other three. Moreover, different types of entities may have different contributions to the ranking of articles. Simply combining all information may not be the best way.
Finally, the accuracy of \ewpr in the Leaderboard of Phase~1 is $0.656$ ({\footnotesize https://wsdmcupchallenge.azurewebsites.net/Home}).
\eat{
\stitle{Summary}.
From these tests we find the following.
\noindent(1) The Time-Weighted PageRank which weights the propagation of authority achieves good effectiveness for ranking scholarly articles, \ie outperforming the traditional PageRank by $1.4\%$ in terms of pairwise accuracy.
\noindent(2) The ensemble model is better than using citation information alone and the mixed model by $3.2\%$ and $3.4\%$ in pairwise accuracy, respectively. We believe that the ensemble model is the best choice in the scenario of noisy heterogeneous graphs.
\noindent(3) Different types of literature entities may have different contributions to the ranking of scholarly articles. Simply combining all information does not mean the best result, \eg affiliation information damages the effectiveness.
}
\begin{table}[t!]
\label{tab-result}
\begin{center}
\begin{scriptsize}
\vspace{1ex}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
{\bf Methods} & \blpr & \blwpr & \blmulrank & \ewpr & \ewprall \\
\hline
${\PairAcc}$ & $0.687$ & $0.701$ & $0.699$ & {\bf 0.733} & $0.711$ \\ \hline
\end{tabular}
\vspace{-4.5ex}
\end{scriptsize}
\end{center}
\caption{Results of pairwise accuracy}
\vspace{-3ex}
\end{table}
\section{Dealing with Missing Data}
\label{sec-impl}
Data quality is one of the most challenging issues in large scale data management, especially for data from open domains and multiple sources, \eg the Microsoft Academic Graph (MAG)~\cite{Sinha15:MAG}.
The early version of MAG has $120$ million scholarly articles, among which we find that there are about $73$ million articles without references and about $77$ million ones without venues. The ranks of those articles with missing information are underestimated by our model \ewpr, since ensembles assign the minimum scores to articles. As a result, data missing seriously impairs the ranking accuracy.
As for references and venues, the later are easier to obtain, and each filled venue can have a direct and substantial impact on the article ranking, \ie $R_v(u)$ of Eq.~(\ref{eq-ensemble}). In contrast, a filled reference only has an indirect and slight impact. Hence, we decide to use external data to fill in missing venues.
\stitle{Data collecting}.
The raw external data is collected from publicly available Digital Libraries, such as IEEE Xplore ({\footnotesize http://ieeexplo-re.ieee.org/gateway/}), PubMed ({\footnotesize http://www.ncbi.nlm.nih.gov/pub-med/}) and DBLP ({\footnotesize http://dblp.uni-trier.de/db/}). In total, we collect $2.8$ million articles with venue information as our external data, in which there are $57,000$ different venues.
\stitle{Data preprocessing}.
The venues in MAG are well processed, and are replaced by their series names. For example, {\em ``9th International Conference on Web Search and Data Mining, 2016''} is replaced with {\em ``Web Search and Data Mining''}. This makes it hard to directly link with the collected raw venue names. Hence, we preprocess raw venue names for the simplification of subsequent venue linking.
We first remove stop words such as {``on''} and common words like {``Conference''}, as well as years and some special characters from collected raw venue names. Then the same venues are merged, and the number of different venue names is reduced to $42,000$.
\stitle{Data linking}.
The final and also the most important step of filling missing venue information is to link each collected venue name to an existing one in MAG. Intuitively, linking based on name similarity is the most effective way such that two venues are linked if their names bear high similarity. We exploit the Jaro metric to evaluate the name similarity, which is based on the number and order of the common characters between two strings, and obtains good results in tasks such as record linkage and name matching~\cite{Cohen03strcompa}. Formally, a collected venue name is linked to an existing one in MAG if their Jaro similarity exceeds a pre-define threshold.
However, such a threshold is nontrivial to determine in practice. A high threshold can guarantee the accuracy of linked pairs, while only a tiny proportion of collected venue names are linked. On the other hand, a low threshold increases the number of linked pairs, which, in the same time, also introduces many errors. In order to reach a good balance between the number of linked pairs and the accuracy, we propose to combine another constraint on topic similarity of venues for linking, and only weaker filter conditions need be used in both constraints.
In MAG, fields of study (FOS) represent research topics of articles, such as {\em Web pages}, and {\em language technology}. Hence, we use FOS to evaluate the topic similarity of two venues. There are about $54,000$ FOS in MAG and most articles are assigned with two or three FOS. Let the set of FOS of each venue be the union of the sets of FOS of articles published in that venue. And the topic similarity of two venues based on FOS is defined as:
\vspace{-1ex}
\begin{small}
\begin{equation} \label{eq-fos}
TS(s,t)=({|F_s\bigcap F_t|})/{\sqrt{|F_s|\cdot|F_t|}},
\end{equation}
\end{small}
\vspace{-1ex}
\noindent in which $s$ and $t$ are two venues, and $F_s$ and $F_t$ are the sets of FOS of $s$ and $t$, respectively.
When we link a collected venue name, it is directly linked to the most similar one in terms of name similarity, if their Jaro similarity exceeds a high threshold $\lambda$. Otherwise, we first use the topic similarity constraint to select several candidates in MAG, \ie venues whose topic similarities with the collected venue exceed a threshold $\theta$. Intuitively, these candidates are in the similar fields of the collected one. We then select the most similar candidate in terms of name similarity as its linked venue, if their Jaro similarity exceeds another threshold $\phi$. Hence, the collected venue is linked to the one to which it is similar in terms of both topics and names.
In our model \ewpr, threshold $\lambda$ is set to $0.95$, while thresholds $\theta$ and $\phi$ need not be very high, which are $0.5$ and $0.7$, respectively. Finally, $6,000$ among the $42,000$ collected venues are linked, resulting in $340,000$ (about $12\%$) articles with enriched venue information. Note that a majority of the collected venue names are not valid venues, such as booktitles and names of workshops, and cannot be linked to any one in MAG.
\section{Introduction}
\label{sec-intro}
Ranking scholarly articles is a critical and challenging task, due to the heterogeneity and dynamism of entities involved~\cite{fcs-biggraph}.
Generally speaking, a ranking is {\em a function that assigns each entity a numerical score}.
Such a ranking plays a key role in literature recommendation systems, especially in the {\em cold start} scenarios.
This paper focuses on ranking the importance of scholarly articles in a query independent way. As scholarly articles involve with authors, venues, affiliations
and references, they indeed form a complex heterogeneous graph.
Hence, this is essentially a problem of assessing the importance of nodes in a heterogeneous graph.
\begin{figure}
\centering
\includegraphics[scale=0.5]{./citation.eps}
\vspace{-2ex}
\caption{Citations \wrt the number of published years}
\label{fig-citation}
\vspace{-3ex}
\end{figure}
Comparing with homogeneous rankings such as of Web pages, ranking scholarly articles in a complex heterogeneous graph is much more challenging from two aspects. Firstly, even if we are only to rank one type of entities, \ie scholarly articles, other types of entities such as venues and authors are typically involved. Moreover, the impacts of different types of entities on the ranking of scholarly articles differ from each other. Secondly, entities are evolving, and the importance of an article varies with time~\cite{Li08TSRanking}. Recently published articles are more likely to have increasing impacts in the next few years, and those published many years ago tend to have decreasing impacts since people potentially care more about the latest results. For instance, the citations of articles typically increase in the first two years after publication, and then decrease after that, as shown by the statistics of the Microsoft Academic Graph (MAG)~\cite{Sinha15:MAG} in Figure~\ref{fig-citation}. Indeed, {\em how to accurately rank nodes in heterogeneous graphs remains a challenging task}.
Currently, structure based methods, such as PageRank~\cite{Brin98:PageRank} and its variant Weighted PageRank~\cite{Xing04:WPR}, are among the most effective ones for ranking scholarly articles. However, as pointed out in~\cite{Li08TSRanking}, these previous methods favour older articles that have accumulated a large number of links (\eg citations). However, recently published articles are often underestimated, and they potentially have an increasing impact. These motivate us to develop a new approach to ranking the importance of scholarly articles.
\stitle{Contributions}.
To this end, we propose Ensemble enabled Weighted PageRank (\ewpr) for ranking the importance of scholarly articles, which is among the first to address those challenges above.
\vspace{0.5ex}
\noindent(1) Firstly, we propose Time-Weighted PageRank that extends PageRank \cite{Brin98:PageRank} by introducing a time decaying factor, inspired by Weighted PageRank \cite{Xing04:WPR}, and the authorities of individual articles are discriminately propagated in terms of their own citation information with time, rather than equally propagated like PageRank.
\vspace{0.5ex}
\noindent(2) Secondly, we develop an ensemble method to assemble the authorities of the heterogeneous entities involved in scholarly articles,
which is much more flexible than the mixed model~\cite{Jiang12-MRank} that simultaneously exploits entities and directly produces the ranking.
\vspace{0.5ex}
\noindent(3) Finally, we propose to use external data sources to enrich data and further improve the ranking accuracy.
\stitle{Organization}. The rest of our paper is organized as follows. Section~\ref{sec-model} introduces the ranking model. Section~\ref{sec-impl} discusses how to deal with missing data using external sources. Experimental results are reported in Section~\ref{sec-exp}, followed by conclusions in Section~\ref{sec-conc}.
\section{Ranking Model}
\label{sec-model}
We first introduce our model for ranking scholarly articles.
\subsection{Time-Weighted PageRank}
\label{subsec-twpr}
PageRank~\cite{Brin98:PageRank} has been extensively applied to the ranking of scholarly articles~\cite{Li08TSRanking,Richardson06:BPR,sayyadi09,Zhou07-CoRank}, as hyperlinks among Web pages can be easily replaced with citation relationships among articles, and citation analysis plays a key role to evaluate the importance of scholarly articles. However, the direct use of PageRank for ranking scholarly articles is problematic in terms of the following:
\noindent(1) First, each article equally distributes its authority to its reference articles in the iteration of PageRank~\cite{Brin98:PageRank}, which essentially assumes that each article is equally influenced by its references. However, scholarly articles typically have different impacts in practice, and there is a need to differentiate the impacts of reference articles.
\noindent(2) Second, citation relationships are significantly different from hyperlinks, as the former are time-evolving, and have been successfully exploited in scholarly article ranking~\cite{Li08TSRanking,Wang13AAAI,sayyadi09}. Such temporal information is supplementary to purely structure based PageRank.
\stitle{Time-Weighted PageRank}. We incorporate time information into our ranking model. While in most previous work, time information is simply exploited in the form of exponential decay~\cite{Li08TSRanking,Wang13AAAI,sayyadi09}. We rethink the usage of time information in terms of the impacts of scholarly articles. Recall that Figure~\ref{fig-citation} illustrates the total number of citations \wrt the number of years after the publication of articles.
Here we use the number of citations to evaluate the impacts of articles. As we can see, the number of citations reaches a peak in two years, and gradually decreases after that. This statistical result conforms to our perception of the impacts of articles.
According to Figure~\ref{fig-citation}, the impacts of articles do depend on time, but not simply in the form of exponential decay. Specifically, if an article is cited after the citation peak, its impact should decay with time. Otherwise, its impact is fixed as a constant number, since we argue that the increment of its citations during this period is mainly due to the increase of its popularity.
Moreover, considering that different articles may reach their citation peaks in different ways, we compute the peak time for each individual article, rather than using the same citation peak for all articles.
Inspired by these properties of scholarly article citations, we present Time-Weighted PageRank that evaluates the authorities of nodes in a directed graph, in which each node is attached with time information. It differs from PageRank by weighting the influence propagation using the {\em impact weights on edges}, which represent the relative amounts of authorities that should be propagated from the edge sources to targets, and which also depend on the time information on nodes, following the same temporal tendency as scholarly article citations discussed above.
Formally, the impact weight on directed edge $(u,v)$, \ie edge from $u$ to $v$, is defined as:
\vspace{-1ex}
\begin{small}
\begin{equation} \label{eq-infl-weights}
w(u,v) = \begin{cases} \hspace{10ex} 1 & T_u < Peak_v \\
1/(\ln(e+T_u-Peak_v))^t & T_u \geq Peak_v,
\end{cases}
\end{equation}
\end{small}
\vspace{-1ex}
\noindent
where $T_u$ is the time information on node $u$, $Peak_v$ is the peak time of node $v$ using the time information of all nodes connecting to $v$, and $t$ is the decaying factor. By default, Eq.~(\ref{eq-infl-weights}) uses years as its time granularity. For the sake of completeness, we further set $w(u,v)$ to $0$ if these does not exist an edge from $u$ to $v$.
The authority update rule in Time-Weighted PageRank is:
\vspace{-1ex}
\begin{small}
\begin{equation}\label{eq-twpr}
PR(v)=(1-d)+d \cdot \sum_{u\in IN(v)} \frac{w(u,v)\cdot PR(u)}{W(u)},
\end{equation}
\end{small}
\vspace{-1ex}
\noindent where $PR(u)$ and $PR(v)$ are the authorities of $u$ and $v$, respectively, $IN(v)$ is the set of nodes having edges to $v$, $W(u)=\Sigma_{v} w(u,v)$ is the sum of impact weights on all edges from $u$, and $d$ is a damping parameter in $[0, 1]$. From Eq.~(\ref{eq-twpr}) we can see that authorities are based on the impact weights, not equally distributed.
\stitle{Remarks}. Note that here Eq.~(\ref{eq-twpr}) is indeed a more general update rule than Weighted PageRank~\cite{Xing04:WPR}, and the name of Time-Weighted PageRank comes from the use of time information in the initial impact weight $w(u,v)$ of Eq.~(\ref{eq-infl-weights}).
\subsection{Ensembles}
\label{subsec-ensemble}
We start this part by thinking about how people evaluate the importance of scholarly articles. In practice, the importance of an article can be evaluated according to many factors such as citations, venues and authors. Only focusing on the citation information limits the accuracy of the results. Consider the case when we are to evaluate a newly published article whose citations are not currently available. In this case citation information fails to give a reasonable rank, but other information such as venues and authors could be used instead to refine the rank. Hence, we propose the use of an ensemble model, in which each ensemble is essentially a ranking based on the authorities of one type of heterogeneous entities, and these ensembles are assembled to produce the final ranking.
\stitle{Citation ensemble}.
The first ensemble is based on the authorities of articles and it is called citation ensemble since we use citation information to evaluate these authorities.
Specifically, it first uses citation information to construct a directed graph, where a node represents an article and an edge $(u,v)$ denotes that $u$ cites $v$ as its reference. The graph is further associated with time information such that (1) the publication years of articles are attached to corresponding nodes, and (2) the peak time of each node is the year with the largest number of citations, in which ties are broken randomly if existing.
After that, Time-Weighted PageRank is run on the graph and each node is assigned its authority. Finally, the ensemble maps each article to the authority of its corresponding node as its rank.
\stitle{Venue ensemble}.
The second ensemble is based on the authorities of venues. It first evaluates the authority of each venue, and then maps each article to the authority of the venue where it is published as its rank.
To do this, we also construct a directed graph, in which a node represents a venue and an edge $(s,t)$ means that there is at least one article published in $s$ citing at least one article published in $t$. We also use {\em impact weights} to denote the weights among venues.
And the impact weights are defined as sums of impact weights between articles published in the corresponding venues:
\vspace{-1ex}
\begin{small}
\begin{equation} \label{eq-infl-weights-v}
w_v(s,t) = \sum_{u\in C(s), v\in C(t)} w(u,v).
\end{equation}
\end{small}
\vspace{-1ex}
Here, $C(s)$ and $C(t)$ are the collections of articles published in venues $s$ and $t$, respectively, and $w(u,v)$ is the impact weight of articles $u$ and $v$ produced in the citation ensemble. It then iteratively computes the authorities of venues using the impact weights of venues and the update rule in Eq.~(\ref{eq-twpr}).
\stitle{Author ensemble}.
The third ensemble is based on the authorities of authors. Similar to the venue ensemble, we could first evaluate the authority of each author and then map each article to the average authority of the author(s) associated with the article as its rank.
However, the resulting graph is too large to handle. Hence, we adopt another way to evaluate the authority of an author, by using the average authority of all articles published by the author, which are produced by the citation ensemble.
\stitle{Affiliation ensemble}.
Recall that articles in our data are also associated with affiliation information. Following the way of the venue or author ensemble, we can derive another ensemble, \ie affiliation ensemble. However, we argue that the use of affiliation ensemble may have negative effects since the correlation between the importance of an article and the average authority of its affiliation(s) is not as strong as others such like authors and venues. As shown by the experimental study in Section~\ref{sec-exp}, the incorporation of the affiliation ensemble impairs the ranking accuracy. Hence, we choose not to use the affiliation ensemble in our model.
\stitle{Remarks}.
Traditional PageRank equally distributes the authorities of nodes, and PageRank based models suffer from the problem that older articles are preferred since they have accumulated a large number of citations~\cite{Li08TSRanking}, and Time-Weighted PageRank based models alleviate the problem to a certain degree by lowering the impact weights of articles when they are cited after their peak time, \ie $T_u\geq Peak_v$. We further propose the venue and author ensembles to improve the ranking accuracy.
\subsection{Ensemble Enabled Ranking}
\label{subsec-eerank}
The aforementioned ensembles (except the affiliation one) are finally assembled to produce the final ranking, referred to as \underline{E}nsemble enabled \underline{W}eighted \underline{P}age\underline{R}ank (\ewpr). Before assembling, each ranking is properly scaled such that the average scores of different rankings are the same. Suppose that the scaled ranking scores of articles $u$ are $R_c(u)$, $R_v(u)$, and $R_a(u)$ from the citation ensemble, venue ensemble and author ensemble, respectively. The final ranking score of $u$ is aggregated as follows:
\vspace{-1ex}
\begin{small}
\begin{equation} \label{eq-ensemble}
R(u) = \frac{R_c(u) + \alpha \cdot R_v(u) + \beta \cdot R_a(u)} {1+\alpha+\beta}.
\end{equation}
\end{small}
\vspace{-1ex}
\noindent Here parameters $\alpha$ and $\beta$ as well as the value $1$ are used to regularize the contributions of the citation, venue and author information. Intuitively, these values indicate the intensity of the correlation between the importance of articles and the specific information.
\begin{figure}
\centering
\includegraphics[scale=0.12]{./framework.eps}
\vspace{-2ex}
\caption{Architecture of our ranking model \ewpr} \label{fig-framework}
\vspace{-3ex}
\end{figure}
We close this section by presenting the architecture of our ranking model \ewpr, illustrated in Figure~\ref{fig-framework}.
Our model \ewpr contains three distinct ensembles, \ie the citation ensemble, venue ensemble and author ensemble. The citation ensemble directly uses Time-Weighted PageRank, while the other two are partially based on Time-Weighted PageRank. These ensembles are further assembled to produce the final ranking. As illustrated in Figure~\ref{fig-framework}, external data is also exploited in \ewpr. How to collect and use external data will be introduced in the coming section.
\section{Conclusions}
\label{sec-conc}
We have proposed a novel model for scholarly article ranking, which combines Time-Weighted PageRank and ensembles. The authority propagation of scholarly articles are weighted based on each article's individual citation information with time, and the ranking of articles is given by assembling the results of citation, venue and author ensembles. We have also proposed to use external data to enhance the quality of data for improving the ranking accuracy.
\balance
\stitle{Acknowledgments}.
This work is supported in part by 973 program ({\small No. 2014CB340300}), NSFC ({\small No. 61322207\&61421003}), and MSRA Collaborative Research Program. We also thank team members Dr. Xuelian Lin and Niannian Wu for their support.
For any correspondence, please refer to Shuai Ma.
\vspace{1ex}
\bibliographystyle{abbrv}
\begin{small}
|
\section{Supplementary material: Wannier-Stark problem}
We want to find solutions of the Wannier-Stark Hamiltonian:
\begin{equation}\label{aa}
H_{WS} = H_0 + F \hat{x},
\end{equation}
where:
\begin{equation}
H_0 = \frac{\hat{p}^2}{2m} + V \cos^2(k_L \hat{x}).
\end{equation}
The standard solutions of $H_0$ are Bloch functionslabelled by quasimomentum $k$ and the band number. for low temperatures the higher bands may be omitted due to a finite energy gap between bands. From the Bloch functions one may construct the basis of exponentially localized states - the Wannier functions given by the formula:
\begin{equation}
\psi_l(x) = \int_{-\pi/2}^{\pi/a} \exp(-i k l a) \phi_k(x) dk,
\end{equation}
where $\psi_l(x)$ is the Wannier function localized in l-th site, $a=\frac{\pi}{k_L}$ is lattice constant and $\phi_k(x)$ is Bloch function from the lowest band with quasimomentum $k$. Wannier functions satisfy relation $\psi_{l+1}(x) = \psi_l(x-a)$, which one would expect from translational symmetry. The Wannier functions are exponentially localized. In second quantization formalism and keeping only the nearest neighbour tunnelings, the standard tight-binding representation of $H_0$ reads
\begin{equation}
H_0 \approx \sum_{l} \epsilon a^\dagger_{l}a_{l} - \sum_{l} {t a^\dagger_{l+1} a_{l} + h.c.}.
\end{equation}
Adding now the linear tilt as in \eqref{aa} and incorporating only the diagonal part of the linear potential into the tight biding model results in the Hamiltonian:
\begin{equation}
H_{TB} = \sum_{l} (\epsilon + aFl) a^\dagger_{l}a_{l} - \sum_{l} {t a^\dagger_{l+1} a_{l} + h.c.}.
\end{equation}
The exact solution of that hamiltonian are Wannier-Stark functions, defined by:
\begin{equation}
\omega_l(x) = \sum_{m} J_{m-l}\left( \frac{2t}{aF} \right) \psi_m (x).
\end{equation}
For more information see \cite{Gluck02}.
\end{document}
|
\section{Introduction}
Topological insulators\cite{hasan2010colloquium,qi2011topological}
(TI) are a new class of materials that present a gap in the
bulk but on their edges they have gapless states that conduct without
dissipation. They have produced a renewed understanding of
the topological properties of the single-particle Hamiltonian that describes
the electronic structure of an insulator.
The topological properties
of the edge states arise due to a finite Chern number as in the quantum Hall
effect,\cite{hatsugai1993chern,thouless1982quantized}
they can be protected by different symmetries of the Hamiltonian,
such as crystalline \cite{fu2011topological}
or time reversal symmetry,\cite{kane2005z}
or they can be arise due to
interaction effects in topological
Mott insulators\cite{tmi_balents}
and topological Kondo insulators.\cite{dzero2010topological}
Among the different possible classes, TI with time reversal (TR) symmetry\cite{hasan2010colloquium}
show spin-momentum
locking in their edge states.
All sorts of peculiar physics has been predicted to arise from them, such as
topological superconductors showing\cite{qi2011topological} Majorana
bound states\cite{ti_majo} with anyonic
statistics.\cite{hastings2013metaplectic}
In terms of applications, those
dissipation-less edge state could help
in transmitting information
without losses,\cite{jiang2009topological}
find applications in spintronic devices\cite{pesin2012spintronics}
or become a key ingredient for topological
quantum computing.\cite{ti_majo,laflamme2014hybrid}
Topological insulators can appear in two or three dimensional systems. In two-dimensions, the first proposal for a topological insulator was the
prediction of a quantum spin Hall effect,\cite{ti_1} a peculiar type of quantum Hall
effect that is not caused by an external magnetic field but by the internal
spin-orbit interaction of the material. This was first proposed to appear in
graphene\cite{ti_2} but not materialized experimentally. The same
ideas were shown to operate in HgTe/CdTe quantum
wells,\cite{hgte_cdte_prediction} where a larger SOC strength
takes place and allowed for its experimental observation.\cite{hgte_cdte_molenkamp}
Oxides provide also an interesting platform for non-trivial topological
properties to appear. Early since the discovery of topological insulators,
several oxides with different structures have been identified as possible
topological insulators: honeycomb-based iridates,\cite{na2iro3}
pyrochlores,\cite{pyrochlores_ti} corundum structure\cite{corundum,afonso}
and rutiles.\cite{vpardo_Dirac,cro2_nano}
The honeycomb lattice has been the simplest to investigate. Based on these concepts, Xiao et al.\cite{xiao} proposed to build an oxide-based honeycomb
lattice, though largely buckled, using perovskite bilayers grown along the
perovskite (111) direction, and other authors have followed suit, showing that
this kind of phases appear ubiquitously, even for the low SOC
limit.\cite{pentcheva_111} It was shown\cite{lado} that the
size of the topological gap depends on the
strength of the trigonal splitting, and only inversely with SOC
strength, hence these phases show up also for 3d electron systems, but there
they compete with Jahn-Teller distortions, charge and magnetic ordering.
\cite{xiao,lado,afonso}
Here we propose a different oxide-based route to realize topologically non-trivial properties. The system under study shows a
low-energy electronic structure
identical to the one of graphene, but with $d_{z^2}$
orbitals dominating the spectrum around the Fermi level instead of the $p_z$ states of graphene.
Our proposal is based of atomic substitution
on the recently synthesized compound
InCu$_{2/3}$V$_{1/3}$O$_3$,\cite{incuv,yan2012magnetic,yehia2010finite} which shows a honeycomb
lattice of S=1/2 Cu$^{2+}$ cations. We propose to substitute Cu by
isoelectronic Ag\cite{bratsch2014aag2}
and isolate the active Ag-rich planes by a Zn-rich layer to
produce a truly two-dimensional electronic structure.
In InZn$_{1/3}$Ag$_{1/3}$V$_{1/3}$O$_3$,
due to the interplay of local
$C_3$ symmetry, crystal field and charge transfer,
close to the Fermi energy
the band structure is composed
by Ag $d_{z^2}$ orbitals
forming an effective honeycomb lattice.
In the absence of SOC, the low-energy spectrum
is described by two gapless Dirac equations close to the $K$ and $K'$
points.
When relativistic
effects are included, the system opens up a
gap and shows a topological invariant $\nu=-1$, realizing
the quantum spin Hall effect as a d$_{z^2}$ version
of the Kane-Mele model.\cite{ti_2}
We
show that the quantum spin Hall effect
is resilient to certain time reversal symmetry breaking, such
as off-plane magnetism and gauge magnetic fields.
We also show that the gauge magnetic field yields
a ferromagnetic quantum spin Hall state, showing
the Landau level structure typical of Dirac fermions.
Finally, we summarize our conclusions.
\section{Electronic structure analysis}
\subsection{Antiferromagnetic Cu based compound}
We start describing the electronic structure
of the recently synthesized compound
InCu$_{2/3}$V$_{1/3}$O$_3$,\cite{incuv} which is formed by a honeycomb
lattice of S=1/2 Cu$^{2+}$ cations developing
an antiferromagnetic order. Even though its structure
(see Fig. \ref{bs_ag}a) looks very
complicated, composed by alternating layers
of In and Cu,V within an oxygen environment,
its electronic structure turns out to be very simple. With a
simple ionic model one can do the following electron count: In$^{3+}$,
V$^{5+}$: d$^0$ (both full/empty shell ions) and Cu$^{2+}$: d$^9$ (S=1/2).
The local environment of Cu is CuO$_5$, with the
z-axis being a C$_3$ symmetry axis and the three in-plane oxygens forming
120$^{\circ}$ (see Fig. \ref{bs_ag}b,c). In this situation the
d$_{z^2}$ orbitals lie
higher in energy, half filled, whereas the other d-orbitals
are completely filled.
Therefore, close to the Fermi level, only the higher-lying
d$_{z^2}$ orbitals are present, substantially separated
and electronically decoupled from all other
d orbitals. This
gives rise
to an effective single orbital weakly coupled pair of honeycomb lattices,
one per each Cu plane in the unit cell
(Fig. \ref{bs_ag}d).
This decoupling of the $d_{z^2}$ bands can be understood by the
local $C_3$ symmetry, which allows to expand the polar dependence
of the local potential as
$V_{C_3}(r,\theta,\phi) = \sum_n a_n e^{i3\phi n} v_n(r,\theta)$, with
$\phi$ the polar angle.
The previous expansion guarantees that
locally there is no mixing between $d_{z^2} (m=0)$ and the other
d-orbitals
$(m =\pm 1$ i.e. the xz/yz orbitals, $\pm 2$ i.e. the xy/$x^2-y^2$ ones),
given that any matrix element of the form
$\langle m=0 | e^{i3\phi n} | m=\pm1,\pm2 \rangle$
is identically zero for $n \neq 0$.
This can be seen in Fig. \ref{bs_ag}e, where the band structure of
the antiferromagnetic Cu-based
system is presented and the bands both just above and below the Fermi level
have Cu $d_{z^2}$ symmetry. Thus, to some extent, such a complicated material
like
InCu$_{2/3}$V$_{1/3}$O$_3$ turns out to be effectively a single-band compound.
Opposite to high-T$_c$ cuprates, where the d$_{x^2-y^2}$ band occurs around the Fermi
level, here it is the d$_{z^2}$ band that is central. The nearest-neighbor
antiferromagnetic coupling produces a 2D-AF, which is the ground state even
without correlations due to the low bandwidth of the Cu $d_{z^2}$ bands. Experimentally, a large J$_{AF}$= 120 K has been estimated, with a magnetic phase transition of some sort occurring at 38 K.\cite{incuv}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\columnwidth]{fig1.pdf}
\caption{(Color online.)
(a) Unit cell of the crystal,
showing alternating planes of Cu(Ag),V and In.
Panel (b) shows a sketch of the Cu(Ag) oxygen environment,
giving rise
to local $C_3$ symmetry. Panel (c) shows the Ag,V plane,
where the Ag atoms
form a (slightly distorted) honeycomb lattice
as shown in (d).
Panel (e) shows the
band structure of InCu$_{2/3}$V$_{1/3}$O$_3$ calculated at the GGA
level, yielding a gapped antiferromagnetic ground state.
In comparison,
for InAg$_{2/3}$V$_{1/3}$O$_3$
the ground state is non magnetic as
shown in (f).
In both cases,
the bands just around the Fermi level are d$_{z^2}$ from Cu for (e)
and Ag for (f).
}
\label{bs_ag}
\end{center}
\end{figure}
\subsection{Non magnetic Ag based compound}
A simple way to remove magnetism in this system is to
substitute directly Cu by isoelectronic Ag. Doing this, we change from
a 3d to a 4d electron system, with the corresponding increase in spin-orbit
coupling strength and band width.
The solution
obtained for the compound InAg$_{2/3}$V$_{1/3}$O$_3$ is non-magnetic and its
band structure is shown in Fig. \ref{bs_ag}f.
The electron count is basically
retained: V$^{5+}$:d$^0$, empty-shell nonmagnetic cations and Ag$^{2+}$:d$^9$
atoms with a single hole in the aforementioned d$_{z^2}$ orbital, that becomes
decoupled by the peculiar AgO$_5$ environment this structure presents. In
this situation, the crystal-field energy that separates the d$_{z^2}$ bands
from the rest is even higher due to the more delocalized nature of the Ag 4d
electrons, and hence this system becomes even more one-orbital-like. We can see
that there is a 1.5 eV energy window around the Fermi level completely dominated by the four Ag
d$_{z^2}$ bands in the unit cell.
However the two Ag-rich planes
show a sizable interaction
due to their proximity, so that
the low-energy properties
are not the ones of two decoupled
honeycomb lattices.
\subsection{Dirac-like Ag-Zn based compound}
In order to produce
a single Dirac point in the band structure and change the topological
properties of the system, one can try to remove this
interaction between the active Ag layers
by
adding a spacing layer between them. Ideally, the
preferred experimental situation would be to grow a single layer of this
honeycomb lattice on the appropriate substrate (typically Al$_2$O$_3$ or some
other hexagonal-based substrate like a wurtzite nitride AlN, e.g.).
Another approach is to
introduce a spacing layer based on full-shell Zn$^{2+}$:d$^{10}$ cations in
the place
of Ag (Figs. \ref{mono_ag}a,b).
That way, interactions between the active Ag orbitals along the z-axis
will be largely reduced
and the problem becomes purely two-dimensional.
Calculation of the orbital resolved density of states in Ag
(Fig. \ref{mono_ag}c)
confirms that in such situation the low-energy levels
are still Ag d$_{z^2}$, showing also
a strong mixing between the other d-orbitals away from the Fermi level (Fig. \ref{mono_ag}d).
In the situation described above, we
can observe in Fig. \ref{mono_ag}a the band structure of
that layered alternation of Zn-rich and Ag-rich planes in the two honeycombs
that form the unit cell. Due to
the lack of interaction between active Ag-rich planes,
the low-energy spectrum are two Dirac cones
close to the K and K' points,
as expected from a simple one-orbital picture on a purely
two-dimensional honeycomb
lattice. Such electronic structure (see Fig. \ref{mono_ag}b) can be easily captured
by means of Wannier procedure projecting
onto the Ag d-orbitals, which gives rise to a
tight binding Hamiltonian of the form
\begin{equation}
H_W = \sum_{i,j} t_{ij} c_{j}^\dagger c_{i}
\end{equation}
where $c_i$ and $c^\dagger_j$ are (spinless) annihilation and creation operators
in the d-like Wannier Ag orbitals, and $t_{i,j}$ are the
hopping parameters obtained with the Wannierization procedure.
It is worth to mention that
the underlying Ag honeycomb lattice
has a small distortion
due to a small dimerization
between pairs of Ag sites.
The effect
of such structural feature is that the Dirac points are
not strictly located in the K and K' points, but
have a very small displacement in
reciprocal space.\cite{pereira2009tight,oliva2016effective}
\section{Topological insulating state in Ag-Zn multilayer}
The calculations presented so far do not consider
relativistic effects, so that the low-energy
Hamiltonian of the Ag-Zn compound are two gapless Dirac equations.
In comparison,
when SOC is included in the ab initio calculations,
we observe that a gap opens up in the Dirac points (Fig. \ref{mono_soc}a).
The former situation is analogous to the behavior of
other Dirac materials as graphene, where an opening of the Dirac points
by a perturbation which does not break time reversal
symmetry nor inversion leads to a topological insulator
of the $Z_2$ topological class.\cite{ti_2}
\subsection{Topological invariant}
In crystals with inversion symmetry,
the non-triviality of the band structure
can be easily checked by calculating the
parities of the Khon-Sham eigenvalues.\cite{fu_kane}
However, the unit cell of
the Ag-Zn compound no longer preserves inversion symmetry,
invalidating the previous approach.
To check that the band structure of this system
is topologically non-trivial, we will instead
take advantage of the tight binding Hamiltonian
obtained previously.
We generate a relativistic
Hamiltonian by taking the Wannier Hamiltonian
and adding SOC
as an atomic term of the form
$H_{SOC} = \lambda_{SOC}\vec L \cdot \vec S$.
This leads to the following
relativistic tight binding Hamiltonian
\begin{multline}
H_0 = H_{W} + H_{SOC} = \\ = \sum_{i,j,s} t_{ij} c_{j,s}^\dagger c_{i,s}
+ \sum_{i,j,s,s'}\lambda_{SOC} \vec L_{i,j} \cdot \vec S_{s,s'} c^\dagger_{j,s} c_{i,s'}
\end{multline}
where $t_{ij}$ are the (spinless) hopping parameters
obtained from the Wannierization technique,
$c_{i,s}$ and $c^\dagger_{j,s'}$ spinful annihilation and creation
operators,
$\vec L$ the orbital
angular momentum in the d-orbitals and $\vec S$ the spin Pauli matrices.
This approach allows to artificially control the strength of the
SOC without having to perform further ab initio calculations.
This is specially useful to understand the mechanism of gap opening
as we will see later.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\columnwidth]{fig2.pdf}
\caption{(Color online.)
(a) Non relativistic band structure of
InZn$_{1/3}$Ag$_{1/3}$V$_{1/3}$O$_3$ calculated at the GGA level, where
the low-energy properties are dominated by the single Ag layer.
We observe that
the Dirac point occurs close to the K point. Panel (b)
shows a comparison of the low-energy bands between DFT and the
tight binding model obtained through Wannierization.
Although the electronic spectrum
is Dirac-like, there is a strong electron-hole asymmetry.
Projected density of states (c) over the Ag atom shows
the dominant d$_{z^2}$ character of the Dirac bands, that
can be understood with the simplified picture (d) considering the three oxygens
in-plane create a local 120$^{\circ}$ rotation symmetry.
}\label{bs_zn}
\label{mono_ag}
\end{center}
\end{figure}
Upon introduction of SOC in the Wannier Hamiltonian
we obtain a band structure showing a band gap (Fig. \ref{model}b),
in agreement with the
fully relativistic ab initio calculations (Fig. \ref{model}a).
With the tight binding model,
the non-triviality of the band gap can be probed
by calculating the $Z_2$ invariant by means
of the flow of Wannier charge centers.\cite{soluyanov2011computing}
Figure \ref{model}c confirms that the system
is non-trivial because an odd number of crossings occurs along the variation of the chosen cyclic parameter, in this case the crystal momentum.
An interesting property of the gap of this system is that it depends
linearly with the SOC strength (Fig. \ref{model}d),
meaning that the channel that opens up
the band gap enters as first order in
SOC.\cite{gmitra2009band}
This is an important
finding. Usually these oxides present a very small band gap since it opens up
via second or third-order perturbations. That is the situation in perovskite
(111) bilayers,\cite{lado} but here we see that the mechanism of gap opening in
principle could allow for larger gaps to appear. It is crucial to understand
what structures are prone to yield larger gaps, since these are required for
applications, but so far these have been quite elusive, at least using oxides
(with the exception of BaBiO$_3$,\cite{yan2013large} which is however a metal
if undoped). Being the band gap produced by SOC strength
mainly, our DFT calculations show that it is quite insensitive to the
exchange-correlation functional used, LDA or GGA yielding a very similar value
close to 18 meV. Comparing with our tight-binding model, we can then extract
the SOC strength by comparing the gap obtained with DFT and
that from the model (see Fig. \ref{model}d), leading to a $\lambda_{SOC}$ of
about 150-170 meV, which is a large value but consistent with the atomic number
of
Ag.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\columnwidth]{fig3.pdf}
\caption{(Color online.)
Ab initio band structure (a) with spin-orbit coupling introduced in a fully relativistic manner by using the {\sc elk} code, showing
a band gap opening close to the Dirac points. Panel (b) shows the band
structure calculated using the tight-binding model obtained through Wannierization plus atomic SOC, which agrees very well
with the one obtained from DFT. With the tight binding model, the flow of the
Wannier charge centers, black dots in (c), is calculated as a function of a
the other crystal momentum, which is used as pumping parameter.
The fact that the Wannier centers cross the dashed green line an odd number
of times proves
the topological non-trivial nature of
the band gap.
Panel (d) shows the evolution of the gap with the SOC strength,
showing a linear dependence and allowing for a determination of $\lambda_{SOC}$ by comparing with the ab initio calculation.
}
\label{model}
\label{mono_soc}
\end{center}
\end{figure}
\subsection{Surface states}
The direct consequence of the non-triviality
of the band structure can be observed in finite geometries,
where
spin-polarized edge states show up.
This can be easily checked with a tight binding model (in our case obtained
from the Wannierization described above) by building up
a finite system and calculating its electronic spectra.
In particular, in a semi-infinite
geometry two branches of edge states appear, which can
be calculated by solving the Dyson equation for each Bloch Hamiltonian
with wavevector parallel to the surface
\begin{equation}
G (k_x,E) = (E - h_0 (k_x) - t(k_x)^\dagger G (k_x,E) t(k_x))^{-1}
\label{sgreen}
\end{equation}
where $h_0(k_x)$ is the k-dependent intracell hopping matrix and
$t(k_x)$ the k-dependent intercell hopping matrix.
From the surface Green function (Eq.\ref{sgreen})
the spectral density can be obtained as
\begin{equation}
\rho(k_x,E) = -\frac{1}{\pi}\text{Tr}[\text{Im} [G(k_x,E+i0^+)]]
\end{equation}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\columnwidth]{fig4.pdf}
\caption{(Color online.)
Surface spectral function (a) in armchair like edge, showing two branches of
gapless edge states plus the gapped bulk band structure. By calculating the spin character of those surface states (panel (b)), it is
observed that they show opposite spin polarization as expected from a quantum spin
Hall insulator due to spin-momentum locking.
When time reversal is broken by a uniform exchange coupling, the
conductance depends on the magnetization direction. For
off-plane exchange $m_z = 2$ meV the edge states remain gapless (c), whereas
for in-plane $m_x=10$ meV they acquire a small gap (d).
}\label{surf_states}
\end{center}
\end{figure}
The previous k-dependent spectral function (see Fig. \ref{surf_states}a)
shows the two surface branches edge states expected
for the semi-infinite geometry of a topological insulator.
This shows the gapless edge states that appear
well separated from the gapped bulk bands. Further insight on the topological
properties of those states
can be achieved by examining its spin polarization, which
can be calculated from the spin spectral
function
\begin{equation}
\rho_z(k_x,E) = -\frac{1}{\pi}\text{Tr}[\text{Im} [S_z G(k_x,E+i0^+)]]
\end{equation}
where $S_z$ is the spin Pauli matrix.
This spin spectral function is shown in Fig. \ref{surf_states}b, and it confirms
the opposite spin polarization of the surface states that is produced by spin-momentum locking, typically obtained in QSHE systems like this.
The spin polarized states can be gapped out by perturbations
that break time reversal symmetry, such as local exchange fields.
That kind of perturbations can arise by interaction with a magnetic substrate,
substitutional magnetic impurities or an external magnetic field.
To understand the electronic properties in this
situation, we study what is the effect on
the surface edge states in the case that time reversal symmetry is broken. Under those circumstances,
the bulk $Z_2$ invariant cannot be defined anymore, but we can gain some
insight by studying
the surface states of the semi-infinite system when
adding a uniform local exchange to the Hamiltonian of the form:
\begin{equation}
H = H_0 + H_Z = H_0 + \sum_{i,s,s'}\vec m \cdot \vec S_{s',s} c_{i,s'}^\dagger c_{i,s}
\end{equation}
where $H_0$ is the Wannier Hamiltonian used previously.
We will consider two cases: when the magnetization is perpendicular to the
Ag hexagonal plane, and when it is parallel. For perpendicular exchange,
the nearly perfect spin polarization of the edge states allows them
to be resilient to a gap opening, still showing
gapless edge states (Fig. \ref{surf_states}c).
When the off-plane exchange field is large enough,
the bulk gap closes leading to a topological
metal, which has gapless edge states that coexist with normal
conduction electrons.
For in-plane exchange fields a gap opens up in the
edge channels\cite{rachel2014giant,lado2014magnetic}
as shown in Fig. \ref{surf_states}d, although
the value of the gap opening is way smaller than the perturbation
in this armchair edge.
The bulk band structure remains gapped even when the in-plane exchange
field becomes
larger than the SOC gap.
Something similar has been reported in oxides
before: VO$_2$/TiO$_2$ multilayers\cite{vpardo_Dirac,sD_banerjee} have been
shown to be Chern insulators\cite{vo2_tio2_vanderbilt} when magnetization is
along the off-plane direction and open a trivial gap when the magnetization
lies inside the plane.\cite{vo2_tio2_mit}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\columnwidth]{fig5.pdf}
\caption{(Color online.) (a) Band structure
of a quantum Hall slab of the Ag-Zn compound, showing the position (a) and
spin flavor (b) of the different eigenstates. The quantum spin Hall states
survive even when time reversal symmetry is broken by a gauge magnetic field,
and turns the system a ferromagnet even without interactions. Panel (c) shows a
sketch of the quantum spin Hall configuration and (d) shows the different
insulating states as function of the zero Landau level
filling, quantum spin Hall at filling 0, and
quantum Hall for filling $\pm 2$.
}
\label{fig:hall}
\end{center}
\end{figure}
\subsection{Quantum Hall effect}
The Dirac-like low-energy spectra suggests that the system will
show an unconventional Landau level (LL) spectrum very much like graphene.
In the absence of SOC, the Dirac dispersion will yield
a set of 4 zero Landau levels: one per spin and valley.
The topological gap is equivalent to a mass in the K and K' points of the Brillouin zone
that follows $m = s_z \kappa $, where $s_z$ labels spin and
$\kappa$ labels valley. When a magnetic field is turned on,
the zero Landau level will develop a splitting following
$\Delta = m \kappa = s_z$, therefore independent on
the valley. The previous splitting is equivalent to the one
obtained in a Dirac ferromagnet\cite{young2012spin,fertig2006luttinger}
but with
the zero LL spin splitting coming from SOC instead of Zeeman,
and automatically realizes
the quantum spin Hall effect.
Using the tight binding model derived before, we build a quantum Hall
bar, where the magnetic field is included by means
of the Peierls substitution $t_{ij} \rightarrow t_{ij}e^{i\phi_{ij}}$,
with $\phi_{ij} = \frac{eB}{\hbar}(x_i -x_j)(y_i+y_j)$ and $x_i,y_j$
the positions of the Wannier charge centers. The Zeeman
term of the magnetic field is neglected provided
that the splittings created by SOC are larger. The band structure obtained
is shown in Fig. \ref{fig:hall}a,b, where the position and spin
are represented by the color of the eigenvalue.
When the system is with the
$d_{z^2}$ manifold half filled, the spin-down zero LL are filled, whereas
the spin-up are empty. In this situation,
the Hall bar is the quantum spin Hall state, characterized
by a spin Chern number $C_S=C_{\uparrow}-C_{\downarrow} = 2$, and
two counter-propagating edge channels (Fig \ref{fig:hall}c). It
is interesting to note that this quantum spin Hall state
is not protected by time reversal symmetry but by conservation of
$S_z$, and that the total Chern number is $C=0$, so that the
system will be vulnerable to perturbations
creating spin mixing. In addition, this QSH state
is adiabatically connected to the one occurring in the absence of SOC.
When the Fermi energy is moved away from half filling (Fig. \ref{fig:hall}d) of the
$d_{z^2}$ states and all
the zero Landau levels are filled, the system no longer realizes the
QSH state. Instead, it enters into a normal quantum Hall state
with $C=2$ and two co-propagating edge states.
\section{Concluding remarks.}
To summarize, in this paper we present ab initio calculations on
the oxide
InZn$_{1/3}$Ag$_{1/3}$V$_{1/3}$O$_3$
based on the structure of the recently synthesized
InCu$_{2/3}$V$_{1/3}$O$_3$.
We have emphasized the Dirac-like low-energy
electronic structure that occurs when magnetism is switched off and
two-dimensionality is enforced.
The low-energy properties, on both sides of the Fermi level, are shown to be
dominated by Ag
$d_{z^2}$ orbitals.
The unexpectedly simple low-energy effective model compares with the rather
complex crystal structure and deep energy electronic mixing.
In the same fashion of other Dirac materials, the inclusion of spin-orbit coupling leads to a topological insulating state, with a band
gap scaling linear with SOC strength. The non-triviality
of the gap was confirmed by the calculation of the $Z_2$ invariant
and edge states spectral functions. The study of this system provides some
clues on how one can design topologicality using oxides, and in particular to
understand how the topological gap can be tuned or even enhanced. We have
compared the results obtained for this system with other oxide structures built
from a honeycomb lattice, which can be even qualitatively very different. We
have also analyzed the material using a tight-binding model obtained through a Wannierization procedure that fits perfectly the DFT
bands around the Fermi level. Our calculations using the tight-binding
Hamiltonian describe the behavior of such system in a quantum Hall experiment,
and show that its properties would be very similar to the response of graphene,
another single-band system (of p$_z$ symmetry in that case) with honeycomb
structure. Thus, this system we propose would be a close d-electron analogue of
graphene.
\acknowledgments
This work was supported by Xunta de Galicia under the Emerxentes Program via the
project no. EM2013/037 and the MINECO via project MAT2013-44673-R. V.P.
acknowledges support from the MINECO of Spain via the Ramon y Cajal program
RyC-2011-09024. J. L. Lado acknowledges financial support
by Marie-Curie-ITN Grant No. 607904-SPINOGRAPH.
\section{Computational procedures}
Ab initio electronic structure calculations based on the density functional
theory (DFT)\cite{hk,ks} have been performed
using two all-electron full
potential codes ({\sc wien2k}\cite{wien} and {\sc
Elk}\cite{elk}) and the pseudopotential-based
Quantum Espresso code.\cite{giannozzi2009quantum}
Unless stated otherwise, structural optimizations and band structure
calculations were performed with GGA-PBE\cite{gga}.
{\sc wien2k} calculations
were performed with a converged k-mesh, a value of
R$_{mt}$K$_{max}$= 7.0, and
spin-orbit coupling was introduced
in a second variational manner using the scalar
relativistic approximation.\cite{singh}
Elk calculations were carried out
with the non-collinear formalism and spin orbit coupling
in order to calculate the
topological gap as well as to check that the Ag based compound is
non magnetic.
Quantum Espresso calculations were carried out using PAW
pseudopotentials.\cite{kresse1999ultrasoft}
The structures were relaxed and the different approaches gave analogous results.
The Wannierization\cite{mostofi2008wannier90,marzari1997maximally,souza2001maximally,marzari2012maximally}
is performed in the non relativistic Quantum Espresso
PAW calculation. The frozen window is chosen
in the interval [-0.9,0.6] eV so that it contains the
low energy dz$^2$ bands, whereas the outer window is chosen
in the interval [-7,0.6] eV around the Fermi energy. The projections
used are the d-manifold of the Ag atoms, giving rise to a $10\times10$
tight binding Hamiltonian. Spin orbit coupling is included
afterwards as an $\vec L \cdot \vec S$ term in the
Wannier tight binding model obtained, turning the Hamiltonian into a
$20\times20$ matrix.
|
\section{Introduction}
\label{Intro}
One of the main goals of computational solid state physics is the simulation
of ``hypothetical'' compounds that have not been synthesized in the
experimental lab. Driven by its remarkable predictive power, Density
functional theory (DFT) is today the most important tool in the field. Together with the high level of sophistication that synthesis
technology has reached, the vision of materials design (i.e. the composition of
functional materials that are tuned for usage in specific devices) seems to
become reality.
Control on the atomic level in materials synthesis e.g. with modern molecular
beam epitaxy and pulsed laser deposition lead to an increasing focus on heterogeneous superstructures and
effects associated with interfaces and surfaces. Especially oxide
superstructures \cite{Mannhart:sci10, Triscone:review2011} attracted lots of attention due to
partially extraordinary physics\cite{Ohtomo:nat04, Mannhart:SCLAOSTO} unknown in the bulk
materials but also adatom lattices or graphene grown on functional substrates
are in the focus of current experimental and theoretical studies.
In our DFT study we concentrate on a particular quality
of functional materials which is largely affected by its surface and crucial
to many applications: the work function $\Phi$. Devices that make use of thermionic electron
emission\cite{Yamamoto:rpp06}, catalytic surface properties\cite{Suntivich:Natc11}, construction of Schottky barriers\cite{Hikita:apl07,
Minohara:prb10,Yajima:natc15}, or the conception of organic electronics
\cite{Greiner:natm12, Greiner:NPGAsia13} are some of the technologies for which knowledge of the work function is
essential. One of the main motivations for our study can be attributed to the very recent conception of so called thermoelectronic
devices \cite{Meir:JRSE13} which rely on the thermionic emission of
an emitter and the subsequent condensation on a collector material. Being, so
to speak, the next evolutionary step following thermionic energy convertors,
the new devices strive for a breakthrough in thermal to electrical energy
conversion. Two main aspects are key for the novel setups: i) stability
towards surface degradation also at elevated temperatures and ii) emitter and
collector materials with work functions tuned to one another. Due to these two
criteria we focus our study on transition metal oxide (TMO) materials: Most TMOs
are thermally very stable and have high melting points. Moreover, we know from an extensive body of
research that TMOs are sensitive to external perturbations (i.e. they are \emph{tunable}) which lead to rich phase
diagrams\cite{Imada:rmp98}.
If we turn to past studies of density functional theory on materials work
functions we find a good amount of research for simple metals
\cite{Skriver:prb92, Methfessel:prb92,Singh-Miller:prb09}, molecular
structures\cite{Rusu:prb06, deBoer:ADVMAT05} and simple oxides like MgO and ZnO
\cite{Giordano:prb05,Giordano:prl08} on metal surfaces, Sc$_2$O$_3$ with adsorbed Ba
\cite{Jacobs:jpcc14}, modified silicon (111) surfaces \cite{Arefi:PhysChem14}, and even
graphene \cite{Giovannetti:PRL08, Young-JunYu:NANOLett09}.
Yet, TMO work functions have been rarely studied and only recently started to attract
attention\cite{Suchitra:jap14, Kumar:jap14}.
In the present study we clarify the sensitivity of TMO work functions with respect to the specific
surface termination, surface relaxation, surface reconstruction, defect
structures (i.e. oxygen vacancies), externally induced surface strain,
electronic interactions on the mean-field level (i.e. inclusion of a Hubbard
U), and most importantly, material trends for a number of perovskite oxides and
superstructures. Our findings will not only serve as a reference for the
presented compounds but especially the observed parameter trends present a first
systematic step towards a broader understanding of how to push a compounds work function to the desired value.
The manuscript is organized in the following way: After reporting details of
our calculation scheme in section \ref{methods} we divide our results in three
sections. In section \ref{resultsA} we report on the sensitivity of the work
function on ``external'' and calculation parameters. While some of the
calculation parameters serve purely as a DFT benchmark (e.g. choice of the
DFT functional or the Hubbard interaction U), others will be quite relevant
for comparison to experimental dependencies (e.g. lattice strain or oxygen
vacancies). In section \ref{resultsB} we explore the material
trends for different ABO$_3$ perovskite oxides. In section
\ref{resultsC} we consider the potential of tuning work functions with
heterostructuring oxide materials.
\section{Background and calculation details}
\label{methods}
The work function is defined as the minimal energy required to remove an
electron from inside the material across its surface into the vacuum.
Conceptually the work function can be devided into a ``bulk'' and ``surface''
dependent part. If there were no charge redistribution at a materials
surface and the vacuum potential would be set to V$_{\rm vac.}=0$, the work function would be
equivalent to $\Phi=V_{\rm vac.}-E_F$ (where $E_F$ is the
materials Fermi energy).
(see e.g. Ref.~\onlinecite{Ashcroft_and_Mermin}).
In reality, however, ionic and electronic charge in the vicinity to the
surface is very different from the
bulk and an additional electric field is generated by the non vanishing dipole
moment of the shifted charge arrangement. It is intuitively clear that
generally such a field, and therefore also the work function, depends on the
specific surface indices and termination. For our studies we
assume clean surfaces with well defined terminations. Let us remark already
here that while there are many materials, in particular simple metals
\cite{Michaelson:jap77}, which show little dependence of the work function on
microscopic details of the surface, TMO work functions are extremely sensitive
to these details.
\begin{figure}[t!]
\includegraphics[width=0.5\textwidth]{Figure1.eps}
\caption{\label{Fig1}
Plane averaged electrostatic potential of a symmetric SrRuO$_3$ slab
consisting of six SrO layers, five RuO$_2$ layers, and a 20 \AA vacuum. The
electrostatic potential is defined with respect to Fermi energy $E_F$, and
converges to a constant value in the vacuum region that is the work function
$\Phi$ of SrRuO$_3$ with SrO surface termination. $\Phi$ therefore indicates
the required energy to remove an electron at Fermi level from the material to
a state at rest in the vacuum nearby the surface.}
\end{figure}
In this study we have selected a number of perovskite transition metal oxides
ABO$_3$ (where A=Ca, Sr, Ba; B=3d, Ti-Co; 4$d$: Zr-Rh) and their
heterostructures. The ABO$_3$ structure can be viewed as simple cubic lattice
of A atoms with a body centered B atom and oxygen atoms in the face centers.
In the following we consider surfaces along the (001) direction which are the
most commonly studied surfaces in thin film or heterostructure compounds. In
this direction the crystal is build up by an alternating stacking of AO and
BO$_2$ layers \citep{Ohtomo:nat04}. The (001) ABO$_3$ then have either AO or
BO$_2$ surface terminations, and thus two intrinsically different work
functions. For the calculation of the work function within a DFT framework we
employ a symmetric slab geometry which is sketched in the top part of
Fig.~\ref{Fig1} for the example of SrRuO$_3$. The SrO terminated SrRuO$_3$
slab consists of six SrO layers, five RuO$_2$ layers, and a vacuum thickness
of 20\AA; for BO$_2$ terminated surface (not shown in Fig.~\ref{Fig1}) we add
an additional BO$_2$ layer on each side of the slab. The work function $\Phi$
is then calculated as the energy difference between the plane averaged
electrostatic potential (excluding the exchange correlation part) of the slab
in the vacuum region and the Fermi level as can be seen in Fig.~\ref{Fig1};
Benchmark calculations confirmed that $\Phi$ is converged with this setup.
For further information on the calculation of the averaged electrostatic
potential we refer to \onlinecite{Giovannetti:PRL08} and \onlinecite{Rusu:prb06}.
Since we are mostly interested in materials and superstructures grown on
substrates, the in-plane lattice constant was fixed to a=3.905\AA which is that
of an assumed undistorted cubic SrTiO$_3$ substrate (it is also
very close to the bulk lattice constant a=3.923\AA of cubic SrRuO$_3$
\cite{Samata2009623}). For all calculations the internal atomic positions were
relaxed. The calculations were performed with the VASP (Vienna ab initio
simulation package) code \citep{Kresse:prb99} using the generalized gradient
approximation GGA -PBE functional \citep{PerdewPRL96} for electronic exchange
and correlation and a 16$\times$16$\times$1 k-point grid including the
$\Gamma$ point. In a set of selected benchmark calculations we compare the GGA
results also to those obtained by a local-density approximation (LDA)
functional\citep{Perdew:prb81}. While the latter one generally leads to
somewhat larger values of the work function, our main conclusions about materials trends
and sensitivities remain unchanged by the choice of the functional.
Let us also explicitly mention that this study is not concerned with the
temperature dependence of the work functions. While motivated by applications
and devices which operate at elevated temperatures the non-trivial
inclusion of finite temperatures in DFT calculations is beyond the scope of
this study.
\section{Results A: External parameters}
\label{resultsA}
\begin{table*}[]
\begin{ruledtabular}
\caption{Work function of SrRuO$_3$ and SrTiO$_3$ with either SrO or
(Ru,Ti)O$_2$ surface terminations. The table summarizes dependencies of the
work function for i) different choices of the DFT functional (columns 3 \& 4
), ii) an unrelaxed lattice (column 5), iii) compressive (a=3.80\AA) or
tensile (a=4.00 \AA) strained substrate (columns 6 \& 7), iv) ferromagnetic
ground state of SrRuO$_3$ with an on-site Coulomb repulsion U=2.0eV (column
8), v) a monolayer film; the low temperature orthorhombic structure of
SrRuO$_3$; SrTiO$_3$ with a TiO$_2$ 2$\times$1 surface reconstruction
(columns 9,10,11), vi) with surface oxygen vacancies in top layer/subsurface
layer (column 12), vii) values observed in experiment}
\label{Tableone}
\begin{tabular}{cl|ll|l|ll|l|lll|l|l}
\multicolumn{1}{l}{} \textbf{$\Phi$ in eV} & term. & LDA & GGA & unrelaxed & \textit{a}=3.80 & \textit{a}=4.00 & U=2eV & monolayer & orthorhombic & reconstruct & Ov & exp \\
\hline
\multirow{2}{*}{SrRuO$_3$} & SrO & 2.80 & \textbf{2.39} & 1.30 & 2.00 & 2.55 & 2.37 & 2.60 & 2.29 & - & 2.05/2.39 & \multirow{2}{*}{5.2 \footnotemark[1]} \\
& RuO$_2$ & 5.01 & \textbf{4.88} & 3.90 & 5.54 & 4.92 & 5.33 & 4,95 & 5.05 & - & 5.03/4.91 & \\
\multirow{2}{*}{SrTiO$_3$} & SrO & 2.52 & \textbf{1.92} & 0.82 & 1.69 & 2.04 & - & 2,02 & - & - & 2.26/1.33 & 2.4 \footnotemark[2] \\
& TiO$_2$ & 4.67 & \textbf{4.48} & 3.70 & 4.47 & 4.51 & - & 4.18 & - & 6.18 & 3.39/3.86 & 4.6 \footnotemark[2]
\end{tabular}
\end{ruledtabular}
\footnotetext[1]{Fang \emph{et al.} Ref. \onlinecite{Fang:99} with unknown surface termination}
\footnotetext[2]{Susaki \emph{et al.} Ref. \onlinecite{Susaki:prb11}}
\end{table*}
Part of our first set of calculations in which we identify key parameters that
alter the work function can be considered as DFT benchmarks. Obviously, if
the parameter in question is experimentally accessible (like e.g. substrate
strain), one can deduce potential tuning parameters of $\Phi$.
The results we report in this section are obtained for SrRuO$_3$ and
SrTiO$_3$. Both materials are well studied and experimental data for their
work functions are available \cite{Fang:99,Susaki:prb11}. SrRuO$_3$ is a 4$d$
system and a ferromagnetic metal \cite{Koster:rmp12}; SrTiO$_3$ on the other hand is a 3$d$ non-magnetic
insulator. At this point we should make some more specific remarks about how
we deal with the calculation of work functions for the insulating SrTiO$_3$. The
difficulty for insulators is the uncertainty of the Fermi energy which needs
to be subtracted from the vacuum potential to yield $\Phi$. Instead, we decided to
consider the bottom of the conduction band as the Fermi energy due to a simple
and pragmatic argument: Our choice corresponds to the electron doped
version of the material which can be realized in experiment by La substituting Sr
\cite{Ohta:jap05}, or by Nb substituting Ti. The later technique was used in a work function study for SrTiO$_3$ by Susaki \emph{et
al.}\cite{Susaki:prb11} and as one can see in Fig.\ref{Fig1} thei calculated
values based on our definitions are in satisfactory agreement with the experimental observations.
Moreover, even without active doping, the (very common) occurrence of oxygen
vacancies in TMO surfaces effectively lead to the same kind of doped
electronic structure. Let us anticipate already here that our
calculations, which include such oxygen vacancies explicitly, do not capture effects
from an insulator to metal transition but rather from very small to very
large concentration of oxygen vacancies.
For these cases of slightly doped insulators, where the work function might rely
sensitively on the size of the gap between conduction and valence band, we
also make sure that DFT-GGA, which is known to underestimate gap sizes and the
Heyd-Scuseria-Ernzerhof (HSE) hybrid functionals \cite{Heyd:jcp03,
Silva:prb07} (known to yield better results for band gaps) yield consistent
results for $\Phi$. We summarize our parameter study in
Table~\ref{Tableone}. Here, we report values for both materials and consider
either a SrO terminated or a (Ru,Ti)O$_2$ terminated (001) surface.
The values of $\Phi$ in the first column of Table~\ref{Tableone} obtained with plain GGA
for relaxed slabs already show an extremely important effect that we observe
in basically all calculations we have performed: Different from simple metals like
tungsten or silver, where the work function shows a surface dependence on the
order of hundreds of meV \cite{Michaelson:jap77}, the work function for perovskite oxides shows a much
more severe modulation with the choice of a specific surface, e.g. if it is AO
or BO$_2$ terminated. From our calculations we observe a difference of
$\Phi_{\rm BO_2}-\Phi_{\rm AO}=2.49$eV ($2.56$eV) for SrRuO$_3$ (SrTiO$_3$)
which prohibits clearly an approximation of $\Phi$ by a single $\Phi_{\rm
ABO_3}$ value for oxide materials and sets the challenge for a theory/experiment
comparison: control of the sample on sub-unit cell scale seems necessary in
synthesis to support/falsify predictions from computer simulations.
If such control is not possible, the samples might have mixed termination and,
hence, display a strong sensitivity of the work function to details of the
sample preparation. Such difficulties might be one of the reasons that as of
yet there are only few experimental studies on TMO work functions\cite{Fang:99,Susaki:prb11,Greiner:afm12}. Moreover,
these complications also affect the conception of interface devices like,
e.g., TMO Schottky barriers at metal/semiconductor interface where the barrier
height is calculated with the work function of the metal \cite{Hikita:apl07}.
We will return to the discussion of termination dependent work function in the
context of building up superstructures (see section \ref{resultsC}).
Let us turn to the comparison of GGA with LDA results. Our test cases are
actually well in line with an extensive study of Singh-Miller and
Marzari\cite{Singh-Miller:prb09}, where the functional dependence of DFT
workfunctions for metallic surfaces is discussed. The differences between GGA
and LDA can be attributed on the one hand to differences in the relaxed
structure (since GGA, e.g. generally overestimates bondlengths when compared
to experiment). On the other hand, when performed for identical lattices LDA tends
to yield always somewhat larger values than GGA. As can be seen in
Table~\ref{Tableone} we observe total differences between $\approx
0.13-0.50$eV. While these differences are surely non negligible and one should
be aware of possible error bars. Relative values and materials trends,
however, are not affected.
More crucial than the choice of the particular DFT functional is, however, the relaxation
of the atomic positions in the unit cell near the surface with respect to an
unrelaxed surface. Work functions calculated with unrelaxed surfaces differ partially more than $1.0$eV from the
relaxed calculations. In our calculations the surface relaxation always
increases the work function which means that the surface dipole-field increases.
It is tempting to attribute this increase just to a surface buckling that features
an outward shift of oxygen ions at the surface (stronger for AO than for
BO$_2$ terminated surfaces). On quantitative levels a purely ionic picture is,
however, misleading since it disregards effects of relaxation of the
electronic charge involving interlayer charge transfer. The sensitivity that
we observe here indicates already the strong dependence of the work function
on microscopic details of the surface as can be seen also in the calculations
for either compressive ($a=3.80$\AA) or tensile ($a=4.00$\AA) strain which can
be achieved with growing the material on specifically chosen substrates. It
turns out that in this way modification of the work function can be achieved over a range of
up to $\approx 0.7$eV. In the SrO terminated compounds we always find a
decrease (increase) of the work function upon compressive (tensile) strain. In
the BO$_2$ terminated compounds we see a clear difference between the metallic
RuO$_2$ layer which shows an increase of the work function upon either
compressive or tensile strain, while the TiO$_2$ terminated systems is
affected by the pressure only on a very small scale compared to the other
cases.
Next we turn to the question whether a Hubbard U interaction parameter on the
B atom d-shell treated on the mean-field level has impact on $\Phi$. Such additional local potential will only
have impact in cases of partially filled shells which is why an interaction
$U=2.0$eV was taken into account only for the ruthenate calculation. We have
carried out the GGA+U calculation \cite{Dudarev:prb57} where we allowed for a
ferromagnetic symmetry broken ground state. It does not come as a big surprise
that the RuO$_2$ terminated surface is more influenced by the U on the Ru
d-shell which results in a work function enhancement by $\approx 0.4$eV while
the AO terminated surface is basically unaffected. For the general case,
however, please note that electronic interaction/correlation (approximated on
the Hartree level or beyond) might trigger phase transitions that result in a
charge redistribution, e.g. surface charge-ordered states
\cite{Hansmann:prl13} which might have significant impact on the surface
dipole field and, hence, its work function. Let us briefly point out that we
did not consider a GGA+U calculation for the band-insulating SrTiO$_3$ with a
practically empty d-shell. U would simply enlarge the gap by pushing
up empty states. Since not much is gained by such a
manual gap renormalization we state that the most reasonable step would
rather be a GW calculations without adjustable parameters which, however, is
beyond the scope of our current study.
The following three calculations consider again more structural effects: In
monolayer setups effects of quantum confinement can alter the electronic
structure \cite{Yoshimat:sci11, zhong:prb13}. Also we remark that for most
ABO$_3$ materials, the low temperature structure is not cubic but often shows
orthorhombic distortions with tilted and rotated BO$_6$ octahedron, which,
however, do not alter $\Phi$ dramatically. Moreover, we argue that
simulations for $\Phi$ should rather consider the materials structure at the operation
temperature of the hypothetical device. Also, depending on the temperature, we take
into account that for real oxide surfaces, various surface reconstructions
exist (for example of SrTiO$_3$ \cite{Kawasaki:sci94, Koster:apl98,
Erdman:nat02,Herger:prl98}). It is reasonable to assume that the different
structure of bonds and hence electronic densities in reconstructed surfaces
will lead to specific dipole fields and, hence, altered work functions. We
confirmed this hypothesis by studying a well-established so called
(2$\times$1) surface reconstructed phase of SrTiO$_3$ \cite{Erdman:nat02},
which can be viewed as a double TiO$_2$ layer. It turns out that the
reconstruction has a great influence on the work function which is, with a
value of $6.18$eV, much higher than the $4.48$eV of the bare TiO$_2$ surface.
Finally, it is a well known issue for oxide surfaces that defects in the form
of oxygen vacancies should not be disregarded \cite{Nakagawa:Natm06,SusakiMgO}.
While there is a certain amount of control over oxygen vacancies in synthesis
(e.g. adjusting the oxygen pressure and annealing) the exact concentration and
distribution is generally unknown and hard to pin down. Such
defects pose a real challenge to comparing different experiments, but also experiment to
an electronic structure calculation of the oxide surfaces. The best one can do
in a calculation is to assume periodic vacancies in supercells. In our case we
assume a $2\times 2$ supercell and introduce for each case considered an
oxygen vacancy in the surface or the first subsurface layer. With this setup
we actually assume a quite high concentration of oxygen vacancies so that our
results for $\Phi$ might be considered as an upper bound of the O vacancy
effect.
As conclusion of this section stands a classification of external parameters
by means of their impact on a materials work function. The first and most
important message is that for transition metal oxide work functions the
microscopic structure of the surface electronic states/density \emph{does}
matter crucially. While our analysis underlines the challenging (but nowadays
feasible) necessity of
experimental control on the atomic scale it also tells us that a materials
\emph{work function can be tuned} with a number of external
parameters. While magnetism (in the tested cases), interaction effects or
even ``quantum confinement'' effects are not major ($\Phi$ converges rather
quickly as a function of thickness), clean terminations and control of
surface reconstructions is absolutely mandatory. The latter parameters can
tune the work function on the scale of electron volts. On a smaller scale
($\approx 0.5$eV) the work function can be modified, i.e. fine tuned, by
exerting control on the oxygen defect structure and/or the choice of
substrate. With these results in mind we will now turn to another type of
``control parameter'': The choice of alkali earth cation A cation and
transition metal element B.
\section{Results B: ABO$_3$ material trends}
\label{resultsB}
As mentioned before the results in this section were obtained from setups
with a fixed in-plane lattice constant of a=3.905\AA which corresponds to
growth on a SrTiO$_3$ substrate. To disentangle trends originating in the
specific choice of cation (A)
and TM (B) from other parameters (see previous section) we consider
defect free, relaxed structures with well defined terminations in a GGA
slab calculation. The results are reported in Fig. \ref{Fig2} and
\ref{Fig3} (and corresponding data tables in appendix A).
\subsection{Termination dependence in different materials}
Overall we find as a first remarkable fact a confirmation of the
crucial dependence of $\Phi$ on the termination, see Fig. \ref{Fig2} and
\ref{Fig3}. Except for a single case (CaZrO$_3$) the work function of the AO
terminated surface is smaller than the BO$_2$ terminated one.
It turns out that this observed materials dependence hints towards a new twist to the
interpretation of the termination sensitivity: We remind ourselves that the
surface dependence of the work function originates in the dipole field created
by polarization of the electronic charge and shifts of atoms/ions close to the
specific surface. Two effects which are obviously coupled in our calculations which
include a self consistent lattice relaxation. While this interplay is quite
involved and cannot be disentangled easily, we observe a clear
correlation between the work function behavior and the
\emph{electronegativity $\chi$} of the cation and transition metal elements \cite{Allen:jacs89}.
The concept of electronegativity is usually used in order to estimate the
character of an ionic bond in a binary compound. Taking the difference of the
two elements electronegativity allows for classification of the bond into
either ionic or covalent. So \emph{$\chi$ reflects the ability of an atom to attract
electron density}. Keeping this in mind it comes not as a big surprise that for
simple metals the compounds work function is linked to the elements
electronegativity\cite{Trasatti}. Our results and analysis show that, remarkably we can
still use the electronegativity concept in our ternary compounds.
Following the Allen scheme\cite{Allen:jacs89} of electronegativity we find a
monotonous increase in $\chi_B$ from Ti ($\chi_{\rm Ti}=1.38$) to Co
($\chi_{\rm Co}=1.84$) for the 3d series, a monotonous increase from
Zr($\chi_{\rm Zr}=1.32$) to Rh ($\chi_{\rm Rh}=1.56$) for the 4d series, and
for the cations we have $\chi_{\rm Ca}=1.03$, $\chi_{\rm Sr}=0.96$, and
$\chi_{\rm Ba}=0.881$ \footnote{For our qualitative discussion the specific
type of the electronegativity (e.g. Pauling, Muliken, Allan, etc.) is not
really important since materials trends are very similar for all schemes.}
Oxides are typically considered as very ionic due to the high electronegativity
of oxygen $\chi_{\rm O}=3.61$.
For our materials it turns out to be useful to introduce the idea of a layer
electronegativity $\chi_{AO}$ and $\chi_{BO_2}$. In all considered compounds
$\chi$ of cation A is smaller than that of transition metal B so that on the
one hand A-O bonds can be considered as more ionic than B-O bonds but also
that the average AO electronegativity is smaller than that of
the BO$_2$ layers $\chi_{AO}<\chi_{BO_2}$. This explains the general tendency
of smaller AO work functions $\Phi_{AO}<\Phi_{BO_2}$ which we already reported.
It turns out that with these rough estimates many of the following materials
trends can be explained.
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{Figure2.eps}
\caption{\label{Fig2} DFT calculated work functions of the perovskite
ABO$_3$ series with A=Sr and \textit{B} being an element of the 3d (red)
or 4$d$ (blue) period, and with AO (circle with solid lines) and
BO$_2$ (square with dashed lines) surface termination. Experimental work
function of SrTiO$_3$ with SrO and TiO$_2$ termination by Susaki
\emph{et al.} \onlinecite{Susaki:prb11}
as well as SrRuO$_3$ with unknown termination by Fang \emph{et al.}
\onlinecite{Fang:99} are shown as stars.}
\end{figure}
\subsection{Control via transition metal B for A=Sr}
We will now discuss tuning of the work function with choice of the transition
metal element B in more detail. As shown in Fig.~\ref{Fig2} we have performed
calculations for a series of 3$d$ and 4$d$ transition metal compounds. The overall
variation of the work function is a remarkable $6$eV ($\approx \Phi^{\rm
BO_2}_{\text{SrCoO}_3}-\Phi^{\rm AO}_{\text{SrZrO}_3}$ ). Continuing the line of
argument from above, the trends we observe can be explained by comparison of
$\chi_B$ of the transition metal element (or the effective BO$_2$
electronegativity $\chi_{BO_2}$). It is no surprise that the
trend is weaker in the case of AO termination since increased $\chi_{BO_2}$
in the subsurface layer has less effect on $\Phi$. The increasing $\Phi$ within
each series (3$d$,4$d$) reflects precisely the increase of $\chi_{BO_2}$ within
the period and so does the decrease of $\Phi_{3d} < \Phi_{4d}$ reflect
$\chi_{3d} < \chi_{4d}$. A closer look at the numbers shows indeed quite low
work functions for AO termination throughout the series and an almost
monotonous increase from $\Phi_{\text{SrTiO}_3}\approx 1.9$eV (3$d$ series)
and $\Phi_{\text{SrZrO}_3}\approx 1.1$eV (4$d$ series) \footnote{Please note
that SrTiO$_3$ as well as SrZrO$_3$ are band insulators such that we took
the position of the lowest conduction band as reference energy for the work function.}
to $\Phi_{\text{SrCoO}_3}\approx 3.2$eV and
$\Phi_{\text{SrRhO}_3}\approx 3.0$eV with the only exception of $\Phi_{\text{SrNbO}_3}$
being slightly higher than $\Phi_{\text{SrMoO}_3}$.
We find the same trend in the BO$_2$ terminated surfaces of these materials
though at values for $\Phi$ which are roughly larger by a factor of 2.
Before we continue our discussion for cation controlled tuning of $\Phi$ let
us compare the results to the few experimentally available data points (shown
as ``stars'' in Fig.~\ref{Fig2}). For
SrTiO$_3$ both values for TiO$_2$ or SrO terminated surfaces \cite{Susaki:prb11} are in very
satisfactory agreement with our calculations. (The comparison shows further
that due to the very likely presence of oxygen vacancies it is a reasonable
ansatz to take the energy of the lowest (electron doped) conduction band as
reference energy for $\Phi$). The only other data point is that of
SrRuO$_3$ for which, however, the precise termination was undetermined\cite{Fang:99}. In
fact one is tempted to conclude by comparison to our results that the termination was
most likely a RuO$_2$ dominated one. Yet, one has to be careful since rough
surfaces are most probably determined not only by a mixture of AO and BO$_2$
domains but also by polarizability of defects like domain walls etc.. We use
this observation to emphasize once more the importance of microscopic control
of the surface structure if comparison or predictions of calculations like
ours should be considered. Moreover, if oxide heterostructures are used in
devices like Schottky barriers \cite{Hikita:apl07,
Minohara:prb10,Yajima:natc15} or for organic electronics
\cite{Greiner:natm12, Greiner:NPGAsia13} the oxide work function is often used
as for simple metals or semiconductors. We emphasize once more, however, that this is very
dangerous since \emph{there is no such thing as a single valued work function for
an oxide material}.
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{Figure3.eps}
\caption{\label{Fig3} DFT calculated work functions of the ATiO$_3$
(red), AZrO$_3$ (black), and ARuO$_3$ (blue) series with A=Ca, Sr and
Ba, and with AO (circle with solid lines) and BO$_2$ (square with dashed
lines) surface termination.}
\end{figure}
\subsection{Control via cation A}
We now turn to control of the work function with the choice of the alkaline
earth cation A. In Fig. \ref{Fig3} we report results obtained for nine
selected compounds (Ca,Sr,Ba)(Ti,Zr,Ru)O$_3$ to study trends with the cation
choice. As a first observation we state that the overall dependence of $\Phi$ is
quite large and that the cation choice is apparently a promising tuning
parameter. The changes in the AO terminated surfaces are more sizable than before and
follow the trend of electronegativity of the cation: Ba has the smallest
$\chi$, and hence the smallest work functions.
However, the trend in BO$_2$ terminated surfaces with cation A is not as
easily explained! While $\chi_A$ decreases from Ca to Ba, the work function for
BO$_2$ terminations actually increases. This points towards an aspect which is
not taken into account by our simple electronegativity argument: \emph{Charge
transfer} between AO and BO$_2$ layers. One possible explanation is that more
ionic AO layers lead to an increased charge transfer to the terminating BO$_2$ layer
which then increases the dipole field and, hence, the work function. On the
other hand we cannot underline this hypothesis with evidence and remain with
reporting the observed trend.
\subsection{Promising materials}
At the end of our materials study we can confirm experimental strategies with our calculations
that have been established on an empirical basis in the past. The
electronegativities $\chi_A$ and $\chi_B$ might be used (keeping the
limitations of the estimate in mind) as a rough guidance to select promising
materials. Our considered materials cover a wide range of work functions
reaching from $6.91$eV (CoO$_2$ terminated SrCoO$_3$) down to $0.82$eV
(BaO terminated BaZrO$_3$). For electron emitting devices and, more specifically,
emitter and collector materials in thermoelectronic setups one needs low work functions.
These are found in the compounds of the early elements in the 3$d$ and 4$d$
series: First of all we point out SrO terminated SrMoO$_3$ ($\Phi^{\rm
SrO}_{\rm SrMoO_3}$)which is the material with the highest conductivity in
our 4$d$ series. It can be grown in thin films by pulsed laser deposition\cite{Radetinac:ape10}
and would be a good candidate, e.g. for the electrode material in
thermoelectronic devices. Another very interesting compound we would like to
highlight is CaZrO$_3$. It is remarkable due to the fact that the work
function is at the same time quite low \emph{and} rather similar for both
terminations (see Fig.\ref{Fig3}). This means, that also less clean
(heterogeneous) CaZrO$_3$ (001) surfaces, which are much less
tedious/expensive to synthesize, will be good electrode materials. Third, our
results for BaBO$_3$ are in agreement with the well known observation that
coverage with BaO will lower a materials work function \emph{if AO terminated
systems are concerned}. This last conclusion leads us to our third and last
section where we discuss how to tune the work function by building
heterostructures.
\begin{figure}[h,t]
\includegraphics[width=0.5\textwidth]{Figure4.eps}
\caption{\label{Fig4} DFT calculated density of states for
SrVO$_3$ capped with one unit cell
of SrTiO$_3$ (upper panel) and SrNbO$_3$ capped with
one unit cell of SrTiO$_3$ (lower
panel). We show only the partial density of states for V, Nb, and Ti
3d-states.$t_{2g}$ and $e_g$ states are indicated by solid and dashed lines respectively. The figure shows a clear material dependence of the charge
transfer between base and capping material.}
\end{figure}
\section{Results C: Oxide heterostructures}
\label{resultsC}
In the previous section we have already made several observations how (and
gave arguments why) changes in surface layer and charge
transfer between layers close to the surface affect the work function.
In the third and final part of our study we turn to even more drastic
surface manipulation: instead of just choosing one of the materials lattice
planes to be the surface layer we build the surface actively by combining
different compounds.
To this end we model symmetric A'B'O$_3$/ABO$_3$ heterostructures by adding
A'B'O$_3$ thin films (grown in the 001 direction) with varying number of unit
cells N on both sides of the ABO$_3$ slab (i.e. capping) with either AO or
BO$_2$ termination. To be more specific, an AO terminated material capped with
A'B'O$_3$ will have an A'O interface with vacuum while a BO$_2$ terminated one
will have a B'O$_2$ interface with vacuum.
\begin{table}[h,t]
\begin{ruledtabular}
\caption{Work function of SrVO$_3$, SrNbO$_3$ and SrRuO$_3$ capped by SrTiO$_3$ thin films. The thickness of SrTiO$_3$ N is varied from 0 to 2}
\label{Tabletwo}
\begin{tabular}{lllllll}
SrTiO$_{3}$ & \multicolumn{2}{l}{SrVO$_3$} & \multicolumn{2}{l}{SrNbO$_3$} & \multicolumn{2}{l}{SrRuO$_3$} \\
& SrO & VO$_2$ & SrO & NbO$_2$ & SrO & RuO$_2$ \\
\hline
N=0 & 1.81 & 4.38 & 1.66 & 2.07 & 2.39 & 4.88 \\
N=1 & 1.88 & 4.61 & \textbf{1.18} & 3.41 & 2.17 & 4.72 \\
N=2 & 1.87 & 4.77 & 1.29 & 3.82 & 2.13 & 4.87
\end{tabular}
\end{ruledtabular}
\end{table}
\begin{figure}[h,t]
\includegraphics[width=0.5\textwidth]{Figure5.eps}
\caption{\label{Fig5} Plane averaged electrostatic potential for a
LaAlO$_3$/SrRuO$_3$ heterostructure calculated with DFT: three layers of
LaAlO$_3$ grown on a SrRuO$_{3}$ substrate with AO termination (upper
panel) or BO$_2$ termination (lower panel). The internal field of the
polar layers tunes the work function depending on its direction.}
\end{figure}
\subsection{Non-polar capping:}
In table \ref{Tabletwo} we present results for capping three example materials with
SrTiO$_3$ of varying thickness of either one or two unit cells (as well as the
reference N=0 for the uncapped material). SrTiO$_3$ is a charge neutral and
non-polar band insulator. The results depend strongly on the specific
case. The most severe change is found in SrNbO$_3$ capped with one unit cell
of SrTiO$_3$. Here the work function is reduced for the SrO termination from
$1.66$eV to $1.18$eV while the NbO$_2$ terminated (i.e. TiO$_2$ as final
layer) compound experiences an increase of the work function from $2.07$eV to
$3.41$eV. By comparison changes in SrVO$_3$ or SrRuO$_3$ are less
pronounced. The reason for this different behavior is found when we study the
charge transfer between base and capping material. In Fig.~\ref{Fig4} we show
the partial density of states for SrTiO$_3$/SrVO$_3$ (upper panel) and
SrTiO$_3$/SrNbO$_3$ (lower panel).
The local potential of Ti d orbitals in SrTiO$_3$ turns out to be comparable
to the Nb d potential in SrNbO$_3$. V d orbitals in SrVO$_3$, however, reside at
a much lower energy \cite{zhong:epl12}. As a consequence we encounter a
substantial charge transfer in SrTiO$_3$/SrNbO$_3$ (as seen in
Fig.\ref{Fig4}), while basically no charge transfer occurs in SrTiO$_3$/SrVO$_3$.
As one can from the lower panel the Ti d-states in the SrTiO$_3$/SrNbO$_3$
system are heavily electron doped by the base material while this is not the
case for the vanadate case. Such doping has severe consequences not only by
shifting the Fermi level into the Ti d-states but also by actually charging
the capping layer which alters the surface dipole field. It is interesting how
such complex interplay of lattice, orbital and charge degrees of freedom
eventually leads to a work function for the AO terminated surface which is
lower than that of the parent compounds SrTiO$_3$ or SrNbO$_3$. This case
confirms, as a proof of principle, the potential of tuning $\Phi$ by
heterostructuring oxide materials.
One should remark that in the example we have just discussed the capping
material was an insulator. If we choose A'B'O$_3$ to be a metal, we see that the
work function is dominated by the A'B'O$_3$ thin films (so that the capped
ABO$_3$ material is almost entirely irrelevant). This is simply due to the
screening properties of the metallic capping compound and we could only study
the strain and quantum confinement effect in A'B'O$_3$.
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{Figure6.eps}
\caption{\label{Fig6} Work function plotted versus thickness of polar
capping material. As already indicated in Fig.~\ref{Fig5} the additional
field produced by polar layers can be used to tune the work function in
either way. Here we show how either termination can be tuned up or down
by choosing the appropriate capping material LaAlO$_3$ or KTaO$_3$ which
have opposite effects due to their opposite polarity.}
\end{figure}
\subsection{Inducing intrinsic fields by polar capping}
When A'B'O$_3$ is a polar insulator, such as LaAlO$_3$, which can be viewed as
an alternating stack of positively charged (LaO)$^+$ layers and negatively
charged AlO$_{2}^{-}$ layers. When it is grown on a non-polar compound like
SrRuO$_3$ or SrTiO$_3$ (charge neutral (SrO)$^0$ and (Ru,Ti)O$_{2}^{0}$
layers), A possible hypothesis that was studied in the past
\cite{Susaki:prb11} is that thin films of a polar material (e.g. LaAlO$_3$)
play the role of a parallel capacitor that introduces an internal electric
field pointing from surface, e.g. LaO, layer to the interface of
e.g. SrO/AlO2. The resulting potential drop should lead, as the polar capping
material thickness increases, to a decrease of the work function. We sketch
this scenario in Fig.~\ref{Fig5}: the result for LaAlO$_3$ (KTaO$_3$) capping
would be a decrease (increase) of the work function for the AO and an increase
(decrease) for the BO$_2$ terminated case. We performed calculations for
SrRuO$_3$ or SrTiO$_3$ for LaAlO$_3$ and KTaO$_3$ capping of different
thickness and present the results in Fig.~\ref{Fig6}. Let us first remark that
the comparison between N=0 and N=1 is always delicate since we
introduce not only the additional field but also the surface relaxation changes most
from the uncapped to the N=1 case. The data from N=1 to N=3 shows that the DFT
calculation confirms the simple hypothesis for all checked cases: LAO capped
AO surface layers show decreasing $\Phi$ with increasing N while LAO capped
BO$_2$ terminated cases show an increase. The opposite is true for capping
with KTaO$_3$. In principle this observation is encouraging. However, these
observations are in contradiction to experiments performed by Susaki et
al.~\cite{Susaki:prb11}. In their article the authors already mention the
discrepancy of the simple ``additional internal field'' picture with their
measured data: Instead of a work function increase they found a remarkable
decrease in TiO$_2$ terminated SrTiO$_3$ capped with LaAlO$_3$.
In summary we report that DFT results agree with the simple picture of
superimposed potentials but not with experiment. A reasonable explanation for
the discrepancy might be found in defect structures not taken into account by
our calculations. In section~\ref{resultsA} we have seen that oxygen vacancies
can alter the work function on a significant scale so that a disregard of
likely defects in the LaAlO$_3$ capped systems is not justified.
\section{Conclusion:}
In conclusion we have presented a systematic work function study for transition
metal compounds which allowed us to classify the sensitivity of the work function
to parameters of DFT calculations as well as to experimentally accessible
conditions like oxygen vacancies, substrate strain, and heterostructuring.
We were able to conclude, on general grounds, that the work function concept is
more complex in oxide materials than in simple metals and that microscopic
details at the materials surface matter crucially. We emphasize that tuning oxide work functions
(predictably) in the experimental lab generally requires synthesis control on
the atomic scale. This challenge comes, however, with the prize that oxide
work functions are indeed highly tunable and, depending on the
parameter can be manipulated on different energy scales. While the choice of
the material and the choice of the terminating layer tunes oxide work
functions over several eV, substrate induced strain or a suitable capping
material can fine tune the work function on the sub-eV scale to the desired
value. We have also uncovered that, like for simple metals, there is a link
between the observed work function $\Phi$ and electronegativities $\chi_A$ and $\chi_B$ of
the elements in the compound which was helpful in explaining the observed
materials trends and might prove to be also useful for extrapolating our
results in other directions. For the manipulations of $\Phi$ with polar
capping layers we conclude from our theory/experiment comparison with a
warning that for real materials an oversimplified electrostatic picture is
highly doubtful.
Let us finally remark that one of the main intentions of this manuscript is to form a
fix point for future studies. The ``phase space'' of materials and tuning
parameters is infinitely large so that any systematic search needs a well
established base. We hope that also experimental colleagues can help to judge
the quality of the many data points we predicted for materials that have not yet
been measured in order to establish such a base.
\section{Acknowledgments}
We thank J. Mannhart, I. Rastegar, G. Giovannetti, N. Spaldin, and L. Giordano for motivation and useful discussions.
|
\section{}
|
\section{Introduction}
The notion of \emph{random rooted graph} (as described by D.~Aldous and
R.~Lyons in~\cite{AL}) arises in substance from the paper \cite{LPP} of
R.~Lyons, R.~Pemantle, and Y.~Peres on Galton-Watson trees. Bernoulli bond
percolation on a Cayley graph $\mathbb{G}$ provides the basic example of random rooted
graph, which is obtained by keeping each edge with constant probability $p$
independently to other edges, see Figure~\ref{fig:cluster}. This kind of random
graphs (and other obtained by bond percolation on Cayley graphs or unimodular
transitive graphs) enjoy the important property of \emph{insertion-tolerance}
introduced by R. Lyons and O. Schramm in Definition 3.2 of \cite{LS}: the
measure of a nonnull Borel set after adding an edge is nonnull.
\medskip
\begin{figure}
\centering
\subfigure[Insertion-tolerant random graph]{
\includegraphics[height=2.3in]{cluster}
\label{fig:cluster}
}
\subfigure[Repetitive random graph]{
\includegraphics[height=2.2in]{Kenyon}
\label{fig:Kenyontree}
}
\caption{Random subgraphs of the Cayley graph of $\mathbb{Z}^2$.}
\end{figure}
However, this property does not hold for other examples of random graphs
arising as orbit closures of repetitive subgraphs of $\mathbb{G}$ in the
Gromov-Hausdorff space (described in \cite{Gh}; see also \cite{ALM} and
\cite{B}), see Figure~\ref{fig:Kenyontree}. The \emph{repetitiveness} of a
subgraph of $\mathbb{G}$ is equivalent to the minimality of its orbit closure.
\medskip
In this paper, we show that the two properties above cannot occur
simultaneously. More precisely, any random subgraph of a Cayley graph $\mathbb{G}$
with insertion-tolerant, ergodic in restriction to the set of infinite states
and quasi-invariant law and the orbit closure of any proper repetitive subgraph
of $\mathbb{G}$ are mutually singular.
\section{A lemma for Borel equivalence relations}
\label{SBorel}
Let $X$ be a Borel space and let $\mathcal{R} \subset X \times X$ be a countable
Borel equivalence relation. For each $x \in X$, we define
$\mathcal{R} [x] = \{ \,y \in X \mid (x,y) \in \mathcal{R} \,\}$, and similarly for each
Borel set $B \subset X$,
\[
\mathcal{R} [B] = \bigcup_{x \in B} \mathcal{R} [x].
\]
A Borel probability measure $\mu$ on $X$ is \emph{$\mathcal{R}$-invariant} if
$\varphi_\ast (\mu|_A) = \mu|_B$ for any partial transformation
$\varphi:A\to B$ of $\mathcal{R}$ (i.e. $\varphi$ is a measurable bijection whose
graph is contained in $\mathcal{R}$), where $\mu|_A$ and $\mu|_B$ are the measures
restricted to $A$ and $B$ respectively. If only $\mu$-null sets are preserved,
$\mu$ is \emph{$\mathcal{R}$-quasi-invariant} and $\mathcal{R}$ is
\emph{$\mu$-nonsingular}. The measure $\mu$ is \emph{$\mathcal{R}$-ergodic} if either
$\mu(\mathcal{R} [B]) = 0$ or $\mu(\mathcal{R} [B]) = 1$ for every Borel set $B \subset X$.
\medskip
If $X$ is also equipped with a topology, a closed subset $Z \subset X$ is
\emph{$\mathcal{R}$-minimal} if every class $\mathcal{R}[x]$ is dense in $Z$. Under some
additional assumptions, we have the following general lemma:
\begin{lemma}
\label{keylemma}
Let $X$ be a first countable Hausdorff space and let $\mathcal{L}\subset X\times X$
be a countable Borel equivalence relation. Let $\mathcal{R}$ be a nonsingular
equivalence subrelation of $\mathcal{L}$ equipped with an ergodic quasi-invariant
probability measure $\mu$. Suppose that there exists a point $x$ in the
support of $\mu$ such that $\mathcal{L}[x] = \{x\}$. If $Z$ is a closed
$\mathcal{L}$-minimal subset of $X$, then either $Z = \{x\}$ or $\mu(Z) = 0$.
\end{lemma}
\begin{proof}
Since $X$ is a first countable Hausdorff space and $x$ belongs to the support
of $\mu$, there is a decreasing sequence of open neighborhoods $U_n$ of $x$
such that
\[
\bigcap_{n \in \mathbb{N}} U_n = \{x\}
\quad \text{and} \quad
\text{$\mu(U_n)>0$ for all $n \in \mathbb{N}$.}
\]
As $\mu$ is $\mathcal{R}$-quasi-invariant and $\mathcal{R}$-ergodic, $\mu(\mathcal{R}[U_n])=1$
for all $n \in \mathbb{N}$ and hence
\[
Y = \bigcap_{n \in \mathbb{N}}\mathcal{R}[ U_n]
\]
is also a conull Borel set. Let $Z$ be a closed $\mathcal{L}$-minimal subset of $X$
and assume that there is a point $z \in Z \cap Y$. Therefore, since
$\mathcal{R} \subset \mathcal{L}$, the point
\[
z \in Z \cap \mathcal{R}[ U_n] \subset Z \cap \mathcal{L}[U_n]
\]
for all $n\in\mathbb{N}$. Thus, for each $n\in\mathbb{N}$, we can find a point
$z_n\in Z\cap U_n$. The sequence of points $z_n \in Z$ converges to $x$, but
as $Z$ is closed, it follows that $x \in Z$. By the $\mathcal{L}$-minimality of
$Z$, we conclude that $Z =\{x\}$. Finally, if $Z \cap Y = \emptyset$, then
$\mu(Z) = 0$ since $Y$ is conull.
\end{proof}
Actually, assuming $X$ is compact, we can replace $\{x\}$ by any
$\mathcal{L}$-minimal set $\mathcal{M}$ contained in the support of $\mu$. If there is a
point $z \in Z \cap Y$, we still obtain a sequence of points
$z_n\in X\cap U_n$. By compactness, passing to a subsequence if necessary, we
can assume that $z_n$ converges to point $x \in \mathcal{M}$. But as before, since
$Z$ is closed, the limit point $x \in Z$ and hence $Z = \mathcal{M}$ by minimality.
\section{Random subgraphs of Cayley graphs}
\label{Srandomsubgraphs}
Let $G$ be a countable group $G$ with a finite and symmetric generating set
$S\subset G$. The \emph{Cayley graph} $\mathbb{G} = (V,E)$ associated to $(G,S)$ is
defined by $V = G$ and $E = \{\,(g,h)\in G\times G\mid hg^{-1}\in S\,\}$. It
has a natural right $G$-action by graph automorphisms that combines the natural
right action of $G$ on itself and the right diagonal action of $G$ on
$E\subset G\times G$. Let $2^E$ the power set of all subsets of $E$, equipped
with the product topology. This is a compact metrizable space, so in particular
first countable and Hausdorff. It also have a natural left $G$-action by
homeomorphisms, which is induced by the natural right $G$-action on $E$, namely
\[
g.\omega = \omega g^{-1}
\]
for all $g \in G$ and all $\omega \in 2^E$. Note that $E$ is a fixed point for
this action. Given any finite and symmetric subset $F \subset E$, we consider
the open set
\[
U_F = \{\,\omega \in 2^E \mid F \subset \omega\,\}.
\]
and thus we have that
\[
\{E\} = \bigcap_{n \in \mathbb{N}} U_{F_n}
\]
for any increasing exhaustion $\{F_n\}$ of $E$ by finite and symmetric subsets.
\medskip
For each $g \in G$ and each $\omega \in 2^E$, we denote by $C_g(\omega)$ the
connected component or \emph{cluster} of the graph $(G,\omega)$ that contains
$g$. If we take $g$ equal to the identity $\mathds{1}$, the map associating to each
element $\omega \in 2^E$ the cluster $C_\mathds{1}(\omega) \subset \mathbb{G}$ is a
continuous map. It takes values in the \emph{Gromov-Hausdorff space} $\mathcal{G}$
formed of all connected subgraphs $\mathbb{H}$ of $\mathbb{G}$ containing $\mathds{1}$ (see
\cite{ALM} and \cite{Gh}). In fact, this space is endowed with the ultrametric
\[
d(\mathbb{H},\mathbb{H}') = 1/\exp(\sup\{\,r \geq 0 \mid B_{\mathbb{H}}(\mathds{1},r)
= B_{\mathbb{H}'}(\mathds{1},r)\,\}),
\]
where $B_{\mathbb{H}}(\mathds{1},r)$ is the combinatorial closed ball of radius $r$ centered
at $\mathds{1}$.
\medskip
The \emph{orbit equivalence relation} $\mathcal{R}^G$ on $2^E$ is defined by
\[
\mathcal{R}^G = \{\,(\omega,\omega') \in 2^E\times 2^E
\mid\exists\,g\in G:\omega'=g.\omega \,\},
\]
and the \emph{cluster equivalence relation}s $\mathcal{R}$ on $2^E$ by
\[
\mathcal{R} = \{\,(\omega,\omega') \in 2^E \times 2^E \mid
\exists \, g \in C_\mathds{1}(\omega) : \omega' = g.\omega \,\}.
\]
On the quotient $\mathcal{G}$, we have another natural equivalence relation, abusively denoted by $\mathcal{R}$, which is defined by
\[
\mathcal{R} = \{\,(\mathbb{H},\mathbb{H}') \in \mathcal{G} \times \mathcal{G}
\mid \exists \, g \in G : \mathbb{H}' = g.\mathbb{H} =\mathbb{H}'g^{-1} \,\}
\]
and induced by the equivalence relation $\mathcal{R}$ on $2^E$. These two compatible
actions are the orbital equivalence relations defined by the left actions of the
pseudogroups generated by the local transformations $\omega \mapsto g.\omega$
and $\mathbb{H} \mapsto g.\mathbb{H}$ defined on the open subset of $2^E$ of all sets
$\omega$ such that $g\in C_\mathds{1}(\omega)$ and the open subset of $\mathcal{G}$ of all
subgraphs $\mathbb{H}$ containing the vertex $g$ respectively.
\medskip
Let $\mu$ be a probability measure on $X=2^E$ that makes the cluster
equivalence relation $\mathcal{R}$ nonsingular. Assume that:
\begin{itemize}
\item[(i)] The $\mathcal{R}$-invariant set
$X_\infty = \{\,\omega \in 2^E \mid card(\mathcal{R}[\omega]) = \infty \,\}$
has positive measure, and the restriction of $\mu$ to this set is
$\mathcal{R}$-ergodic.
\item[(ii)] The measure $\mu$ is \emph{insertion-tolerant}, i.e. for each finite
and symmetric set $F \subset E$, the map $i_F : 2^E \to 2^E$ given by
\[
i_F(\omega) = \omega \cup F
\]
preserves all Borel sets of positive $\mu$-measure.
\end{itemize}
Since $\mu(X_\infty)>0$, we can replace $\mu$ with itself conditioned to
$X_\infty$, which is still denoted abusively by $\mu$. By our second
assumption, if $\{F_n\}$ is an increasing exhaustion of $E$, the open sets
$i_{F_n}(X_\infty) = U_{F_n} \cap X_\infty$
have positive $\mu$-measure for every $n$. It follows that all the open
neighborhoods $U_{F_n}$ of $E$ have also positive $\mu$-measure and hence $E$
belong to the support of $\mu$. Finally, since $E$ is a fixed point of the
$G$-action on $X$ and $C_\mathds{1}(E) = \mathbb{G}$, we conclude that $\mathcal{R}[E] = \{E\}$.
Lemma~\ref{keylemma} applied to the first countable Hausdorff space $X = 2^E$
equipped with the equivalence relation $\mathcal{L} = \mathcal{R}$ and the $G$-fixed point
$x = E$ yields:
\begin{theorem}
\label{thm:maintheorem1}
Let $\mu$ be a probability measure on $X=2^E$. If $\mu$ is quasi-invariant
and ergodic with respect to $\mathcal{R}$ and insertion tolerant, then any closed
$\mathcal{R}$-minimal set $Z \subset X$ is either $Z = \{E\}$ or $\mu(Z) = 0$. \qed
\end{theorem}
In this result, we can replace the configuration space $X=2^E$ with the
Gromov-Hausdorff space $\mathcal{G}$ so that a probability measure $\mu$ makes
$\mathcal{G}$ a \emph{random subgraph} of the Cayley graph $\mathbb{G}$. If $\mu$ is
quasi-invariant with respect to $\mathcal{R}$, we say the random graph $(\mathcal{G},\mu)$
is \emph{nonsingular}. Let $\mathcal{G}_\bullet$ be the locally compact metrizable
space of the isomorphism classes of locally finite connected rooted graphs (see
\cite{BS}). Using the natural continuous map from $\mathcal{G}$ to $\mathcal{G}_\bullet$ sending
each graph $\mathbb{H}$ on its isomorphism class $[\mathbb{H}]$, we can interpret $(\mathcal{G},\mu)$ as a \emph{random graph} in
the sense of \cite{AL}. On the other hand, as pointed out in the introduction,
it is well known (see for example \cite{ALM} and \cite{B}) that the closed
$\mathcal{R}$-minimal subsets of $\mathcal{G}$ are in one-to-one correspondence with the
orbit closures of repetitive subgraphs of $\mathbb{G}$. Now, we can state an
equivalent version of the previous theorem:
\begin{theorem}
\label{thm:maintheorem2}
Any nonsingular random subgraph of a Cayley graph $\mathbb{G}$ whose law is
insertion-tolerant and ergodic in restriction to the set of its infinite
states and any orbit closure of a proper repetitive subgraph of $\mathbb{G}$ are
mutually singular. \qed
\end{theorem}
\section{Examples}
In this section, we assemble some examples to which Theorems~\ref{thm:maintheorem1}~and~\ref{thm:maintheorem2} apply.
\begin{example} \label{ex:Bernoulli}
\emph{Bernoulli bond percolation} on a Cayley graph $\mathbb{G}$ provides the basic
example of random graph, which is obtained by keeping each edge with constant
probability $p$ independently to other edges. More precisely, the power set
$2^E$ is equipped with the Bernoulli measure $\mu$ with constant survival
parameter $p$ (which means that $\mu$ is the product of the measure on
$\{0,1\}$ assigning probabilities $1-p$ and $p$ to $0$ and $1$ respectively).
In the supercritical phase $p_c < p <1$, if we denote by $\mathbb{P}$ the law that
governs the clusters $C_\mathds{1}(\omega)$, the set of infinite clusters has
positive $\mathbb{P}$-measure. According to the cluster indistinguishability theorem
proved by R.~Lyons and O.~Schramm in \cite{LS} (see also \cite{GL}), the
measure $\mathbb{P}$ conditioned to the set of infinite clusters is $\mathcal{R}$-ergodic
and insertion-tolerant. Then the orbit closure of any proper repetitive
subgraph $\mathbb{H}$ of $\mathbb{G}$ is $\mathbb{P}$-null.
\end{example}
\begin{example} \label{ex:percolation}
Since Theorem 3.3 of \cite{LS} is valid for any insertion-tolerant
$\mathcal{R}$-invariant probability measure $\mu$ on $2^E$,
Theorem~\ref{thm:maintheorem2} also applies to random graphs obtained by this
kind of $G$-invariant percolation. Recall that a \emph{$G$-invariant bond
percolation process} on $\mathbb{G}$ is given by a measure preserving $G$-action on
a standard Borel probability space $(X,\mu)$ together with a $G$-equivariant
Borel map $\pi:X\to 2^E$ (see \cite{G}). In the case under consideration,
the map $\pi$ is the identity $id$, but the law of the process is not longer
the Bernoulli measure. However, the cluster indistinguishability theorem also
holds for some $G$-invariant percolation processes with $\pi\neq id$, like
the \emph{percolation process with scenery} introduced in \cite[Remark 3.4]{LS} (see also \cite[Proposition 6]{GL}).
\end{example}
\begin{example} \label{ex:unimodular}
Let $\mathbb{G} = (V,E)$ be a locally finite graph and let $G$ be a unimodular
closed group of automorphisms of $\mathbb{G}$ acting transitively on $V$. Then the
proof of Theorem~\ref{thm:maintheorem1} extends to this case. Thus, if $\mu$
is an insertion tolerant, $\mathcal{R}$-ergodic and $\mathcal{R}$-invariant probability
measure on $2^E$, any closed $\mathcal{R}$-minimal subset $Z$ of $2^E$ is either
$Z = \{E\}$ or $\mu(Z)=0$. But Theorem 3.3 of \cite{LS} is also valid for any
$G$-invariant insertion-tolerant bond percolation process on $\mathbb{G}$, and hence
Theorem~\ref{thm:maintheorem2} applies to all unimodular random graphs
obtained by this kind of percolation.
\end{example}
\section*{Acknowledgements}
The Lemma~\ref{keylemma} and its proof were provided by the referee based on
the first version of the paper. Section~\ref{Srandomsubgraphs} also benefited
from their suggestions and comments. The authors were supported by Spanish
Excellence Grant MTM2013-46337-C2-2-P, Galician Grant GPC2015/006 and the
European Regional Development Fund. Second author was also supported by the
European Social Fund and Diputaci\'on General de Arag\'on (Grant E15
Geometr\'ia).
|
\section{Introduction}
In $1974$, Gabriel \cite{gab} reduced the problem of classifying
bilinear forms over an arbitrary field $\F$ to the problem of
classifying nonsingular bilinear forms. In this paper, we take an analogous step towards the topological classification of
bilinear and sesquilinear forms, reducing it to the nonsingular case.
Unlike the problem of topological classification of forms, which has not yet been considered, the
problem of topological classification of linear operators has been
thoroughly studied. Kuiper and Robbin \cite{Kuip-Robb, Robb} gave a
criterion for topological similarity of real matrices without
eigenvalues that are roots of $1$. Their result was extended to
complex matrices in \cite{bud1}. The problem of topological
similarity of matrices with an eigenvalue that is a root of $1$ was
also considered by these authors \cite{Kuip-Robb, Robb} as well as by Cappell and Shaneson
\cite{Capp-conexamp, Capp-2th-nas-n<=6,Capp-big-n<6,
Cap+sha,Cap+ste}, and by Hambleton and Pedersen \cite{h-p,h-p1}. The
problem of topological classification was studied for orthogonal
operators \cite{Pardon}, for affine operators
\cite{Blanc,bud,bud1,Ephr}, for M\"obius transformations
\cite{ryb+ser_meb}, for chains of linear mappings \cite{ryb+ser},
for matrix pencils \cite{f-r-s}, for oriented cycles of linear
mappings \cite{ryb+ser1}, and for quiver representations \cite{lop}.
A pair $(U,\Phi)$ consisting of a vector space $U$ and a bilinear
form $\Phi$ is called by Gabriel \cite{gab} a \emph{bilinear space}.
Similarly, we call a pair $(U,\Phi)$ a \emph{sesquilinear space} if
$\Phi$ is a sesquilinear form. A pair $(U,\Phi)$ is \emph{singular} or
\emph{nonsingular} if $\Phi$ is so. Two spaces $(U,\Phi)$ are
$(V,\Psi)$ are \emph{isomorphic} if there exists a linear bijection
$\varphi :U\to V$ such that
\begin{equation}\label{fdr}
\Phi(x,y)=\Psi(\varphi (x),\varphi (y))\, ,\qquad\text{for all
}x,y\in U.
\end{equation}
The \emph{direct sum} of pairs is the pair
\[
(U,\Phi)\oplus(V,\Psi):=(U\oplus V,\Phi\oplus\Psi).
\]
A pair is \emph{indecomposable} if it is not isomorphic to a direct sum of pairs with vector spaces of smaller sizes.
Let vector spaces $U$ and $V$ be also topological spaces. For
example, they are subspaces of $\C^m:=\C\oplus\dots\oplus\C$ ($m$
summands) with a usual topology. We say that $(U,\Phi)$ and
$(V,\Psi)$ are \emph{topologically equivalent} if there exists a
homeomorphism $\varphi :U\to V$, i.e., a continuous bijection whose
inverse is also a continuous bijection, such that \eqref{fdr} holds.
The main result of the paper is
the following theorem, which is
proved in Section \ref{ssw}.
\begin{theorem}\label{jus}
Let $\F$ be $\C$ or $\R$. Let $(\F^m,\Phi)$ and $(\F^n,\Psi)$ be two
bilinear or two sesquilinear spaces that are topologically
equivalent. Suppose that
\begin{align*}
(\F^m,\Phi)&=(U_0,\Phi_0)\oplus(U_1,\Phi_1)
\oplus\dots\oplus(U_r,\Phi_r)\\
(\F^n,\Psi)&=(V_0,\Psi_0)\oplus(V_1,\Psi_1)
\oplus\dots\oplus(V_s,\Psi_s),
\end{align*}
where $(U_0,\Phi_0)$ and $(V_0,\Psi_0)$ are nonsingular and the
other summands are indecomposable and singular. Then $m=n$, $r=s$,
$(U_0,\Phi_0)$ and $(V_0,\Psi_0)$ are topologically equivalent, and,
after a suitable reindexing, each $(U_i,\Phi_i)$ is isomorphic to
$(V_i,\Psi_i)$.
\end{theorem}
The equality $m=n$ in Theorem \ref{jus}
holds due to
the following statement:
\begin{equation}\label{svt}
\text{if $\F\in\{\C,\R\}$ and $\F^m$ is homeomorphic to $\F^n$, then
$m=n$}
\end{equation}
(see \cite[Corollary 19.10]{Bred} or \cite[Section 11]{McCl}).
Let us reformulate Theorem \ref{jus} in a matrix form. Each bilinear
or sesquilinear space $(\F^m,\Phi)$ can be given by the pair
$(\F^m,A)$, in which $A$ is the matrix of $\Phi$ in the standard
basis. Changing the basis, we can reduce $A$ by congruence
transformations $SAS^T$ with nonsingular $S\in\F^{n\times n}$ if
$\Phi$ is bilinear, or by *congruence transformations $SAS^*$ with
nonsingular $S\in\F^{n\times n}$ if $\Phi$ is sesquilinear.
Each square matrix $M$ over $\F$ is congruent (resp., *congruent) to a direct sum
\begin{equation}\label{li5}
R\oplus J_{n_1}\oplus\dots\oplus J_{n_p}\, ,\qquad\text{with a
nonsingular $R$,}
\end{equation}
in which the matrix $R$ is uniquely determined by $M$ up to
congruence (resp., *congruence) and the $n_i$-by-$n_i$ singular
Jordan blocks $J_{n_i}$ are uniquely determined up to permutation;
see Theorem \ref{t3}(a). Horn and Sergeichuk \cite{h-s_lin} called
the sum \eqref{li5} the \emph{regularizing decomposition} of $M$,
$J_{n_1},\dots, J_{n_p}$ the \emph{singular summands}, and the
matrix $R$ the \emph{regular part} of $M$. They gave an algorithm
for constructing \eqref{li5} by $M$.
We say that two matrices $A,B\in\F^{n\times n}$ are
\emph{topologically congruent} (resp., \emph{topologically
*congruent}) if the bilinear (resp., sesquilinear) spaces $(\F^n,A)$
and $(\F^n,B)$ are topologically equivalent.
The next theorem is the matricial analogue of Theorem \ref{jus}.
\begin{theorem}\label{jye}
Two square matrices over $\F\in\{\C,\R\}$ are topologically congruent $($resp., *congruent$)$ if and only if their regularizing decompositions coincide up to topological congruence $($resp., *congruence$)$ of their regular parts and permutations of direct summands.
\end{theorem}
The regularizing decomposition \eqref{li5} is the first step towards
reducing a matrix to its canonical form under congruence and
*congruence. Canonical forms under congruence and *congruence over
any field $\F$ of characteristic not 2 were given by Sergeichuk
\cite{ser_izv} (see also \cite{h-s_con}) up to classification of
quadratic and Hermitian forms over finite extensions of $\F$. They
were latter simplified for the case of complex matrices by Horn and
Sergeichuk \cite{hor-ser}. An alternative proof that the canonical
matrices from \cite{hor-ser} are indeed canonical was given by Horn
and Sergeichuk \cite{hor-ser_can}. These authors gave the proof only
for nonsingular matrices, which was sufficient due to the uniqueness
of regularizing decomposition (Theorem \ref{t3}(a)).
\section{The regularizing algorithm}
In this section, we recall the regularizing algorithm for matrices
under congruence and *congruence, which was constructed by Horn and
Sergeichuk \cite{h-s_lin}. An analogous regularization algorithm for
matrix pencils was constructed by Van Dooren \cite{doo}.
Let $\mathbb F$ be any field with a fixed involution $a\mapsto
\tilde{a}$, which can be the identity. We say that a form is
\emph{$^{\star}$\!sesquilinear} (we use a five-pointed star) if it
is sesquilinear with respect to this involution. The transformation
$A\mapsto SAS^{\star}$ of $A\in\F^{n\times n}$, in which
$S\in\F^{n\times n}$ is nonsingular and $S^{\star}:=\tilde S^T$, is
called the \emph{$^{\star}$\!congruence transformation}. We remark
that {$^{\star}$congruence transformations over $\F=\C$ are
*congruence transformations if the involution $a\mapsto \tilde{a}$
is the complex conjugation, and they are congruence transformations
if the involution is the identity.
We denote by $0_n$ the zero matrix of size $n\times n$, $n\ge 0$,
assuming that when $n=0$ we formally have an empty square matrix.
Let $A$ be a singular square matrix over $\mathbb
F$. We reduce it by
$^{\star}$congruence transformations
as follows:
\begin{align}\label{new1a}
A\,&\longmapsto\,SA=\begin{bmatrix}
A_1\\0
\end{bmatrix}\!\! \begin{matrix}
\\ \}{\scriptstyle m_1}
\end{matrix}
\quad
\begin{matrix} \text{($S$ is nonsingular and
the rows of}\\ \text{$A_1$
are linearly independent)}
\end{matrix}
\\ \label{new1b}
&\longmapsto\,\begin{bmatrix} A_1\\0
\end{bmatrix}S^{\star}=\left[\begin{array}{c|c}
B&C\\\hline 0&0_{m_1}
\end{array}\right]\quad
\text{($S$ is the same and
$B$ is square)}
\\ \label{new1c}
&\longmapsto\, (S_1\oplus
I_{m_1})\left[\begin{array}{c|c}
B&C\\\hline 0&0_{m_1}
\end{array}\right](S_1^{\star}\oplus
I_{m_1})
=\left[\begin{array}{cc|c}
D & E & C_1 \\
F&A_2&0\\\hline
\multicolumn{2}{c|}0 &
0_{m_1}
\end{array}\right]
\!\!
\begin{matrix}
\}{\scriptstyle m_2} \\ \\ \}{\scriptstyle m_1}
\end{matrix}
\end{align}
in which $D$ and $A_2$ are square, $S_1$ is nonsingular, and the
rows of $C_1$ are linearly independent. The nonnegative integers
$m_1,m_2$ and the matrix $A_2$ are used in the following theorem.
\begin{theorem}[\cite{h-s_lin}] \label{t3}
Let $\mathbb F$ be a field with involution, which can be the identity.
{\rm(a)} Each square matrix $A$ over $\mathbb F$ is
$^{\star}$\!congruent to a direct sum
\begin{equation*}\label{1.1}
R\oplus J_{n_1}\oplus\dots\oplus J_{n_p}\, ,\quad\text{with $R$
nonsingular,}
\end{equation*}
in which the matrix $R$ is determined by $A$ uniquely up to $^{\star}$\!congruence, and the
$n_i\times n_i$ singular Jordan
blocks $J_{n_i}$ are
determined uniquely up to permutation.
{\rm(b)} Given a singular
square matrix $A$ over $\mathbb F$.
Apply the reduction
\eqref{new1a}--\eqref{new1c} to
$A$ and get $m_1,m_2,A_2$. Apply this reduction to
$A_2$ and get $m_3,m_4,A_4$, and
so on until obtain a nonsingular
$A_{2t}$:
\begin{equation}\label{new2}
A \Longrightarrow
\begin{cases}
\quad A_2 \Longrightarrow
\\ m_1,m_2
\end{cases}
\begin{matrix}
\!\!\!\!\!\!
\begin{cases}
\quad A_4
\Longrightarrow\cdots \Longrightarrow
\\ m_3,m_4
\end{cases}\\
\phantom{A}
\end{matrix} \!\!\!\!\!
\begin{matrix}
\begin{cases}
\text{nonsingular }A_{2t}
\\ m_{2t-1},m_{2t}.
\end{cases}\\
\phantom{A}\\ \phantom{A}
\end{matrix}
\end{equation}
Then $m_1\ge m_2\cdots\dots\ge m_{2t}$ and $A$ is
$^{\star}$\!congruent to
\begin{equation*}\label{eq9}
A_{2t}\oplus
J_{1}^{[m_{1}-m_{2}]}\oplus
J_{2}^{[m_{2}-m_{3}]} \oplus
\dots\oplus
J_{2t-1}^{[m_{2t-1}-m_{2t}]}\oplus
J_{2t}^{[m_{2t}]},
\end{equation*}
in which
$J_i^{[m]}:=J_i\oplus\dots\oplus
J_i$ $(m$ summands, in particular,
$J_i^{[0]}=0_0)$. Thus, $A_{2t}$ is the regular part of $A$.
{\rm(c)} If\/ $\mathbb F=\mathbb
C$ or $\R$, then the reduction
\eqref{new2} can be realized by
unitary or orthogonal transformations, respectively, which
improves the numerical stability
of the algorithm.
\end{theorem}
\section{Proof of Theorem \ref{jus}}\label{ssw}
Theorem \ref{jus} is formulated for forms on $\C^n$ or $\R^n$, but
it is more convenient to prove it for forms on unitary or Euclidean
spaces since their subspaces are also unitary or Euclidean, respectively. A unitary space is also called a complex inner product space.
We consider unitary and Euclidean spaces as topological spaces.
Let $\F$ be $\C$ or $\R$, and let
\begin{equation*}\label{krb}
\Phi: U\times U\to \F,\qquad
\Phi': U'\times U'\to \F
\end{equation*}
be two bilinear or two sesquilinear forms on unitary spaces if
$\F=\C$, or two bilinear forms on Euclidean spaces if $\F=\R$. We
suppose that these forms are topologically equivalent, i.e., there
exists a homeomorphism $\varphi:U\to U'$ such that
\begin{equation}\label{kgr}
\Phi (x,y)=\Phi'(\varphi (x),\varphi (y))\, ,\qquad\text{for all
$x,y\in U$}.
\end{equation}
Let $A$ and $A'$ be matrices of $\Phi $ and $\Phi '$ in orthonormal
bases. Applying to $A$ the reduction \eqref{new1a}--\eqref{new1c}
in which the transforming matrices are unitary if $F=\C$ or
orthogonal if $F=\R$, we obtain $m_1$, $m_2$ and $A_2$. Applying it
to $A'$, we obtain $m'_1$, $m'_2$, and $A'_2$. We need to prove that
\begin{equation}\label{yep}
m_1=m_1',\quad m_2=m_2',\quad \text{and $A_2$ and $A_2'$ are
topologically $^{\star}$congruent}
\end{equation}
(that is, $A_2$ and $A_2'$ are topologically congruent if the forms
$\Phi $ and $\Phi '$ are bilinear; they are topologically *congruent
if the forms are sesquilinear).
Let $S$ be a unitary matrix if $\F=\C$ or an orthogonal matrix if $\F=\R$ such that $SA$ has the form given in \eqref{new1a}. Then
\begin{equation}\label{kie}
SAS^{\star}=\left[\begin{array}{c|c}
B&C\\\hline 0&0_{m_1}
\end{array}\right]\quad
\text{(the rows of $[B\,C]$
are linearly independent)}
\end{equation}
is the matrix of $\Phi$ in a new orthonormal basis.
The basis vectors that correspond to the second horizontal strip of \eqref{kie} generate the vector space
\begin{equation*}\label{jyw}
L:=\{x\in U\,|\,\Phi (x,U)=0\},
\end{equation*}
which is called the \emph{left kernel}, or the left radical, of
$\Phi $. Denote by $L'$ the left kernel of $\Phi '$. If $x\in L$,
then $\Phi' (\varphi (x),U')= \Phi (x,U)=0$ by \eqref{kgr}, hence
$\varphi(L)\subset L'$. The inclusion $\varphi(L)\supset L'$ holds
too since, for each $x'\in L'$ and setting $x:=\varphi ^{-1}(x')$ we
have
$$\Phi (x,U)=\Phi'(\varphi (x),\varphi(U))=\Phi'(x',U')=0\, ,$$ and so
$x\in L$. Thus,
\begin{equation}\label{lwp}
\varphi(L)=L',
\end{equation}
which proves the first equality in \eqref{yep} due to \eqref{svt}.
Let $S_1$ be a unitary matrix if $\F=\C$ or an orthogonal matrix if $\F=\R$ such that
\[
S_1C=\begin{bmatrix}
C_1\\
0\\
\end{bmatrix}
\quad \text{(the rows of $C_1$ are linearly independent)}.
\]
Then
\begin{equation}\label{lke}
(S_1\oplus
I_{m_1})\left[\begin{array}{c|c}
B&C\\\hline 0&0_{m_1}
\end{array}\right](S_1^{\star}\oplus
I_{m_1})
=\left[\begin{array}{cc|c}
D & E & C_1 \\
F&A_2&0\\\hline
\multicolumn{2}{c|}0 &
0_{m_1}
\end{array}\right]
\!\!
\begin{matrix}
\}{\scriptstyle m_2} \\ \\ \}{\scriptstyle m_1}
\end{matrix}
\end{equation}
is the matrix of $\Phi $ in a new orthonormal basis.
Since the basis vectors that correspond to the columns of $C_1$ generate $L$, the basis vectors that correspond to the second and third horizontal strips of the right hand side matrix in \eqref{lke} generate the vector space
\begin{equation*}\label{der}
K:=\{x\in U\,|\,\Phi(x,L)=0\}.
\end{equation*}
Analogously, define $K':=\{x'\in U'\,|\,\Phi'(x',L')=0\}$.
By \eqref{kgr} and \eqref{lwp}, for each $x\in K$ we have
$$\Phi'(\varphi (x),L')=\Phi'(\varphi (x),\varphi (L))=\Phi(x,L)=0\,
,$$ and consequently $\varphi (K)\subset K'$. The inclusion
$\varphi(K)\supset K'$ holds too since for each $x'\in K'$ and
$x:=\varphi ^{-1}(x')$, we have $$\Phi (x,L)=\Phi'(\varphi
(x),\varphi(L))=\Phi'(x',L')=0\, ,$$ and so $x\in K$. Thus,
\begin{equation*}\label{kyt}
\varphi(K)=K'.
\end{equation*}
By \eqref{svt}, $$m_2=\dim U-\dim K=\dim U'-\dim K'=m_2'\, ,$$ which
proves the second equality in \eqref{yep}.
Since the basis in $U$ is orthonormal, the basis vectors that
correspond to the second horizontal strip of the right hand side
matrix in \eqref{lke} generate the vector space
\[
L^{\bot}_K:=\{x\in K\,|\,(x,L)=0\},
\]
which is the orthogonal complement of $L$ in $K$. Analogously, write
$L^{\prime\bot}_{K'}:=\{x'\in K'\,|\,(x',L')=0\}.$ Then $
K=L^{\bot}_{K}\oplus L$ and $K'=L^{\prime\bot}_{K'}\oplus L'. $
Define the maps that
are the compositions of three maps:
\begin{equation}\label{ka}
\begin{matrix}
\psi:&
L^{\bot}_K\a{\iota}L^{\bot}_{K}\oplus L\a{\varphi }L^{\prime\bot}_{K'}\oplus L'\a{\pi' }L^{\prime\bot}_{K'}\\
\psi':&
L^{\prime\bot}_{K'}\a{\iota'}L'^{\bot}_{K'}\oplus L'\a{\varphi^{-1} }L^{\bot}_{K}\oplus L\a{\pi }L^{\bot}_{K}
\end{matrix}
\end{equation}
where $\iota,\iota'$ are the injections and $\pi ,\pi '$ are the orthogonal projections.
\begin{lemma}\label{jje}
The map $\psi: L^{\bot}_{K}\to L^{\prime\bot}_{K'}$ is a homeomorphism and $\psi^{-1}=\psi'$.
\end{lemma}
\begin{proof}
By \eqref{kgr}, for all $x,y\in U$
\[
\Phi'(\varphi (x),\varphi (y))=\Phi (x,y)=\Phi (x+L,y)=\Phi '(\varphi (x+L),\varphi (y)),
\]
hence
$
\Phi'(\varphi (x+L)-\varphi (x),\varphi (y))=0.
$
Since $\varphi $ is a surjection,
each element of $U'$ is represented in the form $\varphi (y)$, and so $\varphi (x+L)-\varphi (x)\subset L'$. Thus,
$\varphi (x+L)\subset \varphi (x)+L'$, which implies
\begin{equation}\label{lke!}
\varphi (x+L)=\varphi (x)+L',\qquad\text{for all $x\in U$}
\end{equation}
because $\varphi $ is a surjection.
Let $x\in L^{\bot}_{K}$ and $\varphi(x)=x'+l'$, where $x'\in
L^{\prime\bot}_{K'}$ and $l'\in L'$. By \eqref{lke!}, $\varphi
(x+L)=x'+l'+L'=x'+L'$, and so there exists $l\in L$ such that
$\varphi (x+l)=x'$. By \eqref{ka},
\begin{align*}
\psi&:\ x\arr{\iota}x+0\arr{\varphi}x'+l'\arr{\pi'}x'\\ \nonumber
\psi'&:\ x'
\arr{\iota'}x'+0\arr{\varphi^{-1}}x+l\arr{\pi}x.
\end{align*}
Hence $\psi'\psi =1$. Analogously, $\psi \psi' =1$, and so
$\psi^{-1}=\psi'$. Since
$\iota,\iota',\pi,\pi',\varphi,\varphi^{-1}$ are continuous, $\psi$
and $\psi'$ are continuous too, which proves that $\psi$ is a
homeomorphism.
\end{proof}
For all $x,y\in L^{\bot}_{K}$, we write $\varphi (x)=\psi(x)+l'$ and
$\varphi (y)=\psi(y)+l''$, in which $l',l''\in L'$. It follows from
the zeros in the matrix \eqref{lke} that
\begin{equation}\label{dlt}
\begin{split}
\Phi(x,y) &= \Phi'(\varphi (x),\varphi (y)) \\
&= \Phi'(\psi(x)+l',\psi (y)+l'') \\
&= \Phi'(\psi(x),\psi(y))+\Phi'(\psi(x),l'')+\Phi'(l',\psi(y))+
\Phi'(l',l'') \\
&= \Phi'(\psi(x),\psi(y)).
\end{split}
\end{equation}
Let
\begin{equation*}\label{dr5}
\Phi_2: L^{\bot}_{K}\times L^{\bot}_{K}\to \F,\qquad
\Phi'_2: L^{\prime\bot}_{K'}\times L^{\prime\bot}_{K'}\to \F
\end{equation*}
be the restrictions of $\Phi$ and $\Phi'$. By Lemma \ref{jje} and
\eqref{dlt}, these forms are topologically equivalent. Moreover,
$A_2$ and $A_2'$ are their matrices in the orthonormal bases. We
have proved \eqref{yep}.
In the same way, we apply to $A_2$ and $A_2'$ the reduction \eqref{new1a}--\eqref{new1c} in which the transforming matrices are unitary if $\F=\C$ or orthogonal if $\F=\R$. We obtain the numbers $m_3=m_3',
m_4=m_4'$ and topologically equivalent forms $\Phi_4,\Phi'_4$ with matrices $A_4,A_4'$
in orthonormal bases (see \eqref{new2}). We repeat this reduction until obtain topologically equivalent forms $\Phi_{2t},\Phi'_{2t}$ with nonsingular matrices $A_{2t},A_{2t}'$.
By Theorem \ref{t3}(b), there exist bases of the spaces $U$ and $U'$, in which the matrices of $\Phi$ and $\Phi'$ have the form
\begin{align*}
&A_{2t}\oplus
J_{1}^{[m_{1}-m_{2}]}\oplus
J_{2}^{[m_{2}-m_{3}]} \oplus
\dots\oplus
J_{2t-1}^{[m_{2t-1}-m_{2t}]}\oplus
J_{2t}^{[m_{2t}]}
\\
&A'_{2t}\oplus
J_{1}^{[m_{1}-m_{2}]}\oplus
J_{2}^{[m_{2}-m_{3}]} \oplus
\dots\oplus
J_{2t-1}^{[m_{2t-1}-m_{2t}]}\oplus
J_{2t}^{[m_{2t}]}
\end{align*}
which completes the proof of Theorem \ref{jus}.
|
\section{Introduction}
Astrophysical turbulence, in spite of the stochastic nature,
allows for a statistical study that can have access to its underlying regularities
\citep{Biskampbook, Brad13, BL15}.
The turbulent spectrum, as a statistical measure of turbulence, contains a great wealth of information on the injection, nonlinear transfer, and
dissipation of turbulent energy, and thus can characterize the essential properties of interstellar turbulence.
The statistics of the turbulent velocity provide a direct diagnostic of the turbulent spectrum, but it is challenging to disentangle velocity and
density contributions when utilizing spectroscopic data to obtain velocity statistics
\citep{Ves03}.
Among the attempts to overcome this difficulty,
new techniques, e.g., the Velocity Channel Analysis and Velocity Coordinate Spectrum, have been developed on a solid theoretical ground
and successfully tested by numerical simulations
(see a review by \citealt{Laz09rev}).
On the other hand, the statistical study of density is rather straightforward and has attracted more attention
\citep{GN85, Spa90, Armstrong95, CL10}.
A Kolmogorov spectrum of the fluctuations in the interstellar plasma density is suggested by observational evidence.
However, density is a passive scalar and the measure of density fluctuations
can only be regarded as an indirect approach of tracing turbulence.
Numerical studies show that
the density spectrum significantly deviates from the velocity spectrum in supersonic turbulence
\citep{CL03, BLC05,KL07}.
Turbulence induces fluctuating magnetic field by the small-scale dynamo, and brings the magnetic energy up to
the injection scale of turbulence through the inverse cascade
\citep{CVB09, Bere11, Zra14}.
The generated magnetic field in turn affects the properties of turbulence and converts the hydrodynamic turbulence
into magnetohydrodynamic (MHD) turbulence, which is a common state of interstellar plasma.
Achieving the spectral profile of turbulent magnetic field is crucial for
studying the processes such as cosmic-ray scattering
\citep{Kota_Jok2000},
star formation
\citep{Mck99},
and magnetic reconnection
\citep{LV99}.
It requires an adequate understanding of MHD turbulence.
The point of contention is whether the scaling law for hydrodynamic turbulence is still valid
in the context of MHD turbulence.
Within theory's reach,
\citet{GS95}
pointed out that the transverse mixing motions of magnetic field lines in MHD turbulence preserve the character of
hydrodynamic turbulent motions,
thus the hydrodynamic turbulence scaling holds in the direction perpendicular to the local magnetic field.
The Kolmogorov-type spectrum of density fluctuations in the magnetized interstellar plasma is in accordance with this theoretical expectation.
Statistical analyses of the magnetic field data produced by numerical simulations support the theory
\citep{CV00, CLV_incomp,MG01}.
However, there is a shortage of observational evidence since
magnetic field statistics are more poorly constrained from observations than velocity and density
statistics.
Magnetic field cannot be measured independently, but is intermixed with other quantities such as densities of relativistic
electrons or thermal electrons.
Only a theoretical model capable of reproducing the detected features of related observables
can give us confidence in eliminating the inherent ambiguities and unveiling the physics in the measurements of turbulent
magnetic fields.
Based on the modern understanding of MHD turbulence,
\citet{LP12, LP15}
carried out comprehensive statistical studies
on fluctuations in synchrotron intensity, synchrotron polarization, and Faraday measure.
They provided a thorough exposition on the quantitative correlations between the
statistics of synchrotron emission and characteristics of the underlying magnetic turbulence.
Their synchrotron studies of turbulence and cosmic magnetic fields
open the avenue to a wide range of astrophysical applications.
In particular,
\citet{LP15} (hereafter LP16)
presented the structure function (SF) analysis of rotation measures (RMs), including both cases with
spatially-coincident and spatially-separated synchrotron emission and Faraday rotation regions.
The latter case can be applied to probing the turbulent magnetic fields
embedded in the diffuse ionized component of the interstellar medium (ISM), when the observed Faraday rotation
only contains the contribution from the Galaxy.
In practice, the SFs of RMs have been attained from a number of independent observations covering
different scales and areas in the Galaxy
(e.g., \citealt{Sim84,Sim86,Cle92,MS96,Hav03,Hav04,SH04,Hav08,Roy08, Stil09, Opp12};
see also, \citealt{Han04,Han09}).
In combination with the SFs of emission measures (EMs), it provides a possibility for determining the properties of turbulent magnetic fields in the ISM.
The observations reveal some common features in the form of RM SFs as a function of angular scales:
(1) The SF has a much shallower slope than that expected from the standard Kolmogorov power law.
(2) The SF follows a broken power spectrum changing from a relatively steeper slope to a shallower one at a scale
on the order of $1$ pc.
(3) The slope of the SF varies from region to region and has a dependence on Galactic latitude.
(4) The slope of the SF tends to flatten at large angular scales.
(5) The SF of EMs shows a similar slope to that of RMs detected from the same region.
There has not been a compelling interpretation for these features in earlier literature.
As an empirical attempt,
\citet{MS96}
suggested
that the broken power-law spectrum may result from the transition from three-dimensional to two-dimensional filamentary turbulent
structure, but this imposed turbulent structure, which may not be a common occurrence,
fails to account for other observational features.
As the major impediment of the problem,
both density and magnetic field fluctuations are imprinted in the observed RM fluctuations.
The relative importance between them determines
whether the behavior of RM SFs can effectively diagnose the turbulent magnetic fields.
In the studies of, e.g.
\citet{Sim84, LP15},
the product of electron density and magnetic field has been treated as a composite quantity.
In
\citet{MS96},
the two are separated, but both fluctuations are assumed to conform to the Kolmogorov spectrum.
In the current work, in order to resolve the respective influence of density and magnetic field fluctuations
and gain a clear insight into the properties of their associated turbulence,
we separate the density and magnetic field components in the statistical analysis of RM fluctuations,
but do not restrict the scalings of their respective power spectra.
Our goal is not only to seek understanding of the observed features of RM SFs,
but also to clarify its relation with the turbulent density field and magnetic field.
In Section 2, we present a statistical analysis of RM and EM fluctuations, and provide the expressions of
their SFs with separate contributions from electron density and magnetic field.
In Section 3, we compare the analytical result with the measured RM and EM SFs from observations.
Conclusions and discussions are given in Section 4.
\section{Statistical analysis of RM and EM fluctuations}
As the fundamental radio propagation measurements, the measure of magnetization (RM) and electron densities
(EM) provide unique information on the magnetized turbulence in the diffusive, ionized component of the ISM.
In astronomically convenient units, they are defined as
\begin{equation}
\text{RM} (\text{rad m}^{-2})= 0.81 \int_0^L n_e (\text{cm}^{-3}) B_z ( \mu \text{G}) dz (\text{pc}),
\end{equation}
and
\begin{equation}
\text{EM} (\text{pc cm}^{-6}) = \int_0^L [n_e (\text{cm}^{-3})]^2 dz (\text{pc}),
\end{equation}
where $n_e$ is electron density, $B_z$ is the line-of-sight (LOS) component of magnetic field, and $L$ is the path length
through the Faraday rotating medium.
\subsection{SFs of RM and EM fluctuations}\label{sec: lp}
For our statistical analysis of RM and EM fluctuations, we follow the approach employed by
LP16
that deals with
RM per unit length along LOS, namely, RM density $0.81(n_eB_z)$, and EM density $(n_e^2)$.
We first treat them as composite quantities.
We consider them to be statistically homogeneous and isotropic.
This ensures that these quantities are invariant with respect to the LOS orientation.
We assume that the RM (EM) density can be described as a sum of its ensemble-average mean and zero mean fluctuations,
\begin{equation}\label{eq: rmdendes}
\phi(\bm{X}, z) = \phi_0 + \delta \phi (\bm{X}, z), ~~ \langle \delta \phi(\bm{X}, z) \rangle =0,
\end{equation}
where $\bm{X}$ denotes the position on the plane of sky and $z$ is the distance along the LOS.
We use $\langle ... \rangle$ to denote an ensemble average.
As the real-space statistical tool, the two-point correlation function (CF) is
\begin{equation}
\xi(\bm{R}, \Delta z) = \langle \phi(\bm{X_1}, z_1) \phi(\bm{X_2}, z_2) \rangle,
\end{equation}
for RM (EM) density, and
\begin{equation}
\widetilde{\xi}(\bm{R}, \Delta z) = \langle \delta \phi(\bm{X_1}, z_1) \delta \phi(\bm{X_2}, z_2) \rangle
\end{equation}
for RM (EM) density fluctuations.
The two are related as
\begin{equation}
\widetilde{\xi}(\bm{R}, \Delta z) = \xi(\bm{R}, \Delta z) - \phi_0^2,
\end{equation}
with $\bm{R} = \bm{X_1}-\bm{X_2}$ and $\Delta z = z_1-z_2$.
The SF of RM (EM) density and RM (EM) density fluctuations are identical,
\begin{equation}\label{eq: rmdsf}
d(\bm{R}, \Delta z) = \widetilde{d}(\bm{R}, \Delta z) =
\langle [\delta \phi(\bm{X_1}, z_1)- \delta \phi(\bm{X_2}, z_2)]^2 \rangle.
\end{equation}
According to the statistical descriptions presented in
LP16,
we adopt a power-law model of CF and SF, which is adequate for characterizing the scaling properties of turbulence.
Their forms are
\begin{subequations}\label{eq: cfsflp}
\begin{align}
& \widetilde{\xi}_\phi (\bm{R},\Delta z) = \sigma_\phi^2 \frac{r_\phi^{m_\phi}}{r_\phi^{m_\phi} + (R^2 + \Delta z^2)^{m_\phi/2}}, \label{eq: molcf}\\
& \widetilde{d}_\phi (\bm{R},\Delta z) = 2 \sigma_\phi^2 \frac{(R^2 + \Delta z^2)^{m_\phi/2}}{r_\phi^{m_\phi} + (R^2 + \Delta z^2)^{m_\phi/2}},
\end{align}
\end{subequations}
with the variance of fluctuations defined as
\begin{equation}
\sigma_\phi^2 = 0.81^2 \langle \delta(n_e B_z)^2 \rangle
\end{equation}
for RM density, and
\begin{equation}
\sigma_\phi^2 = \langle \delta(n_e^2)^2 \rangle
\end{equation}
for EM density,
where $r_\phi$ is the correlation scale of RM (EM) density fluctuations, and
$m_\phi$ is the index of their power-law functions in real space.
Under the condition of statistical homogeneity,
we see from above expressions that CF and SF only depend on the relative separation distance
instead of the separation vector between the two points.
The power spectrum in Fourier space $E(k) \sim k ^\alpha$ is complementary to CF and SF.
Since CF and SF respectively apply to small-scale and large-scale dominated statistics,
a shallow ($\alpha>-3$) spectrum is more properly described by CF, while a steep ($\alpha <-3$) spectrum more favors SF treatment.
Only when the cutoffs of CF at small scales and SF at large scales are both defined, can CF and SF be related and employed simultaneously
(see detailed discussions in \citealt{LP04,LP06}).
From Eq. \eqref{eq: cfsflp}, we find
\begin{equation}
\begin{aligned}
& \widetilde{\xi}_\phi (0) = \sigma_\phi^2, \\
& \widetilde{d}_\phi (\infty) = 2 \sigma_\phi^2,
\end{aligned}
\end{equation}
and thus the CF and SF are related by
\begin{equation}
\begin{aligned}
& \widetilde{d}_\phi (\bm{R},\Delta z) = 2[\widetilde{\xi}_\phi (0) - \widetilde{\xi}_\phi (\bm{R},\Delta z)], \\
& \widetilde{\xi}_\phi (\bm{R},\Delta z) = \frac{1}{2} [\widetilde{d}_\phi (\infty) - \widetilde{d}_\phi (\bm{R},\Delta z)].
\end{aligned}
\end{equation}
The relation between the CF (SF) index $m_\phi$ and the spectral index $\alpha$ depends on whether the turbulent spectrum is shallow or steep
\citep{LP06},
\begin{subnumcases}
{\alpha = \label{eq: malp}}
m_\phi-N, ~~~~~~ \alpha >-3, \label{eq: alaslw} \\
-m_\phi-N, ~~~\alpha <-3,
\end{subnumcases}
where $N$ is the dimensionality of space.
In the case of three-dimensional Kolmogorov turbulence with $\alpha = -11/3$, the corresponding value of $m_\phi$ is $2/3$.
Regarding the correlation scale $r_\phi$, as pointed out by LP16,
it corresponds to the energy dissipation scale for a shallow spectrum at wavenumbers smaller than $1/r_\phi$, and
the injection scale of turbulent energy for a steep spectrum at wavenumbers larger than $1/r_\phi$.
Strictly speaking, the forms of CF and SF given in Eq. \eqref{eq: cfsflp} are only applicable in the inertial range of the spectrum.
For scales below the inner scale $r_\phi$ of a shallow spectrum and above the outer scale $r_\phi$ of a steep spectrum,
the exact forms of CF and SF depend on the specific dissipation and injection processes of turbulent energy.
The total RM and EM are the integrals of RM and EM densities over the path length along the LOS
and thus their SFs have a dependence on the integration path.
In a simple case where there is only one single Faraday rotating screen along the LOS with a thickness $L$, as derived in
LP16,
the SF for the fluctuations of RM (EM) is given by
\begin{equation}\label{eq: sflp}
D_{\Delta{\Phi}}(\bm{R},L,L) = 2 \int_0^L d \Delta z (L-\Delta z) [\widetilde{\xi}_\phi(0,\Delta z)-\widetilde{\xi}_\phi (\bm{R},\Delta z)].
\end{equation}
Notice that the SF for RM (EM) has the same form as above due to
$\widetilde{\xi}_\phi(0,\Delta z)-\widetilde{\xi}_\phi (\bm{R},\Delta z) =\xi_\phi(0,\Delta z)-\xi_\phi (\bm{R},\Delta z)$.
A more complicated expression of $D_{\Delta{\Phi}}$ applicable to the situation with the synchrotron radiation and
Faraday rotation taking place in the same volume is available in
LP16.
We consider a thick Faraday screen with $L>r_\phi$, which is common to extragalactic sources.
After inserting Eq. \eqref{eq: molcf},
the SF from Eq. \eqref{eq: sflp} has asymptotic expressions in different ranges of $R$
(see Appendix C in LP16),
\begin{subnumcases}
{ D_{\Delta{\Phi}}(R) = \label{eq: appc}}
2 \sigma_\phi^2 LR \Big(\frac{R}{r_\phi}\Big)^{m_\phi}, ~~~~~~R<r_\phi<L,\\
2 \sigma_\phi^2 LR \Big(\frac{R}{r_\phi}\Big)^{-m_\phi}, ~~~~ r_\phi<R<L,\\
2 \sigma_\phi^2 L^2 \Big(\frac{L}{r_\phi}\Big)^{-m_\phi}, ~~~~~ r_\phi<L<R.
\end{subnumcases}
Regarding its dependence on $R$, the slope changes from $1+m_\phi$ to $1-m_\phi$ when $R$ reaches the correlation scale $r_\phi$,
and flattens when $R$ exceeds $L$.
It is necessary to point out that, as we discussed above,
the expression of $D_{\Delta{\Phi}}(R)$ at $R<r_\phi$ in the case of a shallow spectrum of RM (EM) density fluctuations
and those at $R>r_\phi$ in the case of a steep spectrum are not robust.
In the following calculations
we assume that the same power-law model of $\widetilde{\xi}_\phi (\bm{R},\Delta z)$ can be extensively applied beyond the inertial range of turbulence,
keeping in mind that $m_\phi$ has an adjustable value in different ranges of scales,
and we will discuss the modifications in a more realistic situation in Section \ref{sec: comobs}.
For an observer sitting in the Galaxy, if the Faraday rotation effect for the extragalactic sources mainly arises from the Galaxy,
the angular separation $\theta$ between a pair of LOSs through the ISM
coincides with the ratio of the projected distance and
geometrical depth of the Galactic Faraday material, i.e., $R/L$.
Thus, Eq. \eqref{eq: appc} can be recast into
\begin{subnumcases}
{ D_{\Delta{\Phi}}(\theta) = \label{eq: comsfthe}}
2 \sigma_\phi^2 L^2 \Big(\frac{L}{r_\phi}\Big)^{m_\phi}\theta^{1+m_\phi}, \label{eq: comrmsa} \nonumber \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\theta<\frac{r_\phi}{L}<1, \\
2 \sigma_\phi^2 L^2 \Big(\frac{L}{r_\phi}\Big)^{-m_\phi} \theta^{1-m_\phi}, \label{eq: comrmla} \nonumber \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ \frac{r_\phi}{L}<\theta<1, \\
2 \sigma_\phi^2 L^2 \Big(\frac{L}{r_\phi}\Big)^{-m_\phi}, \label{eq: comrmlla} \nonumber \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ \theta>1>\frac{r_\phi}{L}.
\end{subnumcases}
If the underlying turbulence conforms to the Kolmogorov scaling, there is $m_\phi = 2/3$, so we expect that the SF exhibits a
broken slope changing from $5/3$ at $\theta < r_\phi/L$ to $1/3$ at $r_\phi/L < \theta <1$,
and remains unchanged at $\theta>1$.
The above conversion from the linear scale $R$ to angular scale $\theta$ should be adjusted
when the LOSs intersect a Faraday screen at a distance from the observer.
The observed angular separation of a source pair becomes
\begin{equation}
\theta = \frac{R}{L+L_f},
\end{equation}
where $L_f$ is the distance from the Faraday screen to the observer.
By substituting
\begin{equation}
L^\prime = L + L_f, ~~\zeta =\frac{L}{L^\prime},
\end{equation}
and inserting
\begin{equation}
R = L^\prime \theta,
\end{equation}
Eq. \eqref{eq: appc} is reformulated as
\begin{subnumcases}
{ D_{\Delta{\Phi}}(\theta) = \label{eq: extrintr}}
2 \sigma_\phi^2 \zeta {L^\prime}^2 \Big(\frac{L^\prime}{r_\phi}\Big)^{m_\phi}\theta^{1+m_\phi}, \nonumber \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\theta<\frac{r_\phi}{L^\prime}<\zeta, \\
2 \sigma_\phi^2 \zeta {L^\prime}^2 \Big(\frac{L^\prime}{r_\phi}\Big)^{-m_\phi} \theta^{1-m_\phi}, \nonumber \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ \frac{r_\phi}{L^\prime}<\theta<\zeta, \\
2 \sigma_\phi^2 L^2 \Big(\frac{L}{r_\phi}\Big)^{-m_\phi}, \nonumber \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ \theta>\zeta>\frac{r_\phi}{L^\prime}.
\end{subnumcases}
The actual value of $\zeta$ can be evaluated from observations as the angular scale beyond which
$D_{\Delta{\Phi}}(\theta)$ has a zero slope.
We notice that Eq. \eqref{eq: comsfthe} corresponds to the specialization of the above equation at $\zeta =1$,
applicable to both extragalactic sources with little internal Faraday rotation and Galactic sources.
In the opposite limit, for an extended extragalactic source with high internal Faraday rotation
(within which the multiple RM components are correlated)
and relatively negligible Galactic contribution,
$\zeta$ can be much less than unity,
from which and the estimated size of the radiation emitting region ($\sim L$),
the location of the Faraday screen, i.e., the distance of the source, can potentially be obtained.
But such an observation requires very high angular resolution due to the large distance of the extragalactic source.
Besides the case with a single thick Faraday screen, which is the focus of this paper,
analyses for other realizations with, e.g., a thin Faraday screen ($L<r_\phi$) or a single LOS ($R=0$) are also
provided in
LP16,
which can be widely applied to different observational situations.
\subsection{SFs with separate contributions from electron density and magnetic field} \label{sec: thw}
The above approach straightforwardly reveals the dependence of the
slope and amplitude of the RM SF on the spectral characteristics of RM density fluctuations.
But it has the disadvantage that from the observed RM SF,
the relative significance between density and magnetic field fluctuations cannot be readily discerned.
We next carry out an analogous derivation of SFs of RM and EM as above, but
separate the contributions from density and magnetic field.
Similarly, we assume that the electron density and LOS component of magnetic field are described by
\begin{equation}
n_e= n_{e0} + \delta n_e, ~~ B_z = B_{z0} + \delta B_z,
\end{equation}
and
\begin{equation}
\langle \delta n_e \rangle =0, ~~ \langle \delta B_z \rangle =0,
\end{equation}
such that we have the product $n_e B_z$
\begin{equation}
\begin{aligned}
n_e B_z &= (n_{e0} + \delta n_e ) (B_{z0} + \delta B_z ) \\
&= n_{e0} B_{z0} + n_{e0} \delta B_z + B_{z0} \delta n_e + \delta n_e \delta B_z,
\end{aligned}
\end{equation}
and the squared $n_e$
\begin{equation}
n_e^2 = n_{e0}^2 + \delta n_e ^2 + 2n_{e0} \delta n_e .
\end{equation}
It follows that the CFs of RM and EM densities become (see detailed calculations given in the Appendix)
\begin{equation}\label{eq: seprm}
\xi_\phi(\text{RM}) = 0.81^2 (n_{e0}^2 B_{z0}^2 + n_{e0}^2 \widetilde{\xi}_B + B_{z0}^2 \widetilde{\xi}_n + \widetilde{\xi}_{n} \widetilde{\xi}_B),
\end{equation}
and
\begin{equation}\label{eq: sepem}
\xi_\phi(\text{EM}) = (n_{e0}^2 + \langle \delta n_e^2 \rangle)^2 + 2 \widetilde{\xi}_n^2 + 4n_{e0}^2 \widetilde{\xi}_n .
\end{equation}
Here $ \widetilde{\xi}_n$ and $ \widetilde{\xi}_B$ are CFs for fluctuations in $n_e$ and $B_z$.
By adopting the same model of CF as introduced in Eq. \eqref{eq: molcf}, we have
\begin{equation}
\begin{aligned}
\widetilde{\xi}_n (\bm{R},\Delta z) &= \langle \delta n_e (\bm{X_1}, z_1)\delta n_e (\bm{X_2}, z_2) \rangle \\
& =\sigma_n^2 \frac{r_n^{m_n}}{r_n^{m_n} + (R^2 + \Delta z^2)^{m_n/2}}, \\
\widetilde{\xi}_B (\bm{R},\Delta z) &=\langle \delta B_z (\bm{X_1}, z_1)\delta B_z (\bm{X_2}, z_2) \rangle \\
& =\sigma_B^2 \frac{r_B^{m_B}}{r_B^{m_B} + (R^2 + \Delta z^2)^{m_B/2}},
\end{aligned}
\end{equation}
with their respective correlation lengths $r_n, r_B$, power-law indices
$m_n, m_B$, and variances of fluctuations
\begin{equation}
\sigma_n^2 = \langle \delta n_e^2 \rangle, ~~~
\sigma_B^2 = \langle \delta B_z^2 \rangle.
\end{equation}
Under the condition of relatively small fluctuations, the linear terms in $\xi_\phi(\text{RM})$ (Eq. \eqref{eq: seprm}) and
$\xi_\phi(\text{EM})$ (Eq. \eqref{eq: sepem}) play a dominant role in determining the resultant SFs of RM and EM.
Combining Eq. \eqref{eq: sflp} with Eq. \eqref{eq: seprm} and \eqref{eq: sepem}, we obtain analytical estimates,
\begin{equation}\label{eq: drmanaes}
\begin{aligned}
& D_\text{RM} \approx \\
& 2 \times 0.81^2 n_{e0}^2 \int_0^L d \Delta z (L-\Delta z) [\widetilde{\xi}_B(0,\Delta z)-\widetilde{\xi}_B (\bm{R},\Delta z)] \\
& + 2 \times 0.81^2 B_{z0}^2 \int_0^L d \Delta z (L-\Delta z) [\widetilde{\xi}_n(0,\Delta z)-\widetilde{\xi}_n (\bm{R},\Delta z)] ,
\end{aligned}
\end{equation}
and
\begin{equation}\label{eq: sepsfem}
D_\text{EM} \approx 2 \times 4n_{e0}^2 \int_0^L d\Delta z (L-\Delta z) \Big[ \widetilde{\xi}_n(0,\Delta z) - \widetilde{\xi}_n(\bm{R},\Delta z)\Big] .
\end{equation}
By comparing Eq. \eqref{eq: sflp}, \eqref{eq: comsfthe}, and \eqref{eq: drmanaes},
the simplified expressions of $D_\text{RM}$ in different asymptotic regimes can then be derived:
\begin{subnumcases}
{ D_\text{RM} (\theta)= \label{eq: msclex}}
2 \times 0.81^2 L^2 \Big [n_{e0}^2 \sigma_B^2 \Big(\frac{L}{r_B}\Big)^{m_B}\theta^{1+m_B} \nonumber \\
~~~~~~~~~~~~~~~~~+ B_{z0}^2 \sigma_n^2 \Big(\frac{L}{r_n}\Big)^{m_n}\theta^{1+m_n} \Big ], \nonumber \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~\theta<\frac{r_n}{L}<\frac{r_B}{L}<1, \\
2 \times 0.81^2 L^2 \Big [n_{e0}^2 \sigma_B^2 \Big(\frac{L}{r_B}\Big)^{m_B}\theta^{1+m_B} \nonumber \\
~~~~~~~~~~~~~~~~~+ B_{z0}^2 \sigma_n^2 \Big(\frac{L}{r_n}\Big)^{-m_n} \theta^{1-m_n} \Big], \nonumber \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~\frac{r_n}{L}<\theta<\frac{r_B}{L}<1, \label{eq: sepdrmb} \\
2 \times 0.81^2 L^2 \Big [n_{e0}^2 \sigma_B^2 \Big(\frac{L}{r_B}\Big)^{-m_B} \theta^{1-m_B} \nonumber \\
~~~~~~~~~~~~~~~~~ + B_{z0}^2 \sigma_n^2 \Big(\frac{L}{r_n}\Big)^{-m_n} \theta^{1-m_n} \Big], \nonumber \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\frac{r_B}{L}<\theta<1, \\
2 \times 0.81^2 L^2 \Big [n_{e0}^2 \sigma_B^2 \Big(\frac{L}{r_B}\Big)^{-m_B} \nonumber \\
~~~~~~~~~~~~~~~~~+ B_{z0}^2 \sigma_n^2 \Big(\frac{L}{r_n}\Big)^{-m_n} \Big], \nonumber \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ \theta>1>\frac{r_B}{L}.
\end{subnumcases}
In the above expression, we consider the condition $r_n < r_B$. The opposite case with $r_B<r_n$ can be similarly formulated:
\begin{subnumcases}
{ D_\text{RM} (\theta)= \label{eq: msclex2}}
2 \times 0.81^2 L^2 \Big [n_{e0}^2 \sigma_B^2 \Big(\frac{L}{r_B}\Big)^{m_B}\theta^{1+m_B} \nonumber \\
~~~~~~~~~~~~~~~~~+ B_{z0}^2 \sigma_n^2 \Big(\frac{L}{r_n}\Big)^{m_n}\theta^{1+m_n} \Big ], \nonumber \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~\theta<\frac{r_B}{L}<\frac{r_n}{L}<1, \\
2 \times 0.81^2 L^2 \Big [n_{e0}^2 \sigma_B^2 \Big(\frac{L}{r_B}\Big)^{-m_B} \theta^{1-m_B} \nonumber \\
~~~~~~~~~~~~~~~~~+ B_{z0}^2 \sigma_n^2 \Big(\frac{L}{r_n}\Big)^{m_n}\theta^{1+m_n} \Big], \nonumber \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~\frac{r_B}{L}<\theta<\frac{r_n}{L}<1, \\
2 \times 0.81^2 L^2 \Big [n_{e0}^2 \sigma_B^2 \Big(\frac{L}{r_B}\Big)^{-m_B} \theta^{1-m_B} \nonumber \\
~~~~~~~~~~~~~~~~~ + B_{z0}^2 \sigma_n^2 \Big(\frac{L}{r_n}\Big)^{-m_n} \theta^{1-m_n} \Big], \nonumber \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\frac{r_n}{L}<\theta<1, \\
2 \times 0.81^2 L^2 \Big [n_{e0}^2 \sigma_B^2 \Big(\frac{L}{r_B}\Big)^{-m_B} \nonumber \\
~~~~~~~~~~~~~~~~~+ B_{z0}^2 \sigma_n^2 \Big(\frac{L}{r_n}\Big)^{-m_n} \Big], \nonumber \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ \theta>1>\frac{r_n}{L}.
\end{subnumcases}
It reveals that the total $D_\text{RM}$ is the superposition of two components that are related to the fluctuations in electron density and the
LOS component of magnetic field, respectively.
In an exceptional situation with $r_n = r_B$ and $m_n = m_B$, namely, density and magnetic turbulence
share the same spectral scaling law,
the expression of $D_\text{RM} (\theta)$ in Eq. \eqref{eq: comsfthe} can be recovered from either
Eq. \eqref{eq: msclex} or Eq. \eqref{eq: msclex2}, and accordingly the CF parameters of RM density fluctuations can be more explicitly written as
\begin{equation}
\begin{aligned}
& \sigma_\phi^2 = 0.81^2 (n_{e0}^2 \sigma_B^2 + B_{z0}^2 \sigma_n^2 ), \\
& r_\phi = r_n = r_B, \\
& m_\phi = m_n = m_B.
\end{aligned}
\end{equation}
This is only valid under rather restrictive circumstances.
More generally, one would expect that the spectra for fluctuating density and magnetic field are not aligned and
the behavior of $D_\text{RM}$ depends on the relative importance between $B_{z0}^2 \sigma_n^2$ and $n_{e0}^2 \sigma_B^2$.
For example, when
$B_{z0}^2 \sigma_n^2$ is significantly larger than $n_{e0}^2 \sigma_B^2$, that is, the relative density fluctuations are
much stronger in comparison with relative magnetic field fluctuations,
\begin{equation}\label{eq: relimp}
\frac{\sigma_n^2}{n_{e0}^2} \gg \frac{\sigma_B^2}{B_{z0}^2},
\end{equation}
density fluctuations dictate the behavior of $D_\text{RM}$.
Both Eq. \eqref{eq: msclex} and \eqref{eq: msclex2} approximately go back to Eq. \eqref{eq: comsfthe}, and
the CF parameters of RM density fluctuations are equivalent to
\begin{equation}\label{eq: rmexp}
\begin{aligned}
& \sigma_\phi^2 = 0.81^2 B_{z0}^2 \sigma_n^2 , \\
& r_\phi = r_n, \\
& m_\phi = m_n.
\end{aligned}
\end{equation}
In this situation, magnetic field fluctuations are basically not responsible for the observed RM SFs.
As regards the SF of EM, combining Eq. \eqref{eq: sflp}, \eqref{eq: comsfthe}, and \eqref{eq: sepsfem} yields
\begin{subnumcases}
{ D_\text{EM}(\theta) = }
8n_{e0}^2 \sigma_n^2 L^2 \Big(\frac{L}{r_n}\Big)^{m_n}\theta^{1+m_n}, \nonumber \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\theta<\frac{r_n}{L}<1,\\
8n_{e0}^2 \sigma_n^2 L^2 \Big(\frac{L}{r_n}\Big)^{-m_n} \theta^{1-m_n}, \nonumber \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ \frac{r_n}{L}<\theta<1, \label{eq: demlang}\\
8n_{e0}^2 \sigma_n^2 L^2 \Big(\frac{L}{r_n}\Big)^{-m_n}, \nonumber \\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ \theta>1>\frac{r_n}{L}.
\end{subnumcases}
The similarity between the form of the above $D_\text{EM}(\theta)$ and Eq. \eqref{eq: comsfthe} suggests that the EM density fluctuations
inherit the turbulence properties from the density field and their CF parameters are related as
\begin{equation}\label{eq: emdd}
\begin{aligned}
& \sigma_\phi^2 = 4n_{e0}^2 \sigma_n^2 , \\
& r_\phi = r_n, \\
& m_\phi = m_n.
\end{aligned}
\end{equation}
If the situation described in Eq. \eqref{eq: relimp} is realized, we expect that
$ D_\text{RM}(\theta)$ and $D_\text{EM}(\theta)$ measured from the same region in the sky
exhibit a similar behavior in terms of the
spectral slope and correlation scale corresponding to the break in slope. Their ratio
\begin{equation}
\frac{D_\text{RM} (\theta)}{D_\text{EM}(\theta) } =\Big(\frac{0.81 B_{z0}}{2 n_{e0}}\Big)^2
\end{equation}
is associated with the mean plasma properties and determined by the ratio between the mean values of $B_z$ and $n_e$.
\section{Comparison with observations}\label{sec: comobs}
Observationally determined SFs of RMs allow a quantitative test of the above analysis.
\citet{MS96} studied both RM and EM measurements for 38 extragalactic sources.
The sample of source components are selected to have insignificant intrasource variations in RM and depolarization effects
for the purpose of studying the Faraday rotation of the Galactic medium only.
Another strong argument that the observed RM is dominated by the Galactic Faraday rotation is the evident dependence on the angular separation
of the RM SF.
The measured RM SF $D_\text{RM}$ increases as a power-law-like function of angular scale and is characterized by a break in its spectrum
around $0.1^\circ$.
The least-squares fit to the data above $0.1^\circ$ gives
\begin{equation} \label{eq: obsdrm}
D_\text{RM} (\theta > 0.1^\circ) = (340 \pm 30) ~ [\theta(^\circ)]^{0.64\pm 0.06} ~\text{rad}^2 \text{m}^{-4}.
\end{equation}
The three data points below $0.1^\circ$ are insufficient for achieving a reliable fitting,
but they clearly indicate a more steepened spectrum.
These observed features of RM SFs appear to be consistent with the theoretical expectations presented in Section 2.
To enable a quantitative comparison, we employ the parameters provided in
\citet{MS96}.
Since the observed Faraday rotation for their sample is dominated by the magnetized ISM of our Galaxy,
we adopt the average path length $2900$ pc through the ISM in the observed region as the depth $L$ of the Faraday screen.
Besides,
the very few measurements at small angular separations do not allow an accurate determination of the break point in the SF spectrum.
We then choose the ``outer scale'' of turbulence of $3.6$ pc suggested in
\citet{MS96}
as an experimental correlation scale $r_\phi$.
The transition angle at the break is thus
\begin{equation}
\theta_\text{tr} = \frac{r_\phi} {L}= 0.0711^\circ.
\end{equation}
Given the coincidence $\theta=R/L$ for this observation, and the determination of $L$ and $r_\phi$,
we can compare Eq. \eqref{eq: obsdrm} with the functional form of $D_\text{RM}$ at $\theta > \theta_\text{tr}$ from Eq. \eqref{eq: comrmla},
which directly yields the CF index and variance of RM density fluctuations,
\begin{equation}\label{eq: commsp}
m_\phi = 1-0.64 = 0.36, ~~~
\sigma_\phi = 5.5 \times 10^{-2} ~ \text{rad}~ \text{m}^{-2} \text{pc}^{-1}.
\end{equation}
Evidently, the measured $m_\phi$ does not coincide with the prediction by Kolmogorov turbulence, which entails
$m_\phi = 2/3$ instead (see Section \ref{sec: lp}).
In fact, since the range $R>r_\phi$ (i.e., $\theta > \theta_\text{tr}$) corresponds to the inertial range of a shallow spectrum of RM density fluctuations,
the resulting spectral index is (Eq. \eqref{eq: alaslw})
\begin{equation}\label{eq: msspesl}
\alpha = m_\phi -3 = 0.36-3 = -2.64.
\end{equation}
Meanwhile, the identification of the inertial range over the scales larger than $r_\phi$
indicates that $r_\phi$ is actually the inner scale rather than the outer scale of turbulence.
Below $r_\phi$, the damping effect efficiently suppresses the fluctuations
and steepens the spectral tail in the dissipation range.
If we assume that the same scaling as in the inertial range can still be used at scales below but in the vicinity of $r_\phi$,
inserting the values in Eq. \eqref{eq: commsp} into the expression of $D_\text{RM}$ at $\theta < \theta_\text{tr}$
from Eq. \eqref{eq: comrmsa} leads to
\begin{equation}\label{eq: drmtos}
D_\text{RM} (\theta < \theta_\text{tr})= 2.28\times10^3~ [\theta(^\circ)]^{1.36} \text{rad}^2 ~\text{m}^{-4}.
\end{equation}
Fig. \ref{figms96} plots $D_\text{RM}$ from Eq. \eqref{eq: drmtos} at $\theta < \theta_\text{tr}$ and Eq. \eqref{eq: obsdrm} at
$\theta > \theta_\text{tr}$, superposed with the observational data points taken from figure 5 in
\citet{MS96}.
It seems that both the amplitude and spectral slope of $D_\text{RM}$ given by Eq. \eqref{eq: drmtos} are in good agreement with the observational result.
However, due to the limited number of close source pairs,
it is difficult to impose a strong constraint on the exact spectral slope in the
dissipation range of turbulence.
Nevertheless, the comparison illustrates that the theoretical model originally constructed by LP16
can satisfactorily interpret the observed RM SFs.
We exclude the possibility that $r_\phi$ is the outer scale of a steep spectrum, as in this case one would expect that the SF saturates at
a constant value and flattens
at $\theta > \theta_\text{tr}$.
Otherwise, the observed SF spectrum over an extended range of angular scales beyond $\theta_\text{tr}$
severely challenges the model for the energy injection of turbulence.
\begin{figure}[htbp]
\centering
\includegraphics[width=9cm]{MS96.eps}
\caption{ SFs of RMs vs. angular separation. The data points are taken from figure 5 in
\citet{MS96}. The solid line corresponds to the theoretical formalisms Eq. \eqref{eq: comrmsa} and \eqref{eq: comrmla}. }
\label{figms96}
\end{figure}
From RM SFs alone, we are unable to identify the separate contributions due to density and magnetic field fluctuations.
Hence, it is necessary to invoke SFs of EMs for an exclusive extraction of the electron density fluctuations.
We now turn to the observationally measured EM SFs.
A power-law representation with a similar slope to that of RM SFs
fits the observations
\citep{MS96},
\begin{equation}\label{eq: obsem}
D_\text{EM} (\theta) = (5.5 \pm 0.6) ~ [\theta(^\circ)]^{0.73\pm 0.08} \text{cm}^{-12} \text{pc}^2.
\end{equation}
Due to the lower angular resolution of the EM data, the measurements of $D_\text{EM}$ at small angular scales are absent,
which prevents the detection of the possible break in its spectrum and determination of the correlation scale of EM density fluctuations.
But the fitting by Eq. \eqref{eq: obsem} is informative on the CF index of EM density fluctuations, which is also approximately the CF index of
density fluctuations (see Eq. \eqref{eq: emdd}).
Comparing Eq. \eqref{eq: comrmla} and \eqref{eq: obsem} gives the value
\begin{equation}
m_\phi = 1-0.73(\pm 0.08) = 0.27(\pm 0.08),
\end{equation}
close to that of RM density fluctuations (Eq. \eqref{eq: commsp}).
The similarity between the slopes of RM and EM SFs strongly
suggests that the density fluctuations take a major part in composing the SF of RM.
Presumably, the CF index and correlation scale obtained from RM SFs also match those quantities of density fluctuations
(see Eq. \eqref{eq: rmexp}),
\begin{equation}\label{eq: presmr}
m_n = 0.36, ~~~r_n = 3.6 ~\text{pc},
\end{equation}
and the EM SFs at $\theta>\theta_\text{tr}$ satisfy
\begin{equation}
D_\text{EM} (\theta) = 5.5 ~ [\theta(^\circ)]^{0.64} \text{cm}^{-12} \text{pc}^2.
\end{equation}
By comparing the above equation with Eq. \eqref{eq: demlang} and using $L= 2900$ pc and values in Eq. \eqref{eq: presmr},
we find
\begin{equation}
n_{e0}\sigma_n = 3.47\times10^9 \text{m}^{-6}.
\end{equation}
Furthermore, under the condition of Eq. \eqref{eq: relimp},
we can safely neglect the terms associated with magnetic fluctuations in Eq. \eqref{eq: msclex} and \eqref{eq: msclex2}
due to their minor contribution.
From Eq. \eqref{eq: rmexp} and \eqref{eq: commsp}, we get
\begin{equation}
0.81 B_{z0} \sigma_n = \sigma_\phi = 5.5 \times 10^{-2} ~ \text{rad}~ \text{m}^{-2} \text{pc}^{-1},
\end{equation}
indicative of the fact that the level of RM density fluctuations is determined by the joint strength of fluctuating
electron density and mean magnetic field along the LOS.
Given the estimate of mean electron density $n_{e0}$, $\sigma_n$ and $B_{z0}$ can be both derived from the above two equations.
However, the properties of the fluctuating component of magnetic field are poorly constrained by SFs of RMs.
The above results suggest that the observed RM SFs
originate from the underlying shallow spectrum of density fluctuations, with the break in the slope of the SF
corresponding to the inner scale of the density spectrum.
Accordingly, the Faraday rotating medium that gives rise to the characteristics of RM SFs
has an excess of dense structures at small scales comparable to $r_\phi$.
As a caveat to the applicability of the theoretical model,
the uncertainties in the amplitude and correlation scale of turbulence obtained from observations
are introduced by the choice of Faraday rotation depth $L$.
The RM fluctuations traced by Galactic sources provide information about the small-scale turbulence within a specific region in
the Galaxy.
The Faraday rotating medium extends from the observer to sources and thus the sources' distances can be used
as the estimate of $L$.
Observations of extragalactic sources with marginal intrinsic Faraday rotation as in
\citet{MS96}
can bring forth large-scale features of ISM turbulence.
The depth of the Faraday screen is the path length throughout the Galaxy, and the sources' distances are irrelevant.
But in the case with dominant intrinsic Faraday rotation,
provided that the multiple RM components within one source can be resolved,
the properties of turbulence in the source region are probed.
Apart from $L$, the source's distance is also involved in the analysis
(see Eq. \eqref{eq: extrintr}).
\section{Conclusions and discussion}
Following the model SF of RM fluctuations put forth by
LP16,
we proceed to carry out the SF analysis of RM and EM by separating the contributions from
fluctuations of electron density and the LOS component of magnetic field, and we
assess their relative importance in determining the form of SFs of RMs as a function of angular scales.
We found that the SF of RM can be considered as a sum of two SFs stemming from the fluctuations of electron density and magnetic field, respectively.
The turbulent spectrum of the density fluctuations can be extracted from the SF measurement of EM
provided that the angular resolution is sufficient.
Applying the analysis to observationally determined SFs of RMs shows that
the model SF can consistently interpret the shape of SFs on different ranges of angular scales.
Similar behavior of RM and EM SFs suggests that the observed SF of RM is mainly guided by fluctuating electron density and thus
incapable of probing the nature of turbulent magnetic field.
As an observational fact,
the changing shape of SFs of RMs with angular scales was earlier interpreted as arising from a transition from three-dimensional to
two-dimensional turbulence
\citep{MS96},
or two spatially separate Faraday screens
\citep{Hav04}.
Without the intervention of two turbulent structures or two Faraday rotation regions,
the power-law CF for RM density
fluctuations introduced in
LP16 can naturally reproduce the observed features and explain the broken power-law shape of the RM SF.
In earlier works, magnetic field is assumed to be frozen in matter, and thus both density and magnetic field fluctuations
adhere to an identical power-law spectrum with the same spectral index and inner and outer scales.
But this conjecture has been definitely rejected by observations on RM fluctuations.
The Kolmogorov spectrum appears to be too steep to account for the shallow slope especially at scales larger than the
spectral break, and it alone cannot serve as a satisfactory turbulent model to fit the
various slopes of SFs from different observations.
In fact, it has been known that the conventional flux-freezing concept breaks down in realistic MHD turbulence as
the diffusion of magnetic field lines is mediated by
fast magnetic reconnection, which is an intrinsic process inherent in MHD turbulence
\citep{Laz05, Laz11,LEC12, EyNa13}.
It is more plausible that density and magnetic field fluctuations conform to distinct power spectra of turbulence.
The dominant one between them is more important in determining the shape of the resultant SF spectrum of RM.
The similarity between the behavior of RM and EM SFs revealed by observations
(e.g. \citealt{MS96, Hav04})
indicates that the major contribution to the measured RM SFs comes from electron density fluctuations,
which tend to follow a shallow spectrum of turbulence down to the dissipation scale corresponding to the spectral break of
the RM SF.
Both theoretical considerations and numerical simulations suggest that a shallow spectrum of density field can arise in
compressible turbulent flows.
Compressibility leads to the formation of clumpy density structures, with condensations embedded in
relatively diffuse regions
\citep{BLC05, Kri07, KL07,Fal14}.
The coupling between this density structure and local turbulent motions results in a steeper
velocity power spectrum with a slope of $\sim -2$, but a much shallower power spectrum of the density field than
the Kolmogorov $-5/3$ scaling for one-dimensional spectra.
Moreover,
the shallowness of the spectral slope of density fluctuations is strongly affected by magnetic field strength in subsonic turbulence
and by sonic Mach number $M_s$ in supersonic turbulence
\citep{KL07}.
For example,
\citet{KL07}
observed a slope of $\sim -0.5$ for the one-dimensional density spectrum obtained from the simulated
supersonic turbulence with $M_s = 7$, which is even shallower than the spectral slope indicated from the
RM SFs measured by
\citet{MS96} (see Eq. \eqref{eq: msspesl}).
The ISM is highly inhomogeneous with dense structures accumulated by shocks on small scales.
Depending on the local compressibility,
the density fluctuations and the resultant RM SFs can have spatially diverse spectral slopes.
In contrast to density, magnetic field is better coupled with turbulent velocity field and regulates the turbulence
properties, e.g., scale-dependent turbulent anisotropy, as the turbulent energy cascades down from large to small scales.
The ``Big Power Law in the Sky" extending from $10^{17}$ m to $10^6$ m
\citep{Armstrong95,CL10}
suggests that the interstellar turbulence has a Galactic-scale ($>100$ pc) energy injection source and cascades toward very small scales.
At the parsec scale of this global turbulent cascade, magnetic field fluctuations can have a lower
amplitude compared with the enhanced amplitude of density fluctuations. These excessive density fluctuations are
thus manifest by dominating the composite RM density fluctuations.
The candidate regions corresponding to the localized enhanced turbulence in density field can be
the extended envelopes of H {\small \uppercase\expandafter{\romannumeral2}} regions and the warm ionized medium of the
\citet{MO77}
model in the Galactic plane
\citep{Span91, Armstrong95}.
\citet{CL10} carried out a remarkable extension of the Big Power Law in the Sky in
\citet{Armstrong95}
up to $10^{17}$ m by using the data of the Wisconsin H$\alpha$ Mapper (WHAM)
and demonstrate a universal spectrum of interstellar density fluctuations consistent with Kolmogorov turbulence.
On the contrary, observations of RM fluctuations at low Galactic latitudes
(e.g. \citealt{Sim86, Cle92, Hav04,Hav08})
imply a shallower spectrum than the Kolmogorov one.
As a possible understanding of this contradiction,
for LOSs that traverse the supersonic turbulent flows through the Galactic plane,
the RM results are mainly governed by density fluctuations, which we do not expect to be compatible with the Kolmogorov spectrum
as they are more independent from turbulent velocity and sensitive to the local compressibility of the medium.
\citet{CL10} analyzed the WHAM data at high Galactic latitudes and
avoided the contamination from the H {\small \uppercase\expandafter{\romannumeral2}} regions in the Galactic plane.
As a result, they achieved one universal turbulent spectrum throughout the Galaxy.
Their result is also in agreement with the study of velocity turbulence by using the velocity coordinate spectrum technique in
\citet{CLS10}.
So great caution is needed when one uses RM fluctuations to probe the turbulent magnetic fields in the Galaxy, not only because
the fluctuating density instead of fluctuating magnetic field can dominate the observed RM SF behavior,
but also because more complexity can be introduced by additional structures of turbulence embedded in the observed region.
They both hinder the recovery of the underlying spectrum of magnetized turbulence.
We identify the transition scale on the order of $1$ pc at the spectral break of RM SFs as the inner scale of
a shallow electron density spectrum over larger scales, instead of the injection scale of RM fluctuations at smaller scales as determined by
\citet{Hav04,Hav08}.
In a partially ionized ISM, the cascade of MHD turbulence is severely damped due to ion-neutral collisions.
Below the scale where neutral fluid decouples from ion-electron fluid,
the MHD turbulence in ion-electron fluid is efficiently suppressed by ion-neutral collisional damping, but the cascade of hydrodynamic turbulence in neutrals proceeds
down to the viscous cutoff.
The dissipation scale inferred from RM SFs we obtained in this work is consistent with the ion-neutral collisional damping scale of MHD turbulence
in the warm neutral phase of the ISM calculated by
\citet{Xu15}.
Plausibly, this consistency could suggest that the inner scale to electron density fluctuations of $\sim 1$ pc characterizes the typical and also minimum
scale of discrete structures of excess electron densities in the observed region,
while smaller condensations mostly consist of neutrals and are driven by gravitational contraction instead of supersonic motions.
\\
\\
The authors thank the anonymous referee for the valuable comments.
This work is partially supported by the National Basic Research Program (973 Program) of China under grant No. 2014CB845800.
|
\section{Introduction} \label{sec1}
The shortest path length between two nodes is the number of edges in the shortest path between them. The distribution of shortest path lengths and the average shortest path length of the network are important measures of the network topology~\cite{Newman10,Dorogovtsev08} as they characterize the efficiency of various spreading processes on networks; the analysis of shortest path lengths is at the centre of the six degrees of separation and the small-world phenomena~\cite{Watts98}. The calculation of shortest path lengths have been also used for estimating the accuracy of analytical approximations for dynamics on networks~\cite{Melnik11}, examining the onset of synchronization~\cite{zhao06} and assessing the resilience of communication networks to attacks and failures~\cite{Albert00}.
Significant effort has been devoted to the development of efficient numerical algorithms for both the exact and approximate calculation of the intervertex distances on a given network (see, for example,~\cite{Zwick01} and related literature). Exact numerical calculation of the probability distribution of distances between a pair of randomly chosen nodes (which requires to solve all-pairs shortest path problem) using the well-known Dijkstra algorithm has running time $O(mN + N^2\log N)$, where $N$ is the number of nodes and $m$ is the number of edges of the graph. Although there has been continuous improvement, the numerical calculation of shortest path lengths in large networks will remain a computationally expensive task.
Relatively little attention has been paid to the analytical calculation of distances on random networks~\cite{Newman01a,Dorogovtsev03b, Fronczak04,Fronczak04b,Dorogovtsev06b,Dorogovtsev08b,Katzav15}. In this paper, we consider undirected unweighted random networks with prescribed degree distribution $p_k$ or joint degree-degree distribution $P(k,k')$~and propose a simple analytical method for calculating the probability distribution of shortest path lengths. Our method is more accurate than some other analytical methods for such classes of networks and allows us to predict (very often more accurately than any other know analytical method) shortest path lengths in real-world networks from their degree distribution $p_k$ or joint degree-degree distribution $P(k,k')$.
This paper is organized as follows. In Sec.~\ref{s:analogy}, we formulate the calculation of shortest path lengths in terms of a susceptible-infected epidemic model. In Sec.~\ref{s:z_regular}, we explain our analytical approach using $z$-regular random networks, and generalize it to random networks with arbitrary degree distribution or with degree-degree correlations in Secs.~\ref{pk_theory} and \ref{Pkk_theory} respectively. We apply our approach to calculate intervertex distances in real-world networks in Sec.~\ref{s:RWNs}, and conclude in Sec.~\ref{s:conclusions}.
\section{Analogy between intervertex distance and time to infection} \label{s:analogy}
Let $D_n$ be the probability that the length of the shortest path (the distance) between two randomly-chosen nodes is equal to $n$. To calculate it numerically for a given network one could go through all pairs of nodes, find the shortest path between each pair and build a histogram as shown in Fig.~\ref{f:Dn_APSP_Hist}.
\begin{figure}[b]
\centering
\includegraphics[width=0.7\columnwidth]{fig1.pdf}
\caption{(Color online) For a given network, the probability $D_n$ that two randomly chosen nodes are at a distance $n$ can be obtained numerically by calculating the shortest path length between each pair of nodes. For example, the shortest path between nodes $i$ and $j$ has length $4$ (blue links), so it contributes to $D_4$.}
\label{f:Dn_APSP_Hist}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=0.7\columnwidth]{fig2.pdf}
\caption{(Color online) (Left) A single node (the seed) is infected at Time 0, and at each subsequent time step, every infected node infects all its susceptible neighbors. In this process, the time when other nodes become infected gives the distance between those nodes and the seed. (Top right) Let $\rho_n$ be the expected fraction of infected nodes at time $n$ as we run this epidemic process for different seeds. Here, $\rho_1 = (\frac34+\frac34+\frac44+\frac24)/4 = 3/4$. (Bottom right) Importantly, $\rho_n$ determines the distribution of distances~between two randomly-chosen nodes: $D_n= \rho_n-\rho_{n-1}$.}
\label{f:SI_epidemic}
\end{figure}
To proceed with our analytical approach, let us first formulate the calculation of distribution of distances~in terms of the following susceptible-infected epidemic model. Let us start with a single infected node $i$ as the source of an epidemic. At each discrete time step $n$, let every infected node infect all of its susceptible neighbors. Infected nodes remain infected indefinitely.\footnote{One can think of this process as susceptible-infected epidemic with infection probability 1, or as the Watts threshold model~\cite{Watts02} with threshold $\phi=0$. In both cases the states of all nodes are updated at each discrete time step.} At step $n=1$, node $i$ infects all its immediate neighbors, at step $n=2$ the second neighbors of $i$ become infected, and so on. The process stops when all nodes in the network become infected, see Fig.~\ref{f:SI_epidemic}. Notice that the distribution of distances from node $i$ to other nodes is given by the distribution of times when other nodes became infected. For example, the nodes who became infected at time $2$ are exactly $2$ steps away from node $i$. Let $\rho^i_n$ be the fraction of network nodes infected at step $n$ when node $i$ is initially infected.
Repeat this epidemic process for every node $i$ in the network (i.e., each time starting from a different initially infected node $i$) as shown in Fig.~\ref{f:SI_epidemic}, and let $\rho_n \equiv \left<\rho^i_n\right>$ be the expected fraction of infected nodes at time $n$. Then the distribution of intervertex distances is given by the difference between $\rho_n$ and $\rho_{n-1}$
\begin{align}
\label{Dn}
D_n = \rho_n - \rho_{n-1},
\end{align}
and the average shortest path length is $\bar D=\sum_n n D_n$.
Therefore, if we analytically calculate the expected fraction of infected nodes at time $n$ for the above epidemic process, we can obtain the distribution of shortest path lengths. Below we explain the main idea of our approach using, for simplicity, regular random graphs (i.e., networks where every node has $z$ neighbors and connections between nodes are random), and in the subsequent sections we generalize the results to random networks with arbitrary degree distribution (the so-called configuration model networks) and to networks with degree-degree correlations.
\section{Explaining our approach using $z$-regular random graphs}\label{s:z_regular}
We consider a randomly chosen node $A$ and calculate its probability of being infected at time $n$ (i.e., after $n$ synchronous updates of states of all nodes). This probability is $\rho_n$ since we chose $A$ uniformly at random. Initially a single node $i$ in the network is infected, which implies that $A$ is initially infected with probability $\rho_0=1/N$, where $N$ is the total number of nodes, or initially susceptible with probability $1-\rho_0$.
We denote by $q_{n}$ the probability that at time $n$ (i.e., immediately before update $n+1$ of node $A$) a random neighbor $B$ of node $A$ is infected, conditioned on node $A$ itself being susceptible. This conditioning accounts for the fact that neighbors of $A$ become infected before $A$, and that $A$ did not infect them. Note that $A$ is susceptible only if it was initially susceptible and none of its neighbors have yet infected it. Thus, the probability that $A$ is infected at time $n+1$ is
\begin{align}
\label{RRG_rhon}
\rho_{n+1} = 1 -(1-\rho_0) (1-q_{n})^z,
\end{align}
where $(1-\rho_0)$ is the probability that $A$ is initially susceptible, and $(1-q_{n})^z$ is the probability that none of $z$ neighbors of $A$ is infected at time $n$.\footnote{In Eq.~\eqref{RRG_rhon}, we have assumed that the states of any two neighbors of node $A$ are independent, which is the case for a graph that is locally tree-like, such as random networks constructed using the configuration model~\cite{Newman10}. Although the tree-like assumption breaks down on real-world networks with high clustering coefficients and/or significant community structure, we have previously demonstrated that results obtained using locally tree-like approximations often remain reasonably accurate on such networks~\cite{Melnik11,Gleeson12}.}
\begin{figure}[h]
\centering
\includegraphics[width=0.8\columnwidth]{fig3.pdf}
\caption{(Color online) Tree-like structure of a 3-regular random network near node $A$. Since $A$ is a randomly chosen node, the probability that $A$ is infected at time $n+1$ is $\rho_{n+1}$. Here, $q_n$ is the probability that node $B$, a child of $A$, is infected by any of its children, given $A$ is susceptible. Similarly, $q_{n-1}$ is the probability that $C$ (a child of susceptible $B$) is infected by any of $C$'s own children.}
\label{f:tree_update}
\end{figure}
In order to calculate $q_n$, we consider node $B$, a neighbor of $A$, and establish a recurrence relation for $q_n$~\cite{Melnik13}. Using similar reasoning as for Eq.~\eqref{RRG_rhon}, we express the probability $q_n$ (that $B$ is infected, given $A$ is susceptible) in terms of probability $q_{n-1}$ that a child of $B$, node $C$, --- defined as a neighbor of $B$ that is one step further away from $A$ --- is infected given $B$ is susceptible, see Fig.~\ref{f:tree_update}:
\begin{align}
\label{RRG_qn}
q_{n} = 1-(1-\rho_0)(1-q_{n-1})^{z-1}.
\end{align}
The power $z-1$ in this equation appears instead of $z$ as in Eq.~\eqref{RRG_rhon} because, by the definition of $q_n$, node $A$ is susceptible and cannot infect $B$ and thus is excluded from consideration (likewise $B$ cannot infect its children, and so on).
Therefore, starting with $q_0 = \rho_0=1/N$, we can iterate Eqs.~\eqref{RRG_rhon}-\eqref{RRG_qn} to obtain values of $\rho_n$, and use them in Eq.~\eqref{Dn} to calculate the distribution of distances. In Fig.~\ref{f:rho_n_Dn_vs_n}, we show the values of $\rho_n$ calculated for a 4-regular random network with $N=500$ nodes, and the corresponding values of $D_n$. Observe the excellent match between numerical and theoretical results.
\begin{figure}[h]
\centering
\includegraphics[width=0.97\columnwidth]{fig4.pdf}
\caption{(Color online) (a) Time evolution of the fraction of infected nodes $\rho_n$ for the discrete-time susceptible-infected epidemic model described in the text, calculated by iterating Eqs.~\eqref{RRG_rhon}-\eqref{RRG_qn} starting with $q_0 = \rho_0 = 1/N$, for 4-regular random network with $N=500$ nodes. (b) Values of $\rho_n-\rho_{n-1}$ together with numerically calculated distribution of distances. Numerical results are averaged over 10 realizations of networks.}
\label{f:rho_n_Dn_vs_n}
\end{figure}
We can analytically solve Eqs.~\eqref{RRG_rhon}-\eqref{RRG_qn} (see Appendix~\ref{App:RRG}) and obtain an explicit formula for $D_n$ for $z$-regular networks:
\begin{align}
D_n^{\rm RRG} = \exp \left[ -\frac{z(z-1)^{n-1}-2}{(z-2)N} \right] - \exp \left[ -\frac{z(z-1)^n-2}{(z-2)N} \right].
\label{LargeNRRG}
\end{align}
We show in Fig.~\ref{RRG_z5} that this expression predicts the numerical results extremely accurately even for $N$ as low as $50$, and outperforms some previous analytical results~\cite{Fronczak05,Dorogovtsev03b,Newman01a} which we present in Appendices~\ref{App:RRG} and \ref{App:Davg}. The accuracy observed in Fig.~\ref{RRG_z5} for regular random networks suggests potential high accuracy of our approach for more complicated network topologies that we consider next.
\begin{figure}[h]
\centering
\includegraphics[width=0.97\columnwidth]{fig5.pdf}
\caption{(Color online) (a) Average shortest path length $\bar D=\sum_n n D_n$ and (b) the probabilities $D_n$ that the distance between a random pair of nodes is $n$ (for $n\in\{2,3,4\}$) as functions of the number of network nodes $N$ for a regular random graph with mean degree $z=5$. We compare the results of our Eq.~\eqref{LargeNRRG} and the analytical results of~\cite{Fronczak05,Dorogovtsev03b,Newman01a} (presented in Appendices~\ref{App:RRG} and \ref{App:Davg}) with direct numerical simulations shown by symbols. In all cases Eq.~\eqref{LargeNRRG} provides the best prediction of numerical results, which remains accurate even for small $N$. The results of Dorogovtsev et al.~\cite{Dorogovtsev03b} match very closely to~Eq.~\eqref{LargeNRRG}, but are more difficult to calculate. Numerical results are averaged over 10 realizations of networks.}
\label{RRG_z5}
\end{figure}
\section{$p_k$-theory: Generalization to random networks with arbitrary degree distribution} \label{pk_theory}
To generalize Eqs.~\eqref{RRG_rhon}-\eqref{RRG_qn} from $z$-regular random networks to networks with an arbitrary degree distribution $p_k$ constructed using the configuration model~\cite{Newman10}, we consider an epidemic started by a degree-$k'$ seed node, and let $\rho_n^{k,k'}$ be the corresponding expected fraction of degree-$k$ nodes infected at time step $n$. We can calculate $\rho_n^{k,k'}$ using the following set of recurrence equations
\begin{align}
\rho_{n+1}^{k,k'} &= 1-(1-\rho_0^{k,k'}) (1-\bar q_n^{k,k'})^k, \label{pk_rhon} \\
q_{n+1}^{k,k'} &= 1- (1-\rho_0^{k,k'})(1-\bar q_n^{k,k'})^{k-1}, \label{pk_qn} \\
\bar q_n^{k,k'} &= \sum_k \frac{k p_k}{z} q_n^{k,k'}, \label{pk_qbar}
\end{align}
with initial values
\begin{align}\label{rho0}
q_0^{k,k'} = \rho_0^{k,k'}=\frac{\delta_{k,k'}}{N p_{k'}}.
\end{align}
Here $q_n^{k,k'}$ is the probability that a degree-$k$ node is infected, given that its parent is susceptible, and $\bar q_n^{k,k'}$ is the probability that a child of a susceptible degree-$k$ node is infected at time step $n$ of an epidemic started from a single infected node of degree $k'$,\footnote{Note that $\bar q_n^{k,k'}$ given by~\eqref{pk_qbar} is independent of $k$; we write it in this general form so that Eqs.~\eqref{pk_rhon}-\eqref{pk_qn} are compatible with Eq.~\eqref{qbar} of Sec.~\ref{Pkk_theory}.} and $z=\sum_k k p_k$ is the mean degree. Note that Eqs.~\eqref{pk_rhon}-\eqref{pk_qbar} reduce to Eqs.~\eqref{RRG_rhon}-\eqref{RRG_qn} for the $z$-regular case of $p_k=\delta_{k,z}$.
Since finite paths only exist in a connected component, we focus our further analysis on the giant connected component (GCC) of the network, and exclude small components and nodes with degree 0. Note that for configuration model networks, GCC is the entire network (in the thermodynamic limit) if and only if the degree distribution $p_k$ does not contain degree-0 and degree-1 nodes. If there are degree-1 nodes in the degree distribution, the generated network will have the fractional GCC size $<1$.\footnote{If $p_1>0$, pairs of connected degree-1 nodes, as well as other nodes or small components surrounded by degree-1 nodes will exist and will not belong to GCC.} For example, for Erd\H{o}s-R\'{e}nyi~networks with low mean degree $z$, GCC size is significantly below 1, and asymptotically approaches $1$ as $z$ increases.
The values of $\rho_n^{k,k'}$ calculated by iterating Eqs.~\eqref{pk_rhon}-\eqref{pk_qbar} tell us the fractions of infected nodes in the entire network, as opposed to that in the GCC. Since all nodes in GCC eventually become infected, the steady state values $\rho_\infty^{k}\equiv\rho_\infty^{k,k'}\le 1$ (that do not depend on the seed node degree $k'$) give us the fraction of degree-$k$ nodes who are part of GCC~\cite{Melnik14,Gleeson08a}. Hence, the fraction of infected degree-$k$ nodes in GCC at time $n$ is $\rho_n^{k,k'} /\rho_\infty^{k,k'}$.
The probability that two nodes (chosen from the entire network) are not connected is $D_\infty = 1-(\sum_k p_k \rho_\infty^{k})^2$, where $\sum_k p_k \rho_\infty^{k}$ is the fractional size of GCC.
Next, assuming that a pair of nodes is chosen from GCC, the probability that two random nodes with degrees $k$ and $k'$ are at a distance $n$ from each other is
\begin{align}
D_n^{k,k'} = \frac{\rho_n^{k,k'} - \rho_{n-1}^{k,k'}}{\rho_\infty^{k,k'}} \label{e:Dnkk};
\end{align}
the probability that a randomly chosen node (of any degree) is at a distance $n$ from a randomly chosen degree-$k$ node is
\begin{align}
D_n^k=\sum_{k'} p_{k'} D_n^{k,k'} \label{e:Dnk},
\end{align}
and
\begin{align}
D_n=\sum_{k} p_{k} D_n^{k} \label{e:Dn}
\end{align}
is the probability that the distance between a pair of random nodes in GCC is $n$.
\begin{figure}[htb]
\flushright
\includegraphics[width=0.95\columnwidth]{fig6a.pdf}
\includegraphics[width=0.97\columnwidth]{fig6b.pdf}
\caption{(Color online) Average shortest path length $\bar D=\sum_n n D_n$ versus the number of nodes $N$ (top panel) and the distribution of distances~$D_n$ (bottom panels) for GCC of Erd\H{o}s-R\'{e}nyi~networks with mean degree $z=3$ and $z=5$. We compare our theoretical results (Eqs.~\eqref{pk_rhon}-\eqref{e:Dn}) and the analytical results of~\cite{Fronczak05,Newman01a,Katzav15} (presented in Appendices~\ref{App:Davg} and \ref{App:Katzav}) with direct numerical simulations. In all cases our theory provides the best prediction of numerical results, which remains accurate even for small $N$. Numerical results are averaged over 50 realizations of networks.}
\label{ERG}
\end{figure}
In Fig.~\ref{ERG}, we illustrate our approach using Erd\H{o}s-R\'{e}nyi~networks and compare the results with numerical simulations, and with some previously obtained analytical results~\cite{Fronczak05,Newman01a,Katzav15} presented in Appendices~\ref{App:Davg} and \ref{App:Katzav}. We use the Poisson degree distribution $p_k = e^{-z}z^k/k!$ of~Erd\H{o}s-R\'{e}nyi~networks in Eqs.~\eqref{pk_rhon}-\eqref{e:Dn} and plot the results in Fig.~\ref{ERG}. Our theory agrees better with numerical simulations than the previous analytical approaches. The prediction of the distribution of distances~is excellent, but not as perfect as in the $z$-regular case in Fig.~\ref{f:rho_n_Dn_vs_n}(b); we explain the reasons for this in Appendix~\ref{s:error_pk}.
\section{$P(k,k')$-theory: Generalization to random networks with degree-degree correlations} \label{Pkk_theory}
Our approach can be easily generalized to random networks specified by the joint degree-degree distribution $P(k,k')$, which is defined as the probability that a randomly chosen network edge connects a degree-$k$ node to a degree-$k'$ node~\cite{Gleeson08a,Melnik11,Melnik14}. The joint distribution $P(k,k')$ determines the degree distribution of the network
\begin{align}
p_k = \frac{\sum_{k'}P(k,k')/k}{\sum_{k',k''}P(k',k'')/k'},
\end{align}
but it also contains additional information about the correlation of node degrees at either end of an edge. Thus, $P(k,k')$ describes the network topology more accurately than $p_k$, and using $P(k,k')$ should improve the accuracy of our approach when we apply it to real-world networks.
Equations~\eqref{pk_rhon}-\eqref{e:Dn} of Sec.~\ref{pk_theory} can be directly applied to networks specified by the joint degree-degree distribution $P(k,k')$, except that Eq.~\eqref{pk_qbar} should be replaced with
\begin{align}
\bar q_n^{k,k'} = \frac{\sum_{k''}P(k,k'')q_n^{k'',k'}}{\sum_{k''}P(k,k'')} \label{qbar}.
\end{align}
In the case of degree-uncorrelated networks, the joint degree-degree distribution factorizes as $P(k,k') = k p_k k' p_{k'}/ z^2$ (since the degrees of nodes at either end of an edge are independent) and Eq.~\eqref{qbar} reduces to Eq.~\eqref{pk_qbar}.
\section{Application to Real-World Networks}\label{s:RWNs}
To apply our approach to real-world networks, we calculate the degree distribution $p_k$ and/or the joint degree-degree distribution $P(k,k')$ from the network adjacency matrix. We then use one or both of these distributions in the equations presented in Secs.~\ref{pk_theory} and \ref{Pkk_theory} to obtain theoretical results. We will refer to the results of our approach where we use $p_k$ of the network and Eq.~\eqref{pk_qbar} as $p_k$-theory, and where we use $P(k,k')$ of the network and Eq.~\eqref{qbar} as $P(k.k')$-theory.
In Fig.~\ref{RWN}, we compare the numerically calculated distribution of distances~for several real-world networks (see Table~\ref{table1}) with the results of $p_k$ and $P(k,k')$-theories. In general, $P(k,k')$-theory predicts the numerical distribution of distances~better than $p_k$-theory as it uses additional information about the degree correlations in the network.
\begin{figure}[h]
\centering
\includegraphics[width=0.97\columnwidth]{fig7.pdf}
\caption{(Color online) Distribution of distances $D_n$ for several real-world networks. Generally, $P(k,k')$-theory predicts the actual $D_n$ better than $p_k$-theory.}
\label{RWN}
\end{figure}
\setlength{\tabcolsep}{0.05cm}
\begin{table}[floatfix]
\begin{center}
\begin{tabular}{l|l r c c |c c c c}
\hline
&&&&& \multicolumn{4}{c}{Relative errors in $\bar D$ for}\\
&Network&$N$&$z$&$\bar D$&$p_k$&$P(k,k')$&FR.& NMN.\\
\hline
\begin{rotate}{90}
\hspace{-2cm}Real world
\end{rotate}
&Word adj. Eng.~\cite{Onnela12}& 7377 & 12 & 2.78 & -0.09 & -0.02 & -0.04 & 0.24\\
&Word adj. French~\cite{Onnela12}& 8308 & 5.7 & 3.22 & -0.02 & 0.00 & 0.06 & 0.27\\
&500 Airports~\cite{Colizza07}& 500 & 11.9 & 2.99 & 0.07 & 0.06 & 0.08 & 0.35\\
&Interact. Prot.~\cite{Colizza05,Colizza06}&4713&6.0 & 4.22 & 0.08 & 0.06 & 0.10 & 0.28\\
&C. Eleg. Met.~\cite{Duch05} & 453 & 8.9 & 2.66 &-0.02 & 0.02 & 0.01 & 0.23\\
&C. Eleg. Neur.~\cite{Watts98}& 297 & 14.5 & 2.46 & 0.00 & 0.03 & 0.00 & 0.22\\
&FB Caltech~\cite{Traud08} & 762 & 43.7 & 2.34 & 0.03 & 0.03 & 0.03 & 0.29\\
&FB Georgetown~\cite{Traud08}& 9388 & 90.7 & 2.76 & 0.10 & 0.08 & 0.09 & 0.31\\
&FB Oklahoma~\cite{Traud08}& 17420 &102.5 & 2.77 & 0.07 & 0.04 & 0.07 & 0.29\\
\hline
\begin{rotate}{90}
\hbox{\hspace{-1.3cm}Synthetic}
\end{rotate}
&$z$-regular &500&4 & 5.02 & 0.00 & 0.00 & 0.07 & 0.00\\
&$z$-regular &500&10 & 2.96 & 0.00 & 0.00 & 0.02 & 0.08\\
&Erd\H{o}s-R\'{e}nyi &500&10 & 2.96 & 0.00 & 0.00 & 0.01 & 0.10\\
&$z$-regular &10000&4 & 7.73 & 0.00 & 0.00 & 0.05 & 0.00\\
&$z$-regular &10000&10 & 4.34 & 0.00 & 0.00 & 0.01 & 0.06\\
&Erd\H{o}s-R\'{e}nyi &10000&10 & 4.26 & 0.00 & 0.00 & 0.01 & 0.07
\end{tabular}
\end{center}
\caption{Comparison of the actual value of the mean distance $\bar D$ for several real-world and synthetic networks with the values calculated using different analytical methods. Here, $N$ is the number of network nodes and $z$ is the mean degree. We show the relative errors $(\bar D - \bar D_{\rm theor.} )/ \bar D$ for the four theories: our $p_k$ and $P(k,k')$-theories of Secs.~\ref{pk_theory} and~\ref{Pkk_theory} respectively, (FR.) Eq.~\eqref{e:Dbar_FR} of Fronczak et al.~\cite{Fronczak05}, and (NMN.) Eq.~\eqref{e:Dbar_NMN} of Newman et al.~\cite{Newman01a}.
}
\label{table1}
\end{table}
In Fig.~\ref{fig:Davg_vs_k}, we plot the expected length of the shortest path between two nodes one of which has degree $k$. We plot $\bar D^{k}_a=\sum_{k',n} n p_{k'} D_{n}^{k',k}$ and $\bar D^k_b=\sum_{k',n} n p_{k'} D_{n}^{k,k'}$. Here, $\bar D^{k}_a$ is calculated based on an epidemic started from a degree-$k$ seed and averaged over the degrees of other nodes, while in $\bar D^{k}_b$ we consider a degree-$k$ node and average over multiple epidemics started with seeds of various degrees. Of course these are the same when one runs numerical simulations of the epidemic process on a given finite network. However, our theoretical approach gives only approximate equality between $D_n^{k,k'}$ and $D_n^{k',k}$ (due to the infinite network assumption as we discuss in Appendix \ref{s:asymmetry}), leading to small differences between $\bar D^{k}_a$ and $\bar D^{k}_b$. We show the results for several real-world and synthetic networks. For real-world networks, $P(k,k')$-theory usually better predicts the exact numerical results than $p_k$-theory. The errors for $\bar D^{k}_a$ and $\bar D^{k}_b$ are similar, and the choice of quantity that best matches the numerical results depends on a particular network. The errors may be attributed in part to clustering~\cite{Faqeeh15a,Melnik11}, modular structure~\cite{Melnik14}, or other topological features~\cite{Faqeeh16} present in these networks. For synthetic networks the results of $P(k,k')$~and $p_k$-theories coincide because of the absence of degree-degree correlations, and are in excellent agreement with numerical results.
\begin{figure}[h]
\flushleft
\includegraphics[width=0.98\columnwidth]{fig8a.pdf}
\includegraphics[width=1\columnwidth]{fig8b.pdf}
\caption{(Color online) The average distance from a degree-$k$ node to all other nodes as a function of $k$. For real-world networks, the result of $P(k,k')$-theory (left column) is more accurate than the result of $p_k$-theory (right column). The bottom panel shows the result for synthetic networks with $N=5000$ nodes: an Erd\H{o}s-R\'{e}nyi~graph with mean degree $z=5$, a random network with a (truncated) power-law degree distribution $p_k\propto k^{-2.5}$ ($2\le k\le 30$), and a random network with a ``flat'' degree distribution $p_k = \rm{const}$ ($2\le k\le 50$).}
\label{fig:Davg_vs_k}
\end{figure}
As our last example, we show in Fig.~\ref{Davg_vs_kk} the expected distance between a random pair nodes as a function of the product of their degrees. We use the same set of networks as in Fig.~\ref{fig:Davg_vs_k}, and plot
\begin{align}
\left<D_{k_i k_j}\right>=\frac{\sum_{k_i'\le k_j'}\delta_{k_i' k_j',k_i k_j} \frac12 \left(\bar D^{k_i',k_j'} + \bar D^{k_j',k_i'}\right) }{\sum_{k_i'\le k_j'} \delta_{k_i' k_j',k_i k_j}}, \label{e:Dkk}
\end{align}
where $\bar D^{k,k'} = \sum_n n D_n^{k,k'}$, and $\delta_{i,j}$ is the Kronecker delta function. As in Fig.~\ref{fig:Davg_vs_k}, $P(k,k')$-theory usually works better than $p_k$-theory for real-world networks. For synthetic networks the theoretical results are again in excellent agreement with the exact numerical calculations.
\begin{figure}[h]
\flushleft
\includegraphics[width=0.97\columnwidth]{fig9a.pdf}
\includegraphics[width=0.97\columnwidth]{fig9b.pdf}
\caption{(Color online) Average shortest path length between a pair of nodes $i$ and $j$ as a function of the product of their degrees $k_i$ and $k_j$. Same set of networks as in Fig.~\ref{fig:Davg_vs_k}.
For real-world networks, the result of $P(k,k')$-theory (left column) is more accurate than the result of $p_k$-theory (right column).}
\label{Davg_vs_kk}
\end{figure}
Finally, we note that theoretical predictions for networks that contain degree-1 nodes can be further improved using the procedure described in Appendix~\ref{App:rectified}.
\section{Conclusions} \label{s:conclusions}
We have presented a simple yet very powerful analytical method for calculating shortest path lengths in networks. Our approach is directly applicable to real-world networks with known degree distribution $p_k$ and/or joint degree-degree distribution $P(k,k')$. It is simpler and yields more accurate predictions for synthetic and real-world networks than some previous analytical methods. Our approach can be extended to modular networks~\cite{Gleeson08a,Melnik14}, to networks with non-zero clustering~\cite{Gleeson09a,Gleeson09b,Hackett11,Miller09b,Newman09}, and also to directed networks~\cite{Gleeson08b}. We hope that these results will be useful for investigating the interdependence of network characteristics and will help advance the understanding of network structure and dynamics.
\section*{ACKNOWLEDGEMENTS}
S.M. acknowledges the INSPIRE fellowship funded by the Irish Research Council co-funded by Marie Curie Actions under FP7. J.P.G. and S.M. acknowledge funding provided by Science Foundation Ireland under programmes 11/PI/1026 and 12/IA/1683. We thank Ali Faqeeh, Dmitri Krioukov, Sergey Dorogovtsev and Alexander Goltsev for useful discussions. We thank Mason Porter, Adam D'Angelo and Facebook for providing the Facebook data used in this study. We also thank Jukka-Pekka Onnela, Alex Arenas, Mark Newman, and Cx-Nets collaboratory for sharing other data sets used in this paper. This work was conceived in part at the Complex Systems Summer School (CSSS) in Santa Fe Institute, NM, USA.
|
\section{Introduction} In the 2000's, following a proposal by Kontsevich, physicists established a precise link between D-branes in B-type topological Landau-Ginzburg models and matrix factorizations of the corresponding Landau-Ginzburg superpotentials \cite{KL1,BHLS,La4}. Even though the matrix factorizations had, by that time, already been an active area of research in mathematics, the physics interpretation very soon yielded new insights into the subject. One of the first examples of that was a simple universal formula, discovered in \cite{KL2} (``the Kapustin-Li formula''), for Calabi-Yau structures on categories of matrix factorizations. The main result of the present work is a refinement of this formula. In the rest of this introduction we recall what the original formula looks like, explain why (and in what sense) it needs to be refined, and outline our approach to the problem.
Let $f$ be a regular function (a ``superpotential'') on an open affine subset $X\subset{\mathbb{C}}^n$. Assume that the only critical value of $f$ is 0 and that all the critical points of $f$ are isolated. A matrix factorization of $f$ is a ${\mathbb{Z}/2}$-graded trivial bundle on $X$ endowed with an odd endomorphism ${\mathrm{D}}$ satisfying ${\mathrm{D}}^2=f\cdot{\rm id}$. Such factorizations can be organized into a differential ${\mathbb{Z}/2}$-graded category, ${\mathscr{MF}(f)}$, as follows: the ${\mathbb{Z}/2}$-graded space of morphisms between two matrix factorizations comprises all homomorphisms between the underlying ${\mathbb{Z}/2}$-graded bundles and the differential on this space is the super-commutator with the corresponding ${\mathrm{D}}$-operators.
The Kapustin-Li formula describes a Calabi-Yau structure on the ${\mathbb{Z}/2}$-graded homotopy category ${\mathfrak{MF}(f)}$ of ${\mathscr{MF}(f)}$. Let us explain what the term ``Calabi-Yau structure'' really means in this context. Given a ${\mathbb{Z}/2}$-graded Hom-finite category ${\mathfrak{A}}$ (all our categories are ${\mathbb{C}}$-linear and small), a Calabi-Yau structure of degree (or parity) $d\in{\mathbb{Z}/2}$ on ${\mathfrak{A}}$ is a non-degenerate degree $d$ trace on ${\mathfrak{A}}$, i.~e. linear maps
$
\theta: {\mathrm{End}}^d_{\mathfrak{A}}(X)\to{\mathbb{C}},
$
for all $X\in{\mathfrak{A}}$, such that the induced pairings
$
{\mathrm{Hom}}^*_{\mathfrak{A}}(X,Y)\otimes{\mathrm{Hom}}^{d-*}_{\mathfrak{A}}(Y,X)\to{\mathbb{C}}
$
are non-degenerate and graded-symmetric for all $X,Y$.
The category ${\mathfrak{MF}(f)}$ is known to be Hom-finite (cf. Proposition \ref{mfpr}). The explicit Calabi-Yau structure (of the same parity as $n=\dim\,X$) on ${\mathfrak{MF}(f)}$ found in \cite{KL2} is given by the formula
\begin{equation}\label{KLtr}
\theta_{\rm KL}: {\mathrm{End}}^n_{{\mathfrak{MF}(f)}}({\mathrm{D}})\to{\mathbb{C}},\quad \theta_{\rm KL}(\Phi):=\frac1{n!}\sum_{x}\mathrm{Res}_x\left[\frac{{{\sf str}}\left((\partial{\mathrm{D}})^{\wedge n}\,\Phi\right)}{\partial_1f\ldots\,\partial_nf}\right].
\end{equation}
Here the summation is over the critical points of $f$, $\mathrm{Res}_x$ stands for the local residue at $x$, ${{\sf str}}$ for the super-trace, $\partial$ for the holomorphic de Rham differential, and $\partial_i:=\frac{\partial}{\partial z_i}$.
(It should be noted that a mathematical proof of the non-degeneracy of the associated pairing -- called by physicists the open-string topological metric -- was first found only few years later \cite{M,DM}.)
From the physics perspective, the formula (\ref{KLtr}) solves a concrete problem: it completes the description of the open topological field theory \cite{La1,Mo} underlying the B-twisted Landau-Ginzburg model with superpotential $f$. It is at the same time the starting point for a new project \cite{Ca} -- that of ``upgrading'' the open topological field theory to an open topological {\it conformal} field theory \cite{Co,HLL}, i.~e. finding a minimal $A_\infty$ model
\[
({\mathfrak{MF}(f)}, \mu_2\!=\!\boldsymbol{\cdot}, \mu_3,\mu_4,\ldots)
\] for ${\mathscr{MF}(f)}$ ($\boldsymbol{\cdot}$ denotes the composition of morphisms in ${\mathfrak{MF}(f)}$) that is cyclic with respect to $\theta_{\rm KL}$. The latter means that for any $l$ and any factorizations ${\mathrm{D}}^{(i)}$ and morphisms $\Phi_i\in{\mathrm{Hom}}^*_{{\mathfrak{MF}(f)}}({\mathrm{D}}^{(i)},{\mathrm{D}}^{(i+1)})$ ($i\in {\mathbb{Z}}/(l+1){\mathbb{Z}}$) the expression $\theta_{\rm KL}(\Phi_{l+1}\cdot\mu_{l}(\Phi_{l}\otimes\ldots\otimes\Phi_1))$ has to be cyclically graded-symmetric.
The cyclicity property is a severe restriction on the minimal model and a generic minimal model does not satisfy it. Moreover, the very existence of a cyclic model is far from obvious. Let ${\mathscr{A}}$ be a differential ${\mathbb{Z}/2}$-graded category whose ${\mathbb{Z}/2}$-graded homotopy category ${\mathfrak{A}}$ is Hom-finite (in this case ${\mathscr{A}}$ is said to be proper) and carries a degree $d$ Calabi-Yau structure $\theta$. How can one be sure that ${\mathscr{A}}$ has a minimal $A_\infty$ model that is cyclic with respect to $\theta$? An exhaustive answer to this question was given in \cite{KS}. The answer involves the broader notion of a Calabi-Yau structure on a proper differential ${\mathbb{Z}/2}$-graded category and can be formulated as follows: A cyclic model exists provided $\theta$ can be ``lifted'' to a Calabi-Yau structure on ${\mathscr{A}}$, i.~e. if there is a functional $\Theta:{\mathsf{H}}^\lambda_d({\mathscr{A}})\to{\mathbb{C}}$ on the $d$-th cyclic homology group of ${\mathscr{A}}$ whose pullback along the composition of certain canonical maps (cf. Section \ref{cypc})
\begin{equation}\label{lift}
\bigoplus_{X\in{\mathfrak{A}}}{\mathrm{End}}^d_{{\mathfrak{A}}}(X)\to {\sf HH}_d({\mathscr{A}})\to{\mathsf{H}}^\lambda_d({\mathscr{A}})
\end{equation}
(${\sf HH}_*$ is the Hochschild homology) coincides with $\theta$. This result was proven in \cite{KS} using tools of formal non-commutative symplectic geometry. The same tools may be used -- at least, in principle -- to actually construct a cyclic minimal $A_\infty$ model of ${\mathscr{A}}$ starting from any lift $\Theta$ of $\theta$ and any, not necessarily cyclic minimal model of ${\mathscr{A}}$ \cite{Ca}.
Thus, we are naturally led to the question: Can $\theta_{\rm KL}$ be lifted to ${\mathsf{H}}^\lambda_n({\mathscr{MF}(f)})$? The answer is known to be yes \cite{DM,Se} but this is not at all straightforward. The naive idea that the same formula (\ref{KLtr}), extended to ${\mathrm{End}}^n_{{\mathscr{MF}(f)}}({\mathrm{D}})$, gives a Calabi-Yau structure on ${\mathscr{MF}(f)}$ is easily seen to be wrong since,
in general,
\[
\theta_{\rm KL}(\Phi''\Phi')\neq(-1)^{|\Phi'||\Phi''|}\theta_{\rm KL}(\Phi'\Phi'')
\]
(here $|\cdot|$ denotes the parity of a morphism). Instead, one can argue as follows: According to \cite{Se}, $\theta_{\rm KL}$ can be extended to a functional on ${\mathsf{HH}}_n({\mathscr{MF}(f)})$ but this already suffices to claim that an extension to ${\mathsf{H}}^\lambda_n({\mathscr{MF}(f)})$ exists as well -- this follows from the degeneration of the Hochschild-to-cyclic spectral sequence for ${\mathscr{MF}(f)}$ \cite{Dy,DM}.
Given the above-mentioned potential practical significance of Calabi-Yau structures, another natural question to ask is: Are there explicit formulas, similar to (\ref{KLtr}), for a lift of $\theta_{\rm KL}$ to ${\mathsf{H}}^\lambda_n({\mathscr{MF}(f)})$? It is this question that we answer in the present work. The simplest of our formulas looks as follows:
\begin{equation}\label{corrKL}
\theta_{\rm KL}(\Phi''\Phi')-(-1)^{|\Phi'||\Phi''|}\theta_{\rm KL}(\Phi'\Phi'')=\widetilde{\theta}(\Phi''\otimes\delta(\Phi'))-(-1)^{|\Phi''|}\widetilde{\theta}(\delta(\Phi'')\otimes\Phi')
\end{equation}
where $\Phi'\in {\mathrm{Hom}}^*_{\mathscr{MF}(f)}({\mathrm{D}}',{\mathrm{D}}'')$, $\Phi''\in {\mathrm{Hom}}^*_{\mathscr{MF}(f)}({\mathrm{D}}'',{\mathrm{D}}')$, $\delta$ denotes the differential on the morphisms in ${\mathscr{MF}(f)}$, and
\begin{multline*}
\widetilde{\theta}(\Psi''\otimes\Psi'):=
\frac{(-1)^n}{(n+1)!}\sum_{x}\sum_{j,\,k=1}^n(-1)^{(k-1)(|\Psi'|+1)}\\
\mathrm{Res}_x\!\left[\frac{{{\sf str}}\left(\Psi''\,(\partial{\mathrm{D}}'')^{\wedge k}\,\Psi'\partial_{j} {\mathrm{D}}'\,(\partial{\mathrm{D}}')^{\wedge(n-k)}+(-1)^{k-1}\Psi''\,\partial_{j} {\mathrm{D}}''\,(\partial{\mathrm{D}}'')^{\wedge(k-1)}\,\Psi'\,(\partial{\mathrm{D}}')^{\wedge (n-k+1)}\right)}{\partial_1f\ldots(\partial_jf)^2\ldots \partial_nf}\right]
\end{multline*}
The functional $\widetilde{\theta}$ is cyclically graded-symmetric:
\[
\widetilde{\theta}(\Psi''\otimes\Psi')=(-1)^{(|\Psi'|+1)(|\Psi''|+1)}\widetilde{\theta}(\Psi'\otimes\Psi''),
\]
which means that $\theta_{\rm KL}+\widetilde{\theta}$ is a kind of ``infinitesimal lift'' of $\theta_{\rm KL}$ to the cyclic homology of ${\mathscr{MF}(f)}$. It is still not a Calabi-Yau structure and needs to be corrected by ``higher order'' terms. Our main result -- Theorem \ref{mainresult} -- provides explicit formulas for all such higher corrections to $\theta_{\rm KL}$, and thereby solves the problem of lifting the latter to a Calabi-Yau structure on ${\mathscr{MF}(f)}$.
In fact, our result is more general, namely, we construct a family of Calabi-Yau structures on ${\mathscr{MF}(f)}$ depending on a holomorphic volume form $\Omega$ on $X$; the Kapustin-Li trace and its lift to ${\mathscr{MF}(f)}$ correspond to the special case $\Omega=dz_1\wedge\ldots\wedge dz_n$. The fact that any volume form gives rise to a Calabi-Yau structure on ${\mathscr{MF}(f)}$ is not surprising and is in agreement with \cite[Thm.8.3.4]{Pr} where the volume forms were shown to determine {\it smooth} Calabi-Yau structures on matrix factorization categories. The notion of a smooth Calabi-Yau structure is dual, in some sense, to the one we are interested in here (the latter is also referred to as a {\it proper} Calabi-Yau structure) and for ``nice'' categories, like ${\mathscr{MF}(f)}$, the two types of Calabi-Yau structures are known to be in bijection \cite[Prop.6.10]{GPS}.
Let us quickly summarize the main ideas of our approach. To begin with, it does not rely on any of the previously mentioned ideas and results. In particular, the non-degeneracy of the pairing associated with $\theta_{\rm KL}$ is a consequence of the construction. It is also independent of the original paper \cite{KL2} in the sense that the trace $\theta_{\rm KL}$ is not part of the input data. What we do here is an ``open-string'' generalization of the approach to the closed-string topological metric in the same setting of Landau-Ginzburg models developed in \cite[Sect.2.2]{LLS}. In \cite{LLS} the authors construct an explicit quasi-isomorphism
\begin{equation}\label{pv}
({\rm PV}^*({X^{\mathtt{h}}}), \lbrace f,\cdot\rbrace)\to({\rm PV}^*({X^{\mathtt{h}}}), \lbrace f,\cdot\rbrace)\otimes({\mathscr{E}}_{\tt c}^{(0,*)}({X^{\mathtt{h}}}), \bar{\partial})
\end{equation}
where ${X^{\mathtt{h}}}$ is the analytification of $X$, ${\rm PV}^*({X^{\mathtt{h}}})$ stands for the space of holomorphic polyvector fields, $\lbrace\cdot,\cdot\rbrace$ is the Schouten-Nijenhuis bracket, and
${\mathscr{E}}_{\tt c}^{(0,*)}({X^{\mathtt{h}}})$ is the space of smooth compactly supported $(0,*)$-forms. Any holomorphic volume form on ${X^{\mathtt{h}}}$ determines a trace on the right-hand side of (\ref{pv}) whose pullback along the quasi-isomorphism can be written in terms of the classical residue trace. This matches the description of the closed-string topological metric found in \cite{Va}. We apply the same technique to the matrix factorization categories, namely, we construct an explicit $A_\infty$ quasi-equivalence
\begin{equation}\label{mf}
{\mathscr{MF}(f)}\to{\mathscr{MF}(f)}\otimes({\mathscr{E}}_{\tt c}^{(0,*)}({X^{\mathtt{h}}}), \bar{\partial}),
\end{equation}
observe that any volume form gives rise to a Calabi-Yau structure on the right-hand side, and then pull back this Calabi-Yau structure along the quasi-equivalence.
We would like to close by pointing out that the category in the right-hand side of (\ref{mf}) -- we denote it by ${\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}$ in the text -- seems to be an interesting object in its own right. Being an exact analog of the Dolbeault realization of the bounded derived category of a Calabi-Yau variety, ${\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}$ may provide a streamlined approach to various aspects of the B-twisted Landau-Ginzburg models. For instance, it should allow one to transfer the constructions of \cite{Co1,Po} to the Landau-Ginzburg setting. We hope to return to this topic in the future.
\medskip
\noindent{\bf Conventions.} For a ${\mathbb{Z}/2}$-graded space $V$ the parity of an element $v\in V$ will be denoted by $|v|$, ${{{\mathit{s}}}} V$ will stand for $V$ with the reversed ${\mathbb{Z}/2}$-grading, and ${{{\mathit{s}}}}$ by itself will stand for the canonical odd map $V\to{{{\mathit{s}}}} V$. If $V=(V,d_V)$ is a ${\mathbb{Z}/2}$-graded complex then ${{{\mathit{s}}}} V$ will also denote the complex $({{{\mathit{s}}}} V,d_{{{{\mathit{s}}}} V})$ where $d_{{{{\mathit{s}}}} V}({{{\mathit{s}}}} v):= -{{{\mathit{s}}}} d_V(v)$. The Koszul sign rule will always be assumed when working with tensors and $[\cdot,\cdot]$ will denote the super-commutator. We will abbreviate ``differential ${\mathbb{Z}/2}$-graded'' and ``Calabi-Yau'' to ``dg'' and ``CY''.
\medskip
\noindent{\bf Acknowledgements.} I would like to thank Manfred Herbst and Daniel Murfet for introducing me to the Kapustin-Li formula and Wolfgang Lerche, Emanuel Scheidegger and Johannes Walcher for inspiring discussions on matrix factorizations and ``open'' mirror symmetry. Special thanks are due to Nils Carqueville for bringing my attention to the problem of constructing an ``off-shell'' version of the Kapustin-Li trace.
\section{Preliminaries and the main result}
\subsection{CY structures on proper dg categories}\label{cypc}
In this section ${\mathscr{A}}$ stands for a small ${\mathbb{C}}$-linear dg category.
Let us first recall the definition of the {\it Hochschild complex} $({\mathsf{C}}_*({\mathscr{A}}),b)$ of ${\mathscr{A}}$.
For $l\geq1$ set
\begin{eqnarray*}
{\mathsf{C}}^{\{l\}}_*({\mathscr{A}}):=\bigoplus
{\mathrm{Hom}}^*_{\mathscr{A}}(X^{(l)},X^{(1)})\otimes {{{\mathit{s}}}}{\mathrm{Hom}}^*_{\mathscr{A}}(X^{(l-1)},X^{(l)})\otimes\ldots\otimes
{{{\mathit{s}}}}{\mathrm{Hom}}^*_{\mathscr{A}}(X^{(1)},X^{(2)})
\end{eqnarray*}
where the sum is over all length $l$ collections $X^{(1)},\ldots, \,X^{(l)}$ of objects of ${\mathscr{A}}$ and the ${\mathbb{Z}/2}$-grading on the left-hand side is the total grading on the tensor products. Then, by definition,
$
{\mathsf{C}}_*({\mathscr{A}})=\bigoplus_{l\geq1} {\mathsf{C}}^{\{l\}}_*({\mathscr{A}}).
$
(For $a_{l}\in{\mathrm{Hom}}^*_{\mathscr{A}}(X^{(l)},X^{(1)})$ and $a_i\in{\mathrm{Hom}}^*_{{\mathscr{A}}}(X^{(i)},X^{(i+1)})$, $i\leq l-1$, the corresponding element of ${\mathsf{C}}^{\{l\}}_*({\mathscr{A}})$ will be written as $a_{l}[a_{l-1}|\ldots |a_1]$.) The differential $b$ is the sum of two anti-commuting differentials $b(\delta)$ and $b(\mu)$ where
\[
b(\delta)(a_{l}[a_{l-1}|\ldots |a_1])=\delta a_{l}[a_{l-1}|\ldots |a_1]+\sum\limits_{i=1}^{l-1}(-1)^{\epsilon_{i+1}}a_{l}[a_{l-1}|\ldots |\delta a_i|\ldots|a_1]
\]
($\delta$ stands for the differential on the Hom-complexes of ${\mathscr{A}}$ and $\epsilon_i:=\sum_{j\geq i}|{{{\mathit{s}}}} a_{j}|$)
and
\begin{multline}\label{bmu}
b(\mu)(a_{l}[a_{l-1}|\ldots |a_1])=
(-1)^{|a_{l}|}a_{l}a_{l-1}[a_{l-2}|\ldots |a_1]-\\-\sum\limits_{i=1}^{l-2}(-1)^{\epsilon_{i+1}}a_{l}[a_{l-1}|\ldots
|a_{i+1}a_{i}|\ldots|a_1]-
(-1)^{|{{{\mathit{s}}}} a_1|(\epsilon_2+1)}a_1a_{l}[a_{l-1}|\ldots |a_{2}].
\end{multline}
The definition of the {\it cyclic complex} $({\mathsf{C}}_*^\lambda({\mathscr{A}}),b)$ of ${\mathscr{A}}$ involves the cyclic permutation
\begin{equation}\label{cycper}
\tau: {\mathsf{C}}_*({\mathscr{A}})\to {\mathsf{C}}_*({\mathscr{A}}),\quad \tau(a_{l}[a_{l-1}|\ldots |a_1])=(-1)^{|{{{\mathit{s}}}} a_{l}|(\epsilon_1-|{{{\mathit{s}}}} a_{l}|)}a_{l-1}[a_{l-2}|\ldots |a_1|a_{l}].
\end{equation}
One can easily check that $b(\delta)(1-\tau)=(1-\tau)b(\delta)$ and $b(\mu)(1-\tau)=(1-\tau)b'(\mu)$
where $b'(\mu)$ is the operator on ${\mathsf{C}}_*({\mathscr{A}})$ given by the second line in (\ref{bmu}). It follows that the image ${\rm Im}(1-\tau)\subset {\mathsf{C}}_*({\mathscr{A}})$ is $b$-invariant. Then $({\mathsf{C}}_*^\lambda({\mathscr{A}}),b)$ is defined as the quotient of $({\mathsf{C}}_*({\mathscr{A}}),b)$ by the subcomplex $({\rm Im}(1-\tau),b)$. The cohomology of this complex -- the cyclic homology of ${\mathscr{A}}$ -- is denoted by ${\mathsf{H}}_*^\lambda({\mathscr{A}})$.
Recall \cite[Sect.8.2]{KS} that a dg category ${\mathscr{A}}$ is said to be {\it proper} if for any pair $X'$ and $X''$ of objects $\dim {\mathrm{H}}^*({\mathrm{Hom}}_{{\mathscr{A}}}(X',X''),\delta)<\infty$.
\begin{definition}{\rm (\cite[Sect.10.2]{KS}) Let ${\mathscr{A}}$ be a proper dg category. A {\it degree $d\in{\mathbb{Z}/2}$ CY structure} on ${\mathscr{A}}$ is a functional
$
{\Theta}: {\mathsf{H}}_d^\lambda({\mathscr{A}})\to{\mathbb{C}}
$
with the property that for any objects $X'$ and $X''$ the pairing
\begin{eqnarray}\label{ndc}
{\mathrm{H}}^*({\mathrm{Hom}}_{{\mathscr{A}}}(X'',X'),\delta)\otimes {\mathrm{H}}^{d-*}({\mathrm{Hom}}_{{\mathscr{A}}}(X',X''),\delta)\to {\mathbb{C}},\quad
a''\otimes a'\mapsto {\Theta}(\iota(a''a'))
\end{eqnarray}
is non-degenerate. Here $\iota: \bigoplus_X{\mathrm{H}}^*({\mathrm{End}}_{{\mathscr{A}}}(X),\delta)\to {\mathsf{H}}_*^\lambda({\mathscr{A}})$ is the map induced by the composition of the embedding $\bigoplus_X({\mathrm{End}}^*_{{\mathscr{A}}}(X),\delta)=({\mathsf{C}}^{\{1\}}_*({\mathscr{A}}),b(\delta))\to ({\mathsf{C}}_*({\mathscr{A}}),b)$ and the projection $({\mathsf{C}}_*({\mathscr{A}}),b)\to ({\mathsf{C}}_*^\lambda({\mathscr{A}}),b)$.
}
\end{definition}
\begin{remark}{\rm We will also speak of {\it chain-level} CY structures. By a chain-level CY structure on ${\mathscr{A}}$ we will understand an even/odd functional
$
{\Theta}: {\mathsf{C}}_*({\mathscr{A}})\to{\mathbb{C}}
$
that descends to ${\mathsf{H}}_*^\lambda({\mathscr{A}})$, meaning
\begin{equation}\label{clcy}
{\Theta}\cdot(1-\tau)=0\quad \text{and}\quad{\Theta}\cdot b=0,
\end{equation}
and induces a CY structure in the above sense. }
\end{remark}
\subsection{Matrix factorizations}\label{smf}
Let $(X,f)$ be as in the Introduction, i.~e.
\begin{itemize}
\item $X$ is an open affine subset of ${\mathbb{C}}^n={\rm Spec}\,{\mathbb{C}}[z_1,\ldots,z_n]$;
\item $f$ is a regular function on $X$ whose only critical value is $0\in{\mathbb{C}}$ and whose critical points are all isolated. (The set of critical points will be denoted by $C_f$.)
\end{itemize}
Let us reiterate the definition of the dg category ${\mathscr{MF}(f)}={\mathscr{MF}}(X,f)$.
Its objects are block matrices of the form
\[
{\mathrm{D}}=\begin{blockarray}{ccc}
\scriptstyle{k \,{\rm columns}}& \scriptstyle{k \,{\rm columns}} \\
\begin{block}{[c|c]c}
0& {\mathrm{D}}_{12} & \scriptstyle{k \,{\rm rows}} \\
\BAhhline{--}
{\mathrm{D}}_{21}& 0 & \scriptstyle{k \,{\rm rows}} \\
\end{block}
\end{blockarray}
\]
where ${\mathrm{D}}_{12}$ and ${\mathrm{D}}_{21}$ are $k\times k$ matrices with entries in ${\mathbb{C}}[X]$ satisfying the conditions
\begin{eqnarray}\label{mfc}
{\mathrm{D}}_{12}{\mathrm{D}}_{21}={\mathrm{D}}_{21}{\mathrm{D}}_{12}=f\cdot {\mathbf{1}}_k \quad(\Leftrightarrow\,\, {\mathrm{D}}^2=f\cdot{\mathbf{1}}_{2k})
\end{eqnarray}
(${\mathbf{1}}_k$ stands for the identity $k\times k$ matrix). The space of {\it even} resp. {\it odd} morphisms between two objects
\begin{eqnarray}\label{2obj}
{\mathrm{D}}'=\begin{blockarray}{ccc}
\scriptstyle{k} & \scriptstyle{k} \\
\begin{block}{[c|c]c}
0& {\mathrm{D}}'_{12} & \scriptstyle{k} \\
\BAhhline{--}
{\mathrm{D}}'_{21}& 0 & \scriptstyle{k} \\
\end{block}
\end{blockarray}\quad \text{and} \quad
{\mathrm{D}}''=\begin{blockarray}{ccc}
\scriptstyle{l} & \scriptstyle{l} \\
\begin{block}{[c|c]c}
0& {\mathrm{D}}''_{12} & \scriptstyle{l}\\
\BAhhline{--}
{\mathrm{D}}''_{21}& 0 & \scriptstyle{l} \\
\end{block}
\end{blockarray}
\end{eqnarray}
is the space (in fact, ${\mathbb{C}}[X]$-module) ${\mathrm{Hom}}^{{\mathrm{ev}}}_{{\mathscr{MF}(f)}}({\mathrm{D}}',{\mathrm{D}}'')$ resp. ${\mathrm{Hom}}^{{\mathrm{od}}}_{{\mathscr{MF}(f)}}({\mathrm{D}}',{\mathrm{D}}'')$ of block matrices of the form
\begin{eqnarray}\label{evodmor}
\Phi=\begin{blockarray}{ccc}
\scriptstyle{k} & \scriptstyle{k} \\
\begin{block}{[c|c]c}
\Phi_{11}& 0 & \scriptstyle{l} \\
\BAhhline{--}
0 & \Phi_{22} & \scriptstyle{l} \\
\end{block}
\end{blockarray}
\quad \text{resp.}\quad
\Phi=\begin{blockarray}{ccc}
\scriptstyle{k} & \scriptstyle{k} \\
\begin{block}{[c|c]c}
0& \Phi_{12} & \scriptstyle{l} \\
\BAhhline{--}
\Phi_{21} & 0 & \scriptstyle{l} \\
\end{block}
\end{blockarray}
\end{eqnarray}
where $\Phi_{ij}$ are arbitrary $k\times l$ matrices with entries in ${\mathbb{C}}[X]$; the composition of morphisms is the usual matrix multiplication. Finally, the differential $\delta: {\mathrm{Hom}}^{{\mathrm{ev}}/{\mathrm{od}}}_{{\mathscr{MF}(f)}}({\mathrm{D}}',{\mathrm{D}}'')\to {\mathrm{Hom}}^{{\mathrm{od}}/{\mathrm{ev}}}_{{\mathscr{MF}(f)}}({\mathrm{D}}',{\mathrm{D}}'')$ is defined by the formula
\begin{eqnarray}\label{diff}
\delta(\Phi):={\mathrm{D}}''\Phi-(-1)^{|\Phi|}\Phi {\mathrm{D}}'.
\end{eqnarray}
(Note that $\delta$ is a morphism of ${\mathbb{C}}[X]$-modules.)
\begin{proposition}\label{mfpr} The dg category ${\mathscr{MF}(f)}$ is proper.
\end{proposition}
\noindent This is a special case of \cite[Thm.8.1.1]{Pr} but can also be easily shown directly. Namely, for any pair ${\mathrm{D}}',{\mathrm{D}}''$ of matrix factorizations the complex of $\mathcal{O}_X$-modules, associated with the complex $({\mathrm{Hom}}^*_{{\mathscr{MF}(f)}}({\mathrm{D}}',{\mathrm{D}}''),\delta)$ of ${\mathbb{C}}[X]$-modules, is (Zariski) locally contractible outside of the set of critical points of $f$. To see it, pick a point $x\in X\setminus C_f$ and assume that $\partial_i f(x)\neq0$ for some $i$. Then it follows from the equality ${\mathrm{D}}''^{2}=f\cdot{\mathbf{1}}$ that in the affine neighborhood $U:=\{\partial_i f\neq0\}$ of $x$
\begin{equation}\label{homot1}
{\mathrm{D}}''\cdot {h}_{{\mathrm{D}}''}+{h}_{{\mathrm{D}}''}\cdot {\mathrm{D}}''=\mathbf{1},\quad {h}_{{\mathrm{D}}''}:=\frac{{\partial_i {\mathrm{D}}''}}{{\partial_i f}}.
\end{equation}
Viewing ${h}_{{\mathrm{D}}''}$ as an element of ${\mathrm{End}}^{\mathrm{od}}_{{\mathscr{MF}}(U,f)}({\mathrm{D}}'')$, (\ref{homot1}) means $\delta({h}_{{\mathrm{D}}''})={id}_{{\mathrm{D}}''}$. As a consequence, for every $\Phi\in {\mathrm{Hom}}^*_{{\mathscr{MF}}(U,f)}({\mathrm{D}}',{\mathrm{D}}'')$ one has $\Phi=\delta {h}(\Phi)+{h}\delta(\Phi)$ where
$
{h}(\Phi):={h}_{{\mathrm{D}}''}\cdot \Phi.
$
Thus, the sheaf of $\mathcal{O}_X$-modules associated with the ${\mathbb{C}}[X]$-module ${\mathrm{H}}^*({\mathrm{Hom}}_{{\mathscr{MF}(f)}}({\mathrm{D}}',{\mathrm{D}}''),\delta)$ is a coherent sheaf supported on a finite subset of $X$ which implies the claim.
\subsection{Main theorem} The main result of the present work is
\begin{theorem}\label{mainresult} For any nowhere vanishing holomorphic top degree form $\Omega$ on the analytic space associated with $X$ the following functional ${\Theta}={\Theta}_{\Omega}$ defines a chain-level CY structure on ${\mathscr{MF}(f)}$ (of the same parity as $n=\dim\,X$):
\begin{equation*}
{\Theta}=\sum_{x\in C_f} {\Theta}_{x}:{\mathsf{C}}_*({\mathscr{MF}(f)})\to {\mathbb{C}}
\end{equation*}
where
\begin{multline}\label{upsx}
{\Theta}_{x}(\Phi_{l}[\Phi_{l-1}|\ldots|\Phi_1]):=\frac1{(n+l-1)!}\sum\limits_{\substack{k_1+\ldots+k_{l}=n-1\\k_1,\ldots,k_l\geq0}}(-1)^{k_1\epsilon_1+\ldots+k_l\epsilon_{l}}
\sum_{i=1}^n(-1)^{i}\\
\sum_{\substack {r_1+\ldots+r_n=l\\ r_j\geq0\, j\neq i, r_i\geq1}}r_1!\,\ldots \,r_n!\sum\limits_{\left(i^{(1)},\ldots,\, i^{(l)}\right)\in \Lambda_n^l(r_1,\ldots, r_n)}\\
\sum\limits_{\left(j^{(l)}_1,\ldots, j^{(l)}_{k_1},\ldots, j^{(1)}_1,\ldots, j^{(1)}_{k_{l}}\right)\in {S}_n^i}{\rm sgn}\left(j^{(l)}_1,\ldots, j^{(l)}_{k_1},\ldots, j^{(1)}_1,\ldots, j^{(1)}_{k_{l}}\right)\\\mathrm{Res}_x\left[\frac{{{\sf str}}\left(\Phi_{l}\partial_{i^{(l)}} {\mathrm{D}}^{(l)}\,\partial_{j^{(l)}_1} {\mathrm{D}}^{(l)}\ldots \partial_{j^{(l)}_{k_{l}}} {\mathrm{D}}^{(l)}\cdot \ldots\cdot \Phi_{1}\partial_{i^{(1)}} {\mathrm{D}}^{(1)}\,\partial_{j^{(1)}_1} {\mathrm{D}}^{(1)}\ldots \partial_{j^{(1)}_{k_{1}}}{\mathrm{D}}^{(1)}\right) \wedge
\Omega}{(\partial_1f)^{r_1+1}\ldots(\partial_if)^{r_i}\ldots (\partial_nf)^{r_n+1}}\right].
\end{multline}
In this formula
\begin{itemize}
\item $\{{\mathrm{D}}^{(i)}\}_{i=1,\ldots, l}$ are arbitrary matrix factorizations, $\Phi_{l}\in{\mathrm{Hom}}^{{\mathrm{ev}}/{\mathrm{od}}}_{{\mathscr{MF}(f)}}(D^{(l)},D^{(1)})$ and $\Phi_i\in{\mathrm{Hom}}^{{\mathrm{ev}}/{\mathrm{od}}}_{{\mathscr{MF}(f)}}(D^{(i)},D^{(i+1)})$, $i\leq l-1$;
\item $\epsilon_i:=\sum_{j\geq i}|{{{\mathit{s}}}} \Phi_{j}|$;
\item $\Lambda_n^l(r_1,\ldots, r_n)$ denotes the subset in $\{1,\ldots,n\}^l$ of those multi-indices $(i^{(1)},\ldots,\, i^{(l)})$ that contain precisely $r_1$ copies of 1, $r_2$ copies of 2 etc.
\item ${S}_n^i$ ($i=1,\ldots, n$) stands for the set of all permutations of $(1,2,\ldots,n)\setminus \{i\}$; given an element $(j_1,\ldots,j_{n-1})\in {S}_n^i$, ${\rm sgn}(j_1,\ldots,j_{n-1})$ denotes the sign of the corresponding permutation;
\item $\mathrm{Res}_x$ is the local residue at $x$ and ${{\sf str}}$ is the supertrace.
\end{itemize}
\end{theorem}
The rest of the paper is devoted to the proof of this theorem. We leave it as an exercise for the reader to show that in the case $l=1$ and $\Omega=dz_1\wedge\ldots\wedge dz_n$ the above formula reproduces the Kapustin-Li trace (\ref{KLtr}). Combining this observation with the condition ${\Theta}(b(\Phi''[\Phi']))=0$, one can derive the formula (\ref{corrKL}).
As yet another exercise, the reader may try to prove (part of) the theorem for functions of one variable ``by hand''. Namely, in the special case $n=1$ the formula (\ref{upsx}) becomes quite simple:
\[
{\Theta}_{x}(\Phi_{l}[\Phi_{l-1}|\ldots|\Phi_1])=-\mathrm{Res}_x\left[\frac{{{\sf str}}\left(\Phi_{l}\partial_{z} {\mathrm{D}}^{(l)}\ldots\Phi_{1}\partial_{z} {\mathrm{D}}^{(1)}\right] \wedge
\Omega}{(\partial_zf)^{l}}\right],
\]
and the properties (\ref{clcy}) can be checked directly.
\section{Proof of the main result}
\subsection{``Dolbeault'' models for the category of matrix factorizations}
\begin{definition}{\rm Let
${\mathscr{MF}^{\tt h}(f)}$, ${\mathscr{MF}^{\mathtt{D}}(f)}$ and ${\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}$ be the dg categories with the same objects as ${\mathscr{MF}(f)}$ and with morphism complexes defined as follows:
\begin{enumerate}
\item
$
{\mathrm{Hom}}^*_{{\mathscr{MF}^{\tt h}(f)}}({\cdot,\cdot}):=\left({\mathrm{Hom}}^*_{{\mathscr{MF}(f)}}({\cdot,\cdot})\otimes_{{\mathbb{C}}[X]} \mathcal{A}({X^{\mathtt{h}}}), \delta\otimes 1\right)
$\\
\noindent where ${X^{\mathtt{h}}}$ is the complex manifold (analytic space) associated with $X$ and $\mathcal{A}({X^{\mathtt{h}}})$ is the algebra of holomorphic functions.
\item
$
{\mathrm{Hom}}^*_{{\mathscr{MF}^{\mathtt{D}}(f)}}({\cdot,\cdot}):=\left({\mathrm{Hom}}^*_{{\mathscr{MF}(f)}}({\cdot,\cdot})\otimes_{{\mathbb{C}}[X]} {\mathscr{E}}^{(0,*)}({X^{\mathtt{h}}}), \bar{\boldsymbol{\delta}}:=\delta\otimes 1+1\otimes \bar{\partial}\right)
$\\
\noindent where $({\mathscr{E}}^{(0,*)}({X^{\mathtt{h}}}),\bar{\partial})$ is the differential ${\mathbb{Z}/2}$-graded algebra of smooth $(0,*)$-forms on ${X^{\mathtt{h}}}$ (the ${\mathbb{Z}/2}$-grading comes from the natural ${\mathbb{Z}}$-grading).
\item
$
{\mathrm{Hom}}^*_{{\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}}({\cdot,\cdot}):=\left({\mathrm{Hom}}^*_{{\mathscr{MF}(f)}}({\cdot,\cdot})\otimes_{{\mathbb{C}}[X]} {\mathscr{E}}_{\tt c}^{(0,*)}({X^{\mathtt{h}}}), \bar{\boldsymbol{\delta}}:=\delta\otimes 1+1\otimes \bar{\partial}\right)
$\\
\noindent where ${\mathscr{E}}_{\tt c}^{(0,*)}({X^{\mathtt{h}}})\subset{\mathscr{E}}^{(0,*)}({X^{\mathtt{h}}})$ denotes the subalgebra of compactly supported forms.
\end{enumerate}
}
\end{definition}
\begin{remark}{\rm
Strictly speaking, ${\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}$ is not a dg category in the conventional sense since it does not have identity morphisms. However, Proposition \ref{main_s1} below implies that it is weakly unital.
}
\end{remark}
Let us comment on the structure of the Hom-complexes in these categories.
The case of ${\mathscr{MF}^{\tt h}(f)}$ is clear: we simply allow arbitrary holomorphic functions as entries of the matrices representing the morphisms. The rest of the structure - namely, the composition and the differential - are the same as in the algebraic case. The structure of the Hom-complexes in ${\mathscr{MF}^{\mathtt{D}}(f)}$ and ${\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}$ is slightly more complicated. It is still convenient to think of the morphisms in ${\mathscr{MF}^{\mathtt{D}}(f)}$ as matrices with entries in the algebra of $(0,*)$-forms on ${X^{\mathtt{h}}}$. For example, a generic even morphism in this category is the sum of matrices of the following two types:
\begin{equation}\label{evod}
\begin{bmatrix}
\Phi^{\mathrm{ev}}_{11}& 0 \\
0 & \Phi^{\mathrm{ev}}_{22}
\end{bmatrix}\quad\text{and}\quad
\begin{bmatrix}
0 & \Phi^{\mathrm{od}}_{12} \\
\Phi^{\mathrm{od}}_{21} & 0
\end{bmatrix}
\end{equation}
where $\Phi^{\mathrm{ev}}_{ij}$ resp. $\Phi^{\mathrm{od}}_{ij}$ are matrices with entries in ${\mathscr{E}}^{(0,{\mathrm{ev}})}({X^{\mathtt{h}}})$ resp. ${\mathscr{E}}^{(0,{\mathrm{od}})}({X^{\mathtt{h}}})$. The odd morphisms have a similar structure. One should be careful with this description though. Note that in these terms the composition of morphism does not always coincide with the matrix multiplication. Similarly, the action of the differential $\delta\otimes 1$ is not always given by the formula (\ref{diff}) and the action of the differential $1\otimes \bar{\partial}$ is not necessarily given by the componentwise action of $\bar{\partial}$. This is a consequence of the Koszul rule of signs. To avoid confusion in the future, we will denote the composition of morphisms and the differentials $\delta\otimes1$ and $1\otimes\bar{\partial}$ in both ${\mathscr{MF}^{\mathtt{D}}(f)}$ and ${\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}$ by $\circ$, ${\boldsymbol{\delta}}$ and ${\boldsymbol{\bar{\partial}}}$, respectively. (Thus, $\bar{\boldsymbol{\delta}}={\boldsymbol{\delta}}+{\boldsymbol{\bar{\partial}}}$.)
For example, the composition of
\[
\Psi=\begin{bmatrix}
\Psi^{\mathrm{od}}_{11}& 0 \\
0 & \Psi^{\mathrm{od}}_{22}
\end{bmatrix}\in {\mathrm{Hom}}^{\mathrm{od}}_{{\mathscr{MF}^{\mathtt{D}}(f)}}({\mathrm{D}}'',{\mathrm{D}}'''), \quad
\Phi=\begin{bmatrix}
0 & \Phi^{\mathrm{ev}}_{12} \\
\Phi^{\mathrm{ev}}_{21} & 0
\end{bmatrix}\in {\mathrm{Hom}}^{\mathrm{od}}_{{\mathscr{MF}^{\mathtt{D}}(f)}}({\mathrm{D}}',{\mathrm{D}}'')
\]
is the negative of the matrix product:
\[
\Psi\circ\Phi=-\begin{bmatrix}
0& \Psi^{\mathrm{od}}_{11}\Phi^{\mathrm{ev}}_{12} \\
\Psi^{\mathrm{od}}_{22}\Phi^{\mathrm{ev}}_{21} & 0
\end{bmatrix}\in{\mathrm{Hom}}^{\mathrm{ev}}_{{\mathscr{MF}^{\mathtt{D}}(f)}}({\mathrm{D}}',{\mathrm{D}}''').
\]
As another example, for an odd $\Phi$ as above
\begin{equation}\label{krs}
{\boldsymbol{\bar{\partial}}}(\Phi)=-\begin{bmatrix}
0 & \bar{\partial}\Phi^{\mathrm{ev}}_{12} \\
\bar{\partial}\Phi^{\mathrm{ev}}_{21} & 0
\end{bmatrix}.
\end{equation}
\medskip
\subsection{Equivalences between the models}
One has obvious dg functors
\begin{eqnarray}\label{dgf}
{\mathscr{MF}(f)}\stackrel{{{\rm I}^{\sf h}}}\longrightarrow
{\mathscr{MF}^{\tt h}(f)}\stackrel{{{\rm I}^{\tt D}}}\longrightarrow{\mathscr{MF}^{\mathtt{D}}(f)}\stackrel{{{\rm I}_{\tt c}}}\longleftarrow{\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}
\end{eqnarray}
induced by the natural embeddings of Hom-complexes
\[
{\mathrm{Hom}}^*_{{\mathscr{MF}(f)}}({\cdot,\cdot})\stackrel{{i^{\sf h}}}\hookrightarrow
{\mathrm{Hom}}^*_{{\mathscr{MF}^{\tt h}(f)}}({\cdot,\cdot})\stackrel{{i^{\tt D}}}\hookrightarrow{\mathrm{Hom}}^*_{{\mathscr{MF}^{\mathtt{D}}(f)}}({\cdot,\cdot})\stackrel{{i_{\tt c}}}\hookleftarrow{\mathrm{Hom}}^*_{{\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}}({\cdot,\cdot}).
\]
\begin{proposition}\label{main_s1}
The dg functor ${{\rm I}^{\sf h}}$, ${{\rm I}^{\tt D}}$ and ${{\rm I}_{\tt c}}$ are quasi-equivalences.
\end{proposition}
\noindent{\bf Proof.} We have to show that the embeddings ${i^{\sf h}}$, ${i^{\tt D}}$ and ${i_{\tt c}}$ are quasi-isomorphisms. In what follows ${\mathrm{D}}'$ and ${\mathrm{D}}''$ are two arbitrary matrix factorizations.
\medskip
\noindent\underline{\it Proof for ${i^{\sf h}}$:} Since $\mathcal{A}({X^{\mathtt{h}}})$ is flat as a ${\mathbb{C}}[X]$-module \cite[Sect.A1.2]{N}, it suffices to prove that the natural map
\[
{\mathrm{H}}^{*}({\mathrm{Hom}}_{{\mathscr{MF}(f)}}({\mathrm{D}}',{\mathrm{D}}''),\delta)\simeq{\mathrm{H}}^{*}({\mathrm{Hom}}_{{\mathscr{MF}(f)}}({\mathrm{D}}',{\mathrm{D}}''),\delta)\otimes_{{\mathbb{C}}[X]}\mathcal{A}({X^{\mathtt{h}}})
\]
is an isomorphism. As we know from the proof of Proposition \ref{mfpr}, the sheaf of $\mathcal{O}_X$-modules, underlying the ${\mathbb{C}}[X]$-module ${\mathrm{H}}^*({\mathrm{Hom}}_{{\mathscr{MF}(f)}}({\mathrm{D}}',{\mathrm{D}}''),\delta)$, is a coherent sheaf supported at the points of $C_f$. In particular, the support is proper and, as a consequence of GAGA, the space of global sections of the sheaf does not change upon the analytification.
\medskip
\noindent\underline{\it Proof for ${i^{\tt D}}$:} That
$
{i^{\tt D}}
$
is a quasi-isomorphism is an elementary consequence of the Dolbeault theorem (and the fact that ${X^{\mathtt{h}}}$ is Stein).
\medskip
\noindent\underline{\it Proof for ${i_{\tt c}}$:} To show that ${i_{\tt c}}$
is a quasi-isomorphism we will construct an explicit inverse-up-to-homotopy
by mimicking an idea from \cite[Sect.2.2]{LLS}.
Let us fix a Hermitian metric $\langle\cdot,\cdot\rangle$ on the bundle of $(1,0)$-forms on ${X^{\mathtt{h}}}$. Then one has the following analog of (\ref{homot1}):
on ${X^{\mathtt{h}}}\setminus C_f$
\begin{equation}\label{homot4}
{\mathrm{D}}''{\mathrm{H}}_{{\mathrm{D}}''}+{\mathrm{H}}_{{\mathrm{D}}''}{\mathrm{D}}''=\mathbf{1},\quad {\mathrm{H}}_{{\mathrm{D}}''}:=\frac{\langle\partial {\mathrm{D}}'',\,\partial f\rangle}{||\partial f||^2}.
\end{equation}
Here $||\cdot||$ stands for the norm associated with the metric and $\langle\partial {\mathrm{D}}'',\,\partial f\rangle$ is the result of applying $\langle {\cdot},\,\partial f\rangle$ to every entry of $\partial {\mathrm{D}}''$ (thus, ${\mathrm{H}}_{{\mathrm{D}}''}$ is a matrix of smooth functions on ${X^{\mathtt{h}}}\setminus C_f$). If we view ${\mathrm{H}}_{{\mathrm{D}}''}$ as an element of ${\mathrm{End}}^{\mathrm{od}}_{{\mathscr{MF}}^{\sf {\mathrm{D}}}({X^{\mathtt{h}}}\setminus C_f,f)}({\mathrm{D}}'')$ then (\ref{homot4}) means
$
{\boldsymbol{\delta}}({\mathrm{H}}_{{\mathrm{D}}''})={id}_{{\mathrm{D}}''}.
$
By analogy with the algebraic case, we obtain an odd endomorphism
\begin{equation}\label{mult1}
{{\fH}}(\Phi):={\mathrm{H}}_{{\mathrm{D}}''}\circ\Phi
\end{equation} of ${\mathrm{Hom}}^*_{{\mathscr{MF}}^{\tt D}({X^{\mathtt{h}}}\setminus C_f,f)}({\mathrm{D}}',{\mathrm{D}}'')$ with the property
\begin{equation}\label{homot3}
\Phi=[{\boldsymbol{\delta}},{{\fH}}](\Phi).
\end{equation}
Furthermore, note that the operator $[{\boldsymbol{\bar{\partial}}},{{\fH}}]$ on ${\mathrm{Hom}}^*_{{\mathscr{MF}}^{\tt D}({X^{\mathtt{h}}}\setminus C_f,f)}({\mathrm{D}}',{\mathrm{D}}'')$ is nilpotent:
\begin{eqnarray}\label{mult2}
[{\boldsymbol{\bar{\partial}}},{{\fH}}](\Phi)=({\boldsymbol{\bar{\partial}}}{{\fH}}+{{\fH}}{\boldsymbol{\bar{\partial}}})(\Phi)={\boldsymbol{\bar{\partial}}}({\mathrm{H}}_{{\mathrm{D}}''})\circ\Phi=-\bar{\partial}{\mathrm{H}}_{{\mathrm{D}}''}\circ\Phi
\end{eqnarray}
(the last equality is a special case of (\ref{krs})).
Together with (\ref{homot3}) this implies that $[\bar{\boldsymbol{\delta}},{{\fH}}]$ is an invertible operator on ${\mathrm{Hom}}^*_{{\mathscr{MF}}^{\tt D}({X^{\mathtt{h}}}\setminus C_f,f)}({\mathrm{D}}',{\mathrm{D}}'')$. Consider the operator
\begin{equation*}\label{Sop}
{\cV}:={{\fH}} \cdot [\bar{\boldsymbol{\delta}},{{\fH}}]^{-1}={{\fH}} \cdot (1+[{\boldsymbol{\bar{\partial}}},{{\fH}}])^{-1}.
\end{equation*}
Obviously, $[\bar{\boldsymbol{\delta}},{\cV}]$ is the identity operator on ${\mathrm{Hom}}^*_{{\mathscr{MF}}^{\tt D}({X^{\mathtt{h}}}\setminus C_f,f)}({\mathrm{D}}',{\mathrm{D}}'')$. As a consequence of (\ref{mult1}) and (\ref{mult2}),
\begin{equation}\label{G}
{\cV}(\Phi)={\mathscr{H}}_{{\mathrm{D}}''}\circ\Phi, \quad {\mathscr{H}}_{{\mathrm{D}}''}=\sum_{i}{\mathrm{H}}_{{\mathrm{D}}''}\circ(\bar{\partial}{\mathrm{H}}_{{\mathrm{D}}''})^{\circ i}.
\end{equation}
We would like to ``extend'' ${\cV}$ the whole of ${X^{\mathtt{h}}}$.
Let us fix a smooth function $\varrho$ on ${X^{\mathtt{h}}}$ such that
\begin{equation}\label{cutoff}
\varrho|_{{{U}}_1} = 1,\quad\varrho|_{{X^{\mathtt{h}}}\setminus {{U}}_2} = 0
\end{equation} for some relatively
compact open neighborhoods (in the analytic topology) ${{{U}}_1}\subset\overline{{{U}}}_1\subset{{{U}}_2}$ of $C_f$ and set
\begin{eqnarray*}
\widehat{\cV}:=(1-\varrho)\cdot{\cV}: {\mathrm{Hom}}^{{\mathrm{ev}}/{\mathrm{od}}}_{{\mathscr{MF}}^{\tt D}({X^{\mathtt{h}}},f)}({\mathrm{D}}',{\mathrm{D}}'')\to {\mathrm{Hom}}^{{\mathrm{od}}/{\mathrm{ev}}}_{{\mathscr{MF}}^{\tt D}({X^{\mathtt{h}}},f)}({\mathrm{D}}',{\mathrm{D}}'')
\end{eqnarray*}
($\widehat{\cV}$ is well-defined on the whole of ${X^{\mathtt{h}}}$ because of (\ref{cutoff})). Thanks to (\ref{G}) $\widehat{\cV}$ is the operator of left multiplication with an element of ${\mathrm{End}}^{\mathrm{od}}_{{\mathscr{MF}}^{\tt D}({X^{\mathtt{h}}},f)}({\mathrm{D}}'')$:
\begin{equation}\label{hod}
\widehat{\cV}(\Phi)=\widehat{\cV}_{{\mathrm{D}}''}\circ\Phi, \quad \widehat{\cV}_{{\mathrm{D}}''}=(1-\varrho)\cdot{\mathscr{H}}_{{\mathrm{D}}''}=(1-\varrho)\cdot\sum_{i}{\mathrm{H}}_{{\mathrm{D}}''}\circ(\bar{\partial}{\mathrm{H}}_{{\mathrm{D}}''})^{\circ i}.
\end{equation}
In particular, $\widehat{\cV}$ preserves the subspace ${\mathrm{Hom}}^*_{{\mathscr{MF}}^{\tt D}_{\tt c}({X^{\mathtt{h}}},f)}({\mathrm{D}}',{\mathrm{D}}'')$. We have
\begin{eqnarray}\label{calc}
[\bar{\boldsymbol{\delta}},\widehat{\cV}]=[\bar{\boldsymbol{\delta}},(1-\varrho)]{\cV} +(1-\varrho)[\bar{\boldsymbol{\delta}},{\cV}]=-\bar{\partial}\varrho\,{\cV}+(1-\varrho)= 1-(\varrho+\bar{\partial}\varrho\,{\cV}).
\end{eqnarray}
Thus, the morphism of complexes
\begin{eqnarray*}
{\pi}:=\varrho+\bar{\partial}\varrho\cdot{\cV}: {\mathrm{Hom}}^*_{{\mathscr{MF}}^{\tt D}({X^{\mathtt{h}}},f)}({\mathrm{D}}',{\mathrm{D}}'')\to {\mathrm{Hom}}^*_{{\mathscr{MF}}^{\tt D}_{\tt c}({X^{\mathtt{h}}},f)}({\mathrm{D}}',{\mathrm{D}}'')
\end{eqnarray*}
satisfies $id-{i_{\tt c}}{\pi}=[\bar{\boldsymbol{\delta}},\widehat{\cV}]$ and $id-{\pi}{i_{\tt c}}=[\bar{\boldsymbol{\delta}},\widehat{\cV}]$ which
finishes the proof.
\hfill$\blacksquare$
\subsection{An $A_\infty$ functor}
Unlike ${i_{\tt c}}$, its homotopy inverse ${\pi}$ that we have just constructed is not compatible with the composition of morphisms, i.~e. does not define a dg functor. Our goal now is to promote ${\pi}$ to an {\it $A_\infty$ functor} ${\boldsymbol{\pi}}:{\mathscr{MF}^{\mathtt{D}}(f)}\to{\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}$. That is, we want to find a collection $\{\pi_l\}_{l\geq2}$ of odd maps
\begin{gather*}
{{{\mathit{s}}}}{\mathrm{Hom}}^*_{{\mathscr{MF}^{\mathtt{D}}(f)}}({\mathrm{D}}^{(l)},{\mathrm{D}}^{(l+1)})\otimes\ldots\otimes{{{\mathit{s}}}}{\mathrm{Hom}}^*_{{\mathscr{MF}^{\mathtt{D}}(f)}}({\mathrm{D}}^{(2)},{\mathrm{D}}^{(3)})
\otimes {{{\mathit{s}}}}{\mathrm{Hom}}^*_{{\mathscr{MF}^{\mathtt{D}}(f)}}({\mathrm{D}}^{(1)},{\mathrm{D}}^{(2)})\\
\Big\downarrow{\pi_l}
\\
{\mathrm{Hom}}^*_{{\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}}({\mathrm{D}}^{(1)},{\mathrm{D}}^{(l+1)})
\end{gather*}
(for arbitrary ${\mathrm{D}}^{(1)},\ldots, {\mathrm{D}}^{(l+1)}$) that together with $\pi_1:={\pi}\cdot {{{\mathit{s}}}}$ satisfy the relations
\begin{multline}\label{ainfrel}
\sum_{i=1}^{l-1} (-1)^{\epsilon_{i+1}}\left(\pi_{l-i}(\Phi_l|\ldots|\Phi_{i+1})\circ\pi_i(\Phi_i|\ldots|\Phi_{1})-\pi_{l-1}(\Phi_l|\ldots|\Phi_{i+1}\circ\Phi_i|\ldots|\Phi_{1})\right)
=\\
=\bar{\boldsymbol{\delta}}\pi_l(\Phi_l|\ldots|\Phi_{1})-\sum_{i=1}^{l} (-1)^{\epsilon_{i+1}}\pi_{l}(\Phi_l| \ldots|\bar{\boldsymbol{\delta}}\Phi_i |\ldots|\Phi_{1})
\end{multline}
where $\Phi_i\in {\mathrm{Hom}}^*_{{\mathscr{MF}^{\mathtt{D}}(f)}}({\mathrm{D}}^{(i)},{\mathrm{D}}^{(i+1)})$,
$
\Phi_l|\ldots|\Phi_{1}:={{{\mathit{s}}}}\Phi_l\otimes\ldots\otimes{{{\mathit{s}}}}\Phi_1
$
($\epsilon_{i+1}$ here and in the rest of the paper has the same meaning as in Section \ref{cypc} and in Theorem \ref{mainresult}).
By (\ref{G}) ${\pi}$ is the operator of left multiplication with an even element, namely
\begin{equation}\label{pid}
{\pi}(\Phi)=\pi_{{\mathrm{D}}''}\circ\Phi, \quad \pi_{{\mathrm{D}}''}:=\varrho+\bar{\partial}\varrho\cdot{\mathscr{H}}_{{\mathrm{D}}''}=\varrho+\bar{\partial}\varrho\cdot\sum_{i}{\mathrm{H}}_{{\mathrm{D}}''}\circ(\bar{\partial}{\mathrm{H}}_{{\mathrm{D}}''})^{\circ i}.
\end{equation}
Recall also the odd elements $\widehat{\cV}_{{\mathrm{D}}''}$ defined in (\ref{hod}).
\begin{proposition}\label{ainffun} The maps
\begin{eqnarray*}\label{ainfty}
\pi_l(\Phi_l|\ldots|\Phi_{1})=\pi_{{\mathrm{D}}^{(l+1)}}\circ\Phi_l\circ \widehat{\cV}_{{\mathrm{D}}^{(l)}}\circ\Phi_{l-1}\circ\ldots \circ\widehat{\cV}_{{\mathrm{D}}^{(2)}}\circ \Phi_1
\end{eqnarray*}
define an $A_\infty$ functor ${\boldsymbol{\pi}}:{\mathscr{MF}^{\mathtt{D}}(f)}\to{\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}$.
\end{proposition}
\noindent{\bf Proof.} By the Leibniz rule
\begin{multline*}
\bar{\boldsymbol{\delta}}\pi_l(\Phi_l|\ldots|\Phi_{1})=\bar{\boldsymbol{\delta}}\left(\pi_{{\mathrm{D}}^{(l+1)}}\circ\Phi_l\circ \widehat{\cV}_{{\mathrm{D}}^{(l)}}\circ\ldots \circ\widehat{\cV}_{{\mathrm{D}}^{(2)}}\circ \Phi_1\right)=\\
=\sum_i (-1)^{\epsilon_{i+1}} \pi_{{\mathrm{D}}^{(l+1)}}\circ \ldots\circ\widehat{\cV}_{{\mathrm{D}}^{(i+1)}}\circ \bar{\boldsymbol{\delta}}\Phi_{i}\circ\widehat{\cV}_{{\mathrm{D}}^{(i)}}\circ \ldots \circ \Phi_1-\\
-\sum_i (-1)^{\epsilon_{i+1}} \pi_{{\mathrm{D}}^{(l+1)}}\circ \ldots\circ \Phi_{i+1}\circ\bar{\boldsymbol{\delta}}\widehat{\cV}_{{\mathrm{D}}^{(i+1)}}\circ\Phi_{i}\circ \ldots \circ \Phi_1
\end{multline*}
and therefore
\begin{multline*}
\bar{\boldsymbol{\delta}}\pi_l(\Phi_l|\ldots|\Phi_{1})-\sum_{i} (-1)^{\epsilon_{i+1}}\pi_{l}(\Phi_l| \ldots|\bar{\boldsymbol{\delta}}\Phi_i |\ldots|\Phi_{1})=\\
=-\sum_i (-1)^{\epsilon_{i+1}}\pi_{{\mathrm{D}}^{(l+1)}}\circ \ldots\circ \Phi_{i+1}\circ\bar{\boldsymbol{\delta}}\widehat{\cV}_{{\mathrm{D}}^{(i+1)}}\circ\Phi_{i}\circ \ldots \circ \Phi_1.
\end{multline*}
By (\ref{calc}) $\bar{\boldsymbol{\delta}}\widehat{\cV}_{{\mathrm{D}}^{(i+1)}}=id_{{\mathrm{D}}^{(i+1)}}-\pi_{{\mathrm{D}}^{(i+1)}}$. Hence
\begin{eqnarray*}
&&-\sum_i (-1)^{\epsilon_{i+1}} \pi_{{\mathrm{D}}^{(l+1)}}\circ \ldots\circ \Phi_{i+1}\circ\bar{\boldsymbol{\delta}}\widehat{\cV}_{{\mathrm{D}}^{(i+1)}}\circ\Phi_{i}\circ \ldots \circ \Phi_1=\\
&&=\sum_i (-1)^{\epsilon_{i+1}} \pi_{{\mathrm{D}}^{(l+1)}}\circ \ldots\circ \Phi_{i+1}\circ(\pi_{{\mathrm{D}}^{(i+1)}}-id_{{\mathrm{D}}^{(i+1)}})\circ\Phi_{i}\circ \ldots \circ \Phi_1=\\
&&=\sum_i (-1)^{\epsilon_{i+1}} (\pi_{{\mathrm{D}}^{(l+1)}}\circ \ldots\circ \widehat{\cV}_{{\mathrm{D}}^{(i+2)}}\circ \Phi_{i+1})\circ(\pi_{{\mathrm{D}}^{(i+1)}}\circ\Phi_{i}\circ \widehat{\cV}_{{\mathrm{D}}^{(i)}}\circ \ldots \circ \Phi_1)-\\
&&-\sum_i (-1)^{\epsilon_{i+1}} \pi_{{\mathrm{D}}^{(l+1)}}\circ \ldots\circ \widehat{\cV}_{{\mathrm{D}}^{(i+2)}}\circ(\Phi_{i+1}\circ\Phi_{i})\circ \widehat{\cV}_{{\mathrm{D}}^{(i)}}\circ \ldots \circ \Phi_1=\\
&&=\sum_{i} (-1)^{\epsilon_{i+1}}\left(\pi_{l-i}(\Phi_l|\ldots|\Phi_{i+1})\circ\pi_i(\Phi_i|\ldots|\Phi_{1})-\pi_{l-1}(\Phi_l|\ldots|\Phi_{i+1}\circ\Phi_i|\ldots|\Phi_{1})\right). \qquad\blacksquare
\end{eqnarray*}
\begin{remark}\label{ainf}{\rm By Proposition \ref{main_s1} ${\boldsymbol{\pi}}$ is not just an $A_\infty$ functor but an $A_\infty$ quasi-equivalence. By precomposing it with the dg quasi-equivalence ${{\rm I}^{\tt D}}\cdot{{\rm I}^{\sf h}}:{\mathscr{MF}(f)}\to{\mathscr{MF}^{\mathtt{D}}(f)}$ (i.~e. by restricting ${\boldsymbol{\pi}}$ to the dg subcategory ${\mathscr{MF}(f)}$) one obtains an $A_\infty$ quasi-equivalence ${\mathscr{MF}(f)}\to{\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}$ which we will still denote by ${\boldsymbol{\pi}}=\{\pi_l\}_{l\geq1}$.}
\end{remark}
\subsection{CY structures on ${\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}$}
Any holomorphic volume form $\Omega$ on ${X^{\mathtt{h}}}$ determines a linear functional
on ${\mathrm{End}}^*_{{\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}}({\mathrm{D}})$ (for every ${\mathrm{D}}$), namely
\begin{equation}\label{trdef}
{\Theta_{\tt c}}(\Phi):=\int_{{X^{\mathtt{h}}}} {{\sf str}}(\Phi)\wedge \Omega
\end{equation}
where ${{\sf str}}$ is the ${\mathscr{E}}_{\tt c}^{(0,*)}({X^{\mathtt{h}}})$-linear extension of the ordinary supertrace
\[
{{\sf str}}:{\mathrm{End}}^*_{{\mathscr{MF}(f)}}({\mathrm{D}})\to{\mathbb{C}}[X], \quad \begin{bmatrix}
\Phi_{11}& \Phi_{12} \\
\Phi_{21} & \Phi_{22}
\end{bmatrix}\mapsto \mathsf{tr}(\Phi_{11})-\mathsf{tr}(\Phi_{22}).
\]
Let us extend the resulting functional on ${\mathsf{C}}_*^{\{1\}}({\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)})$ to the whole of ${\mathsf{C}}_*({\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)})$ by setting ${\Theta_{\tt c}}|_{{\mathsf{C}}_*^{\{l\}}({\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)})}=0$ for $l\geq2$.
\begin{proposition} The extended functional ${\Theta_{\tt c}}$ is a chain-level CY structure on ${\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}$ (of the same parity as $n=\dim\,X$).
\end{proposition}
\noindent{\bf Proof.} The first equality in (\ref{clcy}) is vacuous in this case, while the second one follows from the following easy-to-check properties of ${\Theta_{\tt c}}$:
for any $\Phi\in {\mathrm{End}}^*_{{\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}}({\mathrm{D}})$
\begin{equation}\label{p1}
{\Theta_{\tt c}}({\boldsymbol{\delta}}\Phi)={\Theta_{\tt c}}({\boldsymbol{\bar{\partial}}}\Phi)=0\quad(\Rightarrow\,\, {\Theta_{\tt c}}(\bar{\boldsymbol{\delta}}\Phi)=0)
\end{equation}
and for any $\Phi\in {\mathrm{Hom}}^*_{{\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}}({\mathrm{D}}',{\mathrm{D}}'')$ and $\Psi\in {\mathrm{Hom}}^*_{{\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}}({\mathrm{D}}'',{\mathrm{D}}')$
\begin{equation}\label{p2}
{\Theta_{\tt c}}(\Phi\circ\Psi)=(-1)^{|\Phi||\Psi|}{\Theta_{\tt c}}(\Psi\circ\Phi).
\end{equation}
That the pairing
\begin{equation}\label{indpair}
{\mathrm{Hom}}^*_{{\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}}({\mathrm{D}}'',{\mathrm{D}}')\otimes {\mathrm{Hom}}^*_{{\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}}({\mathrm{D}}',{\mathrm{D}}'')\to {\mathbb{C}},\quad
\Psi\otimes \Phi\mapsto {\Theta_{\tt c}}(\Psi\circ\Phi)
\end{equation}
induces a non-degenerate pairing on the $\bar{\boldsymbol{\delta}}$-cohomology is a consequence of the classical Serre duality \cite{Ser}. Let us sketch the proof.
Thanks to Proposition \ref{main_s1}, it suffices to show that
\[
{\mathrm{Hom}}^*_{{\mathscr{MF}^{\mathtt{D}}(f)}}({\mathrm{D}}'',{\mathrm{D}}')\otimes {\mathrm{Hom}}^*_{{\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}}({\mathrm{D}}',{\mathrm{D}}'')\to {\mathbb{C}},\quad
\Psi\otimes \Phi\mapsto {\Theta_{\tt c}}(\Psi\circ\Phi)
\]
induces a non-degenerate pairing on the $\bar{\boldsymbol{\delta}}$-cohomology.
The spaces ${\mathscr{E}}^{(0,*)}({X^{\mathtt{h}}})$ are Fr\'echet spaces with respect to the topology
of uniform convergence of the forms together with all derivatives on the
compact subsets of ${X^{\mathtt{h}}}$ \cite[Sect.3]{Ser}. Let ${\mathscr{D}}^{(n,*)}({X^{\mathtt{h}}})$ denote the (strong) dual topological vector space of compactly supported currents on ${X^{\mathtt{h}}}$ of type $(n,*)$. Then the complex $({\mathrm{Hom}}^*_{{\mathscr{MF}(f)}}({\mathrm{D}}'',{\mathrm{D}}')\otimes_{{\mathbb{C}}[X]} {\mathscr{E}}^{(0,*)}({X^{\mathtt{h}}}), \bar{\boldsymbol{\delta}})$ can be viewed as a 2-periodic complex of Fr\'echet spaces (obviously, both differentials ${\boldsymbol{\delta}}$ and ${\boldsymbol{\bar{\partial}}}$ are continuous maps) and the dual topological complex (cf. \cite[Def.1.1]{LTL}) can be identified with \[({\mathrm{Hom}}^*_{{\mathscr{MF}(f)}}({\mathrm{D}}',{\mathrm{D}}'')\otimes_{{\mathbb{C}}[X]} {\mathscr{D}}^{(n,*)}({X^{\mathtt{h}}}), \bar{\boldsymbol{\delta}}^\vee={\boldsymbol{\delta}}^\vee+{\boldsymbol{\bar{\partial}}}^\vee)\] where $\vee$ indicates the transposed map.
By \cite[Thm.1.5,1.6]{LTL} the induced pairing
\begin{gather*}
{\mathrm{H}}^*\left({\mathrm{Hom}}_{{\mathscr{MF}(f)}}({\mathrm{D}}'',{\mathrm{D}}')\otimes_{{\mathbb{C}}[X]} {\mathscr{E}}^{(0,*)}({X^{\mathtt{h}}}), \bar{\boldsymbol{\delta}}\right)\otimes {\mathrm{H}}^*\left({\mathrm{Hom}}_{{\mathscr{MF}(f)}}({\mathrm{D}}',{\mathrm{D}}'')\otimes_{{\mathbb{C}}[X]} {\mathscr{D}}^{(n,*)}({X^{\mathtt{h}}}), \bar{\boldsymbol{\delta}}^\vee\right)
\\
\Big\downarrow
\\
{\mathbb{C}}
\end{gather*}
is non-degenerate because ${\mathrm{H}}^*\left({\mathrm{Hom}}_{{\mathscr{MF}(f)}}({\mathrm{D}}'',{\mathrm{D}}')\otimes_{{\mathbb{C}}[X]} {\mathscr{E}}^{(0,*)}({X^{\mathtt{h}}}), \bar{\boldsymbol{\delta}}\right)$ is finite-dimensional (Propositions \ref{mfpr}, \ref{main_s1}) and hence separated.
By (\ref{p1}) the form $\Omega$ gives rise to a natural embedding
\[
\left({\mathrm{Hom}}^*_{{\mathscr{MF}(f)}}({\mathrm{D}}',{\mathrm{D}}'')\otimes_{{\mathbb{C}}[X]}{\mathscr{E}}_{\tt c}^{(0,*)}, \bar{\boldsymbol{\delta}}\right)[n]\to\left({\mathrm{Hom}}^*_{{\mathscr{MF}(f)}}({\mathrm{D}}',{\mathrm{D}}'')\otimes_{{\mathbb{C}}[X]} {\mathscr{D}}^{(n,*)}({X^{\mathtt{h}}}), \bar{\boldsymbol{\delta}}^\vee\right)
\]
and it remains to explain why this embedding is a quasi-isomorphism. In fact, (\ref{p1}) implies more, namely, that the above embedding is a morphism of the underlying {\it double} complexes. Since both double complexes have bounded antidiagonals, it suffices to show that the morphisms
\[
\left({\mathrm{Hom}}^*_{{\mathscr{MF}(f)}}({\mathrm{D}}',{\mathrm{D}}'')\otimes_{{\mathbb{C}}[X]} {\mathscr{E}}_{\tt c}^{(0,*)}, {\boldsymbol{\bar{\partial}}}\right)[n]\to\left({\mathrm{Hom}}^*_{{\mathscr{MF}(f)}}({\mathrm{D}}',{\mathrm{D}}'')\otimes_{{\mathbb{C}}[X]} {\mathscr{D}}^{(n,*)}({X^{\mathtt{h}}}), {\boldsymbol{\bar{\partial}}}^\vee\right)
\]
are quasi-isomorphisms which is a special case of \cite[Thm.1]{Ser}.
\hfill$\blacksquare$
\subsection{CY structures on ${\mathscr{MF}(f)}$}
We want to pull back the above CY structure to ${\mathscr{MF}(f)}$ using the $A_\infty$ quasi-equivalence ${\boldsymbol{\pi}}=\{\pi_l\}_{l\geq1}: {\mathscr{MF}(f)}\to{\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}$ (cf. Remark \ref{ainf}).
Let $\widehat{{\boldsymbol{\pi}}}$ stand for the even linear map ${\mathsf{C}}_*({\mathscr{MF}(f)})\to {\mathsf{C}}_*^{\{1\}}({\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)})$ given by
\[
\widehat{{\boldsymbol{\pi}}}(\Phi_{l}[\Phi_{l-1}|\ldots|\Phi_1]):=\pi_{l}(\Phi_{l}|\Phi_{l-1}|\ldots|\Phi_1).
\]
Let also $N$ denote the endomorphism of ${\mathsf{C}}_*({\mathscr{MF}(f)})$ that acts as $\sum_{i=0}^{l-1}\tau^i$ on the tensors of length $l$ (here $\tau$ is the cyclic permutation (\ref{cycper})).
\begin{proposition} The functional
\begin{equation}\label{cymf}
{\Theta}={\Theta_{\tt c}}\cdot \widehat{{\boldsymbol{\pi}}}\cdot N: {\mathsf{C}}_*({\mathscr{MF}(f)})\to {\mathbb{C}}
\end{equation}
is a chain-level CY structure on ${\mathscr{MF}(f)}$ (of the same parity as $n=\dim\,X$).
\end{proposition}
\noindent{\bf Proof.} The non-degeneracy of the induced pairings
\[
{\mathrm{H}}^*({\mathrm{Hom}}_{{\mathscr{MF}(f)}}({\mathrm{D}}'',{\mathrm{D}}'), \delta)\otimes {\mathrm{H}}^*({\mathrm{Hom}}_{{\mathscr{MF}(f)}}({\mathrm{D}}',{\mathrm{D}}''), \delta)\to{\mathbb{C}}
\]
follows immediately from the non-degeneracy of the pairings induced by (\ref{indpair}) and the fact that ${\boldsymbol{\pi}}$ is an $A_\infty$ quasi-equivalence. The first property in (\ref{clcy}) is obvious since $N\cdot(1-\tau)=0$. The second property in (\ref{clcy}) follows from \cite[Lem.2.4]{FLS} but we supply an independent proof.
It is easy to see that $N\cdot b(\delta)=b(\delta)\cdot N$ and $N\cdot b(\mu)=b''(\mu)\cdot N$
where
\begin{multline*}
b''(\mu)(\Phi_{l}[\Phi_{l-1}|\ldots |\Phi_1])=
(-1)^{|\Phi_{l}|}\Phi_{l}\Phi_{l-1}[\Phi_{l-2}|\ldots |\Phi_1]-\\-\sum\limits_{i=1}^{l-2}(-1)^{\epsilon_{i+1}}\Phi_{l}[\Phi_{l-1}|\ldots
|\Phi_{i+1}\Phi_{i}|\ldots|\Phi_1].
\end{multline*}
Therefore, ${\Theta}\cdot b={\Theta_{\tt c}}\cdot \widehat{{\boldsymbol{\pi}}}\cdot(b(\delta)+b''(\mu))\cdot N$.
Let us calculate the functional ${\Theta_{\tt c}}\cdot \widehat{{\boldsymbol{\pi}}}\cdot(b(\delta)+b''(\mu))$ explicitly:
\begin{eqnarray*}
&&{\Theta_{\tt c}}\cdot \widehat{{\boldsymbol{\pi}}}\cdot(b(\delta)+b''(\mu))(\Phi_{l}[\Phi_{l-1}|\ldots |\Phi_1])=\\
&&={\Theta_{\tt c}}\left(\sum\limits_{i=1}^{l}(-1)^{\epsilon_{i+1}}\pi_{l}(\Phi_{l}|\ldots |\delta\Phi_i|\ldots|\Phi_1)-\sum\limits_{i=1}^{l-1}(-1)^{\epsilon_{i+1}}\pi_{l-1}(\Phi_{l}|\ldots
|\Phi_{i+1}\Phi_{i}|\ldots|\Phi_1\right)=\\
&&\stackrel{\bf (\ref{ainfrel})}={\Theta_{\tt c}}\left(\bar{\boldsymbol{\delta}}\pi_{l}(\Phi_{l}|\ldots|\Phi_{1})-\sum_{i=1}^{l-1} (-1)^{\epsilon_{i+1}}\pi_{l-i}(\Phi_{l}|\ldots|\Phi_{i+1})\circ\pi_i(\Phi_i|\ldots|\Phi_{1})\right)=\\
&&\stackrel{\bf(\ref{p1})}=-\sum_{i=1}^{l-1} (-1)^{\epsilon_{i+1}}{\Theta_{\tt c}}\left(\pi_{l-i}(\Phi_{l}|\ldots|\Phi_{i+1})\circ\pi_i(\Phi_i|\ldots|\Phi_{1})\right).
\end{eqnarray*}
As a result,
\begin{equation*}
{\Theta}\cdot b={\Theta_{\tt c}}\cdot \widehat{{\boldsymbol{\pi}}}\cdot(b(\delta)+b''(\mu))\cdot N=\sum_{i=1}^{l-1}{\Theta_{\tt c}}\cdot\mu\cdot(\pi_{l-i}\otimes\pi_i)\cdot N
\end{equation*}
where $\mu$ denotes the composition of morphisms in ${\mathscr{MF}_{\tt c}^{\mathtt{D}}(f)}$ and
\begin{equation}\label{pipi}
(\pi_{l-i}\otimes\pi_i)(\Phi_{l}[\Phi_{l-1}|\ldots |\Phi_1]):=(-1)^{\epsilon_{i+1}+1}\pi_{l-i}(\Phi_{l}|\ldots|\Phi_{i+1})\otimes\pi_i(\Phi_i|\ldots|\Phi_{1}).
\end{equation}
Since $\tau\cdot N=N$, one has
\[
\sum_{i=1}^{l-1}{\Theta_{\tt c}}\cdot\mu\cdot(\pi_{l-i}\otimes\pi_i)\cdot N=\sum_{i=1}^{l-1}{\Theta_{\tt c}}\cdot\mu\cdot(\pi_{i}\otimes\pi_{l-i})\cdot \tau^{-i}\cdot N.
\]
Observe that
\begin{eqnarray*}
&&{\Theta_{\tt c}}\cdot\mu\cdot(\pi_{i}\otimes\pi_{l-i})\cdot \tau^{-i}(\Phi_{l}[\Phi_{l-1}|\ldots |\Phi_1])=\\
&&=
(-1)^{|{{{\mathit{s}}}} \Phi_1|(\epsilon_1-|{{{\mathit{s}}}} \Phi_1|)+\ldots+|{{{\mathit{s}}}} \Phi_i|(\epsilon_1-|{{{\mathit{s}}}} \Phi_i|)}\cdot{\Theta_{\tt c}}\cdot\mu\cdot(\pi_{i}\otimes\pi_{l-i})(\Phi_{i}[\Phi_{i-1|}|\ldots |\Phi_1|\Phi_{l}|\ldots|\Phi_{i+1}])=\\
&&=
(-1)^{\epsilon_{i+1}(\epsilon_1-\epsilon_{i+1})}\cdot{\Theta_{\tt c}}\cdot\mu\cdot(\pi_{i}\otimes\pi_{l-i})(\Phi_{i}[\Phi_{i-1|}|\ldots |\Phi_1|\Phi_{l}|\ldots|\Phi_{i+1}])=\\
&&\stackrel{\bf (\ref{pipi})}=(-1)^{\epsilon_{i+1}(\epsilon_1-\epsilon_{i+1})}\cdot(-1)^{\epsilon_1-\epsilon_{i+1}+1}\cdot{\Theta_{\tt c}}\left(\pi_{i}(\Phi_{i}|\ldots|\Phi_{1})\circ\pi_{l-i}(\Phi_{l}|\ldots|\Phi_{i+1})\right)=\\
&&=(-1)^{\epsilon_{i+1}\epsilon_1+\epsilon_{1}+1}\cdot{\Theta_{\tt c}}\left(\pi_{i}(\Phi_{i}|\ldots|\Phi_{1})\circ\pi_{l-i}(\Phi_{l}|\ldots|\Phi_{i+1})\right)=\\
&&\stackrel{\bf (\ref{p2})}=(-1)^{\epsilon_{i+1}\epsilon_1+\epsilon_{1}+1}\cdot(-1)^{(\epsilon_{i+1}+1)(\epsilon_1-\epsilon_{i+1}+1)}\cdot{\Theta_{\tt c}}\left(\pi_{l-i}(\Phi_{l}|\ldots|\Phi_{i+1})\circ\pi_{i}(\Phi_{i}|\ldots|\Phi_{1})\right)=\\
&&=(-1)^{\epsilon_{i+1}}\cdot{\Theta_{\tt c}}\left(\pi_{l-i}(\Phi_{l}|\ldots|\Phi_{i+1})\circ\pi_{i}(\Phi_{i}|\ldots|\Phi_{1})\right)=\\
&&\stackrel{\bf (\ref{pipi})}=-{\Theta_{\tt c}}\cdot\mu\cdot(\pi_{l-i}\otimes\pi_{i})(\Phi_{l}[\Phi_{l-1}|\ldots |\Phi_1]).
\end{eqnarray*}
and therefore
\[
\sum_{i=1}^{l-1}{\Theta_{\tt c}}\cdot\mu\cdot(\pi_{l-i}\otimes\pi_i)\cdot N=\sum_{i=1}^{l-1}{\Theta_{\tt c}}\cdot\mu\cdot(\pi_{i}\otimes\pi_{l-i})\cdot \tau^{-i}\cdot N=-\sum_{i=1}^{l-1}{\Theta_{\tt c}}\cdot\mu\cdot(\pi_{l-i}\otimes\pi_i)\cdot N.
\quad \blacksquare
\]
\subsection{Explicit formulas in terms of residues} Apart from the form $\Omega$, the CY structure (\ref{cymf}) depends on the following data:
\begin{itemize}
\item[(a)] the Hermitian metric $\langle\cdot,\cdot\rangle$ on the bundle of $(1,0)$-forms that we used to construct the ``homotopies'' in (\ref{homot4});
\item[(b)] the neighborhoods ${{{U}}_1},{{{U}}_2}$ of $C_f$ and
the smooth function $\varrho$ satisfying the conditions (\ref{cutoff}) for these neighborhoods;
\end{itemize}
The proof of Theorem \ref{mainresult} amounts now to evaluating ${\Theta}$ for a special choice of (a) and (b). The goal of this final section is to perform this calculation. (A similar calculation was carried out in \cite{LLS} -- cf. the proof of Proposition 2.5 therein -- but our case is somewhat more involved.)
Let us start by expanding the formula (\ref{cymf}) in the general case. As before, we fix some matrix factorizations $\{{\mathrm{D}}^{(i)}\}_{i=1,\ldots, l}$ and morphisms $\Phi_{l}\in{\mathrm{Hom}}^*_{{\mathscr{MF}(f)}}({\mathrm{D}}^{(l)},{\mathrm{D}}^{(1)})$ and $\Phi_i\in{\mathrm{Hom}}^*_{{\mathscr{MF}(f)}}({\mathrm{D}}^{(i)},{\mathrm{D}}^{(i+1)})$, $i\leq l-1$.
We have
\begin{multline*}
{\Theta}(\Phi_{l}[\Phi_{l-1}|\ldots|\Phi_1])={\Theta_{\tt c}}\cdot \widehat{{\boldsymbol{\pi}}}\cdot N(\Phi_{l}[\Phi_{l-1}|\ldots|\Phi_1])=\\
=\sum_{i}(-1)^{\epsilon_{i}(\epsilon_1-\epsilon_{i})}{\Theta_{\tt c}}\cdot\pi_{l}(\Phi_{i-1}|\ldots |\Phi_1|\Phi_{l}|\ldots|\Phi_{i})=\\
\stackrel{\bf Prop.\ref{ainffun}}=\sum_{i}(-1)^{\epsilon_{i}(\epsilon_1-\epsilon_{i})}{\Theta_{\tt c}}(\pi_{{\mathrm{D}}^{(i)}}\circ\Phi_{i-1}\circ\widehat{\cV}_{{\mathrm{D}}^{(i-1)}}\circ \ldots \circ \Phi_1\circ \widehat{\cV}_{{\mathrm{D}}^{(1)}}\circ \Phi_{l}\circ \widehat{\cV}_{{\mathrm{D}}^{(l)}}\circ \ldots\circ \widehat{\cV}_{{\mathrm{D}}^{(i+1)}}\circ \Phi_{i})=\\
\stackrel{\bf (\ref{p2})}=\sum_{i=1}^{l}(-1)^{\epsilon_1-\epsilon_{i}}{\Theta_{\tt c}}((\Phi_{l}\circ \widehat{\cV}_{{\mathrm{D}}^{(l)}})\circ \ldots\circ (\Phi_{i}\circ \pi_{{\mathrm{D}}^{(i)}})\circ\ldots\circ (\Phi_1\circ \widehat{\cV}_{{\mathrm{D}}^{(1)}}))=\\
\stackrel{\bf (\ref{hod}),(\ref{pid})}=\sum_{i=1}^{l}(-1)^{\epsilon_1-\epsilon_{i}}{\Theta_{\tt c}}\left((\Phi_{l}\circ{\mathscr{H}}_{{\mathrm{D}}^{(l)}})\circ \ldots\circ\Phi_i\circ\ldots\circ (\Phi_1\circ {\mathscr{H}}_{{\mathrm{D}}^{(1)}}) \varrho(1-\varrho)^{l-1}\right)-\\
-\sum_{i=1}^{l}{\Theta_{\tt c}}\left((\Phi_{l}\circ{\mathscr{H}}_{{\mathrm{D}}^{(l)}})\circ \ldots\circ (\Phi_1\circ {\mathscr{H}}_{{\mathrm{D}}^{(1)}}) \bar{\partial}\varrho(1-\varrho)^{l-1}\right)=\\
\stackrel{\bf (\ref{trdef})}=\sum_{i=1}^{l}(-1)^{\epsilon_1-\epsilon_{i}}\int_{{X^{\mathtt{h}}}} {{\sf str}}\left((\Phi_{l}\circ{\mathscr{H}}_{{\mathrm{D}}^{(l)}})\circ \ldots\circ\Phi_i\circ\ldots\circ (\Phi_1\circ {\mathscr{H}}_{{\mathrm{D}}^{(1)}})\right) \varrho(1-\varrho)^{l-1}\wedge \Omega-\\
-l\int_{{X^{\mathtt{h}}}} {{\sf str}}\left((\Phi_{l}\circ{\mathscr{H}}_{{\mathrm{D}}^{(l)}})\circ \ldots\circ (\Phi_1\circ {\mathscr{H}}_{{\mathrm{D}}^{(1)}})\right) \bar{\partial}\varrho(1-\varrho)^{l-1}\wedge \Omega=\\
=\sum_{i=1}^{l}(-1)^{\epsilon_1-\epsilon_{i}}\int_{{X^{\mathtt{h}}}} {{\sf str}}\left((\Phi_{l}\circ{\mathscr{H}}_{{\mathrm{D}}^{(l)}})\circ \ldots\circ\Phi_i\circ\ldots\circ (\Phi_1\circ {\mathscr{H}}_{{\mathrm{D}}^{(1)}})\right) \varrho(1-\varrho)^{l-1}\wedge \Omega+\\
+\int_{{X^{\mathtt{h}}}} {{\sf str}}\left((\Phi_{l}\circ{\mathscr{H}}_{{\mathrm{D}}^{(l)}})\circ \ldots\circ (\Phi_1\circ {\mathscr{H}}_{{\mathrm{D}}^{(1)}})\right) \bar{\partial}\left((1-\varrho)^{l}\right)\wedge \Omega.
\end{multline*}
From now on,
\begin{itemize}
\item[(a)] $\langle\sum_ig'_idz_i,\sum_jg''_jdz_j\rangle:=\sum_ig'_i\overline{g''_i}$;
\item[(b)] ${{{U}}_2}=\cup_{x\in C_f} B_{2r}(x)$ and ${{{U}}_1}=\cup_{x\in C_f} B_{r}(x)$ where $B_{r}(x)$ stands for the open ball of radius $r$ centered at $x$ (for the standard metric on ${\mathbb{C}}^n$). We assume that $r$ is small enough, so that $\overline{B_{2r}(x)}\subset {X^{\mathtt{h}}}$ and $\overline{B_{2r}(x)}\cap \overline{B_{2r}(y)}=\varnothing$ for two different $x,y\in C_f$.
\end{itemize}
Then
\begin{multline*}
\sum_{i=1}^{l}(-1)^{\epsilon_1-\epsilon_{i}}\int_{{X^{\mathtt{h}}}} {{\sf str}}\left((\Phi_{l}\circ{\mathscr{H}}_{{\mathrm{D}}^{(l)}})\circ \ldots\circ\Phi_i\circ\ldots\circ (\Phi_1\circ {\mathscr{H}}_{{\mathrm{D}}^{(1)}})\right) \varrho(1-\varrho)^{l-1}\wedge \Omega+\\
+\int_{{X^{\mathtt{h}}}} {{\sf str}}\left((\Phi_{l}\circ{\mathscr{H}}_{{\mathrm{D}}^{(l)}})\circ \ldots\circ (\Phi_1\circ {\mathscr{H}}_{{\mathrm{D}}^{(1)}})\right) \bar{\partial}\left((1-\varrho)^{l}\right)\wedge \Omega=\\
=\sum_{i=1}^{l}(-1)^{\epsilon_1-\epsilon_{i}}\int_{\overline{U}_2\setminus U_1} {{\sf str}}\left((\Phi_{l}\circ{\mathscr{H}}_{{\mathrm{D}}^{(l)}})\circ \ldots\circ\Phi_i\circ\ldots\circ (\Phi_1\circ {\mathscr{H}}_{{\mathrm{D}}^{(1)}})\right) \varrho(1-\varrho)^{l-1}\wedge \Omega+\\
+\int_{\overline{U}_2\setminus U_1} {{\sf str}}\left((\Phi_{l}\circ{\mathscr{H}}_{{\mathrm{D}}^{(l)}})\circ \ldots\circ (\Phi_1\circ {\mathscr{H}}_{{\mathrm{D}}^{(1)}})\right) \bar{\partial}\left((1-\varrho)^{l}\right)\wedge \Omega
\end{multline*}
which by the Stokes theorem equals
\begin{multline}\label{finfor}
\sum_{i=1}^{l}(-1)^{\epsilon_1-\epsilon_{i}}\int_{\overline{U}_2\setminus U_1} {{\sf str}}\left((\Phi_{l}\circ{\mathscr{H}}_{{\mathrm{D}}^{(l)}})\circ \ldots\circ\Phi_i\circ\ldots\circ (\Phi_1\circ {\mathscr{H}}_{{\mathrm{D}}^{(1)}})\right) \varrho(1-\varrho)^{l-1}\wedge \Omega+\\
+(-1)^n\int_{\overline{U}_2\setminus U_1} \bar{\partial}\left({{\sf str}}\left((\Phi_{l}\circ{\mathscr{H}}_{{\mathrm{D}}^{(l)}})\circ \ldots\circ (\Phi_1\circ {\mathscr{H}}_{{\mathrm{D}}^{(1)}})\right) \right)(1-\varrho)^{l}\wedge \Omega+\\
+(-1)^{n-1}\sum_{x\in C_f}\int_{\partial\overline{B_{2r}(x)}-\partial\overline{B_{r}(x)}} {{\sf str}}\left((\Phi_{l}\circ{\mathscr{H}}_{{\mathrm{D}}^{(l)}})\circ \ldots\circ (\Phi_1\circ {\mathscr{H}}_{{\mathrm{D}}^{(1)}})\right) (1-\varrho)^{l}\wedge \Omega.
\end{multline}
Taking into account (\ref{cutoff}), the last sum in (\ref{finfor}) is equal to
\begin{eqnarray*}
(-1)^{n-1}\sum_{x\in C_f}\int_{\partial\overline{B_{2r}(x)}} {{\sf str}}\left((\Phi_{l}\circ{\mathscr{H}}_{{\mathrm{D}}^{(l)}})\circ \ldots\circ (\Phi_1\circ {\mathscr{H}}_{{\mathrm{D}}^{(1)}})\right)\wedge \Omega.
\end{eqnarray*}
\begin{lemma} The first and the second lines in (\ref{finfor}) vanish.
\end{lemma}
\noindent{\bf Proof.} It follows from the definition (\ref{G}) of ${\mathscr{H}}_{{\mathrm{D}}}$ and from the holomorphicity of the $\Phi$'s that the top degree parts of the integrands in the first and the second lines in (\ref{finfor}) are sums of expressions of the form
\begin{eqnarray*}
{{\sf str}}\left(\xi_{l} \circ (\bar{\partial}{\mathrm{H}}_{{\mathrm{D}}^{(l)}})^{\circ k_{l}}\circ \ldots\circ\xi_1\circ (\bar{\partial}{\mathrm{H}}_{{\mathrm{D}}^{(1)}})^{\circ k_{1}}\right) \wedge\ldots
\end{eqnarray*}
where $\xi_i$ are some morphisms and $k_1+\ldots+k_{l}=n$ ($k_i\geq0$). Therefore, it is enough to show that the wedge-product of arbitrary $n$ matrix elements of the matrices $\bar{\partial}{\mathrm{H}}_{{\mathrm{D}}^{(i)}}$ equals 0. In view of (\ref{homot4}), it suffices to prove that for any collection $\{\omega_1,\ldots,\omega_n\}$ of holomorphic 1-forms on ${X^{\mathtt{h}}}$
\[
\bar{\partial}\frac{\langle\omega_1,\partial f\rangle}{||\partial f||^2}\wedge\ldots\wedge \bar{\partial}\frac{\langle\omega_n,\partial f\rangle}{||\partial f||^2}=0.
\]
Note that the left-hand side of the latter equality is $\mathcal{A}({X^{\mathtt{h}}})$-multilinear and skew-symmetric in the $\omega$'s. Thus, we only need to prove it for $\omega_i=dz_i$, $i=1,\ldots n$. We have
\begin{multline*}
\bar{\partial}\frac{\langle dz_1,\partial f\rangle}{||\partial f||^2}\wedge\ldots\wedge \bar{\partial}\frac{\langle dz_n,\partial f\rangle}{||\partial f||^2}=\bar{\partial}\frac{\overline{\partial_1 f}}{||\partial f||^2}\wedge\ldots\wedge \bar{\partial}\frac{\overline{\partial_n f}}{||\partial f||^2}
=\frac1{\prod_i\partial_i f}\bar{\partial}\frac{|\partial_1 f|^2}{||\partial f||^2}\wedge\ldots\wedge \bar{\partial}\frac{|\partial_n f|^2}{||\partial f||^2}=\\
=\frac1{\prod_i\partial_i f}\bar{\partial}\frac{|\partial_1 f|^2}{||\partial f||^2}\wedge\ldots\wedge \bar{\partial}\frac{|\partial_{n-1} f|^2}{||\partial f||^2}\wedge \bar{\partial}\left(1-\frac{|\partial_1 f|^2+\ldots+|\partial_{n-1} f|^2}{||\partial f||^2}\right)=0.
\end{multline*}
\hfill$\blacksquare$
\bigskip
The conclusion so far is that
\begin{eqnarray}\label{fin1}
{\Theta}(\Phi_{l}[\Phi_{l-1}|\ldots|\Phi_1])=(-1)^{n-1}\sum_{x\in C_f}\int_{\partial\overline{B_{2r}(x)}} {{\sf str}}\left(\Phi_{l}\circ{\mathscr{H}}_{{\mathrm{D}}^{(l)}}\circ \ldots\circ \Phi_1\circ {\mathscr{H}}_{{\mathrm{D}}^{(1)}}\right)\wedge \Omega.
\end{eqnarray}
By (\ref{G}) and (\ref{homot4})
\begin{multline*}
(-1)^{n-1}\int_{\partial\overline{B_{2r}(x)}} {{\sf str}}\left(\Phi_{l}\circ{\mathscr{H}}_{{\mathrm{D}}^{(l)}}\circ \ldots\circ \Phi_1\circ {\mathscr{H}}_{{\mathrm{D}}^{(1)}}\right)\wedge \Omega=\\
=(-1)^{n-1}\sum\limits_{\substack{k_1+\ldots+k_{l}=n-1\\k_1,\ldots,k_l\geq0}}\int_{\partial\overline{B_{2r}(x)}}\\ {{\sf str}}\left(\Phi_{l}\circ{\mathrm{H}}_{{\mathrm{D}}^{(l)}}\circ(\bar{\partial}{\mathrm{H}}_{{\mathrm{D}}^{(l)}})^{\circ k_{l}}\circ \ldots\circ \Phi_1\circ {\mathrm{H}}_{{\mathrm{D}}^{(1)}}\circ(\bar{\partial}{\mathrm{H}}_{{\mathrm{D}}^{(1)}})^{\circ k_1}\right)\wedge \Omega=\\
=(-1)^{n-1}\sum\limits_{\substack{k_1+\ldots+k_{l}=n-1\\k_1,\ldots,k_l\geq0}}\int_{\partial\overline{B_{2r}(x)}} \\
{{\sf str}}\left(\frac{\langle\Phi_l\partial {\mathrm{D}}^{(l)},\partial f\rangle}{||\partial f||^2}\circ\left(\bar{\partial}\frac{\langle\partial {\mathrm{D}}^{(l)},\partial f\rangle}{||\partial f||^2}\right)^{\circ k_{l}}\circ \ldots\circ \frac{\langle \Phi_1\partial {\mathrm{D}}^{(1)},\partial f\rangle}{||\partial f||^2}\circ\left(\bar{\partial}\frac{\langle\partial {\mathrm{D}}^{(1)},\partial f\rangle}{||\partial f||^2}\right)^{\circ k_{1}}\right)\wedge \Omega.
\end{multline*}
Setting $\zeta_j:=\frac{\overline{\partial_j f}}{||\partial f||^2}$ and unfolding the definition of $\circ$, the latter equals
\begin{multline*}
-(-1)^{\frac{n(n+1)}2}\sum\limits_{\substack{k_1+\ldots+k_{l}=n-1\\k_1,\ldots,k_l\geq0}}(-1)^{k_1\epsilon_1+\ldots+k_l\epsilon_{l}}\sum\limits_{\left(j^{(1)}_1,\ldots, j^{(1)}_{k_1},\ldots, j^{(l)}_1,\ldots, j^{(l)}_{k_{l}}\right)}\sum\limits_{\left(i^{(1)}, \ldots, i^{(l)}\right)}\\
\int_{\partial\overline{B_{2r}(x)}} \zeta_{i^{(l)}}\ldots\zeta_{i^{(1)}}\bar{\partial}\zeta_{j^{(l)}_1}\wedge\ldots\wedge\bar{\partial}\zeta_{j^{(l)}_{k_{l}}}\wedge\ldots
\wedge \bar{\partial}\zeta_{j^{(1)}_1}\wedge\ldots\wedge\bar{\partial}\zeta_{j^{(1)}_{k_{1}}}\wedge\\
\wedge{{\sf str}}\left(\Phi_{l}\partial_{i^{(l)}} {\mathrm{D}}^{(l)}\,\partial_{j^{(l)}_1} {\mathrm{D}}^{(l)}\ldots \partial_{j^{(l)}_{k_{l}}} {\mathrm{D}}^{(l)}\cdot \ldots\cdot \Phi_{1}\partial_{i^{(1)}} {\mathrm{D}}^{(1)}\,\partial_{j^{(1)}_1} {\mathrm{D}}^{(1)}\ldots \partial_{j^{(1)}_{k_{1}}}{\mathrm{D}}^{(1)}\right) \wedge
\Omega
\end{multline*}
where $i^{(s)}$ and $j^{(s)}_r$ run from 1 to $n$. Since $\bar{\partial}\zeta_{j^{(l)}_1}\wedge\ldots\wedge\bar{\partial}\zeta_{j^{(l)}_{k_{l}}}\wedge\ldots
\wedge \bar{\partial}\zeta_{j^{(1)}_1}\wedge\ldots\wedge\bar{\partial}\zeta_{j^{(1)}_{k_{1}}}$ is non-trivial only for
$\left(j^{(l)}_1,\ldots, j^{(l)}_{k_1},\ldots, j^{(1)}_1,\ldots, j^{(1)}_{k_{l}}\right)\in {S}_n^i$ (see Theorem \ref{mainresult} for the definition of ${S}_n^i$), the above expression is equal to
\begin{multline*}
-(-1)^{\frac{n(n+1)}2}\sum\limits_{\substack{k_1+\ldots+k_{l}=n-1\\k_1,\ldots,k_l\geq0}}(-1)^{k_1\epsilon_1+\ldots+k_l\epsilon_{l}}\sum_{i=1}^n \\
\sum\limits_{\left(j^{(l)}_1,\ldots, j^{(l)}_{k_1},\ldots, j^{(1)}_1,\ldots, j^{(1)}_{k_{l}}\right)\in {S}_n^i} {\rm sgn}\left(j^{(l)}_1,\ldots, j^{(l)}_{k_1},\ldots, j^{(1)}_1,\ldots, j^{(1)}_{k_{l}}\right)
\sum\limits_{\left(i^{(1)}, \ldots, i^{(l)}\right)}\\
\int_{\partial\overline{B_{2r}(x)}} \zeta_{i^{(l)}}\ldots\zeta_{i^{(1)}}\bar{\partial}\zeta_{1}\wedge\ldots\wedge\widehat{\overline{\partial}\zeta_{i}}\wedge\ldots
\wedge \bar{\partial}\zeta_{n}\wedge\\
\wedge{{\sf str}}\left(\Phi_{l}\partial_{i^{(l)}} {\mathrm{D}}^{(l)}\,\partial_{j^{(l)}_1} {\mathrm{D}}^{(l)}\ldots \partial_{j^{(l)}_{k_{l}}} {\mathrm{D}}^{(l)}\cdot \ldots\cdot \Phi_{1}\partial_{i^{(1)}} {\mathrm{D}}^{(1)}\,\partial_{j^{(1)}_1} {\mathrm{D}}^{(1)}\ldots \partial_{j^{(1)}_{k_{1}}}{\mathrm{D}}^{(1)}\right) \wedge
\Omega
\end{multline*}
where $\widehat{\,\,\,}$ means the term is omitted. As it is shown in \cite[Ch.5, Sect.1]{GH}, for any $i=1,\ldots,n$
\begin{eqnarray*}
\bar{\partial}\zeta_{1}\wedge\ldots\wedge\widehat{\overline{\partial}\zeta_{i}}\wedge\ldots
\wedge \bar{\partial}\zeta_{n}=(-1)^{i-1} \cdot \partial_if \cdot\boldsymbol{\eta}
\end{eqnarray*}
where $\boldsymbol{\eta}$ stands for the $(0,n-1)$-form
\begin{eqnarray*}
\frac1{||\partial f||^{2n}}\cdot \sum_{s=1}^n(-1)^{s-1} \overline{\partial_s f} \,\overline{\partial (\partial_1 f)}\wedge\ldots\wedge\widehat{\overline{\partial (\partial_s f)}}\wedge\ldots\wedge\overline{\partial(\partial_n f)}.
\end{eqnarray*}
Thus,
\begin{multline}\label{pref}
{\Theta}(\Phi_{l}[\Phi_{l-1}|\ldots|\Phi_1])=-(-1)^{\frac{n(n+1)}2}\sum\limits_{\substack{k_1+\ldots+k_{l}=n-1\\k_1,\ldots,k_l\geq0}}(-1)^{k_1\epsilon_1+\ldots+k_l\epsilon_{l}}\sum_{i=1}^n(-1)^{i-1}\\
\sum\limits_{\left(j^{(l)}_1,\ldots, j^{(l)}_{k_1},\ldots, j^{(1)}_1,\ldots, j^{(1)}_{k_{l}}\right)\in {S}_n^i} {\rm sgn}\left(j^{(l)}_1,\ldots, j^{(l)}_{k_1},\ldots, j^{(1)}_1,\ldots, j^{(1)}_{k_{l}}\right)
\sum\limits_{\left(i^{(1)}, \ldots, i^{(l)}\right)}\\
\int_{\partial\overline{B_{2r}(x)}} \frac{\overline{\partial_{i^{(l)}} f}}{||\partial f||^2}\ldots\frac{\overline{\partial_{i^{(1)}} f}}{||\partial f||^2} \cdot \boldsymbol{\eta}\\
\wedge\partial_if\cdot{{\sf str}}\left(\Phi_{l}\partial_{i^{(l)}} {\mathrm{D}}^{(l)}\,\partial_{j^{(l)}_1} {\mathrm{D}}^{(l)}\ldots \partial_{j^{(l)}_{k_{l}}} {\mathrm{D}}^{(l)}\cdot \ldots\cdot \Phi_{1}\partial_{i^{(1)}} {\mathrm{D}}^{(1)}\,\partial_{j^{(1)}_1} {\mathrm{D}}^{(1)}\ldots \partial_{j^{(1)}_{k_{1}}}{\mathrm{D}}^{(1)}\right) \wedge
\Omega.
\end{multline}
Finally, using the decomposition
\[
\{1,\ldots,n\}^l=\coprod_{\substack {r_1+\ldots+r_n=l\\ r_1,\ldots,r_n\geq0}}\Lambda_n^l(r_1,\ldots, r_n).
\]
(see Theorem \ref{mainresult} for the definition of $\Lambda_n^l(r_1,\ldots, r_n)$) and the formula (\ref{GGH}) with $g_i:=\partial_if$, (\ref{pref}) can be written as
\begin{multline*}
-(-1)^{\frac{n(n+1)}2}(2\pi i)^n\sum\limits_{\substack{k_1+\ldots+k_{l}=n-1\\k_1,\ldots,k_l\geq0}}(-1)^{k_1\epsilon_1+\ldots+k_l\epsilon_{l}}\sum_{i=1}^n(-1)^{i-1}\\
\sum\limits_{\left(j^{(l)}_1,\ldots, j^{(l)}_{k_1},\ldots, j^{(1)}_1,\ldots, j^{(1)}_{k_{l}}\right)\in {S}_n^i}{\rm sgn}\left(j^{(l)}_1,\ldots, j^{(l)}_{k_1},\ldots, j^{(1)}_1,\ldots, j^{(1)}_{k_{l}}\right)\\\sum_{\substack {r_1+\ldots+r_n=l\\ r_1,\ldots,r_n\geq0}}\frac{r_1!\,\ldots \,r_n!}{(n+l-1)!}\sum\limits_{\left(i^{(1)},\ldots,\, i^{(l)}\right)\in \Lambda_n^l(r_1,\ldots, r_n)}\\
\mathrm{Res}_x\left[\frac{{{\sf str}}\left(\Phi_{l}\partial_{i^{(l)}} {\mathrm{D}}^{(l)}\,\partial_{j^{(l)}_1} {\mathrm{D}}^{(l)}\ldots \partial_{j^{(l)}_{k_{l}}} {\mathrm{D}}^{(l)}\cdot \ldots\cdot \Phi_{1}\partial_{i^{(1)}} {\mathrm{D}}^{(1)}\,\partial_{j^{(1)}_1} {\mathrm{D}}^{(1)}\ldots \partial_{j^{(1)}_{k_{1}}}{\mathrm{D}}^{(1)}\right) \wedge
\Omega}{(\partial_1f)^{r_1+1}\ldots(\partial_if)^{r_i}\ldots (\partial_nf)^{r_n+1}}\right]
\end{multline*}
which equals, up to a constant, the right-hand side of (\ref{upsx}).
|
\chapter{Three qubit gates}
\label{chap:3qubit}
In this chapter, we describe different ways to create 3-qubit Toffoli and Fredkin gates, which are universal for classical \emph{reversible} computation. The origins of reversible computation stem from Landauer's erasure principle; every bit of information erased results in $kT \log 2$ heat being irreversibly lost to the environment~\cite{Landauer1961}. It follows that in order to perform computation without dissipating heat, no information must be erased, and so only a reversible computation will not dissipate heat. In 1973, Bennett~\cite{Bennett1973}, building on work by Lecerf~\cite{Lecerf1963}, showed that reversible computation can be performed just as efficiently as irreversible computation. The field was then further enhanced by Fredkin and Toffoli~\cite{Fredkin1982} (around the same time that quantum computation was being first thought of), in which they introduced the Fredkin and Toffoli gates.
The Toffoli and Fredkin gates are also useful primitives for quantum computation. The Toffoli gate is universal for quantum computation with the addition of any basis-changing single qubit gate such as a Hadamard~\cite{Shi}, and is used in some quantum error correction schemes (e.g.~\cite{Sarovar,Cory}). The Fredkin gate can be useful in fusing W states~\cite{Ozaydin2014}, implementing Deustch's algorithm~\cite{Chuan1995}, and is also used in error-correcting schemes~\cite{Barenco1997}. Although Toffoli and Fredkin gates can be synthesised by two-qubit gates and single qubit operations, being able to perform these gates in a more direct way may reduce the size or complexity of certain circuits (e.g.\ a Toffoli gate requires a minimum of five 2-qubit gates~\cite{Yu2013}).
Given the ever-increasing density of components on a chip, and the apparent absence of a Moore's law for cooling techniques, it seems likely that using a reversible architecture for classical computation will become important over the next 20 years, so that chip heating rates and power usage do not become unmanageable. Not only \emph{logical} reversibility, but \emph{physical} reversibility will be important to reduce the heat dissipated per logic gate. Quantum computation is of course inherently reversible, since transformations occur through reversible unitary transformations. However, in the absence of any operational quantum computers in the near future, it may be useful instead to use quantum gates to perform classical logic; this would replace the traditional irreversible dissipative components such as CMOS technology (which can be made physically reversible but at the expense of increasing the gate operation time~\cite{Merkle1993}) by inherently physically reversible quantum components. In addition, since the requirements for classical computation are less stringent (there are only bit flip errors, and gates only have to be coherent for the length of the gate, not the length of the computation), creating classical reversible gates with quantum components could be a good halfway point en route to full quantum computation, perhaps allowing faster gates or more miniaturised components.
With this as motivation, we consider methods to create Toffoli and Fredkin gates. We propose three different ways of constructing 3-qubit gates, using time-independent Hamiltonians (in the sense that the Hamiltonian does not change during the operation of the gate). Proposals already exist to implement Toffoli or Fredkin gates, such as qubits interacting through cavities~\cite{Chen2012,Chen2006,Xiao2007,Shao2013}, using photon interference~\cite{Deng2007}, or using systems with extra `hidden' states or higher dimensional Hilbert spaces~\cite{Fedorov2012,Lanyon2008,Zheng2013,Ivanov2011}, or using extra `mediator' qubits~\cite{Fei2012}. Toffoli gates and Fredkin gates have also been successfully experimentally implemented using `hidden' states in superconductors~\cite{Fedorov2012}, and using interactions controlled by laser pulses in ion traps~\cite{Monz2009}. In our approach, we aim to implement the gate in a single step, using interactions that are fixed over the length of the gate, in the hope that minimal control of the system will enable better isolation of the system, and thus limit decoherence. We also focus on performing these gates on systems of qubits, rather than higher-dimensional systems, and without the aid of an external systems such as cavities or ancillas. In addition, we aim to find gates that can be easily implemented experimentally, and we will give examples of how these gates could be implemented in ion traps or in doped silicon
This chapter is structured as follows; in Sec.~\ref{sec:Toff} we describe a method to implement a (locally equivalent) Toffoli gate using Ising interactions and magnetic fields. In Sec.~\ref{sec:FredHeis} we describe a method to perform a locally equivalent Fredkin gate using Ising and Heisenberg exchange interactions and magnetic fields. In Sec.~\ref{sec:XX} we describe a method to perform a locally equivalent Fredkin gate using Ising and XX interactions with magnetic fields.
\section{Toffoli gate using uniform Ising interactions}
\label{sec:Toff}
In this section we will describe how to implement a Toffoli gate using Ising interactions, with a transverse field on the target qubit (see Fig.~\ref{fig:ZZ}). A Toffoli gate performs a controlled-controlled-\textsc{not} operation $\ket{ij}\ket{k} \to \ket{ij}\ket{k \oplus ij}$. To construct this gate using a single fixed Hamiltonian, with qubits 1 and 3 as controls and qubit 2 as target, we consider a Hamiltonian of the form
\begin{align}\label{eqn:HamToff1}
H_{\textsc{tof}} = \frac{1}{2}J_{zz} ( Z_1Z_2 + Z_2 Z_3) + \frac{1}{2}\sum_n \omega_n Z_n + \frac{1}{2}\Omega X_2.
\end{align}
The Hamiltonian is assumed to be in units of frequency, and for the rest of this chapter we will work in these units. This is similar to the type of Hamiltonian used in~\cite{Kumar2013} to implement an $N$-qubit controlled unitary, however this protocol requires non-uniform interactions and interactions between all qubits, which we would like to avoid. This Hamiltonian has the convenient property that it commutes with both $Z_1$ and $Z_3$, so that the eigenstates are formed from pairs of states with the same $z$-value on qubits 1 and 3 but different value on qubit 2 (e.g. $\{ \ket{010} , \ket{000} \}$.
To produce a Toffoli gate, the gate needs to flip the target qubit (qubit 2 in this case) when each of the control qubits (1 and 3) are in the $\ket{1}$ state. For this to be possible it is sufficient that $H_{\textsc{tof}}$ has eigenstates of the form $\frac{1}{\sqrt{2}}\ket{1} (\ket{0} \pm e^{i\phi}\ket{1})\ket{1}$. By multiplying $H_{\textsc{tof}}$ by states of this form, it can easily be confirmed that the states $\frac{1}{\sqrt{2}}\ket{1} (\ket{0} \pm \ket{1})\ket{1}$ are eigenstates when $\omega_2 = 2J_{zz}$ (intuitively, the effect of the Ising coupling and local $Z_2$ field cancel out, so that the only remaining operator acting on qubit 2 is $X_2$).
\begin{figure}[b]
\begin{center}
\includegraphics[scale = 0.5]{LinearZZ-eps-converted-to.pdf}
\caption{Setup for creating a 3-qubit Toffoli gate.}
\label{fig:ZZ}
\end{center}
\end{figure}
The eigenstates and energies of the full Hamiltonian were then found using Mathematica with $\omega_2 = 2J_{zz}$, giving
\begin{align}
\ket{\pm}_{ab} &= \frac{1}{\mathcal{N}^{\pm}_{ab}}\text{ } \ket{a} \left[(d_{ab} \pm \sqrt{1 + d_{ab}^2})\ket{0} + \ket{1} \right] \ket{b},\\
E_{ab}^{\pm} /\hbar &= \frac{1}{2} \left[ (-1)^{a}\omega_1 + (-1)^{b} \omega_3 \pm \Omega \sqrt{d_{ab}^2 + 1} \right],
\end{align}
where $a,b\in \{0,1\}$ and we have introduced the parameters $d_{11} = 0$, $d_{01} = d_{10} = \omega_2/\Omega$, $d_{00} = 2\omega_2 /\Omega$, and $\mathcal{N}^{\pm}_{ab}$ are normalising factors.
The $\ket{101} \leftrightarrow\ket{111}$ swapping will occur when $\exp\left[ -iH_{\textsc{tof}}t\right] \ket{101} = e^{i\theta} \ket{111}$, which occurs at a time $t = \tau_n = (2n+1)\pi \hbar /|E_{110}^{+} - E_{110}^{-}| = (2n+1)\pi / \Omega $ (where $n$ is an integer, and assuming without loss of generality that $\Omega$ is positive).
The remaining computational states $\{ \ket{000},\ket{010},\ket{001},\ket{011},\ket{100},\ket{110} \}$ will also evolve during this time, so we now look at how to engineer the Hamiltonian such that this swapping happens an even number of times during the gate time $\tau_n$. The fidelities of these swapping events occurring is measured by the fidelity $f_{lmn \rightarrow xyz} := \bra{xyz} e^{-i H_{\textsc{tof}}\tau_n } \ket{lmn}$, which we obtain by first expressing each computational state in terms of the eigenstates of the system, then multiplying each eigenstate $\ket{\pm}_{ab}$ by $e^{-iE_{ab}^{\pm} \tau_n / \hbar}$ and simplifying using Mathematica and the Quantum Mathematica add-on~\cite{QuantumMathematica}. Following this process, and recalling that the Hamiltonian commutes with $Z_1$ and $Z_3$ so that transitions only occur between states with the same values of $m_z$ on qubits 1 and 3, we obtain the following fidelities:
\begin{align}\label{eqn:Fid}
&f_{abc \rightarrow a \bar{b} c} = \frac{-ie^{-i\phi_{ac}} (2n+1) \pi}{2}\mbox{sinc} \left( \frac{(2n+1)\pi}{2} \sqrt{d_{ac}^2 + 1} \right) \nonumber\\
&f_{abc \rightarrow abc} = -ie^{-i\phi_{ac}} \left[ \cos \left(\frac{(2n+1)\pi}{2} \sqrt{d_{ac}^2 + 1} \right) - \frac{(2n+1) i \pi}{2} d_{ac} \text{ }\mbox{sinc} \left( \frac{(2n+1)\pi}{2} \sqrt{d_{ac}^2 + 1} \right) \right],
\end{align}
where $a,b,c \in \{0,1\}$, $(a,c) \neq (1,1)$, $\bar{b} := b \oplus 1$, and
\begin{align}
\phi_{ac} &= \frac{(2n+1)\pi (E^+_{ac} + E^-_{ac})}{4 \hbar \Omega} = \frac{(2n+1)\pi ((-1)^a \omega_1 + (-1)^c \omega_3)}{2\Omega}.
\end{align}
Note that $|f_{101\leftrightarrow 111}| = 1$ by our choice of gate time.
The condition these fidelities need to satisfy to realise a Toffoli gate is $|f_{a0b}\rightarrow f_{a1b}| = 0$ for $(a,b)\ne (1,1)$. To achieve this the phase inside the sinc function in (\ref{eqn:Fid}) must be an integer multiple of $\pi$ for each fidelity. Noting that the phases for $f_{100 \rightarrow 110}$ and $f_{001 \rightarrow 011}$ are the same, this condition leads us to the following two coupled equations;
\begin{align}
\frac{1}{2} \sqrt{\frac{\omega_2^2}{\Omega^2} + 1} = \frac{m_1}{(2n+1)}, \label{eqn:Diop1}\\
\frac{1}{2} \sqrt{\frac{4\omega_2^2}{\Omega^2} + 1} = \frac{m_2}{(2n+1)}, \label{eqn:Diop2}
\end{align}
where $m_1,m_2$ are non-zero integers. Eliminating $\frac{\omega_2}{\Omega}$ from these equations gives an equation that $m_1$, $m_2$ and $n$ must satisfy for a Toffoli gate to be possible:
\begin{align}
16 m_1^2 - 4m_2^2 = 3(2n+1).
\end{align}
The left hand side of this equation is even, and the right side is odd, so there is no assignment of integer $m_1$, $m_2$ and $n$ which satisfies this.
Whilst this is proof that there is no choice of $\omega_2,$ $\Omega$ that achieves a perfect gate, we can still find parameters which achieve an approximate gate. Assume that $n=0$ for simplicity (i.e.\ only one swap of $\ket{101} \leftrightarrow \ket{111}$ occurs during the gate). Then rearranging eqn.\ (\ref{eqn:Diop1}) for $\frac{1}{2} \sqrt{\frac{\omega_2^2}{\Omega^2} + 1} =m_1$ with $m_1 \geq 1$ and inserting into eqn.\ (\ref{eqn:Diop2}) gives
\begin{align} \label{eq:Cond1}
&\frac{1}{2} \sqrt{\frac{4\omega_2^2}{\Omega^2} + 1} = 2m_1 - \frac{3}{16m_1} + O\left( \frac{1}{m_1^3} \right).
\end{align}
Therefore we find that $m_2 \approx 2m_1$ with this choice of parameters, so that the phases inside the sinc functions in (\ref{eqn:Fid}) are all either zero or approximately zero up to order $1/m_1$. Substituting these expressions into (\ref{eqn:Fid}), we obtain the fidelities (making $m_1$ even for simplicity, but the calculation for odd $m_1$ is similar and doesn't change the form of the error terms)
\begin{align} \label{eqn:Fid2}
&f_{000 \rightarrow000} =f_{010 \rightarrow010} = -ie^{-i\phi_{00}} \left[ \cos \left( \pi \sqrt{16 m_1^2 - 3}\right) - i \pi \sqrt{4 m_1^2 - 1} \text{ sinc} \left( \pi \sqrt{16 m_1^2 - 3} \right) \right]\nonumber\\
&f_{000 \rightarrow 010} = f_{010 \rightarrow 000} =\frac{-ie^{-i\phi_{00}} \pi}{2} \text{sinc} \left( \pi \sqrt{16 m_1^2 - 3} \right),\nonumber\\
& f_{001 \rightarrow 001} =f_{011 \rightarrow 011} = -i e^{- i \phi_{01} },\nonumber\\
& f_{001 \rightarrow 011} = f_{011 \rightarrow 001} = 0, \nonumber\\
&f_{100 \rightarrow 100} = f_{110 \rightarrow 110}=-i e^{- i \phi_{10} },\nonumber\\
& f_{100 \rightarrow 110} = f_{110 \rightarrow 100} = 0,\nonumber\\
&f_{101 \rightarrow101} = f_{111 \rightarrow111} = 0,\nonumber\\
&f_{101 \rightarrow 111} = f_{111 \rightarrow 101} = -ie^{-i\phi_{11}}.
\end{align}
To evaluate how close this approximation brings us to an exact Toffoli gate, we use the process trace distance, which for a 3-qubit system is defined as (see Sec.~\ref{sec:DistMeas} and~\cite{Gilchrist2005}):
\begin{align}
\mathcal{D}_{pro} = \frac{1}{16} \| \chi(U) - \chi(T) \|_{tr},
\end{align}
where $T$ is an ideal Toffoli gate, $U$ is the gate we can achieve with the above setup, and $\| X \|_{tr} = \mbox{tr} ( \sqrt{X^\dagger X} )$ is the trace norm. For simplicity, we ignore any phases and define $U_{|f|}$ such that $ \bra{y}U_{|f|}\ket{x} = |f_{x \rightarrow y }|$ so this will only measure how closes we are to a Toffoli gate apart from some local operations, or how close we are to a classical gate. By inspecting the phases in (\ref{eqn:Fid2}), we can see that the local rotations needed to get rid of the phases on the leading order terms are $i\exp [ i \pi \omega_1 Z_1 / \Omega] \exp [ i \pi \omega_3 Z_2 / \Omega ]$.
Using the terms in (\ref{eqn:Fid2}), and defining $f_0 := f_{000 \rightarrow010}$, $U_{|f|}$ is then
\newlength{\mycolwd}
\settowidth{\mycolwd}{10000em}
\begin{align*}
U_{|f|} = \left(\begin{array}{C{\mycolwd} C{\mycolwd} C{\mycolwd} C{\mycolwd} C{\mycolwd} C{\mycolwd} C{\mycolwd} C{\mycolwd} }
\sqrt{1 - |f_{0}|^2 } & |f_{0}| & 0 & 0 & 0 & 0& 0 & 0\\
|f_{0}| & \sqrt{1 - |f_{0}|^2 } & 0 & 0 & 0 & 0& 0 & 0\\
0 & 0 & 1& 0 & 0 & 0& 0 & 0\\
0 & 0 & 0 & 1 & 0 & 0& 0 & 0\\
0 & 0 & 0 & 0 & 1&0& 0 & 0\\
0 & 0 & 0 & 0 & 0 & 1 & 0 & 0\\
0 & 0 & 0 & 0 & 0 & 0& 0 & 1\\
0 & 0 & 0 & 0 & 0 & 0& 1 & 0
\end{array} \right).
\end{align*}
Calculating $\chi(U_{|f|})$ can be done straightforwardly by flattening $U_{|f|}$ into a $64 \times 1$ vector $V$ containing the elements of $U_{|f|}$, and defining $\chi_{mn} = V_m^* V_n$. This was evaluated using Mathematica, giving $\mathcal{D}_{pro} =\frac{1}{4} \sqrt{6 + |f_0|^2 - 6 \sqrt{1 - |f_0|^2}}$. Expanding $|f_0|$ in a Taylor series for $m_1>1$, and using $\text{sinc}^2(2 m_1 \pi + \delta ) = \frac{\delta^2}{2 m_1^2 \pi^2} + O \left( \frac{\delta^3}{m_1^3} \right)$:
\begin{align}
&|f_0|^2 = |f_{000 \rightarrow010}|^2 = \pi^2 \mbox{sinc}^2 \left( \frac{\pi}{2} \sqrt{16m_1^2 - 3 } \right)= \pi^2 \mbox{sinc}^2 \left( 2 m_1 \pi\sqrt{1 - \frac{3}{16m_1^2} } \right) \nonumber\\
&= \pi^2 \mbox{sinc}^2 \left( 2m_1 \pi - \frac{3 \pi}{16m_1} + O \left( \frac{1}{m_1^3} \right) \right) = \frac{9\pi^2 }{1024 m_1^4} + O \left( \frac{1}{m_1^6} \right).
\end{align}
Substituting this into the expression for $\mathcal{D}_{pro}$, we find that $\mathcal{D}_{pro} = \frac{3 \pi}{16 m_1^2} + O\left( \frac{1}{m_1^4} \right)$. $\mathcal{D}_{pro}$ gives an upper bound on the average probability $\bar{p}_e$ that the gate fails~\cite{Gilchrist2005}, so
\begin{align}
\label{eqn:p}
\bar{p}_e \lesssim \frac{3 \pi}{16m_1^2} \approx \frac{3 \pi}{4} \left(\frac{\Omega}{\omega_2} \right)^2 = \frac{3 \pi}{16} \left(\frac{\Omega}{J_{zz}} \right)^2 .
\end{align}
In summary, a gate locally equivalent to a Toffoli gate can be achieved with average failure error of $\bar{p}_e \lesssim \frac{3 \pi}{16} \left(\frac{\Omega}{J_{zz}} \right)^2$, provided that $J_{zz} = \omega_2/2$ and $\frac{\omega_2}{\Omega} =\sqrt{4 m_1^2 - 1}$, with $m_1 \geq 1$.
\subsection{Sequences of gates}
\begin{figure}[h]
\begin{center}
\includegraphics[scale = 0.45]{HalfAdder-eps-converted-to.pdf}
\caption{a) A half-adder circuit. b) Setup for creating a half adder, using two pulses and with $J_{13} = J_{23} \neq J_{12}$.}
\label{fig:HAcirc}
\end{center}
\end{figure}
We can take advantage of the fact that the Toffoli gate outlined above is activated by an external field to compose it together with other gates. One of the simplest circuits that could be performed is a Toffoli gate followed by a \textsc{cnot} gate, which performs a half-adder circuit (Fig.~\ref{fig:HAcirc}). To perform this, starting with an arrangement of qubits as shown in Fig.~\ref{fig:HAcirc}, we can first perform the Toffoli gate in Section~\ref{sec:Toff} with qubits 1 and 2 as controls. Then to apply the controlled-\textsc{not} gate, we can apply fields such that the field is resonant with qubit 2 only if qubit 1 is in a $\ket{1}$ state, but is independent of the state of qubit 3. This can be done by applying a Hamiltonian of the form
\begin{align}
H = \frac{1}{2}\sum_{n=1}^3 \omega_n Z_n+ \frac{1}{2}\sum_{mn} J_{mn} Z_m Z_n + \frac{1}{2}\Omega_1 X_2 \cos ( \omega_+ t ) + \frac{1}{2}\Omega_2 X_2 \cos (\omega_+ t),
\end{align}
where $\omega_{\pm} = \frac{1}{2}( \omega_3 -J_{13} \pm J_{23}) $, and $\omega_n$ is the resonant frequency of qubit $n$. The first of these fields flips qubit 2 when qubits 1 and 3 are in a $\ket{10}_{12}$ state, and the second field flips when 1 and 3 are in the $\ket{11}_{12}$ state, overall resulting in a \textsc{cnot} gate.
\section{Fredkin gate with Ising and Heisenberg interactions}
\label{sec:FredHeis}
\begin{figure}[b]
\begin{center}
\includegraphics[scale = 0.5]{FredkinHeis-eps-converted-to.pdf}
\caption{Setup for creating a 3-qubit Fredkin gate, using Ising and Heisenberg coupling.}
\label{fig:FredHeis}
\end{center}
\end{figure}
We now consider creating a quantum Fredkin gate (controlled-\textsc{swap}), using Ising and anisotropic Heisenberg interactions. A Fredkin gate performs the operation $\ket{k}\ket{lm} \to \ket{k}\textsc{swap}^k\ket{lm}$. To realise this gate we consider a Hamiltonian of the form
\begin{align}\label{eqn:HamFred}
H_{\textsc{fred}} = \frac{1}{2} J_{12} Z_1 Z_2 + \frac{1}{2} J_{23} (X_2 X_3 + Y_2 Y_3 + Z_2 Z_3) +\frac{1}{2}\sum_{j=1}^3 \omega_j Z_j,
\end{align}
where qubit 1 is the control qubit, and qubits 2 and 3 are to be swapped (see Fig.~\ref{fig:FredHeis}). The intuition is that the swapping induced by the Heisenberg interaction between qubits 2 and 3 will only occur when qubit 1 is in a state that makes the energy splitting of qubits 2 and 3 match.
As in the previous section, we need to choose parameters such that states of the form $\frac{1}{\sqrt{2}} (\ket{110} \pm \ket{101} )$ are eigenstates of the Hamiltonian, which requires $J_{12} = \omega_2 - \omega_3$. With this settings $\ket{100},\ket{111},\ket{011},\ket{000} $ are all eigenstates, and the only eigenstates of the Hamiltonian which are not computational basis states are
\begin{eqnarray}
&\ket{\psi}_{110}^{\pm} = &\frac{1}{\sqrt{2}} (\ket{110} \pm \ket{101} )\nonumber\\
& \ket{\psi}_{010}^{\pm} = &\frac{1}{\mathcal{N}^{\pm}_{010}} \left[ (J_{12} \pm \sqrt{J_{12}^2 + J_{23}^2}) \ket{001} + J_{23}\ket{010} \right],
\end{eqnarray}
where $\mathcal{N}^{\pm}_{010}$ are normalising factors. The eigenenergies of these states are
\begin{align}
&E_{110}^{\pm} /\hbar= - \frac{1}{2} \left[ \omega_1 - J_{23} \pm 2 J_{23} \right],\; E_{010}^{\pm} /\hbar = \frac{1}{2} \left[ \omega_1 - J_{23} \pm 2 \sqrt{J_{12}^2 + J_{23}^2} \right].
\end{align}
The swap $\ket{110} \leftrightarrow \ket{101}$ is complete at a time $\tau_n = (2n+1) \pi \hbar/|E_{110}^+ - E_{110}^-| = (2n+1)\pi /2J_{23}$ (with $n \in \{0,1,2,...\}$, and assuming $J_{23}> 0$, without loss of generality). By expanding $\ket{001},\ket{010}$ in terms of the eigenstates $\ket{\psi}_{010}^{\pm}$, multiplying these by $e^{-iE_{010}^{\pm} \tau_n /\hbar}$ and simplifying we find that the fidelity for swapping $\ket{010} \leftrightarrow \ket{001}$ at time $\tau_n$ is
\begin{align}
f_{010 \leftrightarrow 001} =\frac{i \pi e^{-i \phi }}{ J_{23}} \mbox{sinc} \left[ \frac{ (2n+1)\pi \sqrt{J_{12}^2 + J_{23}^2} }{2J_{23}} \right],
\end{align}
where $\phi = (J_{23}-\omega_1)(2n+1) /2 J_{23}$. For this fidelity to be zero, we need $J_{23}^2 \left(\frac{m^2}{(2n+1)^2}-1 \right) =J_{12}^2 = (\omega_2 - \omega_3)^2$, where $m$ is an integer and $\frac{m}{(2n+1)} > 1$.
\section{Fredkin gate using XX interactions combined with Ising interactions}
\label{sec:XX}
Now we consider performing a Fredkin gate, based on the swapping protocol in~\cite{Benjamin2003}, where a swap operation is achieved between the ends of an $XX$ spin chain. Somewhat intuitively, a Fredkin gate can be achieved simply by adding a qubit connected the middle of this chain, with couplings that effectively fix the middle of the chain and block the swap. We start with the simplest case, three qubits coupled by an XX interaction with an Ising coupling of the middle qubit to an external qubit A
\begin{align}
H_{\textsc{xx}} = \frac{1}{2}J_{xx} (X_1X_2 +Y_1Y_2 + X_2X_3 + Y_2 Y_3) +\frac{1}{2} J_{zz} Z_A Z_2 + \frac{1}{2}\omega_A Z_A + \frac{1}{2}\omega_2 Z_2.
\end{align}
\begin{figure}[h]
\begin{center}
\includegraphics[scale = 0.5]{FredkinXX-eps-converted-to.pdf}
\caption{Setup for creating a Fredkin gate, using Ising and $XX+YY$ interactions.}
\label{fig:FredXX}
\end{center}
\end{figure}
See Fig.~\ref{fig:FredXX} for an illustration of this. Here we define qubits 1 and 3 as the qubits to be swapped, whilst qubit A is the control, and the qubits are written in the order $A,1,2,3$ so e.g. $\ket{0000} = \ket{0}_A \ket{0}_1 \ket{0}_2 \ket{0}_3$. We will also only consider protocols in which qubit 2 is initialised in the $\ket{0}$ state. For simplicity, we will express all quantities relative to $J_{xx}$. Then defining $\alpha = (\omega_2 + J_{zz})/J_{xx}$, $\beta = (\omega_2 - J_{zz})/J_{xx}$, $F(x) = \sqrt{8 + x^2}$, the eigenstates of the Hamiltonian are
\begin{align}
\begin{array}{c c}
\ket{\phi}_{110} = \frac{1}{\sqrt{2}} (\ket{1100} - \ket{1001}), \; &\ket{\phi}_{010} = \frac{1}{\sqrt{2}} (\ket{0100} - \ket{0001}) \\
\ket{\pm}_{110} = \frac{2\ket{1001} - (\beta \pm F(\beta))\ket{1010} + 2\ket{1100}}{\sqrt{8 + (F(\beta) + \beta)^2}} , \; &\ket{\pm}_{010} = \frac{2\ket{0001} - (\alpha \pm F(\alpha))\ket{0010} + 2\ket{0100}}{\sqrt{8 + (F(\alpha) + \alpha)^2}} \\
\ket{\pm}_{111} = \frac{2\ket{1011} + (\beta\pm F(\beta) )\ket{1101} + 2\ket{1110}}{\sqrt{8 + (F(\beta) + \beta)^2}}, \; &\ket{\pm}_{011} = \frac{2\ket{0011} + (\alpha\pm F(\alpha) )\ket{0101} + 2\ket{0110}}{\sqrt{8 + (F(\alpha) + \alpha)^2}} \\
\ket{\phi}_{100} = \ket{1000}, \; &\ket{\phi}_{000} = \ket{0000}.
\end{array}
\end{align}
The energies of these states are, respectively:
\begin{align}
&E_{110} = -\frac{\omega_A}{2J_{xx}} + \frac{1}{2}\beta, \; E_{110}^{\pm} = -\frac{\omega_A}{2J_{xx}} \mp \frac{1}{2}F(\beta), \; E_{010} = \frac{\omega_A}{2J_{xx}} + \frac{1}{2}\alpha, \; E_{010}^{\pm} = \frac{\omega_A}{2J_{xx}} \mp \frac{1}{2}F(\alpha) \nonumber\\
&E_{111}^{\pm} = -\frac{\omega_A}{2J_{xx}} \pm \frac{1}{2}F(\beta), \; E_{011}^{\pm} = \frac{\omega_A}{2J_{xx}} \pm \frac{1}{2}F(\alpha) , \; E_{100} = -\frac{\omega_A}{2J_{xx}} + \frac{1}{2}\beta, \; E_{000} = \frac{\omega_A}{2J_{xx}} + \frac{1}{2}\alpha.
\end{align}
We are now interested in finding the how these states evolve under the Hamiltonian.
Similarly to Sec.~\ref{sec:Toff}, we define the fidelities as $f_{x \leftrightarrow y} := \bra{y} e^{-iH_{\textsc{xx}} t} \ket{x} $ where $x,y$ are 3-digit binary strings containing the states of qubits A,1, and 3 (qubit 2 is initialised in the $\ket{0}$ state so is left out). By expressing each of the possible input states $\ket{0001},\ket{0100},\ket{1001},\ket{1100}$ in terms of the eigenstates of $H_{\textsc{xx}} $, and evolving these in time, we find that the relevant fidelities are (excluding phases, and excluding transitions involving $\ket{0000}$ and $\ket{1000}$ as they are eigenstates)
\begin{align}\label{eq:Fid}
&|f_{010 \leftrightarrow 001}|^2 = G^+(\alpha,t) , \; |f_{010\rightarrow 010}|^2= |f_{001\rightarrow 001}|^2 =G^- (\alpha,t) \nonumber\\
&|f_{110\leftrightarrow 101}|^2 = G^+(\beta,t) , \; |f_{110\rightarrow 110}|^2= |f_{101\rightarrow 101}|^2= G^- (\beta,t) \nonumber\\
&|f_{011 \rightarrow 011} |^2= \frac{1}{F(\alpha)^2} [\alpha^2 + 8\cos^2 F(\alpha)t ], \; |f_{111\rightarrow 111} |^2 = \frac{1}{F(\beta)^2} [\beta^2 + 8\cos^2 F(\beta)t ] \nonumber\\
&f_{|100 \rightarrow 100}|^2 = |f_{000\rightarrow 000} |^2= 1,
\end{align}
where
\begin{align}
&G^{\pm} (\omega,t) := \frac{1}{2} \left( 1 \mp \cos \omega t \cos F(\omega) t - \frac{\sin F(\omega)t}{F^2(\omega)} \left[ 4 \sin F(\omega)t \pm \omega F(\omega) \sin \omega t \right] \right).
\end{align}
To construct a gate locally equivalent to a Fredkin gate, we require $|f_{1xy \leftrightarrow 1yx}|^2 = 1$, $|f_{0xy \leftrightarrow 0yx}|^2 = 0$, $|f_{0xy \leftrightarrow 0xy}|^2 = 1$ for all $x,y \in \{0,1\}$. These will be satisfied when $G^+(\alpha,t) = G^-(\beta,t) = 0$, $G^-(\alpha,t) = G^+(\beta,t) = 0$, $\cos^2 F(\beta)t=\cos^2 F(\alpha)t =1$. We look for solutions to these conditions in 3 regimes:
\begin{itemize}
\item[(1)] $\omega_2, J_{zz} \sim J_{xx}$ ($\alpha,\beta \sim 1$).
\item[(2)] $|\omega_2| \approx |J_{zz}| >> J_{xx}$ ($|\alpha| >> 1, \beta = 0$ or $|\beta| >> 1, \alpha = 0$).
\item[(3)] $|\omega_2| >>|J_{zz}|$ or $|J_{zz}| >> |\omega_2|$ ($|\alpha| >> 1, |\beta| >> 1$).
\end{itemize}
Note that this is not intended to cover all of the possible conditions, but covers a few regimes where results can be found.
\subsection{Regime 1: $\omega_2, J_{zz} \sim J_{xx}$}
We first consider the regime where $\omega_2, J_{zz} \sim J_{xx}$, so that $\alpha,\beta$ are small, and we look for parameters which yield exact solutions. To make $|f_{011 \rightarrow 011} |^2 = |f_{111\rightarrow 111} |^2 = 1$, we must set $F(\alpha)t = n_{\alpha} \pi$, $F(\beta)t = n_{\beta} \pi$. Substituting these into $G^{\pm}(\alpha,t)$ and $G^{\pm}(\beta,t)$ gives:
\begin{align}
&G^{\pm} (\alpha,t) = \frac{1}{2} \left[ 1 \mp (-1)^{n_\alpha}\cos \alpha t \right] , \;G^{\pm}(\beta,t) = \frac{1}{2} \left[ 1 \mp (-1)^{n_\beta}\cos \beta t \right].
\end{align}
To satisfy the remaining conditions, we require $G^+(\alpha,t) =G^-(\beta,t) = 0$, $G^-(\alpha,t) = G^+(\beta,t) = 1$, which occurs when
\begin{align}\label{eqn:Parity}
&\alpha t = m_\alpha \pi, \; \beta t = m_\beta \pi , \; P(m_\alpha) = P(n_\alpha), \; P(m_\beta) = 1 \oplus P(n_\beta),
\end{align}
where $P$ is a parity function, which equals 1 if a number is even and 0 if odd. These conditions are not satisfiable, since
\begin{align}
&F(\alpha)^2 t^2 - F(\beta)^2 t^2 = \alpha^2 t^2 - \beta^2 t^2\; \longrightarrow \; n_\alpha^2 - m_\alpha^2 = n_\beta^2 - m_\beta^2,
\end{align}
which cannot be satisfied by integers that also satisfy the conditions in (\ref{eqn:Parity}). So there are no exact solutions of $\alpha$ and $\beta$ which create a Fredkin gate. In this regime, there may still be an approximate way to make a Fredkin gate, however to date the author has been unsuccessful in finding one.
\subsection{Regime 2: $|\omega_2| \approx |J_{zz}| >> J_{xx}$}
We now investigate the regime where $|\omega_2| \approx |J_{zz}| >> J_{xx}$, so that either $|\alpha| >> 1, \beta \sim 0$ or $|\beta| >> 1, \alpha \sim 0$. To simplify the analysis, we will consider only $|\beta| >> 1, \alpha \sim 0$, since we expect both scenarios to have similar behaviour. To satisfy the condition $|f_{011 \rightarrow 011} |^2 = 1$, we pick $\alpha = 0$ (i.e. $\omega_2 = - J_{zz}$) so that $F(\alpha) = \sqrt{8}$. Under these conditions, $G^\pm(\alpha,t)$ becomes
\begin{align}
G^{\pm}(\alpha=0,t)= \frac{1}{2} [ \cos 2t\sqrt{2} \mp 1].
\end{align}
The conditions $G^+(\alpha = 0,t)=0$ and $G^-(\alpha=0,t) = 1$ are satisfied at times $t = \tau_n = (2n+1) \pi/ 2\sqrt{2}$, so we consider evolution at integer multiples of $\tau_n$. With $|\beta| \gg 1$, $F(\beta) \approx |\beta| + O(\frac{1}{\beta})$, and
\begin{align}
G^{\pm}(\beta,t) &= \frac{1}{2}(1\mp \cos(F(\beta) - \beta)t) + O\left( \frac{1}{\beta^2} \right),
\end{align}
where $\cos( F(\beta) - \beta)t) $ has not been expanded since $t$ may also be large. Substituting these expression into eqn.\ (\ref{eq:Fid}) gives
\begin{align}
&|f_{110\leftrightarrow 101} |^2 = \sin ^2\frac{(F(\beta) - \beta)\tau_n}{2} - O\left( \frac{1}{\beta^2} \right), \;|f_{110\rightarrow 110}|^2 = \cos ^2\frac{(F(\beta) - \beta)\tau_n}{2} + O\left( \frac{1}{\beta^2} \right)\nonumber\\
&|f_{111 \rightarrow 111}|^2 = 1 - O \left(\frac{1}{\beta^2} \right), \; |f_{101\rightarrow 101}|^2 = |f_{110\rightarrow 110}|^2.
\end{align}
This is satisfied for values of $\beta = \frac{2 \tau_n}{(2m+1) \pi} - \frac{(2m+1) \pi}{\tau_n}$, for some integers $m,n$ such that either $\tau_n \gg m \pi$ or $m \pi \gg \tau_n$. Thus up to an error of $O(1/\beta^2) = O( J_{xx}^2 / (2J_{zz})^2 )$ we can achieve approximate solutions.
\subsection{Regime 3: $|\omega_2| >>|J_{zz}|$ or $|J_{zz}| >> |\omega_2|$ }
We now consider the regime where $|\omega_2| >>|J_{zz}|$ or $|J_{zz}| >> |\omega_2|$, such that $|\alpha| >> 1, |\beta| >> 1$ and $F(\alpha) \approx |\alpha| + O(\frac{1}{\alpha})$, $F(\beta) \approx |\beta| + O(\frac{1}{\beta})$. Following a similar analysis to the previous subsection, the fidelities under these conditions are
\begin{align}
&|f_{010 \leftrightarrow 001}|^2 = \sin^2\frac{(F(\alpha) - \alpha)t}{2} + O\left( \frac{1}{\alpha^2} \right), \; |f_{110\leftrightarrow 101}|^2= \sin^2 \frac{(F(\beta) - \beta)t}{2} - O\left( \frac{1}{\beta^2} \right) \nonumber\\
&|f_{010\rightarrow 010}|^2 = \cos^2\frac{(F(\alpha) - \alpha)t}{2} - O\left( \frac{1}{\alpha^2} \right), \; |f_{110\rightarrow 110}|^2 = \cos ^2\frac{(F(\beta) - \beta)t}{2} + O\left( \frac{1}{\beta^2} \right)\nonumber\\
&|f_{011 \rightarrow 011}|^2 = 1 - O\left( \frac{1}{\alpha^2} \right), \; |f_{111\rightarrow 111}|^2 = 1 - O\left( \frac{1}{\beta^2} \right)\nonumber\\
&|f_{001\rightarrow 001} |^2 = |f_{010\rightarrow 010}|^2, \; |f_{100 \rightarrow 100}|^2 = |f_{000\rightarrow 000}|^2 = 1.
\end{align}
These can be satisfied by $\alpha = \frac{2t}{m\pi} - \frac{4m\pi}{t}$, $\beta = \frac{4t}{(2n+1)\pi} - \frac{(2n+1)\pi}{2t}$, where $m$ and $n$ are integers such that $|\alpha|,|\beta| \gg 8$.
\section{Experimental implementations}
Having seen three theoretical proposals for 3-qubit gates, we now focus on how experimentally attainable these would be using current technology. We focus on using linear ion traps and bismuth donors in silicon, although in there are many systems that these gates could be applied to.
\subsection{Toffoli gate using ion traps}
Here we propose a possible way to achieve the Toffoli gate in Sec.~\ref{sec:Toff}, using trapped ions. We consider three $^{171}\mbox{Yb}^+$ ions in a linear Paul trap with secular frequency $\nu = 2\pi \times 100 \mbox{kHz}$. The qubits are encoded in the ground state hyperfine levels which are separated by approximately $12.6 \text{GHz}$~\cite{Monroe2013}. To create Ising interactions between the qubits, we consider using the scheme presented in~\cite{Mintert2001} and discussed in Sec.~\ref{sec:IntroIonTraps}. For three ions in a linear trap with secular frequency $\nu$, the vibrational eigenstates of $H_D$ are $\frac{1}{\sqrt{3}}(1,1,1)$, $\frac{1}{\sqrt{3}}(-1,0,1)$, $\frac{1}{\sqrt{6}}(1,2,1)$ with eigenvalues $\nu_1 = \nu$, $\nu_2 =\nu\sqrt{3}$ and $\nu_3 =\nu\sqrt{29/5}$ respectively~\cite{James1998}. Inserting these motional eigenstates into equation (\ref{eqn:IonJzz}), and with an axial magnetic field gradient of $250 \text{Tm}^{-1}$ one can realise exchange couplings $J_{12} =J_{23} = 2\pi \times 9.98 \mbox{kHz} $, $J_{13} =2\pi \times 7.07 \mbox{kHz}$. The extra $J_{13}$ coupling results in introducing extra phases to the gate, which makes it more complicated to make this into an exact quantum Toffoli gate but has no effect if the gate is used for classical computation.
The $X_2$ field can be achieved by applying a resonant microwave pulse, giving rise to a Hamiltonian of the form
\begin{align}\label{eqn:IonTrap1}
H &= \frac{1}{2}\sum_{j=1}^3 \omega_j Z_j + \frac{1}{2}J_{zz}( Z_1 Z_2 + Z_2 Z_3) + \frac{1}{2}\Omega \cos( \omega_x t )X_2.
\end{align}
We can transform into the interacting frame with respect to $H_{0} = \frac{1}{2}\sum_{j=1}^3 \omega_j Z_j - J_{zz} Z_2$. States in this rotating picture evolve according to the Hamiltonian
\begin{align}
&H_I = e^{-i H_{0}t} (H-H_{0}) e^{i H_{0}t} \nonumber\\
&= J_{zz} Z_2 + \frac{1}{2}J_{zz}( Z_1 Z_2 + Z_2 Z_3) + \frac{1}{2}\Omega e^{-i(\omega_2 - 2J_{zz})Z_2 t } \cos(\omega_x t )X_2 .
\end{align}
Setting $\omega_x = \omega_2 - 2J_{zz}$ and applying identity~\ref{eqn:Iden1} from Appendix~\ref{app:IntPic} to this, we find that the evolution due to eqn.\ (\ref{eqn:IonTrap1}) is approximately the same as the evolution due to this Hamiltonian:
\begin{align}
\label{eqn:HIntIon}
H_{I} = J_{zz} Z_2 +\frac{1}{2} J_{zz} ( Z_1 Z_2 + Z_2 Z_3) + \frac{1}{4} \Omega X_2 +O\left( \frac{\Omega}{2(\omega_2 -2J_{zz})} \right).
\end{align}
With $\Omega = 2 \pi \times 3.5 \mbox{kHz} $, the systematic gate error is around $0.02$ and the gate time is $\sim0.2\text{ms}$, giving an error due to decoherence of around $(1- e^{-t_{gate}/T_2}) = 0.02$ ($T_2 \approx 10 \mbox{ms}$~\cite{Piltz2013} using dynamical decoupling). The error in the approximation in (\ref{eqn:HIntIon}) is negligible since $\omega_2 \sim 12.6 \text{GHz}$. Overall we therefore expect gate with error $\sim 0.04$ to be achievable with current technology, which does not fall below the thresholds required for fault-tolerant quantum computation.
\subsection{Toffoli gate using donors in Silicon}
To create the Toffoli gate in Sec.~\ref{sec:Toff} we can consider three donors arranged in a line. To achieve Ising couplings between two qubits, we propose placing the donors close to each other but preparing the nuclear spins in different states~\cite{Morley2010}, resulting in different hyperfine couplings at each site which convert the natural Heisenberg coupling to an approximate Ising coupling (see e.g.~\cite{Benjamin2003}) and allow individual addressability of the electrons. This is a similar method to that proposed in~\cite{Kalra2014}, where the nuclear states of P atoms are set in different states to realise a two-qubit entangling gate. Such an initialisation of the nuclear spins could be performed starting from a nuclear spin polarised sample (polarisations of up to 90\% were reported in~\cite{Sekiguchi2010}). The process of flipping the nuclear spin from $\frac{9}{2}$ to $-\frac{9}{2}$ would take 9 steps of $\sim$10$\mu s$ each, so around $90\mu s$. However, the nuclear spins are stable on a timescale of hours~\cite{Feher1959,Castner1962}, so this process could be done once for many operations of the gate, and so does not significantly affect the gate time.
We must also apply a magnetic field much larger than the hyperfine splitting to ensure that the nuclear spins do not evolve. As mentioned in Sec.~\ref{sec:DonIntro}, optimal working points~\cite{Mohammady2010,Mohammady2012,Balian2012,Morley2013} cancellation resonances~\cite{Mohammady2010} and clock transitions~\cite{Wolfowicz2013} are not useful here because these schemes only work at lower magnetic fields. The nuclear spins of the donors are initialised in different states $I_1^{z},I_2^{z},I_3^{z}$, and an AC field is applied on qubit 2 in the $x$-direction (which could be applied globally since the hyperfine splittings are different). The effective spin Hamiltonian is then (see Sec.~\ref{sec:DonIntro})
\begin{align}
H &= \omega_L \sum_{n=1}^3 S_n^z + \sum_{n=1}^3 A I_n^z S_n^z + \Omega \cos ( \omega_x t )S_2^x+ \sum_{n=1}^2 J \bm{S}_n \cdot \bm{S}_{n+1},
\end{align}
where the $\sum_{n=1}^3 A I_n^{z} Z_n$ terms account for the different hyperfine splittings and the $\omega_L$ terms account for the large magnetic field in the $z$-direction. In a similar manner to the previous section, we switch to the interaction picture of this Hamiltonian with respect to $H_0 :=\omega_L\sum_{n=1}^3 S_n^z - JS_2^z + \sum_{n=1}^3 A I_n^z S_n^z $:
\begin{align}
H_I &= e^{-i H_0t} (H - H_0) e^{iH_0 t}\nonumber\\
&= \frac{1}{2} e^{-i\tilde{\omega}_2Z_2t}\Omega \cos ( \omega_x t )X_2 + \frac{1}{4} J \sum_{n=1}^2 \left[ e^{-i(Z_n\tilde{\omega}_n + Z_{n+1}\tilde{\omega}_{n+1})t} ( X_n X_{n+1} + Y_n Y_{n+1} ) + Z_n Z_{n+1} \right]
\end{align}
where $\tilde{\omega}_1 := \omega_L + A I_1$, $\tilde{\omega}_2 := \omega_L + A I_2 - 2J$ and $\tilde{\omega}_3 :=\omega_L+ A I_3 $. By setting $\omega_x = \tilde{\omega}_2$ and using the identities from Appendix~\ref{app:IntPic}, the evolution in this interaction picture is approximately equivalent to the evolution by the following Hamiltonian:
\begin{align}\label{eqn:Rot1}
&H_I = \frac{1}{2}J Z_2 + \frac{1}{4}J(Z_1 Z_2 + Z_2 Z_3) +\frac{1}{4} \Omega X_2 + O\left(\sum_{n=2}^3 \frac{J}{A I_n^z - A I_{n-1}^z \pm 2J} + \frac{\Omega}{2(\omega_L -2J + AI_2^z)} \right),\end{align}
which is in the same form as the Hamiltonian in (\ref{eqn:HamToff1}). Note that the additional factor of $-2JZ_2$ added into the rotating frame means that a different local $z$-rotation must be applied to achieve an exact Toffoli gate.
To achieve the minimum error we set $I_1^z = \frac{9}{2}$, $I_2^z=-\frac{9}{2}$, and $I_3^z =\frac{9}{2}$. The large nuclear spin and hyperfine coupling for bismuth (A=1.475 GHz) justifies the use of bismuth over other donors. $\Omega$ must be small enough that the errors in (\ref{eqn:Rot1}) and the systematic error in (\ref{eqn:p}) are small, and large enough to allow a gate time within the decoherence time of the bismuth donors ($\sim 0.5 \text{ms}$ in natural silicon at the high magnetic fields where the nuclear spin is a good quantum number~\cite{Morley2010,George2010}, or 700ms in isotopically pure silicon~\cite{Wolfowicz2012}).
$J$ must be small enough to give a small error in (\ref{eqn:Rot1}), but large enough to reduce the systematic error in (\ref{eqn:p}). The bandwidth of the $X_2$ pulse must be large enough compared to the linewidth of the dopant energy levels (around 0.2MHz in natural silicon~\cite{Tyryshkin2003} and around 20kHz in isotopically pure silicon~\cite{Wolfowicz2012}) so that it excites the whole level. It also must be small enough that it can accurately resolve only one resonance and so only flip qubit 2 under the right conditions (this second condition is already implicitly taken into account in the systematic gate error in (\ref{eqn:p})). In this scheme we assume a square pulse, giving a sinc function of width $1 /t_{gate}$ in the frequency domain. See Fig.~\ref{fig:ToffEnergy} for an illustration of the energy levels of the Toffoli gate and the linewidth and bandwidth.
With settings of $\Omega = 1$MHz $J = 30$MHz, the gate time is $2\mu s$. The systematic gate error term in (\ref{eqn:p}) for these settings is $\sim 10^{-3}$, the errors due to decoherence are roughly $1-e^{-t_{gate}/T_2} \sim 10^{-6}$ and the errors in the Hamiltonian in (\ref{eqn:Rot1}) become $\sim 10^{-3}$, so we expect a gate with probability of failure $\lesssim 10^{-3}$ to be achievable. The bandwidth is around 4 MHz which is much larger than the linewidth of pure silicon. At the expense of greater gate error, the gate time can be as low as 50ns whilst still keeping all of the errors below 1\% (e.g. with $\Omega = 40MHz$, $J = 300MHz$). Since fault tolerant thresholds can be as high as 1\%~\cite{Knill2005}, and fault tolerant thresholds for classical computation are larger, we see that fault-tolerant gates could be achieved using today's technology for both classical and quantum computation. For an exact quantum Toffoli gate, local rotations would need to be performed, however we would not expect these operations to change the result significantly since they can in principle be applied in a short time.
\begin{figure}[h]
\begin{center}
\includegraphics[scale = 0.45]{ToffoliEnergy-eps-converted-to.pdf}
\caption{Visual guide to the energy levels of the Toffoli gate. $c_1$ and $c_3$ indicate the states of qubits 1 and 3, and $E_2$ indicates the corresponding shift in the resonance of qubit 2. The inset illustrates the frequency profile of the square $X$ pulse, with bandwidth $\sim\Omega$. }
\label{fig:ToffEnergy}
\end{center}
\end{figure}
\subsection{Fredkin gate using bismuth donors in silicon}
To create the Fredkin gate in Sec.~\ref{sec:FredHeis} we consider starting with a similar Hamiltonian to the Toffoli gate
\begin{align}
H &= \omega_L \sum_{n=1}^3 S_n^z + \sum_{n=1}^3 A I_n^z S_n^z +\sum_{n=1}^2 J \bm{S}_n \cdot \bm{S}_{n+1},
\end{align}
Transforming into the interaction picture with respect to $H_0= \omega_L \sum_{n=1}^3 S_n^z + A I_1^z S^z_1 + A I_2^z(S_2^z + S_3^z)$ gives
\begin{align}
e^{-iH_0t}(H-H_0)e^{-iH_0t} &=\frac{1}{4} J_{12}e^{-it ( \tilde{\omega}_1 Z_1 + \tilde{\omega}_2 Z_2 )} (X_1 X_2 + Y_1 Y_2)+\frac{1}{4} J_{23}e^{-it ( \tilde{\omega}_2 Z_2 + \tilde{\omega}_3 Z_3)}( X_2 X_3 + Y_2Y_3) \nonumber\\
&+ \frac{1}{4} J_{12}Z_1 Z_2 + \frac{1}{4} J_{23} Z_2 Z_3 + \frac{1}{2}A(I_3 - I_2)Z_3
\end{align}
where $\tilde{\omega}_1 := \omega_L + I_1^z A$, $\tilde{\omega}_2 =\tilde{\omega}_3 := \omega_L + I_2^z A$. Using identity (\ref{eqn:Iden2}), this Hamiltonian behaves the same as
\begin{align}\label{eqn:Rot2}
H_I &= \frac{1}{4} J_{12} Z_1 Z_2 +\frac{1}{4} J_{23}(X_2 X_3 + Y_2 Y_3+ Z_2 Z_3 ) + \frac{1}{2}(A I_3^z - A I_2^z)\sigma_3^z +O\left( \frac{J_{12}}{|A I_1^z - A I_2^z|}\right),
\end{align}
which is similar to the Hamiltonian in (\ref{eqn:HamFred}). To satisfy the conditions of the gate found in Sec.~\ref{sec:FredHeis}, $J_{12} = 2(A I_3^z - A I_2^z)$ and $J_{23}^2 (\frac{m^2}{(2n+1)^2}-1) =4(A I_3^z - A I_2^z)^2$. For the minimal gate time ($n=1$), this would give $J_{23} = \frac{1}{\sqrt{3}} J_{12}$, or alternatively $J_{23}$ could be tuned to different fractions of $J_{12}$ by altering $m$ and $n$. For example, we can achieve $J_{23} = J_{12}(1+10^{-6} )$ by setting $n = 675$ and $m = 2340$. The resulting gate time with $J_{12} = 2(A I_3^z - A I_2^z)$ would be $675\times 0.5\text{ns} \sim 0.5 \mu\text{s} = 10^{-3}T_2$, so such tuning still would not lead to large decoherence errors.
With $J_{12}=2(A I_3^z - A I_2^z)$, the error in (\ref{eqn:Rot2}) is $O( 2(I_2^z - I_3^z) / (I_1^z - I_2^z))$, which is minimal when $ I_1^z = -\frac{9}{2}, I_2^z = \frac{9}{2}, I_3^z = \frac{7}{2}$ ($I_2^z \neq I_3^z$ since this would mean $J_{12} = 0$). To decrease this error, a fourth qubit could be included, as shown in Fig.~\ref{fig:FredHeisT}. Adding this qubit with Ising coupling $J_{E2}$ effectively adds another local magnetic field to qubit 2 and so changes the resonance condition to $J_{12} +J_{E2}= \omega_3 - \omega_2$ (assuming that qubit E is prepared in the $\ket{0}$ state). Note however that this coupling also has an error associated to it, and so the optimal situation is with $J_{E2} = J_{12} =(A I_3^z - A I_2^z) = 1.475$GHz. This results in an error of around 11\%. Adding yet another control qubit $E'$ with coupling $J_{E'2}$ and choosing couplings such that $J_{12} = J_{E2} = J_{E'2} = \frac{2}{3}(A I_3^z - A I_2^z)$ results in errors of around 7\% (adding any more becomes unrealistic as the control qubits may begin to interact significantly). The gate time in this situation would be $\sim1.5\text{ns}$ which is still significantly smaller than $T_2$.
Additionally, setting $J_{12} = 2(A I_3^z - A I_2^z)$ is not straightforward, since the donor position is not continuously tunable. There are oscillations in the exchange interaction that depend on the separation between donors and the orientation of the donors relative to the crystal~\cite{Koiller2001}, although these oscillations can be minimised if donors are aligned along a particular axis, or if strain is applied~\cite{Koiller2002}. Atomically-precise positioning of the donors is difficult but possible~\cite{OBrien2001,Schofield2003,Fuechsle2012}, so a possible approach is to position the donors at separations of around 15-20nm such that $J_{12} \approx 2(A I_3^z - A I_2^z)$ and then use a magnetic field gradient to change the resonances of ions 2 and 3 to match $J_{12}$ (this magnetic field would also decrease the error in (\ref{eqn:Rot2})). Since it would be possible to bring a magnetic tip close to the sample, large magnetic field gradients of up to $10^7 \text{ Tm}^{-1}$ should be possible~\cite{Mamin2007}, which allows a tuning of up to $\pm 0.5\text{ GHz}$, although this could introduce extra heating of the qubits. Alternatively, we could adopt the method in~\cite{Kane1998}, and use electric gates to tune the inter-donor couplings and increase hyperfine interactions, however this might introduce extra noise due to charge fluctuations.
\begin{figure}[t]
\begin{center}
\includegraphics[scale = 0.5]{FredkinHeisT-eps-converted-to.pdf}
\caption{Adding a fourth qubit to the original setup, in order to relax the constraints on $J_{23}$ and provide a control to turn the gate on and off.}
\label{fig:FredHeisT}
\end{center}
\end{figure}
\subsection{Experimental implementation of T-shaped Fredkin gate}
Achieving an experimental implementation of this gate using ion traps or donors is more challenging. One option using ions would be to use different pulses which rapidly oscillate qubit $A$ and then rapidly oscillate qubits $1$ and $3$. Such a rapid oscillation has the effect of decoupling a qubit from its neighbouring spins, allowing us to perform the $Z_AZ_2$ interaction followed by a $X_1X_2 + Y_1Y_2+ X_2X_3 +Y_2Y_3$ (also obtained via stroboscopic means, by rapidly changing the orientation of the field), which approximates to the desired interaction. However, such a technique is cumbersome and deviates from the philosophy in this work of using just a single time-independent pulse.
\section{Conclusions}
We have outlined three different protocols to realise 3-qubit gates, with uniform fixed interaction without any extra states or ancillas, such that these gates are implemented in one step (i.e.\ using only a single pulse). A Toffoli gate can be realised using a linear array of qubits interacting via uniform Ising interactions and with a transverse magnetic field applied to the middle qubit. A Fredkin gate can be achieved using a system of three qubits with Ising and Heisenberg interactions. Another Fredkin gate can also be achieved using four qubits in a T shape, with Ising and XY interactions. Note that all of the gates realised in this work are only equivalent to exact 3-qubit quantum gates up to local $Z$-rotations. It would be interesting and useful to find more general results on which types of uniform interactions on 3-qubits could achieve a Toffoli gate. Other approaches could also be taken such as taking logarithms of Toffoli and Fredkin gates.
We have investigated whether or not current technology can be used to perform these gates. We find that using ion traps we would not be able to create a gate that exceeds the quantum computation threshold, but by using bismuth donors in silicon fault-tolerant Toffoli gates should be possible. For the linear Fredkin gate, we find that bismuth donors in silicon would yield a gate with $\sim 10 \%$ error, which could be reduced by the addition of extra qubits to around 4\%. We have not yet discovered an appropriate experimental scheme for the T-shaped Fredkin gate. Whilst some of the experimental realisations here are limited in success probability, and gate times and gate size are similar to existing classical technology, there is no reason why technological improvements may change this in the future, and other experimental realisations could give better results.
Further work could include using more sophisticated pulses that are more selective than a square wave pulse, such as a Gaussian profile~\cite{Vandersypen2005,Vasilev2004}, and including the effects of realistic noise on the gates. It would also be interesting to see if the idea behind the T-shaped Fredkin gate in Sec.~\ref{sec:XX} could be applied to spin chains that perform perfect state transfer.
\chapter{Adiabatic graph-state quantum computation}
\label{chap:AGQC}
In Chapter~\ref{chap:Intro} we introduced the three most widely used models of computation; the circuit model, measurement-based quantum computation (MBQC) and adiabatic quantum computation (AQC). These models are all ``equivalent'' to each other in the following sense; for a given computation, the number of gates required in the circuit model is \emph{polynomially equivalent} to the number of measurements required in MBQC~\cite{Raussendorf2001}, whilst the number of gates required in the circuit model is \emph{polynomial} in the inverse energy gap of the equivalent AQC computation~\cite{Aharonov2007}.
These results are incredibly important, but from a more practical point of view we are not only interested in polynomial equivalence, but also in the differences that each model has which may make them more attractive when building a quantum computer, and which may give insight into what makes quantum computation powerful. A striking example of these differences is for a circuit consisting of Clifford gates on $N$ qubits: In the circuit model the computation takes at least $\log N$ gates, while in MBQC the measurements can be done in a single constant time step, with a classical processing time that scales with $\log N$~\cite{Browne2011}. So, while it is not possible to use one computational model to (exponentially) outperform another, in MBQC it is possible to move computational time from one ``form'' to another. This presents a potential advantage of MBQC over the circuit model when classical processing is a cheap resource.
Bacon and Flammia proposed the direct translation of MBQC on a cluster state into an adiabatically driven evolution~\cite{Bacon2010}, which they call \emph{adiabatic cluster-state quantum computation} (ACSQC). In this model, the discontinuous measurements of MBQC are replaced with continuous adiabatic transformations. Whilst the evolution proceeds adiabatically, this is not AQC in the typical sense, as the Hamiltonian has a degenerate ground space. It is in fact a type of \emph{open-loop holonomic quantum computation} (see Section~\ref{sec:Holon}), in which the time-varying Hamiltonian performs geometric transformations on information encoded within the degenerate subspace~\cite{Zanardi1999,Pachos1999,Kult2006}. There are also several other works combining ideas from MBQC and AQC. In ancilla-controlled adiabatic evolution~\cite{Wiebe2012,Kieferova2014,Hen2014}, computation is carried out by a combination of adiabatic passage and measurement. In adiabatic topological quantum computation~\cite{Cesare2014}, defects in a topological code are adiabatically deformed to perform logical operations. There are also examples of adiabatically driven computations on non-stabiliser states such as symmetry-protected states of matter~\cite{Williamson2014} and generalised cluster states~\cite{Brell2014}.
In this chapter we discuss the extension of ACSQC to more general graph states which have \emph{gflow}, a property that is fundamentally linked to the adaptivity of MBQC and determines if a given measurement pattern can produce a deterministic computation~\cite{Danos2006,Browne2007}. This new \emph{adiabatic graph-state quantum computation} (AGQC) allows us to investigate how the properties of MBQC translate into a continuous adiabatically driven evolution. Such a connection between two quite different models could yield interesting insight into what underlies the power of quantum computation, and allows ideas, insight and techniques from one model to be shared with the other. In particular, we mentioned above that in MBQC it is sometimes possible to decrease quantum computation time whilst increasing classical processing time, so that there is a trade-off between quantum and classical time. We are interested in how this will manifest itself in an adiabatically driven model.
Since there is also an essential requirement in MBQC that measurements are adaptive, as measurement outcomes are non-deterministic, another natural question to ask is what happens if the adiabatic `measurements' in AGQC are performed in a different order to that prescribed in MBQC. At first glance it appears as if we are really post-selecting the correct measurement outcome by using an adiabatic process, so it is not obvious how the causal structure of the measurements should necessarily remain.
We begin by introducing the adiabatic cluster-state quantum computation model, before generalising it to adiabatic graph-state quantum computation in Section~\ref{sec:GeneralAGQC}. We then investigate what trade-off exists between the adiabatic time, the classical computing time for MBQC and the degree of the initial Hamiltonian in Section~\ref{sec:AGQCexample} and finally discuss the role of the ordering of measurements in adiabatic graph-state quantum computation in Section~\ref{sec:Order}. This chapter is based on the work that was published in~\cite{Antonio2014}.
\section{Adiabatic cluster-state quantum computation}\label{sec:ACSQC}
In Section~\ref{sec:MBQC} it was shown how it is possible to achieve universal MBQC and how stabilisers can be used to describe the measurement process. Using these tools, we will now review \emph{adiabatic cluster-state quantum computation} (ACSQC) which converts MBQC on a cluster state into an adiabatically driven holonomic quantum computation~\cite{Bacon2010}. First consider a 1D cluster state of $N$ qubits. The twisted stabiliser generators for this state are
\begin{eqnarray}\label{Stab1D}
K_v^{\theta_{v}} &=Z_{v-1} X^{\theta_{v}}_{v} Z_{v+1} \quad v = 2,...,N-1; \; \;K_N &=Z_{N-1} X^{\theta_{N}}_{N}.
\end{eqnarray}
To fit with the notation used in~\cite{Bacon2010}, we will use a slightly different definition of these stabilisers generators. In this notation, stabiliser generators are defined as $T_v := K^{\theta_{v+1}}_{v+1}$, i.e.\ the same as the $\{ K^{\theta_{v}}_v \}$ stabilisers except the indices are shifted by 1 site.
The first important thing to note is that, since the stabilisers by definition commute and act as identity on the cluster state, the cluster state is the ground state of the following Hamiltonian:
\begin{eqnarray}
H_0 \equiv - \gamma\sum_{v= 1}^{N-1} T_v,
\end{eqnarray}
where $\gamma$ parametrises the strength of the interactions (in this work we assume that the highest coupling strength possible is chosen, so that $\gamma$ is already set at its highest value and therefore is a constant that does not scale with the system size). The eigenspace is spanned by the two states which can be used to encode a qubit:
\begin{eqnarray}
\ket{0}_L = \prod_{v \sim w} \textsc{cz}_{(v,w)} \ket{0}_1 \bigotimes_{n > 1} \ket{+_{\theta_n}}_n, \; \; \ket{1}_L = \prod_{v \sim w} \textsc{cz}_{(v,w)} \ket{1}_1 \bigotimes_{n > 1} \ket{+_{\theta_n}}_n
\end{eqnarray}
The logical $X$, $Y$, $Z$ operators of this qubit ($X_L$, $Y_L$ and $Z_L$, respectively) are the same as in MBQC;
\begin{eqnarray}\label{eqn:LogOp}
X_L \equiv X_1 Z_2, \; Y_L \equiv Y_1 Z_2 , \; Z_L \equiv Z_1.
\end{eqnarray}
The computation begins by preparing the ground state of $H_0$, which could be done by starting in the ground state of a uniform magnetic field and adiabatically evolving to $H_0$ (encoding an arbitrary input state could then be done by adding a field $\alpha_x X_L + \alpha_y Y_L + \alpha_z Z_L$ to $H_0$). Then instead of measuring qubits as in MBQC, the first stabiliser generator in the Hamiltonian $T_1$ is adiabatically replaced with $X_1$:
\begin{align} \label{Ht}
H(s) &=- (1-s)\gamma T_1 - s \gamma X _1 - \gamma \sum_{n =2}^{N-1} T_n
\end{align}
where $s = t/\tau$, $\tau$ is the total run time and $0 \le t \le \tau$. To see what happens after this adiabatic substitution, the logical operators can be multiplied by stabilisers (since they act as identity, as shown in Section~\ref{sec:MBQC} on MBQC) until they commute with $X_1$ (i.e.\ the logical operators are put in a form in which they are conserved during the adiabatic transformation). $X_L$ already commutes with $X_1$, but $Z_L$ doesn't so it must be multiplied by $T_1$:
\begin{align}
Z_L \rightarrow Z_L T_1 = X_2^{\theta_2} Z_3.
\end{align}
Similarly $Y_L$ also doesn't commute with $X_1$, so must be multiplied by $T_1$:
\begin{align}
&Y_L \rightarrow Y_L T_1 =(Y_1 Z_1) (Z_2 X_2^{\theta_2}) Z_3 \nonumber\\
&=iX_1 (-i)e^{i\pi Z_2/2} e^{-i \theta_2 Z_2} X_2 Z_3= X_1 e^{-i \theta_2 Z_2} e^{i\pi Z_2/2} X_2 Z_3 = X_1 e^{-i \theta_2 Z_2} Y_2 Z_3.
\end{align}
Then, after setting $X_1 = \idop$ since the system is in the $+1$ eigenstate of $X_1$, the logical operators are transformed into
\begin{eqnarray}\label{eqn:LogOp2}
X_L \rightarrow Z_2, \;Y_L\rightarrow e^{-i \theta_2 Z_2} Y_2 Z_3,\; Z_L \rightarrow X_2^{\theta_2} Z_3.
\end{eqnarray}
If new logical operators $\{ \tilde{X}_L, \tilde{Y}_L, \tilde{Z}_L \}$ are defined in the same way as in eqn.\ (\ref{eqn:LogOp}):
\begin{eqnarray}
\tilde{X}_L \equiv X_2 Z_3, \; \tilde{Y}_L \equiv Y_2 Z_3, \;\tilde{Z}_L \equiv Z_2.
\end{eqnarray}
Then expressing the operators in (\ref{eqn:LogOp2}) in this basis gives:
\begin{align}\label{eqn:LogOp3}
X_L &\rightarrow Z_2 = \bar{U}_{2} \bar{\text{H}}_L \tilde{X}_L \bar{\text{H}}_L\bar{U}_{2} \nonumber\\
Y_L &\rightarrow e^{-i \theta_2 Z_2} Y_2 Z_3 \equiv \bar{U}_{2} \bar{\text{H}}_L \tilde{Y}_L \bar{\text{H}}_L\bar{U}_{2} \nonumber\\
Z_L &\rightarrow X_2^{\theta_2} Z_3.\equiv \bar{U}_{2} \bar{\text{H}}_L \tilde{Z}_L \bar{\text{H}}_L\bar{U}_{2},
\end{align}
where $\bar{\text{H}}_L$ is a Hadamard operation acting in the logical subspace, and $\bar{U}_{v} :=\exp[ -i\theta_{v} Z_L / 2]$ is a logical $Z$ rotation. So the effect of the evolution on the logical information is to move it one step down the chain, and to apply $\bar{U}_{2} \bar{\text{H}}$ to the information (exactly the same as in the MBQC example in Sec.~\ref{sec:MBQC}). Then at the next step $T_2$ is replaced with $X_2$, and so on from left to right down the chain (see Fig.~\ref{tab:Normal} for an illustration of this). This results in the information being encoded in the $N^{th}$ qubit, with the information transformed by the operation $\bar{U}_{tot} = \bar{H} \prod\nolimits_{v = N-2}^2(\bar{U}_{v} \bar{\text{H}})$, and since $\bar{U}_{v} = \exp[ -i\theta_{v} Z_L / 2]$ this will depend on the sequence of angles $\{\theta_v \}$ used (following the convention in~\cite{Bacon2010} we set $\theta_1 = 0$). Note that, due to the Hadamard operations, this results in alternating $X$ and $Z$ rotations, which allow universal single-qubit rotations.
When replacing one stabiliser by a Pauli $X$ operator, the speed of this adiabatic transformation is limited by the ratio of the minimum energy gap and $\Vert \dot{H}(s) \Vert$ (see Sec.~\ref{sec:AdTheor}). For the above case, the minimal energy gap of $H(s)$ is $\sqrt{2}$, and $\| \dot{H} \| = 1$, which means the time taken is
\begin{eqnarray}\label{eqn:OneStep}
\tau \geq \tau_0 :=\frac{1}{\varepsilon 2^{1 + \delta/2} \gamma}
\end{eqnarray}
for some fixed $0<\delta\leq1$ and $\varepsilon$. For the remainder of this chapter we will compare adiabatic evolution time to this time $\tau_0$, so that this is our definition of one unit of time for the adiabatic computation, and we say that the adiabatic computation takes one `step' if the time satisfies the same condition as above.
\begin{table}[t]
\begin{eqnarray}
\begin{array}{l | cccccc}
\mbox{Start}& T_1 & T_2 & T_3 & \dots & T_{N-2} & T_{N-1}\\
& \downarrow & & & & & \\
\mbox{Step 1} & X_1 & T_2 & T_3 & \dots & T_{N-2} & T_{N-1}\\
& & \downarrow & & & & \\
\mbox{Step 2} & X_1 & X_2 & T_3 & \dots & T_{N-2} & T_{N-1}\\
& & & \downarrow & & & \\
\mbox{Step 3} & X_1 & X_2 & X_3 & \dots & T_{N-2} & T_{N-1}\\
\hspace{10mm} \vdots&& & \vdots \\
& & & & & \downarrow & \\
\mbox{Step } (N- 2) & X_1 & X_2 & X_3 & \dots & X_{N-2} & T_{N-1}\\
& & & & & & \downarrow \\
\mbox{Step } (N- 1) & X_1 & X_2 & X_3 & \dots & X_{N-2} & X_{N-1}
\end{array}
\end{eqnarray}
\caption{An illustration of the method to perform single qubit operations in adiabatic cluster-state quantum computation; arrows indicate an adiabatic transition from one operator to the other.}
\label{tab:Normal}
\end{table}
\subsection{Two-qubit gates}\label{sec:CNOT}
We have just seen how to perform arbitrary single qubit rotations in the adiabatic cluster-state model. In order to perform universal quantum computations, an entangling gate such as a \textsc{cnot} gate must also be possible. Consider the graph discussed in Sec.~\ref{sec:MBQC}, with two rows of qubits, labelled $a$ and $b$, with 3 columns numbered from left to right (see Fig.~\ref{fig:CNOT}). The stabilisers are $T_{a1} = K_{a2}$, $T_{a2} = K_{a3}$ and similarly for $T_{b1}$, $T_{b2}$. Then the protocol starts with an initial Hamiltonian in which all angles $\{ \theta_n \}$ are 0, and which the inputs are on qubits $a_1$ and $b_1$, and the outputs are on qubits $a_3$ and $b_3$:
\begin{align}
H_{\textsc{cnot}} &= -T_{a1} -T_{a2} - T_{b1} - T_{b2} \nonumber\\
&=-Z_{a1}X_{a2}Z_{a3} Z_{b2} - Z_{a2} Z_{b1} X_{b2} Z_{b3} - Z_{a2} X_{a3} - Z_{b2 }X_{b3}.
\end{align}
Following a similar procedure as in the 1D case above, we see that the effect of adiabatically replacing all of these stabilisers with $X$ operators results in a \textsc{cnot} gate acting on the encoded information. Each replacement of a single stabiliser by a local Pauli operator still has the same adiabatic time as in the previous section, i.e.\ the adiabatic evolution time takes the form $\tau_0$ (provided these substitutions are done in the correct order). Note that also a \textsc{cz} gate can easily be achieved using an adiabatic scheme based on the gates in~\cite{Browne2007}, in which only 3 qubits are required to perform a gate on two qubits.
\begin{figure}[h]
\centering
\includegraphics[scale = 0.7]{CNOT-eps-converted-to.pdf}
\caption{An illustration of the graph used to perform a \textsc{cnot} gate in adiabatic cluster-state quantum computation.}
\label{fig:CNOT}
\end{figure}
\section{Adiabatic graph-state quantum computation}
In the previous section we saw that universal quantum computation is possible using adiabatic substitutions, instead of measurements, on the cluster state. Here this approach is generalised to other graph states using tools from MBQC. This is an extension of adiabatic cluster-state quantum computation, which we call \emph{adiabatic graph-state quantum computation} (AGQC).
\subsection{General translation of an MBQC pattern to a holonomic computation}
\label{sec:GeneralAGQC}
Here we will look at converting any general measurement pattern on a graph state into a holonomic computation. Consider an open graph state with vertices $V$, \emph{gflow} $(g,<)$, input vertices $I$ and output vertices $O$. For simplicity we take the number of inputs to be equal to the number of outputs $|I|=|O|$, but all statements and proofs can be easily extended to the cases $|O|>|I|$ ($|O|<|I|$ is not allowed as this would mean information is lost). The AGQC protocol starts in the ground state of the initial Hamiltonian
\begin{eqnarray}
H_0 =-\gamma \sum_{v \not\in O} T_v,
\end{eqnarray}
where the $T_v$ operators are products of the twisted stabiliser generators,
\begin{align}\label{eqn:GflowStab}
T_v := \prod_{w \in g(v)} K_w^{\theta_w},
\end{align}
$\forall v \notin O$. The ground state of $H_0$ corresponds to the open graph state. Reaching the ground state of $H_0$ could be achieved by starting in the ground state of a uniform magnetic field and adiabatically changing the Hamiltonian to $H_0$. The final Hamiltonian will be $H_f=-\gamma\sum_v X_v$, and the transition will be done in steps, as in the cluster-state case (for which $g(v)$ contains only one element and so $T_v = K_{g(v)}$, as seen in Section~\ref{sec:ACSQC}). The initial Hamiltonian can be computed efficiently from the graph and \emph{gflow}, since each $T_v$ operator can be found through multiplication of the $K_w^{\theta_w}$ generators which can be performed in a time that scales as $O(\log (\mbox{max}_v |g(v)|))$ using the methods in~\cite{Aaronson2004}.
Notice that the $T_v$ operators can cover many sites; this is an important property that we must keep track of, so we define the number of sites an operator covers as \emph{degree} of the operator:
\begin{defin}[Degree of an operator]
The degree of an operator is the number of non-trivial (i.e. non-identity) terms when the operator is written as a tensor product of 2 dimensional Pauli matrices.
\end{defin}
So, for example, the operator $Z_1 \otimes X_2 \otimes \idop_3 \otimes \idop_4$ has degree 2, and the degree of the 1D cluster state stabilisers are $\leq 3$. For a Hamiltonian which is a sum of many operators, we say the Hamiltonian $H$ has degree $k$ if all of the operators in $H$ have degree $\leq k$.
It will also be useful to define the size of a \emph{gflow}, in terms of its stabilisers
\begin{defin}[$|g(v )|$]
The size $|g( v )|$ of a \emph{gflow} is the number of qubits for which the product of the stabilisers over $g( v )$, $\prod_{v\in g(v )}K_v$, is non-trivial.
\end{defin}
E.g. for the cluster-state, there is only one vertex in $g(v)$, and since each stabiliser generator acts on up to 5 qubits, the size of the \emph{gflow} is 5.
To perform the computation, the operators $\{T_v \}$ can be replaced one-by-one in an order that doesn't violate the \emph{gflow}, or those in the same layer can be replaced all at once. One may choose instead to ignore the ordering or \emph{gflow} and replace all the stabilisers in one adiabatic step, as shown in~\cite{Bacon}. The downside of this approach is that the energy gap will most likely shrink polynomially in the the number of qubits, as suggested by the numerical results in~\cite{Bacon}, which means the gap protection is lost, and additionally analytical or numerical results become very difficult to find except for a few specific cases. Thus we consider stepwise computations with fixed energy gap, which still tell us many interesting things about adiabatically driven versions of MBQC. We would also expect that replacing all of the stabilisers will only result in non-zero energy gap when the graph has \emph{gflow}; this is supported by the re-ordering results in Sec.~\ref{sec:Order}, and can be verified for the example in Fig.~\ref{fig:NoGflow}.
Replacing the stabilisers one-by-one will take $N$ steps of time $\tau_0$. When replacing layer by layer, the adiabatic transition for the $k$th step is governed by the interpolation Hamiltonian
\begin{equation}
H_{L_k}(s) = (1-s) H_{k-1} + s H_k,
\end{equation}
where
\begin{equation}
H_k= \gamma \left(-\sum_{v\leq L_k}X_v - \sum_{v>L_k}T_v \right).
\end{equation}
What is the time needed to perform the $k^{th}$ step? To determine this we will need to find the form of the energy gap and the form of $\Vert \dot{H} \Vert$ to put into the adiabatic formula given in eqn.\ (\ref{eqn:RunTime}). First we split the Hamiltonian into two parts
\begin{align}
\label{eqn:HamSplit}
H_{L_k}(s) = - \gamma \left( \sum_{v < L_k}X_v + \sum_{v>L_k}T_v \right) - \gamma \left(\sum_{v\in L_k} s X_v + (1-s)T_v \right).
\end{align}
The terms in the sum on the left commute with each other, which can be seen from the definitions of the $\{ T_v \}$. Each of the summands on the right also commutes with each other and all of the summands on the left. To see this, note that from the conditions of \emph{gflow}, a product of stabiliser generators $\prod_{w \in g(v)} K_w$ has an even number of connections to vertices in the same layer or in previous layers (which means an even number of $Z$ operators, which cancel to give identity) and an odd number of connections to vertex $v$ (which means there is one $Z_v$ term in $\prod_{w \in g(v)} K_w$. This does not change when we use twisted stabiliser generators). Therefore $\{T_v,X_v\} = 0$, and $[T_w,X_v] = 0$ for all $w \geq v$.
A full analysis of the eigenvalues of this Hamiltonian is given in Appendix~\ref{sec:LinAlg}; the result is that the eigenvalues of $H_{L_k}$ are $-|L_k|\gamma\eta,-(|L_k|-2)\gamma\eta,...,(|L_k|-2)\gamma\eta,|L_k|\gamma\eta$ plus factors of $\pm \gamma$, where $\eta := \sqrt{(1-s)^2 + s^2} $. The factors of $\pm \gamma$ do not contribute to the minimum energy gap since $\gamma \eta\le \gamma$. The energy gap between the ground state and first excited state is therefore $2 \gamma \eta$, which is minimal at $s=\frac{1}{2}$, at which point the gap is $\sqrt{2}\gamma$.
The time-derivative of the Hamiltonian is
\begin{eqnarray}
\dot{H}(s)=\gamma\sum_{u \in L_k} ( T_u - X_u).
\end{eqnarray}
Using the same methods as in Appendix~\ref{sec:LinAlg}, the eigenvalues of this derivative are $\{ -|L_k|\gamma,-(|L_k|-2)\gamma,...,(|L_k|-2)\gamma,|L_k|\gamma \}$, so the spectral norm of $\dot{H}(s)$ is $|L_k|\gamma$. Inserting the minimal energy gap and $\Vert \dot{H}(s) \Vert$ into equation (\ref{eqn:RunTime}) for an adiabatic evolution, we then find that after adiabatically replacing the $k^{th}$ layer using a linear interpolating function, the final state will be in the ground space with some fixed error $\varepsilon$ if the adiabatic time $\tau$ scales as
\begin{eqnarray}
\tau = \Omega \left( \frac{ \Vert \dot{H}(s) \Vert^{1+\delta}}{\Delta(s)_{min}^{2+\delta}} \right) = \Omega \left( |L_k|^{1 + \delta} \right)
\end{eqnarray}
for some $0 < \delta \leq 1$.
Now consider what happens to the information when we replace all $T_v$ operators defined above with an $X_v$ operator, in the order given by \emph{gflow}. Following the method in~\cite{Bacon2010}, the effects on the information can be seen by multiplying logical operators with the $T_v$ operators in such a way that the logical operators commute with all the adiabatic `measurements'. This means that it is possible to update the logical operators $\alpha_L$, to $\tilde{\alpha}_L$ such that $[\tilde{\alpha}_L,X_v] = 0$ for all $v$, and where $\alpha = X,Y,Z$. This then tells us what state is output by the computation. For any graph state with \emph{gflow}, it is always possible to do this, since all we have to do is multiply any $Z_v$ or $Y_v$ operators by the $T_v$ defined above; the \emph{gflow} conditions guarantee that after this replacement, $Z_v \to \idop_v$ and $Y_v \to X_v$, and the new logical operator will commute with any $X_w$ such that $v\ge w$. For graphs without \emph{gflow}, such a process is not generally possible, since when multiplying a $Z_v$ or $Y_v$ operator by a $T_v$ stabiliser which violates \emph{gflow}, the result may contain an operator $Z_w$ such that $w < v$. This property can be seen in the example in Fig.~\ref{fig:NoGflow}; to make the logical operator $Z_1$ commute with $X_1$, we multiply $Z_1$ by $K_3$ giving $Z_1 \to \tilde{Z}_1 = X_3 Z_2$. Now to make this commute with $X_2$, we can only multiply by $K_4$, leaving $\tilde{Z}_1 = Z_1 X_3 X_4$ which no longer commutes with $X_1$.
To see that a holonomic quantum computation on a graph state with \emph{gflow} performs the same computation as the equivalent measurement pattern in MBQC, consider starting with a twisted graph state (i.e.\ the stabilisers are twisted by angles $\{ \theta_v \}$, and we make all measurements in the $X$ basis). If we start with the logical operators of the MBQC resource state, and update these logical operators $\alpha_L$, to $\tilde{\alpha}_L$ such that $[\tilde{\alpha}_L,X_v] = 0$ for all $v$, then if we start in the $+1$ eigenstate of a logical operator $\tilde{\alpha}_L$, after the adiabatic substitution (or measurement) we will still be in the $+1$ eigenstate of $\tilde{\alpha}_L$. This is just the same as simulating the MBQC computation in the Heisenberg picture.
These results are summarised in the following theorem:
\begin{theorem}\label{theor:AGQCTheor}
Any measurement based pattern on a graph $G$ of $N$ qubits which has \emph{gflow} g and depth $d$ can be efficiently converted into an equivalent holonomic computation for which
\begin{itemize}
\item The holonomic computation can be done in $d$ steps, where the minimum energy gap for each step is fixed at $\sqrt{2} \gamma$, and $\Vert \dot{H} \Vert = |L_j|$ for the $j^{th}$ step. Thus the time to perform the $j^{th}$ step is $\Omega( |L_j|^{1 +
\delta})$, where $0 < \delta \leq 1$.
\item The resulting computation is the same.
\item The maximum degree of the initial Hamiltonian $k_{\max}$ is equal to the maximum \emph{gflow} size : $k_{\max} = |g(\nu)|_{max}$,
\item The initial Hamiltonian can be computed efficiently.
\end{itemize}
\end{theorem}
As an example, consider the graph from~\cite{Browne2007} shown in fig.~\ref{fig:gflow}, which has \emph{gflow} $g(a_1) = \{b_1 \}, g(a_2) = \{ b_2 \}, g(a_3) = \{ b_3,b_1\}$. Following the prescription given above, we start with stabilisers $T_{a_1} = K_{b_1}, T_{a_2} = K_{b_2}, T_{a_3} =K_{b_3}K_{b_1}$ which have degree at most 3. The adiabatic substitutions are then performed in the order $a_1 < a_2 < a_3$. Using these stabilisers, Theorem~\ref{theor:AGQCTheor} tells us how to carry out the AGQC in adiabatic time proportional to the depth, which is 3 in this case. Also for adiabatic cluster-state quantum computation on a rectangular graph qubits with $r$ rows and $d$ columns, $g(v) = v + r$, $T_v = K_{v+r}$, $|g(v) | \leq 5$ and there are $d$ layers, with each layer containing $r$ qubits, so the time for each layer scales as $\Omega(|r|^{1 +
\delta})$.
Note that, for all known universal graphs there is a \emph{gflow} for which $|g(v)|$ is bounded, so typically the degree will also be bounded. However this is not necessarily the case for all families of graphs; in some cases it can scale with the number of inputs (an example of this is shown in Section~\ref{sec:Example}). This raises an important question: Is this increase in degree regarded as a free resource, or is there a cost associated with it? The evidence to date would suggest that the latter is true, since typically in nature we only see 2-body interactions, with higher degree interactions resulting as a low energy approximation. We take this approach in our model, and assume that degree is bounded and such high degree Hamiltonians can only result as an approximation to a 2-body Hamiltonian.
A few fundamental results are known about high-degree Hamiltonians. For instance, it is known that that there are no 2-body Hamiltonians for which the ground state is exactly the same as that of a $k$-body Hamiltonian, where $k >2$~\cite{Siu2005,Haselgrove2004,VanDenNest2008}. If we try to simulate this $k$-body interaction with a 2-body interaction, then it was shown in \cite{Siu2005,Haselgrove2004} that the energy gap will scale as $E_{gap} < O(1-F^2)$, where $F= \sprod{\psi}{E_0}$ is the fidelity between the actual ground state $\ket{\psi}$ and the $k$-body ground state $\ket{E_0}$ (i.e.\ the accuracy of the approximation). However, as far as the author is aware, there are no known results which say that there is a \emph{fundamental} reason why many body interactions should be difficult to create. What is known is that, for the currently known methods of simulating large-degree Hamiltonians such as perturbation gadgets~\cite{Kempe2004,Oliveira2008,Bartlett2006,Jordan2008}, the Hamiltonian has an energy gap which shrinks with the degree. For example, we can create k-local Hamiltonians $H_k$ using a perturbative Hamiltonian acting on $rk$ ancilla qubits and n computational qubits (r is the number of terms in Hamiltonian with degree $k$, which we consider as being fixed)~\cite{Jordan2008}. The result is that the effective Hamiltonian $H_{\textrm{eff}}$, apart from some overall energy shift, is
\begin{eqnarray}
H_{\textrm{eff}} = \frac{-k(-\lambda)^k}{(k-1)!} H_{k} \otimes P_{+} + O(\lambda^{k+1}),
\end{eqnarray}
where $P_+$ is a projector on the space of $r$ ancilla qubits, projecting each one into the $\ket{+}$ state, and the perturbation converges provided that $\lambda < \frac{k-1}{4k}$. This energy gap decreases exponentially with $k$, therefore making the reasonable assumption that interactions in nature are limited to interactions between 2 (or a finite number of) bodies, then if the degree is allowed to scale with $N$, this imposes a prohibitive cost in that the minimum energy gap of the system shrinks exponentially, and so therefore the adiabatic time grows exponentially. Thus as much as possible it is best to avoid large degree Hamiltonians in adiabatically driven schemes.
Note that we have implicitly assumed that the energy scale is fixed; if it were not, then since the adiabatic runtime scales as $\Omega( | L_k|^{1+\delta} / \gamma)$ for each layer, then increasing the interaction strengths will compensate for the increase in adiabatic runtime. However, scaling the energy freely is likely to be a difficult task in any realistic system, so we consider the case where the interactions are already set at the strongest they can be.
\subsection{Trade-offs in AGQC}
\label{sec:AGQCexample}
Above we have seen that it is possible to transform any measurement pattern on a graph state with \emph{gflow} into an adiabatic computation. Here we investigate how the trade-off between quantum and classical processing in MBQC manifests itself in AGQC. To begin with, consider the graph shown in Fig.~\ref{fig:ZigZag}. Recall from Section~\ref{sec:FlowGflow} that it is possible to define many different \emph{gflows} $g^r(v)$ on this graph
\begin{eqnarray}
g^r(v) = \left\{
\begin{array}{c c}
\{N+v,...,N+v+r -1 \}, & \mbox{if} \hspace{2mm} v+r-1 \le N \\
\{N+v,...,2N \}, & \mbox{if} \hspace{2mm} v+r-1 >N
\end{array} \right.
\end{eqnarray}
where $1 \le r \le N$. The measurement depth $d_r$ is given by $d_r = \left\lceil \frac{N}{r} \right\rceil$, and the size of the \emph{gflow} is $|g^r(v)| \leq r+2$. Following the the application of Theorem~\ref{theor:AGQCTheor} to this graph, the computation can be performed in $d_r$ adiabatic steps, where the time to perform the $j^{th}$ layer scales as $\Omega( |L_j|^{1+ \delta} )$, and the maximum degree is given by the influencing volume. Thus $g^r$ is converted into an adiabatic computation which has $d^r$ steps, and each step takes $\Omega(r^{1+\delta})$ time, and the Hamiltonian degree $k = r+2$. In the most extreme case, $g^{N}$ is converted into an adiabatic computation which takes 1 step, but this step takes $\Omega(N^{1 + \delta})$ time to complete, and the Hamiltonian degree is proportional to $N$.
Thus by playing with the \emph{gflow} in this way, there is a trade-off in this zig-zag graph between the number of adiabatic steps, the size of $\| \dot{H} \|$ and the degree of the Hamiltonian, in direct analogy to the quantum/classical trade-off seen in MBQC. This shows an interesting contrast between MBQC and AGQC: since classical computation is a cheap, low-error resource it is desirable to shift as much computation into classical processors as possible, so the optimal flow to choose for MBQC would be $g^N$ (also called the maximally-delayed gflow) since it maximises classical computation. However, in AGQC the total computation time is always approximately linearly in $N$, so there is no advantage in using large delayed \emph{gflow}, and given the expense of creating high degree operators discussed above, it even seems prohibitive to use the maximally delayed \emph{gflow}, and so the optimal situation seems to be using $g^1$ (the minimally delayed \emph{gflow}).
In MBQC, it is only possible to perform measurements in one step for examples like the zig-zag graph where the measurements are in the Pauli basis, since for more general measurements the random outcomes of the measurements cannot be corrected at the end. Given this fact, it is natural to ask whether or not the trade-off seen above for the zig-zag graph applies to all graphs, and all computations. It is interesting partly for the sake of rigorousness, but also since if it was possible to find a way to perform any computation in one constant time step at the expense of degree, this would have profound implications, giving a fundamental bound on the energy gap for simulating large degree operators (under the reasonable assumption that all computation cannot be done in constant time). Even if this isn't the result, exploring the particular way in which AGQC stops us being able to perform computations in one step is illuminating in itself.
In the following we will investigate this question whilst keeping the energy gap constant, and by manipulating the terms in the Hamiltonian. As mentioned in the previous subsection, this has the advantage of keeping the energy gap protection and allowing analytical results to be found. The numerical results in~\cite{Bacon} allude to another complementary trade-off, perhaps more intuitive, between the energy gap and the number of steps in the computation. Since it is harder to find exact results, or even numerical results, for systems other than 1-dimensional chains, we concentrate on the method below which allows analytical results.
To perform the computation in one step with constant energy gap, following the logic leading to Theorem~\ref{theor:AGQCTheor} we must find a transformation of the stabilisers $T_v \to \tilde{T}_v$ such that all of the summands in the following Hamiltonian commute:
\begin{align}\label{eqn:AGQCallatonce}
H(s) = -\gamma\sum_{v \in V \setminus O} (1-s)\tilde{T}_v + s X_v
\end{align}
These summands will commute when the $\tilde{T}_v$ stabilisers satisfy the commutation relations
\begin{eqnarray}
[\tilde{T}_v, X_w] = 0\; \mbox{for all} \; v \ne w \; \mbox{in} \; V \setminus O
\end{eqnarray}
If such a transformation is possible, then all of the stabilisers $\tilde{T}_v$ can be replaced at the same time with fixed energy gap $2\gamma \eta$, provided the adiabatic runtime scales as $\Omega(N^{1+\delta})$ (see Appendix~\ref{sec:LinAlg}). For a graph with \emph{gflow}, such a procedure is always possible by multiplying the $\tilde{T}_v$ stabilisers together, since any $Z_v$ or $Y_v$ operator can be transformed to identity or $X_v$ by multiplication by $T_v$, and this is guaranteed to commute with all vertices $w < v$. To update stabiliser $T_v$, we start in layer $L_{n+1}$, where $L_n = L(v)$, and update as follows; (1) If $T_v$ contains a $Z_w$ or $Y_w$ term in layer $L_{n+1}$, multiply by $T_w$. (2) Proceed to the next layer, as determined by the time order. Iterate until the outputs are reached. (3) The final updated stabilisers are denoted $\tilde{T}_v$.
However, creating a Hamiltonian in such a way is polynomially equivalent to simulating the computation (see e.g.~\cite{Markham13}), a fact which has already been used in the discussion of Theorem~\ref{theor:AGQCTheor}. For instance, starting with the logical operator $\bar{X}_1 = X_1 Z_2$ and multiplying this by $\tilde{T}_2$ gives a new logical operator $\bar{X}' = X_1 Z_2 \tilde{T}_2$ which commutes with all $X$ operators at non-outputs. Since it commutes with all $X$-measurements, the eigenstate of $\bar{X}'$ is preserved throughout the adiabatic evolution, so if the system starts in the $+1$ eigenstate of $\bar{X}'$ we end in the $+1$ eigenstate of $\bar{X}'$. Then to solve the computation all that is required is to ignore all the $X$ operators on non-outputs, and find the $+1$ eigenstate of the operators in $\bar{X}$ that act on the outputs.
Or to put it in a different way, we can look at how efficient the update procedure is compared to the efficiency of the computation before the procedure. For Clifford operations the procedure takes polynomial time: if we have $N$ qubits in total, we will have to perform one updating sweep per qubit, each sweep involves a search over at most $N$ stabilisers to see whether they commute or anti-commute, and the cost of testing if a stabiliser commutes or not will be $O(N)$ since each stabiliser will contain at most $N$ terms. So the procedure takes $O(N^2)$ steps. For general angles the procedure takes an exponential amount of time, since at every step a $Z_v$ operator is replaced with a $Z_vT_v = e^{-i \theta_{v+1} Z_{v+1}} X_{v+1} Z_{v+2}$ term, so every update converts a $Z$ term into 3 new terms. If we start off with a stabiliser containing $n$ $Z$ operators, then after $r$ sweeps we will have $O(3^r n)$ operators to search through.
In addition, during this update procedure, we multiply every $Z_w$ operator by $T_v = \prod_{w\in g(v)} K_{w}$. So following this procedure, each $\tilde{T}_v$ operator will have Pauli $X$'s in positions $g(w)$ for every $Z_w$ that has been have corrected for. The only parts that will contribute to the degree of $\tilde{T}_v$ are situated at vertices which can be arrived at by following a path on which a non-gflow line is preceded and followed by a gflow line. Following the discussion in Sec.~\ref{sec:GeneralAGQC} we expect the simulation of these Hamiltonians to be very prohibitive.
So we have seen that the trade-off seen for the zig-zag graph can be extended to more general computations, such that the number of adiabatic steps can be decreased at the expense of increasing $\| \dot{H} \|$ and the degree of the Hamiltonian. In addition, the process of finding the initial Hamiltonian becomes inefficient for non-Clifford operations, since the description of the Hamiltonian is polynomially equivalent to the solution of the computation. Using this method, it is not possible to reduce the overall adiabatic time, only the number of steps. Perhaps surprisingly, there is also no apparent advantage in having free access to unbounded-degree Hamiltonians. Indeed, given the prohibitive cost of the known ways to achieve high-degree operators, our result suggests that if it is desirable to keep the energy gap fixed it is best to perform the computation in as many steps as possible (or in the language of MBQC, it is best to use the minimally-delayed \emph{gflow}). This result highlights the subtlety in using the adiabatic theorem; although the minimum energy gap is the same for graphs of different flow, the size of $\Vert \dot{H} \Vert$ changes and leads to a dependence of the adiabatic time on the number of elements in a \emph{gflow}.
\section{Reordering the computation} \label{sec:Order}
In MBQC, doing measurements in the wrong order can result in random outcomes to the computation. In some cases however, such as when all qubits are measured in the Pauli basis, the correct state can still be recovered by using local Pauli corrections. Given that the non-deterministic measurements are absent in AGQC, it is natural to wonder whether or not the same ordering rules apply, and if they do, what factors collude to make it so. One approach to studying this is using the same method as in the previous sections, by manipulating terms in the Hamiltonian and keeping the energy gap fixed. However, as we have seen previously, whilst this does allow the ordering to be changed, it is generally at the expense of high degree Hamiltonians. In this section we will go back to the original formulation of adiabatic cluster-state quantum computation, and look at what order the adiabatic substitutions can be performed in, with the constraint that we will only use the stabilisers used in~\cite{Bacon2010}. As a physical motivation we may imagine that we have some experimental apparatus which is limited to only applying operators below a certain size, but we can turn them on in any combination. Firstly we will consider the most obvious case, where we replace one stabiliser by one Pauli operator, and then we consider a less constrained method, and discuss how these two approaches lead to different behaviour.
\subsection{Re-ordering the adiabatic computation with fixed number of terms in Hamiltonian}\label{sec:Order2}
We start with an $N$-qubit 1D chain with $N-1$ stabilisers (i.e.\ one qubit is encoded in the chain), and we consider replacing one stabiliser by one Pauli operator in a different order to the order in MBQC. We would like to find orders of replacements which preserve information in the 2-dimensional logical subspace, and we might expect from MBQC that changing the order will in some way disrupt the computation. To begin with, consider a 4-qubit chain with all angles set to zero (i.e.\ with un-twisted stabilisers). We start in the ground state of the initial Hamiltonian $H_0$:
\begin{eqnarray}
H_0 = -T_1 - T_2 - T_3.
\end{eqnarray}
First notice that replacing $T_2 \to X_2$ does not affect the energy gap of this Hamiltonian, since $[X_2 , T_1] = 0$ in this case. Thus we are free to go out-of-order in this way, provided $\theta_2 = n \pi$. However, if instead we start by replacing $T_3$ with $X_3$ out-of-order, the time-dependent Hamiltonian becomes
\begin{eqnarray}
H(s) = -T_1 - T_2 - (1-s)T_3 - sX_3,\; \mbox{ with } 0\le s \le 1.
\end{eqnarray}
At $s=1$, this Hamiltonian has a ground state degeneracy of 4, i.e.\ the degeneracy doubles during the evolution. However, there is a constraint in that the operator $T_1T_3$ commutes with $H(s)$ for all $s$ and so the eigenstate of $T_1T_3$ is conserved throughout the evolution. This means that since we start in the $+1$ eigenspace of $T_1T_3$, we will also end in the $+1$ eigenspace of $T_1T_3$. So although the gap closes, the transitions between these degenerate eigenstates are forbidden and so the subspace is preserved.
Now consider replacing $T_1 \to X_1$. Throughout this evolution, $T_1T_3$ no longer commutes with $H(s)$, and so unless there is a stabiliser that $T_1T_3$ can be multiplied with to make it commute with $H(s)$, the system is no longer constrained to the $+1$ eigenspace of $T_1T_3$ . We can easily check that there are no stabilisers which we could multiply $T_1T_3$ with in order to make it commute, since this requires an operator with a $Z_1$ term, which only $T_1$ contains, but we cannot multiply by $T_1$ since this leaves $T_1 T_1 T_3 = T_3$ which doesn't commute with $H(s)$. Since at the very start of the $T_1 \to X_1$ evolution, the energy gap is zero and the subspace is no longer preserved, this means the system will now leak out of the subspace unless the evolution time $\tau \to \infty$. Thus the computation fails at the last step for this ordering. Note that this happens whether or not we replace $T_2 \to X_2$; the key part was the fact that there were no stabilisers available to multiply $T_1 T_3$ with to make it commute with $H(s)$.
Extending this to larger chains, we can see that whenever $T_n$ is replaced with $X_n$ for $n > 2$, the system will still be constrained to the $+1$ eigenspace of $T_{n-2}T_n$. If this `hidden' stabiliser anticommutes with the next Pauli replacement $X_m$, then provided we can multiply by $T_{m-2}$ the subspace is still preserved. But since at some point it will be necessary to replace $T_{1}$ or $T_2$, and there are no other stabilisers to multiply with to make sure that the `hidden' stabiliser still commutes with $H(s)$, the computation fails as there will be zero energy gap and the subspace is no longer preserved.
Similar behaviour can be seen when performing $Y$ measurements, again using untwisted stabilisers. Consider a Hamiltonian on 3 qubits, $H = T_1 + T_2$. Now after replacing $T_2 \to Y_2$, the system is still confined to a 2-dimensional subspace since $[T_1T_2,H(s)]=0$ and so the subspace where $T_1T_2$ has eigenvalue $+1$ is preserved. However, after replacing $T_{1} \to Y_1$, there are no stabilisers which can be chosen to make $T_1T_2$ commute with $H(s)$, and so the computation fails at the last step. This argument can similarly be extended to larger chains, and just like in the above case we will find that once the boundary is reached the computation time must diverge to avoid losing information.
This argument can easily be extended to more general measurement angles. Consider starting with a Hamiltonian made up of untwisted stabilisers as above, and replacing the $n^{th}$ untwisted stabiliser with $X_n^{\theta_n}$. The interpolating Hamiltonian has the form
\begin{align}\label{eqn:HamReorder}
H(s) = -\sum_{j < n -2} T_j - \left( T_{n-2} + T_{n-1} + (1-s) T_n + sX_n^{\theta_n} \right) - \sum_{k > n} T_k.
\end{align}
The terms inside the two sums commute with the terms in brackets, so we can ignore these and focus on the terms in brackets (see Appendix~\ref{sec:LinAlg}). The eigenvalues at $s =0$ and $s=1$ can be calculated in Mathematica, with the result that the degeneracy at $s=1$ is double the degeneracy at $s=0$ for $n > 2$. Just as above, this degeneracy does not initially mean leakage out of the logical subspace, as there are terms that are conserved (in this case, $T_{n-2}T_n$ and $T_{n-1}T_n$ are conserved quantities). However, at the point where the boundary stabilisers need to be replaced, there will be leakage out of the logical subspace.
This argument works provided $n > 2$, but for $n=2$ it is in fact possible to perform a computation without leakage out of the logical subspace, for certain angles. For $n=2$, the eigenenergies are doubly degenerate for all $s$ and take the form:
\begin{eqnarray}
E_{1} (\theta_{2},s) &=\pm \sqrt{ 2(1 - s + s^2) \pm \Gamma(\theta_{2},s)},
\end{eqnarray}
where $\Gamma (\theta,s) \equiv \sqrt{ 2s^2 \cos{2\theta_{2}} + (4 - 8s + 6s^2)}$. The energy gap $\Delta_1$ is given by
\begin{eqnarray}
\Delta_1 (\theta_{2},s) &= \sqrt{ 2(1 - s + s^2) + \Gamma(\theta_{2},s)} - \sqrt{ 2(1 - s + s^2) - \Gamma(\theta_{2},s)}.
\end{eqnarray}
For a given $\theta_{2}$, this reaches a minimum at $s = (1- \frac{1}{2}\cos \theta_{2})$. For all $\theta_{2}$ then, the minimum energy gap $\Delta_1^{min}$ is given by:
\begin{eqnarray}
\Delta_1^{min}(\theta_{2})= \Delta_1 \left(\theta_{2} ,1- \frac{1}{2} \cos \theta_{2}\right).
\end{eqnarray}
Over all possible $\theta_{2}$ values, $\Delta_1^{min}(\theta_{2})$ is largest for $\theta_{2} = l\pi, s = 1/2$, and goes to zero when $\theta_{2}= (2l+1)\pi/2, s=1$, where $l$ is an integer.
So we see that, for angles $\theta_2 \neq \frac{(2n+1)\pi}{2}$, it is possible to perform computations out of order, in a very limited sense.
While the information remains in a protected subspace, it is also necessary to check if the information is transformed in the expected way. To see this, we start with the logical operators
\begin{eqnarray}
X_L \equiv X_1 Z_2, \; Z_L \equiv Z_1.
\end{eqnarray}
($Y_L$ isn't included since $Y_L = iX_LZ_L$). Using the method in~\cite{Bacon2010}, these can be multiplied by stabilisers to make them commute with the `measurements'. The transformations that would be applied to the logical operators above would be
\begin{eqnarray}
&X_L \to X_L T_2 = X_1 X_3^{\theta_3}Z_4 \nonumber\\
&Z_L \to Z_LT_1 = e^{-i\theta_2 Z_2} X_2 Z_3 \to e^{-i\theta_2 Z_2 T_2} X_2 Z_3 = e^{-i\theta_2 X_3^{\theta_3}Z_4} X_2 Z_3
\end{eqnarray}
Note that this is exactly the same transformation we would make in the normal computation; we have just done two steps at once. Both these logical operators are in a form which commute with $X_1$ and $X_2$, so they commute with the time-dependent Hamiltonian, and so, for example, if the system starts in the $+1$ eigenstate of $X_L$, it will end up in the $+1$ eigenstate of $X_3^{\theta_3}Z_4$.
In summary, we have seen that when replacing $T_n$ operators with $X_n$ operators one-by-one on a 1D chain, the orderings that preserve the ground subspace are limited, and reordering is only possible using measurement angles which are not odd multiples of $\frac{\pi}{2}$. In particular, in a 1D chain, replacing the stabiliser at the first site or the second site appears to be the only other allowed ordering. The information stored in the chain is transformed in the same way in both cases, however the latter only works for $\theta_2 \ne (2n+1) \frac{\pi}{2}$, and the energy gap depends on $\theta_2$ so the speed of the adiabatic substitution must vary.
\subsection{Re-ordering without a fixed number of terms in Hamiltonian}
So far we have seen that the initial approach to re-ordering the adiabatic substitutions works only for a limited case. This is perhaps expected, as the out-of-order measurements considered so far are not a proper reflection of what happens in MBQC; in MBQC measuring a qubit destroys any entanglement on edges connected to that qubit. This is clearly not true above, since the final Hamiltonians contain terms like $T_1 + T_2 + X_3$. So a more natural way to perform the out-of-order measurements would be to remove all entanglement to measured qubits, or just remove all anticommuting terms entirely~\cite{Fitzsimons}. Take the chain considered above:
\begin{eqnarray}
H_0 = -T_1 - T_2 - T_3
\end{eqnarray}
Now if we replace $T_3 \to X_3$, we also remove any other operators which anticommute with $X_3$. The result is
\begin{eqnarray}\label{eqn:ReOrder1}
H_1 = - T_2 - X_3
\end{eqnarray}
Notice that the operator $T_1 T_3$ is still conserved, just like before, but this time when we want to measure $X_1$, there is no problem, since the system is always constrained to be within a 2-dimensional subspace. This would also work if we had instead just removed all entanglement to site $3$ (except $T_2$ can be left untouched since it commutes with $X_3$), i.e.
\begin{eqnarray}\label{eqn:ReOrder2}
H_1 = -Z_1 X_2 - T_2 - X_3
\end{eqnarray}
The minimum energy gap will also be the same when we introduce $X_1$ in equation (\ref{eqn:ReOrder1}) to when we replace $Z_1X_2\to X_1$ in equation (\ref{eqn:ReOrder2}).
Similar results hold for $Y$ measurements or the \textsc{cnot} gate, so that Clifford operations can be performed in any order. Note that more than one operator is replaced at the same time, the number of anticommuting terms increases, and based on the numerical studies in~\cite{Bacon}, if the number of anticommuting terms scales with $n$ we would expect the energy gap to be polynomial in $1/n$.
In summary, we have seen that Clifford operations can be done in any order, provided we either change the stabilisers so that the entanglement to `measured' sites is destroyed, or remove any stabilisers which anticommute with the measurement operator entirely. This is in contrast to the first method in Sec.~\ref{sec:Order2}, in which the stabilisers did not reflect what happens in MBQC, and in which the re-ordering is limited to performing alternating measurements from left to right. In both of the approaches considered, we cannot perform Clifford operations in one step, since the energy gap decreases inversely with the number of simultaneous measurements made.
\section{Conclusions} \label{sec:conc}
Using the tools of MBQC, such as \emph{flow} and \emph{gflow}, we have shown that the method of implementing measurements on a cluster state in an adiabatic, continuous fashion~\cite{Bacon} can be generalised to graph states with \emph{gflow}, resulting in adiabatic graph-state quantum computation (AGQC). Specifically, any measurement pattern with \emph{gflow} can be converted into an adiabatic evolution such that the number of adiabatic steps is equal to the depth in MBQC, and each step takes time proportional to the size of each layer.
This work opens up the possibility of future results about efficiency and trade-offs in MBQC being transferred to holonomic computation. For example, there is still little understood about how to view the efficient simulatability of Clifford gates from the perspective of \emph{gflow}. The framework developed here offers a natural route to translate future possible results in this area, which may have implications for the efficiency of AGQC as well as the simulation of many-body Hamiltonians.
We have also seen that, in analogy to the trade-off between quantum and classical time in MBQC, there is a trade-off between the number of adiabatic steps taken and the size of $\Vert \dot{H} \Vert$, together with the degree of the initial Hamiltonian. The adiabatic time is proportional to $\Vert \dot{H} \Vert$, so this means a trade-off between the number of steps and the step time. One interesting point is that the trade-off is only between the number of steps and $\Vert \dot{H} \Vert$, but does not involve the energy gap, which highlights the subtleties of using the adiabatic theorem.
This suggest that, in this model of computation, freely available high degree operators are not a useful resource, which is perhaps counter-intuitive as we might expect that access to unbounded degree operators would give some kind of advantage. Conversely, under the assumption that simulating high degree operators shrinks the energy gap (which is the case for all known methods of simulating k-body operators, e.g.~\cite{Jordan2008,Bravyi2008}), this means that the optimal way of performing AGQC is in fact with as many steps as possible (the minimally-delayed \emph{gflow}), in contrast to in MBQC.
We considered the influence of adiabatic measurements in a different order from the one corresponding to the MBQC pattern. When the original stabilisers were replaced with the same local operators as before, just in a different order, the information generally leaked out of the logical subspace and the computation failed with the exception of a few, special cases. When however the continuous measurement is performed in such a way that the terms that anticommute with the measurement, e.g.\ $X_3$, are all adiabatically removed, then all Clifford operations can be performed in any order.
Finally we note that these results do not cover all possible methods of performing a measurement-based quantum computation, as there are some graph states without \emph{gflow} which still yield deterministic computations, and more general states such as those investigated in~\cite{Gross2007}. Understanding these could lead to further insights into holonomic quantum computation.
\section{Notation and conventions used}
\chapter{Eigenvalues of commuting matrices}
\label{sec:LinAlg}
Here we prove a useful result for Chapter~\ref{chap:AGQC}, on the eigenvalues of commuting operators. We begin by presenting the proof found in Michael Nielsen's blog\footnote{\url{http://michaelnielsen.org/blog/archive/notes/fermions_and_jordan_wigner.pdf}}, and which was used in~\cite{Antonio2014} to calculate the minimum energy gap. However, there is a flaw in our interpretation of this proof (thanks to Daniel Burgarth for pointing this out). To amend this, we present another proof of the form of the minimum energy gap for the Hamiltonians considered in~\ref{chap:AGQC}, but the first misinterpreted proof is included since it may be of interest to other researchers in the field.
The spectral decomposition theorem states that normal matrices (of which Hermitian matrices are a special case) can be diagonalised, and therefore a Hermitian matrix $H$ can be written as
\begin{eqnarray}
H = \sum_j E_j P_j,
\end{eqnarray}
where $E_j$ are the eigenvalues of $H$ (real numbers if $H$ is Hermitian), and $P_j \neq \idop $ are operators projecting into the subspace with energy $E_j$, satisfying
\begin{eqnarray}\label{eqn:OrthBasis}
P_jP_k = \delta_{jk} P_j, \; \sum_j P_j = \idop, \; P_j^2 = P_j.
\end{eqnarray}
The projectors $P_j$ are said to form a complete orthonormal basis if they satisfy these properties.
If two normal matrices $H_1$ and $H_2$ commute, then they can be diagonalised in the same basis, since if $\{ |E^{(1)}_n \rangle \}$ is a basis in which $H_1$ is diagonal, then
\begin{eqnarray}
H_2H_1 | E^{(1)}_n \rangle = E_n^{(1)} \left( H_2 | E^{(1)}_n \rangle \right)= H_1 \left( H_2 | E^{(1)}_n \rangle \right).
\end{eqnarray}
So $H_2 | E^{(1)}_n \rangle = E_n^{(2)} | E^{(1)}_n \rangle$, and so $H_j$ and $H_k$ have common eigenstates. Therefore the eigenvalues of $H_1 + H_2$ will be $E_n^{(1)} + E_n^{(2)}$ for certain combinations of $m$ and $n$; which combinations will depend on the structure of the matrices involved.
Now consider a set of commuting normal matrices $\{ H_j\}$. We can spectrally decompose each one of these matrices:
\begin{eqnarray}
H_j = \sum_k E_{jk} P_{jk}
\end{eqnarray}
Since all of the $H_j$ commute, $[P_{jk},P_{lm}] = 0$ for any $j,k,l,m$. This follows since if two commuting matrices $A$ and $B$ have spectral decompositions $A = \sum_k a_{k} P_{k}^{(A)}$ and $B = \sum_k b_{k} P_{k}^{(B)}$, then the projectors can be written $P_{k}^{(A)} = poly(A)$, $P_{k}^{(B)} = poly(B)$. This can be seen since $A^n = \sum_k a^n_k P_k^{(A)} = M \mathbf{P}$ where $M_{mn} := a^m_n$ and $\mathbf{P} : = (P_1^{(A)},P_2^{(A)},...,P_D^{(A)})$ where $D$ is the dimension of $A$. The determinant of $M$ is then related to the determinant of a Vandermonde matrix (assuming that none of the $a_k$ are 0, which can be made true by excluding any projectors for which this is the case), which is non-zero, and so $M$ can be inverted so that $\mathbf{P} = M^{-1} \mathbf{A}$, where $\mathbf{A} := ( A,A^2,...,A^D)$. In other words, $P_k^{(A)} = poly(A)$ (this part of proof courtesy of Daniel Burgarth).
This means we can construct a complete orthonormal basis $P_{\vec{k}}$ where
\begin{eqnarray}\label{eqn:Pk}
P_{\vec{k}} = P_{1k_1} P_{2 k_2} ... P_{mk_m}
\end{eqnarray}
To prove it is a complete orthonormal basis, we must show that it satisfies the same properties in eqn.(\ref{eqn:OrthBasis}). Using the fact that $[P_{jk},P_{lm}] = 0$,
\begin{align}
P_{\vec{k}} P_{\vec{k'}} & = \left( P_{1k_1} P_{2 k_2} ... P_{mk_m} \right) \left(P_{1k'_1} P_{2 k'_2} ... P_{mk'_m} \right) = \left( P_{1k_1} P_{1 k'_1} \right)\left( P_{2k_2} P_{2 k'_2} \right)...\left( P_{mk_m} P_{m k'_m} \right) \nonumber\\
&=\left( P_{1k_1} \delta_{k_1 k'_1} \right)\left( P_{2k_2} \delta_{k_2 k'_2} \right)...\left( P_{mk_m} \delta_{k_m k'_m} \right) = P_{\vec{k}} \delta_{\vec{k} \vec{k'}}. \\
\sum_{\vec{k}} P_{\vec{k}} &= \sum_{k_1} \sum_{k_2}...\sum_{k_m} P_{\vec{k}} = \left( \sum_{k_1} P_{1k_1} \right)\left( \sum_{k_2} P_{2k_2} \right)...\left( \sum_{k_m} P_{mk_m} \right) = \idop.
\end{align}
So the $P_{\vec{k}}$ form an orthonormal basis, and the eigenvalues of $\sum_j H_j$ are formed from adding the eigenvalues of each of the $H_j$ together. Since each $P_{\vec{k}}$ projects into a different combination of eigenvalues of the individual $H_j$ matrices, we (incorrectly) interpreted this as equivalent to proving that the eigenvalues of $\sum_jH_j$ of a set of commuting matrices $\{H_1,H_2,...,H_m\}$ are $\{ \sum_{\vec{k}} E^{(1)}_{k_1} + E^{(2)}_{k_2} + ... + E^{(m)}_{k_m} \}$, which contains all possible combinations of the individual eigenvalues of the $H_j$. Clearly this cannot be true since, if the vector space has dimension $d$, this implies there are $d^m$ eigenvalues. The resolution to this is to note that if any adjacent pair of the projectors in $P_{\vec{k}} $ are orthogonal, then $P_{\vec{k}} $ becomes 0. Thus to find the eigenvalues of a sum of commuting Hermitian matrices $H = H_1 + H_2 + ... H_n$, we need to find the combinations of projectors in $P_{\vec{k}} $ such that no two projectors are orthogonal.
We now present a proof that, for Hamiltonians of the form in (\ref{eqn:HamSplit}), the eigenvalues are given by all combinations of the eigenvalues of the commuting terms. This proof is based on the explanation given in~\cite{PreskillNotes} for why stabilisers each halve the number of encoded qubits in a stabiliser code. Say we have a set of Hermitian operators $\mathcal{M} = \{ M_n \}$ which have eigenvalues $\{ \pm \lambda_n\}$ and eigenvectors $\{ \ket{\pm \lambda_n}\}$. Consider that we also have a set of unitary operators $\mathcal{Q} = \{ Q_n \} $ such that $\{ Q_n , M_n \} = 0$, $[Q_n,M_m]_{n \ne m} = 0$ for all $n,m$. Then $Q_n\ket{+\lambda_n}$ is an eigenstate of $M_n$ with eigenvalue $-\lambda_n$. Since $Q_n$ is a unitary operator, it is a one-to-one mapping, so that there must be an equal number of eigenstates of $M_n$ with eigenvalue $+\lambda_n$ or $-\lambda_n$ (i.e.\ they have the same degeneracy).
Now extending this, consider a state $\ket{\psi}$ which is a simultaneous eigenstate of all the $M_n$ operators. Then
\begin{eqnarray}
\left( \sum_{n=1}^{| \mathcal{M} |} M_n \right) Q_m \ket{\psi} =\left( \sum_{n=1}^{| \mathcal{M} |} \lambda_n(-1)^{\delta_{nm}} \right) Q_m \ket{\psi}.
\end{eqnarray}
So $Q_m \ket{\psi}$ is an eigenstate of $\left( \sum_{n=1}^{|\mathcal{M}|} M_n \right)$, with eigenvalue $\sum_{n=1}^{|\mathcal{M}|} \lambda_n(-1)^{\delta_{nm}} $, and with the same degeneracy as $\ket{\psi}$. Therefore if there exists a unitary operator $Q_n$ for every operator $M_n$ such that $\{ Q_n , M_n \} = 0$, $[Q_n,M_m]_{n \ne m} = 0$ for all $m$, and if each $M_n$ has eigenvalues $\pm \lambda_{n}$, the spectrum is $\pm \lambda_{1} \pm \lambda_{2} \pm... \pm\lambda_{N}$. If the dimension of the space on which these operators act is $D$, then the degeneracy of each of these eigenvalues will be $\frac{D}{ 2^{|\mathcal{M}|}}$.
We now apply this argument to the Hamiltonian in (\ref{eqn:HamSplit}):
\begin{align}
H_{L_k}(s) = - \gamma \left( \sum_{v < L_k}X_v + \sum_{v>L_k}T_v \right) - \gamma \left(\sum_{v\in L_k} s X_v + (1-s)T_v \right).
\end{align}
where the operators $\{ T_v \}$ are stabilisers of the form in (\ref{eqn:GflowStab}) for a graph with \emph{gflow}. Each of the individual summands has two eigenvalues of equal magnitude and opposite sign, so the analysis above can be applied. For the summands in the $\sum\nolimits_{v < L_k} X_v$ term, operators that satisfy the conditions above are $Q_v = Y_v$. For any summand in the $\sum\nolimits_{v > L_k} T_v$ term, we can select $Q_v = X_v^{\theta_v}$ (chosen so that it will commute with any $T_w$ such that $v \in g(w)$). Finally, for any of the terms in the $\sum_{v\in L_k} [(1-s) T_v + s X_v]$ sum, we can select $Q_v = Y_v$. This follows since $Y_v$ anticommutes with $[(1-s) T_v + s X_v]$ and commutes with $[(1-s) T_u + s X_u]_{u \neq v}$, since the stabiliser $T_u$ has an identity operator acting on all vertices in the same layer as $u$ (from the rules of \emph{gflow}, there is an even connection between $g(u)$ and $v$ if $u$ and $v$ are in the same layer. Thus if $u$ and $v$ are in the same layer and $u \neq v$, then $[(1-s) T_u + s X_u , Y_v] = 0$). Similarly $[Y_v,\sum\nolimits_{u > L_k} T_u]=0$ since each $T_u$ operator has identity acting on vertex $v$, and $[Y_v,\sum\nolimits_{u < L_k} X_u] = 0$.
Thus, given that it is possible to find a set of $\{ Q_n \}$ operators for each of the individual commuting summands $M_n$ in $H_{L_k}(s)$ that satisfy the commutation rules $\{ Q_n , M_n \} = 0$, $[Q_n,M_m]_{n \ne m} = 0$, the eigenvalues of $H_{L_k}(s)$ are all combinations of the eigenvalues of the commuting parts. The eigenvalues of the individual $[(1-s) T_u + s X_u]$ terms are $\pm \sqrt{ (1-s)^2 +s^2 } := \pm \eta$, so the eigenvalues of $H_{L_k}(s)$ are $-|L_k|\gamma\eta,-(|L_k|-2)\gamma\eta,...,(|L_k|-2)\gamma\eta,|L_k|\gamma\eta$ plus some integer multiples of $\pm \gamma$. The total number of summands in $H_{L_k}$ is equal to the number of qubits encoded into the graph $N_q$, which means that the degeneracies of each eigenvalue are the same and equal to $\frac{2^N}{2^{N-N_q}} = 2^{N_q}$.
Similar reasoning can be applied to the time derivative $\dot{H}_{L_k}(s)$, with the result that the eigenvalues of $\dot{H}_{L_k}(s)$ are $\{ -|L_k|\gamma,-(|L_k|-2)\gamma,...,(|L_k|-2)\gamma,|L_k|\gamma \}$. The spectral norm of $\dot{H}_{L_k}(s)$ is therefore $\| \dot{H}_{L_k}(s) \| = |L_k| \gamma$.
A similar reasoning can also be applied to the Hamiltonian in eqn.~\ref{eqn:AGQCallatonce}:
\begin{align}
H(s) = -\gamma\sum_{v \in V \setminus O} (1-s)\tilde{T}_v + s X_v
\end{align}
where $[\tilde{T}_v, X_w] = 0\; \mbox{for all} \; v \ne w \; \mbox{in} \; V \setminus O$. In this case, for a summand $ (1-s)\tilde{T}_v + s X_v$ we can choose $Q_v = Y_v$. Since we are not just replacing stabilisers in the same layer as in the case above, other stabilisers may either have an identity operator or $X$ operator located at site $v$, and so there may be other stabilisers $\tilde{T}_u$ that anticommute with $Y_v$. This can be remedied by multiplying $Q_v$ by $X_u$ for any stabilisers $\tilde{T}_u$ that anticommute with $Y_v$, since $[(1-s)\tilde{T}_u + s X_u,Y_vX_u] = 0$ if $\{ T_u,Y_v\} = 0$. The eigenvalues of each summand will still be $\pm \eta$, which we can see since squaring a summand gives $\gamma^2[(1-s)\tilde{T}_v + s X_v]^2 = \gamma^2[(1-s)^2 + s^2]\idop = \gamma^2 \eta^2 \idop$. Thus the eigenvalues of the Hamiltonian in eqn.~\ref{eqn:AGQCallatonce} are multiples of $\pm\gamma \eta$, and the energy gap is $2 \gamma \eta$. Following similar reasoning, the maximum eigenvalue of $\| \dot{H} \|$ scales with $N \gamma$.
Finally we show how to arrive at the energy gap of the Hamiltonian given in equation (\ref{eqn:HamReorder}). The Hamiltonian has the form:
\begin{align}
H(s) = -\sum_{j < n -2} T_j - \left( T_{n-2} + T_{n-1} + (1-s) T_n + sX_n^{\theta_n} \right) - \sum_{k > n} T_k.
\end{align}
Using the identity $\textsc{cz}_{m,n}^\dagger X_n Z_m \textsc{cz}_{m,n} = X_n$, then conjugating with \textsc{cz} operators acting between sites $(n-2,n-1)$ and $(n+1,n+2)$, gives
\begin{align}
H'(s) &= \textsc{cz}_{n-2,n-1}^\dagger\textsc{cz}_{n+1,n+2}^\dagger H(s) \textsc{cz}_{n+1,n+2}\textsc{cz}_{n-2,n-1} \nonumber\\
&= -\left( \sum_{j < n -3} T_j + Z_{n-3}X_{n-2} \right) - \left( X_{n-1} Z_{n} + T_{n-1} + (1-s) Z_n X_{n+1} + sX_n^{\theta_n} \right) \nonumber\\
&- \left( X_{n+2} Z_{n+3} + \sum_{k > n} T_k \right).
\end{align}
$H'(s)$ has the same spectrum as $H(s)$, but is in a particularly useful form, as it is a \emph{Kronecker sum}. Given two square matrices $A$ and $B$ acting on vector spaces of dimension $d_A$ and $d_B$ respectively, we define a Kronecker sum as $A \oplus B : = A \otimes \idop_{d_B} + \idop_{d_A} \otimes B$. Thus $H'(s)$ can be written:
\begin{align}\label{eqn:KronSum}
-H'(s) &= \left( \sum_{j < n -3} T_j + Z_{n-3}X_{n-2} \right) \oplus \left( X_{n-1} Z_{n} + T_{n-1} + (1-s) Z_n X_{n+1} + sX_n^{\theta_n} \right) \nonumber \\
&\oplus \left( X_{n+2} Z_{n+3} + \sum_{k > n} T_k \right).
\end{align}
A particularly useful property of Kronecker sums is that the eigenvalues of $A \oplus B$ are all combinations of the eigenvalues of $A$ and $B$~\cite{HornJohnson2}. Thus the eigenstates of $H'(s)$ are all combinations of the eigenvalues of the individual terms in the Kronecker sum. The terms on the left and right in (\ref{eqn:KronSum}) have integer eigenvalues, whilst the middle term has energy gap $\leq 1$, and so the energy gap is given by the energy gap of the middle term.
\chapter{Interaction picture and the rotating wave approximation}
\label{app:IntPic}
In quantum mechanics, typically we consider states $\ket{\psi(t)}$ that evolve according to the Schr\"{o}dinger equation $i\hbar | \dot{\psi}(t) \rangle= H\ket{\psi(t)}$, so that the state vector carries the information of how the system changes, whilst operations are performed. This picture of quantum mechanics is usually called the Schr\"{o}dinger picture. For some applications, it is often more informative to use different ways of depicting the evolution other than the Schr\"{o}dinger picture. One alternative is the \emph{Heisenberg picture}, in which states are stationary but the operators carrying the time dependence (see e.g.\ Chapter~\ref{chap:AGQC} for an example of using the Heisenberg picture), for which we can express the Schr\"{o}dinger equation as an equation involving operators $O(t)$ rather than states:
\begin{align}
i\hbar \frac{d}{dt} O(t) = H(t) O(t)
\end{align}
In cases where we have a Hamiltonian $H(t) = H_0(t) + V(t)$ which can be split up into a part which is easily solvable $H_0$, and an additional part $V$, it is often convenient to use the \emph{intermediate} or \emph{Dirac} picture to follow the dynamics, which is effectively halfway between the Schr\"{o}dinger and Heisenberg pictures. To transform into the intermediate picture, we represent operators and states relative to the evolution due to $H_0(t)$, which can be done by operating with $A(t) := e^{-iH_0t/\hbar}$. States in this picture then become $\ket{\psi_I(t)} := A^{\dagger}(t) \ket{\psi(t)}$, and operators are $O_I(t) := A^{\dagger}(t) O(t) A(t)$. If we insert these into the time-dependent Schr\"{o}dinger equation, we see that:
\begin{align}
\frac{d}{dt} \ket{\psi_I(t)} &= \frac{d}{dt} [ A(t) \ket{\psi_I(t)}] = \frac{d}{dt} [ e^{iH_0t/\hbar} \ket{\psi_I(t)}] \nonumber\\
&=\frac{i}{\hbar}H_0 \ket{\psi_I(t)} + e^{iH_0t/\hbar} \left[ \frac{1}{i\hbar} H(t) \ket{\psi(t)} \right]= - \frac{ie^{iH_0t/\hbar} }{\hbar} V \ket{\psi(t)}=- \frac{i}{\hbar} V_I \ket{\psi_I(t)}
\end{align}
So that
\begin{align}
i\hbar \frac{d}{dt} \ket{\psi_I(t)} = V_I(t) \ket{\psi_I(t)}
\end{align}
which can be a useful way to frame a particular problem.
A useful tool when using the interaction picture is the rotating wave approximation, in which terms in the interaction picture with a high frequency time dependence can be ignored. To see this, start with the Dyson series for a time-dependent Hamiltonian $H_I$ (see e.g.~\cite{DalessandroBook})
\begin{align}
U(t,0) & = \mathcal{T} e^{-\frac{i}{\hbar} \int_0^t H_I(t_1) dt_1} =\idop - \frac{i}{\hbar}\int_0^t H_I(t_1) dt_1 + \frac{1}{2\hbar^2}\int_0^t \int_0^{t} \mathcal{T}[H_I(t_1) H_I(t_2)] dt_1 dt_2 + ...
\end{align}
where $\mathcal{T}$ is the time-ordering operator, $\mathcal{T}[x(t_1)x(t_2)] = \Theta(t_1-t_2)x(t_1) x(t_2) + \Theta(t_2 - t_1) x(t_2)x(t_1)$, and $\Theta(t)$ is the Heaviside step function. Consider a Hamiltonian of the form $H_I = V \cos(\omega t)$ where $V$ is time-independent and Hermitian. If it is the case that $\omega$ is a high frequency, this means that $V \cos(\omega t)$ is a fast oscillating term which may have negligible effect on the dynamics. Using the identity $\int_0^t \cos (\omega t') dt' = \frac{\sin (\omega t')}{ \omega} $, the evolution operator can be written
\begin{align}
U(t,0) &= \idop - \frac{i}{\hbar}\int_0^t V \cos(\omega t_1) dt_1 + \frac{1}{2\hbar^2}\int_0^t \int_0^{t} \mathcal{T}[V \cos(\omega t_1) V \cos(\omega t_2)] dt_1 dt_2 + ... \nonumber \\
& = \idop + O \left( \frac{\Vert V \Vert}{\hbar \omega} \right)
\end{align}
provided that $\Vert V \Vert < \hbar\omega$. This means that, whenever we have terms in a Hamiltonian which have a high frequency time dependence, these can be ignored with an error of order $\Vert V \Vert / \hbar \omega$.
We now derive some identities that will be useful in the next subsections:
Consider a Hamiltonian term of the form $C\Omega e^{-i\omega Z t} \cos (\omega_x t) X$. Setting $\omega_x = \omega$ gives
\begin{align}\label{eqn:Iden1}
& C e^{-i\omega Z t} \cos (\omega t) X =\frac{C}{2} e^{-i\omega Z t} (e^{i\omega t} + e^{-i\omega t})X \nonumber\\
&=\frac{C}{2}(e^{i\omega t} + e^{-i\omega t})(e^{-i\omega t } \sigma^+ + e^{i\omega t } \sigma^- )= \frac{C}{2}\left( \sigma^+ + e^{ 2i\omega t } \sigma^- + e^{-2i\omega t} \sigma^+ + \sigma^- \right)\nonumber\\
& = \frac{C}{2} X + O \left( \frac{C}{\hbar\omega} \right)
\end{align}
provided that $C < \hbar \omega$. Strictly speaking the last line is incorrect: the Hamiltonian itself does not have terms of order $C / \hbar \omega$, but the evolution due to the Hamiltonian does. We will use this type of notation in Chapter~\ref{chap:3qubit}.
For the second identity, consider the following Hamiltonian:
\begin{align}\label{eqn:Iden2}
&e^{-it [\omega_1 Z_1 + \omega_2 Z_2]} J ( X_1X_2 + Y_1 Y_2 + Z_1 Z_2)e^{it [\omega_1 Z_1 + \omega_2 Z_2]} =2J e^{-2 it [\omega_1 Z_1 + \omega_2 Z_2]} (\sigma^+_1 \sigma^-_2 + \sigma^-_1 \sigma^+_2) + JZ_1 Z_2 \nonumber\\
&=2J\sigma^+_1 \sigma^-_2 e^{2it[\omega_1 - \omega_2]} + 2J\sigma^-_1 \sigma^+_2 e^{-2it[\omega_1 - \omega_2]} + JZ_1 Z_2= JZ_1 Z_2 + O \left( \frac{2J}{\hbar( \omega_1 - \omega_2)} \right)
\end{align}
provided $2 J < \hbar(\omega_1 - \omega_2)$.
\chapter{Density matrix renormalisation group}
\label{sec:DMRG}
The Hilbert space of a spin chain grows exponentially with $N$, and so for even a modest number of spins, exactly diagonalising a Hamiltonian can take a very long time. Fortunately, several tools have been developed to approximate the behaviour of a system, by ignoring the states in a system which have the smallest effect. This is the basis of the density matrix renormalisation group (DMRG) approach~\cite{White92}.
\subsubsection{Time-independent DMRG}
This algorithm can be split into two parts: the `infinite' algorithm and the `finite' algorithm, which we outline here for a simple Hamiltonian with nearest-neighbour interactions. The overall process is that we build up the ground state of a chain, starting off with two sites, and sequentially adding on 2 sites at a time, choosing a truncated basis in which to represent this ground state (if necessary). Say we start with a chain which we divide into two `blocks', left and right, which both have dimension M, and which have individual Hamiltonians $H_L$ and $H_R$ (i.e.\ for the moment we assume the two halves of the chain are non-interacting)
\begin{equation}
|\psi \rangle = \sum_{L,R} A_{L,R} | L \rangle | R \rangle.
\end{equation}
Now we add sites of dimension d to the left and the right sides to get
\begin{eqnarray}
|\psi \rangle &=& \sum_{L,R,l,r} A_{L,R,l,r} | L \rangle | l \rangle | r \rangle |R \rangle := \sum_{i,j} B_{i,j} | i \rangle |j \rangle,
\end{eqnarray}
so that the left and right sides now have dimensions $Md$, and we have a new `extended' Hamiltonian for both sides, $H_{eL} $ and $H_{eR}$, and an overall Hamiltonian $H_S$:
\begin{eqnarray}
H_{eL} &=& H_{L} \otimes I_d + J \vec{\sigma}_{N_L} \cdot \vec{\sigma_l} \nonumber\\
H_{eR} &=& I_d \otimes H_{R} + J \vec{\sigma}_r \cdot \vec{\sigma}_{N_R} \nonumber\\
H_S &=& H_{L} \otimes I_{Md} + I_{Md} \otimes H_{R} + I_M \otimes (J \vec{\sigma}_l \cdot \vec{\sigma}_r) \otimes I_M,
\end{eqnarray}
where $\vec{\sigma}_{N_L}$ and $\vec{\sigma}_{N_R}$ indicate, respectively, the Pauli spin operator of the right or leftmost site of L or R (expressed in the basis of L or R), and $\vec{\sigma}_{l}$, $\vec{\sigma}_{r}$ are the Pauli spin operators acting on the added sites. We now ask: what is the best way to express the states $|i \rangle$ and $| j \rangle$ in a basis of dimension M? The answer is that if we find the ground state $| \psi_G \rangle$ of $H_S$, and then find the partial trace of the this wavefunction for the right and left side
\begin{eqnarray}
\rho_{eL} = tr_R(|\psi_G \rangle \langle \psi_G |), \; \rho_{eR} = tr_L(|\psi_G \rangle \langle \psi_G |),
\end{eqnarray}
and keep only the eigenstates corresponding to the M largest eigenvalues, then this will give the state with least error \cite{SchollwockThesis}. Then after we have discarded the lowest eigenvectors, we have a reduced basis in which to express the right and left sides of the chain, and we transform all the Hamiltonians and operators relevant to our calculations into this basis.To do this we construct operators $T_L$ and $T_R$ made up of the first m eigenvectors of $\rho_{eL}$ and $\rho_{eR}$, and express the extended Hamiltonians in terms of this basis, to form new block Hamiltonians, $H_{L}'$ and $H_{R}'$, which are truncated into the $M \times M$ Hilbert space.
\begin{eqnarray}
H_{L}' = T_L^{\dag} H_{eL} T_L , \; H_{R}' = T_R^{\dag} H_{eR} T_R.
\end{eqnarray}
We are now back to where we started, with a left and right block, and we can repeat the algorithm, adding sites onto these blocks, and repeating until the desired system size is reached, but avoiding the exponential growth in Hilbert space (and therefore the growth in computing resources needed). By truncating the Hilbert space we will introduce an error $\epsilon$, which is just given by the sum of eigenvalues of the eigenstates we have chosen to exclude
\begin{equation}
\epsilon = \sum_{j \not \in m} \lambda_j = 1 - \sum_{j \in m} \lambda_j,
\end{equation}
where the $\lambda_j$ are the eigenvalues of $\rho_{eL}$ or $\rho_{eR}$. This is the end of the `infinite' DMRG algorithm; however to achieve a really accurate ground state we need to use a second algorithm, called the `finite' DMRG algorithm, where the chain length is kept fixed. The first part is the same as for the infinite case. We find $\rho_{eL}$ and find $H_{L}'$ in terms of the eigenstates of $\rho_{eL}$, but we leave $H_{R}$ unchanged. In the next step we use $H_{L}'$ as our new left block Hamiltonian, which means we have increased the size of the left part by one site, and so we use a right block Hamiltonian which is smaller by one site (for this reason we have to keep in memory all of the transformation matrices $T_L$, $T_R$ that we found in the infinite algorithm). Then the chain length is unchanged, but we have advanced towards the right by one site (but we have only made changes to the left hand side), and we can repeat the process again until we reach the right hand end of the chain
\begin{equation}
|\psi_{End} \rangle = \sum_{L,l,r} A_{L,l,r} | L \rangle | l \rangle | r \rangle.
\end{equation}
The process is reversed so that the right blocks are updated, and so on back and forth until there is some convergence (e.g. convergence of the ground state energy).
\subsubsection{DMRG with time evolution}
The next step is to evolve the system in time. For systems with nearest-neighbour interactions, we would normally use the Suzuki-Trotter time-step method \cite{Vidal2003} \cite{White2004}, whereby the evolution operator can be expressed as
\begin{eqnarray}
e^{-i H \tau} \approx e^{-i H_{12} \tau/2} e^{-i H_{23} \tau/2} ... e^{-i H_{23} \tau/2} e^{-i H_{12} \tau/2} + \mathcal{O}(\tau^3),
\end{eqnarray}
where $H_{ij}$ is the Hamiltonian acting between sites i and j only, and we operate on each pair up to $(N-1,N)$ and back. However, since we are using next-nearest neighbour interactions, this pairwise method is not suitable here, and instead we use the time-step targeting method, introduced in \cite{Luo2003} and developed further in \cite{Feiguin2005}, which is generally less accurate than using the Suzuki-Trotter time-step method but more suitable for our calculations.
To do this, we start at one end of the chain (i.e. after the finite algorithm has reached one end). Then we sweep back to the other end, updating the basis of the block+site at each step as before, but this time updating them with respect to some target states \cite{Feiguin2005}, which are given by the fourth order Runge-Kutta method, and will have some dependence on time. To perform this we construct a set of four vectors:
\begin{eqnarray}
|k_1 \rangle &=& -i \tau \tilde{H}(t) | \psi (t) \rangle \nonumber\\
|k_2 \rangle &=& -i \tau \tilde{H}(t+\tau/2) [| \psi (t) \rangle + \frac{1}{2} | k_1 \rangle ] \nonumber\\
|k_3 \rangle &=& -i \tau \tilde{H}(t+\tau/2) [| \psi (t) \rangle + \frac{1}{2} | k_2 \rangle] \nonumber\\
|k_4 \rangle &=& -i \tau \tilde{H}(t+\tau) [| \psi (t) \rangle + | k_3 \rangle],
\end{eqnarray}
Where $\tilde{H}(t) = H(t) - E_0$, and in our case the Hamiltonian is time invariant. We then find expressions for the state at times $(t+\frac{\tau}{3})$, $(t+\frac{2\tau}{3})$ and $(t+\tau)$
\begin{eqnarray}
|\psi(t+\frac{\tau}{3}) \rangle &\approx& |\psi(t) \rangle + \frac{1}{162}[ 31|k_1 \rangle + 14|k_2\rangle + 14|k_3\rangle -5|k_4 \rangle] \nonumber \\
|\psi(t+\frac{2\tau}{3}) \rangle &\approx& |\psi(t) \rangle + \frac{1}{81}[ 16 |k_1 \rangle + 20 |k_2\rangle + 20|k_3\rangle -2 |k_4 \rangle] \nonumber \\
|\psi(t + \tau) \rangle &\approx& |\psi(t) \rangle + \frac{1}{6}[ |k_1 \rangle + 2|k_2\rangle + 2|k_3\rangle + |k_4 \rangle].
\end{eqnarray}
We then construct a reduced density matrix using a weighted sum over the states at these four times e.g. for the left block:
\begin{equation}
\rho_L = \sum_{n=0}^3 w_n tr_R(| \psi(t + n \tau / 3) \rangle \langle \psi(t + n \tau / 3) |),
\end{equation}
where the optimal weightings were determined in \cite{Feiguin2005} to be $w_0=1/3, w_1=1/6, w_2=1/6, w_3=1/3$. And as before we can diagonalise this and discard a certain number of states to find a new more efficient basis in which to do our time evolution (without doing this kind of basis transformation, we would quickly run into errors \cite{GarciaRipoll2006}). Then we repeat this process along the chain, updating the basis at each step (and importantly, updating any operators, Hamiltonians, or transformation matrices that we may need to use later on) but without advancing in time. Then at the end of this half-sweep, when we have updated each basis and thus found the optimal basis in which to perform our time evolution, we evolve the state by a single time step (and since the computational time required to do this last part is small, we can use a more accurate procedure than that above, so instead we use 10 Runge-Kutta iterations each with time step $\tau / 10$). Then we repeat the process, making sweeps along the chain, advancing in time only at the end of each half sweep, until we have evolved the state for the required length of time.
\chapter{Conclusions}
In this thesis, we have introduced novel results aimed at simplifying and exploring the possibilities of quantum computation. A scheme to implement two-qubit quantum gates on decoherence-free qubits has been demonstrated, giving high fidelities above known error thresholds even with reasonable levels of noise. This scheme makes use of many-body interactions that are predicted to occur in certain systems, and requires far fewer pulses than existing gates, so could be useful in implementations where there is limited control, or where control errors are an issue. Further work could be done to explore encoded qubits on more than 4 qubits, and include effects due to Aharanov-Bohm phases from magnetic fields, either as a way to tune the interactions or as another source of error.
Three schemes to realise 3-qubit gates on arrays of qubits have also been proposed, using Hamiltonians that are time-independent during the gate operation, and using interactions that are as uniform as possible. The viability of implementing these using ion traps or bismuth donors in silicon has also been explored, finding that a fault-tolerant Toffoli gate could be realised using bismuth donors with today's technology, and it is hoped that experimental realisation of these gates could happen in the near future. These gates could be useful as primitives for quantum computation, or could prove useful as classical gates, making use of the many advances in quantum technologies before a quantum computer is within our reach, potentially allowing smaller or faster classical gates.More could be done to explore how much dissipation we would expect in a practical implementation of these gates including e.g.\ the effects of heating from laser fields, and it would also be interesting to explore the possibility of creating larger time-independent automated structures that could perform more than just a single gate.
An extension of the adiabatic cluster-state model of computation has also been presented, using the tools of measurement-based quantum computation (MBQC) to transform adaptive measurements into adiabatically driven holonomic transformations. This translation has allowed a study of how properties of MBQC are translated into a holonomic setting, showing interesting trade-offs between properties of the Hamiltonian and the number of adiabatic steps. The allowed ordering of operations in this model was explored, finding that that for some approaches the allowed ordering is very limited compared to MBQC. It is hoped that this work opens up new channels of communication between the adiabatic and measurement-based schools of thought, providing a fruitful comparison and sharing of concepts. This line of thought could also possibly be extended to other MBQC patterns for which the tools used here are not valid. It would also be useful to explore practical implementations of this computational method, to see whether or not there is a significant practical advantage in using adiabatic processes rather than measurements.
The practicality of using a Wigner crystal as a quantum communication channel was explored using the semiclassical instanton approximation, yielding average fidelities of up to 0.92 with nine or ten electrons, with a transfer time significantly below the typical decoherence time of electrons in GaAs quantum dots. Whilst the parameters we use are for electrons in GaAs, the results are quite general and are applicable to any system in which a Wigner crystal can be created. We hope that this study inspires future technological applications using Wigner crystals, and hope that it might be developed into an experimentally realisable scheme in the near future. There are many parameters and additional effects such as noise, magnetic fields and image charges that we have not included in this analysis, so future work will include these effects and look for better ways of transferring information. There is also scope to go further, to use a Wigner crystal to perform gates between qubits, or even to use a Wigner crystal to perform universal quantum computation.
Finally, methods of communicating using more idealised spin chains were also studied, extending a previous study to include next-nearest-neighbour interactions, finding an interesting although as yet not fully explained peak close to the Majumdar-Ghosh point. This demonstrates that previous spin chain communication protocols are robust to next-neighbour interactions, and provides interesting questions for future work. Novel schemes to transfer information using edge-locking and ground state transfer effects were also briefly studied. These schemes lay the foundation for further analysis exploring how these effects can best be put to use.
\chapter{Creating gates for decoherence-free qubits}
\label{chap:DFS}
In Section~\ref{sec:DFSintro} we introduced the concept of decoherence-free subspaces and subsystems as ways of encoding qubits to reduce the detrimental effects of the environment on the information. The possible ways of encoding qubits to protect against strong and weak collective noise were also discussed, revealing that the 4-qubit strong DF subspace and a 3-qubit strong subsystem are the smallest encodings. Whilst these protected qubits could be useful as a quantum memory, it is desirable to perform universal quantum computation using them. To achieve this, it must be possible to perform single qubit rotations and interactions between pairs of qubits. Performing single qubit rotations in the 4-qubit DF subspace and 3-qubit DF subsystem can be realised using exchange interactions between pairs of physical qubits in the encoded qubit (see e.g.~\cite{BaconThesis,Hsieh2003}), however creating two-qubit gates is harder; the methods found so far to perform two-qubit gates involve either interactions that could be difficult to create in an experiment~\cite{BaconThesis,Bacon2000}, or involve large numbers of gates to be switched on and off sequentially (e.g.\ 22 gates in 13 time steps for the 3-qubit DF subsystem in~\cite{Fong2011}), or use perturbative/complicated control sequences to create these gates~\cite{Jiang2009}. As an example of why we might want to simplify the existing two-qubit gates in the 4-qubit DF subspace, consider one of the interactions used in~\cite{BaconThesis,Bacon2000} to create a gate:
\begin{align}
H = 3 {E}_{12} + \frac{2}{3}({E}_{24} + {E}_{23} + {E}_{34}).
\end{align}
Given the asymmetry in these interactions, this could be a challenging gate to realise (note that although this only acts within one encoded qubit, this Hamiltonian is turned on whilst we are out of the logical subspace, and so is not simply a local unitary operation). The main other alternative to realising this gate is shown in Fig.~\ref{fig:ExchGate}; here many 2-body exchange interactions are used. Whilst this gate requires very simple interactions, it also requires many steps, so is perhaps susceptible to control errors, and could perhaps be simplified by using less constrained interactions.
\begin{figure}[h]
\[
\Qcircuit @C=1em @R = 0.6em @!R {
\lstick{A_1}& \qw & \qw & \qw & \qw & \qw & \qw & \qw & \qw & \qw & \qw & \qw & \qw & \qw & \qw\\
\lstick{A_1}& \qw & \qw & \qw & \qw & \multigate{1}{1} & \qw & \multigate{1}{-\frac{1}{2}} & \qw & \multigate{1}{1} & \qw & \qw & \qw& \qw & \qw\\
\lstick{A_1}& \qw & \multigate{1}{\frac{1}{2}} & \qw & \multigate{1}{-\frac{1}{2}} & \ghost{1} & \multigate{1}{-\frac{1}{2}} & \ghost{-\frac{1}{2}} & \multigate{1}{-\frac{1}{2}} & \ghost{1} & \multigate{1}{-\frac{1}{2}} & \qw &\multigate{1}{\frac{1}{2}}& \qw & \qw\\
\lstick{B_1}& \multigate{1}{p_1} & \ghost{\frac{1}{2}} & \multigate{1}{1} & \ghost{-\frac{1}{2}} & \multigate{1}{-\frac{1}{2}} & \ghost{-\frac{1}{2}} & \multigate{1}{\frac{1}{2}} & \ghost{-\frac{1}{2}} & \multigate{1}{-\frac{1}{2}} & \ghost{-\frac{1}{2}} &\multigate{1}{1} & \ghost{\frac{1}{2}} & \multigate{1}{-p_1} & \qw\\
\lstick{B_1}& \ghost{p_1} & \multigate{1}{p_2} & \ghost{1} & \multigate{1}{-\frac{1}{2}} & \ghost{-\frac{1}{2}} & \multigate{1}{1} & \ghost{\frac{1}{2}} & \multigate{1}{1} & \ghost{-\frac{1}{2}} & \multigate{1}{-\frac{1}{2}} & \ghost{1} & \multigate{1}{1-p_2}& \ghost{-p_1} & \qw\\
\lstick{B_1}& \qw & \ghost{p_2} & \qw & \ghost{-\frac{1}{2}} & \qw & \ghost{1} & \qw & \ghost{1} & \qw & \ghost{-\frac{1}{2}} &\qw & \ghost{1-p_2}& \qw & \qw\\
}
\]
\caption{\label{fig:ExchGate} The sequence of exchange gates used in~\cite{Fong2011} to realise a 2-qubit gate between 3-qubit DF subsystem qubits. The number in each box represents the time for which the exchange gate is applied.}
\end{figure}
In this chapter we outline alternative ways to perform 2-qubit gates in the 3-qubit DF subsystem and 4-qubit DF subspace, using less operations that require minimal control, so that realisation of decoherence-free qubits might be more attainable in an experiment (motivated by the recent experimental advances in realising quadruple quantum dots in a square configuration~\cite{Thalineau2012}).
With practicality in mind, we also only consider physical qubits arranged in a regular formation; three spins in an equilateral triangle for the 3-qubit DF subsystem, and four spins in a square for the 4-qubit DF subspace. The interactions we consider for both encodings are interactions between the middle 4 spins (see Fig.~\ref{fig:Plaquette} for an illustration of this). This work is a more detailed account of the work published in~\cite{AntonioPRA2013}.
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.35]{QubitArray-eps-converted-to.pdf}
\caption{An illustration of the two geometries of qubits we consider when constructing logical qubits. Filled circles represent physical qubits, and solid lines illustrate the kind of interactions we consider when looking for a gate in this chapter. The top diagram shows the layout for the 4-qubit encoding and the bottom diagram shows the layout for the 3-qubit encoding. }
\label{fig:Plaquette}
\end{center}
\end{figure}
\subsection{Ring Exchange Interactions} \label{sec:Rexchange}
When constructing a two-qubit DF gate, we will include `ring exchange' interactions. These interactions, which are used to explain excitations in La$_2$CuO$_4$~\cite{Katanin02,Coldea01} and become important in electrons forming a Wigner crystal~\cite{Bernu2001,Voelker2001}, appear as corrections in the exchange Hamiltonian due to higher order hopping processes between different physical qubits in the extended Hubbard Hamiltonian~\cite{Takahashi1977}. They have also been investigated in the context of quantum computing~\cite{Scarola2005,Mizel2004,Mizel2004B}, and it is clear from these works that ring exchange processes should not be ignored. We will not derive the full ring exchange Hamiltonian here, but will present the main results. Consider electrons confined to quadratic potentials; when a single band tight-binding model is appropriate we can describe the physics of the system using an extended Hubbard model with fermions
\begin{align}
H = -\sum_{ \substack{ i < j,\\ \sigma,\sigma' \in \{\uparrow, \downarrow\} }} t_{ij} c^\dagger_{i,\sigma}c_{j,\sigma'}^{} + U \sum_{i} n_{i,\uparrow}n_{i,\downarrow}^{} + V \sum_{i; \sigma, \sigma' \in \{ \uparrow , \downarrow \} }n^\dagger_{i,\sigma} n_{j,\sigma'},
\end{align}
where $c^\dagger_{i,\sigma}$ ($c_{i,\sigma}$) is the creation (annihilation) operator for an electron with spin $\sigma$ at site $i$, $t_{ij}$ is the tunnelling coefficient between sites $i$ and $j$, $U$ is the on-site Coulomb interaction, and $V$ is the inter-site repulsion. In the regime where the tunnelling is weak compared to the on-site repulsion, and where the system is half-filled (i.e.\ one electron per site), the Hamiltonian for four spin-1/2 particles located at sites 1,2,3 and 4 is~\cite{Takahashi1977}:
\begin{align}\label{Rexchange1}
{H}&= \sum\nolimits_{i < j} J_{ij} {E}_{ij} +C_{1234} [ {E}_{12} {E}_{34} + {E}_{14} {E}_{23} - {E}_{13} {E}_{24}] \nonumber \\
&+ C_{1324} [ {E}_{13} {E}_{24} + {E}_{14} {E}_{23} - {E}_{12} {E}_{34}]+C_{1342} [ {E}_{13} {E}_{24} + {E}_{12} {E}_{34} - {E}_{14} {E}_{23}] +O \left(\frac{t^5}{U^4} \right),
\end{align}
where $t = \max_{i,j}(t_{ij})$, $E_{ij}$ is the exchange interaction defined in Sec.~\ref{sec:CircuitModel}, and
\begin{align}
J_{ij} = \frac{t_{ij}^2}{U} - \frac{4 t_{ij}^4}{U^3} + \sum_k \frac{t_{ik}^2 t_{kj}^2}{U^3} + \sum_{ \substack{ k\neq l,i,j \\ l \neq i,j \\ i \neq j}} \frac{t_{ik}t_{kl}t_{lj}t_{ji}}{U^3};\text{ }C_{ijkl} =\frac{t_{ik}t_{kl}t_{lj}t_{ji}}{U^3}.
\end{align}
From the form of $C_{ijkl}$ we can see that ring exchange terms will appear when there are exchange terms present which form a loop (i.e.\ when a physical qubit is indirectly coupled to itself). In~\cite{Scarola2005}, it was also found that the presence of magnetic fields changes the coefficients $C_{ijkl}$ and introduces three-body terms, with couplings that depend on the magnetic flux passing through 3 or 4-site loops. Here we assume that these magnetic fields are low enough so that the magnetic flux has negligible impact and three body terms can be ignored, and we leave the effects of non-zero magnetic fields to later work.
With four qubits arranged on a square, it is possible to make all of the exchange couplings uniform, and since this will simplify things considerably, whenever we need to include ring-exchange terms in this chapter we will assume that this is the case. Taking uniform interactions in eqn.\ (\ref{Rexchange1}) results in a `symmetric' version of the Hamiltonian, ${H}_S$, which takes the same form as derived in~\cite{Mizel2004}:
\begin{align}\label{Rexchange2}
{H}_S &= J_{\Box}{H}_{\Box} + J_{\times} {H}_{\times} + J_{\circlearrowleft} {H}_{\circlearrowleft} = J ({H}_{\Box} + {H}_{\times}) + J_{\circlearrowleft} {H}_{\circlearrowleft} =J \left( {H}_{\Box} + {H}_{\times} + \alpha {H}_{\circlearrowleft} \right) \nonumber\\
&= J \left( \sum_{n=1}^4 {E}_{n,n+1} + \sum_{n=1}^2 {E}_{n,n+2} + \alpha [{E}_{12} {E}_{34} + {E}_{14} {E}_{23} + {E}_{13} {E}_{24}] \right),
\end{align}
where ${H}_{\Box}$, ${H}_{\times}$, ${H}_{\circlearrowleft}$ represent the nearest-neighbour, next-nearest-neighbour and ring-exchange Hamiltonians, respectively, and $J_{\Box}$, $J_{\times}$, $J_{\circlearrowleft}$ are the corresponding interaction strengths for these Hamiltonians. Since we have equal coupling between all sites, $J_{\Box} = J_{\times} = J$ in this equation, and in the final line we have defined $\alpha := J_{\circlearrowleft} /J_{\Box} \equiv J_{\circlearrowleft} /J$ as the ratio of ring exchange terms to the nearest/next-nearest neighbour terms. Restricting ourselves to symmetric Hamiltonians of this form simplifies things considerably, since ${H}_{\circlearrowleft}$ commutes with ${H}_{\Box}$ and ${H}_{\times}$, and as argued in~\cite{Mizel2004}, this form of Hamiltonian contains enough degrees of freedom to fix all of the eigenvalues, and so we do not need to take into account any higher order terms. Since single qubit rotations of the encoded qubits given in (\ref{eqn:3states}) and (\ref{eqn:4States}) can be performed by exchange interactions between physical qubits, a two-qubit interaction on these encoded qubits is equivalent to a 4-body interaction of the form $E_{ij}E_{kl}$, and since the ring exchange interactions contain terms similar in nature to these, it seems plausible that we can use these to simplify the 2-qubit gates acting on the encoded qubits. Investigating whether or not the presence of ring-exchange interactions can lead to simpler gates on the encoded qubits forms part of the motivation for this work.
\section{Two-qubit gates}\label{sec:Two}
We now look at creating a controlled-${Z}$ gate ($\textsc{cz}$) between two encoded qubits, which, along with certain single qubit gates (e.g.\ Hadamard and $U_z(\pi/4)$ gates), enables us to perform universal quantum computation \cite{Barenco95}. An important point to note is that, in this case if we can find a two-qubit gate which works for the 3-qubit DF subsystem, this may also be a valid gate for the 4-qubit DF subspace. To see this, first observe that we can rewrite the states in (\ref{eqn:4States}) in this form:
\begin{align}\label{eqn:App1}
\ket{\bar{0}^{(4)}} = \frac{1}{\sqrt{2}} \left[ |\bar{0}^{(3)}_{+1}\rangle\ket{1} - |\bar{0}^{(3)}_{-1}\rangle\ket{0} \right], \;\ket{\bar{1}^{(4)}} = \frac{1}{\sqrt{2}} \left[ |\bar{1}^{(3)}_{+1}\rangle\ket{1} -|\bar{1}^{(3)}_{-1}\rangle \ket{0} \right].
\end{align}
A valid pulse sequence for the 3-qubit DF subsystem will perform a gate which is (locally equivalent to) a $\textsc{cz}$ gate on the two 3-qubit states up to a gauge transformation. In general, this gauge transformation may mean that this pulse sequence does not work for the 4-qubit DF subspace (it may not amount to a simple local rotation in the 4-qubit case). However, in certain situations, a valid pulse sequence for the 3-qubit DF subsystem will work for the 4-qubit DF subspace as well. For example, since exchange interactions commute with $\mathbf{S}^2$ and $\mathbf{S}_z$ over all the qubits, any interactions made up of exchange coupling only (such as the gate found in~\cite{Fong2011}, shown in Fig.~\ref{fig:ExchGate}) will not couple different gauge states. So pulse sequences for the 3-qubit DF subsystem which are made up of exchange interactions enable us to perform a $\textsc{cz}$ gate on the 4-qubit DF subspace as well.
The converse is also not necessarily true; a valid pulse sequence on the 4-qubit DF subspace will not necessarily work on the 3-qubit DF subsystem, simply because in the 4-qubit case, interactions can be over 8 physical qubits rather than 6. However, if all interactions are confined to 3 physical qubits on each logical qubit (which is the case in our protocol) then a valid pulse sequence for the 4-qubit DF subspace is also a valid gate on the 3-qubit DF subsystem. This can be seen since any gate locally equivalent to a $\textsc{cz}$ gate is also locally equivalent to the interaction $\bar{Z}_A \bar{Z}_B$ acting between two logical qubits $A$ and $B$, where $\bar{Z}_k$ is the logical $Z$ operator acting on logical qubit $k$. Since the logical $Z$ operator for the 4-qubit encoding is the same as for the 3-qubit encoding (an exchange interaction between two qubits, e.g.\ $-E_{1,2}$ using the above definitions of the states), then this gate is also locally equivalent to a $\textsc{cz}$ gate for the 3-qubit encoding. This simplifies the search considerably, as only one pulse sequence needs to be found for both of these encodings.
Performing two qubit gates is, predictably, not as straightforward as performing single qubit gates, mainly because the system can explore a much larger Hilbert space as soon as interactions between the two qubits are turned on. Since $[ \mathbf{S}^2,{E}_{ij}] = 0$ (provided $i$,$j$ are both qubits that $\mathbf{S}^2$ operates on), if two encoded qubits are initialised in the logical subspace with a total spin number $S$, then when they are interacted together using exchange interactions the system will still have overall spin number $S$, so we can use this property to restrict calculations to a small region of the full Hilbert space, to speed up calculations. We also constrain ourselves to couplings which are realistically possible; we only couple sites which can realistically be placed near each other (by e.g.\ placing two encoded qubits side-by-side, see Fig.~\ref{fig:Plaquette}). The method used to search for a quantum gate, once we had chosen a certain set of interactions to use, uses the invariant quantities found by Makhlin~\cite{Makhlin2002} (see Sec.~\ref{sec:DistMeas}).
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.3]{AllHams-eps-converted-to.pdf}
\caption{An illustration of the interactions used to construct the gates, excluding ring exchange interactions which are difficult to represent in this form. Filled circles represent physical qubits, and solid lines represent exchange interactions.}
\label{fig:AllHams}
\end{center}
\end{figure}
The Hamiltonians that we used to construct a gate with are the following:
\begin{align} \label{HamSeq}
{H}_{\asymp} &:= {E}_{12} + {E}_{34}, \text{ } {H}_{\parallel} := {E}_{14} + {E}_{23}, \text{ }{H}_{\times} := {E}_{13} + {E}_{24}, \hspace{2mm}\nonumber\\
{H}_{\Box}& := {H}_{\asymp} + {H}_{\parallel}, \text{ } {H}_{\circlearrowleft} :={E}_{12} {E}_{34} + {E}_{14} {E}_{23} + {E}_{13} {E}_{24}.
\end{align}
The convention for numbering the qubits is shown in Fig.~\ref{fig:AllHams}, along with an illustration of these Hamiltonians (excluding $H_{\circlearrowleft}$ which is hard to represent in pictorial form). For each Hamiltonian in this set we define a corresponding unitary operator:
\begin{align}
{U}_{n} &= \exp \left(-i \frac{J_n {H}_{n} \tau_n}{ \hbar} \right) \equiv \exp \left(-i {H}_{n} \theta_n \right)\text{, e.g. } {U}_{\asymp} = \exp(-i {H}_{\asymp} \theta_{\asymp}), \; \mbox{etc.} \nonumber
\end{align}
where $\theta_n := J_n \tau_n / \hbar$. Note that all of the $\theta_n$ are phases, so we are free to add multiples of $2 \pi$. However, since $\theta_{n} = J_n t_{n}/\hbar$, the trade-off is that increasing $\theta_{n}$ corresponds to either increasing $J_n$ or increasing $t_{_n}$, and under the assumption that we are using the strongest couplings possible, this really means increasing the gate time by $2\pi \hbar / J$ every time we add a multiple of $2 \pi$. This is not ideal, but at least provides some more flexibility. the full gate operator is then arrived at by multiplying all of these unitaries together:
\begin{align}
{U}_{tot} = \prod_n {U}_n = \prod_n \exp(-i {H}_{n} \theta_{n}).
\end{align}
Since the method described above requires a $4 \times 4$ matrix to compare with the ideal $\textsc{cz}$ gate, the unitary evolution $U_{tot}$ must first be projected into the logical subspaces of the 3- or 4-qubit encodings, so $U_{tot} \to U'_{tot} = P_j^{\dagger} U_{tot}P_j$, where $P_j$ is a projector into the logical subspace $\mathcal{L}_j$ of the two encoded qubits, with the subscript $j$ indicating whether it is the 3-qubit encoding ($j=3$) or the 4-qubit encoding ($j=4$). We define these projection operators as
\begin{align}
P_4 = \sum_{x,y \in \{0,1\}} |\bar{x}^{(4)} \rangle | \bar{y}^{(4)} \rangle \langle \bar{x}^{(4)}|\langle \bar{y}^{(4)}|,\text{ }P_3 = \sum_{i,j = \pm 1 } \sum_{x,y \in \{0,1\}} |\bar{x}^{(3)}_{i} \rangle | \bar{y}^{(3)}_{j} \rangle \langle \bar{x}^{(3)}_{i}| \langle \bar{y}^{(3)}_{j}|.
\end{align}
Note that when combining two 3-qubit states with $S = \frac{1}{2}$ together, there are four possibilities; three $S=1$ states and one $S=0$ states. Each of these states is 4-fold degenerate, due to the gauge choices, so $P_3$ projects into a 16-dimensional subspace overall. We can define the projectors onto each of these 4-dimensional subspaces as
\begin{align}
&P_3^{(1,1)} = \sum_{x,y \in \{0,1\} } |\bar{x}^{(3)}_{+1}\rangle |\bar{y}^{(3)}_{+1} \rangle \langle \bar{x}^{(3)}_{+1} | \langle \bar{y}^{(3)}_{+1}|,\; P_3^{(1,0)} = \frac{1}{2}\sum_{i,j= \pm 1}\sum_{x,y \in \{0,1\} } | \bar{x}^{(3)}_{i}\rangle |\bar{y}^{(3)}_{-i}\rangle \langle \bar{x}^{(3)}_{j}| \langle \bar{y}^{(3)}_{-j}| \nonumber\\
&P_3^{(1,-1)} := \sum_{x,y \in \{0,1\}} |\bar{x}^{(3)}_{-1} \rangle | \bar{y}^{(3)}_{-1} \rangle \langle \bar{x}^{(3)}_{-1} | \langle \bar{y}^{(3)}_{-1}| \nonumber\\
&P_3^{(0 ,0)} = \frac{1}{2} \sum_{i,j = \pm1}\sum_{x,y \in \{0,1\} } (-1)^{\frac{i+j+2}{2}} | \bar{x}^{(3)}_{i} \rangle | \bar{y}^{(3)}_{-i} \rangle \langle \bar{x}^{(3)}_{j}| \langle \bar{y}^{(3)}_{-j} |.
\end{align}
where $P_3^{(S,m_S)}$ projects into the subspace with quantum numbers $S,m_S$. In order to find the Makhlin invariant we must project into a 4-dimensional subspace; for the 3 -qubit encoding, we could project onto each of the $S=1$ subspaces and the $S=0$ subspace individually, and verify that the gate works in each subspace. However, this is not necessary, since $[\sum_{i=1}^N \sigma_i^\pm,E_{nm}] = 0$ if $n,m \in \{ 1,2,...,N\}$, which means that if we have a $\textsc{cz}$ gate which works in one of the $S=1$ subspaces, then this gate also works in any of the other two $S=1$ subspaces. So if we can find a gate which is locally equivalent to a $\textsc{cz}$ in any one of the $S=1$ subspaces and also the $S=0$ subspace, it will be locally equivalent to a $\textsc{cz}$ gate when acting on the overall 3-qubit DF subsystem up to some gauge transformation.
As shown above, any parameters which work for the 4-qubit encoding will also work for the 3-qubit encoding (since we use exchange interactions restricted to the middle four physical qubits), so rather than searching for two separate parameters for the 4- and 3-qubit encodings, we can just search for gates which work for the 4-qubit encoding, which simplifies this process.
So we take the projected unitary $ U'_{tot} = P_4^{\dagger} U_{tot}P_4$, find the corresponding Makhlin invariants $m_1({U}'_{tot}),m_2({U}'_{tot})$, and compare these to the Makhlin invariants $m_1(\textsc{cz}),m_2(\textsc{cz})$ of an ideal $\textsc{cz}$ gate using the function $f_m(\textsc{cz},U_{tot})$ defined in Sec.~\ref{sec:DistMeas}, which gives a measure of how close $U_{tot}'$ is to a $\textsc{cz}$ gate, excluding local unitary rotations. Minimising over $f_m$ gives possible ways to implement a gate (or will give evidence that it isn't possible for the considered type of interactions). The local operations required to transform our result to a $\textsc{cz}$ gate will be easy to find compared to the difficulty of minimising over $f_m$, and so we focus on finding sequences which minimise $f_m$ without searching for the local operations. We also need to consider how far out of the logical subspace the system is. One way to do this is to take a state inside the logical subspace $\mathcal{L}$, evolve it with the unitary $U_{tot}$, and then find what the probability of being in the logical subspace is. This can be written $\sum\nolimits_{\ket{m} \in \mathcal{L}} |\bra{m} U_{tot} \ket{n}|^2$ for some input state $\ket{n} \in \mathcal{L}$. Since we are interested in how the gate behaves for all possible input states, this can be turned into an average probability $\bar{p}$ of staying in the logical space by summing over the inputs $\ket{n}$ and dividing by the size of $\mathcal{L}$
\begin{align}
\bar{p} := \frac{1}{|\mathcal{L}|}\sum\nolimits_{\ket{m},\ket{n} \in \mathcal{L}} |\bra{m} U_{tot} \ket{n}|^2 = \frac{1}{|\mathcal{L}|} \Vert P_\mathcal{L} U_{tot} P_\mathcal{L} \Vert_F^2,
\end{align}
where $\| U \|_F$ is the Frobenius norm of $U$ defined as $\| A \|_F = \sqrt{\sum_{i=1}^m \sum_{i=1}^n |A_{ij}|^2}$ for an $m \times n$ matrix $A$, and $P_\mathcal{L}$ projects into $\mathcal{L}$. Then we define the leakage parameter as $L = 1-\bar{p}$, giving for the 3 and 4-qubit encodings respectively:
\begin{align}
\label{eqn:Leakage}
L_3 := 1 - \frac{1}{16}\| P_3 {U}'_{tot} P_3 \|_F^2,\text{ } L_4 := 1 - \frac{1}{4}\| P_4 {U}'_{tot} P_4 \|_F^2.
\end{align}
Note the division by 16 and 4 since we are considering systems of two encoded qubits. Using these measures, we searched for suitable parameters to make a \textsc{cz} gate using a genetic algorithm (see e.g.~\cite{GeneticBook}), followed by a Nelder-Mead simplex search~\cite{NelderMead} once the search had been narrowed down to a sufficient level. The same method has been used in~\cite{DiVincenzo2000} and~\cite{Hsieh2003}.
Finally we note an identity which makes this search over parameters easier to make. As noted in Section~\ref{sec:Rexchange}, we consider ${H}_{\circlearrowleft}$ such that it commutes with ${H}_{\times}$ and ${H}_{\Box}$ (and ${H}_{\times}$ and ${H}_{\Box}$ also commute), which allows us to rearrange ${U}_{\times} {U}_{\Box} {U}_{\circlearrowleft}$ as:
\begin{align}\label{eqn:DFSIden1}
{U}_{\times} {U}_{\Box} {U}_{\circlearrowleft} &= e^{-i {H}_{\times} \theta_{\times}} e^{-i {H}_{\Box} \theta_{\Box}} e^{-i {H}_{\circlearrowleft} \theta_{\circlearrowleft}} = e^{-i ({H}_{\times} \theta_{\times} + {H}_{\Box} \theta_{\Box} + {H}_{\circlearrowleft} \theta_{\circlearrowleft})} \nonumber \\
&= e^{-i ({H}_{\times} (\theta_{\times} - \theta_{\Box}))} e^{(-i {H}_{\times}\theta_{\Box} + {H}_{\Box}\theta_{\Box} + {H}_{\circlearrowleft}\theta_{\circlearrowleft} )} = e^{-i ({H}_{\times} (\theta_{\times} - \theta_{\Box}))} e^{-i J\tau_{\Box}( {H}_{\times} + {H}_{\Box} + (J_{\circlearrowleft}/J) {H}_{\circlearrowleft} )} \nonumber\\
&= e^{-i ({H}_{\times} (\theta_{\times} - \theta_{\Box}))} e^{-i \theta_{\Box} ( {H}_{\times} + {H}_{\Box} + \alpha {H}_{\circlearrowleft} )} = {U}_{\times}' {U}_S,
\end{align}
where
\begin{align}
{U}_S = e^{-i[{H}_{\Box} + {H}_{\times} + \alpha {H}_{\circlearrowleft} ]} \equiv e^{-i {H}_S \theta_S},\text{ }{U}_{\times}' = e^{-i ({H}_{\times} (\theta_{\times} - \theta_{\Box}))} =e^{-i {H}_{\times} \theta_{\times}'} .
\end{align}
In the above we set $\tau_{\Box} = \tau_{\circlearrowleft}$ since $H_{\Box}$ and $H_{\circlearrowleft}$ operate at the same time, and recall that $\alpha = J_{\circlearrowleft}/J $ and $J_{\Box} = J_{\times} = J$ in order to use the symmetric form of the ring exchange interaction, $H_{\circlearrowleft}$. Then using this identity, if there are within the gate sequential applications of the Hamiltonians ${H}_{\times}$, ${H}_{\Box}$ and ${H}_{\circlearrowleft}$ with parameters $\theta_{\Box}$, $\theta_{\times}$, $\theta_{\circlearrowleft}$ respectively, these can be expressed as the Hamiltonian in eqn.\ (\ref{Rexchange2}) plus an additional next-nearest-neighbour term ${U}_{\times}' $ with parameter $\theta_{\times}'$ out in front (if $\theta_{\times}'$ is negative we can just add $2 \pi$ since it is just a phase). Also note that, since we have set $t_{\Box} = t_{\circlearrowleft}$, then $\alpha = \theta_{\circlearrowleft} / \theta_{\Box}$. This identity means that for the purposes of the minimisation we can treat ${H}_{\times}$, ${H}_{\Box}$ and ${H}_{\circlearrowleft}$ as if they were separate interactions with independent parameters and then combine them together at the end using the identity in eqn.\ (\ref{eqn:DFSIden1}), and when we do combine them the value of $\alpha$ is set by $\theta_{\circlearrowleft} / \theta_{\Box}$.
\subsection{Results}
Starting with the simplest Hamiltonian which might result with ring exchange interactions, we tried many combinations of the Hamiltonians given above, performing a genetic search for each one followed by a Nelder-Mead search (typically with around 200 iterations of genetic search, then to within tolerance of $10^{-12}$ with the Nelder Mead search, and stopping if nothing was found after 100 attempts). Using the following sequence of interactions did not yield any gates with $f_m < 0.001$
\begin{align}
{U}_{1} = {U}_{\times} {U}_{\Box} {U}_{\circlearrowleft}, \; {U}_{2} = {U}_{\asymp}^{(1)} {U}_{\times} {U}_{\Box} {U}_{\circlearrowleft} {U}_\asymp^{(2)}.
\end{align}
The reason for choosing these sequences in particular is that we are interested in forming a gate from the ring exchange, and these are the simplest sequences that contain a ring exchange interaction (with local rotations before and after the ring exchange). By adding one more interaction, we can find a sequence which does work:
\begin{align} \label{Usequence}
{U}_{gate}= {U}_{\asymp}^{(1)} {U}_{\parallel} {U}_{\times} {U}_{\Box} {U}_{\circlearrowleft} {U}_\asymp^{(2)} = {U}_{\asymp}^{(1)} {U}_{\parallel} {U}_{\times}' {U}_S {U}_{\asymp}^{(2)},
\end{align}
with corresponding parameters $\{ \theta_{\asymp}^{(1)}, \theta_{\parallel}, \theta_{\times}, \theta_{\Box}, \theta_{\circlearrowleft}, \theta_\asymp^{(2)} \}$, where the superscripts (1) and (2) are used to differentiate between the interactions used at the beginning and at the end of the gate. This is a gate which requires only $5$ separate pulses to perform (since identity (\ref{eqn:DFSIden1}) can be used). Note that we also attempted to form a gate using the sequence in (\ref{Usequence}) but with $\theta_\times = \theta_\Box$ in an effort to reduce the number of pulses, but no solutions were found.
Using this combination of interactions, with the parameters given in Table~\ref{thetaVals}, we were able to find a gate with a value of $f_m = O(10^{-16})$ and leakage $L = O(10^{-16})$ (i.e.\ both around the machine precision of $O(10^{-16})$, suggesting the existence of an exact solution). If we assume that for each interaction the coupling strength is limited to some maximum possible value $J_{max}$, then we can find the total gate time $T$ in units of $\hbar /J_{max}$, as an indicator of how long this gate would take compared to other gates. Note that we do not simply add the parameters in Table~\ref{thetaVals}, since we apply the identity in eqn.\ (\ref{eqn:DFSIden1}) first, so in fact the true gate time $T$ is
\begin{align}
T =\left( \sum_n \theta_n \right) - \theta_{\times} + (\theta_{\times}-\theta_{\Box})\textrm{mod}\;{2\pi}.
\end{align}
This gives a gate time of $16.7\; \hbar /J_{max}$. The gate in~\cite{Fong2011} has a total time of 9.9 in these units, and although we found other parameter sets which yielded similar gate times to this, we have picked the parameters with the most realistic ring exchange couplings (see Sec.~\ref{sec:Constraints}).
\begin{table}[htbp]
\begin{centering}
\begin{tabular}{c c c c c c | c c c}
$\theta_{\asymp}^{(1)}$ & $\theta_{\parallel}$ & $\theta_{\times}$ & $\theta_{\Box}$& $\theta_{\circlearrowleft}$& $\theta_{\asymp}^{(2)}$& $f_m$ & $L$ & $\alpha$ \\[5pt] \hline
2.748894 & 4.319690 & 2.552544 & 3.730678 & 0.589049 & 0.785361 & $O(10^{-16})$ & $O(10^{-16})$ & 0.158
\end{tabular}
\caption{\label{thetaVals}Parameters which realise a $\textsc{cz}$ gate, up to local rotations.}
\end{centering}
\end{table}
\section{Constraints on the ring exchange strength} \label{sec:Constraints}
We are constrained in our choices of parameters, as the relative sizes of the nearest-neighbour couplings, $J$ and ring-exchange coupling $J_{\circlearrowleft}$, are set by the ratio $\alpha = J_{\circlearrowleft}/J = \theta_{\circlearrowleft} /\theta_{\Box}$, since as we have seen in eqn.\ (\ref{eqn:DFSIden1}) we end up turning on the Hamiltonian $ {H}_{\times}+{H}_{\Box}+ \alpha {H}_{\circlearrowleft}$ for some time $t_{\Box} =t_{\circlearrowleft}=\theta_{\Box} \hbar / J$. This means we are constrained to situations where we can set $\alpha$ to $\theta_{\circlearrowleft} /\theta_{\Box}$. We would expect $\alpha \lesssim 0.17$ (see e.g.~\cite{Katanin02,Coldea01,Mizel2004}), which is why we have chosen to use the particular parameter set shown in Table~\ref{thetaVals} which has $\alpha = 0.158$.
We are not completely constrained to this value, since all of the $\theta$ values are just phases. As discussed in Sec.~\ref{sec:Two}, we are free to add factors of $2 \pi$ to any of them, with the gate time increasing by $2\pi \hbar / J$ every time we add a multiple of $2 \pi$. However, even with this freedom, we still seem to be tied down to a few precise values of $\alpha$. To get around this, we notice that we can split up the ring exchange into two parts:
\begin{align}
{U}_S &= \exp \left(-i {H}_{\times} \theta_{\Box} \right) \exp \left(-i {H}_{\Box} \theta_{\Box} \right) \exp \left(-i {H}_{\circlearrowleft} \theta_{\circlearrowleft} \right) \nonumber \\
&= \exp \left(-i {H}_{\times} (\theta_{\Box}^a + \theta_{\Box}^b) \right) \exp \left(-i {H}_{\Box} (\theta_{\Box}^a + \theta_{\Box}^b)\right) \exp \left(-i {H}_{\circlearrowleft} (\theta_{\circlearrowleft}^a + \theta_{\circlearrowleft}^b) \right) \nonumber\\
&= \exp \left(-i {H}_{\times} \theta_{\Box}^a \right)\exp \left(-i {H}_{\Box} \theta_{\Box}^a \right) \exp \left(-i {H}_{\circlearrowleft} \theta_{\circlearrowleft}^a \right) \exp \left(-i {H}_{\times} \theta_{\Box}^b \right)\exp \left(-i {H}_{\Box} \theta_{\Box}^b \right) \exp \left(-i {H}_{\circlearrowleft} \theta_{\circlearrowleft}^b \right) \nonumber \\
&= {U}_{\times}^a {U}_{\Box}^a {U}_{\circlearrowleft}^a {U}_{\times}^b{U}_{\Box}^b {U}_{\circlearrowleft}^b:={U}_S^a {U}_S^b
\end{align}
so we now have two ring exchange interactions, ${U}_S^a$ and ${U}_S^b$ which have the same form as $H_S$ but with different values of $\alpha$. Now since $\theta_n := J_n t_n / \hbar$, and since ${U}_S^a{U}_S^b = {U}_S$ and all of the terms commute, this means that:
\begin{align}\label{eqn:Constr}
&J^a t_a + J^b t_b = \theta_{\Box}, \; J_{\circlearrowleft}^a t_a + J_{\circlearrowleft}^b t_b = \theta_{\circlearrowleft}.
\end{align}
where $\theta_{\Box}$, $\theta_{\circlearrowleft}$ are the parameters we found in the search in Sec.~\ref{sec:Two}, and we have defined $\theta_{\Box}^a = J^a t_a$, $\theta_{\Box}^b = J^b t_b$. For simplicity, we take $J^a = J^b = J$, without loss of generality, since we are free to scale these parameters as we wish, provided we scale the corresponding $J_{\circlearrowleft}$ values correctly. Since $\theta_{\Box}$ and $\theta_{\circlearrowleft}$ are phases, we are free to add multiples of $2\pi$ to these values, so we replace $\theta_{\Box}$ with $\theta_{\Box}^{(n)}$, and rearrange equation (\ref{eqn:Constr}) to give
\begin{align}
&t_a = \frac{\theta_{\Box}^{(n)}}{J}\left[ \frac{\alpha^{(n)} - \alpha_b}{\alpha_a - \alpha_b} \right] , \;t_b = \frac{\theta_{\Box}^{(n)}}{J}\left[ \frac{\alpha_a- \alpha^{(n)} }{\alpha_a - \alpha_b} \right],
\end{align}
where $\alpha_a := \theta_{\circlearrowleft}^a / \theta_{\Box}^a =J_{\circlearrowleft}^a / J^a $, $\alpha_b := \theta_{\circlearrowleft}^b / \theta_{\Box}^b = J_{\circlearrowleft}^b / J^b$, and $\alpha^{(n)}$ is defined above. For $t_a$ and $t_b$ to be positive, we need couplings such that $\alpha_a > \alpha^{(n)} >\alpha_b$. So this tells us that if we are able to control the relative strengths of the ring and nearest-neighbour terms, and if we could get them such that $\alpha_a >\alpha^{(n)} >\alpha_b$ is satisfied, then regardless of what the actual values of $\alpha_a$ and $\alpha_b$ are, we can create the $\textsc{cz}$ gate (at the expense of adding more interactions and thus increasing the time of the gate).
\section{Including noise}\label{sec:Performance}
We now look at the performance of this gate under the influence of noise. The two types of noise we consider are errors in coupling strengths or timing of the gates (i.e.\ random errors in the coupling parameters when implementing each of the Hamiltonians in (\ref{Usequence})) and fluctuations in magnetic fields acting on the qubits during the gate implementation (which would normally be protected by the decoherence-free or supercoherent properties, but are not while the gates are being implemented). The reasons for picking these particular types of errors are that one of the most promising systems in which to implement these encoded qubits is in arrays of quantum dots, and there have been several advances towards achieving these, e.g.~\cite{Laird2010,Thalineau2012}. These quantum dots are susceptible to errors in exchange coupling due to charge fluctuation~\cite{Burkard1999} (which we are modelling by fluctuations in the $\{ \theta \}$ values) and fluctuations in external magnetic field due to the nuclear spin bath or stray magnetic fields (see e.g.~\cite{Taylor2007}). When looking at these magnetic fluctuations, we look at two cases; in the first case we assume that magnetic field fluctuations are roughly uniform over each encoded qubit (so that there is collective decoherence acting on it), but the magnetic fields acting on different encoded qubits are different in magnitude and direction. In the second case, we consider magnetic field fluctuations acting independently on each physical qubit (i.e.\ a situation where a supercoherent qubit would be more appropriate). For all errors, we assume that the time scale for the fluctuations is large compared to the time to perform the gate, which is typically the case~\cite{Hu2006,Merkulov2002,Taylor2007}. We also compare the susceptibility of the 3-qubit and 4-qubit encodings to errors, since there may be difference.
To measure the effects of noise, we calculate the gate fidelity using the techniques of quantum process tomography discussed in Section~\ref{sec:OSR}. The unitary transformation applied to the encoded qubits is $e^{i\Phi} {V}^{\dagger}{U} {V}$, where $V$ is a local rotation, $\Phi$ is a global phase, and $U$ is the unitary found in Sec.~\ref{sec:Two} which is locally equivalent to a $\textsc{cz}$ gate. To find the local operations $V$, we take the gate with zero noise ${U}$ corresponding to the gate sequence found in Sec.~\ref{sec:Two}, and find the matrix ${V}$ which diagonalises it, and then find $\Phi$ such that e.g. $e^{i\Phi} \bra{\bar{0}^{(4)}}{V}^{\dagger}{U} {V} \ket{ \bar{0}^{(4)}} =1$ for the 4-qubit encoding. Once we have a gate with noise added, ${U}'$, ${V}$ and $\Phi$ are still used to convert ${U}'$ into a noisy $\textsc{cz}$ gate (i.e. $e^{i\Phi} {V}^{\dagger}{U}' {V} \approx \textsc{cz}$), since we are restricted to always using the same local operations (i.e.\ we cannot change our local operations since we don't know what the noise is doing to our system).
Following this unitary evolution, we imagine performing a measurement on each of the encoded qubits to test whether or not they are still in the logical subspace, which corresponds to applying the projectors $P = P_3 $ or $P_4$ depending on whether we are considering the 3- or 4-qubit encoding. The probability of the qubit being in the logical subspace is given by the leakage defined in (\ref{eqn:Leakage}). Overall then, a state $\rho$ is transformed by this process according to $\mathcal{E} (\rho) = \sum_i {E}_i \rho {E}_i^{\dagger}$ where the Kraus operators take the form
\begin{align}
E_1= \sqrt{1-L}\text{ } P e^{i\Phi} {V}^{\dagger}{U} {V}, \; E_2=\sqrt{ L} \text{ } (1-P) e^{i\Phi} {V}^{\dagger}{U} {V}
\end{align}
where $P = P_3 $ or $P_4$.
Rather than finding the fidelity of this process as a whole, it is more useful to decompose the analysis of this noisy gate into two parts: the probability $L$ of being outside the logical subspace, and the process fidelity of the gate when projected into the logical subspace, since we don't mind what happens to states outside the logical subspace / subsystem. Therefore we construct a process matrix $\chi$ from a process $\mathcal{E}'$ involving $E_1$ alone:
\begin{align}
\mathcal{E}' (\rho) &= \frac{1}{(1-L)}{E}_1 \rho {E}_1^{\dagger} = \sum_{mn} \chi_{mn} {A}_m \rho {A}_n^{\dagger}
\end{align}
where $\{ {A}_m \}$ is a set of $4 \times 4$ matrices that form an orthogonal basis, and we have divided by $(1-L)$ to account for the probabilistic nature of the measurement. We use the process fidelity $F_p(\chi,\chi_{\textsc{cz}})$ to compare this to the process matrix of an ideal \textsc{cz} gate, and then $(1-(1-L)F_{p})$ has the interpretation as the upper bound of the average failure probability $\bar{p}_e$~\cite{Gilchrist2005}, making this a natural choice for measuring the accuracy of the gate. Note that a subtlety arises when dealing with the 3-qubit DF subsystem, in that $P_3$ projects onto a 16-dimensional space, so we cannot directly compare this to the 4-dimensional $\textsc{cz}$ gate. To get around this, we can instead project onto only one of the gauge degrees of freedom, so for instance we could replace $P_3$ with $P_3^{(1,1)}$ (defined in Sec.~\ref{sec:Two}),
provided $P_3^{(1,1)} U P_3^{(1,1)} \ne 0$, otherwise we could project onto any of the other 3 gauge subspaces. Then instead of normalising $\chi$ by dividing by $(1-L)$, we can just divide by $\textrm{tr}(\chi)$ since $\textrm{tr}(E_1 \rho E_1) = \textrm{tr}(\chi)$, and the effect is the same. Note that the leakage is still calculated in the same way as in Sec.~\ref{sec:Two}, independently of the method used to project into the 4-dimensional subspace.
\subsection{Coupling Errors}
To simulate random errors in values of the couplings between physical qubits, charge fluctuations or timing when implementing gates, we added random noise to each of the $\theta_n$ parameters:
\begin{equation}
\theta_n \rightarrow \theta_n + \delta\theta_n,
\end{equation}
where $\delta\theta_n$ is sampled from a Gaussian distribution. Over the range $\varepsilon \in [0,0.05]$ of $\varepsilon$, we calculated the process fidelity over 250 iterations taken from a Gaussian distribution with standard deviation $\varepsilon$ and mean 0, finding the average over all of these iterations (note that all the interactions commute with $\mathbf{S}^2$, so the system is confined to a subspace of constant $S$). The results are shown in Fig.~\ref{TimingErrors}, with only one set of results shown since both the 3- and 4-qubit encoding give very similar results. The process fidelity falls off slowly and stays above $0.9$ for $\epsilon \lesssim 0.05$. A reasonable estimate of these fluctuations in gate couplings would be around $0.01$~\cite{Burkard1999}, at which point both gates have fidelity $\sim 0.99$, so we can see that these gates still have high fidelity even with this level of noise. Over the entire range of $\epsilon$, the leakage stayed below $0.003$, and the leakage at $\epsilon \sim 0.01$ is around $O(10^{-6})$. So overall we can achieve an overall average gate failure probability of $\bar{p}_e \lesssim 0.01$ even with a reasonable level of coupling errors.
\begin{figure}[h]
\centering\includegraphics[width = 0.6\textwidth]{TErrorsFid-eps-converted-to.pdf}
\caption{The average process fidelity $\bar{F}_p$ when performing the gate with random fluctuations in gate times, where the fluctuations have standard deviation $\varepsilon$. Only results for the 4-qubit encoding are shown, since both encodings showed a similar behaviour. A fit to a curve of the form $y = 1 - c \varepsilon^2$ is also shown, with $c=35.4$.}
\label{TimingErrors}
\end{figure}
\subsection{Magnetic fluctuations}
\begin{figure}[h!]
\centering
\subfloat[]{\includegraphics[width = 0.7\textwidth]{MagFieldErrors_Fid-eps-converted-to.pdf}} \\
\subfloat[]{\includegraphics[width = 0.7\textwidth]{MagFieldErrors_SingleSite-eps-converted-to.pdf}}
\caption{\label{fig:MagEvolveLong}Average process fidelity $\bar{F}_p$ as a function of $B_{\textrm{nuc}} / J$, the strength of magnetic field fluctuations relative to the exchange coupling, with a) random uniform fields acting on each logical qubit and b) random fields on each physical qubit. The range is over $B_{\textrm{nuc}} / J= 0 \rightarrow 0.02$, and results corresponding to a 3-qubit or 4-qubit encoding are both shown. A fit of all data up to $B_{\textrm{nuc}} / J = 0.01$ to a curve of the form $y = 1 - c ( B_{\textrm{nuc}} / J)^2$ is also shown, with $c_4=95.6$, $c_3=252.9$ for the 4-qubit and 3-qubit encoding respectively. Note that we use this fit to see roughly when the fidelity starts to deviate from a quadratic decay, not as a best fit to the data.}
\end{figure}
If we were to implement this supercoherent qubit in a quantum dot, then there could be random magnetic field fluctuations due to the nuclear spins in the substrate material, or stray magnetic fields. We studied two scenarios which could occur; in one case we looked at the effects of having random field fluctuations which are uniform over all physical qubits, but which may vary between encoded qubits (so that the collective decoherence assumption is valid for single encoded qubits, but when we interact two of these the assumption is not valid). In the other case we considered having independent random magnetic field fluctuations on each physical qubit. For both of these cases, we followed the arguments in~\cite{Merkulov2002, Taylor2007}, using the quasi-static approximation in which the magnetic field from the nuclei $\vec{B}$ stays constant over the time we perform the gate, has random direction and has magnitude $| \vec{B} |$ following a Gaussian probability distribution
\begin{align}
P(| \vec{B} |) = \frac{1}{(2 \pi B_{\textrm{nuc}}^2)^{3/2}} \exp(- | \vec{B} |^2 / 2B_{\textrm{nuc}}^2),
\end{align}
where $B_{\textrm{nuc}}$ is the standard deviation in fluctuations of magnetic field. We took the average over 250 iterations, with each iteration having a different magnitude and direction of magnetic field sampled from the above Gaussian distribution. The magnetic field strength was taken relative to the nearest-neighbour coupling strength $J$. In each case we found how the process fidelity varied for the 3-qubit and 4-qubit encoding. The results are shown in Fig.~\ref{fig:MagEvolveLong} a) and b), together with a fit of all data up to $B_{\textrm{nuc}} / J= 0.01$ to a curve of the form $\bar{F}_p = 1 - c \left( B_{\textrm{nuc}} / J \right)^2$ (we use this fit to see roughly when the fidelity starts to deviate from a quadratic decay, not as a best fit to the data).
Based on the exchange values given in~\cite{Loss1998} and the values for $B_{\textrm{nuc}}$ given in~\cite{Taylor2007}, we would expect $B_{\textrm{nuc}}/J$ to be at most $\sim 0.01$. With this kind of nuclear field present, the 4-qubit encoding achieves a process fidelity of $\bar{F}_p \sim 0.99$ and a leakage probability of $L \sim 0.002$ if there are errors across the encoded qubits, or $\bar{F}_p \sim 0.97, L \sim 0.002$ if there are errors on each individual physical qubit. The 3-qubit encoding achieves $\bar{F}_p\sim 0.97,L \sim 0.002$ if there are errors across encoded qubits, or $\bar{F}_p \sim 0.95,L \sim 0.002$ if there are errors on physical qubits.
So overall, as we might expect, the qubits are much more robust to errors which are uniform on the physical qubits, but less robust to errors which vary between physical qubits. Also, the gate on the 3-qubit encoding performs slightly worse than the 4-qubit one, which we are currently unable to explain. Using a 4-qubit encoding in the presence of magnetic fluctuation across encoded qubits, with a strength we might expect in a realistic system, we can achieve process fidelities of $\bar{F}_p > 0.99$ and leakages of $L \sim 0.002$ (leading to an overall average gate failure probability of $\bar{p}_e \lesssim 0.01$). The results for the situation with errors on physical qubits are unsurprisingly worse, but we can still achieve $\bar{F}_p \sim 0.97$, $L \sim 0.002$, giving an overall average gate failure probability of $\bar{p}_e \lesssim 0.03$. These results could be improved if we reduced the effects of fluctuations in nuclear spin, such as the methods presented in~\cite{Imamoglu2003} and~\cite{Reilly2008}.
\section{Conclusions}
We have demonstrated a simple way to implement a \textsc{cz} gate in the 4-qubit decoherence-free subspace and the 3-qubit decoherence-free subsystem, using a sequence of 5 operations (excluding local operations), and including ring exchange interactions. The gate we have found minimises the Makhlin invariant function $f_m$ to within machine precision, suggesting the existence of an exact solution. The simplicity of this gate also suggests that interactions involving more than 2 physical qubits can be used to achieve simplified gates, which we might intuitively expect since a direct $\textsc{cz}$ gate on these encoded qubits would involve a four-body interaction, which is present in the ring exchange terms. Investigating this more rigorously could pose an interesting question for future research.
We introduced errors when performing these gates, to simulate errors in coupling strength or gate times, and to simulate fluctuations in magnetic field due to some external environment, e.g.\ nuclear spins in a quantum dot. The 4-qubit gate maintained an average failure probability of $\bar{p}_e \lesssim 0.01$ even with nuclear fluctuations of around $1\%$ of $J$ over the encoded qubits, or with timing errors of up to around $1 \%$ of $J/\hbar$, where $J$ is the strength of the nearest-neighbour exchange coupling. Whilst this maintains the gate below the most generous threshold condition~\cite{Knill2005}, it is likely that it will be advantageous to keep these errors errors significantly lower in order to implement this gate in a fault tolerant manner. Techniques such as dynamical decoupling and leakage reduction, previously applied to 3-qubit encoded qubits in~\cite{West2012,Fong2011}, could perhaps be used to strengthen the performance of this gate in the presence of noise.
Such a gate could be useful in systems where ring exchange is particularly prominent, or in situations where it is particularly important to keep the number of pulses to a minimum, or where the control is limited. Further studies could be done to investigate the effects of magnetic flux on the couplings as reported in~\cite{Scarola2005}, or using more general forms of the ring exchange interaction rather than the symmetric one we have used here.
\chapter{Introduction}
\label{chap:Intro}
A study in 2007 estimated the total information storage capacity across all computer devices in the world to be around $295\times 10^{18}$ bytes (295 exabytes)~\cite{Hilbert2011}, roughly the same information contained by paperback novels stretching to the sun and back 200 times. This is symptomatic of our ever growing appetite for information and computing power, and how much computers have become integral to our lives over the last 50 years or so. Computers are now used in phones, coffee machines, glasses, predicting the weather, even creating new forms of currency.
The cause, or result, of this has been a reliable exponential growth in computational power over the years, as first noticed by Moore in his famous law~\cite{Moore1965}, which states that the number of transistors that can be put inexpensively onto a square metre approximately doubles every 2 years. The numbers are impressive; each year we gain enough computational power to perform around 60\% of the computations that could have possibly been executed by all existing general-purpose computers in the years before~\cite{Hilbert2011}. The Apollo 11 guidance computers, weighing 32kg and costing hundreds of thousands of dollars, are now outperformed by a Raspberry Pi, weighing 0.05kg and available for less than \pounds 25~\cite{NASA}. This steady trend has been possible thanks to extraordinary feats in physics and engineering, from integrated circuits to the development of complementary metal-oxide-semiconductor (CMOS) technology. However there is only so far that this technology can go, and it doesn't seem long before we reach fundamental limitations of fabrication and device reliability~\cite{Frank2002,Frank2001,Markov2014,Chien2013}.
Whilst historically our predictions about the limitations of computational power seem to be about as accurate as predictions of the end of the world, and we aren't sure of the future economics behind new computer research, it all points to a future where information is stored at the scale of molecules and atoms. With information stored at this scale, we must constantly battle the effects of quantum mechanics, or else embrace it and start to use the features of quantum mechanics to make an entirely new form of computation. This is the paradigm of \emph{quantum computation} (as opposed to the \emph{classical} computation we have today), first proposed by Benioff and Feynman~\cite{Benioff1982, Feynman1982}, and later made more concrete by Deutsch~\cite{Deutsch1989}.
Aside from satisfying our appetite for smaller computers, the idea of quantum computation raises many fundamental questions, and provides an interesting link between the seemingly different disciplines of physics and computer science. Intuitively, if we really do believe that quantum mechanics is a better description of reality than classical physics, then we expect that a quantum computer must be able to do things that a classical computer cannot. This is at odds with an enduring law of computation, the strong Church-Turing thesis, which states that any task that can be efficiently computed on a physically reasonable device can be simulated efficiently by a (probabilistic) Turing machine~\cite{Turing1936,Church1936}. Can a quantum computer be efficiently simulated by a Turing machine, or is there a violation of this rule? Whilst there are no outright proofs that this is the case, there is much evidence. The most prominent example is Shor's algorithm~\cite{Shor1994}, building on work by Simon~\cite{Simon1994}, which gives an efficient quantum algorithm for prime factorisation (a problem for which the best known classical algorithm takes close to exponential time). This is an important problem in areas such as cryptography where the hardness of prime factorisation underpins the hardness of breaking RSA~\cite{Rivest1978}, the cryptographic protocol at the heart of most secure transactions online.
Other examples of differences between quantum computers and classical computers are Grover's quadratic speed up for searching through unsorted lists~\cite{Grover1997} and the apparent difficulty (assuming that commonly held beliefs in computer science are true) of classically simulating circuits of commuting operations~\cite{Bremner2011} or linear optical networks~\cite{Aaronson2011}. Even more claims are made in the media, with the promise that quantum computers will do anything from winning elections to discovering planets~\cite{Time}.
Whether or not there is a difference between classical and quantum computing, the outcome will be positive for physics; we either unleash untapped computational power (potentially also allowing fast computer simulations of quantum systems), or we find some (perhaps fundamental) limits to our ability to control the world, as well as the numerous other offshoots that have already stemmed from quantum information theory, such as quantum cryptography, quantum sensing and random number generation. And regardless of what we do actually know about how quantum computation will change things, the history of classical computers tells us that we won't know until we try it.
But even if we accept that it is a good idea to build a quantum computer, is it \emph{practically} possible to build a quantum computer? Are there any insurmountable obstacles which prevent us from constructing a universal quantum computer? Recent work shows that in order to simulate the ground state of a molecule that is twice the size as the current largest classically simulatable molecule, a quantum computer would need on the order of $10^{18}$ gates~\cite{Wecker2014}, and to factorise a 200 digit number requires approximately 100,000 qubits~\cite{Steane1999}. Given that current state-of-the-art programmable quantum computers have reached around 20 qubits~\cite{Monz2011}, and quantum effects are often washed away in the macroscopic world, this seems like an insurmountable task. Fortunately, it is possible to encode information in ways such that the errors are correctable so, provided there are no unexpected experimental brick walls, creating a quantum computer only seems like a matter of time and effort.
This thesis sits somewhere in between the practical and theoretical side of quantum computing, aiming to push the proposals of quantum computing just a little bit closer to the experimentalists' reach. Perhaps reflecting the diverse range of topics that make up quantum information theory, this thesis is also split into different projects. These broadly fall into three categories: work aimed at finding new ways to create quantum gates, or classical gates using quantum components, work aimed at comparing different models of computation, and work aimed at investigating how we might communicate in a quantum computer. The unifying theme behind these is the desire to make quantum computation more attainable.
In Chapter~\ref{chap:DFS}, we propose a way in which 2-qubit gates could be performed in a decoherence-free subspace. We focus on using a small number of minimal-control pulses, with minimally engineered interactions, including many-body ring exchange interactions that are predicted to be present in many systems. The gate we find has less pulses than any previous scheme, but takes roughly the same time as the best known scheme. During the 2-qubit decoherence-free gate, the system leaks out of the decoherence-free subspace, and the effects of this are studied by simulating the effect of control errors and magnetic field fluctuations. We find that the gate error remains below 1\% for errors typical of gated quantum dots in GaAs, which is below the most generous fault-tolerant threshold.
Chapter~\ref{chap:3qubit} similarly deals with a proposal to make gates, this time 3-qubit Toffoli and Fredkin gates, using linear arrays of qubits. Such gates are not only useful for quantum circuits, but are also a useful primitive for classical reversible computation. We focus on using only a single pulse, with interactions that are as uniform as possible. We find one method to perform a Toffoli gate, and two methods to perform a Fredkin gate. Experimental schemes for realising two of these gates in ion traps and donors in silicon are also proposed, along with an assessment of how viable these schemes could be, for both quantum and classical computation, using current technology. The reachable fidelities of some of these gates with current technology are lower than the quantum error correction threshold, suggesting that they could be a useful technological innovation for either quantum or classical computation, perhaps providing a stepping stone en route to full quantum computation.
Chapter~\ref{chap:AGQC} combines the traditional measurement-based quantum computation (MBQC) protocol with adiabatic evolution, in an extension of the adiabatic cluster-state model. We show that any MBQC pattern which satisfies a set of graphical rules called \emph{gflow} can be converted into a holonomic quantum computation where the measurements are replaced by adiabatic evolutions, such that the computation is the same. We investigate how the trade-offs in MBQC manifest in this adiabatically driven model, and explore which orderings of operations are possible.
In Chapters~\ref{chap:Wigner} \&~\ref{chap:NNN}, the ability of spin chains to transfer information is studied. The viability of using a 1D (or quasi-1D) Wigner crystal as a quantum communication channel is investigated using the semi-classical instanton approximation, finding that high average fidelities are possible for the particular couplings that we would expect to be present. This is followed by a study of the effects of next-neighbour interactions on state transfer through an ideal Heisenberg spin chain, and preliminary results for using edge locking and ground state correlations to transfer information.
\section{Quantum Computation and notation}\label{sec:CircuitModel}
In this section we give a broad introduction to quantum computing, giving a brief overview of the nomenclature and notation used in the ensuing chapters of this thesis. The fundamental building block of quantum computation is the quantum bit, or \emph{qubit}, in which quantum information can be stored. A qubit can be any two-level quantum system, such as two energy levels in an atom, two spin states of an electron, or two polarisations of a photon, and in this thesis all of the results will use qubits (although computation is of course possible using higher dimensional systems).
Qubits are represented by a two-dimensional normalised complex vector (more formally, it is a vector in the Hilbert space $\mathcal{H} = \mathbb{C}^2$, where $\mathbb{C}$ is the set of complex numbers). The two states of the qubit are usually denoted $\ket{0}$ and $\ket{1}$ in analogy to the bits in a classical computer. Unlike classical bits, the postulates of quantum mechanics permit arbitrary superpositions of these qubit states to exist, written as $\ket{\phi} = \alpha \ket{0} + \beta\ket{1}$, where $|\alpha|^2 + |\beta|^2 = 1$. Several qubits can be combined together using the tensor product $\otimes$, obtaining a $2^n$ dimensional complex vector for $n$ qubits (and a corresponding Hilbert space $\mathcal{H} = \mathbb{C}^2 \otimes \mathbb{C}^2 \otimes ... \otimes \mathbb{C}^2 = (\mathbb{C}^2)^{\otimes n}$) written as e.g. $\ket{\phi}\otimes \ket{\psi}$ or $\ket{\phi \psi}$. A particularly important basis of many qubits is the computational basis, defined as all possible tensor products of $\ket{0},\ket{1}$ over $N$ qubits, i.e. $\{ \ket{000...0},\ket{100...0},\ket{010...0}...,\ket{111...1} \}$.
A more general quantum state is represented by a Hermitian, non-negative, trace one complex density matrix, allowing for the possibility that the preparation of a quantum state may be uncertain, and so is a classical ensemble of quantum states. In general these states have the form $\rho = \sum_n p_n \ket{\phi_n}\bra{\phi_n}$, where $p_n$ is the probability that the system is in state $\ket{\phi_n}$. When the density matrix can be written $\rho = \ket{\phi}\bra{\phi}$ (or equivalently $\text{tr} (\rho^2) = 1$), it is said to be a pure state, and otherwise it is said to be a mixed state ($\text{tr} (\rho^2) < 1$).
A quantum computation then typically proceeds in the following steps
\begin{itemize}
\item[1)] Qubits are prepared in an initial pure state, which is the input to the computation.
\item[2)] A sequence of operations are performed on these qubits.
\item[3)] The qubits are measured at the end of the computation, giving the outcome of the algorithm with suitably high probability.
\end{itemize}
This is the essence of the most commonly used model of quantum computation, the \emph{circuit model}, first formulated by Deutsch~\cite{Deutsch1989}. Such computations are typically represented as circuit diagrams, where qubits are represented as horizontal lines, and operations on these qubits are shapes on these lines, with time going forward from left to right. For example, a circuit with two qubits prepared in states $\ket{\phi}$, $\ket{0}$ which performs a two qubit $\textsc{cnot}$ gate between them is
\[
\Qcircuit @C=1em @R = 2.5em @!R {
\lstick{\ket{\phi}} & \qw & \ctrl{1} & \qw & \qw \\
\lstick{\ket{0}} & \qw & \targ{U} & \qw & \qw \\
}
\]
Operations are performed by turning on interactions between qubits for times prescribed by the algorithm. In the standard formalism of quantum mechanics, the interactions in the system are given by the (generally time-dependent) Hamiltonian $H(t)$, and the evolution of the qubits due to this Hamiltonian is given by the unitary matrix $U(t)$
\begin{align}
U(t) = \mathcal{T} \exp \left( -\frac{i}{\hbar} \int_0^t H(t') dt' \right)
\end{align}
where $\mathcal{T}$ is the time-ordering operator. Single qubit rotations can be achieved by turning on a Hamiltonian consisting of one of the 3 Pauli matrices:
\begin{align}
X = \left(\begin{array}{cc}
0&1\\
1&0
\end{array}\right)\text{, }
Y = \left(\begin{array}{cc}
0&-i\\
i&0
\end{array}\right)\text{, }
Z = \left(\begin{array}{cc}
1&0\\
0&-1
\end{array}\right).
\end{align}
To perform rotations about the $x$, $y$ or $z$ axis respectively, a Hamiltonian $h_x X,h_y Y$ or $h_zZ$ must be turned on, resulting in the unitary rotations
\begin{align}
U_x(\theta_x) &= \exp \left(i\theta_x X/2 \right) = \cos\left(\frac{ \theta_x}{2} \right) \idop + i\sin\left(\frac{ \theta_x}{2} \right)X\\
U_y(\theta_y) &= \exp \left(i\theta_y Y /2 \right) =\cos\left(\frac{ \theta_y}{2}\right)\idop + i\sin\left(\frac{ \theta_y}{2}\right)Y\\
U_z(\theta_z) &= \exp \left( i \theta_z Z/2\right) =\cos\left(\frac{ \theta_z}{2} \right) \idop + i\sin\left(\frac{ \theta_z}{2} \right)Z,
\end{align}
where $\theta_\alpha = -h_\alpha t/\hbar$, with $h_\alpha$ as the strength of the interaction, and the equality follows since $X^2 = Y^2 = Z^2 = \idop$. Using any two of these rotations is enough to perform arbitrary rotations of a single qubit, as a general rotation can be written as e.g.\ $U_x(t_1)U_z(t_2)U_x(t_3)$~\cite{NielsenChuang}. We will use the shorthand $X_n$, $Y_n$, $Z_n$ to denote single qubit operations acting on qubit $n$ only and identities elsewhere, so for example $Z_1 = Z\otimes \idop$. We also define the Pauli group on $n$ qubits, denoted $\mathcal{G}_n$, as $\mathcal{G}_n = \langle X_1,Y_1,Z_1,X_2,Y_2,Z_2,...,X_n,Y_n,Z_n\rangle$, where the notation $\langle \cdot \rangle$ indicates that these operators generate the full group. We will often use the Heisenberg exchange interaction between qubits, which we denote $E_{nm} := X_n X_m + Y_n Y_m + Z_n Z_m$.
Other important single qubit operations that will be used extensively in this thesis are the Hadamard ($\text{H}$) and raising and lowering operators $\sigma^\pm$ defined as
\begin{align}
\text{H} = \frac{1}{\sqrt{2}}\left(\begin{array}{cc}
1&1\\
1&-1
\end{array}\right), \; \sigma^\pm = \frac{1}{2} ( X \pm iY).
\end{align}
Clearly to perform general quantum computations, we will also need interactions between qubits. To perform universal quantum computation, it is only necessary to be able to perform any entangling two-qubit gate plus single qubit rotations~\cite{Barenco95,DiVincenzo1995,Bremner2002}. In fact not even general single qubit rotations are necessary, but only certain single qubit operations are required. The two most commonly entangling two-qubit gates are the controlled-\textsc{not} (\textsc{cnot}) and controlled-$Z$ (\textsc{cz}) gates. Here the two qubits are split into a control and target qubit, and the gate implements a Pauli $X$ (\textsc{cnot}) or $Z$ (\textsc{cz}) on the target qubit when the control qubit is in the $\ket{1}$ state. In matrix form they are
\begin{align}
\textsc{cnot} = \left( \begin{array}{cccc}
1&0&0&0\\
0&1&0&0\\
0&0&0&1\\
0&0&1&0
\end{array}\right)\text{, }
\textsc{cz}= \left( \begin{array}{cccc}
1&0&0&0\\
0&1&0&0\\
0&0&1&0\\
0&0&0&-1
\end{array}\right).
\end{align}
Two other important gates, which also act as universal gates when combined with single qubit rotations, and which are universal for classical reversible computation are the Toffoli (controlled-controlled-\textsc{not}) gate and Fredkin (controlled-\textsc{swap}) gate~\cite{Fredkin1982}:
\begin{align}
T = \left( \begin{array}{cccccccc}
1&0&0&0&0&0&0&0\\
0&1&0&0&0&0&0&0\\
0&0&1&0&0&0&0&0\\
0&0&0&1&0&0&0&0\\
0&0&0&0&1&0&0&0\\
0&0&0&0&0&1&0&0\\
0&0&0&0&0&0&0&1\\
0&0&0&0&0&0&1&0\\
\end{array}\right)\text{, }
F = \left( \begin{array}{cccccccc}
1&0&0&0&0&0&0&0\\
0&1&0&0&0&0&0&0\\
0&0&1&0&0&0&0&0\\
0&0&0&1&0&0&0&0\\
0&0&0&0&1&0&0&0\\
0&0&0&0&0&0&1&0\\
0&0&0&0&0&1&0&0\\
0&0&0&0&0&0&0&1\\
\end{array}\right).
\end{align}
An important subgroup of all possible quantum operations is the \emph{Clifford group}. An important property of this group is that conjugation of any member of the Pauli group by a Clifford gives another member of the Pauli group, or more formally
\begin{align}
MPM^\dagger \in \mathcal{G}_n \quad \forall M \in \mathcal{C}_n, P \in \mathcal{G}_n,
\end{align}
where $\mathcal{C}_n$ is the Clifford group for $n$ qubits. Because of this property, any circuit made up entirely of Clifford gates can be simulated efficiently by a classical computer, a result known as the Gottesman-Knill theorem~\cite{NielsenChuang}. The Clifford group can be generated by $\langle \textsc{cz}, H, \sqrt{Z} \rangle$, and contains the Pauli group.
\section{Noise and decoherence}
In the previous section, we saw the basic framework for quantum computation in the absence of noise. In reality, nature is never as idealised as our models, and inevitably the control we have over the quantum states is imprecise and is never completely isolated from the outside world, so there are always ways for unwanted errors to occur. Thus a crucial part of computation is devising a scheme by which we can keep the effects of errors below a tolerable level. Before we discuss how to deal with errors, we first review the tools which are typically used to model the errors that occur.
We start with a simple example which demonstrates how quantum information stored in a qubit is affected by interactions with its surroundings. Consider a single qubit prepared in a pure state $\ket{\phi(0)} = (\alpha\ket{0} + \beta \ket{1})$ (which we call the `system') coupled to another qubit (the `environment') prepared in a $\ket{+}_E$ state. Consider coupling these two qubits via a \textsc{cz} interaction with strength $J$ for time $t$. Following the evolution the system + environment ends up in a state $\ket{\phi(t)}$
\begin{align}
\ket{\phi(t)} &= e^{-i J\; \textsc{cz} \; t / \hbar} (\alpha\ket{0} + \beta \ket{1}) \otimes \ket{+}_E := e^{-i \theta \; \textsc{cz} } (\alpha\ket{0} + \beta \ket{1}) \otimes \ket{+}_E \nonumber\\
& = (\cos \theta \idop -i\sin \theta \textsc{cz})(\alpha\ket{0} + \beta \ket{1}) \otimes \ket{+}_E \nonumber\\
&= \cos \theta (\alpha\ket{0}\ket{+}_E + \beta \ket{1}\ket{+}_E) -i \sin \theta (\alpha\ket{0}\ket{+}_E + \beta \ket{1} \ket{-}_E),
\end{align}
where we have defined $\theta = J t / \hbar$.
We can see that evolving the system + environment for times such that $\theta \neq n\pi$, there is entanglement between the system and the environment, so that to recover the information originally stored in the system we need a combined measurement of both system and environment. However, since the environment is by definition inaccessible, we cannot do this and instead have to reflect our ignorance of the environment by tracing it out altogether (effectively averaging out the effects of the environment). After tracing out the environment, we find that the system is in the state
\begin{align}\label{eqn:DecEx}
\rho_S(t) = \text{tr}_E (\ket{\phi(t)} \bra{\phi(t)}) = | \alpha|^2 \ket{0} \bra{0} + | \beta |^2 \ket{1} \bra{1} + \alpha^* \beta \cos^2\theta \ket{1}\bra{0} + \alpha \beta^* \cos^2\theta \ket{0}\bra{1}.
\end{align}
The result has off-diagonal elements that vary with $\theta$, and go to zero when $\theta = \pi/2$, and the purity of the system varies as $\text{tr} (\rho_S(t)^2) = |\alpha|^4 + |\beta|^4 + 2 |\alpha|^2 |\beta|^2 \cos^4 \theta$.
Thus, depending on the strength of the system-environment coupling, the relative phase between the $\ket{0}$ and $\ket{1}$ states is affected, and the state becomes more mixed. This example is slightly contrived in that the information about the initial state of the system $\ket{\phi}$ is still accessible at times where $\theta = n\pi$, but in a realistic situation the environment will be large and there will be randomness present in the couplings $J$, so that this type of process leads to an irreversible degradation of the information.
This is one way in which connection to an external environment can degrade the quantum information; another way is where the diagonal terms are affected. This is covered in more detail in the next subsection where we look at a useful formalism that captures the behaviour of the system-environment interaction.
\subsection{The operator sum representation}
\label{sec:OSR}
In general we will have a system coupled to a large environment that we have no access to. For any system coupled to an environment, the Hamiltonian $H$ governing the evolution of the qubits plus environment can generally be written
\begin{align}
H = H_{sys} \otimes \idop_E + \idop_{sys}\otimes H_{E} + H_I,
\end{align}
where $H_{sys}$, $H_E$ and $H_I$ are the system, environment and system-environment interaction Hamiltonians, respectively. If we assume that that our starting state has no entanglement between system and environment then the initial state is $\rho(0) = \rho_S \otimes \rho_E$ (this is a reasonable assumption if we assume that it is possible to prepare the system in a pure state, which is a standard requirement for quantum computation anyway). Following the evolution due to $H$, the state at time $t$ is $\rho(t) = e^{-iHt/\hbar} (\rho_S \otimes \rho_E)e^{iHt/\hbar}$. For an observer who only has access to the system, the state at time $t$ is
\begin{align}\label{eqn:rhoS}
\rho_S(t) &= \mbox{tr}_E \left( e^{-iHt/\hbar} (\rho_S(0) \otimes \rho_E(0))e^{iHt/\hbar} \right),
\end{align}
where $\mbox{tr}_E$ indicates tracing out the environment degrees of freedom, appropriate since we assume no access to the information stored in the environment, and so we must average over the possible states of the environment. The transformation in (\ref{eqn:rhoS}) is cumbersome, but through some manipulation we can arrive at a much more useful form.
Assume that the environment is initially in a pure state $\rho_E(0) = \ket{\nu_0} \bra{\nu_0}$ - this is without loss of generality since if it was not in a pure state we could always purify it by adding another system~\cite{NielsenChuang}. We now rewrite equation (\ref{eqn:rhoS}) in terms of $\ket{\nu_0}$ and a complete orthonormal basis of the environment $\{ \ket{\mu} \}$
\begin{align}
\rho_S(t) &= \sum_{\mu} \bra{\mu} e^{-iH t/\hbar} \rho_S(0) \otimes \ket{\nu_0}\bra{\nu_0} e^{iH t/\hbar} \ket{\mu} = \sum_{\mu} \bra{\mu} e^{-iH t/\hbar}\ket{\nu_0} \rho_S(0) \bra{\nu_0}e^{iH t/\hbar} \ket{\mu}\nonumber \\
& := \sum_{\mu} E_{\mu}(t) \rho_S(0) E_{\mu }^{\dagger}(t),
\end{align}
where $E_{\mu }(t) = \bra{\mu} e^{-iH t/\hbar}\ket{\nu_0}$.
This is the \emph{operator sum representation} (OSR) of the evolution of $\rho_S$, and is useful in that it reduces the overall system-bath dynamics to the operators $E_{\mu}$ (usually called Kraus operators) acting only on the system. This transformation is usually written $\rho_S(t) = \mathcal{E} (\rho_S(0) )$, and is a completely positive trace-preserving map, which means that $\sum_\mu E_{\mu}(t)^\dagger E_{\mu }(t) = \idop$ and $\mathcal{E}$ must preserve the non-negativity of the density matrix, even when applied to a subsystem~\cite{NielsenChuang}.
Using this formalism, we can now introduce many different types of noisy channels without worrying about the specific dynamics of the environment. We can think of the qubit interaction with the environment as being like a noisy quantum communication protocol, in which $\rho_S(0) $ is input in one end and $\rho_S(t)$ is output at the other.
One important type of noise is \emph{amplitude damping}, which has Kraus operators~\cite{NielsenChuang}
\begin{align}
&E_0 =\sqrt{p} \left[ \begin{array}{cc}
1 & 0 \\
0 & \sqrt{1 - \gamma}
\end{array} \right]\text{, }
E_1 = \sqrt{p}\left[ \begin{array}{cc}
0 & \sqrt{\gamma} \\
0 & 0
\end{array} \right]\text{, }
E_2 = \sqrt{1-p}\left[ \begin{array}{cc}
\sqrt{1-\gamma} & \\
0 & 1
\end{array} \right] \nonumber\\
&E_3 = \sqrt{1-p}\left[ \begin{array}{cc}
0 & 0 \\
\sqrt{\gamma} & 0
\end{array} \right],
\end{align}
and which has the effect of lowering the population of the $\ket{1}$ state with probability $\gamma$, corresponding to a process such as interaction with a thermal environment (in which case $p$ corresponds to the Boltzmann occupation factor). $\gamma$ typically depends on time, and allows a characteristic time $T_1$ for the decay to be found. If $\gamma$ is an exponential or Gaussian decay then $T_1$ is the time at which $\gamma = 1/e$. In general, the way of experimentally defining $T_1$ will depend on the particular form of the decay.
Another important channel is the phase damping channel, which has Kraus operators
\begin{align}
E_0 =\left[ \begin{array}{cc}
1 & 0 \\
0 & \sqrt{1 - \gamma}
\end{array} \right]\text{, }
E_1 = \left[ \begin{array}{cc}
0 & 0 \\
0 & \sqrt{\gamma}
\end{array} \right]\text{.}
\end{align}
This is in fact the operator-sum form of the example seen in the example in (\ref{eqn:DecEx}), and describes a process where the phase between $\ket{0}$ and $\ket{1}$ is affected (this can happen by e.g.\ random shifting of the energy levels due to the environment). Like the amplitude damping example above, $\gamma$ determines a characteristic dephasing time $T_2$, which is usually taken as the $1/e$ time for Gaussian or exponential decay. $T_2 \leq 2T_1$, and typically $T_1 \gg T_2$, so $T_2$ is the limiting factor for realising quantum computers~\cite{Ladd2010}.
In order to compare different processes later on in this thesis, a slightly modified form of the operator sum representation is needed. First the Kraus operators can be expressed in terms of a complete basis $\{ A_n \}_{n=1}^{2^N}$ which form an orthogonal basis under the Hilbert-Schmidt inner product, i.e.\ $\text{tr}({A}_m^{\dagger} {A}_n) = \delta_{mn}$. For example, we could choose a basis formed from outer products of a set of orthogonal vectors$\{ A_n \} = \{ \ket{l}\bra{m} \}$, or the Pauli group $\mathcal{G}_N$. Expressing the Kraus operators as $E_{\mu} = \sum_n a_{n\mu} A_n$, where $a_{n \mu} = \text{tr}( A_n^\dagger E_{\mu})$, a map $\mathcal{E}$ can be written
\begin{align}
\mathcal{E} (\rho) &=\sum_\mu {E}_\mu \rho {E}_\mu^{\dagger} = \sum_\mu \sum_{mn} (a_{m\mu} a_{n \mu}^{*}) {A}_m \rho {A}_n^{\dagger} := \sum_{mn} \chi_{mn} {A}_m \rho {A}_n^{\dagger},
\end{align}
where in the last line we have defined the \emph{process matrix} $\chi_{mn} := \sum_{\mu} a_{m\mu} a_{n \mu}^{*}$.
\subsection{Quantum error correction}
\label{sec:QEC}
Having introduced some of the tools of quantum error correction, we will look at some of the methods to deal with quantum errors. There are two approaches to this problem: the `hardware' approach, in which we try to minimise the amount of errors that occur in the first place by building the computer in a way that is insensitive to the particular noise, or improving the quality of the components. There is also the `software' approach, where we encode our information in such a way that any errors can be recognised and corrected.
For an illuminating example of the differences between the software and hardware approaches, consider an `$n$-bit-flip' channel through which we can send an $N$-bit message, where the channel flips $n$ bits of the message with probability $p$. One software approach would be to use these $N$ bits to encode a single bit, using a repetition code. Provided $n < N/2$, at the other end of the channel we can just decode the noisy information by taking a majority vote. However, if the channel flips \emph{all} of the bits with probability $p$, then this repetition code is no longer useful. Instead of using a repetition code, we can encode the logical bit into the parity of the $N$ qubits, so that if the channel flips all of the bits the parity is conserved (assuming that $N$ is even, but if $N$ is odd the same can be achieved by ignoring the $N^{th}$ bit). This is the hardware approach, where the encoding of the information itself is naturally immune to errors.
Clearly the sensible tactic in making quantum computers is to develop both approaches simultaneously, since environments will rarely be as predictable as the example give above. In this section we will introduce aspects of both approaches that are relevant to this thesis. We begin with the software approach; one of the simplest examples of this approach is copying the information, and hoping that all copies of the information will not suffer the same errors (a technique familiar to anyone who backs up their computer). Whilst this is a good starting point for classical error correction, in quantum computation we are stymied by the no-cloning theorem~\cite{Dieks1982,Wootters1982}, which prohibits copying of arbitrary quantum states. Fortunately, quantum error correcting codes are possible; Shor developed a quantum analogue of the classical repetition code~\cite{Shor1995} whilst Steane developed a quantum analogue of the classical Hamming code~\cite{Steane1996}. Following these developments, a five-qubit code was discovered by Bennett et.\ al.~\cite{Bennett1996}, and Calderbank, Shor and Steane discovered a scheme whereby any classical code can be made into a quantum code~\cite{Calderbank1996,Steane1996a}. A more general framework for error correction was developed by Gottesman~\cite{Gottesman1997} using stabilisers, which are discussed in more detail in Sec.~\ref{sec:MBQCStab}.
The general feature of error-correcting codes is that there are a number of \emph{physical} qubits, which are the basic physical systems that we have at our disposal (e.g. ions, electrons), and which can be combined together to give \emph{encoded} or \emph{logical} qubits, which are 2-level systems existing in a subspace or subsystem of the physical qubits. As an example, the 3-qubit bit flip code has 3 physical qubits for every encoded qubit: $\ket{0}_L = \ket{000}\text{, } \ket{1}_L = \ket{111}$, and can withstand bit-flip errors on 1 physical qubit.
Using these error correcting codes, we can derive approximate bounds on how accurate each logic gate must be such that the accumulation of errors during a computation is not too large. To calculate this, we can consider a fault-tolerant circuit to construct a fault tolerant encoded \textsc{cnot} gate on two encoded qubits. The gate itself is performed using many imperfect \textsc{cnot} gates acting between physical qubits inside the encoded qubit. For example, a \textsc{cnot} gate for the 3-qubit bit flip code between two encoded qubits $Q_1$, $Q_2$ is:
\[
\Qcircuit @C=1em @R = 1em @!R {
& & &&&& & & \ctrl{4} & \qw & \qw & \qw \\
\lstick{Q_1}&\ctrl{4} &\qw &&&& &\lstick{Q_1} &\qw & \ctrl{4} & \qw & \qw \\
& & &&&& & & \qw & \qw & \ctrl{4} & \qw \\
& & &&:=&& & & & & & \\
& & &&&& & & \targ{X} & \qw & \qw & \qw \\
\lstick{Q_2}&\targ{X} &\qw &&&& & \lstick{Q_2}& \qw & \targ{X} & \qw & \qw \\
& & &&&& & & \qw & \qw & \targ{X} & \qw
\gategroup{2}{2}{6}{2}{.7em}{--}
}
\]
Following the gate, we perform a syndrome measurement to detect any errors, followed by a recovery step to correct these errors:
\[
\Qcircuit @C=1em @R = 2em @!R {
\lstick{Q_1} & \ctrl{1} & \qw & \gate{\text{Syndrome}} & \qw & \gate{\text{Recovery}} & \qw & \qw & \qw \\
\lstick{Q_2} & \targ{X} & \qw & \gate{\text{Syndrome}} & \qw & \gate{\text{Recovery}} & \qw & \qw & \qw \gategroup{1}{2}{2}{2}{.7em}{--}
}
\]
The probability of such a code failing can be calculated given the probability of each individual gate on the physical qubits failing, and finding the number of ways in which two or more errors could occur during the gate, syndrome or recovery steps (see~\cite{NielsenChuang} for an example of this). The result is that the total gate error is $c p^2$, where $c$ is a factor accounting for the number of possible ways two errors on an encoded qubit can occur.
Given that we will need to combine many gates to construct a quantum computer, the question is whether or not a particular gate on encoded qubits is scalable, so that a large number of these gates does not produce an uncorrectable error. In fact it is possible to show~\cite{Preskill1998,Kitaev1997,Aharonov1997,Knill1998,Knill2005} that scalable quantum computation is possible provided the physical gates have an error below a certain value; this is the \emph{threshold theorem}. The particular value of the threshold will depend on the encoding used, the scheme used to combine gates together, and the error model, and such a scheme is called a \emph{fault-tolerant} circuit. The currently known thresholds range from $p \lesssim 10^{-6}$ to up to around $0.01$~\cite{Preskill1998,Kitaev1997,Gottesman1997,Aharonov1997,Knill1998,Steane2003,Knill2005}.
\subsection{Decoherence-free subspace}
\label{sec:DFSintro}
Above we have seen a brief description of the `software' approach to correcting quantum errors. We will now discuss an example of the `hardware' approach. We will assume that we have some noisy physical qubits, and we will encode a single encoded bit of information over several of these physical qubits such that the errors have little or no effect on the information. Such an encoding is called a \emph{decoherence-free subspace} (DF subspace) or more generally a \emph{decoherence-free subsystem} (DF subsystem)~\cite{ChuangYamamoto1995,DuanGuo1997,Lidar1998,ZanardiRasetti1997}.
Consider a system-environment evolution that has OSR operators $E_{\mu} = \bra{\mu} e^{-iH_I t/\hbar}\ket{\nu_0}$, where $H_I$ is the system-environment interaction. If there exists a subspace $\mathcal{S}$ such that for all of the OSR operators
\begin{align}
E_{\mu} \ket{j} = e^{i\phi_{\mu}}\ket{j} \; \; \forall \; \; \ket{j} \in \mathcal{S},
\end{align}
then any state encoded in this subspace at time $t=0$ will not be affected by the system-environment interaction, since the only effect is a global phase which has no noticeable effect. This is a \emph{decoherence-free subspace}, where the full Hilbert space is partitioned into logical states and non-logical states, and every logical state is represented by one orthonormal state. An example of a decoherence free subspace would be the bit flip channel described above which flips all bits together with probability $p$; if $N=2$ then one choice of decoherence-free subspace would be spanned by $\{\frac{1}{2}(\ket{01} + \ket{10}),\frac{1}{2}(\ket{00} + \ket{11} )\}$.
Since we also want to do useful computations on this subspace, we also require that the Hamiltonian used to control the encoded qubit preserves the subspace, which will be true provided the eigenstates of this Hamiltonian are either completely inside or completely outside the decoherence free subspace~\cite{BaconThesis}.
This is not the most general way of encoding information; the most general way to encode information is in a \emph{subsystem}, in which there are also gauge degrees of freedom which can change without affecting the encoded information (so one logical state can be represented by many states of the physical qubits).
The most trivial choice of subsystem encoding for 2 qubits is simply encoding the information in one physical qubit and having the second qubit in an arbitrary state. So $\ket{0_L} := \ket{0}\ket{\phi}$ and $\ket{1_L} :=\ket{1}\ket{\phi}$ are logical operators in the subsystem of qubit 1, and the gauge degree of freedom is the arbitrary choice of the state of qubit 2. Another example is for the `all-bit-flip' channel described above; if $N=2$, the logical $\ket{0}$ states is any combination of the states $\{ \ket{00}, \ket{11} \}$ which have even parity whilst the $\ket{1}$ states are any combination of the states $\{ \ket{10}, \ket{01} \}$ with odd parity. In this case the logical space is the parity of the qubits, whilst the gauge freedom is in the actual values of the physical qubits. For a slightly more abstract example, consider the Bell states on 2 qubits: $\ket{\Phi^{\pm}} = (1/\sqrt{2})(\ket{01} \pm \ket{10} )$, $\ket{\Psi^{\pm}} = (1/\sqrt{2})(\ket{00} \pm \ket{11} )$. We can rewrite these in a tensor product structure as $\ket{\chi^{\lambda}} = \ket{\chi} \ket{\lambda}$, where $\chi = \Phi,\Psi$, $\lambda = +,-$, and then we can think of storing information in the $\chi$ degree of freedom, so that $\lambda$ is the gauge degree of freedom. A subsystem $\mathcal{H}_L$, in which a quantum state $\ket{\phi}$ can be encoded in the state $\rho_{\phi}$ is a \emph{decoherence-free subsystem} if, for all of the OSR operators $E_{\mu}$
\begin{align}
\mbox{tr}_{G} \left[ E_{\mu} \rho_{\phi} (E_{\mu})^\dagger \right] = \ket{\phi} \bra{\phi},
\end{align}
where $\mbox{tr}_{G}$ means tracing out all of the gauge degrees of freedom.
In Chapter~\ref{chap:DFS}, we will focus on a particular kind of system-environment interaction where each physical qubit interacts identically with the environment (e.g.\ by being close enough together with respect to the environment). Such a decoherence model is called \emph{collective decoherence}. We consider collective decoherence acting on $N$ spin-$\frac{1}{2}$ particles, such that the system-environment interaction Hamiltonian $H_I$ can be written as
\begin{align}
H_I = \sum_{\alpha = x,y,z} a_{\alpha} \mathbf{S}_{\alpha}\otimes B_{\alpha},
\end{align}
where $\mathbf{S}_{\alpha} := \sum_{n=1}^N \frac{1}{2}\alpha_{n}$, and ${\alpha}_{n} \in \{ X_n,Y_n,Z_n\}$ is a Pauli operator acting on the $n^{th}$ physical qubit, and $B_{\alpha}$ are operators acting on the environment. Strictly speaking, the decoherence-free subspace/subsystem should be defined using the operator-sum representation, but the algebra of the operator-sum representation for $H_I$ is generated by $\{ \mathbf{S}_{\alpha} \}_{\alpha = x,y,z}$ plus $\idop$~\cite{BaconThesis}, so these definitions cover the operator sum form as well. We have already seen an example of collective decoherence, in the `n-bit flip' channel discussed in Sec.~\ref{sec:QEC}. For the case where the channel probabilistically flips all of the qubits, this is equivalent to collective decoherence where $a_y = a_z = 0, a_x \neq 0$. This form of decoherence, where the collective spin operator only acts along one direction, is called \emph{weak} collective decoherence. The more general case, where $a_x,a_y,a_z \neq 0$ is called \emph{strong} collective decoherence~\cite{BaconThesis}. We will refer to a DF subspace (subsystem) that protects against weak or strong collective decoherence as a weak or strong DF subspace (subsystem) respectively. Clearly a strong DF subspace or subsystem can also be used as a weak DF subspace or subsystem.
To explore how to construct weak and strong DF subspaces and subsystems we first briefly review some relevant spin physics, restricted to spin-$\frac{1}{2}$ particles for this thesis, but the arguments can be extended to other systems. A system with spin is characterised by two quantum numbers: the total spin $s$ and the projection of this spin along one of the $x,y$ or $z$ axes, denoted $m_{x},m_y,m_z$ respectively. Since the spin operators $X$, $Y$ and $Z$ don't commute, only one projection can be specified at a time, and we take this to be $m_z$ without loss of generality. $m_z \hbar$ is the eigenvalue of the operator $\frac{1}{2} Z$ and can take any of the $(2s+1)$ values in the range $-s,-s+1,...,s-1,s$, whilst $s(s+1)\hbar^2$ is the eigenvalue of the operator $\frac{1}{4}(X^2+Y^2+Z^2)$. When combining a number of particles with total spin $s_1,s_2,s_3,...,s_N$ together the resultant total spin $S$ can take any of the values of the combinations of the spins, $S = | s_1 \pm s_2... \pm s_N|$. Just as for single spins, for a collection of spins there are spin operators $\mathbf{S}_{x},\mathbf{S}_y,\mathbf{S}_z$ (defined above) and $\mathbf{S}^2 = \mathbf{S}_x^2 + \mathbf{S}_y^2 + \mathbf{S}_z^2$, so that states can be labelled as $\ket{S,m_S}$ such that $\mathbf{S}_z \ket{S,m_S} = \hbar m_S$ and $\mathbf{S}^2 \ket{S,m_S} = \hbar^2 S(S+1) \ket{S,m_S}$.
There can also be more than one way to reach the same value of $S$, and so there can be degeneracy in the system with respect to the $S$ value. A useful way of depicting the degeneracies from combining spin in this way was given in~\cite{BaconThesis}, whereby addition of spin is represented as addition of vectors (see Fig.~\ref{fig:SpinAddition}), and the degeneracies of each spin number $S$ is given by Pascal's triangle. Note that for each value of $S$, there are $(2S+1)$ states with different values of the quantum number $m_S$.
To construct a weak DF subspace, where there is only one operator such as $\mathbf{S}_z$ in the system-environment interaction, all that is required are states with the same value of $m_S$. So for example we can construct a 2-qubit weak DF subspace using the $\ket{S=1,m_S=0} = \frac{1}{\sqrt{2}}(\ket{10} + \ket{01})$ and $\ket{S=0,m_S=0} = \frac{1}{\sqrt{2}}(\ket{10} - \ket{01})$ states, and for three qubits we could use the following states used in~\cite{DiVincenzo2000}:
\begin{align}
&\ket{S = 1/2, m_S = 1/2 , S_{1,2} = 0 } = \ket{\psi^-}_{12}\ket{0}_3 \nonumber\\
&\ket{S =1/2,m_S =1/2, S_{1,2} = 1} = \frac{1}{\sqrt{3}}( \sqrt{2} \ket{T_+}_{12}\ket{1}_3 - \ket{T_0}_{12}\ket{0}_3)
\end{align}
Here we have used the singlet states on qubits $a$ and $b$, defined as $\ket{\psi^-}_{ab} := (\ket{01}_{ab} - \ket{10}_{ab})/\sqrt{2}$ and the triplet states $\ket{T_+}_{ab} = \ket{00}_{ab}$, $\ket{T_0}_{ab} = (\ket{01}_{ab} + \ket{10}_{ab})/\sqrt{2}$. Note that, since the states are degenerate with respect to the $S$ and $m_S$ value, we distinguish between them using a third quantum number $S_{1,2}$ defined as the total spin quantum number of spins 1 and 2 only.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.45\textwidth]{SpinAddition-eps-converted-to.pdf}
\caption{\label{fig:SpinAddition} An illustration of the degeneracies in total spin value $S$ when combining $N$ spins. }
\end{center}
\end{figure}
Finding strong DF subspaces and subsystems is more difficult. With all three of $\mathbf{S}_x$, $\mathbf{S}_y$ and $\mathbf{S}_z$ present in the system-environment interaction, since $[ \mathbf{S}_\alpha, \mathbf{S}_\beta] = i \epsilon_{\alpha \beta \gamma} \mathbf{S}_\gamma$ and $[ \mathbf{S}^2, \mathbf{S}_\alpha] = 0$ for $\alpha,\beta, \gamma \in \{ x,y,z\}$, the noise acts to mix states with the same values of $S$ but different values of $m_S$. Thus the 2- and 3-qubit encodings shown above are not suitable, as leakage into other states needs to be taken into account. Note that also degenerate states with the same value of $S$ are not mixed, since the total spin number $S_{q_1,q_2,...,q_n}$ on some subset of size $n <N$ of the qubits will differ between different degenerate states, and following the same commutation arguments the eigenvalue of $\mathbf{S}_{q_1,q_2,...,q_n}$ will not be changed by collective decoherence.
To construct a strong DF subspace requires two states with the same value of $m_S$, with no other states that can be evolved to under the collective Pauli operators. This will only be the case when $S=0,m_S=0$, and one can see from Fig.~\ref{fig:SpinAddition} that $N=4$ is the lowest $N$ for which there are two $S=0,m_S=0$ states, so this is the smallest number of spins over which a qubit can be encoded in a strong DF subspace, with logical states of the following form:
\begin{align}\label{eqn:4States}
\ket{\bar{0}^{(4)}} & := \ket{ S = 0, m_S=0, S_{1,2} = S_{3,4} = 0} = \ket{\psi^-}_{12} \ket{\psi^-}_{34}\nonumber\\
\ket{\bar{1}^{(4)}} & := \ket{ S = 0, m_S=0, S_{1,2} = S_{3,4} = 1} = \frac{1}{\sqrt{3}} \left[ \ket{T_+}_{12} \ket{T_-}_{34} - \ket{T_0}_{12} \ket{T_0}_{34}+ \ket{T_-}_{12} \ket{T_+}_{34}\right].
\end{align}
This 4-qubit encoding has one additional desirable property; it also functions as a \emph{supercoherent qubit}~\cite{Bacon2001}. Supercoherence would allow resistance to errors acting on individual physical qubits, with a mechanism as follows: When an error along any direction is applied to the physical qubits in the $S=0$ states in eqn.\ (\ref{eqn:4States}), it is accompanied by a change in the $S$ value by 1~\cite{Bacon2001}. In order to use this to create a supercoherent qubit, we could switch on the Hamiltonian ${H}_{SC}$, defined as
\begin{equation}\label{eqn:Hsc}
{H}_{SC} = J_{SC}\sum_{m,n} {E}_{mn},
\end{equation}
where ${E}_{mn} := X_m X_n +Y_m Y_n + Z_m Z_n$, and the sum is over all pairs of the 4 physical qubits. With this Hamiltonian switched on, the $S=0$ states are degenerate and lowest in energy, with an energy gap between the $S=0$ states and any other states. Thus any decoherence process acting on individual physical qubits in the $S = 0$ state involves an increase in energy of the encoded qubit, and will lead to a transfer of energy from the environment to the system, which we can inhibit by cooling the environment. Thus supercoherent qubits would be very useful as quantum memories, and it was argued in~\cite{Bacon2001} that computation with supercoherent qubits could be performed provided the interaction strength between qubits was small enough compared to $J_{SC}$ (leading to a trade-off between the speed of operations and the robustness against errors). In this chapter, we will not aim to make our interactions supercoherent as well (i.e.\ we envisage a protocol in which we use the supercoherent mechanism as a means to reliably store information, but turn off the supercoherent Hamiltonian ${H}_{SC}$ when we interact encoded qubits).
In~\cite{BaconThesis} it is shown that strong DF subsystems can be constructed whenever there is a degeneracy in the $S$ value. From Fig.~\ref{fig:SpinAddition} we can see that the smallest number of qubits with this property is $N=3$, for which there are two $S = \frac{1}{2}$ states. Thus we can encode a qubit into the strong DF subsystem on the following states:
\begin{align}\label{eqn:3states}
&\ket{\bar{0}^{(3)}_{+1} }=\ket{S = 1/2, m_S = 1/2, S_{1,2} = 0 } = \ket{\psi^-}_{12}\ket{0}_3\nonumber\\
&\ket{\bar{0}^{(3)}_{-1}}=\ket{S =1/2,m_S =-1/2, S_{1,2}=0} = \ket{\psi^-}_{12}\ket{1}_3\nonumber\\
&\ket{\bar{1}^{(3)}_{+1}}=\ket{S =1/2,m_S =1/2, S_{1,2}=1} = \frac{1}{\sqrt{3}}( \sqrt{2} \ket{T_+}_{12}\ket{1}_3 - \ket{T_0}_{12}\ket{0}_3)\nonumber\\
&\ket{\bar{1}^{(3)}_{-1}}=\ket{S =1/2,m_S =-1/2, S_{1,2}=1} = \frac{1}{\sqrt{3}}( \ket{T_0}_{12}\ket{1}_3 - \sqrt{2} \ket{T_{-}}_{12}\ket{0}_3).
\end{align}
where $\ket{T_-}_{ab} = \ket{11}_{ab} $. The logical zero state $(\ket{\bar{0}^{(3)}})$ in this 3-qubit subsystem is defined to be an arbitrary combination of the first two states, i.e.\ $\ket{\bar{0}^{(3)}} := \zeta |\bar{0}^{(3)}_{+1} \rangle + \gamma | \bar{0}^{(3)}_{-1} \rangle $, whilst the logical one state $(\ket{\bar{1}^{(3)}})$ is a superposition of the last two states with the same coefficients, $\ket{\bar{1}^{(3)}} := \zeta |\bar{1}^{(3)}_{+1} \rangle + \gamma |\bar{1}^{(3)}_{-1}\rangle$. The action of collective decoherence on this encoding will act identically on the two states, and so will change the values of $\zeta$ and $\gamma$ identically for both logical states, but will not mix the different logical states since they have different values of $S_{1,2}$. The arbitrary choice of $\zeta$ and $\gamma$ is the gauge degree of freedom, and any transformation which only changes the values of $\zeta$ and $\gamma$ is called a gauge transformation (in this case a gauge transformation is an operation which couples to the $m_S$ degree of freedom). We refer to the separate subspaces with different $m_S$ values as {\it gauge subspaces}. An illustration of this subsystem encoding is shown in Fig.~\ref{fig:DFsubsystem}, comparing the the action of strong collective noise on all of the possible states with 3 physical qubits.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.8\textwidth]{DFsubsystem-eps-converted-to.pdf}
\caption{\label{fig:DFsubsystem} An illustration of the action of strong collective decoherence on the degenerate $S = \frac{1}{2}$ states and the non-degenerate $S = \frac{3}{2}$ states, for 3 qubits. The split lines indicate the different quantum numbers characterising different states, and the dashed ellipsoids indicate which states are mixed by the action of strong collective decoherence.}
\end{center}
\end{figure}
It has been shown~\cite{Lidar1998,ZanardiRasetti1997,Bacon2000} that universal quantum computation can be performed inside a DF subspace or subsystem. To do this, we must be able to perform certain single qubit rotations as well as gates between two encoded qubits (such as a controlled-Z gate)~\cite{Barenco95}. For collective decoherence, explicit gate sequences for two-qubit gates have been found for the 3-qubit DF subsystem and 4-qubit DF subspace~\cite{Bacon2000,DiVincenzo2000,BaconThesis,Hsieh2003,Fong2011}. Searching for these gates will be the subject of Chapter~\ref{chap:DFS}. For a more in depth discussion of DF subspaces and subsystems, the reader is referred to~\cite{Lidar1998,Lidar2003,BaconThesis,DuanGuo1998,PalmaSuominen1996,Kempe2001}.
\section{Distance measures}
\label{sec:DistMeas}
Later on in this thesis we will be attempting to create two-qubit and three-qubit gates, and so it will be useful to have some measures to evaluate how close a quantum state is to another quantum state, and how close a given unitary is to an ideal gate. Performing these two different tasks can be done using some of the same machinery we have introduced above for error correction.
One of the standard measures of distance between two quantum states $\rho_1$ and $\rho_2$ is the fidelity $F(\rho_1,\rho_2) $, which measures how close these quantum states are:
\begin{align}
F(\rho_1,\rho_2) = \left( \text{tr} \sqrt{ \sqrt{\rho_1} \rho_2 \sqrt{\rho_1} } \right)^2,
\end{align}
which is 1 iff $\rho_1 = \rho_2$. For pure states $F(\ket{\phi_1}\bra{\phi_1},\ket{\phi_2}\bra{\phi_2}) = |\sprod{\phi_1}{\phi_2}|^2$. Note that we have squared the trace, which is different from the form seen in e.g.~\cite{NielsenChuang}. Both forms with and without the square are used in the literature, but when we later on discuss how to construct useful metrics for gate distance, this squaring will allow the fidelity to have the interpretation of a probability~\cite{Gilchrist2005}, which is more intuitive than the square root of probability. In cases where the literature uses the non-squared version, we will also use it.
The second important measure is the trace distance, which measures how far apart two quantum states $\rho_1$ and $\rho_2$ are
\begin{align}
D(\rho_1,\rho_2) = \frac{1}{2} \| \rho_1 - \rho_2 \|_{tr},
\end{align}
where $\| A \|_{tr} := \text{tr}(\sqrt{A A^\dagger})$ is the trace norm.
To use these metrics to measure the distance between two gates, we can use the Jamiolkowski isomorphism~\cite{JamiolKowski1972}. If $\mathcal{E}$ is a quantum operation acting on a $d$-dimensional system, and $\ket{\Phi} = \frac{1}{\sqrt{d}} \sum_{j=1}^{d }\ket{j}\ket{j}$ is an entangled state of two $d$-dimensional systems ($d=2$ for qubits), then we can construct a state $\rho_{E}$ where $\rho_{\mathcal{E}} =[ \idop \otimes \mathcal{E} ](\ket{\Phi}\bra{\Phi})$, i.e.\ where the first system is left untouched but the second system transforms by $\mathcal{E}$. If we express $\mathcal{E}$ in terms of a basis $A_n = \ket{j_n}\bra{j'_n}$ such that $(\ket{j_n},\ket{j_n'})$ is a unique pairing of the possible orthogonal computational basis states of $N$ qubits, then $\mathcal{E}(\rho) = \sum_{mn} \chi_{mn} A_m \rho A_n^\dagger$ (see Sec.~\ref{sec:OSR}), and $\rho_{\mathcal{E}}$ becomes
\begin{align}
\rho_{\mathcal{E}} &= \frac{1}{d} \sum_{jkmn}\chi_{mn} (\idop \otimes A_m) \ket{jj}\bra{kk} (\idop \otimes A_n^\dagger) \nonumber\\
&=\frac{1}{d} \sum_{jkmn}\chi_{mn} \delta_{j j'_m} \ket{jj_m}\bra{kk_n} \delta_{k k_n'}= \frac{1}{d} \sum_{mn}\chi_{mn} \ket{j'_m j_m}\bra{k'_n k_n}.
\end{align}
Since $(j_n ,j_n')$ and $(k_n,k_n')$ are unique pairs, by construction, this means that each element of $\chi$ is mapped to a unique element in $\rho_{\mathcal{E}}$. If in addition $A_n = \ket{x_1 x_2 ... x_N}\bra{x_{N+1}x_{N+2}...x_{2N} }$ such that the binary string $x_1 x_2 ... x_Nx_{N+1}x_{N+2}...x_{2N}$ is $n$ in binary notation, then $(\rho_{\mathcal{E}})_{mn} = \frac{1}{d}\chi_{mn}$.
Using this isomorphism, we can then compare the process matrix $\chi_{id}$ of the ideal process with the actual process $\chi$ using the fidelity and trace distance~\cite{Gilchrist2005}. Thus we define the process fidelity $F_{pro}$ between two process matrices $\chi_1$, $\chi_2$ as
\begin{align}
F_{pro} = \frac{1}{d^2} \left( \text{tr} \sqrt{ \sqrt{\chi_1} \chi_2 \sqrt{\chi_1} } \right)^2,
\end{align}
whilst the process distance is defined as
\begin{align}
D_{pro} = \frac{1}{d}\| \chi_1 - \chi_2 \|_{tr}.
\end{align}
$D_{pro}$ and $1- F_{pro}$ are particularly useful as they act as upper bounds on $\bar{p}_e$, the average probability that the evolution does not produce the desired output~\cite{Gilchrist2005}.
Often it is only important to know when two processes are equivalent up to some local single-qubit operations (since these are typically easy to perform). Such gates which are equivalent up to local single-qubit operations are called \emph{locally equivalent}. $D_{pro}$ and $F_{pro}$ are not appropriate for detecting gates that are locally equivalent, so in certain situations the number of search parameters can be significantly reduced if we can use a distance measure which ignores local operations. One method to do this due to Makhlin~\cite{Makhlin2002}. For a two qubit operation $M$, we first transform ${M}$ into the Bell basis, ${M}\rightarrow {M}_B = {Q}^{\dagger}{M} {Q}$ where
\begin{align}
{Q} = \frac{1}{\sqrt{2}} \left( \begin{array}{c c c c} 1 &0& 0& i \\ 0 &i &1 &0 \\ 0 &i &-1& 0\\ 1& 0& 0& -i \end{array} \right).
\end{align}
Makhlin showed that for two gates $L$ and $M$ expressed in the Bell basis as $L_B$ and $M_B$, the gates are equivalent up to local operations iff the spectra of $L_B^T L_B$ and $M_B^T M_B$ are the same. This follows since local operations correspond to orthogonal operations in the Bell basis, and so for any matrix $M_B$ in the Bell basis the spectrum of $M_B^T M_B$ after a local transformation $M_B \to O_1 M_B O_2$ is conserved. The spectrum can also be characterised by two quantities $m_1$ and $m_2$ defined as:
\begin{align}
m_1({M}) &= [\textrm{tr}\; (M_B^{T}M_B)]^2/16\det {M}^{\dagger},\nonumber \\
m_2({M}) &=[(\textrm{tr}\; (M_B^{T}M_B))^2 -\textrm{tr}((M_B^{T}M_B)^2) ]/4\det {M}^{\dagger}.
\end{align}
so that two gates $M$ and $L$ are locally equivalent iff $\{m_1(L),m_2(L)\}=\{m_1(M),m_2(M)\}$. To use these invariants, it is useful to combine them to get a single number which is 0 for identical gates and larger for very different gates. To do this we define
\begin{equation}
f_m(L,M) = \sum_{i=1}^2 \left|m_i(L) - m_i(M) \right|.
\end{equation}
This distance measure compares the spectra of the two different gates $M$ and $L$ up to some local operations, so can be thought of in a similar manner to the trace distance. The same measure of distance between gates has been used in e.g.~\cite{DiVincenzo2000}.
\section{Measurement-based quantum computation}\label{sec:MBQC}
So far we have discussed quantum computation in the `standard' circuit model picture, where operations are performed sequentially on some systems containing the information, and the result is read out at the end. This is analogous to the standard way in which classical computation is performed, and is perhaps the most intuitive way to envision a quantum computer. Unsurprisingly, quantum computation allows computations to be performed in ways that have no analogy in classical computation. For instance, instead of measurements only occurring after the logical operations have been performed, we can use the measurements to perform the computation. This is the paradigm of \emph{measurement-based quantum computation} (MBQC)~\cite{Raussendorf2001}, in which measurements on a highly entangled state are used to perform computations, rather than unitary gates.
As an illuminating example, consider the simplest MBQC setup: a system of two qubits, A and B. A contains some unknown quantum information $\ket{\chi} = \alpha \ket{0} + \beta \ket{1}$, whilst qubit B is prepared in the state $\ket{+}$. To ensure the measurement on A has an effect on B, the qubits are first entangled using a \textsc{cz} gate, resulting in the state $\ket{\psi_{AB}}= \alpha \ket{0}\ket{+} + \beta \ket{1} \ket{-}$. This state $\ket{\psi_{AB}}$ can be rewritten as
\begin{eqnarray}
\label{eqn:2qubitEx}
\ket{\psi_{AB}}= &\frac{1}{\sqrt{2}} \ket{+_{\phi}}( \alpha\ket{+} + e^{-i\phi} \beta \ket{-}) + \frac{1}{\sqrt{2}} \ket{-_{\phi}}( \alpha\ket{+} -e^{-i\phi} \beta \ket{-}) \nonumber\\
&= \frac{e^{-i\phi/2} }{\sqrt{2}} \ket{+_{\phi}} \text{H} U_z(\phi) \ket{\chi} +\frac{e^{-i\phi/2} }{\sqrt{2}} \ket{-_{\phi}} X \text{H} U_z(\phi) \ket{\chi},
\end{eqnarray}
where $ \ket{\pm_{\phi}} := \frac{1}{\sqrt{2}} ( \ket{0} + e^{i\phi} \ket{1})$ and $\text{H}$ is a Hadamard gate. Now consider performing a measurement on qubit A, such that the measurement has eigenstates $\ket{\pm_{\phi}}=1/ \sqrt{2} ( \ket{0} \pm e^{i\phi} \ket{1})$. If the outcome is $\ket{+_{\phi}}$, the state of qubit B is $e^{-i\phi/2}\text{H} U_z(\phi) \ket{\chi}$, and if the outcome is $\ket{-_{\phi}}$ the state of qubit B is $e^{-i\phi/2} X \text{H} U_z(\phi) \ket{\chi}$. This can be written in a more compact form by assigning a variable $m$ to the outcome of the measurement on $A$, such that $m=0$ if the outcome is $\ket{+_{\phi}}$ and $m=1$ if the outcome is $\ket{-_{\phi}}$. Then the overall result is that the information is moved to site $B$ and transformed by $e^{-i\phi/2} X^{m}\text{H} U_z(\phi) $. By applying a Pauli $X^m$ correction on qubit $B$, both outcomes are the same, and so a deterministic $\text{H} U_z(\phi)$ gate can be obtained (up to a global phase) despite the randomness of the measurement. Note that by preparing A in state $\ket{\chi}$ instead of $\ket{+}$, information can be encoded in the system, so we call A the input, and since B is the system where the information will be at the end of the computation, we call this the output.
Thus we have seen how to perform the operation $\text{H} U_z(\phi)$ on the encoded information in this 1D chain whilst teleporting the information from one qubit to the other. This simple example shows the main features of the MBQC protocol:
\begin{itemize}
\item[1.] Non-input qubits are prepared in $\ket{+}$ states.
\item[2.] Entangling \textsc{cz} gates are performed between certain pairs of qubits.
\item[3.] Non-output qubits are measured. The measurement basis determines which operations are performed.
\item[4.] To make the outcome deterministic, corrections that depend on previous measurement outcomes are needed.
\end{itemize}
To make the following discussion easier to understand, we introduce a graphical representation of MBQC which is commonly used in the literature. We represent qubits as vertices $V$ in a graph $G$. The input qubits $I$ are represented as vertices contained within squares, whilst output qubits $O$ are represented as hollow circles. Non-inputs are prepared in the $\ket{+}$ state, and the edges of this graph $E$ represent which pairs of qubits have been acted on by the \textsc{cz} gate (we use the notation $v \sim w$ to indicate that $v$ and $w$ are connected by an edge). The state resulting from these operations is called a {\it graph state}. We label each non-output vertex with a pair of angles $(\theta,\phi)$ corresponding to the type of measurement performed on that qubit (where applicable); the angles give the measurement axis in the Bloch sphere (so e.g.\ $(\pi/2,0)$ denotes measurement in the $X$ basis, $(\pi/2,\pi/2)$ in the $Y$-basis, etc.). An illustration of the simple 2-qubit example expressed as a graph is shown in Fig.~\ref{fig:SingleRotation}. After the measurement, the entanglement between qubits A and B is `used up', so that there is no longer an edge between them.
\begin{figure}
\begin{center}
\subfloat[]{\includegraphics[width=0.2\textwidth]{SingleRotation1-eps-converted-to.pdf}}
\qquad
\subfloat[]{\includegraphics[width=0.5\textwidth]{SingleRotation2-eps-converted-to.pdf}}
\caption{\label{fig:SingleRotation} a) A graphical depiction of the state in eqn.\ (\ref{eqn:2qubitEx}). b) A graphical depiction of a state that can be used to perform single qubit rotations.}
\end{center}
\end{figure}
The 2-qubit example above can be extended to perform arbitrary single qubit rotations. Starting with a linear chain of four qubits, entangled to nearest-neighbours only, the first three qubits are measured, resulting in the transformation
\begin{align}
&X^{m_3} \text{H} U_z(\phi_3)X^{m_2}\text{H} U_z(\phi_2)X^{m_1}\text{H} U_z(\phi_1) = X^{m_3}H U_z(\phi_3)X^{m_2}U_x(\phi_2)Z^{m_1}U_z(\phi_1) \nonumber\\
&=X^{m_1+m_3}Z^{m_2}H U_z((-1)^{m_2}\phi_3)U_x((-1)^{m_1}\phi_2)U_z(\phi_1),
\end{align}
where we have used the identities $\text{H}\text{H} = \idop$, $\text{H}X = Z\text{H}$, $U_z(\phi)X = XU_z(-\phi)$, $U_x(\phi)Z = ZU_x(-\phi)$. In this last line, we recognise the form of a general rotation in the Bloch sphere, with some added terms to account for the randomness of the measurement.
The factors such as $(-1)^{m_1}$ mean that the measurements must be performed \emph{adaptively}, and also it imposes an order in which the measurements can be done ($1 \to 2 \to 3$). This is a key property of MBQC, which prohibits simply measuring all the qubits at the same time, and we call the particular order of measurements on a graph state a \emph{measurement pattern}. The terms such as $X^{m_1}$ amount to corrections on the output state, and can be treated as extra classical processing that must be done in order to get a deterministic output.
A $\textsc{cnot}$ gate can be performed starting with qubits arranged as shown in Fig.~\ref{fig:CNOT1}. After measuring the non-output qubits from left to right, a $\textsc{cnot}$ gate is performed, plus some Hadamard operations (see~\cite{Browne2011} for details). We therefore have building blocks with which universal quantum computation is possible, and so by piecing these blocks together a graph-state can be built that is a universal resource for quantum computation. Clearly any graph state which can be made by entangling this universal resource with extra qubits is also universal; for instance starting with a graph which is a 2-dimensional grid of qubits, it is possible to strip away some of the qubits (by measuring them and performing corrections after these measurements) to realise a universal resource. This 2-dimensional rectangular graph-state is known as a \emph{cluster state}, and is the most commonly used universal resource for MBQC, due to its simplicity. There are many other regular graphs that are known to be universal, such as hexagonal and triangular graphs~\cite{VanDenNest2006}, however a complete classification of all graph states which are universal does not exist, and is part of the motivation for this work. We will see in the next section that there are tools that allow us to efficiently determine whether certain graphs can used for universal deterministic MBQC or not.
\begin{figure}[h]
\begin{center}
\subfloat[]{\includegraphics[width=0.25\textwidth]{CNOT-eps-converted-to.pdf}} \hspace{5mm}
\subfloat[]{\includegraphics[width=0.5\textwidth]{ClusterState-eps-converted-to.pdf}}
\caption{\label{fig:CNOT1} a) A graph state to achieve a \textsc{cnot} gate. b) A cluster state.}
\end{center}
\end{figure}
\subsection{\emph{Flow} and \emph{gflow}}\label{sec:f&g}
\label{sec:FlowGflow}
In the previous section we have demonstrated that measurements can be performed adaptively on certain simple graph states, such that the outcome of the computation will always be the same regardless of the measurement outcomes (i.e.\ the outcome is deterministic). There are many more possible graphs, however, and it is not easy to tell whether a given graph will allow such a procedure or not. Fortunately, there are simple graphical tools called \emph{flow}, or more generally, \emph{generalised-flow} (\emph{gflow})~\cite{Danos2006,Browne2007}, which allow an efficient characterisation of whether or not there is an adaptive measurement pattern on a graph state that produces a deterministic outcome (although they do not reveal whether or not the graph allows universal quantum computation). \emph{Flow} and \emph{gflow} provide sufficient (although not necessary) conditions for the existence of such a measurement pattern, and have proved incredibly useful in studying parallelism~\cite{Broadbent2009,Browne2011}, the translation between MBQC and the circuit model~\cite{Broadbent2009,Silva13} and causal order in MBQC~\cite{Raussendorf2011}.
For the purposes of this thesis, we will only use \emph{gflow} since it incorporates \emph{flow}. \emph{Gflow} has two components: a time ordering over the vertices (using the notation $v > w$ to represent that vertex $v$ is measured after vertex $w$, and $v=w$ to indicate that $v$ and $w$ can be measured at the same time), and a \emph{gflow function} $g$ for each vertex. $g(v)$ is a list of the vertices that are affected by the measurement outcome of vertex $v$ (and is therefore a list of the qubits that can be used to correct for this measurement outcome). We say that a set of vertices $U$ is oddly (evenly) connected to a vertex $v$ if there is an odd (even) number of edges connecting $U$ and $v$. The definition of \emph{gflow} is then~\cite{Browne2007}
\begin{defin}[Gflow]
Given an open graph state $G$ with inputs $I$, outputs $O$, edges $E$ and vertices $V$, we say it has \emph{gflow} if there exists a \emph{gflow function} $g$ and a time ordering $<$ over $V$ such that, for all $v \in V$ which are not outputs:
\begin{itemize}
\item[(G1)] All qubits $w$ in $g(v)$ are in the future of $v$, i.e. $v < w$ for all $w \in g(v)$.
\item[(G2)] if $w \leq v$, and $v \ne w$, then $w$ is evenly connected to all qubits in $g(v)$.
\item[(G3)]
\begin{itemize}
\item For measurements in the $(X,Y)$ plane: $v\notin g(v)$, and $g(v)$ is oddly connected to $v$.
\item For measurements in the $(X,Z)$ plane: $v\in g(v)$, and $g(v)$ is oddly connected to $v$.
\item For measurements in the $(Y,Z)$ plane: $v\in g(v)$, and $g(v)$ is evenly connected to $v$.
\end{itemize}
\end{itemize}
\end{defin}
In this thesis, we will only use measurements in the $(X,Y)$ plane, as results for measurements in other planes should follow in a similar fashion. Note that finding out whether or not a graph has \emph{gflow} can be done in polynomial time~\cite{Mhalla}.
To represent \emph{gflow}, arrows can be superimposed on a graphical representation of a graph state, such that an arrow connects $v$ to $w$ if $w \in g(v)$. We call these arrows \emph{gflow lines} (see for example Fig.~\ref{fig:gflow}).
We have seen above that \emph{gflow} requires a time ordering in the measurement pattern, which leads to the definition of {\em layers}, which are groups of vertices which can be measured at the same time:
\begin{defin}[Layers]
A layer of a computation is defined as any non-output qubits in a measurement pattern which can be measured at the same time.
\end{defin}
We denote the layers as $L_k$, and we use $L(v)$ to denote the layer that vertex $v$ is in. For example, for the \emph{gflow} defined on the graph in Fig.~\ref{fig:gflow}, the layers are $L_1 = \{ 1\}$, $L_2 = \{2\}$, $L_3 = \{3\}$. In Fig.~\ref{fig:Example}, the layers are given by $L_k = \{ a_k,b_k,c_k,d_k,e_k\}$, for $k < 6$, and $L(a_k) = L(b_k) = L_k$ etc.
Using the definition of layers, we can also define the depth for MBQC;
\begin{defin}[Depth]
The depth of an MBQC with \emph{gflow} is the number of rounds of measurements in the measurement pattern, or equivalently the number of layers in a measurement pattern.
\end{defin}
In general this depth will be different depending on which \emph{gflow} we are using (there can be more than one - indeed we will see below an example where many \emph{gflow}s can be realised). Since we can think of \emph{gflow} as a directed graph superimposed on an undirected graph, an equivalent and perhaps more intuitive definition is that the depth is the longest possible path along the \emph{gflow} lines. For example, in Fig.~\ref{fig:Example} the depth is 5.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.45\textwidth]{InfluencingVolume2-eps-converted-to.pdf}
\caption{\label{fig:Example} An illustration of the definitions in Sec.~\ref{sec:f&g}, applied to a cluster state with depth 5. Inputs are represented by vertices with squares, and outputs are vertices with hollow circles. Arrows indicate \emph{gflow} lines and the dotted circle indicates a single layer of qubits.}
\end{center}
\end{figure}
\subsection{Graph states in the stabiliser formalism}
\label{sec:MBQCStab}
Graph states can also be described using \emph{stabilisers}~\cite{Gottesman1997}. Stabilisers form a subgroup of the Pauli group $\mathcal{G}$ (excluding $- \idop$) which act as identity on a particular state or subspace. We say a set of stabilisers $\mathcal{K}_{\mathcal{S}}$ stabilise the subspace $\mathcal{S}$ if
\begin{eqnarray} \label{eqn:GSStabEQN}
K \ket{\phi} = \ket{\phi} \text{ } \forall K \in \mathcal{K}_{\mathcal{S}}, \text{ } \forall \ket{\phi} \in \mathcal{S}.
\end{eqnarray}
From this definition we can see that stabilisers must also commute with each other, so they form an Abelian subgroup of the Pauli group. The smallest set of stabilisers that generate the group of stabilisers for a particular subspace are called the \emph{generators} of that group. Each stabiliser generator constrains the space by a half; with $n-k$ generators for the stabiliser the dimension of the codespace is therefore $2^n / 2^{n-k} = 2^k$, i.e.\ the codespace contains $k$ qubits~\cite{PreskillNotes}. Note that the choice of generators of a stabiliser group are not unique.
The stabiliser generators for a graph state are
\begin{eqnarray} \label{eqn:GSStab}
K_{v} &= X_{v} \prod_{w \sim v} Z_w,\text{ } \forall v \notin I,
\end{eqnarray}
where $X_v$, $Y_v$, $Z_v$ are the Pauli matrices acting on site $v$. In this case, since we do not have stabilisers with indices $v \in I$ where $I$ are the input vertices, the number of encoded qubits equals the number of input qubits, as required. This type of graph, where the inputs are unconstrained is referred to as an \emph{open} graph-state.
The evolution of a computation can be followed in terms of these stabiliser operators, rather than by following the evolution of the state itself. This is the Heisenberg picture of quantum computation, and provides an alternative perspective on quantum computation allowing some results, such as the Gottesman-Knill theorem, to be readily apparent.
To follow how MBQC progresses in the stabiliser picture we first define some logical operators that determine what state the system is prepared in. In the case of open graph states, logical $X$ operators can be taken from the stabilisers acting on the inputs that were left out in (\ref{eqn:GSStab}). For example, a 1D cluster state, has stabiliser generators
\begin{align}
K_v &=Z_{v-1} X_{v} Z_{v+1} \quad v = 2,...,N-1; \; \;K_N =Z_{N-1} X_{N}.
\end{align}
The logical $X$ operator can then be made from the missing stabiliser generator $X_L := K_1 = X_1 Z_2$. $Z_L$ can be chosen as $Z_1$ so that it obeys the commutation relations with $X_L$ and the other stabiliser generators ($Y_L$ is usually left out since it can be inferred from the product $X_LZ_L$). Preparing the input in a logical $\ket{\pm}_L$ state is then the same as including $\pm X_L$ as one of the stabilisers.
Now suppose we perform a measurement $M$ on this cluster state in the $X-Y$ basis; this measurement will not commute with all of the logical operators, and so the form of the logical operators will not be preserved after this measurement. However, since stabilisers act as identity on the codespace, the logical operators can be multiplied by any stabilisers which anticommute with the measurement, to put them in a form which is conserved during the measurement. So for example, consider measuring the first qubit in the 1D cluster state in the $X$-basis, so that $\{Z_L,M\}=\{Y_L,M\}=0$. The stabiliser $K_2$ also anticommutes with $M$ so multiplying $Z_L$ and $Y_L$ by $K_2$ gives new logical operators $\tilde{Z}_L = X_2 Z_3$ and $\tilde{Y}_L = X_1 Y_2 Z_3$ which do commute with $M$, and so are conserved during the measurement (i.e.\ if the system is prepared in a particular eigenstate of $\tilde{Z}_L$, it will remain in this eigenstate after the measurement). Overall the transformation of the stabilisers is
\begin{align}
X_L \to X_1 Z_2\text{, }Y_L \to X_1 Y_2 Z_3\text{, }Z_L \to X_2 Z_3.
\end{align}
Qubit 1 will either be in a $\ket{+}$ state or a $\ket{-}$ depending on the outcome of the measurement. We can therefore replace $X_1$ by $\pm 1$, and then following a $X_2^m$ correction, the logical operators are
\begin{align}
X_L \to Z_2\text{, }Y_L \to Y_2 Z_3\text{, }Z_L \to X_2 Z_3,
\end{align}
so we see that the result of the measurement is to move the information down the chain, and apply a Hadamard gate to it, as seen in the beginning of this section. More general measurements in the $\{ \frac{\pi}{2},\phi \}$ basis which neither commute or anticommute with the logical operators can be dealt with in a similar way, by multiplying any non-commuting terms by stabilisers, so e.g.\ $\exp(i\theta Z_2) \to \exp( i\theta Z_2 K_{g(2)})$, where $g$ is the \emph{gflow} function defined earlier. We will see examples of this in Chapter~\ref{chap:AGQC}.
An alternative and equivalent picture of MBQC is to start with a cluster state in which the qubits are prepared in $\ket{\pm_\theta}$ states, and all measurements are performed in the $\{ \ket{+} \bra{+} , \ket{-}\bra{-} \}$ basis. Thus the measurement angles are encoded in the graph state instead, and such a graph state is called a \emph{twisted} graph state. Using the notation $ X_{v}^{\theta_v }:= e^{-i\theta_v Z_v /2} X_{v}e^{i\theta_v Z_v /2}$, the stabiliser generators for the twisted graph state are
\begin{eqnarray} \label{eqn:GSStabTwist}
K_{v}^{\theta_v} &= X_{v}^{\theta_v} \prod_{w \sim v} Z_w,\text{ } \forall v \notin I,
\end{eqnarray}
where $\theta_v$ is the measurement angle on qubit $v$.
\subsubsection{Examples}
\label{sec:Example}
We now look at some examples of graphs, to illustrate the definitions above and since they will be useful in Chapter~\ref{chap:AGQC}. The first is the graph in Fig.~\ref{fig:gflow} from ref.~\cite{Browne2007}. The \emph{gflow} that satisfies all the \emph{gflow} rules is $g(a_1) = \{b_1 \}, g(a_2) = \{ b_2 \}, g(a_3) = \{ b_3,b_1\}$ with an ordering of $a_1 < a_2 < a_3$.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.25\textwidth]{gflow-eps-converted-to.pdf}
\caption{\label{fig:gflow} The \emph{gflow} is given by $g(a_1) = \{b_1\}, g(a_2) = \{ b_2\}, g(a_3) =\{b_1, b_3\}$ and the time ordering is $\{a_1\} < \{a_2\} < \{a_3\}$. The gflow is represented graphically as gflow lines (arrows), and the dotted line represents a gflow line not in the original graph, since there is no edge connecting vertices $a_3$ and $b_1$ but $g(a_3) = \{b_1,b_3\}$. }
\end{center}
\end{figure}
The second is the graph in Fig.~\ref{fig:ZigZag}, also studied in~\cite{Browne2007}. Many different possible \emph{gflows} can be defined on this graph. For example, a family of \emph{gflows} can be defined as
\begin{eqnarray}
g^r(v) = \left\{
\begin{array}{c c}
\{N+v,...,N+v+r -1 \}, & \mbox{if} \hspace{2mm} v+r-1 \le N \\
\{N+v,...,2N \}, & \mbox{if} \hspace{2mm} v+r-1 >N,
\end{array} \right.
\end{eqnarray}
where $1 \le r \le N$. If we were to perform measurements on this graph in MBQC, we can use any of these \emph{gflows}, provided we perform the right corrections on the outputs. For $g^r$, $r$ measurements can be performed simultaneously, interspersed by classical processing. Corrections on qubits will be of the form $X^{s_1 + s_2 + ... s_m}$ where the $s_m$ are binary variables accounting for the measurement outcomes. Thus the classical processing involves evaluating the binary sum $(s_1 + s_2 + ... s_m)$ of $r$ measurement outcomes, and so takes time $O(\log r)$~\cite{Furst84}. Since we can perform $r$ measurements simultaneously, the measurement depth $d_r$ is given by
\begin{eqnarray}
d_r = \left\lceil \frac{N}{r} \right\rceil.
\end{eqnarray}
The two extreme cases are where $r=1$ or $r=N$. The former is just where we can perform each measurement one-by-one ($d_1 =N$), with no addition of binary variables in between. The latter is where we can perform all measurements simultaneously ($d_N = 1$), but we must perform corrections on the outputs which take time $O(\log N)$. $g^N$ is called the \emph{maximally delayed flow} associated with this graph~\cite{Mhalla}. Note also that the zig-zag graph is not a universal resource for quantum computation, but combining several layers of the zig-zag graph we can make a universal resource.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.25\textwidth]{ZigZag-eps-converted-to.pdf}
\caption{\label{fig:ZigZag} A graph for which many different \emph{gflows} are applicable.}
\end{center}
\end{figure}
Finally, to aid the discussions later on, an example of a graph without \emph{gflow} is shown in Fig.~\ref{fig:NoGflow}. It can easily be verified that there are no assignments of $g(v)$ and ordering that satisfy the rules of \emph{gflow}.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.25\textwidth]{NoGflow-eps-converted-to.pdf}
\caption{\label{fig:NoGflow} A graph which violates the rules of \emph{gflow}.}
\end{center}
\end{figure}
\section{Quantum computation through adiabatic evolution}
\label{sec:AQC}
In the previous section we saw how computation can proceed through non-deterministic adaptive measurements on an entangled resource state. We will now a method of computation which is the opposite of this; quantum computation through \emph{adiabatic} evolution, in which operations are performed slowly and deterministically.
\subsection{The adiabatic theorem}
\label{sec:AdTheor}
The word `adiabatic' has a long history in thermodynamics, where it is taken to mean any process that occurs without heat passing into or out of a system (the word \emph{adiabatic} in Greek means `not passable'). For a physical process such as compression of an ideal gas, we might achieve adiabaticity by performing the compression fast enough that the system has not had time to equilibrate with its surroundings, or by insulating it from the outside world.
The way the word adiabatic is used in quantum mechanics typically is more analogous to (isothermal) reversible adiabatic processes. For a process to be truly reversible in classical thermodynamics, infinitesimal changes must be made such that the system is always equilibrated. In other words, the process of change must be slower than the process of equilibration. Of course, to be perfectly reversible the change must take an infinite amount of time, but the dynamics can be reversible up to some error if the dynamics of equilibration is fast enough.
As an example, imagine pulling a tablecloth from a table covered in glasses. If the cloth is pulled incredibly slowly (assuming zero friction between cloth and table and infinite friction between cloth and glasses), such that the glasses only wobble a negligible amount and so are constantly in equilibrium, then the glasses slide along with the table cloth; if the cloth is not pulled slowly enough, then the glasses wobble, with energy being released irreversibly in the form of heat as they rock from side to side on the table. Here what is `slow enough' is dictated by the properties of the glasses, and how fast the wobbling motion is damped.
Quantum adiabatic processes are much the same: the word `adiabatic' has connotations with a process that is performed \emph{slowly} with respect to the energy scale of the system, such that the probability of finding a system in the $n^{th}$ highest energy level is small during the process. Thus it is much like an isothermal reversible adiabatic process in which work can be performed to change the energy levels of the system, but the occupancy of the levels does not change (the entropy is fixed).
So what determines what is `slow enough' for a quantum system? The first treatment of adiabatic processes in quantum mechanics dates back almost 100 years ago, to Ehrenfest's work on adiabatic invariants~\cite{Ehrenfest1916} followed by Born and Fock~\cite{Born1928}. Since then, it has found applications in quantum computation~\cite{Farhi2001}, molecular dynamics~\cite{Born1927} and the quantum Hall effect~\cite{Niu1984,Avron1987}.
We first consider the simplest form of the adiabatic theorem, valid for systems where the energy levels are non-degenerate. Consider a Hamiltonian $H(t)$ with non-degenerate eigenvectors $\ket{E_n(t)}$ and energies $E_n(t)$. If we prepare the system in state $\ket{\psi(0)} = \ket{E_m(0)}$ and evolve it with the time-dependent Hamiltonian for some time $\tau$, then the system will end in the state $\ket{E_m(\tau)}$ with probability $1-p_{err}$, where~\cite{Messiah}
\begin{align}\label{eqn:AdiabC1}
p_{err} = O \left( \max_{m \neq n} \left| \frac{\max_{0 \leq s \leq 1} \bra{E_m(s)} \frac{dH(s)}{ds} \ket{ E_n (s)}}{ \min_{0 \leq s \leq 1}|E_m(s) - E_n(s)|^2} \right|^2, \right)
\end{align}
where $s = t /\tau$. In 1950, Kato derived a more general adiabatic condition that dropped the assumption of non-degenerate eigenvectors, and without assuming anything about the spectrum other than how it behaved near the region of interest~\cite{Kato1950}. This laid the foundations for the proof by by Albert Messiah~\cite{Messiah}, and provided the framework upon which later generalisations were built. A particularly useful result from the point of adiabatic quantum computation was the result in~\cite{Reichardt2004}, building on the work in~\cite{Avron1987}, which gives an expression for the error of an adiabatic transformation for the case when the energy levels can be degenerate (again using the notation $s = t /\tau$):
\begin{theorem}[Adiabatic theorem~\cite{Reichardt2004}]
Let $H(s)$ be a finite-dimensional, $(\kappa+1)$-times differentiable Hermitian matrix, with $\kappa \ge 1$, and with eigenvalues $E_j(s)$. Let $U(s)$ be the unitary evolution due to $H(s)$. Let $P(s)$ be the projection onto the (possibly degenerate) subspace of $H(s)$, with eigenvalue $E_0(s)$. Then after the evolution $U$, the amount by which a state initialised in the ground subspace of $H(0)$ has leaked into the excited states at time $s$ is given by
\begin{align}
p_{err} =\Vert (\idop - P(s)) U P(0) \Vert \le c(\kappa) \left( \left. \frac{1}{\tau} \frac{h(s)}{\Delta(s)^2}\right|_{b.c.} + \max_{0 \le s \le 1} \frac{1}{\tau^\kappa} \frac{h(s)^{\kappa+1} }{\Delta(s)^{2\kappa+1}} \right),
\end{align}
where $\Delta(s) = \min_{n \ne 0} |E_n(s) - E_0(s) |$, $h(s)$ satisfies $\Vert \left(\frac{d}{ds}\right)^l H(s) \Vert \le h(s)$ for all $l \le \kappa$ and $|_{b.c.}$ indicates that these quantities are evaluated at the boundaries $s=0,1$.
\end{theorem}
The norm used here is the spectral norm, which for a square Hermitian matrix $X$ is just the largest absolute eigenvalue of $X$. $p_{err} = \Vert (\idop - P(s)) U P(0) \Vert^2$ gives the probability of having leaked out of the subspace of $E_0$ at time $\tau$, and we can see a strong resemblance to the adiabatic theorem in (\ref{eqn:AdiabC1}). In this work, we consider the simplest case, where $H(s)$ is a linear interpolation of the form $(1-s)H_0 + sH_p$, so that $H(s)$ is infinitely differentiable. It is tempting then to set $\kappa \to \infty$, however this isn't possible without raising the adiabatic time to $\infty$, since $c(\kappa) = O(2^{\kappa}\kappa!)$~\cite{Reichardt2004}, which goes to $\infty$ as $\kappa \to \infty$. So we consider $\kappa$ as begin some fixed, large finite number. We also consider $\varepsilon$ to be fixed. With a linear interpolating Hamiltonian, $\Vert \dot{H}(s) \Vert = \Vert H_p - H_0 \Vert$ is independent of $s$, and the error in the adiabatic evolution becomes
\begin{align}
p_{err} \le c(\kappa) \left( \left. \frac{1}{\tau} \frac{\Vert \dot{H}(s) \Vert}{\Delta(s)^2}\right|_{b.c.} + \frac{\Vert \dot{H}(s) \Vert^{\kappa+1}}{\tau^k\Delta(s)^{2\kappa+1}_{min}} \right),
\end{align}
where $\Delta(s)_{min} := \min_{0 \le s \le 1}\left( \Delta(s) \right)$. By choosing a runtime that satisfies
\begin{align}\label{eqn:RunTime}
\tau \ge \left( \frac{ c(\kappa) \Vert \dot{H}(s) \Vert^{1+1/\kappa}}{\varepsilon\Delta(s)_{min}^{2+1/\kappa}} \right),
\end{align}
the probability of error becomes
\begin{align}
p_{err} \le \left. \frac{\varepsilon \Delta(s)_{min}^{2+1/\kappa} }{\Vert \dot{H}(s) \Vert^{1/\kappa} \Delta(s)^2 } \right|_{b.c.} + \frac{\varepsilon^\kappa }{c(\kappa)^{\kappa-1}}.
\end{align}
In all of the situations considered in this work, $\Vert \dot{H}(s) \Vert \geq \gamma$, and $\Delta(s) \leq \gamma$, where $\gamma$ is the energy scale set by the interactions in the particular system we implement the computation in. This means that $\Delta(s)_{min} / \Vert \dot{H}(s) \Vert \leq 1 $, and since $\Delta(s)_{min} \leq \Delta(s)$,
\begin{align}
p_{err} \le \varepsilon + \frac{\varepsilon^\kappa }{c(\kappa)^{\kappa-1}}.
\end{align}
So with a time scale of the form in (\ref{eqn:RunTime}), the error is of the order $\varepsilon$ (assuming that $c(\kappa) >1$, which is true for large enough $\kappa$). Following the terminology used in~\cite{Aharonov2008}, we say that the final state is $\varepsilon$ close to a state in the ground subspace of $H_p$, and ignoring the terms which are independent of $N$ we write the adiabatic runtime as
\begin{align}
\tau = \Omega \left( \frac{ \Vert \dot{H}(s) \Vert^{1+\delta}}{\Delta(s)_{min}^{2+\delta}} \right),
\end{align}
where $\delta := 1/\kappa$, and we use the `big Omega' notation $f(n) = \Omega(g(n) )$ to indicate that there is some constant $c$ and value $n_0$ such that $f(n) \geq cg(n)$ for $n > n_0$. Note also that in some cases the energy gap may close but properties of the Hamiltonian may prevent mixing between the degenerate states, so the adiabatic runtime will only diverge if the gap closes and transitions between the degenerate states are possible. We will see an example of this behaviour in chapter~\ref{chap:AGQC}.
The form of the adiabatic theorem given above is in fact not the most general, but it will be appropriate for all of the situations we consider in this thesis. A more general proof has been done by Jansen, Ruskai and Seiler in~\cite{Jansen2007}, generalising to situations where the projector $P(s)$ can project into several different energy levels, all of which can be degenerate. As a historical note, it is worth noting the inconsistency of the adiabatic theorem as pointed out by Marzlin and Sanders~\cite{Marzlin2004}, in which they consider adiabatic evolution including a time-varying field. Following this there was significant efforts to derive necessary and sufficient conditions~\cite{Tong2010,Boixo2010,Sarandy2004,Ortigoso2012}, with sufficient criteria subsequently defined in~\cite{Chueng2011}.
\subsection{Adiabatic quantum computation}
Using the adiabatic theorem, it is possible to perform computation in a very different way to the circuit model or measurement-based models we have seen before: \emph{Adiabatic quantum computation}. The standard adiabatic quantum computation (AQC) protocol~\cite{Farhi} starts with an initial Hamiltonian $H_0$, which is easy to prepare, such as a uniform magnetic field $\sum_n X_n$, where $X_n$ is a Pauli X operator acting on the $n^{th}$ qubit. The system is prepared in the ground state of this Hamiltonian, $\ket{E_0(0)}$, and then the Hamiltonian is slowly changed to a new `problem Hamiltonian' $H_p$ which is typically non-uniform and encodes the problem to be solved. Provided the evolution time obeys the adiabatic theorem, the system will be in the ground state of $H_p$ with high probability.
A natural example is solving a satisfiability problem~\cite{Farhi}, where we have a boolean formula made up of logical clauses (OR, AND,$\neg$) which we would like to be satisfied, e.g. $(x_1 \; \mbox{OR} \; x_2 )$ is satisfied by $(x_1,x_2) = (1,0),(0,1),(1,1)$, and $(x_1 \; \mbox{OR} \; x_2 ) \mbox{AND} (x_1 \; \mbox{OR} \; \neg x_2 )$ is satisfied by the assignment $(x_1,x_2) = (1,0),(1,1)$ etc. Such a problem can be framed as finding the ground state of a Hamiltonian, by replacing the logical clauses with energy penalties, that penalise when the clause is not satisfied. If we replace the boolean states $0,1$ with qubit states $\ket{0},\ket{1}$, then satisfying $(x_1 \; \mbox{OR} \; x_2 )$ is equivalent to finding the lowest energy configuration of $1 - \frac{1}{4}(1 + Z_1)(1+Z_2)$, and satisfying $(x_1 \; \mbox{OR} \; x_2 ) \mbox{AND} (x_1 \; \mbox{OR} \; \neg x_2 )$ is equivalent to finding the lowest energy configuration of $[1 - \frac{1}{4}(1 + Z_1)(1+Z_2)] +[1 - \frac{1}{4}(1 + Z_1)(1- Z_2)]$. Using this method, any satisfiability problem can be framed as finding the ground state of an Ising Hamiltonian, although this doesn't mean that an adiabatic quantum computer could solve all satisfiability problems as the energy gap may shrink faster than polynomially (for example, certain satisfiability problems are NP complete, and it is not generally believed that a quantum computer could solve NP-hard problems efficiently).
AQC has been shown to be \emph{polynomially equivalent} to the standard circuit model of quantum computation, in that it is possible to implement any quantum circuit using an adiabatic protocol, such that the inverse of the energy gap is polynomially equivalent to the number of gates in the circuit~\cite{Aharonov2008}. AQC has also gained much recent attention due to the appearance of D-wave, billing itself as the world's first \emph{quantum annealer} (quantum annealing has a similar philosophy to AQC, except the system is not assumed to be closed). However, whether or not this machine truly is a quantum annealer, or will provide any quantum speed up, is a subject of much debate (see e.g.~\cite{Boixo2014,Shin2014}).
The time to perform the computation in AQC is sensitive to the particular method of interpolating between the simple initial Hamiltonian $H_0$ and the problem Hamiltonian $H_p$. Often we will assume a linear interpolation of the form $H(t) = (1-t/T)H_0 + (t/T)H_p$, but clearly this cannot be optimal; we would expect the best algorithm to depend on the particular form of the gap as a function of $t$. Such a protocol, where the evolution speed can change in response to the time-dependent energy gap, is called a \emph{local} evolution (in contrast to the \emph{global} evolution we have considered so far). These differences lead to significant changes in computation speed; for instance, in~\cite{Roland2002} it was found that an adiabatic implementation of Grover's algorithm only achieved the $O(\sqrt{N})$ scaling if local adiabatic evolution was used. However, in this thesis the distinction between local and global evolution will not be important, so a linear interpolation will be used.
\subsection{Holonomic quantum computation}\label{sec:Holon}
Our discussion of the adiabatic theorem so far has focused on the probability of ending in a particular state. However, much interesting behaviour can also occur due to the phase the system picks up during an adiabatic evolution. Consider a system with Hamiltonian $H(t)$, that is prepared in the ground state $\ket{E_0(0)}$ at $t=0$ and evolved adiabatically. At some time $t$ later, the state is generally $\ket{\psi(t)} = e^{i\phi(t)} \ket{E_0(t)}$, assuming perfect adiabaticity. Inserting this into the Schr\"{o}dinger equation $i\hbar | \dot{\psi}(t)\rangle = H \ket{\psi(t)}$ gives
\begin{align}
(E_0(t) + \hbar \dot{\phi}(t) ) e^{i\phi(t)} \ket{E_0(t)} - i \hbar e^{i\phi(t)} | \dot{E}_0(t) \rangle = 0,
\end{align}
which can be rearranged, after left-multiplying by $\bra{E_0(t)}$ to give
\begin{align}\label{eqn:Berry1}
\phi(t) = i \int_0^t \sprod{E_0(t')}{\dot{E}_0(t')} dt' - \frac{1}{\hbar} \int_0^t E_0(t') dt'.
\end{align}
The term on the right is called the dynamical phase, whilst the term on the left is the \emph{Berry} or \emph{geometrical} phase~\cite{Berry1984}. This type of phase contribution was first pointed out by Pancharatnam in the context of classical optics~\cite{Pancharatnam} (and for this reason the Berry phase is also sometimes called the \emph{Pancharatnam phase}), and we can understand the form of $\gamma_n$ in equation (\ref{eqn:Berry1}) as being the combination of many small contributions due to the slight difference in phase between states at adjacent times. The Berry phase produces observable physical effects, and is used in describing many phenomena in condensed matter systems, such as topological insulators~\cite{Thouless1982,Kane2005}.
Berry phases can also be extended to degenerate Hamiltonians. If instead the Hamiltonian has a $d$-fold degenerate ground space of energy $E_0$ with states labelled as $\ket{E_0^{\alpha}}$, $1 \leq \alpha \leq d$, then following an adiabatic transformation, rather than a phase being applied the ground space is transformed by the $d \times d$ matrix $U_{\alpha \beta}$~\cite{Wilczek1984}, where
\begin{align}
U_{\alpha \beta}(t) = \mathcal{T} \exp \left( i\gamma^{\alpha \beta}(t) \right) = \mathcal{T} \exp \left( - \int_0^t \bra{E_0^{\alpha}(t')}\frac{d}{dt'} |E_0^{\beta}(t') \rangle dt' \right).
\end{align}
$\gamma_{\alpha \beta}$ is the \emph{Wilczek-Zee connection}, and $U_{\alpha \beta}$ is called a \emph{holonomy}. Thus a state prepared in a state $\ket{E_n^\alpha(0)}$ in the degenerate subspace will be transformed into a state $ e^{ \frac{-i}{\hbar} \int_0^t E_0(t) dt' }U_{\alpha \beta}(t) \ket{E_n^\alpha(t)}$, in which the holonomy creates transitions between different states in the degenerate subspace (or in other words, it becomes a non-Abelian generalisation of the Berry phase, since transformations no longer necessarily commute with one another). Provided the system is adequately controllable, it is possible to harness this so that information encoded in the degenerate energy levels of $H(0)$ can be arbitrarily transformed using these holonomies, and so universal quantum computation can be performed. This is known as \emph{holonomic quantum computation} (HQC)~\cite{Zanardi1999,Pachos1999}. The holonomy can be generated either using a closed loop in parameter space such that $H(0) = H(\tau)$, or can more generally have $H(0) \neq H(\tau)$, in which case it is usually called \emph{open-loop} HQC~\cite{Kult2006}. Also note that, although HQC is commonly discussed in terms of adiabatic evolutions, this is not a necessary requirement, and schemes for achieving holonomies via non-adiabatic processes have been proposed~\cite{Aharanov1987,Sjokvist2012,XiangBin2001,Burgarth2013}.
HQC has desirable properties which lead a natural `hardware' resistance to certain types of errors. To see this, consider a Hamiltonian $H(\mathbf{R})$ that depends on a set of parameters $\mathbf{R}[t] = (r_1(t),r_2(t),...,r_n(t))$. Then the Berry phase around a path $\Gamma$ becomes
\begin{align}
\gamma_n &= i \oint_\Gamma \bra{E_n(\mathbf{R})} \nabla_{\mathbf{R}} \ket{E_n(\mathbf{R})} \cdot d\mathbf{R} := \int \int_{S} \mathbf{F}_n \cdot d \mathbf{\sigma},
\end{align}
where we have used Stokes' theorem, and defined $\mathbf{F}_n := \nabla_{\mathbf{R}} \times \mathbf{A}_n $, $\mathbf{A}_n = \bra{E_n(\mathbf{R})} \nabla_{\mathbf{R}} \ket{E_n(\mathbf{R})}$, and $S$ is the surface contained by the path $\Gamma$. In this form, the phase has the interpretation as the integral of a curvature over an area, which highlights one of the most useful and interesting thing about these acquired phases; they only depend on the area contained within the evolution path. Thus quantum computations using such evolutions could be naturally resistant to control errors that change the path taken in parameter space but which don't change the area. This property has been tested theoretically, confirming that HQC operations are less sensitive to dephasing noise~\cite{Carollo2003, DeChiara2003}, and error processes where the gate time is either much longer or much shorter than the noise correlation time~\cite{Solanis2004}. Further protection against noise can in principle be achieved with HQC in decoherence-free subspaces and subsystems as shown in~\cite{Wu2005,Oreshkov2009prl}. HQC does not only provide `hardware' improvements; a scheme for fault-tolerant HQC on stabiliser codes has also been developed~\cite{Oreshkov2009prl2,Oreshkov2009} (unlike AQC, where developing fault tolerant computation is an ongoing problem~\cite{Young2013}), making HQC a promising method to perform quantum computation.
\section{Spin chains for communication}
\label{sec:IntroSpinChain}
As well as being able to initialise, store, and process quantum information in a quantum computer, it is also important to be able to transport quantum information between different parts of the computer, or perform gates between distant qubits, since geometrical constraints will prevent all qubits from being directly connected. Methods to do this with minimal loss of coherence are thus incredibly important. Many proposals to perform this task involve transferring information from one physical system to another (e.g.\ from a trapped ion to a photon) or shuttling a physical system from one part of the computer to another. But in certain situations, it may be beneficial to avoid transfer between mediums or moving qubits around, as these can be error-prone processes. Instead we could create communication channels such that the information is stored and transferred in the same medium, by creating chains of qubits along which quantum information can travel. Then to transfer information from one qubit to another, all that is required is to turn on interactions with the chain, and then turn them off again when the information has arrived at the other end. This is the concept behind using chains of qubits, or `spin chains' as a quantum wire to transfer information in quantum computers~\cite{Bose03}. A spin chain is a 1-dimensional chain of spins where the spin degree of freedom of nearby physical systems are coupled. In this thesis we will just consider spin chains formed of spin-$\frac{1}{2}$ particles, so really they can be thought of as a 1 dimensional array of qubits, but we use the terminology spin chain since it is commonly used in the literature. The interactions are typically confined to nearest or next-nearest neighbours, and typically have the form
\begin{eqnarray}
H= \sum_{m<n} J_{mn} (X_m X_{n} +Y_m Y_{n}+ \Delta_{mn} Z_n Z_n) + \sum_m B_m Z_m,
\end{eqnarray}
where $X, Y,Z$ are the Pauli matrices, $\Delta_{mn}$ is the anisotropy, $B_m$ is the magnetic field strength, and $J_{mn}$ is the coupling parameter (negative for ferromagnetic interactions which favour alignment of spins, and positive for anti-ferromagnetic interactions which favour anti-alignment). Some particularly important and well-used spin chains are where $\Delta_{mn}=1$ (isotropic or Heisenberg spin chain), $\Delta_{mn} = 0$ (XY spin chain) and $\Delta \to \infty, J<0$ (Ising spin chain). Spin chains appear in nature~\cite{Hammar1999}, and can also be realised in physical systems such as Josephson arrays~\cite{Romito2005} or atoms in optical lattices~\cite{Bloch2011}.
Whilst it would be trivial to take a chain of qubits and perform perfect state transfer by turning on a \textsc{swap} interaction between successive pairs of qubits, we are more interested in the case where there is minimal control over the chain, and where the interactions are time-independent and ideally uniform. This is firstly since if we did have a large degree of control over the chain we may as well use it as part of the quantum computer, and secondly since minimal control means less contact with the outside world and thus less decoherence.
The first spin chain communication protocol in~\cite{Bose03} used $\Delta_{mn} = 1$, $J_{mn} = J<0$ and nearest-neighbour interactions. The qubits in the chain are initialised in the $\ket{\mathbf{0}} = \ket{000...0}$ state, and at time $t=0$ qubit $s$ is encoded with a quantum state $\ket{\psi} = \cos \theta/2 \ket{0} + e^{i\phi} \sin \theta/2 \ket{1}$. At some time $\tau$ later, the state of the chain is $\ket{\Psi(\tau)}$, and we calculate the fidelity of a receiving qubit $r$ with the original information, where the fidelity is given by
\begin{align}
F_\psi = \bra{\psi} \text{tr}_{\hat{r}} ( \ket{\Psi(\tau)} \bra{\Psi(\tau)} ) \ket{\psi},
\end{align}
where $\text{tr}_{\hat{r}}$ means tracing out all qubits except qubit $r$, and for now we do not square the fidelity as in Section~\ref{sec:DistMeas}, to fit with the conventions used in the spin chain literature.
Since the fidelity is sensitive to the particular state $\ket{\psi}$ input at $s$, and we require a general metric of how good this channel is, $F_\psi$ must be averaged over all possible input states, giving the overall averaged fidelity $F_{av}$
\begin{align}
F_{av}= \frac{1}{4 \pi} \int F_\psi \; \sin \theta d \phi d\theta.
\end{align}
At some time $\tau_{opt}$, $F_{av}$ will be maximised (with a cut-off after a time $\tau_{max}$, chosen as $4000/J$ in~\cite{Bose03}). When $s=1$ and $r=N$, this simple protocol can achieve near perfect state transfer for $N=4$ spins, and achieves $F_{av} > 2/3$ for up to 80 spins (which is an important benchmark since it is the maximum possible average fidelity achievable by purely classical means~\cite{Horodecki1999}).
After this initial proposal, there followed many more schemes, all using constantly coupled qubits but with different Hamiltonians.
Particularly important results are the perfect state transfer channel~\cite{Christandl04}, which achieves perfect transfer by using a mirror symmetric spin chain with engineered couplings $J_{n,n+1} = \sqrt{N(N-n)}$. Other notable protocols include coupling two qubits weakly to the chain such that the chain acts as an effective qubit~\cite{Plenio2005}, optimised couplings at the ends of an $XY$-chain to make the excitations fall in a region of linear dispersion~\cite{Banchi2011}, and a heralded dual rail protocol~\cite{Burgarth05}. For a more in depth discussion of spin chains for quantum communication, see~\cite{Bose2008}.
In order to measure how well the channel communicates information, we will often use the average fidelity defined above, which can be put in a particularly simple form that is easier to calculate~\cite{Horodecki1999}. For a channel that transforms the input information by $\mathcal{E}$, with systems of dimension $d$, the average fidelity is
\begin{align}\label{eqn:FavrgHorod}
F_{av}[\mathcal{E}] = \frac{d F_e[\mathcal{E}] +1 }{d +1},
\end{align}
where $F_e[\mathcal{E}] = \bra{\psi} [ \idop \otimes \mathcal{E} ](\ket{\psi}\bra{\psi} ) \ket{ \psi}$, and $\ket{\psi}$ is a maximally entangled state such as $\ket{\psi} = \frac{1}{\sqrt{2}} (\ket{01} - \ket{10} )$ (this is often called the \emph{singlet fraction}). For a classical channel, through which only product states can be sent, the largest overlap between a singlet and a product state is $1/\sqrt{d}$, so a classical channel has $F_{av}[\mathcal{E}] \leq \frac{2}{3}$.
Apart from communicating quantum information, using a spin chain also serves as a useful method of sharing entanglement between two different qubits (and in fact a channel that is good for sending quantum information is also a good entanglement sharer~\cite{Bayat10,Horodecki1999}). The protocol follows in a similar fashion as above, except qubit $s$ is maximally entangled with another ancilla qubit $a$. After evolving in time, $a$ will be entangled with qubit $r$. If the protocol is not perfect, it is possible to repeat it many times and use \emph{entanglement distillation} to convert multiple copies of non-maximally entangled states into one maximally entangled state~\cite{Bennett1996}. Once such an entangled state is shared between two parties, it can be used to teleport quantum information between them (provided there is additional classical communication, which is usually considered as a plentiful resource). For a pure state $\ket{\psi_{AB}}$ over two qubits $A$ and $B$, the measure of bipartite entanglement is the Von Neumann entropy $S(\ket{\psi_{AB}})$~\cite{Popescu1997}, defined as
\begin{align}
S(\ket{\psi_{AB}}) = - \text{tr} (\rho_A \log_2 \rho_A) = - \text{tr} ( \rho_B \log_2 \rho_B),
\end{align}
where $\rho_A:=\text{tr}_{B} (\ket{\psi_{AB}}\bra{\psi_{AB}})$ and similarly for $\rho_B$.
In the case when the two qubits are in a mixed state $\rho_{AB}$, a suitable measure of entanglement is the \emph{entanglement of formation}~\cite{Bennett1996,Wootters98}. First, note that there can be many decompositions of $\rho_{AB}$ into an ensemble of pure states, $\rho_{AB} = \sum_n p_n^{\xi} \ket{\psi_n}^{\xi}\bra{\psi_n}^{\xi}$, where the $\xi$ superscript denotes a particular decomposition. The entanglement of formation is then defined as
\begin{align}
E_f(\rho) = \min_{\xi} \sum\nolimits_n p_n^{\xi} S(\ket{\psi_n}^\xi).
\end{align}
In other words, it is the minimum of the expected entanglement over all the possible pure state ensembles that could make up $\rho_{AB}$.
A particularly useful form of $E_f(\rho)$ was derived in~\cite{Wootters98}. If we define $\tilde{\rho} = (Y_1 \otimes Y_2) \rho^* (Y_1 \otimes Y_2)$, then $E_f(\rho) = \Lambda(C)$, where
\begin{align}
\Lambda(C) = H\left(\frac{1 + \sqrt{1-C^2}}{2} \right), \; H(x) = -x \log_2 x - (1-x) \log_2 (1-x)
\end{align}
and $C = \max \{ 0, \lambda_1 - \lambda_2 - \lambda_3 -\lambda_4 \}$, where $\{ \lambda_n \}$ are the eigenvalues of $\sqrt{ \sqrt{\rho} \tilde{\rho} \sqrt{\rho} }$. $C$ is usually called the \emph{concurrence}.
This use of spin chains is not only restricted to transferring information or entanglement in a quantum computer. One dimensional systems can exhibit a rich variety in behaviour, and testing how quantum information is transferred through certain systems could reveal information about this. The quality of transfer could also be used to probe the `quantumness' of the individual qubits in the chain: if any members of the chain behave classically, then we would expect the information transmitting ability of the chain to be impaired.
\section{Some experimental candidates for a quantum computer}
Having seen some of the theoretical ideas behind a quantum computer, we now discuss the more practical side of creating a quantum computer, introducing some important experimental proposals and reviewing some of the experimental progress. The standard framework by which experimental proposals are judged is the \emph{DiVincenzo criteria}~\cite{DiVincenzo2000}, which is essentially a checklist of desirables for an experimental scheme. The checklist is as follows:
\begin{itemize}
\item The qubits should be scalable and well-characterised (meaning that the couplings to other qubits and external influences should be known).
\item It must be possible to initialise qubits in a pure state such as $\ket{00...0}$ (for computation and error correction).
\item The qubits must have decoherence times significantly longer than gate operation times.
\item It must be possible to perform a universal set of gates on the qubits.
\item The state of the qubits must be measurable.
\item It must be possible to transmit qubits between specific locations.
\end{itemize}
In this section we will focus on two particular proposals which could fulfil these criteria; ion traps and donors in silicon. There are of course many more experimental systems which are also promising candidates, such as NMR~\cite{Cory1997,Gershenfeld1997}, linear optics~\cite{Knill2001b}, quantum dots~\cite{Loss1998}, and superconducting qubits~\cite{Devoret2013}, but we will focus on ion traps and donors because of their relevance to Chapter~\ref{chap:3qubit}. For a more comprehensive overview of the current candidates for a quantum computer, see e.g.~\cite{NielsenChuang,Buluta2011,Ladd2010,RoadMap}
\subsection{Ion traps}
\label{sec:IntroIonTraps}
The first scheme for using trapped ions to process quantum information was by Cirac and Zoller~\cite{Cirac1995}. The essential idea is to trap ions in an electromagnetic field, and encode information in internal states of these ions. Typically we consider ions trapped in a linear Paul trap~\cite{Paul1953}, with a combination of an axial DC field and a radial AC field creating a fluctuating potential that approximates a harmonic confining potential $\frac{1}{2}M \nu_x^2 x^2 + \frac{1}{2}M \nu_y^2 y^2 + \frac{1}{2}M \nu^2 z^2$ for each ion, such that $\nu_x,\nu_y \gg \nu$ ($\nu$ is often called the secular frequency of the trap). Note that it is not possible to simply create a static confining potential due to Earnshaw's theorem~\cite{Earnshaw1842}.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.65\textwidth]{IonTrap-eps-converted-to.pdf}
\caption{\label{fig:IonTrap} A schematic diagram of a linear quadrupole Paul trap. Two diagonally opposed electrodes produce a radio frequency oscillating electric field which approximates a harmonic trapping potential. The ions are confined in the $z$-axis by two endcap electrodes. }
\end{center}
\end{figure}
Initialisation of the trapped ion qubits could be achieved using the other excited levels of the ion; if there are excited states which can be reached from the qubit states by application of a suitable laser field, and which both decay into the qubit $\ket{0}$ state, then applying laser fields which are resonant with all transitions from the $\ket{1}$ state result in the qubit being prepared in the $\ket{0}$ state.
Measurement of the qubit state can be performed by applying laser radiation that is resonant with a transition from the $\ket{1}$ state to an excited state that decays back into the $\ket{1}$ state, so that fluorescence is observed only if the qubit is in the $\ket{1}$ state. Single qubit rotations can be performed by pairs of lasers focused to a spot significantly smaller than the separation of the ions; single qubit $X$ rotations can be performed using a resonant laser pulse to drive Rabi oscillations, whilst single qubit $Z$ rotations can be performed using off-resonant laser pulses resulting in an AC Stark shift. $T_2$ times of ions can be up to $\sim 10 \text{ms}$, whilst $T_1$ times can be on the order of minutes~\cite{Buluta2011}.
To perform gates between qubits, the proposal in~\cite{Cirac1995} uses the motional degree of freedom of the ion chain. When the system is sufficiently isolated and cooled so that $k_BT \ll \hbar \nu$, the ion motion in the $z$-direction is restricted to the ground state, which is a coherent phonon corresponding to coordinated oscillation of all of the ions. When the motion is confined in this way, we call this phonon mode the `bus qubit'. We label states of the ion/bus system as $\ket{m}_I\ket{n}_B$, where $m=0,1$ is the state of the ion and $n=0,1$ corresponds to the phonon mode being unoccupied/occupied respectively. When ions absorb a photon, there is a resultant change in momentum of the ion, which translates to an excitation of the phonon mode. If the energy splitting between all of the ion qubit states is different along the chain (which can be achieved by e.g.\ applying a magnetic field gradient) then using a global field we can couple individual ions to the bus qubit.
This can then be used to perform a gate between two ions in the chain (say ions $1$ and $2$). To begin, the internal states of ion $1$ is swapped with the bus qubit by selecting a laser frequency that causes oscillations between the $\ket{0}_1\ket{1}_B$ and $\ket{1}_1\ket{0}_B$ states. Then it is possible to select a laser frequency that is resonant with a transition from the $\ket{1}_2\ket{1}_B$ state to another (non-computational) state, resulting in a $-1$ phase applied to state $\ket{1}_2\ket{1}_B$ but no evolution of the other states. The bus qubit can be swapped back into ion $1$, thereby achieving a \textsc{cz} gate between the two qubits, mediated by the bus qubit.
An alternative method to create interactions between ions, which we will use later on in Chapter~\ref{chap:3qubit}, is magnetic gradient induced coupling~\cite{Mintert2001}. The essential idea is to use a magnetic field gradient to create an effective Ising interaction between two ions, rather than using a laser field to swap ions in and out of the bus qubit. In this magnetic field gradient, the energy of the ion depends on the internal state of the ion, and so the equilibrium position of the ion will also depend on the internal state of the ion. Thus the two qubit states yield different equilibrium positions of the ions, $\bar{z}_n$ and $\bar{z}_n + d_n$. The difference in equilibrium position $d_n$ leads to conditional dynamics; roughly speaking, when the $n^{th}$ ion is in a $\ket{1}_n$ state, it becomes closer to the $(n+1)^{th}$ ion and thus has an effect on the state of this ion. Thus the internal states of the ions affect the interactions between the ions.
The full analysis is given in~\cite{Mintert2001,Wunderlich2001}, in which it is shown that the magnetic field gradient produces couplings of the form $\frac{1}{2}\hbar J_{zz}^{ij} Z_i Z_j$ with coupling strength given by
\begin{align}\label{eqn:IonJzz}
J_{zz}^{ij} = \sum_{n=1}^N \nu_n \varepsilon_{ni} \varepsilon_{nj},
\end{align}
where $\varepsilon_{ni}$ is the effective Lambe-Dicke parameter and $\nu_n$ is the angular frequency of the $n^{th}$ vibrational mode of the ion chain. The effective Lambe-Dicke parameter here depends on the shift in equilibrium position of the $j^{th}$ ion $d_j$ relative to the spatial extent of the $n^{th}$ harmonic oscillator wavefunction, $\Delta z_n = \sqrt{\hbar/2m\nu_n}$, and is given by
\begin{align}
\varepsilon_{nj} =S_{nj} \frac{\Delta z_n \partial_z \omega_j}{\nu_n},
\end{align}
where $\omega_j := \omega_0(z_j)$. The matrix $S$ is a matrix whose columns are the normal modes of the chain, which take into account the extent to which each ion is involved in the $n^{th}$ vibrational mode.
Aside from the above two methods of coupling qubits, the two most prominent alternatives are M{\o}lmer-S{\o}rensen gates~\cite{Sorensen1999} using detuned laser fields, and using a state dependent shifting of the ion wavepacket~\cite{Cirac2000}. Both of these schemes have the additional advantage of not depending on being in the motional ground state, so are less susceptible to decoherence of the bus qubit.
The linear trap architecture is suitable for performing computations on up to $\sim 100$ qubits~\cite{Zhu2006}. Thus scaling up to thousands of qubits or more requires a different architecture. The most prominent proposals to date include shuttling ions between linear traps where logic gates can be performed~\cite{Kielpinski2002}, using a dedicated movable ion to mediate interactions between different ions~\cite{Cirac2000}, or linking distant traps using photons~\cite{Cirac1997},\cite{Pellizzari1997}. To realise these schemes, planar ion traps are a promising candidate~\cite{Chiaverini2005,Seidelin2006}, in which ions are trapped above the surface of a 2D micro-fabricated electrode array.
The first successful implementation of a \textsc{cz} gate on ion traps was by Monroe et.\ al.~\cite{Monroe1995}. Since then there have been significant advances such as realisation of high fidelity two-qubit gates~\cite{Benhelm2008}, Toffoli gates~\cite{Monz2009}, demonstration of a 7-qubit topological error correcting code~\cite{Nigg2014}, demonstrations of ion shuttling~\cite{Blakestad2011,Moehring2011}, and demonstrations of micro-fabricated planar traps~\cite{Moehring2011,Amini2010,Merrill2011,Sterling2014}, making ion traps a strong contender for a quantum computer, or at least a candidate for an early prototype. For a more in depth review of ion traps, see e.g.~\cite{Steane1996c,Buluta2011,Ladd2010,Monroe2013,Haffner2008}.
\subsection{Donors in silicon}
\label{sec:DonIntro}
The use of dopants in silicon to perform quantum information processing was first proposed by Kane~\cite{Kane1998}. The central idea in this scheme is to replace a silicon atom in crystalline silicon by a group V atom such as phosphorus or bismuth. Group V elements have a spare electron, which is donated to the conduction band at room temperature (and hence is called a donor), a technique used in current electronics to negatively dope silicon to create n-type semiconductors. At low temperatures ($\lesssim 15 K$) the electron stays close to the donor nucleus, allowing information to be stored and processed by manipulating this electron. Such a proposal benefits from the vast research efforts into silicon-based technology for classical computation, as well as the low spin-orbit coupling in silicon and low abundance of isotopes with non-zero nuclear spin. Denoting $\{ S_k^x,S_k^y,S_k^z\} = \{ \frac{1}{2} X_k,\frac{1}{2} Y_k,\frac{1}{2} Z_k\}$ and $\{I_k^x,I_k^y,I_k^z\}$ as the spin operators of the $k^{th}$ electron or $k^{th}$ nucleus respectively, the effective spin Hamiltonian for a system of two donors in the presence of a magnetic field in the $z$-direction is~\cite{Kane1998}
\begin{align}
H_{don} = g_e \mu_B B( S_1^z +S_1^z) - g_N \mu_N B( I^z_{1} +I^z_{2} )+ A (\underline{I}_1 \cdot \underline{S}_1+ \underline{I}_2 \cdot \underline{S}_2)+ J \underline{S}_{1}\cdot \underline{S}_{2},
\end{align}
where $\underline{S}_k = S_k^x+S_k^y+S_k^z$, $\underline{I}_k = I_k^x+I_k^y+I_k^z$. The parameter $A$ is the strength of the electron-nucleus hyperfine coupling (which depends on the group V atom used), and $J$ is the strength of the electron-electron exchange coupling. Typical values for $A$ are 117 MHz in phosphorus (nuclear spin $\frac{ 1}{2}$) and 1.475 GHz in bismuth (nuclear spin $\frac{9}{2}$) and J depends on the separation, but is typically on the order of GHz. Magnetic fields used in experiment tend to be up to $<1$T (up to $g_e \mu_B B < 20$GHz). We consider donors with spacings $\lesssim 20$nm, so that the electron-electron exchange is significantly larger than the electron dipole-dipole interactions and interactions between donor nuclear spins~\cite{Zwanenburg2013}.
In the Kane proposal, the hyperfine coupling between each electron and their respective nucleus is controlled by modifying the electron wavefunction overlap with the nucleus via a single electrode placed above each donor (the `A gate', see Fig.~\ref{fig:Donors}) as recently demonstrated in~\cite{Wolfowicz2014}. By manipulating the hyperfine coupling in this way, the energy splitting of the electron spin changes, which performs a $Z$ rotation. $X$-rotations could then be achieved by applying an AC field which matches the resonant frequency of the electron spin. Individual addressability of spins could be achieved by varying the resonances of the electron spins, by e.g.\ applying a magnetic field. Measuring the electron spin could be done using spin-dependent tunnelling~\cite{Morello2009}, which has recently been demonstrated~\cite{Morello2010}, and could also be used to initialise the spins. The interactions between donor electrons are exchange interactions which depend on the wavefunction overlap between the electrons, and could be controlled using an electrode between the donors (also called a `J' gate, see Fig.~\ref{fig:Donors}) which controls this overlap and thus allows the two-qubit interactions to be switched on and off. Alternative schemes for performing two-qubit gates also exist, such as using a control atom that can be optically excited to create an interaction between two donor electrons~\cite{Stoneham2003}, or using different nuclear spin states to create interactions that are activated by an AC field~\cite{Kalra2014}.
Measurements of decoherence times for an ensemble of P atoms in $^{28}\text{Si}$ have shown nuclear $T_2$ times on the time scale of minutes, using dynamical decoupling~\cite{Saeedi2013,Steger2012} and electron $T_2$ times of up to $10\text{s}$ using dynamical decoupling~\cite{Tyryshkin2012}. For individual donors, the most recent study has shown nuclear $T_2$ times of around 30s and electron $T_2$ times of up to 0.5s using dynamical decoupling in isotopically pure silicon~\cite{Muhonen2014}. Since the nuclear spin of the donor has a much larger coherence time than the electron, the optimal strategy would most likely be to use the nuclei for long term storage of information, and use the hyperfine interaction to transfer information to the electrons so that it can be read out or manipulated quickly (as demonstrated in~\cite{Morton2008}).
For the work in this thesis, we will focus on bismuth donors in silicon, because of its large nuclear spin and hyperfine coupling relative to the other group V elements, which will will prove useful in Chapter~\ref{chap:3qubit} for creating quantum gates. Another advantage of bismuth is the presence of optimal working points \cite{Mohammady2010,Mohammady2012,Balian2012,Morley2013,Balian2014}, at which $T_2$ times can be enhanced up a time scale of seconds~\cite{Wolfowicz2013}. However in high magnetic fields, where optimal working points cannot be reached, the best achieved $T_2$ times are $\sim 0.5 \text{ms}$ in natural silicon~\cite{Morley2010,George2010} or around 700ms for isotopically pure silicon~\cite{Wolfowicz2012} For a more in depth review of donors in silicon, see~\cite{Mohammady2013,Morley2014,Zwanenburg2013,Simmons2013}.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.7\textwidth]{Donors-eps-converted-to.pdf}
\caption{\label{fig:Donors} A schematic diagram of the Kane proposal for a quantum computer made from donors in silicon. }
\end{center}
\end{figure}
\chapter*{Acknowledgements}
\include{Introduction/Intro}
\include{DFS/DFSpaper}
\include{AGQC/AGQC_Thesis}
\include{3Qubit/3Qubit_Thesis}
\include{Wigner/Wigner}
\bibliographystyle{plain}
\chapter*{Declaration}
I, \theauthor{}, confirm that the work presented in this thesis is my own. Where information has been derived from other sources, I confirm that this has been indicated in the thesis. Chapter~\ref{chap:DFS} is based on work done in collaboration with Sougato Bose. Chapter~\ref{chap:AGQC} is based on work done in collaboration with Damian Markham and Janet Anders. Chapter~\ref{chap:3qubit} is based on work done in collaboration with Joseph Randall, Gavin Morley Winfried Hensinger and Sougato Bose. Chapter~\ref{chap:Wigner} is based on work done in collaboration with Abofazl Bayat, Sanjeev Kumar, Michael Pepper and Sougato Bose. Section~\ref{sec:NNN} is based on work done in collaboration with Abofazl Bayat and Sougato Bose. Section~\ref{sec:EdgeLock} is based on work done in collaboration with Masudul Haque and Sougato Bose. Parts of this dissertation have been published, or submitted for publication, as follows:
\begin{itemize}
\item Bobby Antonio and Sougato Bose. Two-qubit gates for decoherence-free qubits using a ring
exchange interaction. \emph{Phys. Rev. A}, 88:042306, 2013.
\item Bobby Antonio, Damian Markham and Janet Anders, Adiabatic graph-state quantum computation. \emph{New J. Phys.} 16 113070, 2014.
\item Bobby Antonio, Abolfazl Bayat, Sanjeev Kumar, Michael Pepper and Sougato Bose. Self-Assembled Wigner Crystals as Mediators of Spin Currents and Quantum Information. \emph{Phys. Rev. Lett.}, 115 216804, 2015.
\end{itemize}
\vfill
\hfill \theauthor
\hfill \monthyear\displaydate{thesis_date}
\newpage\null\thispagestyle{empty}
\newpage
\setcounter{page}{3}
\chapter*{Abstract}
\input{Abstract.tex}
\newpage\null\thispagestyle{empty}
\newpage
\setcounter{page}{4}
\IfFileExists{Acknowledgements.tex}{%
\chapter*{Acknowledgements}
\input{Acknowledgements.tex}
}{}
\newpage\null\thispagestyle{empty}
\newpage
\setcounter{page}{5}
\setboolean{@twoside}{true}
\tableofcontents
\listoffigures
\listoftables
\pagestyle{fancy}
\fancyfoot[C]{\thepage}
\fancyhead[RO,LE]{}
\fancyhead[LO]{\leftmark}
\fancyhead[RE]{\leftmark}
\include{Intro}
\include{DFSpaper}
\include{3Qubit_Thesis}
\include{AGQC_Thesis}
\include{Wigner}
\chapter{Summary and Outlook}
\input{Conclusions}
\chapter{Quantum communication through a Wigner crystal}
\label{chap:Wigner}
\section{Quantum communication through a Wigner crystal}
\label{sec:Wigner}
In the introduction, we motivated the need for quantum communication channels to share quantum information or to distribute entanglement between different parties, and we introduced the concept of using spin chains to perform this task. In this chapter, we will investigate a novel way to realise these `quantum wires', using a \emph{Wigner crystal}~\cite{Wigner1934}. Wigner crystals form in an electron gas, in the regime where the electron density is below a certain critical value such that coulomb interactions dominate over the kinetic energy of the electrons. Under these conditions, the electrons are extremely localised near to the classical lowest energy configuration, and so are said to have formed a crystal.
The regimes in which it is possible to see a Wigner crystal can be parameterised using the Wigner-Seitz radius $r_s$, defined as the average distance between electrons in units of the Bohr radius, which for 2D systems is $r_s^2 = (\pi a_B^2n)^{-1}$ where $n$ is the density of electrons per unit area, $a_B$ is the Bohr radius $a_B = 4 \pi \epsilon \hbar^2 / m^* e^2$, and $\epsilon,m^*$ are the permittivity and effective mass of the material being used. To arrive at the conditions for a Wigner crystal, we can think of each electron as sitting in a potential well due to the Coulomb repulsion from other electrons. Modelling this as a harmonic potential with strength $\omega$, the electron wavefunction has width roughly $\Delta r \sim \omega^{-1/2}$. Approximating $\omega^2 \sim \frac{1}{m} \frac{\partial^2 V}{\partial r^2 }$, gives $r_s = \omega^{-2/3} $ when $V$ is a Coulomb potential. Thus $\Delta r / r_s \sim r_s^{-1/4}$, which means the Wigner crystal picture (where the spatial localisation is small compared to the separation of electrons) becomes valid for large values of $r_s$. Monte Carlo simulations support this rough argument~\cite{Tanatar1989}, suggesting that 2D Wigner crystals would occur for $r_s > 37 \pm 5$. For 1D Wigner crystals (where the electrons are strongly confined in two directions), Monte Carlo simulations suggest that the the melting point is much lower, allowing Wigner crystals at $r_s := (n_{\textsc{1d}} a_B)^{-1}> 4$ \cite{Egger1999}, where $n_{\textsc{1d}}$ is the number of electrons per unit length .
A classical Wigner crystal was first realised experimentally on the surface of liquid $^3$He~\cite{Grimes1979}. Since then, there have been many promising experimental observations. In solid state architectures, evidence for a Wigner crystal has been observed in a Si-MOSFET~\cite{Okamoto1998}, and in a 2-dimensional hole structure formed at a GaAs/AlGaAs boundary~\cite{Yoon1999}, which also confirmed the transition to a Wigner crystal at $r_s \sim 37$ predicted in~\cite{Tanatar1989}. In addition there have been many promising advances in the fabrication of quantum wires in GaAs/AlGaAs structures, in studies investigating the Tomonaga-Luttinger phase of electrons, where spin and charge excitations are independent~\cite{Auslaender2005,Steinberg2007,Jompol2009,Laroche2014}. Carbon nanotubes suspended over a substrate are also promising candidates, as they appear to allow high values of $r_s$ to be achieved, with small effects due to impurities~\cite{Tans1997,Jarillo2004}.
There are also other proposals for quantum technologies using Wigner crystals. The most prominent is a proposal for quantum computation using electrons on the surface of liquid helium~\cite{Platzman1999}. In this scheme, metal electrodes are placed below the electron gas so that image charges form which provide a trapping potential. The two logical qubit states are similar to Rydberg states in an atom, and have different average displacements from the surface which could be used to create a 2-qubit interaction via the image charges. Another existing scheme is to use a Wigner crystal made from trapped ions~\cite{Baltrusch2011,Taylor2008}, where qubit operations are performed using a laser field. The use of Wigner crystals as quantum communication channels has been mentioned as a possibility in carbon nanotubes~\cite{Deshpande2008}, but to date as far as the author is aware there have been no studies that have explored the viability of such a proposal.
\subsection{Instanton approach to calculate exchange coupling}
\label{sec:Instanton}
For a system of $N$ electrons constrained to move in 2D, we define collective spatial and spin coordinates $\mathbf{R} = (\mathbf{r}_1,\mathbf{r}_2,...,\mathbf{r}_N)=(x_1,y_1,x_2,y_2...,x_N,y_N)$, $\underline{\sigma} = (\sigma_1,\sigma_2,...,\sigma_N)$, and we use $R_j$ to denote the $j^{th}$ element of $\mathbf{R}$, such that $R_{2j-1} = x_j$, $R_{2j} = y_j$. We assume that the electrons are in the Wigner crystal regime, and that the temperature is low enough compared to the Debye temperature so there are negligible phonon excitations, so the electrons are localised near to $N$ `lattice sites' $\mathbf{\bar{R}} = (\bar{x}_1,\bar{y}_1,...,\bar{x}_{N},\bar{y}_N )$.
Under these conditions we approximate the electron wavefunction using localised orbital wavefunctions $\{ \ket{ \phi_n(\mathbf{r})} \}$, where orbital $n$ is centred around lattice site $\bar{\mathbf{r}}_n$, and such that there is still enough overlap between adjacent orbital wavefunctions to allow the electrons to exchange. We label one configuration of the orbital wavefunction as $\ket{ \Phi( \bar{\mathbf{R}} )} := \ket{ \phi_1(\mathbf{r}_1)} \ket{ \phi_2(\mathbf{r}_2)} ... \ket{ \phi_N(\mathbf{r}_N)}$, corresponding to when electron $n$ occupies the $n^{th}$ orbital wavefunction. Clearly this is not the only way of distributing the electrons; the electron positions can be permuted in $N!$ different ways, and we denote spatial permutations of $\ket{\Phi( \bar{\mathbf{R}} )}$ as $\ket{\Phi( P^{\textsc{r}}\bar{\mathbf{R}} )} = \hat{P}^{\textsc{r}}\ket{\Phi( \bar{\mathbf{R}} )}$, where $P^{\textsc{r}}$ is a spatial permutation of the electrons. Within this space of localised orbitals, the Hamiltonian effectively performs
permutations of the electron positions, since it only contains terms acting on the spatial degree of freedom. Thus the Hamiltonian takes the form
\begin{align}
H_\textsc{r} \propto \sum_{\hat{P}^{\textsc{r}} \neq \idop} C_P \hat{P}^{\textsc{r}}.
\end{align}
To determine the coupling constants $C_P$ for each permutation, we follow the argument of Thouless~\cite{Thouless1965}, referring to each permutation $|\Phi( \hat{P}^{\textsc{r}} \bar{\mathbf{R}} ) \rangle$ of the electrons in orbitals as a `cavity'. We imagine that it is possible to set the system up so that no tunnelling between different cavities can occur. Then suppose we arrange the system so that tunnelling between $\ket{\Phi( \bar{\mathbf{R}} )}$ and $|\Phi( \hat{P}^{\textsc{r}}\bar{\mathbf{R}} ) \rangle$ can occur. This can be treated exactly the same as a particle in a double potential well; there are symmetric and antisymmetric combinations of the states $\ket{\psi^\pm_P} = \frac{1}{\sqrt{2}}\ket{\Phi( \bar{\mathbf{R}} )} \pm |\Phi( \hat{P}^{\textsc{r}}\bar{\mathbf{R}} )\rangle$ with energies $E^\pm_P$. It can be shown~\cite{Thouless1965,Roger1984} that $E^+_P < E^-_P$ so that defining $J_P : = \frac{1}{2} (E^-_P - E^+_P)$ the effective Hamiltonian between the cavities $\ket{\Phi( \bar{\mathbf{R}} )}$ and $|\Phi( \hat{P}^{\textsc{r}}\bar{\mathbf{R}} )\rangle$ can be written
\begin{align}
\langle \Phi( \bar{\mathbf{R}} ) | H_\textsc{r} | \Phi( \hat{P}^{\textsc{r}}\bar{\mathbf{R}} ) \rangle = - J_P
\end{align}
Similarly for all pairs of cavities, such a term in the Hamiltonian will appear, and under the assumption that the connections between these cavities does not significantly alter the cavities themselves (i.e.\ assuming that the wavefunctions overlap without being distorted too much) the coupling for each permutation is roughly given by the singlet-triplet splitting $J_P$. Thus the effective Hamiltonian constrained to the localised orbitals can be written
\begin{align}
H_\textsc{r} \simeq - \sum_{\hat{P}^{\textsc{r}} \neq \idop} J_P \hat{P}^{\textsc{r}}.
\end{align}
The effects of this spatial Hamiltonian can also be expressed as an effective Hamiltonian acting on the spin degree of freedom, by recalling that for fermions the permutation $P$ of fermion labels results in a sign of $(-1)^{m_P}$ being applied to the state, where $m_P$ is the number of pairwise swaps that the permutation can be expressed as. The full permutation operator $\hat{P}$ can be decomposed as $\hat{P} =\hat{P}^{\textsc{r}} \hat{P}^\sigma$, where $\hat{P}^\sigma$ is an operator that permutes the spin degree of freedom, so that $\hat{P}^{\textsc{r}} = (-1)^{m_P} (\hat{P}^\sigma)^{-1}$, and by redefining $(\hat{P}^\sigma)^{-1} \to \hat{P}^\sigma$ we find that the effective spin Hamiltonian is the multi-spin exchange (MSE) Hamiltonian (derived more formally in~\cite{Thouless1965,Roger1984}):
\begin{align}\label{eqn:HThouless}
H_\textsc{mse} \simeq - \sum_{\hat{P}^\sigma \neq \idop} (-1)^{m_P} \hat{P}^{\sigma} J_{P}.
\end{align}
Thus the effective Hamiltonian of the Wigner crystal can equivalently be described in terms of the spin degrees of freedom evolving by the Hamiltonian $H_{\textsc{mse}}$, and this is the Hamiltonian that we will use in this chapter to transfer quantum information encoded in the spin degree of freedom. The operators $\hat{P}^\sigma$ can be related to standard Pauli spin matrices, for example the operator that permutes spins 1 and 2 can be written $\hat{P}^\sigma_{1,2} = \frac{1}{2}(\idop + E_{12})$, and more complicated permutations can be arrived at using the identities in~\cite{Klein1980} (we will see examples of these in Section~\ref{sec:WignInfTrans}). The difficulty lies in calculating the exchange coupling $J_P$, which is generally very hard to do except by using approximations. In this study we will use the path integral instanton (or many-dimensional WKB) method (for an overview of this method, see for example~\cite{Altland2010}). This method has been used extensively in the literature, first by Roger~\cite{Roger1983,Roger1984}, which was successful in qualitatively describing the behaviour of the Weiss temperature on the surface of solid $^3$He, and subsequently by many others e.g.~\cite{Ashizawa2000,Katano2000,Voelker2001,Hirashima2001,Fogler2005,Klironomos2005,Klironomos2007,Meyer2009,Candido2011}. The instanton approach assumes that quantum effects are small and that most of the evolution occurs near the classical trajectories, so that a semiclassical description is appropriate. This reproduces path-integral Monte Carlo results for values of $r_s$ far from the Wigner crystal melting point where the dynamics is close to the classical dynamics~\cite{Bernu2001}, but is not known to be accurate near the melting point where quantum effects can become more dominant. Also it is not likely to be accurate for Wigner crystals formed as a result of impurity pinning, since these may have small $r_s$ values and thus large quantum effects.
In order to apply the path integral instanton approximation, we first derive a particularly useful expression which puts $J_P$ in terms of path integrals. We start with the \emph{propagator} $G( \bar{\mathbf{R}}; P^{\textsc{r}} \bar{\mathbf{R}}; T)$ defined as
\begin{align}\label{eqn:prop1}
G( \bar{\mathbf{R}}; P^{\textsc{r}} \bar{\mathbf{R}}; T) := \bra{P^{\textsc{r}}\bar{\mathbf{R}} } e^{- i H_{\textsc{r}} T/ \hbar} \ket{ \bar{\mathbf{R}} }.
\end{align}
where $\langle \mathbf{R} \ket{\bar{\mathbf{R}}}: = \delta(\mathbf{r}_1 - \bar{\mathbf{r}}_1) \delta(\mathbf{r}_2 - \bar{\mathbf{r}}_2)... \delta(\mathbf{r}_N - \bar{\mathbf{r}}_N)$ is an eigenstate of the position operator, where electrons lie exactly on the lattice sites.
To simplify the propagator, we express $H_\textsc{r}$ in terms of the eigenstates $\ket{\psi^\pm_P}$ given above that mix the $\ket{\Phi( \bar{\mathbf{R}} )}$ and $\ket{\Phi( P^{\textsc{r}} \bar{\mathbf{R}} )}$ states. Expressing $ e^{- i H_{\textsc{r}} T/ \hbar}$ in this subspace, and defining $E_0$ such that $E^+_P = E_0 - J_P$, $E^-_P = E_0 + J_P$, gives
\begin{align}\label{eqn:WigEvo}
\left. e^{-iH_\textsc{r} T / \hbar} \right|_P &= \ket{\psi^+_P}\bra{\psi^+_P} e^{-iH_\textsc{r} T / \hbar} \ket{\psi^+_P}\bra{\psi^+_P} + \ket{\psi^-_P}\bra{\psi^-_P} e^{-iH_\textsc{r} T / \hbar}\ket{\psi^-_P}\bra{\psi^-_P} \nonumber \\
&= e^{i (J_P - E_0) T/\hbar} \ket{\psi^+_P}\bra{\psi^+_P} + e^{-i (J_P + E_0) T/\hbar} \ket{\psi^-_P}\bra{\psi^-_P}.
\end{align}
where we use the notation $\left. U \right|_P$ to indicate operator $U$ constrained to the subspace spanned by $\left\{ \ket{\Phi( \bar{\mathbf{R}} )} , \ket{\Phi(P^{\textsc{r}} \bar{\mathbf{R}} )} \right\}$.
We make the assumption that the amplitude of each orbital wavefunction is negligible at other lattice sites, so $\int \delta(\mathbf{r}_k - \bar{\mathbf{r}}_l) \phi_m(\mathbf{r}_k) d \mathbf{r}_k \simeq \delta_{lm} c$, where $c$ is a complex number. From this assumption it follows that $\left\langle \bar{\mathbf{R}} \right. \left| \psi^\pm_P \right\rangle \simeq \left\langle \bar{\mathbf{R}} \right. \left| \Phi( \bar{\mathbf{R}} ) \right\rangle := \frac{C}{\sqrt{2}}$, $\left\langle P^{\textsc{r}} \bar{\mathbf{R}} \right. \left| \psi^\pm_P \right\rangle \simeq \pm \left\langle P\bar{\mathbf{R}} \right. \left| \Phi( P^{\textsc{r}} \bar{\mathbf{R}}) \right\rangle := \pm\frac{C}{\sqrt{2}}$. Inserting the evolution operator $\left. e^{-iH_\textsc{r} T / \hbar} \right|_P$ into the propagator in (\ref{eqn:prop1}) and simplifying using these identities gives
\begin{align}
G( \bar{\mathbf{R}}; P^{\textsc{r}} \bar{\mathbf{R}}; T)&= \bra{P^{\textsc{r}} \bar{\mathbf{R}} } e^{- i H_{\textsc{r}} T/ \hbar} \ket{ \bar{\mathbf{R}} }= \bra{P^{\textsc{r}} \bar{\mathbf{R}} } \left. e^{-iH_\textsc{r} T / \hbar} \right|_P \ket{ \bar{\mathbf{R}} } \nonumber\\
&\simeq e^{-i E_0 T/\hbar} |C|^2 A_P \sin \left( \frac{J_PT}{\hbar} \right),
\end{align}
where $A_P$ accounts for permutations that are self-inverse, and so are counted twice in the Hamiltonian (for our purposes only two-body exchange processes are self-inverse, so that $A_P = 2$ for 2-body exchange and $A_P = 1$ otherwise). A similar calculation yields $G( \bar{\mathbf{R}}; \bar{\mathbf{R}}; T) \simeq e^{-i E_0 T/\hbar} |C|^2 \cos \left( \frac{J_P T}{\hbar} \right)$. Dividing these two expressions gives
\begin{align}
\frac{G( \bar{\mathbf{R}}; P^{\textsc{r}} \bar{\mathbf{R}}; T)}{G( \bar{\mathbf{R}}; \bar{\mathbf{R}}; T) } \simeq A_P \tan \left( \frac{J_P T}{\hbar} \right) .
\end{align}
When the time $T$ is sufficiently small with respect to $J/\hbar$, then $\tan \left( \frac{J_P T}{\hbar} \right) \simeq \frac{J_P T}{\hbar}$, so that
\begin{align}\label{eqn:J}
J_P \simeq \frac{\hbar A_P}{T}\frac{G( \bar{\mathbf{R}}; P^{\textsc{r}} \bar{\mathbf{R}}; T) }{G( \bar{\mathbf{R}};\bar{\mathbf{R}}; T)},
\end{align}
The expression for $J_P$ in eqn.\ (\ref{eqn:J}) is in a suitable form that approximation techniques based on path integrals can be used. The path integral form of the propagator can be written, assuming a Hamiltonian of the form $H = \sum_i \frac{\mathbf{p}_i^2}{2m} + V(\mathbf{R})$ (see e.g.~\cite{Altland2010}):
\begin{align}\label{eqn:FullProp}
&G( \bar{\mathbf{R}},P^{\textsc{r}} \bar{\mathbf{R}} ; T) =\bra{P^{\textsc{r}} \bar{\mathbf{R}} } e^{-i H T} \ket{ \bar{\mathbf{R}}} =\int_{\mathbf{R}(0) = \bar{\mathbf{R}}}^{\mathbf{R}(T) = P^{\textsc{r}} \bar{\mathbf{R}}} D\mathbf{R}\text{ exp}\left[ - \frac{i}{\hbar} \int_0^{T} dt \frac{m}{2} \left(\frac{d \mathbf{R}}{d t} \right) ^2+ V(\mathbf{R}) \right] \nonumber\\
&=\int_{\mathbf{R}(0) = \bar{\mathbf{R}}}^{\mathbf{R}(T) = P^{\textsc{r}} \bar{\mathbf{R}}} D\mathbf{R}\text{ exp}\left[ - \frac{i}{\hbar} S[\mathbf{R}] \right].
\end{align}
where $S[\mathbf{R}]$ is the action along the path $\mathbf{R}(t)$. To evaluate this path integral, we first perform a Wick rotation $t \to -i \tau$, after which the propagator becomes an imaginary time propagator:
\begin{align}
&G( \bar{\mathbf{R}},P^{\textsc{r}} \bar{\mathbf{R}} ; T_\tau) =\bra{P^{\textsc{r}} \bar{\mathbf{R}} } e^{- H T_\tau } \ket{ \bar{\mathbf{R}}} =\int_{\mathbf{R}(0) = \bar{\mathbf{R}}}^{\mathbf{R}(T_\tau) = P^{\textsc{r}} \bar{\mathbf{R}}} D\mathbf{R}\text{ exp}\left[ - \frac{1}{\hbar} S[\mathbf{R}] \right].
\end{align}
We then use a semiclassical (stationary phase) approximation to approximate this imaginary-time propagator, whereby the quantum dynamics are approximated by quantum fluctuations around the classical path of least action $\mathbf{R}_{cl}(t)$ which satisfies the boundary conditions $\mathbf{R}_{cl}(0) = \bar{\mathbf{R}}$, $\mathbf{R}_{cl}(T_\tau) = P^{\textsc{r}} \bar{\mathbf{R}}$. This classical solution that connects $\bar{\mathbf{R}}$ and $P^{\textsc{r}} \bar{\mathbf{R}}$ is known as an \emph{instanton}. The name instanton reflects the fact that this solution is a kink in the temporal profile of the particles (see Fig.~\ref{fig:Instanton}), and so can be interpreted as a type of particle in time. The width of the instanton depends on the second derivative of the potential at $\bar{\mathbf{R}}$. In theory, any number of instantons connecting $\bar{\mathbf{R}}$ and $P^{\textsc{r}} \bar{\mathbf{R}}$ are solutions to the equation of motion. To get around this, we make an assumption (often called the `dilute gas' approximation), that the distance between instantons $\Delta t$ is much larger than the instanton width $\delta t$. Then if the propagator is evaluated at times $T_\tau$ that satisfy $\delta t \ll T_\tau \ll \Delta t$, we can assume that the contribution from multi-instanton solutions is negligible and so only single instanton solutions need to be taken into account. The point in time at which these instantons occur is also arbitrary, which will be an an important factor later on.
\begin{figure}[h]
\begin{center}
\includegraphics[width = 0.5\textwidth]{Instanton-eps-converted-to.pdf}
\caption{\label{fig:Instanton} A schematic diagram of an `instanton' solution between the two configurations $\bar{\mathbf{R}}$ and $P^{\textsc{r}} \bar{\mathbf{R}}$. For simplicity the second derivative of the potential is denoted as $\omega^2$, but in general there will be many frequencies for higher dimensional problems.}
\end{center}
\end{figure}
We assume that the dilute gas approximation is valid, so that the classical path of least action is a single instanton. We then consider Gaussian fluctuations about this classical instanton path; for a fluctuation $\mathbf{R} (t)= \mathbf{R}_{cl}(t) + \mathbf{v}(t)$ such that $\mathbf{v}(0) = \mathbf{v}(T_\tau) =0$, we can Taylor expand the action to second order (see e.g.~\cite{Riley2006,Altland2010})
\begin{align}
S[\mathbf{R} ] \simeq S[ \mathbf{R}_{cl} ] + \int_0^{T_\tau} dt \mathbf{v}^T(\tau) \cdot \left. \mathbf{A}(\tau)\right|_{\mathbf{R} = \mathbf{R}_{cl}} \cdot \mathbf{v}(\tau),
\end{align}
where
\begin{align}\label{eqn:A}
\mathbf{A}(\tau) = -m \partial_{\tau}^2 \idop + \mathbf{H}(\tau),
\end{align}
and $\mathbf{H}$ is the time-dependent Hessian matrix defined as
\begin{align}
\mathbf{H}_{ij}(\tau) := \left. \frac{\partial^2 V}{\partial R_i(\tau) \partial R_j(\tau)} \right|_{\mathbf{R} = \mathbf{R}_{cl}},
\end{align}
where we use the notation $R_{2n-1 }(\tau) = x_n(\tau)$, $R_{2n}(\tau) = y_n(\tau)$.
The stationary phase approximation then approximates the full action in (\ref{eqn:FullProp}) by the second order Taylor expansion, giving
\begin{align}\label{eqn:Fdefin}
G( \bar{\mathbf{R}};P^{\textsc{r}} \bar{\mathbf{R}} ; T_\tau ) &\simeq e^{-\frac{1}{\hbar}S[\mathbf{R}_{cl}]} \int_{\mathbf{v}(0) = 0}^{\mathbf{v}({T_\tau}) = 0} D\mathbf{v}(\tau) \exp \left[ -\frac{1}{2\hbar} \int_0^{T_\tau} d\tau \mathbf{v}^T(\tau) \cdot \left. \mathbf{A}(\tau)\right|_{\mathbf{R} = \mathbf{R}_{cl}} \cdot \mathbf{v}(\tau)\right] \nonumber\\
&:= e^{-\frac{1}{\hbar}S[\mathbf{R}_{cl}]} F[\mathbf{R}_{cl} ].
\end{align}
and similarly $G( \bar{\mathbf{R}};P^{\textsc{r}} \bar{\mathbf{R}} ; T_\tau ) = F[\mathbf{R}_{cl} ]$, where $F[\bar{\mathbf{R}}]$ is defined with $\mathbf{R}(t) = \bar{\mathbf{R}}$, and we have defined $V(\mathbf{R})$ to be zero when the electrons are at positions $\bar{\mathbf{R}}$ so that the classical action for $G( \bar{\mathbf{R}} ;\bar{\mathbf{R}} ;T_\tau)$ is zero.
Putting this all into equation (\ref{eqn:J}), the expression for $J_P$ becomes
\begin{align}\label{eqn:JP}
J_P \simeq \frac{\hbar A_P}{T_\tau}\frac{ F[\mathbf{R}_{cl} ] }{ F[\bar{\mathbf{R}} ] } e^{-\frac{1}{\hbar}S[\mathbf{R}_{cl}]},
\end{align}
for a process which permutes the electron positions by $P^{\textsc{r}}$, and for sufficiently small $T_\tau$. It appears counter-intuitive that a factor of time should appear in the denominator, however later on we will see that this will cancel out. The factors of $F[\mathbf{R}_{cl} ]$ in eqn.\ (\ref{eqn:JP}) are still quite unwieldy and can be put in a much more convenient form by noting it is a multidimensional Gaussian integral, and so using standard functional integration results it can be written~\cite{Altland2010}
\begin{align}
F[\mathbf{R}_{cl}] &= K \frac{1}{\sqrt{\det( \mathbf{A} )}} = K \frac{1}{\sqrt{\det( -m \partial_{\tau}^2 \idop + \mathbf{H}(\tau) )}},
\end{align}
where any constants have been absorbed into the prefactor $K$. A similar calculation for $F[ \bar{\mathbf{R}} ]$ gives
\begin{align}
F[\bar{\mathbf{R}} ] = K \frac{1}{\sqrt{\det( -m \partial_{\tau}^2 \idop + \mathbf{H}^{(0)}(\tau) )}} ,
\end{align}
where $\mathbf{H}^{(0)}(\tau) := \left. \mathbf{H} \right|_{\mathbf{R}(\tau) = \bar{\mathbf{R}} }$.
We must be careful that all the eigenvalues of $-m \partial_{\tau}^2 \idop + \mathbf{H}(\tau)$ are non-zero, to avoid $F[\mathbf{R}_{cl}]$ diverging. As noted earlier, the position of the instanton in time is arbitrary, and this leads to a eigenvector with zero eigenvalue (or small eigenvalue for finite systems) which corresponds to translation of the instanton in time. Since the positions of the instantons are not important for this calculation, we can ignore this zero mode and only consider the effects of the remaining non-zero modes. This can be done by a suitable change of variables in the functional integration, introducing an extra multiplicative factor~\cite{Zinn2002}
\begin{align}
F[ \mathbf{R}_{cl}] = \frac{ K T_\tau}{\hbar} \left( \frac{S[\mathbf{R}_{cl}]}{2 \pi \hbar m} \right)^{1/2} \frac{1}{\sqrt{\det'( -m \partial_{\tau}^2 \idop + \mathbf{H}(\tau) )}},
\end{align}
where $\det'$ indicates the determinant without the lowest eigenvalue. Combining these results all together, we find that
\begin{align}\label{eqn:JPfinal}
J_P \simeq A_P \left( \frac{S[\mathbf{R}_{cl}]}{2 \pi \hbar m} \right)^{1/2} \sqrt{ \frac{ \det( -m \partial_{\tau}^2 \idop + \mathbf{H}^{(0)}(\tau) ) }{ \det'( -m \partial_{\tau}^2 \idop + \mathbf{H}(\tau) ) } } e^{-\frac{1}{\hbar}S[\mathbf{R}_{cl}]}.
\end{align}
This expression is in a convenient form since it can, in principle, be calculated once the classical path of least action is known. For this study, we will find this path numerically, however before discussing these numerical methods, we first introduce the particular setup we will use to simulate the Wigner crystal, and introduce some important length scales.
\subsection{Proposed setup}
For this scheme, we consider a 2-dimensional system in the Wigner crystal regime, at suitably low temperatures that the MSE model is applicable. We consider a parabolic confining potential $\frac{1}{2}m^* \Omega^2 y^2$ in the $y$ direction, with barriers placed at $y = \pm d/2$, and $N$ electrons placed between these barriers such that the density of electrons is $N/d$ (such a setup is also often called a \emph{Wigner molecule}). Since this potential is mirror symmetric around the $x$-axis, more potentials must be introduced to break this symmetry and make the ground state unique. In our scheme this is done by trapping the two end electrons such that they are moved off centre (in opposite directions or the same directions for odd and even chains respectively), and isolated from the rest of the chain using an internal barrier (see Fig.~\ref{fig:Potential}). We refer to these end electrons as quantum `dots'. Not only does this lift the degeneracy, these two end electrons isolated or coupled to the chain depending on the internal barrier height, and so can act as sender and receiver when sending quantum information through the chain.
\begin{figure}[h]
\begin{center}
\includegraphics[width = 0.8\textwidth]{PotentialExample.jpg}
\caption{\label{fig:Potential} An illustration of the potential given in eqn.\ (\ref{eqn:Potential})}
\end{center}
\end{figure}
We model the potentials as quadratic potentials modulated by Gaussians, with the full potential $V(\mathbf{R})$ given by
\begin{align}\label{eqn:Potential}
&V(\mathbf{R}) = \sum_{i=1}^N \left[ \sum_{j < i}\frac{e^2}{4 \pi \epsilon |\mathbf{r}_i - \mathbf{r}_j|} + \frac{1}{2}m^* \Lambda^2 \left( e^{-\frac{ (x_i - x_0)^2}{2 \sigma^2}} (y_i + (-1)^{N+1} y_0)^2 +e^{-\frac{(x_i + x_0)^2}{ 2 \sigma^2}} (y_i+y_0)^2 \right) \right. \nonumber\\
&\left. + \frac{1}{2}m^* \Omega^2 y_i^2 + h_{L} e^{ - \frac{(x_i + d/2 - l)^2 }{ 2 w^2}} + h_{R} e^{ - \frac{(x_i - d/2 + l)^2 }{ 2 w^2} }+ h_{0} \left( e^{ - \frac{ (x_i - d/2)^2 }{ 2 w_{out}^2}} + e^{ - \frac{(x_i + d/2)^2 }{ 2 w_{out}^2}} \right) \right]
\end{align}
For an illustration of this potential, see Fig.~\ref{fig:Potential}. Note that we have included different barrier heights $h_L,h_R$ for the left and right dots, to allow the coupling between these dots and the chain to be adjusted. There is also a factor of $(-1)^{N+1}$ so that the offset lifts the degeneracy appropriately for odd or even electrons. The possible experimental implementation that we envisage is a 2-dimensional electron gas in GaAs, using surface electrodes to create the potential, but our results extend to any realisation of a Wigner crystal.
The dot displacement $y_0$ is chosen so that the mirror inversion symmetry is broken sufficiently to prevent excitations into the reflected configuration. Let the two degenerate ground states be $\psi_1$ and $\psi_2$. One way to arrive at a rough condition for $y_0$ is to make sure that the probability of thermal excitation from $\psi_1$ to $\psi_2$ is sufficiently low. The probability of excitation from $\psi_1$ to $\psi_2$, in the presence of the symmetry breaking potential $\frac{1}{2}m^* \Lambda^2 (y-y_0)^2$, is roughly $\varepsilon$, where
\begin{align}
\varepsilon \sim \exp ( -\Delta E \beta) = \exp \left(- 2 m^*\Lambda^2 y_0^2 \beta \right)
\end{align}
and $\beta$ is the inverse temperature. To obtain an estimate, we assume that the system is realised in GaAs/AlGaAs at a temperature of $T\simeq 5 \text{mK}$. The minimum size of $\hbar \Omega$ can be estimated from the band splitting in GaAs/AlGaAs, which can be as low as $0.3$meV~\cite{Daneshvar1997}. Inserting this into the expression above, we find $\varepsilon \sim \exp( -2\times 10^{19} y_0^2)$, so choosing $y_0 \sim 10$nm can give a sufficiently low error, and additionally should be within the precision available in experiments (using the characteristic length units $r_0$ defined below, 10nm$\simeq 0.1 r_0$ for GaAs).
The remaining parameters are chosen based on the electron configuration without the dot potentials present (i.e.\ when $h_L = h_R = \Lambda = 0$). For each particular situation, the minimum energy configuration $\bar{ \mathbf{R}} = (\bar{x}_1,\bar{y}_1,\bar{x}_2,\bar{y}_2,...,\bar{x}_N,\bar{y}_N)$ was found without the dot potentials, and then the parameters were chosen such that the internal barriers would sit roughly halfway between the end pairs of electrons, and such that the internal barriers have as little effect as possible on the separation between the dot electron and the chain. Thus $l = \frac{d}{2} - \frac{1}{2}(\bar{x}_1 + \bar{x}_2)$, and empirically $w = \frac{l}{8}$, $w_{out} =\frac{l}{8}$ was found to have little effect on the electron equilibrium positions. We set $\sigma = l/2$, so that the effect of the two quadratic potentials for the dots is confined to within the barriers. The size of $\Lambda$ in eqn.\ (\ref{eqn:Potential}) was chosen to be $4\Omega$, so that it is much stronger than the confining potential of the chain. This was also empirically found to help the numerical algorithm converge to one minimum, and not a local minimum.
It is also useful to introduce a characteristic length scale of the system, to ease the notation, and so that the calculations in the following sections are not dependant on the particular parameters used, but are instead more general. Following~\cite{Klironomos2007}, we define $r_0$ as the length scale at which the Coulomb repulsion between electrons is comparable to the potential energy of the parabolic confinement in the $y$-direction:
\begin{align}\label{eqn:r0}
\frac{1}{2}m^* \Omega^2 r_0^2 = \frac{e^2}{4 \pi \epsilon r_0} \Longrightarrow r_0 = \left( \frac{ e^2}{4 \pi \epsilon m^* \Omega^2} \right)^{1/3}
\end{align}
We also define $r_\Omega$, which is $r_0$ in units of the Bohr radius $a_B = 4 \pi \epsilon \hbar^2 / m^* e^2$:
\begin{align}\label{eqn:rOmega}
r_{\Omega} = 2 \left( \frac{e^4 m^*}{32 \pi^2 \epsilon^2 \hbar^3 \Omega} \right)^{2/3}
\end{align}
Using these units, we also define a dimensionless density of the electrons $\nu = nr_0$ as the number of electrons in a length $r_0$. Thus in these units, $r_s = r_0/a_B \nu = r_\Omega /\nu$.
By making a change of variables of $R_n \to R_n r_0$, and $\tau \to \tau \frac{\sqrt{2}}{\Omega}$ the action then becomes
\begin{align}
S[ \mathbf{R} ] &= \int_0^{T_\tau} d\tau \left[ \frac{m\dot{\mathbf{R}}^2}{2} + V(\mathbf{R}) \right]=\int_0^{T_\tau} d\tau \left[ \sum_j \frac{m\dot{\mathbf{r}}_j^2}{2} + \frac{1}{2}m \Omega^2 y_j^2 + \sum_{j < i} \frac{e^2}{4 \pi \epsilon |\mathbf{r}_i - \mathbf{r}_j|} + \sum_j v(\mathbf{r}_j) \right]\nonumber\\
&\to \hbar \sqrt{r_{\Omega}} \int_0^{T_\tau} d\tau \left[ \sum_j \frac{\dot{\mathbf{r}}_j^2}{2} + y_j^2 + \sum_{j < i} \frac{1}{|\mathbf{r}_i - \mathbf{r}_j|} + \frac{1}{\hbar \sqrt{r_\Omega}}\sum_j v(\mathbf{r}_j)\right] :=\hbar \sqrt{r_{\Omega}} \text{ } \eta[\mathbf{R} ],
\end{align}
where $\eta[\mathbf{R} ]$ is the dimensionless action, $v(\mathbf{r}_j)$ contains all of the remaining terms in $V(\mathbf{R})$, and we have used the identity $\frac{1}{\sqrt{2}} m \Omega_y r_0^2 = \frac{1}{\sqrt{2}} m \Omega_y r_{\Omega}^2 a_B^2 = \hbar \sqrt{r_{\Omega}}$. Performing a similar change of variables on the prefactor $F[\mathbf{R}_{cl}]$, we obtain
\begin{align}
F[\mathbf{R}_{cl}] = \int_{\mathbf{v}(0) = 0}^{\mathbf{v}({T_\tau}) = 0} D\mathbf{v}(\tau) \exp \left[ -\frac{\sqrt{r_{\Omega}}}{2} \int_0^{T_\tau} d\tau \text{ }\mathbf{v}^T(\tau) \cdot \left[ -\partial_{\tau}^2\idop + \frac{1}{\hbar\sqrt{r_{\Omega}}} \mathbf{H}(\tau) \right] \cdot \mathbf{v}(t)\right].
\end{align}
So overall in these units the exchange energy is
\begin{align}
\label{eqn:Jreduced}
J_P \simeq \frac{ e^2 }{4 \pi \epsilon a_B} r_\Omega^{-5/4} A_P \left( \frac{\eta[\mathbf{R}_{cl}] }{2 \pi } \right)^{1/2} \sqrt{ \frac{ \det( -\partial_{\tau}^2 \idop + \frac{1}{\hbar \sqrt{r_\Omega}}\mathbf{H}^{(0)}(\tau) ) }{ \det'( - \partial_{\tau}^2 \idop + \frac{1}{\hbar \sqrt{r_\Omega}} \mathbf{H}(\tau) ) } } e^{-\sqrt{r_\Omega} \eta[\mathbf{R}_{cl}]}.
\end{align}
This is the form of $J_P$ which will be used in the numerical simulations below. Notice that the factors of $\hbar / \sqrt{r_\Omega}$ will disappear if $\mathbf{H}^{(0)}(\tau)$ is expressed with a change of variables of $R_n \to R_n r_0$, and $\tau \to \tau \frac{\sqrt{2}}{\Omega}$. For the remainder of this chapter, we will also give coupling strengths in units of $\Omega$, unless explicitly stated otherwise.
\subsection{Calculating the exchange numerically}
We now describe the process of numerically evaluating eqn.\ (\ref{eqn:JPfinal}) for the potential given in the previous section. The first step is to find the configuration of positions $\bar{\mathbf{R}}$ that minimises $V(\mathbf{R})$, which was done numerically using the built-in MATLAB algorithm \texttt{fminunc}. The minimisation was performed first with $h_L = h_R = \Lambda=0$, to find the spacing between the electrons with a pure harmonic potential in the $y$-direction, so that appropriate values of $l,x_0$ could be inferred. Then the full minimisation was performed with these parameters, starting with the electrons evenly spaced along the $x$-axis with a slight zig-zag perturbation in the $y$-direction.
\begin{figure}[h!]
\subfloat[]{\includegraphics[width=0.5\textwidth]{PlotD=50-eps-converted-to.pdf}}
\subfloat[]{\includegraphics[width=0.5\textwidth]{PlotD=90-eps-converted-to.pdf}}\\
\subfloat[]{\includegraphics[width=0.5\textwidth]{PlotD=100-eps-converted-to.pdf}}
\subfloat[]{\includegraphics[width=0.5\textwidth]{PlotD=120-eps-converted-to.pdf}}
\caption{\label{fig:Density} Classical equilibrium positions for 9 electrons, at densities of a) $\nu =$0.5, b) $\nu =0.9$, c) $\nu = 1$ and d) $\nu = 1.2$. Distances are shown in units of $r_0$, and note the different scales on the axes as density is increased.}
\end{figure}
Results for different density are shown in Fig.~\ref{fig:Density}. Around $\nu=0.9$ there is a transition from a linear chain to a zig-zag chain (since at $r_0=1$ the strength of the confining potential becomes similar to the Coulomb repulsion), and towards higher densities we see that there are two electrons almost equidistant to each dot electron.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.6\textwidth]{WignerExchanges-eps-converted-to.pdf}
\caption{\label{fig:WigExch} An illustration of the types of exchange processes that we consider in this work. $J^{(1)}_{mn}$ indicates pairwise exchange between electrons $m$ and $n$. $J^{(R3}_{n}, J^{(R4)}_n$, $J^{(R5)}_n$, $J^{(R6)}_n$ indicates 3- 4- 5- and 6-body ring exchange starting clockwise from electron $n$.}
\end{center}
\end{figure}
The electrons are typically arranged near the vertices of triangles, and we considered permutation processes involving up to 6 electrons and up to 4$^{th}$-nearest neighbour pairwise exchanges, as shown in Fig.~\ref{fig:WigExch}. To find the strength of each of these permutation processes, we began by numerically finding the classical path of least action for each. To do this, we first discretised the action integral into $M$ equal time steps $\Delta \tau = T_\tau/M$, so that the action integral becomes a Riemann sum:
\begin{align}
\eta[\mathbf{R}] = \int_0^{T_\tau} d\tau \left[ \sum_j \frac{\dot{\mathbf{r}}_j^2}{2} + \frac{1}{\hbar \sqrt{r_\Omega}}V(\mathbf{R})\right] \rightarrow \sum_{m=1}^M \sum_j \frac{1}{\Delta \tau} \frac{ (\mathbf{r}_j(\tau_m) - \mathbf{r}_j(\tau_{m-1}))^2}{2} + \frac{\Delta \tau}{\hbar \sqrt{r_{\Omega}} } V(\mathbf{R}_m)
\end{align}
where $\tau_m = m\Delta \tau$ and $\mathbf{R}_m =(r_1(\tau_m),r_2(\tau_m),...,r_{2N}(\tau_m))$. Notice that this is the same equation as the energy of $M$ particles joined together with springs of stiffness $\frac{1}{ \Delta \tau}$, in a potential $\sum_{n=1}^M \frac{\Delta \tau}{\hbar \sqrt{r_{\Omega}} } V( \mathbf{R}_n)$, and with endpoints attached to $\bar{\mathbf{R}}$ and $P^{\textsc{r}} \bar{\mathbf{R}}$. Thus minimising the action in this discretised picture is the same as minimising the energy of this chain. This can be minimised using a 2-point steepest descent algorithm~\cite{Barzilai1988} which proceeds as follows, with the time-dependent positions of all of the particles at the $n^{th}$ iteration labelled as $\mathbf{R}^{(n)} (\tau)= \{ \mathbf{R}_1^{(n)}, \mathbf{R}_2^{(n)},...,\mathbf{R}_M^{(n)} \}$:
\begin{itemize}
\item Make an initial guess $\mathbf{R}^{(0)}(\tau)$.
\item For each step, $\mathbf{R}^{(n+1)}(\tau) = \mathbf{R}^{(n)}(\tau)- \alpha_n \mathbf{G}_n$, where $\mathbf{G}_n = \nabla_{ \mathbf{R} }\eta[\mathbf{R}^{(n)}(\tau)] $, $\alpha_0$ is a small step size (we use $\alpha_0 = 0.001$) and $\alpha_n = \mathbf{G}_n \cdot (\mathbf{R}^{(n)}(\tau) - \mathbf{R}^{(n-1)}(\tau)) / \|\mathbf{G}_n\|^2$ for $n > 0$~\cite{Barzilai1988}.
\item Stop if $\| \mathbf{G}_n \| < \varepsilon_c$ (convergence to a minimum), or if $\| \mathbf{R}^{(n)}(\tau) \| < \varepsilon_p$ (convergence to a point).
\end{itemize}
where $\| \cdot \|$ indicates the Euclidean norm. For our calculations, we used $T_\tau = 30$ and $M=70$, based on the work in~\cite{Katano2000} that shows reasonable convergence with these parameters for 27 electrons in a Wigner crystal (suggesting that for these parameters the instanton assumptions made in Sec~\ref{sec:Instanton} are valid). The accuracy of discretising time in this way has also been studied in~\cite{Richardson2011} by comparing numerical results to an analytically solvable system. Here they find that for $M=64$, $T_\tau = 30$ the error is on the order of 1\%. The tolerance used for the steepest descent algorithm was $\varepsilon_c = 1\times 10^{-6}$.
Once the classical path $\mathbf{R}_{cl}$ has been found using this algorithm, the prefactor must be calculated by diagonalising the matrix $\mathbf{A}$ in (\ref{eqn:Jreduced}) so that the determinant can be found. This matrix is calculated along the classical path, and in discretised form becomes a $2N(M+1) \times 2N(M+1)$ matrix of the form
\begin{align}
(\mathbf{A})_{R_i(\tau_m)R_j(\tau_n)} &= -\frac{1}{\Delta \tau^2} \left( -2\delta_{mn} + \delta_{m,n-1} + \delta_{m,n+1} \right) \delta_{ij} + \frac{\hbar}{\sqrt{r_\Omega}}\mathbf{H}_{ij}(\tau_n)\delta_{mn}
\end{align}
This matrix takes the following form
\begin{align}
\left[ \begin{array}{c c c c c c }
\frac{\hbar}{\sqrt{r_\Omega}}\mathbf{H}_{11}(\tau_1) +\frac{2}{\Delta \tau^2} &\frac{\hbar}{\sqrt{r_\Omega}}\mathbf{H}_{12}(\tau_1) & \dots & -\frac{1}{\Delta \tau^2}&0 & \\
\frac{\hbar}{\sqrt{r_\Omega}}\mathbf{H}_{21}(\tau_1) & \frac{\hbar}{\sqrt{r_\Omega}}\mathbf{H}_{22}(\tau_1) + \frac{2}{\Delta \tau^2} & \dots & 0 &-\frac{1}{\Delta \tau^2} & \\
\vdots & \vdots & \ddots & & & \\
-\frac{1}{\Delta \tau^2} & 0 & & \frac{\hbar}{\sqrt{r_\Omega}}\mathbf{H}_{11}(\tau_2) + \frac{2}{\Delta \tau^2}& \frac{\hbar}{\sqrt{r_\Omega}}\mathbf{H}_{12}(\tau_2) & \dots \\
0& -\frac{1}{\Delta \tau^2}& & \frac{\hbar}{\sqrt{r_\Omega}}\mathbf{H}_{21}(\tau_2) & \frac{\hbar}{\sqrt{r_\Omega}}\mathbf{H}_{22}(\tau_2) + \frac{2}{\Delta \tau^2} & \dots \\
& & &\vdots & \vdots & \ddots
\end{array}\right]
\end{align}
\subsection{Information transfer}
\label{sec:WignInfTrans}
To send information through the chain, we start with the quantum dots isolated from the rest of the chain (i.e.\ the internal barriers are high), then the internal barriers are dropped to turn on the interactions. This could be done either suddenly (a `quench') or slowly with respect to the couplings in the chain. For this work, we will study the viability of performing a quench, and leave the study of other protocols to later work.
The protocol used here depends on the number of electrons (see Fig.~\ref{fig:TransferProto}). For an odd number of electrons, the system is initialised with one barrier up and one barrier down, so that there are an even number of electrons coupled together (prepared in the ground state) and one isolated dot, which is prepared in the quantum state to be sent. The barrier is then lowered, coupling the isolated spin to the chain, and some time later the opposite barrier is raised to isolate the output electron. For an even number of electrons, the system is initialised with both barriers up, with one dot prepared in the quantum state to be transferred and the opposite dot prepared in a state such as $\ket{ \uparrow }$. The barriers are then simultaneously lowered and raised at the appropriate times. The reason for these different protocols is that intuitively we expect a chain of even electrons to have a unique ground state, since each electron can pair with one nearest neighbour, whereas for an odd chain there can be unpaired electrons which increases the degeneracy.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\textwidth]{TransferProtocol-eps-converted-to.pdf}
\caption{\label{fig:TransferProto} An illustration of the transfer protocol, a) using an odd number of electrons and b) using an even number of electrons. 1) The left barrier is up and the right barrier is down. The isolated (left) dot is prepared in a quantum state $\ket{\phi}$ and the remaining electrons are prepared in the ground state. For the even case, the opposite is also isolated and prepared in an $\ket{\uparrow}$ state. 2) The barriers are lowered, allowing information to travel through the chain. 3) The right barrier is raised at the optimal point to remove the transferred information.}
\end{center}
\end{figure}
For a quench, the potential problem is that such a sudden change will cause the electrons to move out of equilibrium, which could cause phonon excitations leading to decoherence and fluctuations in the exchange interactions. To assess how good this approach would be, we calculated the average change in equilibrium positions of the electrons after lowering the barriers quickly, starting with a barrier height of $5\Omega$ as this was found to reduce the dot-chain coupling to around $10^{-6}$ relative to the rest of the chain. To quantify the average change in equilibrium positions after the barriers drop ($\mathbf{R}$) relative to the equilibrium positions with the barriers up ($\mathbf{R}_0$), we use the following distance measure
\begin{align}\label{eqn:bard}
\bar{d} = \frac{1}{N}\frac{ \| \mathbf{R} - \mathbf{R}_0 \|}{\Delta \mathbf{R}_0},
\end{align}
where $\|\mathbf{R} \| $ is the Euclidean norm, and $\Delta\mathbf{R}_0$ is the average nearest-neighbour separation of the electrons.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.8\textwidth]{QuenchN=9_1Barr-eps-converted-to.pdf}
\caption{\label{fig:Quench9_2} $\bar{d}$ for a chain of 9 electrons, where $\bar{d}$ is defined in eqn.\ (\ref{eqn:bard}). Only the left barrier is lowered here, with $h_R=5$.}
\end{center}
\end{figure}
Results for lowering one barrier with $N=9$ are shown in Fig.~\ref{fig:Quench9_2}. In general $\bar{d}$ stays below about $10^{-2}$, and there are clear dips in $\bar{d}$ around $\nu = 0.6$ and $\nu = 1.2$, and little change over the range of $h_L$. There appears to be some instability around $\nu = 1$, which could be an artefact of the particular potentials used here or could be because this is the length scale where the coulomb repulsion becomes roughly equal to the trapping potential (from the definition of $r_0$). The results for lowering two barriers with either $N=10$ or $N=9$ were similar, but without the instability at $\nu = 1$, probably because two barriers are being lowered simultaneously, so the situation is more symmetric throughout.
The magnitude of the exchange coupling $J_P$ was then found using the instanton technique described above, with settings of $M=70$, $T_\tau = 30$, $h_R =h_L= 0.1$, $r_{\Omega} = 10$ (this value of $r_\Omega$ is chosen based on the band splitting of $\sim 0.3$meV in GaAs). Since the chains are symmetric, to speed up calculations only results for one half of the chain were taken, after confirming that the simulations were giving symmetric results (up to around 0.1 \% error, which confirms that the results are converging well).
Results are shown for up to 4$^{th}$-nearest neighbour and 6-body exchange processes in Figs.~\ref{fig:CouplingsN=9} \& \ref{fig:CouplingsN=10}. The couplings are in meV and are labelled $J^{(1)}_{mn}$ for pairwise exchange between electrons $m$ and $n$, and $J^{(R3)}_n,J^{(R4)}_n,J^{(R5)}_n,J^{(R6)}_n$ to indicate 3-, 4- , 5- and 6-body ring exchanges starting from electron $n$. The general features are low couplings at the boundary (due to the barrier between the dots and the chain) and a `U'-shaped coupling along the chain for low densities, becoming peaked at the middle for higher densities. We have excluded $\nu = 0.9$ and $\nu = 1$ for $N=9$, since for some couplings the algorithm did not converge to a minimum, which we attribute to the same instability seen for the quench in Fig.~\ref{fig:Quench9_2}.
Average couplings are shown in Fig.~\ref{fig:Javrg} for ten electrons, showing roughly which couplings dominate at different densities. Clearly the nearest-neighbour coupling dominates up to around $\nu = 0.8$, and the 3-body ring exchange is next highest, which is a qualitative agreement with other studies of Wigner crystals such as~\cite{Klironomos2007}. At around $\nu = 0.8$ the couplings all become of a similar magnitude, and there seems to be a plateau in coupling strength beyond this point. Although the average coupling shows some of the main behaviour, it is perhaps more informative to study the full coupling profile in Figs.~\ref{fig:CouplingsN=9} \& \ref{fig:CouplingsN=10}, from which it seems that including up to 2$^{nd}$ nearest neighbour and 5-body interactions captures the leading behaviour for $\nu \leq 0.8$. For $\nu > 0.8$, 6-body interactions and longer range pairwise interactions become significant. To be sure that our evolution is accurate, we therefore restrict ourselves to the regime $\nu \leq 0.8$ as it is not clear from our results whether higher order terms such as 7-body exchanges will also become important beyond this point.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.8\textwidth]{Javrg10-eps-converted-to.pdf}
\caption{\label{fig:Javrg} $\bar{J}$ for a chain of 10 electrons, for the different exchange processes. The notation is such that 3NN indicates 3$^{rd}$-nearest neighbour etc.\ and R3 means 3-body ring exchange etc. The distribution of couplings for $N=9$ follows a similar trend.}
\end{center}
\end{figure}
With this restriction, the Hamiltonian that we use to simulate the evolution is
\begin{align}\label{eqn:PermHam}
H =& \sum_{m \sim n } J_{mn}^{(2)} \hat{P}_{mn}^\sigma - \sum_{ \substack{ l,m,n \in \triangle \\ l < m < n} } J_{l}^{(R3)} \hat{P}_{lmn}^\sigma + \sum_{ \substack{ k,l,m,n \in \Box \\ k < l< n < m}} J_{k}^{(R4)} \hat{P}_{klmn}^\sigma + \sum_{ \substack{ j,k,l,m,n \in \pentagon \\ j < k <n < l<m}} J_{j}^{(R5)} \hat{P}_{jklmn}^\sigma,
\end{align}
where e.g.\ $\hat{P}_{klmn}^{\sigma}$ means a cycle of length 4 which permutes the spins according to $(\sigma_k, \sigma_l,\sigma_m,\sigma_n) \to (\sigma_l, \sigma_m,\sigma_n,\sigma_k)$. We use the notation $m \sim n$ to mean $m$ and $n$ are up to 2$^{nd}$-nearest neighbours, $l,m,n \in \triangle$ means that $l,m,n$ are on the vertices of a triangle, $k,l,m,n \in \Box$ means that $k,l,m,n$ are on the vertices of a parallelogram and $j,k,l,m,n \in \pentagon$ means that $j,k,l,m,n$ are on the vertices of trapezium (see Fig.~\ref{fig:WigExch}).
To convert this into a Hamiltonian, the spin permutation operators can be written in terms of pairwise spin exchange interactions, using the following identities~\cite{Klein1980}:
\begin{align}
&\hat{P}_{12}^\sigma = \frac{1}{2}(\idop + E_{12}), \\
&\hat{P}_{123}^\sigma + \hat{P}_{321}^\sigma = \hat{P}_{12}^\sigma + \hat{P}_{23}^\sigma + \hat{P}_{13}^\sigma - 1, \\
&\hat{P}_{1234}^\sigma + \hat{P}_{4321}^\sigma = \hat{P}_{12}^\sigma \hat{P}_{34}^\sigma + \hat{P}_{14}^\sigma \hat{P}_{23}^\sigma - \hat{P}_{13}^\sigma \hat{P}_{24}^\sigma + \hat{P}_{13}^\sigma + \hat{P}_{24}^\sigma -1, \\
&\hat{P}_{12345}^\sigma + \hat{P}_{54321}^\sigma = \frac{1}{2}(\hat{P}^\sigma_{12}\hat{P}^\sigma_{34} + \hat{P}^\sigma_{14}\hat{P}^\sigma_{23} - \hat{P}^\sigma_{13}\hat{P}^\sigma_{24})+ \frac{1}{2}(\hat{P}^\sigma_{15}\hat{P}^\sigma_{23} + \hat{P}^\sigma_{12}\hat{P}^\sigma_{35} - \hat{P}^\sigma_{13}\hat{P}^\sigma_{25} )\nonumber\\
&+\frac{1}{2}(\hat{P}^\sigma_{15}\hat{P}^\sigma_{24} + \hat{P}^\sigma_{12}\hat{P}^\sigma_{45} - \hat{P}^\sigma_{14}\hat{P}^\sigma_{25})+ \frac{1}{2}(\hat{P}^\sigma_{15}\hat{P}^\sigma_{34} + \hat{P}^\sigma_{13}\hat{P}^\sigma_{45} - \hat{P}^\sigma_{14}\hat{P}^\sigma_{35} ) +\frac{1}{2}(\hat{P}^\sigma_{25}\hat{P}^\sigma_{34} + \hat{P}^\sigma_{24}\hat{P}^\sigma_{35} - \hat{P}^\sigma_{25}\hat{P}^\sigma_{34}) \nonumber\\
&- \frac{1}{2}(\hat{P}^\sigma_{12}+\hat{P}^\sigma_{15}+\hat{P}^\sigma_{23}+\hat{P}^\sigma_{34}+\hat{P}^\sigma_{45}) +\frac{1}{2}(\hat{P}^\sigma_{13}+\hat{P}^\sigma_{35}+\hat{P}^\sigma_{25}+\hat{P}^\sigma_{24}+\hat{P}^\sigma_{14})- \frac{1}{2}.
\end{align}
where we have used $E_{mn} := X_m X_n + Y_m Y_n + Z_m Z_n$.
Inserting these into eqn.\ (\ref{eqn:PermHam}), this gives the overall Hamiltonian, ignoring the factors of $\idop$ since they do not affect the dynamics:
\begin{align}
H &= \frac{1}{2} \sum_{m \sim n }J_{mn}^{(2)} E_{mn} - \frac{1}{2}\sum_{ \substack{ l,m,n \in \triangle \\ l < m < n} } J_{l}^{(R3)} ( E_{lm} + E_{mn} + E_{ln}) + \frac{1}{4} \sum_{ \substack{ k,l,m,n \in \Box \\ k < l< n < m}} J_{k}^{(R4)} \left[ E_{kl} + E_{km} + E_{kn} + \right. \nonumber\\
&\left. E_{lm} + E_{ln} + E_{mn} + \Upsilon_{klmn} \right] -\frac{1}{8} \sum_{ \substack{ j,k,l,m,n \in \pentagon \\ j < k <n < l<m}} J_{j}^{(R5)} \left[ \Upsilon_{jklm} + \Upsilon_{jkln} + \Upsilon_{jnkm} + \nonumber\right.\\
&\left. \Upsilon_{jnlm} + \Upsilon_{knlm}
+ E_{jk}+E_{kl}+E_{lm}+E_{mn} +E_{jn}+E_{jl}+E_{km}+E_{ln}+E_{kn}+E_{jm} \right]
\end{align}
where $\Upsilon_{jklm} := E_{jk}E_{lm} + E_{jm}E_{kl} - E_{jl}E_{km}$. Note the minus sign in front of the $J^{(R3)}$ and $J^{(R5)}$ terms since they involve permutations of an odd number of particles, .
\begin{figure}[t]
\begin{center}
\subfloat[]{\includegraphics[width=0.6\textwidth]{EvolutionFid-eps-converted-to.pdf}}\\
\subfloat[]{\includegraphics[width=0.6\textwidth]{Evolution_topt-eps-converted-to.pdf}}
\caption{\label{fig:Evo} a) Maximum value of the average fidelity $F_{av}[\mathcal{E}]$ with different densities for $N=9$ and $N=10$. The red line indicates the average fidelity achievable with a classical channel. b) $t_{opt}$, where $t_{opt}$ is the time at which maximum fidelity is reached, in seconds. The red line indicates the typical $T_2$ time in GaAs.}
\end{center}
\end{figure}
To measure the information transferring ability of this channel, the average fidelity was calculated using the formula in (\ref{eqn:FavrgHorod}) for qubits:
\begin{align}
F_{av}[\mathcal{E}(t)] = \frac{2 F_e[\mathcal{E}(t)] +1 }{3},
\end{align}
where $\mathcal{E}(t)[\rho] = \text{tr}_{\hat{N}} \left( e^{-iHt / \hbar} \rho e^{iHt / \hbar} \right)$. The average fidelity was calculated up to a timescale such that one peak had arrived at the other end, so does not necessarily represent the best achievable fidelity. The time at which the peak reaches the highest value of $F_{av}[\mathcal{E}(t)]$ is denoted $t_{opt}$. Results for the maximum achieved fidelity $F_{av}[\mathcal{E}(t_{opt})]$ are shown in Fig.~\ref{fig:Evo}, showing remarkably high fidelities generally increasing as $\nu$ increases. $t_{opt}$ ranges from around 0.1ns to 10$\mu$s; the decoherence timescale of an electron spin in GaAs is roughly 10ns at present~\cite{Petta2005} so this indicates that $\nu \simeq 0.6$ is the optimal density at which to create a Wigner spin chain in GaAs/AlGaAs. Clearly there is little difference in fidelity in using even or odd electrons in our protocol.
\begin{figure}[h]
\begin{center}
\subfloat[]{\includegraphics[width=0.49\textwidth]{CouplingsN=9_J1-eps-converted-to.pdf}}
\subfloat[]{\includegraphics[width=0.49\textwidth]{CouplingsN=9_J2-eps-converted-to.pdf}}\\
\subfloat[]{\includegraphics[width=0.49\textwidth]{CouplingsN=9_J7-eps-converted-to.pdf}}
\subfloat[]{\includegraphics[width=0.49\textwidth]{CouplingsN=9_J8-eps-converted-to.pdf}}\\
\subfloat[]{\includegraphics[width=0.49\textwidth]{CouplingsN=9_J3-eps-converted-to.pdf}}
\subfloat[]{\includegraphics[width=0.49\textwidth]{CouplingsN=9_J4-eps-converted-to.pdf}}\\
\subfloat[]{\includegraphics[width=0.49\textwidth]{CouplingsN=9_J5-eps-converted-to.pdf}}
\subfloat[]{\includegraphics[width=0.49\textwidth]{CouplingsN=9_J6-eps-converted-to.pdf}}
\caption{\label{fig:CouplingsN=9} $J_P$ for a Wigner crystal with 9 electrons, with $h_L = h_R = 0.1$: a) nearest-neighbour, b) 2$^{nd}$ neighbour, c) 3$^{rd}$ neighbour d) 4$^{th}$ neighbour, e) 3-body f) 4-body, g) 5-body and h) 6-body. $J_P$ is in units of meV. Note that there is a lower limit on the colour bar, which means that very small couplings are rounded up to $10^{-16}$.}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\subfloat[]{\includegraphics[width=0.49\textwidth]{CouplingsN=10_J1-eps-converted-to.pdf}}
\subfloat[]{\includegraphics[width=0.49\textwidth]{CouplingsN=10_J2-eps-converted-to.pdf}}\\
\subfloat[]{\includegraphics[width=0.49\textwidth]{CouplingsN=10_J7-eps-converted-to.pdf}}
\subfloat[]{\includegraphics[width=0.49\textwidth]{CouplingsN=10_J8-eps-converted-to.pdf}}\\
\subfloat[]{\includegraphics[width=0.49\textwidth]{CouplingsN=10_J3-eps-converted-to.pdf}}
\subfloat[]{\includegraphics[width=0.49\textwidth]{CouplingsN=10_J4-eps-converted-to.pdf}}\\
\subfloat[]{\includegraphics[width=0.49\textwidth]{CouplingsN=10_J5-eps-converted-to.pdf}}
\subfloat[]{\includegraphics[width=0.49\textwidth]{CouplingsN=10_J6-eps-converted-to.pdf}}
\caption{\label{fig:CouplingsN=10} $J_P$ for a Wigner crystal with 10 electrons, with $h_L = h_R = 0.1$: a) nearest-neighbour, b) 2$^{nd}$ neighbour, c) 3$^{rd}$ neighbour d) 4$^{th}$ neighbour, e) 3-body f) 4-body, g) 5-body and h) 6-body. $J_P$ is in units of meV. Note that there is a lower limit on the colour bar, which means that very small couplings are rounded up to $10^{-16}$.}
\end{center}
\end{figure}
\subsection{Discussion and Conclusions}
We have calculated exchange couplings for a quasi-2D Wigner crystal using the semi-classical instanton approximation, and investigated the feasibility of using such a chain to transfer quantum information. Our results suggest that quantum information could be transmitted through a Wigner crystal with high fidelity (up to around 0.92 for chains of 9 or 10 electrons) using a quench, with transfer times that are below the typical decoherence times in GaAs, and using parameters typical of today's GaAs technology. This spin chain would be useful perhaps in conjunction with quantum dot qubits~\cite{Loss1998}, or could perhaps be extended to a full proposal of quantum computation, similar to the scheme in~\cite{Benjamin2003} with always-on interactions. Our scheme has many variables that can be altered without affecting the viability of the scheme, such as the barrier heights and the offset of the quantum dots. Magnetic fields are also known to alter the exchanges via the Aharonov-Bohm effect~\cite{Bernu2001,Zhu1995,Hirashima2001}. With some fine tuning of these parameters, we would expect even higher fidelities to be attainable. Using a quench may also not be the ideal protocol, and we would be interested to see whether other approaches such as adiabatically lowering the barriers is also viable.
Further work will need to be done to find a detailed experimental setup for realising this scheme, using either electrons or holes in GaAs/AlGaAs or other systems in which large $r_\Omega$ values can be achieved. The full effect of phonon excitations following the quench also needs to be taken into account to see how much error this leads to. There are also other effects that we have not taken into account, such as magnetic field fluctuations due to impurities, how the transfer is affected by a finite barrier dropping rather than an immediate quench, image charges produced by metal boundaries, dipole-dipole interactions and spin-orbit interactions that can produce anisotropic exchange terms in the Hamiltonian~\cite{Tserkovnyak2009}. It should also be noted that this scheme uses a semi-classical approximation, so is only accurate away from the Wigner crystal melting point ($r_\Omega \simeq 4 $ for 1D Wigner crystals).
The best results seem to occur when there is a combination of nearest-neighbour, next-nearest neighbour and 3-body ring exchange present, which is perhaps because the 3-body exchange acts as an effective ferromagnetic interaction that works against the nearest- and next-nearest neighbour interactions and so makes the coupling strength more uniform throughout the chain. It would be interesting to investigate the roles of each of the exchange processes in more depth, to see if there are any processes which hinder or help information transfer, in order to search for the optimum regime and understand the underlying physical mechanisms. It would also be interesting to take this analysis to higher densities, including higher-order exchange processes to see if higher densities also allow high fidelity transfer.
\chapter{Information transfer with idealised spin chains}
\label{chap:NNN}
In the previous chapter, we explored a potential way to create a spin chain, through which information can be transferred. In this final chapter, we present three smaller studies, each exploring different protocols for sending information through an idealised spin chain, to see how these perform and how they can be optimised. It is hoped that each of these will provide a basis for more in-depth study in the future.
\section{Information transfer with next-neighbour interactions}
\label{sec:NNN}
In this section we investigate the effect of next-neighbour interactions on transfer through a uniform spin chain. The aim of this investigation is to test the robustness of existing nearest neighbour protocols to the addition of next-nearest neighbour (NNN) coupling, which may naturally be present, and to give new insights into what kind of systems are best for information transmission. The work follows a similar approach to the work by Bayat et.\ al.~\cite{Bayat10}, where the entanglement transferral across a uniform spin chain with Hamiltonian
\begin{eqnarray}
H= \sum_{n=1}^{N-1} J (X_n X_{n+1} +Y_n Y_{n+1}+ \Delta Z_n Z_{n+1}).
\end{eqnarray}
was investigated, for different values of $\Delta$. The protocol consists of attaching one half of a singlet to the end of a spin chain in the ground state, and waiting for some time $t$ when the entanglement with the other end of the chain is maximised (see Section~\ref{sec:Prot} below). In this investigation they found that $\Delta = 1$ achieved the highest and fastest end-to-end entanglement
Following~\cite{Bayat10}, we investigate how next-nearest-neighbour (NNN) coupling in isotropic spin chains affects the quality of entanglement transferred using this protocol. The Hamiltonian we consider is the isotropic next-nearest-neighbour Hamiltonian (or `$J_1-J_2$' Hamiltonian):
\begin{eqnarray}
\label{eqn:HamNNN}
H_{\textsc{nnn}}^{(N)} = J_1\sum_{n=1}^{N-1} (X_n X_{n+1} +Y_n Y_{n+1}+ Z_n Z_{n+1}) +J_2 \sum_{n=1}^{N-1} (X_n X_{n+2} +Y_n Y_{n+2}+ Z_n Z_{n+2}),
\end{eqnarray}
where we have picked the isotropic model ($\Delta = 1$) since it already optimises the nearest-neighbour case, and the superscript $N$ indicates that this Hamiltonian acts on $N$ spins. Whilst the phase diagram of this $J_1-J_2$ model is qualitatively well known, it is still not exactly solvable except at certain points (see e.g.~\cite{SchollwockBook}). With $J_1,J_2 > 0$ the model is in the antiferromagnetic frustrated regime. Below a critical value of $J_2 < J_{2c} \approx 0.241J_1$, there is no energy gap between the ground state and excited states, and a gap opens up for $J_2 > J_{2c}$. At $J_2 = 0.5J_1$, the \emph{Majumdar-Ghosh} point, the ground state is analytically solvable~\cite{Majumdar69}. Here the ground state is completely dimerised, and is known as a \emph{dimer product state}:
\begin{eqnarray}
|\phi_{\textsc{mg}} \rangle= |\psi^- \rangle_{1,2} |\psi^- \rangle_{3,4} ... |\psi^- \rangle_{N-1,N},\: \mathrm{where} \quad |\psi^{-} \rangle = \frac{1}{\sqrt{2}}(|\uparrow \downarrow \rangle - |\downarrow \uparrow \rangle ).
\end{eqnarray}
Otherwise, as $J_2$ grows and becomes larger than $J_1$ we effectively end up with two decoupled chains, so in this study we mainly focus on the region $0 < J_2 < 0.6 J_1$. In what follows we will set $J_1 = 1$ unless otherwise stated.
In the case of perfect state transfer, some investigations have already been done into the effects of different types of errors~\cite{Ronke11}, including small amounts of NNN interactions. Other work has also found that NNN coupling is detrimental to state transfer, for a protocol that transfers information by weakly coupling qubits to the two ends of a spin chain in the Kondo regime~\cite{Sodano10}. The differences between these studies and the work in this chapter is that we consider a spin chain with uniform interactions, and for larger NNN interactions.
\subsection{Protocol}
\label{sec:Prot}
We consider Hamiltonians of the form shown in (\ref{eqn:HamNNN}), where $J_1<0$ for the ferromagnetic case and $J_1>0$ for the antiferromagnetic case. We consider open chains (i.e.\ open boundary conditions) with uniform couplings, to minimise engineering requirements and to keep the number of variable parameters low.
Firstly we diagonalise $H_{\textsc{nnn}}^{(N)}$ on $N$ spins to find the ground state $\ket{\psi_G}$ (if there is degeneracy in the ground state a small magnetic field term $h\sum_n \sigma_n^z$ can be applied). We then couple a singlet state $ \ket{ \psi^{-} }_{0'0}$ of two qubits $0'$ and $0$ to one of the ends, so that the initial state is:
\begin{equation}
\ket{\psi(t=0)}= \ket{ \psi^{-} }_{0'0} \ket{\psi_G }.
\end{equation}
to do this, we imagine a system in which there are barriers separating spins 0,0' and 1, so that the coupling can be selectively turned on or off. Spin $0'$ is completely uncoupled from the chain at all times, whilst spin $0$ interacts with the chain with the same couplings as in the chain (we have the choice of selecting weaker or stronger couplings to connect this singlet to the chain, however since faster transfer is likely to arise from stronger couplings, we choose the chain couplings to be the highest possible). Thus the Hamiltonian for the chain$+$singlet system is $H_{tot} = I_{0'} \otimes H_{\textsc{nnn}}^{(N+1)}$. From this we construct the evolution operator $U=e^{-iH_{tot}t}$, and use this to find the state after a time t, $|\psi(t) \rangle =U(t) |\psi(t=0) \rangle$.
Since we have started with the singlet, which has two qubits maximally entangled, we might expect that after allowing the system to evolve there would be some entanglement between the $0'^{th}$ spin and the $N^{th}$ spin. To quantify this, first of all we trace out all of the qubits in the system except for the $0'^{th}$ and $N^{th}$ spins, leaving a density matrix:
\begin{equation}
\rho_{0'N}(t)= tr_{\widehat{0'N}} [|\psi(t) \rangle \langle \psi(t) | ],
\end{equation}
which we can then use to find the entanglement of formation $E_f$ between qubits $0'$ and $N$, where the entanglement of formation is defined in Section~\ref{sec:IntroSpinChain}. We then look for the optimal time where there is maximal entanglement between the 0' and $N^{th}$ spin, and find how this maximum entanglement varies with NNN coupling.
To practically calculate the time-evolved state and the entanglement, we used exact diagonalisation in MATLAB, using a stepwise time evolution, where at each time step the evolution operator was applied to the state and then the entanglement at each time step could be calculated. Ideally we want to minimise the transmission time, as in a realistic system this would decrease the effects of decoherence. With this in mind, our results were limited to the first peak, which usually arrived on a timescale less than $N/2$, and we define $\max_1(E_f)$ as the entanglement of formation at the highest point of the first peak.
Exact diagonalisation works well for $N \leq 16$, but for more spins the computation time is very prohibitive, and so we use the technique of Density Matrix Renormalisation Group (DMRG) for larger systems (using the MATLAB code of Abolfazl Bayat, modified to include next-nearest neighbour coupling. See Appendix~\ref{sec:DMRG} for an overview of DMRG). Note that although $\sum_n Z_n$ commutes with Hamiltonians with the form that we use, so that evolution between states with different eigenvalues of $\sum_n Z_n$ is not allowed, this is not always incredibly useful here since the system is always initialised in the ground state, and so the major bottleneck is diagonalising the large matrices.
\subsection{Results}
\label{sec:J1J2Results}
\begin{figure}
\begin{center}
\subfloat[]{\includegraphics[width=0.6\textwidth]{Even_N-eps-converted-to.pdf} \label{fig:EvenNNN}} \\
\subfloat[]{\includegraphics[width=0.6\textwidth]{Odd_N-eps-converted-to.pdf} \label{fig:OddNNN}}
\end{center}
\caption{a) Maximum end-to-end entanglement $\max_1(E_f)$ vs. $J_2 / J_1$ for chains of even length $N=8$ to $N=16$ with a singlet added on. b) Maximum end-to-end entanglement $\max_1(E_f)$ vs. $J_2 / J_1$ for chains of odd length $N=9$ to $N=15$ with a singlet added on.}
\end{figure}
The results for maximum entanglement versus $J_2$ are shown in Figs.~\ref{fig:EvenNNN} \&~\ref{fig:OddNNN} for $N$ = 8 to 16. For even values of $N$, there is a slight decrease in entanglement up to $J_{2c}$, followed by a significant rise until $J_2 \approx 0.4$, and then a sharp decline near the Majumdar-Ghosh point. For odd values of $N$, a rise is seen up to $J_{2c}$, and the peak around $J_2$ =0.4 is absent.
\begin{figure}
\begin{center}
\subfloat[]{\includegraphics[trim=1.5cm 0cm 2.8cm 0cm, clip=true, width=0.5\textwidth]{Dimerisation-eps-converted-to.pdf} \label{fig:Dimers}}
\subfloat[]{\includegraphics[trim=1.5cm 0cm 2.8cm 0cm, clip=true, width=0.5\textwidth]{EnergyGap-eps-converted-to.pdf}\label{fig:EgapJ2}}
\end{center}
\caption{a)Dimerisation, defined as the overlap between the chain ground state (before addition of a singlet) and the fully dimerised state (the Majumdar-Ghosh ground state). b) Energy gap as $J_2$ varies - note that our system is not perfectly gapless as we are looking at finite chains, but we expect the chains to converge to a gapless system as the size increases.}
\end{figure}
We might expect that this gain in entanglement is to do with oscillations between degenerate states which have a single unpaired spin, i.e.\ we might expect there to be oscillations between degenerate states of the form:
\begin{eqnarray}
|\psi_1 \rangle=|\phi \rangle_1 |\psi_{chain} \rangle, \;|\psi_2 \rangle=|\psi_{chain} \rangle |\phi \rangle_{N},
\end{eqnarray}
where $| \phi \rangle$ is an unpaired single qubit state. We refer to this as a `resonant' mechanism for transferring entanglement. This would only be possible for even $N$ values, for which the actual evolving part of the chain is odd. This is very similar to a result found in \cite{Gualdi2008}, where they found that perfect state transfer is achieved when there were two eigenstates of the Hamiltonian that had maximal overlap with the two ends of the chain but zero overlap with other sites in the chain (and is quite an intuitive result). We can try to look at the plausibility of this explanation in quite a rough way by looking at how the dimerisation and the energy gap vary with $J_2$, where the dimerisation is defined as the fidelity between the ground state of the chain and the Majumdar-Ghosh ground state $\ket{\psi^-}^{\otimes \frac{N}{2}}$. If this was the dominant effect we would expect high dimerisation and a high energy gap to give us maximal entanglement. Results are shown in Figs.~\ref{fig:Dimers} $\&$~\ref{fig:EgapJ2}. Using this model we would expect to see a peak in entanglement transfer at around $J_2=0.5$, which is not the case in our simulations.
The difference between odd and even chains could perhaps be explained more or less using this idea; when $N$ is even, then after we have coupled the $0^{th}$ site to it our starting state has only one unpaired spin. Then as we go towards the Majumdar-Ghosh point where the ground state is dimerised, we would expect that the states with unpaired spins at either end might become more and more isolated, and so might oscillate only between themselves. Then for odd values of $N$, there would be more than two unpaired sites to begin with (since we add our $0^{th}$ spin to an odd chain) and so there may be more degeneracy and more states in which the evolution can disperse.
It is also in some ways analogous to the work done in \cite{Venuti06}. Here they looked at end-to-end entanglement of a chain with an explicitly dimerised Hamiltonian of the form
\begin{equation}
H = \sum\nolimits_{n=1}^{N} [1+(-1)^n \delta] (X_n X_{n+1} +Y_n Y_{n+1}+ Z_n Z_{n+1}),
\end{equation}
which creates alternating strong and weak bonds along the chain, and has an energy gap $\Delta E \propto \delta^{2/3}$. They found that the end-to-end entanglement of this chain increased past a critical value of $\delta$. Then the behaviour we see may be of a similar type. Unfortunately due to the alternating weak and strong bonds in this Hamiltonian it is hard to create the same scenario as we have here, i.e.\ the two states:
\begin{eqnarray} \label{eq:OscStates}
|\psi_1 \rangle=|\phi \rangle_1 \ |\psi^{-} \rangle |\psi^{-} \rangle ... |\psi^{-} \rangle, \; |\psi_2 \rangle=|\psi^{-} \rangle |\psi^{-} \rangle |\psi^{-} \rangle ... |\phi \rangle_{N}
\end{eqnarray}
are not degenerate, because by changing the positions of the singlets we have put them over bonds of different strength, and so we have different energies. In addition this work is based around having weak coupling at either end (the idea being that stronger coupling would mean stronger entanglement to sites within the chain, and therefore less entanglement with the site we actually want to be entangled to, due to the monogamy of entanglement).
Similar ideas to the explanation presented above are looked at by Hartmann et.\ al in \cite{Hartmann06}. Here they argue that the more eigenstates that are involved in the evolution, the more dispersive the information transmission. If we start with a state with energy expectation $\langle E \rangle$, then there is a variance in the energy of the state given by
\begin{equation}
\Delta E = \sqrt { \langle H^2 \rangle - \langle H \rangle}.
\end{equation}
So we can try to quantify the number of eigenstates involved in the evolution by counting the number of eigenstates, $N_S$, within the energy range $\langle E \rangle - \Delta E \leq E_S \leq \langle E \rangle + \Delta E$. This energy range will be constant during the state's evolution, and we can see if this explains the behaviour seen above. The results are shown in Fig.~\ref{fig:DoS8} and~\ref{fig:DoS12} for $N = 8$ and $N = 12$, respectively. Using this measure, there does appear to be some correlation, although not enough to justify using it as an explanation, and it appears to get worse as $N$ is increased.
Since the effects we see do not seem to match up with the model of resonance being the only effect here, we hypothesise that there is information propagation through the chain as well as these resonating effects outlined above, and optimal entanglement transfer is for values of $J_2$ which allow some propagation and some resonance, in such a way that the resonance effects constructively add to the propagation through the chain. Then past this optimal point, propagation through the chain becomes more difficult (perhaps due in part to the argument above involving number of accessible eigenstates), and the resonating effects then become the only method of transferring entanglement. Currently this is as far as we have progressed in finding an explanation, which is not an entirely satisfactory one. To investigate this further we plan to look in more depth at the eigenstates involved in the evolution, to look for evidence that the states involved are really facilitating end-to-end resonance (which would most likely manifest itself as states with large overlap at the ends but not much in the middle, in a similar manner to \cite{Gualdi2008}).
\begin{figure}
\subfloat[]{\includegraphics[trim=1.5cm 0cm 2.8cm 0cm, clip=true, width=0.475\textwidth]{DoS_Simple_8-eps-converted-to.pdf} \label{fig:DoS8}}
\qquad
\subfloat[]{\includegraphics[trim=1.5cm 0cm 2.8cm 0cm, clip=true, width=0.475\textwidth]{DoS_Simple_12-eps-converted-to.pdf}\label{fig:DoS12}}
\caption{Maximum end-to-end entanglement $\max_1(E_f)$, plotted alongside $N_{states}$ (the number of states within $\pm \Delta E $, rescaled to fit on this graph). a) $N = 8$, b) $N = 12$.}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.7\textwidth]{DMRG_all-eps-converted-to.pdf}
\caption{Maximum end-to-end entanglement $\max_1(E_f)$, for different initial states, for $N = 18$, 20 and 22, found using DMRG.}
\label{fig:DMRG}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\textwidth]{Full_all-eps-converted-to.pdf}
\caption{Maximum end-to-end entanglement $\max_1(E_f)$ for positive and negative $J_2$, for chains of length $N = 8,10,12$}
\label{All}
\end{center}
\end{figure}
So far, all of the results have been found using exact diagonalisation of the Hamiltonian, however we can extend to larger $N$ by using the Density Matrix Renormalisation Group (DMRG) method (see Appendix~\ref{sec:DMRG}). Using this method, we extended the exact diagonalisation results up to values of $N=$ 18 and 24. From this we still see the characteristic rise after the critical value $J_{2c}$. Past chain lengths of around $N \sim 26$ the results appeared to diverge (i.e.\ there were sudden dramatic changes in the behaviour around these lengths, particularly for larger $J_2$ values). We attributed this to inaccuracies in the DMRG rather than any change in physical behaviour, since larger chains increase the time we have to wait to receive a signal at the other end, and it is likely that the simulation takes longer than the timescale over which the DMRG is accurate. This would also explain why the results start to diverge for larger values of $J_2$, since larger values of $J_2$ increase the time taken for the signal to reach the end.
\subsection{Extending to negative $J_2$}
\label{sec:FMtransfer}
When we extend this study to negative values of $J_2$ (whilst keeping $J_1=1$) we also see an increase (Fig.~\ref{All}). In this region increasing $J_2$ makes the chain less frustrated. In general then, there seem to be no large drops in entanglement transfer for a range of $-1<J_2<0.4$. This is reassuring, and suggests that generating end-to-end entanglement via this method is quite robust to the introduction of next nearest neighbour coupling.
\section{Edge locking}
\label{sec:EdgeLock}
We now investigate a protocol, based on the work in~\cite{Haque2010}, to transfer information, using certain states that are effectively stationary and so are said to be `locked'. We consider a nearest-neighbour XXZ Hamiltonian:
\begin{align}
H_{\textsc{xxz}} = \sum_n X_n X_{n+1} +Y_n Y_{n+1} + \Delta Z_n Z_{n+1}.
\end{align}
Provided the anisotropy $\Delta$ is large enough, this Hamiltonian allows two types of locked states; the first type is where the leftmost $n$ spins are up whilst the remaining spins are down, such as the following states
\begin{align}
\ket{\uparrow\downarrow \downarrow \downarrow ... \downarrow },\; \ket{\uparrow \uparrow \downarrow \downarrow ... \downarrow } ,\;\ket{\uparrow \uparrow \uparrow \downarrow... \downarrow }.
\end{align}
These are locked since any movement of the spins introduces an extra $\downarrow \uparrow$ boundary (a `domain wall') which increases the energy (provided the $ZZ$ coupling is significantly larger than the $XY$ coupling). The second type of locked states are slightly less intuitive, having blocks of $\uparrow$ surrounded by $\downarrow$, which are sensitive to the number of spin-ups and where they are positioned. For example, the following states are edge-locked
\begin{align}
\ket{\downarrow \uparrow \uparrow \uparrow \downarrow... \downarrow } ,\; \ket{\downarrow \uparrow \uparrow \uparrow \uparrow\downarrow... \downarrow }
\end{align}
and in general, a state will be locked if it contains a block of $N_\uparrow$ spins, with the leftmost spin at position $k$ where $N_\uparrow \geq (2k-1)$ (and similarly for the right hand side of the chain). Following the notation in~\cite{Haque2010} we will label these states $\ket{L_{N_\uparrow,(k)}}$ ($\ket{R_{N_\uparrow,(k)}}$) for a block of $N_\uparrow$ spins with the leftmost (rightmost) spin at position $k$. The explanation in~\cite{Haque2010} for why these states are locked is that, provided $\Delta^{-1}$ is small but non-zero, performing degenerate perturbation theory on the Hamiltonian we find that terms in the Hamiltonian that connect states $\ket{L_{N_\uparrow,(k)}}$ to other states with $k'>k$ are of order $O(\Delta^{-N_\uparrow})$. Additionally one finds that the energy of the $\ket{L_{N_\uparrow,(k)}}$ states is modified by $O(\Delta^{-2(k-1)})$ relative to the other states with $k'>k$. Thus if $\Delta^{-N_\uparrow} > \Delta^{-2(k-1)}$ evolution of $\ket{L_{N_\uparrow,(k)}}$ to other states with $k' > k$ is dominant, and conversely if $\Delta^{-N_\uparrow} < \Delta^{-2(k-1)}$ the evolution is not allowed due to conservation of energy. So the states are approximately locked when $N_\uparrow \geq 2(k-1)$.
\subsection{Applications of edge locking}
This mechanism appears to be very useful as means to control quantum information in spin chains. Here we outline several possible applications, before exploring how robust the transfer is with more realistic conditions.
\subsubsection{Signal amplification}
Consider preparing an edge-locked state, e.g.\ $\ket{L_{3,(1)}}$, and attaching a spin $\ket{\phi}$ to the left edge:
\begin{equation}
\ket{\phi}\ket{\uparrow \uparrow \uparrow\downarrow ... \downarrow \downarrow}
\end{equation}
If $\phi = \uparrow$, then the chain is still locked. However if $\phi = \downarrow$, the chain becomes $\ket{L_{3,(2)}}$ and so is no longer locked, and the state is amplified by a ratio of roughly 2:1. Attaching a spin to the end like this is limited to ratios of 2:1, since if we tried this with larger number the resultant state would be edge-locked for both inputs. This can be extended to larger ratios if we can have spins to the left of our input spins, e.g.\ for a ratio of 6:1 we can start with:
\begin{equation}
\ket{\downarrow \downarrow} \ket{\phi} \ket{\uparrow \uparrow \uparrow \uparrow \uparrow \uparrow \downarrow ... \downarrow \downarrow},
\end{equation}
Such a protocol could also perhaps be used as a method to detect a spin current. Note that the higher the ratio of input to output, the smaller the energy gap, and the more sensitive the scheme will be to fluctuations in energy (e.g.\ there may be additional energy added to the system in the process of coupling the state $\ket{\phi}$ to the end of the chain).
\subsubsection{\textsc{nand} gates}
A classical \textsc{nand} gate ($\textsc{nand}(a,b) = 1 \oplus ab$) can be constructed using a chain with an edge-locked state at each end, e.g.:
\begin{equation}
\ket{\phi_1} \ket{\uparrow \downarrow \downarrow ... \downarrow \downarrow \uparrow} \ket{\phi},
\end{equation}
where $\phi_1$ and $\phi_2$ are our input states. Then for the cases where $\ket{\phi_1\phi_2} = \ket{\downarrow \downarrow},\ket{\uparrow \downarrow}$ and $ \ket{\downarrow \uparrow}$, we will measure a signal travelling through the middle of the chain. If we can measure the presence or absence of this signal, and choose $\ket{0} = \ket{\downarrow},\ket{1} = \ket{\uparrow}$, this performs a \textsc{nand} gate.
\subsubsection{Heralded entanglement}
With the same setup as for the \textsc{nand} gate, we can create heralded entanglement: Firstly, we would expect that the signal travelling through the chain when $\ket{\phi_1\phi_2} = \ket{\uparrow \downarrow}$ to be the same as the case where $\ket{\phi_1\phi_2} =\ket{\downarrow \uparrow}$, and we would expect both of these signals to be different to the case where $\ket{\phi_1\phi_2} = \ket{\downarrow \downarrow}$. If so, then if it is possible to find a measurement that differentiates between the two types of signal, we could distinguish between the $(\ket{\uparrow \downarrow}, \ket{\downarrow \uparrow})$ cases and the $\ket{\downarrow \downarrow}$ case. If the end states are then initialised as $ \ket{\phi_1} = \ket{\phi_2} = \ket{+}$, this creates a superposition of different signals, and if a measurement is made that distinguishes between the $(\ket{\uparrow \downarrow}, \ket{\downarrow \uparrow})$ cases and the $\ket{\downarrow \downarrow}$ case, we know we have projected the end qubits into the superposition $\frac{1}{\sqrt{2}}(\ket{\uparrow \downarrow}+ \ket{\downarrow \uparrow})$.
\subsubsection{Charge qubit to spin qubit transfer}
We could also use edge locking as a means to interface between charge and spin qubits. An example of a charge qubit is an electron in one of two quantum wells, with presence in the left well indicating $\ket{0}$ and the right well indicating $\ket{1}$. If the spin state of this electron is set to $\ket{\downarrow}$, then depending on if it is on the left or the right well, the spin chain will be unlocked or remain locked, and thus the state of the charge qubit can be transferred into spin degrees of freedom.
\subsection{Release under more realistic conditions}
We have just seen several examples of how to use this edge locking state, must of which involve coupling a state to one edge of the chain, and measuring at the middle or the ends. In a perfect world, these operations could be done instantaneously, but of course it is more realistic to have operations that take a finite amount of time. In this subsection we first of all explore the effect of having less instantaneous operations when releasing an edge locked state (for example when performing the signal amplification shown above). We then explore how well a signal can be captured at the opposite end of the chain in the next subsection, by using the reverse strategy. Our motivation is to explore how realistic the applications of the edge locking could be, as well as seeing if there is some optimisation that can be done.
We use an XXZ Hamiltonian with anisotropy $\Delta$:
\begin{eqnarray}\label{XXZ}
H_{\textsc{xxz}} = J \sum_{n=1}^{N-1} (X_n X_{n+1} + Y_n Y_{n+1}) + \Delta Z_n Z_{n+1}
\end{eqnarray}
where $X_n$, $Y_n$ and $Z_n$ are the Pauli matrices acting on site n. To make it easier to sweep the whole range of $0 < \Delta \rightarrow \infty$, we parameterise it as $\Delta = \tan \theta_{\Delta}$, so $H_{XXZ}$ becomes:
\begin{eqnarray}
H_{\textsc{xxz}} &=& J \sum_{n=1}^{N-1} \sin \theta_{\Delta}(X_n X_{n+1} + Y_n Y_{n+1}) + J\cos \theta_{\Delta} Z_n Z_{n+1}.
\end{eqnarray}
To perform the release, we start with the edge locked state:
\begin{equation}\label{eqn:Init}
\ket{\phi_0} \equiv \ket{\uparrow \uparrow \downarrow ... \downarrow \downarrow}.
\end{equation}
Ideally, we would be able to switch off all other interactions and perform a perfect $X$ rotation on the first spin, to give us the state $\ket{L_{1,(1)}} \equiv \ket{\downarrow \uparrow \downarrow ... \downarrow}$ which is no longer locked and so will travel along the chain. However, this may not be realistic, so we model the release as having a mixture of a rotation on the first site (magnetic field in the $X$ direction) along with $H_{XXZ}$, with the relative strengths of these two operators given by $\theta_r$, giving the following Hamiltonian:
\begin{eqnarray}
H_{r}(\theta_r) &=& J \cos \theta_r X_1 + J\sin \theta_r H_{\textsc{xxz}} \nonumber\\
&=& J \cos \theta_r X_1 + J \sin \theta_r \left[ \sum_{n=1}^{N-1} \sin \theta_{\Delta}(X_n X_{n+1} + Y_n Y_{n+1}) + \cos \theta_{\Delta} Z_n Z_{n+1} \right]
\end{eqnarray}
This means we have some ratio $\tan \theta_r$ between the applied magnetic field and the couplings in the chain, with $\theta_r=0$ corresponding to a pure magnetic field (the ideal case) and $\theta_r = \pi/2$ corresponding to no magnetic field (no release). To perform the flip we would then evolve the state in eqn.\ (\ref{eqn:Init}) via:
\begin{eqnarray}
\ket{\phi_f} &=& e^{-iH_{r} t_{r} } \ket{\phi_0} = e^{-iH_{r} \pi / 2 } \ket{\phi_0}
\end{eqnarray}
where we have set $t_{r} = \pi / (2J\cos \theta_r)$, since the maximum overlap of $\ket{\phi_f}$ with the ideal state $\ket{L_{1,(2)}}$ occurs at or very close to this time. Unless otherwise noted, we will express time in units of $\hbar/J$.
We then evolve the state with $H_{\textsc{xxz}}$, so that the time-evolved state is:
\begin{eqnarray}
\ket{\phi(t)} &=& e^{-iH_{\textsc{xxz}} t }\ket{\phi_f} = e^{-iH_{\textsc{xxz}} t } e^{-iH_{r} \pi / 2 } \ket{\phi_0}
\end{eqnarray}
Following this evolution, we calculate the fidelity of $\ket{\phi(t)}$ with state $\ket{R_{1,(2)}} = \ket{\downarrow \downarrow...\downarrow \uparrow \downarrow}$, giving us the `release fidelity', $F_r$:
\begin{equation}
F_{r}(t) = |\sprod{{R_{1,(N-1)}}}{\phi(t)}|^2.
\end{equation}
To begin with, we looked at evolving the state until $F_r$ reaches its first peak, since in an application it would probably be best to wait for the shortest time possible, to limit the effects of decoherence. We thus define $t_{max}$ as the time taken for $F_r$ to reach the first peak. Firstly we looked at the variation of the size of $F_r(t_{max})$ with $\theta_{\Delta}$ and $\theta_{r}$ (see Fig.~\ref{Fid1}); evidently the best signal is for $\theta_{\Delta}$ and $\theta_r$ close to 0, as we might expect. Notice the rapid increase in time taken to reach the first maximum (see Fig.~\ref{tmaxDelta}), which limits how close to zero we can get as the simulations take too long (the minimum value of $\theta_{\Delta}$ in Figs.~\ref{Fid1} and \ref{tmaxDelta} is 0.01 radians, which means the $Z$ terms in $H_{XXZ}$ are roughly 100 times stronger than the $X$ or $Y$ terms). This suggests that there will be an optimal value of $\theta_{\Delta}$, depending on the particular experimental parameters such as decoherence time. There are also some strange discontinuities in Fig.~\ref{Fid1} for values of $\theta_{\Delta} > 0.4$, which we are currently unable to explain.
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=0.7\textwidth]{Delta_VS_Couplings_Long-eps-converted-to.pdf}
\caption{Contour plot showing the maximum of $F_r$ as a function of $\theta_{\Delta}$ and $\theta_1$, with $N=6$, and using exact evolution. Note that the minimum value of $\theta_\Delta$ is 0.01}
\label{Fid1}
\end{center}
\end{figure}
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=0.7\textwidth]{tmax_Delta-eps-converted-to.pdf}
\caption{Time taken to reach the first peak in $F_r$ after the release (in units of $\hbar/J$), for $\theta_r=0$. A similar variation is seen for all other values of $\theta_r$}
\label{tmaxDelta}
\end{center}
\end{figure}
Next we look at how $F_r$ varies with the length of the chain $N$, and $\theta_{\Delta}$. This was done similarly to above by finding the height of the first peak, which shows a steady decrease (see Fig.~\ref{VaryingN}) that appears to decay as $\sim 1/N$ for all $\theta_\Delta$, as we would intuitively expect.
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=0.7\textwidth]{VaryingN_thetaDelta-eps-converted-to.pdf}
\caption{Change in $F_r$ with $N$, with $\theta_r = 0$, where $F_r$ is taken at the first peak. The different coloured lines are for different values of $\theta_{\Delta}$ (in radians)}
\label{VaryingN}
\end{center}
\end{figure}
\subsection{Capture under more realistic conditions}
We now explore flipping the $N^{th}$ spin at some point in the evolution, with the hope of catching the pulse and keeping it locked in place. In a similar manner to the release protocol above, we capture the state by evolving with the Hamiltonian:
\begin{eqnarray}
H_{c}(\theta_c) = J\cos \theta_c X_N + \sin \theta_c H_{\textsc{xxz}}.
\end{eqnarray}
Then, after the release protocol above has been followed and the state has been evolved by $H_{\textsc{xxz}}$ for a time $\tau$, we switch on $H_{c}$ for a time $t_c$ and find the fidelity between this state and the ideal captured state, $\ket{R_{2,(1)}} = \ket{\downarrow \downarrow ... \downarrow \uparrow \uparrow}$. We define this as the capture fidelity, $F_c$:
\begin{equation}
F_c = \bra{R_{2,(1)}} e^{-i H_c } \ket{\phi(t)}, \; t > \tau .
\end{equation}
For smaller values of $\theta_c$, $t_c$ could be set to $\pi/2$ and this would contain the maximum value of $F_c$, however for larger $\theta_c$ the evolution was more chaotic and so longer times had to be used.
An example of release and capture is shown in Fig.~\ref{Capture1}, for which the capturing magnetic field is turned on at $t_{on} =t_{max}$ (the point where $F_r$ reaches a maximum) and then is turned off again when $F_c$ reaches a maximum. We see that capture is possible, with $F_c$ bounded by the maximum value of $F_r$, as we would expect. We also tried varying the timing of this magnetic field; results are shown in Fig.~\ref{Offset}, showing the variation $F_c$ as $t_{on}$, is varied relative to $t_{max}$, and also the variation with $\theta_c$. Clearly the best results happen where $t_{on} = t_{max}$, and with $\theta_c = 0$.
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=0.7\textwidth]{Capture1-eps-converted-to.pdf}
\caption{An example of the capturing protocol, with $N=6$, $\theta_{\Delta} = 0.1$, $\theta_{r} = 0.1$, $\theta_{c} = 0.1$.}
\label{Capture1}
\end{center}
\end{figure}
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=0.7\textwidth]{Offset_VS_Couplings_N=6_HighQ-eps-converted-to.pdf}
\caption{The maximum achievable capture fidelity $\max (F_c)$ as a function of $t_{on} - t_{max}$ and $\theta_c$, with $\theta_{\Delta} = 0.1$.}
\label{Offset}
\end{center}
\end{figure}
\section{Ground state transfer}
\label{sec:MagFieldTransfer}
In this section we outline a potential information transferring protocol, based on the work in~\cite{Khajetoorians2012}. where classical information was shown to be transferred through an Ising chain. The experiment uses an even or odd chain of Fe atoms adsorbed to a copper surface, and through changing the local magnetic field experienced by one end of the chain and then preparing the system in the ground state, the spin of the atoms at opposite ends of the chain are aligned (anti-aligned) for odd (even) chains.
The studies in~\cite{Khajetoorians2012} are performed with Ising interactions. Here we investigate the transferring ability of chains with different Hamiltonians (but not necessarily those which are achievable using chains of Fe atoms). We will use a Hamiltonian of the form
\begin{align}
H_{\pm} &= \pm B_1 Z_1 +J_1\sum_{n=1}^{N-1} (\sin \theta_\Delta (X_n X_{n+1} + Y_n Y_{n+1} )+ \cos\theta_\Delta Z_n Z_{n+1} ) \nonumber \\
&+ J_2 \sum_{n=1}^{N-2} ( \sin \theta_\Delta( X_n X_{n+2} + Y_n Y_{n+2} )+ \cos\theta_\Delta Z_n Z_{n+2} )
\end{align}
To perform the protocol, we select either $H_+$ or $H_-$, and prepare the system in the ground state, then trace out all but the leftmost $n_{out}$ qubits, which acts as the output. These $n_{out}$ output qubits will in general be in a mixed state $\rho_N^{\pm}$. If the probability distribution of the inputs is such that $H_+,H_-$ are chosen with probabilities $p_+,p_-$, the state of the output qubits is $\sum_{x=\pm} p_x \rho_N^{x}$. To assess the ability of this chain to send information, we require a measure of how distinguishable the output qubits are for the ground states $H_+$ and $H_-$. An appropriate way to measure this is to use the Holevo information $\chi$, which for a system prepared in one of the states $\{ \rho_1,\rho_2,...,\rho_n\}$ with probabilities $\{p_1,p_2,...,p_n\}$, is
\begin{align}
\chi_{H} : = S(\rho) - \sum_{x} p_x S(\rho_x)
\end{align}
where $\rho = \sum_x p_x \rho_x$. From Holevo's theorem, this is an upper bound on the amount of information that can be extracted from the output qubits using any possible measurement~\cite{NielsenChuang}.
\begin{figure}[h]
\begin{center}
\subfloat[]{\includegraphics[width=0.49\textwidth]{Holevo1N=8-eps-converted-to.pdf}}
\subfloat[]{\includegraphics[width=0.49\textwidth]{Holevo2N=8-eps-converted-to.pdf}}
\caption{\label{fig:MagFieldTrans_Even}. Information transferring ability using the ground state of a chain with $N=8$ qubits. a) Reading out the end qubit ($n_{out} = 1$) b) reading out from the two end qubits ($n_{out} = 2$).}
\end{center}
\end{figure}
We parameterise $J_1 = \sin\theta_M \cos \theta_J$, $J_2 = \sin\theta_M \sin \theta_J$, $B = \cos \theta_M$, so that the full range of $J_1 / J_2$ and $J_1 / B$ can be covered ($\theta_M = 0$ corresponds to a large magnetic field and weak coupling, whilst $\theta_J=0$ corresponds to $J_1 = 1,J_2 = 0$ and $\theta_J = \frac{\pi}{2}$ corresponds to $J_1 = 0,J_2 = 1$. The Holevo information was then found over a range of parameters by finding the ground states of $H_{\pm}$. Results are shown in Fig.~\ref{fig:MagFieldTrans_Even} a) for a chain of length $N=8$ qubits, with $\theta_M = 0.1$ and for $\theta_\Delta \lesssim \pi /4$ (for larger $\theta_\Delta$, $\chi_{H}$ continued to decay to even smaller values). Fig.~\ref{fig:MagFieldTrans_Even}b) shows the same set-up but reading out the last $n_{out} = 2$ qubits.
The main features are high transferring ability for small $\theta_J$ or $\theta_J \simeq \frac{3 \pi}{8}$, which information transferral decreasing rapidly with increasing $\theta_\Delta$. There is also very poor information transferral around $\theta_J = 0.5$, which is most likely to be a result of the Majumdar-Ghosh point that occurs at $\theta_J = \pi/4, \theta_\Delta = \pi/4$, at which point the ground state is fully dimerised and so there is no long range order. Information is also poorly transferred for $\theta_J \simeq \frac{\pi}{2}$, which is likely to be because at this point there is pure next-neighbour coupling, which for an even chain means no correlation between inputs and outputs.
Expanding the output to include more than one qubit predictably leads to better transfer of information. In future work, we would like to understand the theoretical basis for this, and perhaps derive a formula for the expected Holevo information as a function of the number of end qubits used.
\section{Conclusions}
At the end of this section, we are left with more questions than answers; we have seen that state transfer protocols through anisotropic spin chains can be robust to next-neighbour interactions, and there is a peak near to the Majumdar-Ghosh point where nearest-neighbour and next-nearest-neighbour interactions are equal, although the mechanism by which this peak occurs is not clear.
We have also seen protocols for sending information using edge-locking in XXZ Hamiltonians, for which there do not seem to be any other parameter regimes except the ideal cases (where the spins are flipped instantaneously, and the Hamiltonian is very close to an Ising Hamiltonian) for which edge locking could be used transfer states at high fidelity. Given the long time for the evolution close to $\Delta^{-1} = 0$, the value of $\Delta$ will be limited by the decoherence rates in the system of interest, and future work could involve finding the optimum regime for a particular experimental set up. This study is limited in that we only use square pulses: more sophisticated pulses could also be tried, with which it is likely that near perfect transfer could be achieved with realistic parameters. Non-uniform spin chains may also give an advantage.
Finally we investigate and the natural alignment of the ground state of a XXZ with a boundary magnetic field, and the use of this to transfer information. The information transfer is somewhat robust to the presence of XX and YY interactions, and next-nearest neighbour interactions. Better fidelity can be achieved by measuring more of the end qubits. There are many avenues that these preliminary results could eventually go down, and our hope is that the small results presented here will lay the foundations for future work.
|
\section{Introduction}
\label{sec:intro}
We study imitation learning, i.e.,~the problem of learning to perform a task from the sample trajectories generated by an expert. There are three main approaches to this problem: {\bf 1)} behavioral cloning (e.g.,~\citealt{Pomerleau91ET}) in which the agent learns a policy by solving a supervised learning problem over the state-action pairs of the expert's trajectories, {\bf 2)} inverse reinforcement learning (IRL)~\citep{Ng00AI} followed by reinforcement learning (RL), a process also referred to as RL$\circ$IRL~\citep{GAIL}, where we first find a cost function under which the expert is optimal (IRL part) and then return the optimal policy w.r.t.~this cost function (RL part), and {\bf 3)} generative adversarial imitation learning (GAIL)~\citep{GAIL} that frames the imitation learning problem as occupancy measure matching w.r.t.~either the Jensen-Shannon divergence (GAIL)~\citep{GAIL} or the Wasserstein distance (InfoGAIL)~\citep{INFOGAIL}. Behavioral cloning algorithms are simple but often need a large amount of data to be successful. IRL does not suffer from the main problems of behavioral cloning~\citep{RossB10,RossGB11}, since it takes entire trajectories into account (instead of single time-step decisions) when learning a cost function. However, IRL algorithms are often expensive to run because they require solving a RL problem in their inner loop. This issue had restricted the use of IRL to small problems for a long while and only recently scaleable IRL algorithms have been developed~\citep{Levine12CI,pmlr-v48-finn16}. On the other hand, the nice feature of the GAIL approach to imitation learning is that it bypasses the intermediate IRL step and directly learns a policy from data, as if it were obtained by the RL$\circ$IRL process. The resulting algorithm is closely related to generative adversarial networks (GAN)~\citep{GAN} that has recently gained attention in the deep learning community.
In many applications, we may prefer to optimize some measure of risk in addition to the standard optimization criterion, i.e.,~the expected sum of (discounted) costs. In such cases, we would like to use a criterion that incorporates a penalty for the variability (due to the stochastic nature of the system) induced by a given policy. Several risk-sensitive criteria have been studied in the literature of risk-sensitive Markov decision processes (MDPs)~\citep{Howard72RS} including the expected exponential utility~\citep{Howard72RS,Borkar01SF,Borkar02QR}, a variance-related measure~\citep{Sobel82VD,filar1989variance,tamar2012policy,Prashanth13AC}, or the tail-related measures like value-at-risk (VaR) and conditional value-at-risk (CVaR)~\citep{Filar95PP,Rockafellar00OC,saferl:cvar:chow2014,Tamar15OC}.
In risk-sensitive imitation learning, the agent's goal is to perform at least as well as the expert in terms of one or more risk-sensitive objective(s), e.g.,~$\text{mean}+\lambda\text{CVaR}_\alpha$, for one or more values of $\lambda\geq 0$. This goal cannot be satisfied by risk-neutral imitation learning. As we will show in Section~\ref{subsec:W-RS-GAIL}, if we use GAIL to minimize the Wasserstein distance between the occupancy measures of the agent and the expert, the distance between their CVaRs could be still large.~\citet{RAIL} recently showed empirically that the policy learned by GAIL does not have the desirable tail properties, such as VaR and CVaR, and proposed a modification of GAIL, called risk-averse imitation learning (RAIL), to address this issue. We will discuss about RAIL in more details in Section~\ref{sec:RAIL} as it is probably the closest work to us in the literature. Another related work is by~\citet{Singh18RI} on risk-sensitive IRL in which the proposed algorithm infers not only the expert's cost function but her underlying risk measure, for a rich class of static and dynamic risk measures (coherent risk measures). The agent then learns a policy by optimizing the inferred risk-sensitive objective.
In this paper, we study an imitation learning setting in which the agent's goal is to learn a policy with minimum expected sum of (discounted) costs and with $\text{CVaR}_\alpha$ that is at least as well as that of the expert. We first provide a mathematical formulation for this setting and derive a GAIL-like optimization problem for our formulation, which we call it risk-sensitive GAIL (RS-GAIL), in Section~\ref{subsec:PF}. In Sections~\ref{subsec:JS-RS-GAIL} and~\ref{subsec:W-RS-GAIL}, we define cost function regularizers that when we compute their convex conjugates and plug them into our RS-GAIL objective function, the resulting optimization problems aim at learning the expert's policy by matching occupancy measures w.r.t.~Jensen-Shannon (JS) divergence and Wasserstein distance, respectively. We call the resulting optimization problems JS-RS-GAIL and W-RS-GAIL and propose our risk-sensitive generative adversarial imitation learning algorithm based on these optimization problems in Section~\ref{sec:algo}. It is important to note that unlike the risk-neutral case in which the occupancy measure of the agent is matched with that of the expert, here in the risk-sensitive case, we match two sets of occupancy measures that encode the risk profile of the agent and the expert. This will become more clear in Section~\ref{sec:RSIL}. We present our understanding of RAIL and how it is related to our work in Section~\ref{sec:RAIL}. In Section~\ref{sec:experiments}, we evaluate the performance of our JS-RS-GAIL algorithm and compare it with GAIL and RAIL in two MuJoCo~\citep{Todorov12MuJoCo} tasks that have also been used by~\citet{GAIL} and~\citet{RAIL}. Finally in Section~\ref{sec:conclu}, we conclude the paper and list a number of future directions.
\section{Preliminaries}
\label{sec:Prelim}
We consider the scenario in which the agent's interaction with the environment is modeled as a Markov decision process (MDP). A MDP is a tuple $\mathcal{M} = \{\mathcal{S}, \mathcal{A}, c, p, p_0, \gamma\}$, where $\mathcal{S}$ and $\mathcal{A}$ are state and action spaces; $c: \mathcal{S} \times \mathcal{A} \to \mathbb{R}$ and $p:\mathcal{S}\times\mathcal{A}\rightarrow\Delta_{\mathcal{S}}$ are the cost function and transition probability distribution, with $c(s,a)$ and $p(\cdot|s,a)$ being the cost and next state probability of taking action $a$ in state $s$; $p_0:\mathcal{S}\rightarrow\Delta_\mathcal{S}$ is the initial state distribution; and $\gamma\in[0,1)$ is a discounting factor.
A stationary stochastic policy $\pi:\mathcal{S}\rightarrow\Delta_{\mathcal{A}}$ is a mapping from states to a distribution over actions. We denote by $\Pi$ the set of all such policies.
We denote by $\tau=(s_0,a_0,s_1,a_1,\ldots,s_T)\in\Gamma^\pi$, where $a_t\sim\pi(\cdot|s_t),\;\forall t\in\{0,\ldots,T-1\}$, a trajectory of the fixed horizon $T$ generated by policy $\pi$, by $\Gamma$ the set of all such trajectories, and by $C(\tau)=\sum_{t=0}^{T-1}\gamma^tc(s_t,a_t)$ the {\em loss} of trajectory $\tau$. The probability of trajectory $\tau$ is given by $\mathbb{P}(\tau|\pi)=p^\pi(\tau)=p_0(s_0)\prod_{t=0}^{T-1}\pi(a_t|s_t)p(s_{t+1}|s_t,a_t)$. We denote by $C^\pi$ the random variable of the loss of policy $\pi$. Thus, when $\tau\sim p^\pi$, $C(\tau)$ is an instantiation of the random variable $C^\pi$. The performance of a policy $\pi$ is usually measured by a quantity related to the loss of the trajectories it generates, the most common would be its expectation, i.e.,~$\mathbb{E}[C^\pi]=\mathbb{E}_{\tau\sim p^\pi}[C(\tau)]$. We define the occupancy measure of policy $\pi$ as $d^\pi(s,a) = \sum_{t=0}^{T} \gamma^t \mathbb{P}(s_t = s, a_t = a | \pi)$, which can be interpreted as the unnormalized distribution of the state-action pairs visited by the agent under policy $\pi$. Using occupancy measure, we may write the policy's performance as $\mathbb{E}[C^\pi]=\mathbb{E}_{p^\pi}[C(\tau)]=\mathbb{E}_{d^\pi}[c(s,a)]=\sum_{s,a}d^\pi(s,a)c(s,a)$.
\subsection{Risk-sensitive MDPs}
\label{subsec:RS-MDP}
In risk-sensitive decision-making, in addition to optimizing the expectation of the loss, it is also important to control the variability of this random variable. This variability is often measured by the variance or tail-related quantities such as value-at-risk (VaR) and conditional value-at-risk (CVaR). Given a policy $\pi$ and a confidence level $\alpha \in (0,1]$, we define the VaR at level $\alpha$ of the loss random variable $C^\pi$ as its (left-side) $(1-\alpha)$-quantile, i.e.,~$\nu_\alpha[C^\pi] := \inf\{t\in\mathbb{R}\;|\;\mathbb{P}(C^\pi\le t) \ge 1-\alpha \}$ and its CVaR at level $\alpha$ as $\rho_\alpha[C^\pi]=\inf_{\nu\in\mathbb{R}}\big\{\nu+\frac{1}{\alpha} \mathbb{E}\big[(C^\pi - \nu)_+\big]\big\}$, where $x_+ = \max(x,0)$. We also define the {\em risk envelope} $\mathcal{U}^\pi = \big\{\zeta : \Gamma^\pi \to [0,\frac{1}{\alpha}]\;|\;\sum_{\tau \in \Gamma}\zeta(\tau)\cdot p^\pi(\tau)= 1\big\}$, which is a compact, convex, and bounded set. The quantities $p^\pi_\zeta=\zeta\cdot p^\pi,\;\zeta\in\mathcal{U}^\pi$ are called {\em distorted probability distributions} and we denote by $\mathcal{P}^\pi_\zeta=\big\{p^\pi_\zeta\;|\;\zeta\in\mathcal{U}^\pi\big\}$ the set of such distributions. The set $\mathcal{P}^\pi_\zeta$ induces a set of {\em distorted occupancy measures} $\mathcal{D}^\pi_\zeta$, where each element of $\mathcal{D}^\pi_\zeta$ is the occupancy measure induced by a distorted probability distribution in $\mathcal{P}^\pi_\zeta$. The sets $\mathcal{P}^\pi_\zeta$ and $\mathcal{D}^\pi_\zeta$ characterize the risk of policy $\pi$. Given the risk envelope $\mathcal{U}^\pi$, we may define the dual representation of CVaR as $\rho_\alpha[C^\pi] = \sup_{\zeta\in\mathcal{U}^\pi}\mathbb{E}_{\tau\sim p^\pi}\big[\zeta(\tau)C(\tau)\big]$, where the supremum is attained at the density $\zeta^*(\tau) = \frac{1}{\alpha} \mathbf{1}_{\{C(\tau) \ge \nu_\alpha[C^\pi]\}}$. Hence, CVaR can be considered as the expectation of the loss random variable, when the trajectories are generated from the distorted distribution $p^\pi_{\zeta^*}=\zeta^*\cdot p^\pi$, i.e.,~$\rho_\alpha[C^\pi]=\mathbb{E}_{\tau\sim p^\pi_{\zeta^*}}[C(\tau)]$. If we denote by $d^\pi_{\zeta^*}\in\mathcal{D}^\pi_\zeta$ the distorted occupancy measure induced by $p^\pi_{\zeta^*}$, then we may write the CVaR as $\rho_\alpha[C^\pi]=\mathbb{E}_{p^\pi_{\zeta^*}}[C(\tau)]=\mathbb{E}_{d^\pi_{\zeta^*}}[c(s,a)]$.
\subsection{Generative Adversarial Imitation Learning}
\label{subsec:GAIL}
As discussed in Section~\ref{sec:intro}, generative adversarial imitation learning (GAIL)~\citep{GAIL} is a framework for directly extracting a policy from the trajectories generated by an expert policy $\pi_E$, as if it were obtained by inverse RL (IRL) followed by RL, i.e.,~RL$\circ$IRL$(\pi_E)$. The main idea behind GAIL is to formulate imitation learning as occupancy measure matching w.r.t.~the Jensen-Shannon divergence $D_{\text{JS}}$, i.e.,
\begin{equation*}
\min_\pi \big(D_{\text{JS}}(d^\pi,d^{\pi_E})-\lambda H(\pi)\big),
\end{equation*}
where $H(\pi)=\mathbb{E}_{(s,a)\sim d^\pi}[-\log\pi(a|s)]$ is the $\gamma$-discounted causal entropy of policy $\pi$, $\lambda\geq 0$ is a regularization parameter, and $D_{\text{JS}}(d^{\pi},d^{\pi_E}) := \sup_{f:\mathcal{S}\times\mathcal{A}\to (0,1)} \mathbb{E}_{d^\pi}[\log f(s,a)] + \mathbb{E}_{d^{\pi_E}}[\log(1-f(s,a))]$.~\citet{INFOGAIL} proposed InfoGAIL by reformulating GAIL and replacing the Jensen-Shannon divergence $D_{\text{JS}}(d^\pi,d^{\pi_E})$ with the Wasserstein distance $W(d^\pi,d^{\pi_E}) := \sup_{f \in \mathcal{F}_1} \mathbb{E}_{d^\pi}[f(s,a)] - \mathbb{E}_{d^{\pi_E}}[f(s,a)]$, where $\mathcal{F}_1$ is the set of $1$-Lipschitz functions over $\mathcal{S} \times \mathcal{A}$.
\section{Risk-sensitive Imitation Learning}
\label{sec:RSIL}
In this section, we describe the risk-sensitive imitation learning formulation studied in the paper and derive the optimization problems that our proposed algorithms solve to obtain a risk-sensitive policy from the expert's trajectories.
\subsection{Problem Formulation}
\label{subsec:PF}
As described in Section~\ref{sec:intro}, we consider the risk-sensitive imitation learning setting in which the agent's goal is to learn a policy with minimum loss and with CVaR that is at least as well as that of the expert. Thus, the agent solves the optimization problem
\begin{equation}
\label{eq:agent-objective-1}
\min_\pi \; \mathbb{E}[C^\pi] \qquad , \qquad \text{s.t.}\;\; \rho_\alpha[C^\pi] \le \rho_\alpha[C^{\pi_E}],
\end{equation}
where $C^\pi$ is the loss of policy $\pi$ w.r.t.~the expert's cost function $c$ that is unknown to the agent. The optimization problem~\eqref{eq:agent-objective-1} without the loss of optimality is equivalent to the unconstrained problem
\begin{equation}
\label{eq:agent-objective-2}
\min_\pi\;\sup_{\lambda \ge 0}\;\mathbb{E}[C^\pi] - \mathbb{E}[C^{\pi_E}] + \lambda\big(\rho_\alpha[C^\pi] - \rho_\alpha[C^{\pi_E}]\big).
\end{equation}
Note that $\pi_E$ is a solution of both~\eqref{eq:agent-objective-1} and~\eqref{eq:agent-objective-2}. However, since the expert's cost function is unknown, the agent cannot directly solve~\eqref{eq:agent-objective-2}, and thus, considers the surrogate problem
\begin{equation}
\label{eq:agent-objective-3}
\min_\pi\;\sup_{f\in\mathcal{C}}\;\sup_{\lambda \ge 0}\; \mathbb{E}[C^\pi_f] - \mathbb{E}[C^{\pi_E}_f] + \lambda \big(\rho_\alpha[C^\pi_f] - \rho_\alpha[C^{\pi_E}_f]\big),
\end{equation}
where $\mathcal{C}=\{f:\mathcal{S}\times\mathcal{A}\to\mathbb{R}\}$ and $C^\pi_f$ is the loss of policy $\pi$ w.r.t.~the cost function $f$. We employ the Lagrangian relaxation procedure~\citep{Bertsekas99NP} to swap the inner maximization over $\lambda$ with the minimization over $\pi$ and convert~\eqref{eq:agent-objective-3} to the problem
\begin{equation}
\label{eq:agent-objective-4}
\sup_{\lambda \ge 0}\;\min_\pi\;\sup_{f\in\mathcal{C}}\; \mathbb{E}[C^\pi_f] - \mathbb{E}[C^{\pi_E}_f] + \lambda \big(\rho_\alpha[C^\pi_f] - \rho_\alpha[C^{\pi_E}_f]\big).
\end{equation}
We adopt maximum causal entropy IRL formulation~\citep{Ziebart08ME,Ziebart10MI} and add $-H(\pi)$ to the optimization problem~\eqref{eq:agent-objective-4}. Moreover, since $\mathcal{C}$ is large, to avoid overfitting when we are provided with a finite set of expert's trajectories, we add the negative of a convex regularizer $\psi:\mathcal{C}\to\mathbb{R}\cup\{\infty\}$ to the optimization problem~\eqref{eq:agent-objective-4}. As a result we obtain the following optimization problem for our risk-sensitive imitation learning setting, which we call it RS-GAIL:
\begin{equation}
\label{eq:agent-objective-5}
\textbf{(RS-GAIL)} \qquad\qquad \sup_{\lambda \ge 0}\;\min_\pi\;-H(\pi) + \mathcal{L}_\lambda(\pi,\pi_E),
\end{equation}
where $\mathcal{L}_\lambda(\pi,\pi_E):=\sup_{f\in\mathcal{C}}\;(1+\lambda)\big(\rho_\alpha^\lambda[C^\pi_f] - \rho_\alpha^\lambda[C^{\pi_E}_f]\big)- \psi(f)$, with $\rho_\alpha^\lambda[C^\pi_f]:=\frac{\mathbb{E}[C^\pi_f]+\lambda\rho_\alpha[C^\pi_f]}{1+\lambda}$ being the coherent risk measure for policy $\pi$ corresponding to mean-CVaR with the risk parameter $\lambda$. The parameter $\lambda$ can be interpreted as the tradeoff between the mean performance and risk-sensitivity of the policy. The objective function $\mathcal{L}_\lambda(\pi,\pi_E)$ can be decomposed into three terms: {\bf 1)} the difference between the agent and expert in terms of mean performance, $\mathbb{E}[C^\pi_f]-\mathbb{E}[C^{\pi_E}_f]$, which corresponds to the standard generative imitation learning objective, {\bf 2)} the difference between the agent and the expert in terms of risk $\rho_\alpha[C^\pi_f]-\rho_\alpha[C^{\pi_E}_f]$, and {\bf 3)} the convex regularizer $\psi(f)$ that encodes our belief about the expert cost function $f$.
For the risk-sensitive quantity $\rho_\alpha^\lambda[C^\pi]$, we define the distorted probability distributions $p^\pi_\xi=\xi\cdot p^\pi$, where $\xi=\frac{1+\lambda\zeta}{1+\lambda},\;\zeta\in\mathcal{U}^\pi$. We denote by $\mathcal{P}^\pi_\xi$ the set of such distorted distributions and by $\mathcal{D}^\pi_\xi$ the set of distorted occupancy measures induced by the elements of $\mathcal{P}^\pi_\xi$. Similar to CVaR in Section~\ref{subsec:RS-MDP}, we may write the risk-sensitive quantity $\rho_\alpha^\lambda[C^\pi]$ as the expectation $\rho_\alpha^\lambda[C^\pi]=\mathbb{E}_{p^\pi_{\xi^*}}[C(\tau)]=\mathbb{E}_{d^\pi_{\xi^*}}[c(s,a)]$, where $\xi^*=\frac{1+\lambda\zeta^*}{1+\lambda}$ with $\zeta^*$ defined in Section~\ref{subsec:RS-MDP} and $d^\pi_{\xi^*}\in\mathcal{D}^\pi_\xi$ is the distorted occupancy measure induced by $p^\pi_{\xi^*}\in\mathcal{P}^\pi_\xi$.
In Theorem~\ref{thm:convex-conjugate}, we show that the maximization problem $\mathcal{L}_\lambda(\pi,\pi_E)$ over the cost function $f\in\mathcal{C}$ can be rewritten as a sup-inf problem over the distorted occupancy measures $d\in\mathcal{D^\pi_\xi}$ and $d'\in\mathcal{D}^{\pi_E}_\xi$.
\begin{theorem}
\label{thm:convex-conjugate}
Let $\psi:\mathcal{C}\to\mathbb{R}\cup\{\infty\}$ be a convex cost function regularizer. Then,
\begin{equation}
\label{eq:inner_problem}
\mathcal{L}_\lambda(\pi,\pi_E) = \sup_{f \in \mathcal{C}}\;(1+\lambda)\big(\rho^\lambda_\alpha[C^\pi_f] - \rho^\lambda_\alpha[C^{\pi_E}_f]-\psi(f)\big) = \sup_{d\in\mathcal{D}^\pi_\xi} \inf_{d'\in\mathcal{D}^{\pi_E}_\xi} \psi^*\big((1+\lambda)(d-d')\big),
\end{equation}
where $\psi^*$ is the convex conjugate function of $\psi$, i.e.,~$\psi^*(d) = \sup_{f\in\mathcal{C}}d^\top f-\psi(f)$
\end{theorem}
\begin{proof}
See Appendix~\ref{app:thm1}.
\end{proof}
From Theorem~\ref{thm:convex-conjugate}, we may write the RS-GAIL optimization problem~\eqref{eq:agent-objective-5} as
\begin{equation}
\label{eq:agent-objective-6}
\textbf{(RS-GAIL)} \qquad\qquad \sup_{\lambda \ge 0}\;\min_\pi\;-H(\pi) + \sup_{d\in\mathcal{D}^\pi_\xi} \inf_{d'\in\mathcal{D}^{\pi_E}_\xi} \psi^*\big((1+\lambda)(d-d')\big).
\end{equation}
Comparing the RS-GAIL optimization problem~\eqref{eq:agent-objective-6} with that of GAIL (see Eq.~4 in~\citealt{GAIL}), we notice that the main difference is the $\sup_{\mathcal{D}^\pi_\xi}\inf_{\mathcal{D}^{\pi_E}_\xi}$ in RS-GAIL that does not exist in GAIL. In the risk-neutral case, $\lambda=0$, and thus, the two sets of distorted occupancy measures $\mathcal{D}^\pi_\xi$ and $\mathcal{D}^{\pi_E}_\xi$ are singleton and the RS-GAIL optimization problem is reduced to that of GAIL.
\begin{example}
Let $\psi(f) = \begin{cases} 0 & \text{if} \;\; ||f||_{\infty} \le 1 \\ +\infty & \text{otherwise}\end{cases}$, then $\mathcal{L}_{\lambda}(\pi, \pi_E) = (1+\lambda)\sup_{d \in \mathcal{D}^\pi_\xi} \inf_{d' \in \mathcal{D}^{\pi_E}_\xi} ||d-d^{\prime}||_{\text{TV}}$, where $||d-d'||_{\text{TV}}$ is the total variation distance between $d$ and $d'$. Note that similar to GAIL, our optimization problem aims at learning the expert's policy by matching occupancy measures. However, in order to take risk into account, it now involves matching two sets of occupancy measures (w.r.t.~the TV distance) that encode the risk profile of each policy.
\end{example}
\subsection{Risk-sensitive GAIL with Jensen-Shannon Divergence}
\label{subsec:JS-RS-GAIL}
In this section, we derive RS-GAIL using occupation measure matching via Jensen-Shannon (JS) divergence. We define the cost function regularizer $\psi(f):=\begin{cases} (1+\lambda)\big(-\rho_\alpha^\lambda[C^{\pi_E}_f]+\rho_\alpha^\lambda[G_f^{\pi_E}]\big) & \text{if} \;\;f<0 \\ +\infty & \text{otherwise}\end{cases}$, where $C_f^{\pi_E}$ and $G_f^{\pi_E}$ are the loss random variables of policy $\pi_E$ w.r.t.~the cost functions $c(s,a)=f(s,a)$ and $c(s,a)=g\big(f(s,a)\big)$, respectively, with $g(x):=\begin{cases} -\log(1-e^x) & \text{if} \;\;x<0 \\ +\infty & \text{otherwise}\end{cases}$. To clarify, $G_f^{\pi_E}$ is a random variable whose instantiations are $G_f(\tau)=\sum_{t=0}^{T-1}\gamma^tg\big(f(s_t,a_t)\big)$, where $\tau\sim p^{\pi_E}$ is a trajectory generated by the expert policy $\pi_E$. As described in~\citet{GAIL}, this regularizer places low penalty on cost functions $f$ that assign negative cost to expert's state-action pairs. However, if $f$ assigns large costs (close to zero, which is the upper-bound of the regularizer) to the expert, then $\psi$ will heavily penalize $f$.
In the following theorems, whose proofs are reported in Appendix~\ref{app:JS-RS-GAIL}, we derive the optimization problem of the JS version of our RS-GAIL algorithm by computing~\eqref{eq:agent-objective-6} for the above choice of the cost function regularizer $\psi(f)$.
\begin{theorem}
\label{thm:JS-RS-GAIL}
With the cost function regularizer $\psi(f)$ defined above, we may write
\begin{equation}
\label{eq:JS1}
\mathcal{L}_\lambda(\pi,\pi_E) \geq (1+\lambda)\sup_{f:\mathcal{S}\times\mathcal{A}\to (0,1)}\rho^\lambda_\alpha[F_{1,f}^\pi] - \rho^\lambda_\alpha[-F_{2,f}^{\pi_E}],
\end{equation}
where $F_1^\pi$ and $F_2^{\pi_E}$ are the loss random variables of policies $\pi$ and $\pi_E$ w.r.t.~the cost functions $c(s,a)=\log f(s,a)$ and $c(s,a)=\log\big(1-f(s,a)\big)$, respectively.
\end{theorem}
\begin{corollary}
\label{coro:JS-RS-GAIL}
We may write $\mathcal{L}_{\lambda}(\pi, \pi_E)$ in terms of the Jensen-Shannon (JS) divergence as
\begin{equation}
\label{eq:JS2}
\mathcal{L}_{\lambda}(\pi, \pi_E) \geq (1+\lambda)\sup_{d \in \mathcal{D}^\pi_\xi} \inf_{d' \in \mathcal{D}^{\pi_E}_\xi} D_{\text{JS}}(d,d').
\end{equation}
\end{corollary}
From Theorem~\ref{thm:JS-RS-GAIL}, we write the optimization problem of the JS version of our RS-GAIL algorithm as
\begin{equation}
\label{eq:JS-RS-GAIL}
\textbf{(JS-RS-GAIL)} \qquad \sup_{\lambda \ge 0}\;\min_\pi\;-H(\pi) + (1+\lambda)\sup_{f:\mathcal{S}\times\mathcal{A}\to (0,1)}\rho^\lambda_\alpha[F_{1,f}^\pi] - \rho^\lambda_\alpha[-F_{2,f}^{\pi_E}].
\end{equation}
Hence in JS-RS-GAIL, instead of minimizing the original GAIL objective, we solve the optimization problem~\eqref{eq:JS-RS-GAIL} that aims at matching the sets $\mathcal{D}^\pi_\xi$ and $\mathcal{D}^{\pi_E}_\xi$ w.r.t.~the JS divergence.
\subsection{Risk-sensitive GAIL with Wasserstein Distance}
\label{subsec:W-RS-GAIL}
In this section, we derive RS-GAIL using occupation measure matching via the Wasserstein distance. We define the cost function regularizer $\psi(f):=\begin{cases} 0 & \text{if} \;\; f \in \mathcal{F}_1\\ + \infty & \text{otherwise}\end{cases}$.
\begin{corollary}
\label{coro:W-RS-GAIL}
For the cost function regularizer $\psi(f)$ defined above, we may write
\begin{equation}
\label{eq:W}
\mathcal{L}_{\lambda}(\pi, \pi_E) = (1+\lambda)\sup_{d \in \mathcal{D}^\pi_\xi} \inf_{d' \in \mathcal{D}^{\pi_E}_\xi} W(d,d').
\end{equation}
\end{corollary}
\begin{proof}
See Appendix~\ref{app:W-RS-GAIL}.
\end{proof}
From~\eqref{eq:inner_problem} and the $\psi(f)$ defined above, we have $\mathcal{L}_\lambda(\pi,\pi_E)=\sup_{f\in\mathcal{F}_1}\;\rho_\alpha^\lambda[C^\pi_f]-\rho_\alpha^\lambda[C^{\pi_E}_f]$, which gives the following optimization problem for the Wasserstein version of our RS-GAIL algorithm:
\begin{equation}
\label{eq:W-RS-GAIL}
\textbf{(W-RS-GAIL)} \qquad \sup_{\lambda \ge 0}\;\min_\pi\;-H(\pi) + (1+\lambda)\sup_{f\in\mathcal{F}_1}\rho^\lambda_\alpha[C^\pi_f] - \rho^\lambda_\alpha[C^{\pi_E}_f].
\end{equation}
We conclude this section by a theorem that shows if we use a risk-neutral imitation learning algorithm to minimize the Wasserstein distance between the occupancy measures of the agent and the expert, the distance between their CVaRs could be still large. Thus, new algorithms, such as those developed in this paper, are needed for risk-sensitive imitation learning.
\begin{theorem}
\label{th:worst-case risk difference}
Let $T=1$ be the horizon of the decision problem and $\Delta$ be the worst-case risk difference between the agent and expert, given that their occupancy measures are $\delta$-close ($\delta>0$), i.e.,
\begin{equation}
\begin{split}
\Delta = & \sup_\pi\;\sup_{f \in \mathcal{F}_1}\;\rho_{\alpha}[C_{f}^{\pi}] - \rho_{\alpha}[C_{f}^{\pi_E}] \qquad , \qquad \text{s.t.} \;\; W(d^\pi, d^{\pi_E}) \le \delta.
\end{split}
\end{equation}
Then, $\Delta = \frac{\delta}{\alpha}$.
\end{theorem}
Theorem~\ref{th:worst-case risk difference}, whose proof has been reported in Appendix~\ref{app:W-RS-GAIL}, indicates that the difference between the risks can be $1/\alpha$-times larger than that between the occupancy measures (in terms of Wasserstein).
\section{Practical version of RS-GAIL: Algorithm}
\label{sec:algo}
We now present a practical version of RS-GAIL for model-free imitation in large environments. The goal is to find a saddle-point $(\pi, f)$ of the objective $-H(\pi) + (1+\lambda) [\rho_{\alpha}^{\lambda}[C_{f}^{\pi}] - \rho_{\alpha}^{\lambda}[C_{f}^{\pi} \psi(f)]$. We introduce function approximation for $\pi$ and $f$. Let $w \mapsto f_w$ be a parameterization of the cost function/discriminator $f$ and $\theta \mapsto \pi_{\theta}$ a parameterization of the policy $\pi$. The algorithm alternates between an Adam~\citep{adam} gradient ascent step for the cost function/discriminator parameter $w$ and KL-constrained gradient descent step with respect to a linear approximation of the objective.
\begin{algorithm}[H]
\label{alg:Algorithm1}
\DontPrintSemicolon
\SetKwInOut{Input}{Input}
\Input{Expert trajectories $\tau^E_{j} \sim \pi_E$ for $j=1,\dots,N_E$, risk level $\alpha \in (0,1]$ and initial policy and cost function parameters $\theta_0, w_0$.}
\For{$i=0,1,2,\dots$}{
Sample trajectories with current policy $\tau_j \sim \pi_{\theta_i}$ with $j=1,\dots,N$.\\
(JS-RS-GAIL) Compute estimates of the $(1-\alpha)$-quantiles $\hat{\nu}_{\alpha}(F_{1, f_{w_i}}^{\pi})$ and $\hat{\nu}_{\alpha}(-F_{2, f_{w_i}}^{\pi_E})$.\\
(W-RS-GAIL) Compute estimates of the $(1-\alpha)$-quantiles $\hat{\nu}_{\alpha}(C_{f_{w_i}}^{\pi})$ and $\hat{\nu}_{\alpha}(C_{f_{w_i}}^{\pi_E})$.\\
(JS-RS-GAIL) Update the discriminator parameters from $w_i$ to $w_{i+1}$ by computing a gradient ascent step with respect to the objective $w \mapsto (1+\lambda) \left( \rho_{\alpha}^{\lambda}[F_{1,f_w}^{\pi_{\theta_i}}] - \rho_{\alpha}^{\lambda}[-F_{2,f_{w}}^{\pi_E}] \right)$ (see Appendix \ref{sec:grad}).\\
(W-RS-GAIL) Update the discriminator parameters from $w_i$ to $w_{i+1}$ by computing a gradient ascent step with respect to the objective $w \mapsto (1+\lambda) \left( \rho_{\alpha}^{\lambda}[C_{f_w}^{\pi_{\theta_i}}] - \rho_{\alpha}^{\lambda}[C_{f_{w}}^{\pi_E}] \right)$ (see Appendix \ref{sec:grad}).\\
(JS-RS-GAIL) Update the policy parameters from $\theta_i$ to $\theta_{i+1}$ using a KL-constrained gradient descent step with respect to the objective $\theta \mapsto -H(\pi_{\theta}) + (1+\lambda) \rho_{\alpha}^{\lambda}[F_{1,f_{w_{i+1}}}^{\pi_{\theta}}]$ (see Appendix \ref{sec:grad}).\\
(W-RS-GAIL) Update the policy parameters from $\theta_i$ to $\theta_{i+1}$ using a KL-constrained gradient descent step with respect to the objective $\theta \mapsto -H(\pi_{\theta}) + (1+\lambda) \rho_{\alpha}^{\lambda}[C_{f_{w_{i+1}}}^{\pi_{\theta}}]$ (see Appendix \ref{sec:grad}).\\
}
\caption{JS-RS-GAIL}
\end{algorithm}
\subsubsection*{Practical choice of the mean-risk trade-off parameter $\lambda$}
Consider the task of deploying a robot in a risky environment with several types of random and dangerous disturbances. The environment being complex to describe, we rely on demonstrations from a human expert to perform the task. In particular, we assume the human expert's demonstrations to perform the task well (good mean performance) and to be sufficiently risk-sensitive (good CVaR). Importantly, we want to make sure that when the robot is deployed, then the risks criteria of the policy and the expert are comparable, i.e., the policy does not encounter dangerous states. Hence, we want to choose a parameter $\lambda$ that puts enough emphasis on the risk. On the other hand, we still want to have acceptable performance even at the first shot. The worst-case stated by Theorem \ref{th:worst-case risk difference} gives the designer an order for the value of $\lambda$: $|\rho_{\alpha}[C^{\pi}] - \rho_{\alpha}[C^{\pi_E}]| \simeq \frac{1}{\alpha} |\mathbb{E}[C^{\pi}] - \mathbb{E}[C^{\pi_E}]|$. Hence, a value $\lambda^* \simeq \alpha$ ensures that the risk and mean terms are, in a worst-case, of the same order. Roughly, one can start by choosing a conservative value $\lambda$ slightly greater than $\alpha$, deploy the system and then retrain an imitation policy with smaller and smaller values of $\lambda$ until reaching a satisfying mean-risk trade-off based on real empirical observations.
\section{Related Work: Discussion about RAIL}
\label{sec:RAIL}
We start this section by comparing the RAIL optimization problem (Eq.~9 in~\citealt{RAIL}) with that of our JS-RS-GAIL reported in Eq.~\ref{eq:JS-RS-GAIL}, i.e.,
\begin{align*}
\textbf{(RAIL)} \qquad &\min_\pi\;-H(\pi) + (1+\lambda)\sup_{f:\mathcal{S}\times\mathcal{A}\to (0,1)}\rho^\lambda_\alpha[F_{1,f}^\pi] - \mathbb{E}[-F_{2,f}^{\pi_E}], \\
\textbf{(JS-RS-GAIL)} \qquad &\min_\pi\;-H(\pi) + (1+\lambda)\sup_{f:\mathcal{S}\times\mathcal{A}\to (0,1)}\rho^\lambda_\alpha[F_{1,f}^\pi] - \rho^\lambda_\alpha[-F_{2,f}^{\pi_E}].
\end{align*}
If we write the above optimization problems in terms of the JS divergence, we obtain
\begin{align}
\label{eq:RAIL007}
\textbf{(RAIL)} \qquad &\min_\pi\;-H(\pi) + (1+\lambda)\sup_{d \in \mathcal{D}^\pi_\xi} D_{\text{JS}}(d,d^{\pi_E}), \\
\textbf{(JS-RS-GAIL)} \qquad &\min_\pi\;-H(\pi) + (1+\lambda)\sup_{d \in \mathcal{D}^\pi_\xi} \inf_{d' \in \mathcal{D}^{\pi_E}_\xi} D_{\text{JS}}(d,d'). \qquad\quad \textit{(see Eq.~\ref{eq:JS2})}
\label{eq:JS007}
\end{align}
Note that while JS in~\eqref{eq:JS007} matches the distorted occupancy measures (risk profiles) of the agent and expert, the JS in~\eqref{eq:RAIL007} matches the distorted occupancy measure (risk profile) of the agent with the occupancy measure (mean) of the expert. This means that RAIL does not take the expert's risk into account in its optimization.
Moreover, the results reported in~\citet{RAIL} indicate that GAIL performs poorly in terms of optimizing the risk (VaR and CVaR). By looking at the RAIL's GitHub~\citep{RAIL-GIT}, it seems they used the GAIL implementation from its GitHub~\citep{GAIL-GIT}. Although we used the same GAIL implementation, we did not observe such a poor performance for GAIL, which is not that surprising since the MuJoCo domains used in the GAIL and RAIL papers are all deterministic and the policies are the only source of randomness there. This is why in our experiments in Section~\ref{sec:experiments}, we inject noise to the reward function of the problems. Finally, the gradient of the objective function reported in Eq.~(A.3) of~\citet{RAIL} is a scalar, which does not seem to be correct. We corrected this in our implementation of RAIL in Section~\ref{sec:experiments}.
\section{Experimental Results}
\label{sec:experiments}
We evaluated Algorithm~\ref{alg:Algorithm1} against GAIL~\citep{GAIL} and RAIL~\citep{RAIL} on 2 physics-based high-dimensional continuous control tasks (Hopper-v1 and Walker2d-v1), solved efficiently by model-free reinforcement learning~\citep{TRPO},~\citep{duan2016benchmarking}. All environments were simulated with MuJoCo~\citep{todorov2012mujoco}. Each task comes with a true cost function $c(s,a)$, defined in the OpenAI Gym~\citep{OpenAIgym}. For each task, a stochastic expert has been trained~\citep{GAIL} on these true cost functions to minimize the expected cumulative cost. Also, all transitions and costs are deterministic. We transform the environment such that (i) the costs and/or transitions are random and (ii) the expert is risk-sensitive with respect to its environment. For simplicity, we opted for a cost transformation. Let $(s,a)$ be any state-action pair. We randomly transform the cost $c(s,a)$ into $c_M(s,a) := g(\omega, d_{\pi_E}(\hat{s},\hat{a})) c(s,a)$ where $\omega$ indicates that $g$ is random and $(\hat{s}, \hat{a})$ is the closest state-action pair to $(s,a)$ such that $d_{\pi_E}(\hat{s}, \hat{a}) > 0$. For high values of $d_{\pi_E}(\hat{s},\hat{a})$, we choose $g(d_{\pi_E}(\hat{s},\hat{a})) \simeq 1$ with high probability and, for small values of $d_{\pi_E}(\hat{s},\hat{a})$, we choose $g(d_{\pi_E}(\hat{s},\hat{a}))$ to either leave the cost unchanged or decrease it significantly, both with non-negligible probability. Then, the behavior of the expert becomes risk-sensitive with respect to the random cost function $c_M$. Indeed, in regions where the original cost $c(s,a)$ is large, she has small occupancy measure $d_{\pi_E}(s,a)$. Even though $c_M(s,a)$ can be small, it might be as large as $c(s,a)$ with non-negligible probability. On the other hand, she has higher occupancy measure in regions where the cost $c(s,a)$ is relatively small and the modified cost $c_M(s,a)$ is concentrated around $c(s,a)$. For more details about the implementation, we refer the reader to Appendix~\ref{app:detail-experiments}.
For each task, we used JS-RS-GAIL, GAIL and RAIL to train policies of the same neural network architecture, with two layers and tanh nonlinearities in between. The first, respectively second, layer contains a number of neurons on the order of the observation space dimension, respectively action space dimension. Hence, we have a faster training procedure compared to~\citep{GAIL} that uses $100$ neurons for each layer. Moreover, it has recently been shown~\citep{mania2018simple} that policies parameterized by spaces with such dimensions can be trained to achieve state-of-the-art performance on MuJoCo tasks. The discriminator networks of Algorithm~\ref{alg:Algorithm1} also used the same architecture. For each task, we gave to all algorithms the same amount of environment interaction for training. Table~\ref{environmentsparams} lists the names and version of used environments, the dimension of their observation and action spaces, the number of training iterations (same for each algorithm) and the amount of environment interaction. Table~\ref{performance} shows the exact experimental performance with respect to the mean, $\text{VaR}_{\alpha}$, $\text{CVaR}_{\alpha}$ and $\rho_{\alpha}^{\lambda}$ of the random cumulative cost with respect to the modified cumulative cost function. Due to the increasing amount of samples required to estimate $\text{VaR}_{\alpha}$ and $\text{CVaR}_{\alpha}$ when $\alpha$ decreases, we chose $\alpha = 0.3$, meaning that we are interested in the performance for the $30$\% worst-case outcomes. For each task, each algorithm is run using 5 different random seeds. For each run, we sample $1000$ trajectories using the trained policy. We report the average estimates of each criteria.
On the two high-dimensional control tasks, our algorithm produced policies that (i) perform at least as well as GAIL w.r.t. mean criteria and (ii) ourperform GAIL w.r.t. the risk criteria $\rho_{\alpha}^{\lambda}$. The risk performance of JS-RS-GAIL is actually much closer to the expert one than GAIL. We also observed slight improvements on RAIL for Hopper-v1 and significant improvements for Walker-v1.
\begin{table}[!h]
\caption{Environments and parameters}
\label{environmentsparams}
\centering
\begin{tabular}{lclclclcl}
\toprule
Task & Observation-action space & Training iterations & State-action pairs per iteration \\
\midrule
Hopper-v1 & (11, 3) & 500 & 50000 \\
Walker-v1 & (17, 6) & 500 & 50000 \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[!h]
\caption{Learned policy performance, $\alpha=0.3$, $\lambda=0.05$.}
\label{performance}
\centering
\begin{tabular}{lclclclclcl}
\cmidrule(r){1-5}
Criteria & Expert & GAIL & RAIL & Ours \\
\hline
\multicolumn{5}{c}{Hopper-v1}\\
\midrule
Mean & - 6096 & -5853 & -6064 & -6105 \\
$\text{VaR}_{\alpha}$ & -6129 & -6019 & -6125 & -6124 \\
$\text{CVaR}_{\alpha}$ & -5590 & -4958 & -5493 & -5657 \\
$\rho_{\alpha}^{\lambda}$ & -6375 & -6100 & -6338 & -6387\\
\bottomrule
\end{tabular}
\quad
\begin{tabular}{lclclclclcl}
\cmidrule(r){1-5}
Criteria & Expert & GAIL & RAIL & Ours \\
\hline
\multicolumn{5}{c}{Walker-v1}\\
\midrule
Mean & -7651 & -7231 & -7363 & -7572 \\
$\text{VaR}_{\alpha}$ & -7875 & -7274 & -7773 & -7909 \\
$\text{CVaR}_{\alpha}$ & -6440 & -5353 & -5505 & -5926 \\
$\rho_{\alpha}^{\lambda}$ & -7973 & -7498 & -7638 & -7868\\
\bottomrule
\end{tabular}
\end{table}
\begin{figure}[h]
\caption{Training curves, averaged over $5$ runs.}
\begin{subfigure}
\centering
\includegraphics[width=0.5\columnwidth]{means_hopper.eps}
\end{subfigure}
\begin{subfigure}
\centering
\includegraphics[width=0.5\columnwidth]{cvars_hopper.eps}
\end{subfigure}
\begin{subfigure}
\centering
\includegraphics[width=0.5\columnwidth]{means_walker.eps}
\end{subfigure}
\begin{subfigure}
\centering
\includegraphics[width=0.5\columnwidth]{cvars_walker.eps}
\end{subfigure}
\end{figure}
\section{Conclusions and Future Work}
\label{sec:conclu}
In this paper, we first formulated a risk-sensitive imitation learning setting in which the agent's goal is to have a risk profile as good as expert's. We then derived a GAIL-like optimization problem for our formulation, which we called it risk-sensitive GAIL (RS-GAIL). We proposed two risk-sensitive generative adversarial imitation learning algorithms based on two variations of RS-GAIL that match the agent and expert's risk profiles w.r.t.~Jensen-Shannon (JS) divergence and Wasserstein distance. We experimented with our JS-based algorithm and compared its performance with that of GAIL~\citep{GAIL} and RAIL~\citep{RAIL} in two MuJoCo tasks.
Future directions include {\bf 1)} extending our results to other popular risk measures, such as expected exponential utility and the more general class of coherent risk measures, {\bf 2)} investigating other risk-sensitive imitation learning settings, especially those in which the agent can tune its risk profile w.r.t.~the expert, e.g.,~being a more risk averse/seeking version of the expert, and {\bf 3)} more experiments, particularly with our Wasserstein-based algorithm and in problems with higher intrinsic stochasticity.
\newpage
\subsubsection*{\bibname}}
\begin{document}
\twocolumn[
\aistatstitle{Risk-Sensitive Generative Adversarial Imitation Learning}
\aistatsauthor{ Jonathan Lacotte \And Mohammad Ghavamzadeh \And Yinlam Chow \And Marco Pavone }
\aistatsaddress{ Stanford University \And Facebook AI Research \And DeepMind \And Stanford University } ]
\begin{abstract}
\vspace{-0.2in}
We study risk-sensitive imitation learning where the agent's goal is to perform at least as well as the expert in terms of a risk profile. We first formulate our risk-sensitive imitation learning setting. We consider the generative adversarial approach to imitation learning (GAIL) and derive an optimization problem for our formulation, which we call it risk-sensitive GAIL (RS-GAIL). We then derive two different versions of our RS-GAIL optimization problem that aim at matching the risk profiles of the agent and the expert w.r.t. Jensen-Shannon (JS) divergence and Wasserstein distance, and develop risk-sensitive generative adversarial imitation learning algorithms based on these optimization problems. We evaluate the performance of our algorithms and compare them with GAIL and the risk-averse imitation learning (RAIL) algorithms in two MuJoCo and two OpenAI classical control tasks.
\end{abstract}
\vspace{-0.275in}
\section{Introduction}
\vspace{-0.125in}
\label{sec:intro}
We study imitation learning, i.e.,~the problem of learning to perform a task from the sample trajectories generated by an expert. There are three main approaches to this problem: {\bf 1)} behavioral cloning (e.g.,~\cite{Pomerleau91ET}) in which the agent learns a policy by solving a supervised learning problem over the state-action pairs of the expert's trajectories, {\bf 2)} inverse reinforcement learning (IRL) (e.g.,~\cite{Ng00AI}) followed by reinforcement learning (RL), a process also referred to as RL$\circ$IRL, where we first find a cost function under which the expert is optimal (IRL) and then return the optimal policy w.r.t.~this cost function (RL), and {\bf 3)} generative adversarial imitation learning (GAIL)~\citep{GAIL} that frames the imitation learning problem as occupancy measure matching w.r.t.~either the Jensen-Shannon divergence (GAIL)~\citep{GAIL} or the Wasserstein distance (InfoGAIL)~\citep{INFOGAIL}. Behavioral cloning algorithms are simple but often need a large amount of data to be successful. IRL does not suffer from the main problems of behavioral cloning~\citep{RossB10,RossGB11}, since it takes entire trajectories into account (instead of single time-step decisions) when learning a cost function. However, IRL algorithms are often expensive to run as they require solving a RL problem in their inner loop. This issue had restricted the use of IRL to small problems for a long while and only recently scalable IRL algorithms have been developed~\citep{Levine12CI,pmlr-v48-finn16}. On the other hand, the nice feature of the GAIL approach to imitation learning is that it bypasses the intermediate IRL step and directly learns a policy from data, as if it were obtained by RL$\circ$IRL. The resulting algorithm is closely related to generative adversarial networks (GAN)~\citep{GAN} that has recently gained attention in the deep learning community.
In many applications, we may prefer to optimize some measure of risk in addition to the standard optimization criterion, i.e.,~the expected sum of (discounted) costs. In such cases, we would like to use a criterion that incorporates a penalty for the variability (due to the stochastic nature of the system) induced by a given policy. Several risk-sensitive criteria have been studied in the literature of risk-sensitive Markov decision processes (MDPs)~\citep{Howard72RS} including the expected exponential utility (e.g.,~\cite{Howard72RS,Borkar01SF}), a variance-related measure (e.g.,~\cite{Sobel82VD,tamar2012policy,Prashanth13AC}), or the tail-related measures like value-at-risk (VaR) and conditional value-at-risk (CVaR) (e.g.,~\cite{Rockafellar00OC,saferl:cvar:chow2014,Tamar15OC}).
In risk-sensitive imitation learning, the agent's goal is to perform at least as well as the expert in terms of one or more risk-sensitive objective(s), e.g.,~$\text{mean}+\lambda\text{CVaR}_\alpha$, for one or more values of $\lambda\geq 0$. This goal cannot be satisfied by risk-neutral imitation learning. As we will show in Section~\ref{subsec:W-RS-GAIL}, if we use GAIL to minimize the Wasserstein distance between the occupancy measures of the agent and the expert, the distance between their CVaRs could be still large.~\citet{RAIL} recently showed empirically that the policy learned by GAIL does not have the desirable tail properties, such as VaR and CVaR, and proposed a modification of GAIL, called risk-averse imitation learning (RAIL), to address this issue. We will discuss about RAIL in more details in Section~\ref{sec:RAIL} as it is probably the closest work to us in the literature. Another related work is by~\citet{Singh18RI} on risk-sensitive IRL in which the proposed algorithm infers not only the expert's cost function but her underlying risk measure, for a rich class of static and dynamic risk measures (coherent risk measures). The agent then learns a policy by optimizing the inferred risk-sensitive objective.
In this paper, we study an imitation learning setting in which the agent's goal is to learn a policy with minimum expected sum of (discounted) costs and with $\text{CVaR}_\alpha$ that is at least as well as that of the expert. We first provide a mathematical formulation for this setting and derive a GAIL-like optimization problem for our formulation, which we call it risk-sensitive GAIL (RS-GAIL), in Section~\ref{subsec:PF}. In Sections~\ref{subsec:JS-RS-GAIL} and~\ref{subsec:W-RS-GAIL}, we define cost function regularizers that when we compute their convex conjugates and plug them into our RS-GAIL objective function, the resulting optimization problems aim at learning the expert's policy by matching occupancy measures w.r.t.~Jensen-Shannon (JS) divergence and Wasserstein distance, respectively. We call the resulting optimization problems JS-RS-GAIL and W-RS-GAIL and propose our risk-sensitive generative adversarial imitation learning algorithms based on these optimization problems in Section~\ref{sec:algo}. It is important to note that unlike the risk-neutral case in which the occupancy measure of the agent is matched with that of the expert, here in the risk-sensitive case, we match two sets of occupancy measures that encode the risk profile of the agent and the expert. This will become more clear in Section~\ref{sec:RSIL}. We present our understanding of RAIL and how it is related to our work in Section~\ref{sec:RAIL}. In Section~\ref{sec:experiments}, we evaluate the performance of our algorithms and compare them with GAIL and RAIL in two MuJoCo tasks~\citep{Todorov12MuJoCo} that have also been used in the GAIL~\citep{GAIL} and RAIL~\citep{RAIL} papers, as well as two OpenAI classical control problems~\citep{brockman2016openai}. Finally in Section~\ref{sec:conclu}, we conclude the paper and list a number of future directions.
\vspace{-0.1in}
\section{Preliminaries}
\vspace{-0.1in}
\label{sec:Prelim}
We consider the scenario in which the agent's interaction with the environment is modeled as a Markov decision process (MDP). A MDP is a tuple $\mathcal{M} = \{\mathcal{S}, \mathcal{A}, c, p, p_0, \gamma\}$, where $\mathcal{S}$ and $\mathcal{A}$ are state and action spaces; $c: \mathcal{S} \times \mathcal{A} \to \mathbb{R}$ and $p:\mathcal{S}\times\mathcal{A}\rightarrow\Delta_{\mathcal{S}}$ are the cost function and transition probability distribution, with $c(s,a)$ and $p(\cdot|s,a)$ being the cost and next state probability of taking action $a$ in state $s$; $p_0:\mathcal{S}\rightarrow\Delta_\mathcal{S}$ is the initial state distribution; and $\gamma\in[0,1)$ is a discounting factor.
A stationary stochastic policy $\pi:\mathcal{S}\rightarrow\Delta_{\mathcal{A}}$ is a mapping from states to a distribution over actions. We denote by $\Pi$ the set of all such policies.
We denote by $\tau=(s_0,a_0,s_1,a_1,\ldots,s_T)\in\Gamma$, where $a_t\sim\pi(\cdot|s_t),\;\forall t\in\{0,\ldots,T-1\}$, a trajectory of the fixed horizon $T$ generated by policy $\pi$, by $\Gamma$ the set of all such trajectories, and by $C(\tau)=\sum_{t=0}^{T-1}\gamma^tc(s_t,a_t)$ the {\em loss} of trajectory $\tau$. The probability of trajectory $\tau$ is given by $\mathbb{P}(\tau|\pi)=p^\pi(\tau)=p_0(s_0)\prod_{t=0}^{T-1}\pi(a_t|s_t)p(s_{t+1}|s_t,a_t)$. We denote by $C^\pi$ the random variable of the loss of policy $\pi$. Thus, when $\tau\sim p^\pi$, $C(\tau)$ is an instantiation of the random variable $C^\pi$. The performance of a policy $\pi$ is usually measured by a quantity related to the loss of the trajectories it generates, the most common would be its expectation, i.e.,~$\mathbb{E}[C^\pi]=\mathbb{E}_{\tau\sim p^\pi}[C(\tau)]$. We define the occupancy measure of policy $\pi$ as $d^\pi(s,a) = \sum_{t=0}^{T} \gamma^t \mathbb{P}(s_t = s, a_t = a | \pi)$, which can be interpreted as the unnormalized distribution of the state-action pairs visited by the agent under policy $\pi$. Using occupancy measure, we may write the policy's performance as $\mathbb{E}[C^\pi]=\mathbb{E}_{p^\pi}[C(\tau)]=\mathbb{E}_{d^\pi}[c(s,a)]=\sum_{s,a}d^\pi(s,a)c(s,a)$.
\vspace{-0.1in}
\subsection{Risk-sensitive MDPs}
\vspace{-0.1in}
\label{subsec:RS-MDP}
In risk-sensitive decision-making, in addition to optimizing the expectation of the loss, it is also important to control the variability of this random variable. This variability is often measured by the variance or tail-related quantities such as value-at-risk (VaR) and conditional value-at-risk (CVaR). Given a policy $\pi$ and a confidence level $\alpha \in (0,1]$, we define the VaR at level $\alpha$ of the loss random variable $C^\pi$ as its (left-side) $(1-\alpha)$-quantile, i.e.,~$\nu_\alpha[C^\pi] := \inf\{t\in\mathbb{R}\;|\;\mathbb{P}(C^\pi\le t) \ge 1-\alpha \}$ and its CVaR at level $\alpha$ as $\rho_\alpha[C^\pi]=\inf_{\nu\in\mathbb{R}}\big\{\nu+\frac{1}{\alpha} \mathbb{E}\big[(C^\pi - \nu)_+\big]\big\}$, where $x_+ = \max(x,0)$. We also define the {\em risk envelope} $\mathcal{U}^\pi = \big\{\zeta : \Gamma^\pi \to [0,\frac{1}{\alpha}]\;|\;\sum_{\tau \in \Gamma}\zeta(\tau)\cdot p^\pi(\tau)= 1\big\}$, which is a compact, convex, and bounded set. The quantities $p^\pi_\zeta=\zeta\cdot p^\pi,\;\zeta\in\mathcal{U}^\pi$ are called {\em distorted probability distributions}, and we denote by $\mathcal{P}^\pi_\zeta=\big\{p^\pi_\zeta\;|\;\zeta\in\mathcal{U}^\pi\big\}$ the set of such distributions. The set $\mathcal{P}^\pi_\zeta$ induces a set of {\em distorted occupancy measures} $\mathcal{D}^\pi_\zeta$, where each element of $\mathcal{D}^\pi_\zeta$ is the occupancy measure induced by a distorted probability distribution in $\mathcal{P}^\pi_\zeta$. The sets $\mathcal{P}^\pi_\zeta$ and $\mathcal{D}^\pi_\zeta$ characterize the risk of policy $\pi$. Given the risk envelope $\mathcal{U}^\pi$, we may define the dual representation of CVaR as $\rho_\alpha[C^\pi] = \sup_{\zeta\in\mathcal{U}^\pi}\mathbb{E}_{\tau\sim p^\pi}\big[\zeta(\tau)C(\tau)\big]$, where the supremum is attained at the density $\zeta^*(\tau) = \frac{1}{\alpha} \mathbf{1}_{\{C(\tau) \ge \nu_\alpha[C^\pi]\}}$. Hence, CVaR can be considered as the expectation of the loss random variable, when the trajectories are generated from the distorted distribution $p^\pi_{\zeta^*}=\zeta^*\cdot p^\pi$, i.e.,~$\rho_\alpha[C^\pi]=\mathbb{E}_{\tau\sim p^\pi_{\zeta^*}}[C(\tau)]$. If we denote by $d^\pi_{\zeta^*}\in\mathcal{D}^\pi_\zeta$ the distorted occupancy measure induced by $p^\pi_{\zeta^*}$, then we may write the CVaR as $\rho_\alpha[C^\pi]=\mathbb{E}_{p^\pi_{\zeta^*}}[C(\tau)]=\mathbb{E}_{d^\pi_{\zeta^*}}[c(s,a)]$.
%
%
%
%
\vspace{-0.1in}
\subsection{Generative Adversarial Imitation Learning}
\vspace{-0.1in}
\label{subsec:GAIL}
As discussed in Section~\ref{sec:intro}, generative adversarial imitation learning (GAIL)~\citep{GAIL} is a framework for directly extracting a policy from the trajectories generated by an expert policy $\pi_E$, as if it were obtained by inverse RL (IRL) followed by RL, i.e.,~RL$\circ$IRL$(\pi_E)$. The main idea behind GAIL is to formulate imitation learning as occupancy measure matching w.r.t.~the Jensen-Shannon divergence $D_{\text{JS}}$, i.e.,~$\min_\pi \big(D_{\text{JS}}(d^\pi,d^{\pi_E})-\lambda H(\pi)\big)$, where $H(\pi)=\mathbb{E}_{(s,a)\sim d^\pi}[-\log\pi(a|s)]$ is the $\gamma$-discounted causal entropy of policy $\pi$, $\lambda\geq 0$ is a regularization parameter, and $D_{\text{JS}}(d^{\pi},d^{\pi_E}) := \sup_{f:\mathcal{S}\times\mathcal{A}\to (0,1)} \mathbb{E}_{d^\pi}[\log f(s,a)] + \mathbb{E}_{d^{\pi_E}}[\log(1-f(s,a))]$.~\citet{INFOGAIL} proposed InfoGAIL by reformulating GAIL and replacing the Jensen-Shannon divergence $D_{\text{JS}}(d^\pi,d^{\pi_E})$ with the Wasserstein distance $W(d^\pi,d^{\pi_E}) := \sup_{f \in \mathcal{F}_1} \mathbb{E}_{d^\pi}[f(s,a)] - \mathbb{E}_{d^{\pi_E}}[f(s,a)]$, where $\mathcal{F}_1$ is the set of $1$-Lipschitz functions over $\mathcal{S} \times \mathcal{A}$.
\vspace{-0.125in}
\section{Risk-sensitive Imitation Learning}
\vspace{-0.1in}
\label{sec:RSIL}
In this section, we describe the risk-sensitive imitation learning formulation studied in the paper and derive the optimization problems that our proposed algorithms solve to obtain a risk-sensitive policy from the expert's trajectories.
\vspace{-0.1in}
\subsection{Problem Formulation}
\vspace{-0.1in}
\label{subsec:PF}
As described in Section~\ref{sec:intro}, we consider the risk-sensitive imitation learning setting in which the agent's goal is to learn a policy with minimum loss and with CVaR that is at least as well as that of the expert. Thus, the agent solves the optimization problem
\begin{equation}
\label{eq:agent-objective-1}
\min_\pi \; \mathbb{E}[C^\pi] \qquad , \qquad \text{s.t.}\;\; \rho_\alpha[C^\pi] \le \rho_\alpha[C^{\pi_E}],
\end{equation}
\vspace{-0.15in}
where $C^\pi$ is the loss of policy $\pi$ w.r.t.~the expert's cost function $c$ that is unknown to the agent. The optimization problem~\eqref{eq:agent-objective-1} without the loss of optimality is equivalent to the unconstrained problem
\begin{equation}
\label{eq:agent-objective-2}
\min_\pi\;\sup_{\lambda \ge 0}\;\mathbb{E}[C^\pi] - \mathbb{E}[C^{\pi_E}] + \lambda\big(\rho_\alpha[C^\pi] - \rho_\alpha[C^{\pi_E}]\big).
\end{equation}
\vspace{-0.15in}
Note that $\pi_E$ is a solution of both~\eqref{eq:agent-objective-1} and~\eqref{eq:agent-objective-2}. However, since the expert's cost function is unknown, the agent cannot directly solve~\eqref{eq:agent-objective-2}, and thus, considers the surrogate problem
\vspace{-0.25in}
\begin{small}
\begin{equation}
\label{eq:agent-objective-3}
\min_\pi\;\sup_{f\in\mathcal{C}}\;\sup_{\lambda \ge 0}\; \mathbb{E}[C^\pi_f] - \mathbb{E}[C^{\pi_E}_f] + \lambda \big(\rho_\alpha[C^\pi_f] - \rho_\alpha[C^{\pi_E}_f]\big),
\end{equation}
\end{small}
\vspace{-0.25in}
where $\mathcal{C}=\{f:\mathcal{S}\times\mathcal{A}\to\mathbb{R}\}$ and $C^\pi_f$ is the loss of policy $\pi$ w.r.t.~the cost function $f$. We employ the Lagrangian relaxation procedure~\citep{Bertsekas99NP} to swap the inner maximization over $\lambda$ with the minimization over $\pi$ and convert~\eqref{eq:agent-objective-3} to the problem
\vspace{-0.25in}
\begin{small}
\begin{equation}
\label{eq:agent-objective-4}
\sup_{\lambda \ge 0}\;\min_\pi\;\sup_{f\in\mathcal{C}}\; \mathbb{E}[C^\pi_f] - \mathbb{E}[C^{\pi_E}_f] + \lambda \big(\rho_\alpha[C^\pi_f] - \rho_\alpha[C^{\pi_E}_f]\big).
\end{equation}
\end{small}
\vspace{-0.25in}
We adopt maximum causal entropy IRL formulation~\citep{Ziebart08ME,Ziebart10MI} and add $-H(\pi)$ to the optimization problem~\eqref{eq:agent-objective-4}. Moreover, since $\mathcal{C}$ is large, to avoid overfitting when we are provided with a finite set of expert's trajectories, we add the negative of a convex regularizer $\psi:\mathcal{C}\to\mathbb{R}\cup\{\infty\}$ to the optimization problem~\eqref{eq:agent-objective-4}. As a result we obtain the following optimization problem for our risk-sensitive imitation learning setting, which we call it RS-GAIL:
\vspace{-0.25in}
\begin{equation}
\label{eq:agent-objective-5}
\textbf{(RS-GAIL)} \quad\;\; \sup_{\lambda \ge 0}\;\min_\pi\;-H(\pi) + \mathcal{L}_\lambda(\pi,\pi_E),
\end{equation}
\vspace{-0.25in}
where $\mathcal{L}_\lambda(\pi,\pi_E):=\sup_{f\in\mathcal{C}}\;(1+\lambda)\big(\rho_\alpha^\lambda[C^\pi_f] - \rho_\alpha^\lambda[C^{\pi_E}_f]\big)- \psi(f)$, with $\rho_\alpha^\lambda[C^\pi_f]:=\frac{\mathbb{E}[C^\pi_f]+\lambda\rho_\alpha[C^\pi_f]}{1+\lambda}$ being the coherent risk measure for policy $\pi$ corresponding to mean-CVaR with the risk parameter $\lambda$. The parameter $\lambda$ can be interpreted as the tradeoff between the mean performance and risk-sensitivity of the policy. The objective function $\mathcal{L}_\lambda(\pi,\pi_E)$ can be decomposed into three terms: {\bf 1)} the difference between the agent and the expert in terms of mean performance, $\mathbb{E}[C^\pi_f]-\mathbb{E}[C^{\pi_E}_f]$, which corresponds to the standard generative imitation learning objective, {\bf 2)} the difference between the agent and the expert in terms of risk $\rho_\alpha[C^\pi_f]-\rho_\alpha[C^{\pi_E}_f]$, and {\bf 3)} the convex regularizer $\psi(f)$ that encodes our belief about the expert cost function $f$.
For the risk-sensitive quantity $\rho_\alpha^\lambda[C^\pi]$, we define the distorted probability distributions $p^\pi_\xi=\xi\cdot p^\pi$, where $\xi=\frac{1+\lambda\zeta}{1+\lambda},\;\zeta\in\mathcal{U}^\pi$. We denote by $\mathcal{P}^\pi_\xi$ the set of such distorted distributions and by $\mathcal{D}^\pi_\xi$ the set of distorted occupancy measures induced by the elements of $\mathcal{P}^\pi_\xi$. Similar to CVaR in Section~\ref{subsec:RS-MDP}, we may write the risk-sensitive quantity $\rho_\alpha^\lambda[C^\pi]$ as the expectation $\rho_\alpha^\lambda[C^\pi]=\mathbb{E}_{p^\pi_{\xi^*}}[C(\tau)]=\mathbb{E}_{d^\pi_{\xi^*}}[c(s,a)]$, where $\xi^*=\frac{1+\lambda\zeta^*}{1+\lambda}$ with $\zeta^*$ defined in Section~\ref{subsec:RS-MDP} and $d^\pi_{\xi^*}\in\mathcal{D}^\pi_\xi$ is the distorted occupancy measure induced by $p^\pi_{\xi^*}\in\mathcal{P}^\pi_\xi$.
In Theorem~\ref{thm:convex-conjugate}, we show that the maximization problem $\mathcal{L}_\lambda(\pi,\pi_E)$ over the cost function $f\in\mathcal{C}$ can be rewritten as a sup-inf problem over the distorted occupancy measures $d\in\mathcal{D^\pi_\xi}$ and $d'\in\mathcal{D}^{\pi_E}_\xi$.
\begin{theorem}
\label{thm:convex-conjugate}
Let $\psi:\mathcal{C}\to\mathbb{R}\cup\{\infty\}$ be a convex cost function regularizer. Then,
\vspace{-0.25in}
\begin{align}
\label{eq:inner_problem}
\mathcal{L}_\lambda(\pi,\pi_E) &= \sup_{f \in \mathcal{C}}\;(1+\lambda)\big(\rho^\lambda_\alpha[C^\pi_f] - \rho^\lambda_\alpha[C^{\pi_E}_f]\big)-\psi(f) \nonumber \\
&= \sup_{d\in\mathcal{D}^\pi_\xi} \inf_{d'\in\mathcal{D}^{\pi_E}_\xi} \psi^*\big((1+\lambda)(d-d')\big),
\end{align}
\vspace{-0.225in}
where $\psi^*$ is the convex conjugate function of $\psi$, i.e.,~$\psi^*(d) = \sup_{f\in\mathcal{C}}d^\top f-\psi(f)$
\end{theorem}
\begin{proof}
See Appendix~\ref{app:thm1}.
\end{proof}
From Theorem~\ref{thm:convex-conjugate}, we may write the RS-GAIL optimization problem~\eqref{eq:agent-objective-5} as
\begin{align}
\label{eq:agent-objective-6}
\textbf{(RS-GAIL)} \quad &\sup_{\lambda \ge 0}\;\min_\pi\;-H(\pi) \\
&+ \sup_{d\in\mathcal{D}^\pi_\xi} \inf_{d'\in\mathcal{D}^{\pi_E}_\xi} \psi^*\big((1+\lambda)(d-d')\big). \nonumber
\end{align}
Comparing the RS-GAIL optimization problem~\eqref{eq:agent-objective-6} with that of GAIL (see Eq.~4 in~\cite{GAIL}), we notice that the main difference is the $\sup_{\mathcal{D}^\pi_\xi}\inf_{\mathcal{D}^{\pi_E}_\xi}$ in RS-GAIL that does not exist in GAIL. In the risk-neutral case, $\lambda=0$, and thus, the two sets of distorted occupancy measures $\mathcal{D}^\pi_\xi$ and $\mathcal{D}^{\pi_E}_\xi$ are singleton and the RS-GAIL optimization problem is reduced to that of GAIL.
\begin{example}
Let $\psi(f) = \begin{cases} 0 & \text{if} \;\;||f||_{\infty} \le 1 \\ +\infty & \text{otherwise}\end{cases}$, then $\mathcal{L}_{\lambda}(\pi, \pi_E) = 2(1+\lambda)\sup_{d \in \mathcal{D}^\pi_\xi} \inf_{d' \in \mathcal{D}^{\pi_E}_\xi} ||d-d^{\prime}||_{\text{TV}}$, where $||d-d'||_{\text{TV}}$ is the total variation distance between $d$ and $d'$. Note that similar to GAIL, our optimization problem aims at learning the expert's policy by matching occupancy measures. However, in order to take risk into account, it now involves matching two sets of occupancy measures (w.r.t.~the TV distance) that encode the risk profile of each policy.
\end{example}
\vspace{-0.1in}
\subsection{Risk-sensitive GAIL with Jensen-Shannon Divergence}
\vspace{-0.1in}
\label{subsec:JS-RS-GAIL}
In this section, we derive RS-GAIL using occupation measure matching via Jensen-Shannon (JS) divergence. We define the difference-of-convex cost function regularizer $\psi(f):=\begin{cases} (1+\lambda)\big(-\rho_\alpha^\lambda[C^{\pi_E}_f]+\rho_\alpha^\lambda[G_f^{\pi_E}]\big) & \text{if} \;\;f<0 \\ +\infty & \text{otherwise}\end{cases}$, where $C_f^{\pi_E}$ and $G_f^{\pi_E}$ are the loss random variables of policy $\pi_E$ w.r.t.~the cost functions $c(s,a)=f(s,a)$ and $c(s,a)=g\big(f(s,a)\big)$, respectively, with $g(x):=\begin{cases} -\log(1-e^x) & \text{if} \;\;x<0 \\ +\infty & \text{otherwise}\end{cases}$. To clarify, $G_f^{\pi_E}$ is a random variable whose instantiations are $G_f(\tau)=\sum_{t=0}^{T-1}\gamma^tg\big(f(s_t,a_t)\big)$, where $\tau\sim p^{\pi_E}$ is a trajectory generated by the expert policy $\pi_E$. Similar to the description in~\citet{GAIL}, this regularizer places low penalty on cost functions $f$ that assign negative cost to expert's state-action pairs. However, if $f$ assigns large costs (close to zero, which is the upper-bound of the regularizer) to the expert, then $\psi$ will heavily penalize $f$.
In the following theorems, whose proofs are reported in Appendix~\ref{app:JS-RS-GAIL}, we derive the optimization problem of the JS version of our RS-GAIL algorithm by computing~\eqref{eq:inner_problem} for the above choice of the cost function regularizer $\psi(f)$.
We prove the following results directly from the RS-GAIL optimization problem~\eqref{eq:agent-objective-5}.
\begin{theorem}
\label{thm:JS-RS-GAIL}
With the cost function regularizer $\psi(f)$ defined above, we may write
\begin{equation}
\label{eq:JS1}
\mathcal{L}_\lambda(\pi,\pi_E) = (1+\lambda)\sup_{f:\mathcal{S}\times\mathcal{A}\to (0,1)}\rho^\lambda_\alpha[F_{1,f}^\pi] - \rho^\lambda_\alpha[-F_{2,f}^{\pi_E}],
\end{equation}
where $F_1^\pi$ and $F_2^{\pi_E}$ are the loss random variables of policies $\pi$ and $\pi_E$ w.r.t.~the cost functions $c(s,a)=\log f(s,a)$ and $c(s,a)=\log\big(1-f(s,a)\big)$, respectively.
\end{theorem}
\begin{corollary}
\label{coro:JS-RS-GAIL}
We may write $\mathcal{L}_{\lambda}(\pi, \pi_E)$ in terms of the Jensen-Shannon (JS) divergence as
\begin{equation}
\label{eq:JS2}
\mathcal{L}_{\lambda}(\pi, \pi_E) = (1+\lambda)\sup_{d \in \mathcal{D}^\pi_\xi} \inf_{d' \in \mathcal{D}^{\pi_E}_\xi} D_{\text{JS}}(d,d').
\end{equation}
\end{corollary}
From Theorem~\ref{thm:JS-RS-GAIL}, we write the optimization problem of the JS version of our RS-GAIL algorithm as
\begin{align}
\label{eq:JS-RS-GAIL}
&\textbf{(JS-RS-GAIL)} \quad \sup_{\lambda \ge 0}\;\min_\pi\;-H(\pi) \\
&\qquad\qquad+ (1+\lambda)\sup_{f:\mathcal{S}\times\mathcal{A}\to (0,1)}\rho^\lambda_\alpha[F_{1,f}^\pi] - \rho^\lambda_\alpha[-F_{2,f}^{\pi_E}]. \nonumber
\end{align}
Hence in JS-RS-GAIL, instead of minimizing the original GAIL objective, we solve the optimization problem~\eqref{eq:JS-RS-GAIL} that aims at matching the sets $\mathcal{D}^\pi_\xi$ and $\mathcal{D}^{\pi_E}_\xi$ w.r.t.~the JS divergence.
\vspace{-0.1in}
\subsection{Risk-sensitive GAIL with Wasserstein Distance}
\vspace{-0.1in}
\label{subsec:W-RS-GAIL}
In this section, we derive RS-GAIL using occupation measure matching via the Wasserstein distance. We define the cost function regularizer $\psi(f):=\begin{cases} 0 & \text{if} \;\; f \in \mathcal{F}_1\\ + \infty & \text{otherwise}\end{cases}$.
\begin{corollary}
\label{coro:W-RS-GAIL}
For the cost function regularizer $\psi(f)$ defined above, we may write
\begin{equation}
\label{eq:W}
\mathcal{L}_{\lambda}(\pi, \pi_E) = (1+\lambda)\sup_{d \in \mathcal{D}^\pi_\xi} \inf_{d' \in \mathcal{D}^{\pi_E}_\xi} W(d,d').
\end{equation}
\end{corollary}
\begin{proof}
See Appendix~\ref{app:W-RS-GAIL}.
\end{proof}
From~\eqref{eq:inner_problem} and the cost function regularizer $\psi(f)$ defined above, we have $\mathcal{L}_\lambda(\pi,\pi_E)=\sup_{f\in\mathcal{F}_1}\;\rho_\alpha^\lambda[C^\pi_f]-\rho_\alpha^\lambda[C^{\pi_E}_f]$, which gives the following optimization problem for the Wasserstein version of our RS-GAIL algorithm:
\begin{align}
\label{eq:W-RS-GAIL}
\textbf{(W-RS-GAIL)} \quad &\sup_{\lambda \ge 0}\;\min_\pi\;-H(\pi) \\
&+ (1+\lambda)\sup_{f\in\mathcal{F}_1}\rho^\lambda_\alpha[C^\pi_f] - \rho^\lambda_\alpha[C^{\pi_E}_f]. \nonumber
\end{align}
We conclude this section with a theorem that shows if we use a risk-neutral imitation learning algorithm to minimize the Wasserstein distance between the occupancy measures of the agent and the expert, the distance between their CVaRs could be still large. Thus, new algorithms, such as those developed in this paper, are needed for risk-sensitive imitation learning.
\begin{theorem}
\label{th:worst-case risk difference}
Let $\Delta$ be the worst-case risk difference between the agent and the expert, given that their occupancy measures are $\delta$-close ($\delta>0$), i.e.,
\begin{equation*}
\Delta = \sup_{\pi,p,p_0}\;\sup_{f \in \mathcal{F}_1}\;\rho_{\alpha}[C_{f}^{\pi}] - \rho_{\alpha}[C_{f}^{\pi_E}], \;\; \text{s.t.} \;\; W(d^\pi, d^{\pi_E}) \le \delta.
\end{equation*}
Then, $\Delta \ge \frac{\delta}{\alpha}$.
\end{theorem}
Theorem~\ref{th:worst-case risk difference}, whose proof has been reported in Appendix~\ref{app:W-RS-GAIL}, indicates that the difference between the risks can be $1/\alpha$-times larger than that between the occupancy measures (in terms of Wasserstein distance).
\vspace{-0.15in}
\section{Risk-sensitive Imitation Learning Algorithms}
\label{sec:algo}
\vspace{-0.15in}
Algorithm~\ref{alg:chool1} contains the pseudocode of our JS-based and Wasserstein-based risk-sensitive imitation learning algorithms. The algorithms aim at finding a saddle-point $(\pi, f)$ of the objective function~\eqref{eq:agent-objective-5}. We use the parameterizations for the policy $\theta \mapsto \pi_\theta$ and cost function (discriminator) $w \mapsto f_w$. Similar to GAIL~\citep{GAIL}, the algorithm is TRPO-based~\citep{TRPO} and alternates between an Adam~\citep{adam} gradient ascent step for the cost function parameter $w$ and a KL-constrained gradient descent step w.r.t.~a linear approximation of the objective. The details about the algorithm, including the gradients, are reported in Appendix~\ref{sec:grad}.
\begin{algorithm}[h]
\caption{Pseudocode of JS-RS-GAIL and W-RS-GAIL Algorithms.}
\label{alg:chool1}
\begin{algorithmic}[1]
\begin{small}
\State {\bf Input:} Expert trajectories $\{\tau^E_j\}_{j=1}^{N_E}\sim p^{\pi_E}$, Risk level $\alpha \in (0,1]$, Initial policy and cost function parameters $\theta_0$ and $w_0$.
\For{$i=0,1,2,\dots$}
\State Generate $N$ trajectories using the current policy $\pi_{\theta_i}$, i.e.,~$\{\tau_j\}_{j=1}^N\sim p^{\pi_{\theta_i}}$
\State Estimate VaRs $\;\hat{\nu}_{\alpha}(F_{1, f_{w_i}}^{\pi})$ and $\hat{\nu}_{\alpha}(-F_{2, f_{w_i}}^{\pi_E})$ {\bf (JS)}
\State Estimate VaRs $\;\;\hat{\nu}_{\alpha}(C_{f_{w_i}}^{\pi})\;$ and $\;\hat{\nu}_{\alpha}(C_{f_{w_i}}^{\pi_E})\;\;\;\;$ {\bf (W)}
\State Update the discriminator parameter by computing a gradient ascent step w.r.t.~the objective
\begin{align*}
w_{i+1} &\mapsto (1+\lambda) \left( \rho_{\alpha}^{\lambda}[F_{1,f_{w_i}}^{\pi_{\theta_i}}] - \rho_{\alpha}^{\lambda}[-F_{2,f_{w_i}}^{\pi_E}] \right) \qquad \text{\bf (JS)} \\
w_{i+1} &\mapsto (1+\lambda) \left( \rho_{\alpha}^{\lambda}[C_{f_{w_i}}^{\pi_{\theta_i}}] - \rho_{\alpha}^{\lambda}[C_{f_{w_i}}^{\pi_E}] \right) \qquad\qquad \text{\bf (W)}
\end{align*}
\State Update the policy parameter using a KL-constrained gradient descent step w.r.t.~the objective
\begin{align*}
\theta_{i+1} &\mapsto -H(\pi_{\theta_i}) + (1+\lambda) \rho_{\alpha}^{\lambda}[F_{1,f_{w_{i+1}}}^{\pi_{\theta_i}}] \qquad\qquad \text{\bf (JS)} \\
\theta_{i+1} &\mapsto -H(\pi_{\theta_i}) + (1+\lambda) \rho_{\alpha}^{\lambda}[C_{f_{w_{i+1}}}^{\pi_{\theta_i}}] \qquad\qquad\;\; \text{\bf (W)}
\end{align*}
\EndFor
\end{small}
\end{algorithmic}
\end{algorithm}
In the implementation of our algorithms, we use a grid search and optimize over a finite number of the Lagrangian parameters $\lambda$. This can be seen as the agent selects among a finite number of risk profiles of the form $(\text{mean}+\lambda\text{CVaR}_\alpha)$ when she matches her risk profile to that of the expert.
\vspace{-0.15in}
\section{Related Work: Discussion about RAIL}
\label{sec:RAIL}
\vspace{-0.15in}
We start this section by comparing the RAIL optimization problem (Eq.~9 in~\cite{RAIL}) with that of our JS-RS-GAIL reported in Eq.~\ref{eq:JS-RS-GAIL}, i.e.,
\begin{align*}
&\textbf{(RAIL)} \quad \min_\pi\;-H(\pi) \\
&\qquad\qquad\quad+ (1+\lambda)\sup_{f:\mathcal{S}\times\mathcal{A}\to (0,1)}\rho^\lambda_\alpha[F_{1,f}^\pi] - \mathbb{E}[-F_{2,f}^{\pi_E}], \\
&\textbf{(JS-RS-GAIL)} \quad \min_\pi\;-H(\pi) \\
&\qquad\qquad\quad+ (1+\lambda)\sup_{f:\mathcal{S}\times\mathcal{A}\to (0,1)}\rho^\lambda_\alpha[F_{1,f}^\pi] - \rho^\lambda_\alpha[-F_{2,f}^{\pi_E}].
\end{align*}
If we write the above optimization problems in terms of the JS divergence, we obtain
\begin{align}
\label{eq:RAIL007}
&\textbf{(RAIL)} \quad \min_\pi\;-H(\pi) \\
&\qquad\quad+ (1+\lambda)\sup_{d \in \mathcal{D}^\pi_\xi} D_{\text{JS}}(d,d^{\pi_E}), \nonumber \\
\label{eq:JS007}
&\textbf{(JS-RS-GAIL)} \quad \min_\pi\;-H(\pi) \\
&\qquad\quad+ (1+\lambda)\sup_{d \in \mathcal{D}^\pi_\xi} \inf_{d' \in \mathcal{D}^{\pi_E}_\xi} D_{\text{JS}}(d,d') \quad \textit{(see Eq.~\ref{eq:JS2})}. \nonumber
\end{align}
Note that while the JS in~\eqref{eq:JS007} matches the distorted occupancy measures (risk profiles) of the agent and the expert, the JS in~\eqref{eq:RAIL007} matches the distorted occupancy measure (risk profile) of the agent with the occupancy measure (mean) of the expert. This means that RAIL does not take the expert's risk into account in its optimization.
Moreover, the results reported in~\citet{RAIL} indicate that GAIL performs poorly in terms of optimizing the risk (VaR and CVaR). By looking at the RAIL's GitHub~\citep{RAIL-GIT}, it seems they used the GAIL implementation from its GitHub~\citep{GAIL-GIT}. Although we used the same GAIL implementation, we did not observe such a poor performance for GAIL, which is not that surprising since the MuJoCo domains used in the GAIL and RAIL papers are all deterministic and the policies are the only source of randomness there. This is why in our MuJoCo experiments in Section~\ref{sec:experiments}, we inject noise to the reward function of the problems. Finally, the gradient of the objective function reported in Eq.~(A.3) of~\citet{RAIL} is a scalar, which does not seem to be correct. We corrected this in our implementation of RAIL in Section~\ref{sec:experiments}.
\vspace{-0.1in}
\section{Experiments}
\label{sec:experiments}
\vspace{-0.1in}
In this section, we evaluate the performance of our JS and Wasserstein-based algorithms and compare them with GAIL and RAIL algorithms in two MuJoCo and two OpenAI classical control tasks.
\vspace{-0.1in}
\subsection{Task Specification}
\vspace{-0.1in}
In our experiments, we use two OpenAI classical control tasks: CartPole and Pendulum~\citep{brockman2016openai}, and two MuJoCo tasks: Hopper and Walker~\citep{todorov2012mujoco}. Since these tasks are deterministic and the notion of risk-sensitive decision-making is closely related to the uncertainty in the system, we incorporate stochasticity into the original implementations of these tasks, as described below.
In the OpenAI classical control tasks, we inject stochasticity to the system by adding noise to the actions, which in turn adds noise to both the reward function and the transitions. In the MuJoCo tasks, we first learn a policy by running a RL agent with TRPO~\citep{TRPO} on the risk-neutral version of the original implementation, and then add noise to the costs as a function of the occupancy measure of the learned policy (see Appendix~\ref{app:detail-experiments} for details).
{\bf CartPole:} Our CartPole task is based on the CartPole-v1 environment in~\citet{brockman2016openai} in which at each step the agent can choose one of the two actions: either applying the force $F_x$ (action $a=1$) or the force $-F_x$ (action $a=0$). In our implementation, if the agent selects action $a=0$, it applies the force $-F_x$ w.p.~$0.8$ and the force $-K \, F_x$, where $K$ is an integer uniformly drawn from $\{0,\ldots,8\}$, w.p.~$0.2$.
{\bf Pendulum:} Our Pendulum task is based on the Pendulum-v0 environment in~\citet{brockman2016openai} in which the action space consists of $3$ different torque values $\{-2,0,2\}$ that are applied to the pendulum. In our implementation, we first extend the number of torque values to $5$, and then when the agent selects an action with the torque value $u\in\{-2,-1,0,1,2\}$, w.p.~$0.2$, the value $u$ is multiplied by $(1 + |Z|)$, where $Z\sim\mathcal{N}(0,1)$ is a standard Gaussian random variable truncated to be bounded between $-3$ and $3$.
{\bf Hopper:} Our Hopper task is based on Hopper-v1, a physics-based continuous control task simulated with MuJoCo~\citep{todorov2012mujoco}, which consists of an $11$-dimensional observation space, a $3$-dimensional action space, a deterministic reward function $r(s,a)$, and deterministic dynamics. The goal in Hopper is to make a one-legged robot hop forward as fast as possible.
{\bf Walker2d-v1:} Our Walker task is based on Walker2d-v1, a physics-based continuous control task simulated with MuJoCo~\citep{todorov2012mujoco}, which consists of a $17$-dimensional observation space, a $6$-dimensional action space, a deterministic reward function $r(s,a)$, and deterministic dynamics. The goal in Walker is to make a bipedal robot walk forward as fast as possible.
\vspace{-0.1in}
\subsection{Experimental Setup}
\vspace{-0.1in}
In the OpenAI classical control tasks, the risk-sensitive objective is set to $\rho_\alpha^\lambda=\text{Mean}+0.5\times\text{CVaR}_{0.3}$, which means that the risk-sensitive parameters have been set to $\alpha=0.3$ and $\lambda=0.5$. We set the expert's policy to that learned by the CVaR policy gradient algorithm of~\citet{Tamar15OC}, which is the standard REINFORCE algorithm~\citep{Williams92SS} adapted to the CVaR criteria. In the MuJoCo tasks, the risk-sensitive objective is set to $\rho_\alpha^\lambda=\text{Mean}+0.05\times\text{CVaR}_{0.3}$ and the expert policy is learned by TRPO on the standard (deterministic) implementations of these problems.
In our experiments, we use two different policy (gradient) optimization algorithms for the policy step of RAIL and our JS-RS-GAIL and W-RS-GAIL algorithms: {\bf 1)} the REINFORCE policy gradient algorithm in~\citet{Tamar15OC}, and {\bf 2)} the algorithm implemented in~\citet{RAIL}, which is an extension of TRPO (using KL-constrained gradient step) to risk-sensitive policy optimization. Note that the policy step of GAIL uses TRPO~\citep{TRPO}.
In the OpenAI classical control tasks, using either of these two algorithms in the policy step of RAIL and JS-RS-GAIL did not change their performance. However, in the CartPole task, we did not obtain good results by using the REINFORCE policy gradient algorithm in~\citet{Tamar15OC} for the policy step of W-RS-GAIL, and thus, we conducted the experiments with the extended TRPO. In the MuJoCo tasks, we only obtained good results with the extended TRPO for all the algorithms. We conjecture this is due to the high variance of REINFORCE gradient estimate.
We use $100$ expert trajectories to train all the algorithms, which is a higher number than that used in the experiments of the GAIL paper~\citep{GAIL} (between $1$ and $20$ trajectories). This is normal because the risk-sensitive algorithms require more samples than their risk-neutral counterparts, particularly those that optimize tail-related risk criteria, such as VaR and CVaR. Sample efficiency is one of the most important problems of the tail-related risk-sensitive optimization algorithms and has been reported in the literature (e.g.,~\cite{saferl:cvar:chow2014,Tamar15OC}), and it is mainly due to the fact that these algorithms require to learn a tail-related quantity, often VaR, for which only the trajectories whose return belongs to the tail can be used. There have been work to address this issue and to use the trajectories whose return does not belong to the tail to learn about tail-related quantities (e.g.,~\cite{Bardou09CV,Tamar15OC}), but this is still an open problem and we do not use any of these techniques in this paper.
We pre-train the risk-sensitive algorithms RAIL and JS-RS-GAIL with $100$ iterations of GAIL. As it was noted by~\citet{Tamar15OC}, pre-training risk-sensitive policy gradient algorithms with their risk-neutral counterpart is a useful technique to avoid getting stuck in local minima. In these algorithms, we use the same network architecture as in~\citet{GAIL} and~\citet{RAIL}, which consists of $2$ hidden layers with $32$ units each and tanh activation, for both the policy and discriminator networks. At each iteration, all the algorithms are given the same amount of interaction with the environment by sampling $100$ trajectories. Our algorithms and RAIL use $1$ update step for both the generator and discriminator at each iteration, while GAIL uses $3$ update steps for the generator and $1$ for the discriminator. We found these hyper-parameters by grid search for each algorithm.
We do not pre-train W-RS-GAIL, as we did not observe any improvement due to pre-training, but train it with the same number of iterations (pre-train $+$ train) as the other algorithms. We use a more complex architecture for both the policy and discriminator networks in the Wasserstein-based algorithms: W-GAIL\footnote{Note that the W-GAIL algorithm in our experiments is just the Wasserstein version of GAIL and is simpler than InfoGAIL~\citep{INFOGAIL}.} and W-RS-GAIL. This architecture consists of $3$ hidden layers with $64$, $64$, and $32$ units, tanh activation, and clipping thresholds of $-0.05$ and $0.05$.
\vspace{-0.15in}
\subsection{Experimental Results}
\vspace{-0.13in}
In this section, we compare the performance of our algorithm JS-RS-GAIL with RAIL and GAIL (Tables~\ref{tab:meanperformance}--\ref{tab:top10performanceconfidenceintervalshw}), and our algorithm W-RS-GAIL with W-GAIL (Table~\ref{tab:meanperformanceWasserstein}) in terms of their mean, $\text{VaR}_\alpha$, $\text{CVaR}_\alpha$, and more importantly $\rho_\alpha^\lambda$, which is the main target of our risk-sensitive algorithms. In all of these algorithms, we aim at minimizing the sum of the costs (the lower the better). We also report the performance of the expert and a random policy for reference.
We report the performance of the algorithms in terms of each criterion for the OpenAI control tasks in Table~\ref{tab:meanperformance}. For each task, we run each algorithm for a fixed number of iterations, $200$ for CartPole and $300$ for Pendulum (after $100$ pre-training iterations). After that we run the algorithm for another $100$ iterations and evaluate the performance of each of these $100$ policies by generating $300$ trajectories from that policy. We then average each performance criterion over the $100$ policies. We average over $100$ policies generated after our algorithms stop to show how well each algorithm converges in terms of each performance criterion. We repeat this process for $10$ random seeds and take the average. We then report the average and $95\%$ confidence interval {\em ($\text{empirical mean}\pm 1.96\times{\text{empirical standard deviation}}/{\sqrt{n=10}}$)}.
\begin{table*}[!ht]
\vspace{-0.2in}
\caption{\small{Performance of the policies learned by the algorithms for $\alpha=0.3$ and $\lambda=0.5$. Results are averaged over the last $100$ iterations and $10$ random seeds.}}
\label{tab:meanperformance}
\setlength{\tabcolsep}{2pt}
\centering
\begin{small}
\begin{tabular}{ccccccc|cccccc}
\cmidrule(r){1-13}
Criteria & Random & Expert & GAIL & RAIL & JS-RS-GAIL & & & Random & Expert & GAIL & RAIL & JS-RS-GAIL \\
\midrule
\multicolumn{7}{c}{\textbf{CartPole}} & \multicolumn{6}{c}{\textbf{Pendulum}} \\
\midrule
Mean & -12 & -333 & -296$\pm$12 & -315$\pm$3 & -319$\pm$3 & & & 1410 & 162 & 907$\pm$41 & 1150$\pm$81 & 908$\pm$89 \\
$\text{VaR}_{\alpha}$ & -3 & -301 & -151$\pm$37 & -193$\pm$19 & -231$\pm$16 & & & 1760 & 341 & 1485$\pm$44 & 1517$\pm$59 & 1409$\pm$60 \\
$\text{CVaR}_{\alpha}$ & -2 & -294 & -109$\pm$36 & -163$\pm$19 & -208$\pm$17 & & & 1812 & 401 & 1495$\pm$46 & 1527$\pm$56 & 1419$\pm$58 \\
$\rho_{\alpha}^{\lambda}$ & -13 & -479 & -350$\pm$31 & -398$\pm$11 & -425$\pm$11 & & & 2296 & 362 & 1656$\pm$63 & 1973$\pm$106 & 1616$\pm$109 \\
\bottomrule
\end{tabular}
\end{small}
\end{table*}
Table~\ref{tab:top10performanceconfidenceintervals} contains the exact same results for CartPole and Pendulum, except this time we first average each performance criterion over the top $10$ policies of the last $100$ policies (instead of averaging over the last $100$ policies). Note that the top $10$ policies are different for each performance criterion.
\begin{table*}[!h]
\vspace{-0.2in}
\caption{\small{Performance of the policies learned by the algorithms for $\alpha=0.3$ and $\lambda=0.5$. Results are averaged over the top $10$ policies of the last $100$ iterations and $10$ random seeds.}}
\label{tab:top10performanceconfidenceintervals}
\setlength{\tabcolsep}{2pt}
\centering
\begin{small}
\begin{tabular}{ccccccc|cccccc}
\cmidrule(r){1-13}
Criteria & Random & Expert & GAIL & RAIL & JS-RS-GAIL & & & Random & Expert & GAIL & RAIL & JS-RS-GAIL \\
\midrule
\multicolumn{7}{c}{\textbf{CartPole}} & \multicolumn{6}{c}{\textbf{Pendulum}} \\
\midrule
Mean & -12 & -333 & -326$\pm$3 & -319$\pm$6 & -325$\pm$4 & & & 1410 & 162 & 656$\pm$54 & 961$\pm$135 & 436$\pm$84 \\
$\text{VaR}_{\alpha}$ & -3 & -301 & -269$\pm$6 & -249$\pm$30 & -282$\pm$7 & & & 1760 & 341 & 1403$\pm$27 & 1325$\pm$74 & 1152$\pm$69 \\
$\text{CVaR}_{\alpha}$ & -2 & -294 & -258$\pm$8 & -229$\pm$32 & -278$\pm$8 & & & 1812 & 401 & 1411$\pm$26 & 1335$\pm$81 & 1175$\pm$85 \\
$\rho_{\alpha}^{\lambda}$ & -13 & -479 & -451$\pm$6 & -434$\pm$21 & -465$\pm$6 & & & 2296 & 362 & 1362$\pm$68 & 1629$\pm$188 & 1023$\pm$138 \\
\bottomrule
\end{tabular}
\end{small}
\end{table*}
The results of Tables~\ref{tab:meanperformance} and~\ref{tab:top10performanceconfidenceintervals} show that JS-RS-GAIL achieves the best performance (compared to GAIL and RAIL) in terms of the risk-sensitive criteria, in particular $\rho_{\alpha}^{\lambda}$. This advantage becomes statistically significant when we average over the top $10$ policies (see Table~\ref{tab:top10performanceconfidenceintervals}). We conjecture that if we average over more (than $10$) random seeds, we will see statistically significant advantage for JS-RS-GAIL even when we average over the last $100$ iterations. Note that in Pendulum, none of the algorithms achieve the expert's performance, but they perform better than the random policy. This shows the sign of learning and expert's performance can be achieved with more iterations and parameter tuning.
Tables~\ref{tab:meanperformancehw} and~\ref{tab:top10performanceconfidenceintervalshw} contain the exact same results as in Tables~\ref{tab:meanperformance} and~\ref{tab:top10performanceconfidenceintervals} for the MuJoCo tasks: Hopper and Walker. Similar to the OpenAI classical control problems, here JS-RS-GAIL also achieves the best performance in terms of the risk-sensitive criteria and the advantage becomes statistically significant when we average over the top $10$ policies (see Table~\ref{tab:top10performanceconfidenceintervalshw}).
\begin{table*}[!h]
\vspace{-0.2in}
\caption{\small{Performance of the policies learned by the algorithms for $\alpha=0.3$ and $\lambda=0.05$. Results are averaged over the last $100$ iterations and $10$ random seeds.}}
\label{tab:meanperformancehw}
\setlength{\tabcolsep}{2pt}
\centering
\begin{small}
\begin{tabular}{ccccccc|cccccc}
\cmidrule(r){1-13}
Criteria & Random & Expert & GAIL & RAIL & JS-RS-GAIL & & & Random & Expert & GAIL & RAIL & JS-RS-GAIL \\
\midrule
\multicolumn{7}{c}{\textbf{Hopper}} & \multicolumn{6}{c}{\textbf{Walker}} \\
\midrule
Mean & -10 & -6096 & -5428$\pm$191 & -5638$\pm$220 & -5622$\pm$198 & & & -1 & -7651 & -6542$\pm$252 & -6894$\pm$241 & -6921$\pm$230 \\
$\text{VaR}_{\alpha}$ & -5 & -6129&-5576$\pm$228 & -5621$\pm$202 & -5709$\pm$210 & & & 0 & -7875 & -6674$\pm$187 & -6605$\pm$201 & -6702$\pm$199 \\
$\text{CVaR}_{\alpha}$ & - 3 & -5590 & -4913$\pm$231 & -5141$\pm$215 & -5202$\pm$222 & & & 0 & -6440 & -5341$\pm$352 & -6012$\pm$215 & -6111$\pm$202 \\
$\rho_{\alpha}^{\lambda}$ & -10 & -6375 & -5673$\pm$202 & -5895$\pm$231 & -5882$\pm$209 & & & 1 & -7973 & -6809$\pm$269 & -7194$\pm$251 & -7226$\pm$239 \\
\bottomrule
\end{tabular}
\end{small}
\end{table*}
\begin{table*}[!h]
\vspace{-0.2in}
\caption{\small{Performance of the policies learned by the algorithms for $\alpha=0.3$ and $\lambda=0.05$. Results are averaged over the top $10$ policies of the last $100$ iterations and $10$ random seeds.}}
\label{tab:top10performanceconfidenceintervalshw}
\setlength{\tabcolsep}{2pt}
\centering
\begin{small}
\begin{tabular}{ccccccc|cccccc}
\cmidrule(r){1-13}
Criteria & Random & Expert & GAIL & RAIL & JS-RS-GAIL & & & Random & Expert & GAIL & RAIL & JS-RS-GAIL \\
\midrule
\multicolumn{7}{c}{\textbf{Hopper}} & \multicolumn{6}{c}{\textbf{Walker}} \\
\midrule
Mean & -10 & -6096 & -5743$\pm$145 & -6049$\pm$60 & -6032$\pm$51 & & & -1 & -7651 & -7221$\pm$214 & -7405$\pm$65 & -7621$\pm$63 \\
$\text{VaR}_{\alpha}$ & -5 & -6129&-6130$\pm$91 & -6268$\pm$10 & -6355$\pm$13 & & & 0 & -7875 & -7377$\pm$133 & -7535$\pm$30 & -7925$\pm$30 \\
$\text{CVaR}_{\alpha}$ & -3 & -5590 & -5361$\pm$226 & -5541$\pm$96 & -5595$\pm$83 & & & 0 & -6440 & -5590$\pm$335 & -6172$\pm$136 & -6451$\pm$129 \\
$\rho_{\alpha}^{\lambda}$ & -10 & -6375 & -6011$\pm$156 & -6340$\pm$64 & -6325$\pm$55 & & & -1 & -7973 & -7527$\pm$230 & -7714$\pm$72 & -7953$\pm$70 \\
\bottomrule
\end{tabular}
\end{small}
\end{table*}
Table~\ref{tab:meanperformanceWasserstein} shows the performance of W-RS-GAIL and compares it with that of W-GAIL. We do not compare the Wasserstein-based algorithms with the JS-based ones because they are solving different optimization problems. However, our results indicate that the JS-based algorithms have a better performance than their Wasserstein-based counterparts in terms of the relevant criteria (mean for GAIL and $\rho_{\alpha}^{\lambda}$ for RS-GAIL algorithms) in the CartPole control problem. We conjecture that the reason is the small size of the networks used in these problems. When we use Wasserstein with a small network, we end up having a very limited representation power due to clipping the weights at certain thresholds in order to maintain the Lipschitz smoothness of the network. This is why we think that the Wasserstein-based algorithms could perform better in more complex problems that require more complex networks. Verifying this conjecture requires more experiments that we leave as a future work.
\begin{table*}[!h]
\vspace{-0.2in}
\caption{\small{Performance of the policies learned by the algorithms for $\alpha=0.3$ and $\lambda=0.5$. Results are averaged over the last $100$ iterations and $10$ random seeds (W-GAIL1 and W-RS-GAIL1), as wells as over the top $10$ policies of the last $100$ iterations and $10$ random seeds (W-GAIL2 and W-RS-GAIL2).}}
\label{tab:meanperformanceWasserstein}
\setlength{\tabcolsep}{2pt}
\centering
\begin{small}
\begin{tabular}{ccccccc}
\cmidrule(r){1-7}
Criteria & Random & Expert & W-GAIL1 & W-RS-GAIL1 & W-GAIL2 & W-RS-GAIL2 \\
\midrule
\multicolumn{7}{c}{\textbf{CartPole}} \\
\midrule
Mean & -12 & -333 & -275$\pm$8 & -282$\pm$8 & -284$\pm$5 & -309$\pm$4 \\
$\text{VaR}_{\alpha}$ & -3 & -301 & -43$\pm$26 & -89$\pm$32 & -71$\pm$22 & -171$\pm$31 \\
$\text{CVaR}_{\alpha}$ & -2 & -294 & -30$\pm$14 & -59$\pm$27 & -60$\pm$17 & -149$\pm$31 \\
$\rho_{\alpha}^{\lambda}$ & -13 & -479 & -290$\pm$14 & -312$\pm$12 & -314$\pm$9 & -384$\pm$10\\
\bottomrule
\end{tabular}
\end{small}
\vspace{-0.15in}
\end{table*}
\vspace{-0.15in}
\section{Conclusions and Future Work}
\vspace{-0.15in}
\label{sec:conclu}
In this paper, we first formulated a risk-sensitive imitation learning setting in which the agent's goal is to have a risk profile as good as the expert's. We then derived a GAIL-like optimization problem for our formulation, which we termed it risk-sensitive GAIL (RS-GAIL). We proposed two risk-sensitive generative adversarial imitation learning algorithms based on two variations of RS-GAIL that match the agent's and the expert's risk profiles w.r.t.~Jensen-Shannon (JS) divergence and Wasserstein distance. We experimented with our algorithms and compared their performance with that of GAIL~\cite{GAIL} and RAIL~\cite{RAIL} in two MuJoCo and two OpenAI control tasks. Future directions include {\bf 1)} extending our results to other popular risk measures, such as expected exponential utility and the more general class of coherent risk measures, {\bf 2)} investigating other risk-sensitive imitation learning settings, especially those in which the agent can tune its risk profile w.r.t.~the expert, e.g.,~being a more risk averse/seeking version of the expert, {\bf 3)} reducing variance of the gradient estimate in extended TRPO, and {\bf 4)} more experiments, particularly with our Wasserstein-based algorithm, in more complex problems and in problems with intrinsic stochasticity.
\newpage
|
\section{Introduction}
When the goal is to understand a complex system of interacting agents that are decision makers, there are several different modeling frameworks from which to choose. Often, if the focus is on understanding and capturing each of the interactions and movement, an individual-based approach such as an Agent-Based (AB) model is used \cite{Hinkelmann2011,Holcombe,Laubenbacher2012, Packard1985}. When we are interested in global dynamics, or are more interested in how the field is evolving, density-based approaches are utilized, e.g. difference equations or systems of differential equations. For each of these frameworks, there are different pros and cons with respect to the ability to formalize and analyze a model, as well as the ease with which one can simulate the model \cite{North2014}. There are many challenges, which can arise due to noise, nonlinearities, and other spatial or temporal variations in the system \cite{Roberts2015}.
In AB models, the agents are each individually assessing their surrounding environment, potentially moving or changing state at each time increment based on a given set of rules \cite{Bonabeau7280}. The state-dependent rules could be deterministic or stochastic, and are quite often a nonlinear function based on information (e.g. other agents, states, or other environmental factors) in a locally defined interaction neighborhood \cite{Deutsch}. The movement can be on-lattice with discretely defined locations that are assigned a given probability. For example, if a lattice is a regular, two-dimensional grid, the on-lattice movement of a given agent could be either horizontal or vertical movement to an adjacent lattice node at each time increment \cite{Goldsztein, Stevens}.
Movement can also be off-lattice or continuous in space, where new locations can be determined {via specified rules or determined by solving systems of differential equations.}
Often, questions of interest concern the emergent behavior of a large number of interacting agents, which can be hard to capture at the continuous scale \cite{Holcombe}. We note that since this modeling framework is quite general, the agent could represent any feature of interest in a given system \cite{Pogson2006}, which is why these types of models are frequently used for social, biological, financial, and military applications \cite{An2017,Bonabeau7280,Chaturapruek,Cosgrove,Devitt-Lee,Goldsztein, Interian, Othmer,Stevens}. In terms of biological applications at the cellular level, AB models have been used to investigate tumor growth where the agents are the individual cells that make up the tumor \cite{Interian}, sperm cell motility where the sperm are the individual agents \cite{Burkitt,Burkitt12}, and signaling pathways within and on the membrane of cells where agents are molecules and receptors \cite{Bonabeau7280}.
It is well known that as the number of agents in a system increases, this may cause simulations investigating long term dynamics to become intractable \cite{Holcombe}. Additionally, there is generally a desire to understand how model outcomes change with respect to varying parameter values \cite{Chen}, which again would necessitate possibly thousands of simulations.
As described previously, there is not a universal or agreed upon standard to specify these models and, in many cases, the mathematical description of the rules is also not specified \cite{Hinkelmann2011}.
The focus of this current work is an introduction of a theoretical formalism and subsequent analysis of an off-lattice AB model where agents exhibit stochastic behavior when moving and changing states. {Although we focus on the case of off-lattice AB models, we note that this modeling framework has several similarities to cellular automata (CA) \cite{Deutsch, Hinkelmann2011} and dynamic network models \cite{Starnini12,Gross09}. CA models also have agents but they are generally fixed in space and state changes of an agent are generally determined based on the state of the same neighboring agents. This is in contrast to our model where agents will move and will have proximity to different agents in time, having different interaction neighborhoods. In the case of dynamic network models, agents are generally at particular nodes and interact with other agents at their node or potentially other nodes that are connected with an edge. The dynamic portion could correspond to the dynamic movement of the agents on the network or it could correspond to the creation or deletion of certain edges or nodes in time. Since the AB models we consider have different agents interacting in time, there are similarities. However, in our case, movement and interaction neighborhoods are based on the same specified spatial scale whereas interactions in a dynamic network might occur over a range of spatial scales corresponding to the range of edge lengths. }
The analysis of the off-lattice AB model will rely heavily on a precise definition of the interaction neighborhoods of agents. In contrast to other, more traditional approaches \cite{Deutsch}, we will view the interaction neighborhood as a region where an agent potentially exerts state changes to other agents. Specifically, the necessary notation for the AB model is outlined in Section \ref{notation}, which is similar to previous work on CA models \cite{Deutsch,Hinkelmann2011}.
In Section \ref{grr-basic}, we detail how to derive a Global Recurrence Rule (GRR) to determine the expected value for the number of agents in each state when assuming that an agent's state and movement are solely determined by the agent's current status. To show the applicability of this formalism, in Section \ref{eca-results}, we illustrate how a GRR can be derived for an epidemiological-AB (E-AB) model that captures the spread of an infection such as influenza. The long term behavior and steady state solutions obtained for the infected, recovered, and susceptible states using our GRR have good agreement with simulations and we are able to prove stability of fixed points. In addition, we illustrate with the E-AB how to use additional information about the dynamics to develop a more refined local approximation of the neighborhoods, with reduced error. In Section \ref{eca}, we compare the different models and emphasize which assumptions need to be satisfied in order for the GRR to be a valid approximation for the E-AB model.
\section{Initial definitions}\label{notationsec}
\subsection{AB notation}\label{notation}
We first need to create a precise definition of the properties of the agents and their interactions in order to determine the correct governing equations and hence be able to mathematically analyze the model.
We define a bounded region of interest $\Omega$ in which we track the agents.
We suppose that we have a finite collection, $\mathcal X$, of $N$ agents indexed by $1,...,k-1,k,k+1,...,N$.
At every time $t\geq 0$, each agent $k$ is assigned one and only one state $s_k^t$ and location $\mathbf x_k^t$.
We define the set containing all possible agent states as the state space $\Sigma = \{\mathcal U_1, \mathcal U_2, ... \}$. That is, for each agent $k$ and time $t$, we necessitate that $s_t^k = \mathcal U_i$ for exactly one state $\mathcal U_i \in \Sigma$.
The domain, $\Omega$, may be either a graph or network consisting of a discrete set of nodes connected by edges or it may be a continuous, bounded subset of $\mathbb{R}^M$ for some $M\in \mathbb{N}$.
If $\Omega$ is discrete, we say these agents exist on-lattice{, where they occur on nodes and travel along the graph edges}.
Otherwise, if $\Omega$ is continuous, we say these agents exist off-lattice. In either case, the agents can either be stationary or move.
{The model keeps track of individual agents performing a deterministic or random walk, governed by a model specific probability distribution over a bounded domain} \cite{Chaturapruek,Davis1990,Othmer}.
It is important to note that the scalings and distributions may be spatially, temporally, or state dependent.
An individual-based model is defined by the pair $\mathcal{A} = (f,\mathcal N)$, where we denote the collection of neighborhoods for each agent as
$\mathcal N=\left\{\mathcal N^1,\mathcal N^2,...,\mathcal N^N\right\}$
and $f$ is a local transition rule, which defines how agents change states \cite{Deutsch}. One can visualize the agents and neighborhoods in the insert of Figure \ref{fig: insert}.
If the model specifies that each neighborhood is temporally or state dependent, we define $\mathcal N_t^j$ as the neighborhood of agent $j$ in state $s_t^j$ at iteration $t$.
Since the agents move in time, these neighborhoods are time dependent.
We emphasize that the neighborhood $\mathcal N_t^j$ is the region of influence in which agent $j$ may exert state changes to other agents.
Traditional CA definitions define the neighborhood $\mathcal N_t^j$ as the region in which other agents exert state changes to agent $j$, where density analysis has been studied previously \cite{Deutsch}. However, since we want to focus on locations of moving individuals that can influence state changes to others in a region, we assert the opposite---$\mathcal N_t^j$ is the region in which agent $j$ exerts state changes to other agents. This perspective allows us greater freedom to model more realistic state and property-dependent neighborhoods.
The local transition rule is a function $f:\mathcal X \to \Sigma$. Since each agent belongs to one and only one state, the local transition rule $f$ is well-defined. The function assignment $s_{t+1}^k = f(s_t^k)$ depends conditionally on the neighborhoods in which $\mathbf x_t^k$ is contained.
Define $B_t^{\mathcal V,\mathcal U}$ as the $\mathcal V\to \mathcal U$ transition region
at time $t$. That is, $B_t^{\mathcal V,\mathcal U}$ is the region such that if, at time $t$, there is some agent $k$ such that $s_t^k = \mathcal V$ and $\mathbf x_t^k \in B_t^{\mathcal V,\mathcal U}$, then agent $k$ can transition to state $\mathcal U$ at time $t+1$. In terms of our neighborhoods of influence, we can formally define the transition region as follows.
\begin{definition}
The \underline{$\mathcal V\to \mathcal U$ transition region} at time $t$ is
$B_t^{\mathcal V,\mathcal U} = \bigcup\limits_{j\in \mathcal{C}} \mathcal N^j$,
with indexing set
$\mathcal{C} = \left\{
j \, |\, \exists k \text{ such that } s_t^k = \mathcal V \text{ and } \mathbf x_t^k \in \mathcal N_t^j \Rightarrow \prob\left(f(s_t^k)=\mathcal U\right)>0
\right\}.
$
\end{definition}
We define the transition region in this way, since, in general, agents in states other than $\mathcal U$ can cause agents to transition to state $\mathcal U$.
Our explicit definition of the transition region $B_t^{\mathcal V,\mathcal U}$ for each $\mathcal V, \mathcal U\in \Sigma$ will allow us to clearly define $f(s_t^k)$.
We are interested in finding a GRR, which calculates the expected density of agents in a state at each iteration throughout $\Omega$.
Let $\mathcal U \in \Sigma$, with $U_t$ denoting the number of agents in state $\mathcal U$ at iteration $t$ (that is, $U_t=|\{k: s_t^k=\mathcal U\}|$). We denote $W(\mathcal V \to \mathcal U)$ to be the transition probability that an agent in state $\mathcal V$ at iteration $t$ transitions to state $\mathcal U$ at time $t+1$.
We summarize our AB notation in Glossary \ref{glossary}.
\begin{glossary}[H]
\textit{
\begin{itemize}
\item $\mathcal X$: the collection of agents (indexed as $1,...,k-1,k,k+1,...,N$)
\item $\Sigma$: state space
\item $s_t^k$: the state of agent $k$ at time $t$
\item $\Omega$: the bounded region of interest
\item $\mathbf x_t^k$: the location of agent $k$ at time $t$
\item $\mathcal N_t^k$: neighborhood of agent $k$ at time $t$
\item $f:\mathcal X \to \Sigma$: the transition rule that assigns each agent at time $t$ to a\\ unique state at time $t+1$
\item $U_t$: the number of agents in state $\mathcal U$ at time $t$
\item $B_t^{\mathcal V,\mathcal U}$: the $\mathcal V \to \mathcal U$ transition region
\item $W(\mathcal V\to \mathcal U)$: the probability an agent in state $\mathcal V$ transitions to state $\mathcal U$
\end{itemize}}
\caption{Agent-Based Model Terms}
\label{glossary}
\end{glossary}
\subsection{Global recurrence rule}\label{grr-basic}
We now have the necessary notation to derive the expected density of
{agents in}
a state at any given time.
The state of an agent $k$ at $t+1$ only depends on its state ($s_t^k$) and location ($\mathbf x_t^k$) at time $t$, making this a Markovian process \cite{Markov}. Hence, the probability that agent $k$ is in state $\mathcal U$ at time $t+1$ given that the agent was in state $\mathcal V$ at time $t$ reduces to
\begin{equation}
\prob\left(s_{t+1}^{k} = \mathcal U \Big| s_t^k = \mathcal V\right) = \prob\left(\mathbf x_t^k \in B_t^{\mathcal V,\mathcal U}\right) W(\mathcal V \to \mathcal U),
\end{equation}
the product of the probability that agent $k$ is located in a $\mathcal V \to \mathcal U$ transition region with the probability that agent $k$ transitions from state $\mathcal V$ to state $\mathcal U$.
We can then find $\mathbb E(U_{t+1})$, the expected number of agents in state $\mathcal U$ at iteration $t+1$, by
\begin{align}
\mathbb E(U_{t+1})&=\mathbb E\left(\big|\{k:s_{t+1}^{k}= \mathcal U\} \big|\right)\\
&= \sum_{k=1}^N \prob\left(s_{t+1}^k = \mathcal U\right)
\label{eq:sum_cells}
\\
&= \sum_{\mathcal V \in \Sigma} \sum_{\{k:s_t^{k}= \mathcal V\}} \prob\left(s_{t+1}^k= \mathcal U \Big|s_t^k = \mathcal V\right)
\label{eq:sum_states}
\\
&= \sum_{\mathcal V \in \Sigma} \sum_{\{k:s_t^{k}= \mathcal V\}} \prob\left(\mathbf x_t^k \in B_t^{\mathcal V,\mathcal U}\right)W(\mathcal V \to \mathcal U) .
\label{GeneralGRR}
\end{align}
Note that equality between \eqref{eq:sum_cells} and \eqref{eq:sum_states} is due to the fact that we can partition the collection of agents $\mathcal X$ by the distinct states in $\Sigma$.
This leads us to the definition of the GRR.
\begin{definition}
Let $\mathcal U,\mathcal V \in \Sigma, U_t = |\{k:s_t^k=\mathcal U\}|$, and $B_t^{\mathcal V,\mathcal U}$ be the $\mathcal V\to\mathcal U$ transition region at time $t$. We define the \underline{GRR} as
\[
\mathbb E(U_{t+1}) = \sum_{\mathcal V \in \Sigma} \sum_{\{k:s_t^{k}= \mathcal V\}} \prob\left(\mathbf x_t^k \in B_t^{\mathcal V,\mathcal U}\right)W(\mathcal V \to \mathcal U). \\
\]
\label{defGRR}
\end{definition}
Thus, to find expected values of
{the number of agents in each state}
analytically, one just needs a framework to calculate both the probability of being in a transition region
{as well as }
the probability that an agent in the transition neighborhood will transition to a particular state.
\section{Application to disease dynamics}\label{eca-results}
\subsection{Phenomenological perspective}
Disease dynamics provides an interesting case study to determine the validity of the GRR. Assume there are infected individuals in a population. For simplicity, we can divide the remaining population into two classes, those who are susceptible to infection and those who were infected but cannot currently infect other individuals. We denote these classes ``susceptible'' and ``recovered,'' respectively. Further, suppose that after a finite time the recovered can become susceptible to infection again. That is, an individual in the recovered state is temporarily conferred immunity before returning to the susceptible state. This is often referred to as a Susceptible-Infected-Recovered (SIR) Epidemiological model, where simulations and analysis have been an active research area for many years \cite{Holko16,Prieto2016,Volz09}, especially with respect to endemic equilibrium sizes \cite{Ma, Miller,Volz07} and infectivity wave speed \cite{Fuentes99,Marziano}.
The modeling framework for SIR Epidemiological studies has been based on differential equations, networks, and AB models; each approach has provided some successes. CA models (on-lattice) with no movement have been studied and compared to differential equation models, extending our understanding of disease spread \cite{Fuentes99}. Extensions of CA models to real-world data in a geographic region with movement of people \cite{Holko16} gives more realistic disease spread where intervention strategies can be tested, but analysis of the model is often intractable. Challenges still exist to capture the relevant dynamics and to make time sensitive and accurate predictions with regards to disease outbreaks \cite{BALL201563,Lloyd,Ma,Marziano,Miller,Prieto2016,Roberts2015}.
In terms of the AB framework presented in Section \ref{notationsec}, it is relatively straightforward to implement an E-AB model. There are only three states: $\mathcal S$ (susceptible), $\mathcal I$ (infected), and $\mathcal R$ (recovered). In addition, the only neighborhoods of interest are those belonging to the infected agents since they will influence the state change of susceptible agents in their region of influence.
\subsection{E-AB model}
To simplify, we let the continuous domain,
$\Omega$,
of the E-AB be the
unit square.
The agents remain in the infected and recovered states for $T_i$ and $T_R$ iterations, respectively.
Thus, our state space for the E-AB is $\Sigma = \{\mathcal S,\mathcal I_1,\mathcal I_2,...,\mathcal I_{T_I},\mathcal R_1,R_2,...,R_{T_R}\}$. This dynamic is also referred to as an $SI^{T_I}R^{T_R}$ model \cite{Ma}.
We initialize $N$ agents in $\Omega$ such that $N-1$ agents are in state $\mathcal S$ ($S_0 = N-1$) and one agent is in state $\mathcal I_1$ ($I_0=1$),
where $S_t=|\{k:s_t^k=\mathcal S\}|$ and
$I_t = \cup_{j=1}^{T_I} |\{k:s_t^k= \mathcal I_j\}|$ for each time $t$.
We index the initially infected agent as $k=1$ and initialize its location in the center of the region $\Omega$.
The susceptible agents are randomly initialized following a uniform random distribution (i.e. $\mathbf x_0^k \sim Uniform(\Omega)$ for $k=2,3,\dots,N$).
Each agent\footnote{
For simplicity, every agent in this model moves according to the same rules. However, one could produce a model where each state moves differently. For example, the infected agents could move at a much smaller spatial step than the susceptible or recovered agents.}
moves by a uniform random walk inside $\Omega$.
If $\mathbf x_t^k = (x,y)$, then $\mathbf x_{t+1}^k = (x+\Delta r \cos\theta, y+\Delta r \sin\theta)$,
where $\theta \sim Uniform[0,2\pi)$ and $\Delta r \ll 1$ is the constant radial spatial step.
Reflective boundary conditions are enforced along $\partial \Omega$.
That is, if an agent hits the boundary (or is about to move outside of $\Omega$), it is shifted a distance $\Delta r$ into $\Omega$ along the direction normal to the boundary.
For our E-AB, we assume that the infectivity neighborhood of any infected agent $k$ is radially symmetric with radius $\rho_0$. That is,
\\
$\mathcal N_t^k = \left\{\mathbf y \in \Omega: ||\mathbf y-\mathbf x_t^k||_2\leq \rho_0\right\}$,
the collection of all points of a distance less than $\rho_0$ away from agent $k$, is
the area in which susceptible agents can become infected by agent $k$.
Now consider any agent $k$ such that $s_t^k = \mathcal S$. In order for agent $k$ to become infected, we require $\mathbf x_t^k$ to be in an infected neighborhood, regardless of the iteration of infectivity. We define the $\mathcal S \to \mathcal I_1$ transition region as
$B_t^{\mathcal S,\mathcal I_1} = \bigcup_{\{k: s_t^k = \mathcal I_j, \exists j\}}\mathcal N_t^k$.
Recall that the $\mathcal S$ to $\mathcal I_1$ transition region is the region in which an agent in state $\mathcal S$ can transition to state $\mathcal I_1$. The susceptible agent has the potential to become infected when in at least one neighborhood of an infected agent at any state of the infection (for $j=1,\ldots,T_I$).
In this simple E-AB model, the number of infectivity neighborhoods in which agent $k$ is located does not affect the probability of agent $k$ being infected.
The susceptible agents located inside $B_t^{\mathcal S,\mathcal I_1}$ become infected with probability $1-\kappa$, where $\kappa\in [0,1]$ is the contact tolerance.
For simplification, we will assume that $\rho_0$ and $\kappa$ are scalar constants\footnote{
Our E-AB is a toy example to demonstrate the efficacy of the GRR. For simplification, $\rho_0$ and $\kappa$ are constants. In practice, $\rho_0$ and $\kappa$ should be random variables drawn from specific probability distributions, such as the models found in \cite{Lloyd}.} and the transition rules between states are given below.
\begin{definition}The \underline{local transition rule $f:\mathcal X\to \Sigma$}, such that $s_{t+1}^k = f(s_t^k)$ are given as follows:
{\small
\begin{align}
&\text{If } s_t^k = \mathcal S:
\quad
f(s_t^k) = \begin{cases}
\mathcal I_1&: \text{ if }\mathbf x_t^k \in B_t^{\mathcal S,\mathcal I_1}\text{ and }\kappa>X,\text{ where }X\sim Uniform[0,1],\\
\mathcal S&: \text{otherwise,}
\end{cases}
\label{S_TransitionRule}
\\
&\text{If }s_t^k = \mathcal I_j, \, \text{{for some} } j=1,2,...,T_I:
\quad
f(s_k^t) =
\begin{cases}
\mathcal I_{j+1}&: \text{ if }1\leq j < T_I ,\\
\mathcal R&: \text{ if }j=T_I,
\end{cases}
\label{I_TransitionRule}
\\
&\text{If } s_t^k = \mathcal R_m, \, \text{{for some} } m=1,2,...,T_R:
\quad
f(s_t^k) =
\begin{cases}
\mathcal R_{m+1}&: \text{ if }1\leq m < T_R, \\
\mathcal S&: \text{ if }m=T_R.
\end{cases}.
\label{R_TransitionRule}
\end{align}
}
\end{definition}
Figure \ref{fig: Simulation} illustrates the off-lattice E-AB simulation as outlined above using $N=10000$ agents, where the susceptible, infected, and recovered agents are colored as black, red, and {blue}, respectively. We implemented each iteration by first determining the region of infectivity from a constant infectivity radius of $\rho_0=0.04$. We then updated the agent states according to the above transition rules \eqref{S_TransitionRule}--\eqref{R_TransitionRule} with contact tolerance $\kappa=0.95$. This ``high'' contact tolerance relates to a ``low'' probability of a susceptible agent becoming infected. Moreover, the time spent in the infective state and the time spent in the recovered state are $T_I = T_R = 30$.
Finally, we update the agent location by performing unbiased random walks with $\Delta r = 0.001$.
\begin{figure}
\centering
\subfloat[$t=10$]{ \includegraphics[width=0.38\textwidth, height=0.3\textwidth]{T10Zooming.png} \label{fig: insert}}
\qquad
\subfloat[$t=20$]{ \includegraphics[width=0.35\textwidth, height=0.3\textwidth]{T20Full.png} }\\
\subfloat[$t=30$]{ \includegraphics[width=0.35\textwidth, height=0.3\textwidth]{T30Full.png}}
\qquad
\subfloat[$t=40$]{ \includegraphics[width=0.35\textwidth, height=0.3\textwidth]{T40Full.png}}
\caption{Simulation of 10000 agents at various time steps with a single agent infected initially, which is located at (0.5,0.5).
The states are denoted as \textcolor{black}{$\blacksquare$} Susceptible, \textcolor{red}{$\blacksquare$} Infected, \textcolor{blue}{$\blacksquare$} Recovered.
The simulation parameters are defined as: contact tolerance = $\kappa=0.95$, infectivity radius = $\rho_0=0.04$, infection time = $T_I = 30$, recovered time = $T_R = 30$, and spatial step = $\Delta r = 0.001$.
As time increases, the epidemic spreads as a wave throughout the domain. The inset on (a) shows the radially symmetric neighborhoods of a few infected agents.}
\label{fig: Simulation}
\end{figure}
\subsection{Globally homogeneous GRR}\label{GlobalSection}
To reduce the number of equations, we can assume a Markovian (time-independent) process. We further simplify the number of states in our analysis by defining $\mathcal I = \bigcup_{i=1}^{T_I} \mathcal I_i$ as the infected state, which is independent of the amount of time spent in the infected state.
Similarly, we define $\mathcal R=\bigcup_{i=1}^{T_R} \mathcal R_i$, the total number of recovered agents, regardless of the amount of time spent in the recovered state.
We
{reduce the number of states in this way when calculating the GRR because we }
are primarily interested in calculating the expected total number of infected and recovered agents at each particular iteration $t$, not the particular stage of the infection or recovery.
Adapting equation (\ref{GeneralGRR}) to our E-AB model, we have the system
\begin{align}
S_{t+1} &= \sum_{\{k:s_t^k = \mathcal S\}} W(\mathcal S \to \mathcal S) + \sum_{\{k:s_t^k = \mathcal R\}} W(\mathcal R \to \mathcal S), \\
I_{t+1} &= \sum_{\{k:s_t^k = \mathcal S\}} \prob\left(\mathbf x_t^k \in B_t^{\mathcal S,\mathcal I}\right)W(\mathcal S \to \mathcal I)
+ \sum_{\{k:s_t^k = \mathcal I\}}W(\mathcal I \to \mathcal I), \label{IGRR}\\
R_{t+1} &= \sum_{\{k:s_t^k = \mathcal I\}}W(\mathcal I \to \mathcal R) + \sum_{\{k:s_t^k = \mathcal R\}} W(\mathcal R \to \mathcal R).
\label{RGRR}
\end{align}
{Recall that $W(\mathcal V\to\mathcal U)$ denotes the probability an agent in state $\mathcal V$ transitions to state $\mathcal U$.}
The total number of agents is constant, so $S_t = N - \left(I_t + R_t\right)$. This allows the reduction of the above system to just two equations, namely, \eqref{IGRR} and \eqref{RGRR}. We will now determine the expressions for the probabilities.
We further simplify the derivation by ignoring the effect of the boundary on the infectivity neighborhood, which allows the assumption that the area of the region $\mathcal N_t^k$ is independent of $k$ and $t$. Let $\mu(\mathcal N)$ denote the area of any neighborhood $\mathcal N_t^k$. Since our simulation is two-dimensional, we then make the approximation\footnote{
Agent neighborhoods near the boundary will have smaller area, since, by definition, the neighborhood is contained in $\Omega$. However, we assume that since the initially infected agent is located in the center of the region, there are sufficiently many infected agents away from the boundary to make this simplification reasonable.}
that $\mu(\mathcal N):=\mu(\mathcal N_t^k) = \pi \rho_0^2$ for all $k$ and $t$. It follows that the probability that the $k$th agent is located in the neighborhood of the $j$th agent is
\begin{equation}
\prob(\mathbf x_t^k \in \mathcal N_t^j) = \frac{\mu(\mathcal N)}{\mu(\Omega)}, \quad \forall \j,
\end{equation}
the ratio of the area of the infectivity neighborhood and the area of the region.
For any susceptible agent $k$ to transition to state $\mathcal I$, it is sufficient that $\mathbf x_t^k \in B_t^{\mathcal S,\mathcal I}$.
If we assume that the transition probability $W(\mathcal S \to \mathcal I)$ does not depend on the number of infectivity neighborhoods and that the infectivity neighborhoods are uniformly distributed within $\Omega$, then by the multiplication rule of independent events,
\begin{equation}
\prob(\mathbf x_t^k \notin B_t^{\mathcal S,\mathcal I}) = \left( 1 - \frac{\mu(\mathcal N)}{\mu(\Omega)}\right)^{I_t}.
\end{equation}
It follows that the probability of $\mathbf x_t^k$ being located in an $\mathcal S \to \mathcal I$ transition neighborhood is then
\begin{equation}
\prob(\mathbf x_t^k \in B_t^{\mathcal S,\mathcal I}) = 1 - \left( 1 - \frac{\mu(\mathcal N)}{\mu(\Omega)}\right)^{I_t}.
\label{GlobalS}
\end{equation}
Since $W(\mathcal S \to \mathcal I)$ depends on agent $k$ being located in at least one infectivity neighborhood, and does not change if agent $k$ is located in more than one infectivity neighborhood, it follows that $W(\mathcal S \to \mathcal I) = 1-\kappa$, where $\kappa\in[0,1]$ is the contact rate.
Then, for any $k$ such that $s_t^k = \mathcal S$, by inserting \eqref{GlobalS} into \eqref{IGRR} we have:
\begin{equation}
\begin{split}
\prob(s_{t+1}^k = \mathcal I) &= \prob(\mathbf x_t^k \in B_t^{\mathcal S,\mathcal I})W(\mathcal S \to \mathcal I) \\
&=
\underbrace{ \left\{ 1 - \left( 1 - \frac{\mu(\mathcal N)}{\mu(\Omega)}\right)^{I_t} \right\}}_{\substack{\text{probability in at least}\\ \text{one infectivity region}}}
\underbrace{ (1-\kappa)}_{\substack{\text{probability}\\ \text{becoming}\\ \text{infected}}}.
\label{GlobalSEq}
\end{split}
\end{equation}
Moreover, by assuming the cumulative time spent in infected states is uniformly distributed, we have for any agent $k$ such that $s_t^k = \mathcal I$,
\begin{align}
W(\mathcal I \to \mathcal R) &= 1/T_I, \label{GlobalR}\\
W(\mathcal I \to \mathcal I)&=1-1/T_I,
\label{GlobalI}
\end{align}
where $T_I$ is the time spent in the infective state.
This assumption is valid for a large number of agents and for a sufficiently large number of iterations.
Inserting equations (\ref{GlobalSEq}), (\ref{GlobalR}), and (\ref{GlobalI}) into (\ref{IGRR}), we have
\small{
\begin{equation}
\underbrace{I_{t+1}}_{{\substack{\text{total}\\ \text{infected}\\ \text{agents} \\ \text{at time }t+1}}} =
\underbrace{(N-I_t-R_t)}_{\substack{\text{total}\\ \text{susceptible}\\ \text{agents} \\ \text{at time }t}}
\underbrace{\left\{ 1 - \left( 1 - \frac{\mu(\mathcal N)}{\mu(\Omega)}\right)^{I_t} \right\}}_{ \substack{\text{probability in at least}\\ \text{one infectivity region}}}
\underbrace{(1-\kappa)}_{\substack{\text{probability}\\ \text{becoming}\\ \text{infected}}}
+
\underbrace{\left\{1-\frac{1}{T_I}\right\}}_{\substack{
\text{probability remain} \\
\text{in infected state}}}
\underbrace{I_t}_{\substack{\text{total}\\ \text{infected}\\ \text{agents} \\ \text{at time }t}}.
\end{equation}
}
That is, the expected number of infected agents at $t+1$ is the sum of two terms. The first term is the product of the expected number of susceptible agents at time $t$ multiplied by the probability a susceptible agent transitions to state $\mathcal I$. The second term is the expected number of infected agents at $t$ times the probability that an infected agent remains in state $\mathcal I$.
Similarly, by assuming the time in recovered states is uniformly distributed, we have for any agent $k$ such that $s_t^k = \mathcal R$, the probability of staying in state $\mathcal R$ is
\begin{equation}
W(\mathcal R \to \mathcal R) = 1 - 1/T_R.
\label{GlobalRR}
\end{equation}
Then, inserting (\ref{GlobalR}) and (\ref{GlobalRR}) into (\ref{RGRR}) we have
\begin{equation}
R_{t+1} = \frac{1}{T_I}I_t + \left(1 - \frac{1}{T_R}\right) R_t,
\end{equation}
and the expected number of recovered agents at iteration $t+1$ is the sum of two terms. The first term is the expected number of infected agents at time $t$ times the probability an infected agent transitions to state $\mathcal R$. The second term is the expected number of recovered agents at $t$ times the probability a recovered agent remains in state $\mathcal R$.
Since $T_I$ and $T_R$ will be explicitly defined, we can easily find a $q\in \mathbb{R}$ such that $T_R=qT_I$. We then have the following E-AB GRR:
\begin{align}
I_{t+1} &= (N-I_t-R_t)\left\{ 1 - \left( 1 - \frac{\mu(\mathcal N)}{\mu(\Omega)}\right)^{I_t} \right\}(1-\kappa)
+ \left(1-\frac{1}{T_I}\right)I_t :=H(I_t,R_t),
\label{Global GRRI} \\
R_{t+1} &= \frac{1}{T_I}I_t + \left(1 - \frac{1}{qT_I}\right) R_t:=G(I_t,R_t).
\label{Global GRRR}
\end{align}
Since $S_t = N - \left(I_t + R_t\right)$, we have recurrence formulas for the expected agent densities in each state at each iteration.
With our GRR, we now have a general framework to further analyze the behavior of the system. Note that we identify \eqref{Global GRRI}-\eqref{Global GRRR} as globally homogeneous since we have assumed the infectivity neighborhoods are uniformly distributed in the domain with the same constant area.
\subsection{Fixed point analysis for globally homogeneous GRR}\label{FixedPointSIRSection}
We can now calculate the stability of the fixed points of the globally homogeneous GRR by finding all solutions to the system that simultaneously solve $I_{t+1}-I_t=0$ and $R_{t+1} - R_t=0$.
That is, we need to find all solutions to the system
\begin{equation}
\small
\begin{bmatrix}
(N-I_t-R_t)\left\{ 1-\left(1-\mu(\mathcal N)\right)^{I_t}\right\}(1-\kappa)-\frac{1}{T_I}I_t \\
\frac{1}{T_I}I_t - \frac{1}{qT_I}R_t
\end{bmatrix}
=
\begin{bmatrix} 0 \\ 0 \end{bmatrix}.
\label{SIRFixedSetup}
\end{equation}
We have two fixed points. One is the trivial fixed point, $(I,R)=(0,0)$. The other is the point along the line $R=qI$ that solves the fixed point problem
\begin{equation}
(N-(1+q)I)\left\{ 1-\left(1-\mu(\mathcal N)\right)^{I}\right\}(1-\kappa)-\frac{1}{T_I}I = I.
\end{equation}
This second fixed point has to be computed numerically.
We will analyze the fixed point $(0,0)$ using two-dimensional perturbation theory where details can be found in
\cite{Deutsch,LinSegel}.
The Jacobian matrix of the E-AB GRR is
\begin{align*}
J= {\tiny \begin{bmatrix}
-(1-\kappa)\Big\{(N-I-R)\left( (1-\mu(\mathcal N))^I \ln(1-\mu(\mathcal N))\right) + (1-(1-\mu(\mathcal N))^I)\Big\}-\frac{1}{T_I} & -(1-\kappa) (1-(1-\mu(\mathcal N))^I) \\
\frac{1}{T_I} & 1-\frac{1}{qT_I}
\end{bmatrix} }.
\end{align*}
Evaluating at $(0,0)$ gives us
\begin{equation}
J\Big|_{(I,R)=(0,0)} = \begin{bmatrix}
-N(1-\kappa)\ln\left(1-\mu(\mathcal N)\right)- \frac{1}{T_I} & 0 \\
\frac{1}{T_I} & 1-\frac{1}{qT_I}
\end{bmatrix}.
\end{equation}
The eigenvalues are $\lambda_1 = 1-\frac{1}{qT_I}$ and $\lambda_2=-N(1-\kappa)\ln(1-\mu(\mathcal N))-\frac{1}{T_I}$.
Clearly, since $qT_I = T_R>0$ we know $|\lambda_1|<1$.
Now, suppose $\lambda_2<1$. It follows that $\alpha > 1 - \exp\left(\frac{1+1/T_I}{N(1-\kappa)}\right)$.
That is, $\frac{\mu(\mathcal N)}{\mu(\Omega)} > 1 - \exp\left(\frac{1+1/T_I}{N(1-\kappa)}\right)$.
We know $\mu(\mathcal N)\ll \mu(\Omega)$ and it is reasonable to assume that $N$ is sufficiently large such that $N(1-\kappa)>2$. This contradicts the inequality. It must follow that $\lambda_2>1$. Thus, we have that $(0,0)$ is a saddle point that is only stable along the nullcline $I=0$.
Since we do not have an explicit solution of the second fixed point, we cannot perform the same computation as we did for $(0,0)$. However, we know that the following are bounded: $H$ and all derivatives of $H$,
{the expected number of infected agents at the next time step from equation \eqref{Global GRRI}}, and the domain. Additionally, since $I$ is repelled by $(0,0)$, we can infer that the second fixed point is stable.
Similar to a differential equation SIR model, we are able to obtain fixed points. However, the fixed points are different and there are not direct comparisons since we assume a moving population with dynamic contacts that allow for infection whereas a differential equation assumes a well-mixed population.
\subsection{Locally homogeneous GRR}\label{LocalSection}
When deriving the globally homogeneous E-AB GRR, (\ref{Global GRRI}) and (\ref{Global GRRR}), we assumed that the infectivity neighborhoods were uniformly distributed throughout $\Omega$. For this test case, we initialize one infected agent, $s_0^1 = \mathcal I$, such that it is initially located in the center of the region $\mathbf x_0^1 = (0.5,0.5)$. However, from observation of simulations, such as in Figure \ref{fig: Simulation} (or intuition), we know there is a wave of infectivity propagating from this initial infected agent. The susceptible agents that agent 1 infects must be located in its neighborhood $\mathcal N^{1}$. Future infected agents will be located in those neighborhoods. So rather than generalize a uniform distribution of infected agents, we should modify the GRR to account for the infection wave front.
We then need to create a sequence of regions
$\left\{\tilde B_0^{\mathcal S,\mathcal I}, \tilde B_1^{\mathcal S,\mathcal I}, ...\right\}$
, where $\tilde B_0^{\mathcal S,\mathcal I} = \mathcal N^1$ and $\tilde B_t^{\mathcal S,\mathcal I}$ is the smallest connected region containing the infection front at time $t$.
For notation, we will use tildes above variables to denote variables and functions specifically defined for the locally homogeneous case.
\begin{definition}
$\displaystyle \tilde B_t^{\mathcal S,\mathcal I} = \inf_A \left\{A\subset \Omega: A \text{ is connected and } \cup_{\{k:s_t^k=\mathcal I\}} \mathcal N_t^k \subset A\right\}$.
\end{definition}
Suppose agent $k$ is such that $s_t^k = \mathcal S$ at iteration $t$. We have the following conditional probability that an agent is located in the $\mathcal S\to \mathcal I$ transition neighborhood, $B_t^{\mathcal S,\mathcal I}$, given that the agent is within the infection front, $\tilde B_t^{\mathcal S,\mathcal I}$:
\begin{equation}
\prob\left(\mathbf x_t^k \in
B_t^{\mathcal S,\mathcal I} \Big| \mathbf x_t^k \in \tilde B_t^{\mathcal S,\mathcal I}\right)
= 1-\left(1 - \frac{\mu(\mathcal N)}{\mu\left(\tilde B_t^{\mathcal S,\mathcal I}\right)}\right)^{I_t}.
\label{modify1}
\end{equation}
In general, for regions $\mathcal N$ and $\tilde B_t^{\mathcal S,\mathcal I}$, the probability is given as
$\prob\left(\mathbf x_t^k \in \tilde B_t^{\mathcal S,\mathcal I}\right) = \frac{\mu\left(\tilde B_t^{\mathcal S,\mathcal I}\right)}{\mu(\Omega)}.$
Using Bayes' theorem \cite{Talbott}, we have that the locally homogeneous probability of an agent being in the $\mathcal S\to \mathcal I$ transition neighborhood is
\begin{equation}
\begin{split}
\prob\left(\mathbf x_t^k \in B_t^{\mathcal S,\mathcal I}\right)
&= \prob\left(\mathbf x_t^k \in B_t^{\mathcal S,\mathcal I} \Big| \mathbf x_t^k \in \tilde B_t^{\mathcal S,\mathcal I}\right) \prob\left(\mathbf x_t^k \in \tilde B_t^{\mathcal S,\mathcal I}\right), \\
&= \left\{1-\left(1 - \frac{\mu(\mathcal N)}{\mu\left(\tilde B_t^{\mathcal S,\mathcal I}\right)}\right)^{I_t}\right\}\frac{\mu\left(\tilde B_t^{\mathcal S,\mathcal I}\right)}{\mu(\Omega)}.
\end{split}
\label{modify3}
\end{equation}
Inserting \eqref{modify3} into \eqref{GlobalSEq}, our locally homogeneous E-AB GRR is
\begin{equation}
\begin{split}
\tilde I_{t+1} &= \left(N-\tilde I_t - \tilde R_t\right)
\left\{1-\left(1 - \frac{\mu(\mathcal N)}{\mu\left(\tilde B_t^{\mathcal S,\mathcal I}\right)}\right)^{I_t}\right\}\frac{\mu\left(\tilde B_t^{\mathcal S,\mathcal I}\right)}{\mu(\Omega)}(1-\kappa) \\
&\qquad \qquad +\left(1-\frac{1}{T_I}\right)\tilde I_t =:\tilde H(\tilde I_t, \tilde R_t),\\
\tilde R_{t+1} &= \frac{1}{T_I}\tilde I_t + \left(1 - \frac{1}{qT_I}\right) \tilde R_t =: \tilde G(\tilde I_t, \tilde R_t).
\end{split}
\label{LocalGRRR}
\end{equation}
Recall that the tilde denotes values associated with the locally homogeneous GRR.
We derived this GRR by focusing on early dynamics. But how does the non-uniform assumption of the infection front affect the stability using the locally homogeneous GRR compared with the globally homogeneous GRR?
\begin{theorem}
If $\tilde B_t^{\mathcal S,\mathcal I} \to \Omega$ as $t\to +\infty$ and $\alpha$ is a fixed point of $H$, then $\alpha$ is a fixed point of $\tilde H$. Moreover, $\alpha$ has the same stability conditions for $H$ and $\tilde H$.
\label{LocalSIRTHM}
\end{theorem}
\begin{proof}
Suppose that $\lim_{t\to +\infty}I_t = \alpha$ and suppose that $\lim_{t\to +\infty}\tilde I_t$ exists.
Then, since $\mu\left(\tilde B_t^{\mathcal S,\mathcal I}\right) \to \mu(\Omega)$ as $t \to + \infty$, we have that
$\lim_{t\to +\infty}\left(1-\frac{\mu(\mathcal N)}{\mu\left(\tilde B_t^{\mathcal S,\mathcal I}\right)}\right)^{\tilde I_t} =
\lim_{t\to +\infty}\left(1-\frac{\mu(\mathcal N)}{\mu(\Omega)}\right)^{\tilde I_t}$.
Plugging into \eqref{Global GRRI} and taking the limit,
\begin{small}
\begin{align*}
\lim_{t\to +\infty}\tilde H(\tilde I_t)
&=\lim_{t\to +\infty}\left[ (N-\tilde I_t)\left(1 - \frac{\mu(\mathcal N)}{\mu(\Omega)}\right)^{\tilde I_t}(1-\kappa)+\left(1 - \frac{1}{qT_I}\right)\tilde I_t \right]\\
&=\lim_{t\to +\infty} H(\tilde I_t) = \alpha.
\end{align*}
\end{small}
Moreover, $\mu(\tilde B_{\mathcal S,\mathcal I}^t)\to \mu(\Omega)$ for fixed $\alpha$ and
it is clear that
$\frac{\partial \tilde H}{\partial \tilde I}\Big|_\alpha \to \frac{\partial H}{\partial I}\Big|_\alpha$,
$\frac{\partial \tilde H}{\partial \tilde R}\Big|_\alpha \to \frac{\partial H}{\partial R}\Big|_\alpha$,
$\frac{\partial \tilde G}{\partial \tilde I}\Big|_\alpha \to \frac{\partial G}{\partial I}\Big|_\alpha$, and
$\frac{\partial \tilde G}{\partial \tilde R}\Big|_\alpha \to \frac{\partial G}{\partial R}\Big|_\alpha$.
Since the stability condition depends on the Jacobian, and the Jacobian of the locally homogeneous region approaches the Jacobian of the globally homogeneous region as $t\to +\infty$, the long term stability conditions must be the same.
\end{proof}
From Theorem \ref{LocalSIRTHM} we know that $(\tilde H, \tilde G)$ has the same fixed points as found in Section \ref{FixedPointSIRSection} with the same stability conditions. We thus reduced the problem to capturing an explicit formula for $\mu\left( \tilde B_t^{\mathcal S,\mathcal I} \right)$.
Before, we made the simplifying assumption that the infected agents were distributed uniformly throughout the region, so we did not need to incorporate any spatial characteristics into the globally homogeneous GRR. Now, we need to capture the infection front dynamics in order to explicitly calculate the area of the infectivity region, $\mu\left( \tilde B_t^{\mathcal S,\mathcal I} \right)$, in the locally homogeneous case.
\begin{figure}[htb!]
\centering
\includegraphics[width=5cm]{InfectivityRadius}
\caption{Diagram demonstrating how the locally homogeneous $\tilde B_{t+1}^{\mathcal S,\mathcal I}$ region depends on $\tilde B_t^{\mathcal S,\mathcal I}$ and $\tilde B_{t-1}^{\mathcal S,\mathcal I}$ regions. The initially infected agent at $t=0$ is located in the center of the $\tilde B_t^{\mathcal S,\mathcal I}$ regions, which expand outward with radii $\zeta_t$. We assume newly infected agents lie on the radial center of mass of the region $\tilde B_t^{\mathcal S,\mathcal I}\setminus \tilde B_{t-1}^{\mathcal S,\mathcal I}$, denoted with the dashed line, which is a distance $r$ from the initially infected agent.}
\label{B_kImage}%
\end{figure}
For our formula, we will assume that newly infected agents are expected to be in the region $\tilde B_t^{\mathcal S,\mathcal I}\setminus \tilde B_{t-1}^{\mathcal S,\mathcal I}$ and are moving a fixed distance $\Delta r$. Further, we assume the expected location of the newly infected agents will lie on the circle that is the radial center of mass of $\tilde B_t^{\mathcal S,\mathcal I}\setminus \tilde B_{t-1}^{\mathcal S,\mathcal I}$, a distance $r$ from the initially infected agent location as shown in Figure \ref{B_kImage}. The radius of the infectivity front region $\tilde B_t^{\mathcal S,\mathcal I}$ at time $t$ is denoted as $\zeta_t$. We then have the following expected radius of $\tilde B_{t+1}^{\mathcal S,\mathcal I}$:
\begin{equation}
\zeta_{t+1} = \rho_0 + \sqrt{
\frac{(\zeta_t + \delta_{out}(\Delta r))^2 + (\zeta_{t-1}-\delta_{in}(\Delta r))^2}{2}},
\label{Bkeq}
\end{equation}
where $\delta_{in}$ and $\delta_{out}$ are functions of the expected distance an infected agent travels towards the center of $\tilde B_t^{\mathcal S,\mathcal I}$ and out of the region $\tilde B_t^{\mathcal S,\mathcal I}$, respectively.
For our simulations and derivation, we assume
$\delta_{out}=\Delta r$ and $\delta_{in}=\Delta r$.
Even though the simulation is a Markov process, our analytical solution, which calculates the area $\mu\left( \tilde{B}_{t+1}^{\mathcal S,\mathcal I}\right)$ using $\zeta_{t+1}$, is not, since it relies on information at iterations $t$ and $t-1$. Clearly, the simulation is a Markovian process but the GRR does not have to be for this analysis. In fact, it can belong to a larger class of processes than the underlying AB model.
By the following theorem, we know that equation (\ref{Bkeq}) satisfies the premise of Theorem \ref{LocalSIRTHM}.
\begin{theorem}
If $\rho_0>0$ and for all $j$ such that $\mu\left( \tilde B_j^{\mathcal S,\mathcal I}\right) \neq \mu(\Omega)$ there exists an agent $k$ with $s_j^k = \mathcal S$ such that $\mathbf x_j^k \notin \tilde B_j^{\mathcal S,\mathcal I}$, then $\exists \hat t\in \mathbb{N}$ such that $\mu\left(\tilde B_t^{\mathcal S,\mathcal I}\right) = \mu(\Omega)$ for all $t>\hat{t}$ with radius $\zeta_{t}$ as defined in (\ref{Bkeq}).
\label{Bkthm}
\end{theorem}
The above theorem essentially states that if the infection does not ``die out,'' then the $\mathcal S \to \mathcal I$ transition neighborhood, $\tilde{B}_j^{\mathcal S,\mathcal I}$, eventually covers the entire region of interest $\Omega$.
The proof is clear, since $\Omega$ is bounded.
As we will see in Section \ref{Sec: Numerical}, we are able to approximate early behavior more accurately with the locally homogeneous GRR, while still being able to evaluate and determine the stability of fixed points with the simpler equations of the globally homogeneous GRR.
\subsection{Extensions to more complex neighborhoods}\label{inclusion}
\begin{figure
\begin{center}
\includegraphics[width=0.35\textwidth]{SparseOverview}
\end{center}
\caption{Infectivity neighborhood depends on the number $n$ of newly infected agents. The radial center of mass of the region $\tilde B_t^{\mathcal S,\mathcal I} \setminus \tilde B_{t-1}^{\mathcal S,\mathcal I}$ (dashed line) is a distance $r$ from the initially infected agent. Newly infected agents are distributed uniformly along the radial center of mass and the new infection front radius $\zeta_{t+1}$ will depend on the total area of the infectivity neighborhoods outside $\tilde B_t^{\mathcal S,\mathcal I}$.}
\label{InfectionOverviewInPaper}
\end{figure}
As long as the infection front in the E-AB approaches $\partial \Omega$ and $\tilde B_t^{\mathcal S,\mathcal I} \to \partial \Omega$, Theorem \ref{LocalSIRTHM} holds. One could derive a formula for $\tilde B_t^{\mathcal S,\mathcal I}$ that more closely approximates the initial phases of the infection spread.
Rather than assume that the new infection front extends approximately $\Delta r$ from the mean center of mass as in Figure \ref{B_kImage}, we can assume that the infectivity radius expansion depends on the number of newly infected agents in $\tilde B_t^{\mathcal S,\mathcal I} \setminus \tilde B_{t-1}^{\mathcal S,\mathcal I}$, as shown in Figure \ref{InfectionOverviewInPaper}.
Further details regarding the calculation and derivation for this case of $\tilde B_t^{\mathcal S,\mathcal I}$ can be found in Appendix \ref{Section:Sparse}. In this example, we illustrate that, by relaxing assumptions, one can derive other expressions calculating the area of the infectivity front radius $\zeta_t$ that may decrease the error of the locally homogeneous GRR during the early stages of the epidemic. Moreover, we know that as long as the new formulation of $\zeta_t$ maintains the suppositions of Theorem \ref{LocalSIRTHM}, the long term dynamics will be captured.
\section{Numerical results}\label{eca}
\label{Sec: Numerical}
Results for the E-AB model with $N=10000$ initialized agents are shown in Figure
\ref{SIRCA_Simulations}.
Note that each simulation curve on the plot is the average of 1000 E-AB simulations whereas the curves based on the Globally
\begin{figure}
\centering
\subfloat[$N=10000, \kappa=0.6, T_I=T_R=30$]{{
\includegraphics[width=0.325\textwidth]{SIRSCA_K6r2}
\includegraphics[width=0.3\textwidth]{SIRSCA_K6r4}
\includegraphics[width=0.3\textwidth]{SIRSCA_K6r8}
}} \\
\subfloat[$N=10000, \kappa=0.8, T_I=T_R=30$]{{
\includegraphics[width=0.325\textwidth]{SIRSCA_K8r2}
\includegraphics[width=0.3\textwidth]{SIRSCA_K8r4}
\includegraphics[width=0.3\textwidth]{SIRSCA_K8r8}
}} \\
\subfloat[$N=10000, \kappa=0.8, T_I=30, T_R=45$]{{
\includegraphics[width=0.325\textwidth]{SIRSCA_K8r2v2}
\includegraphics[width=0.3\textwidth]{SIRSCA_K8r4v2}
\includegraphics[width=0.3\textwidth]{SIRSCA_K8r8v2}
}}
\caption{Comparing the average value of 1000 E-AB simulations with results from the globally and locally homogeneous GRRs calculated from \eqref{Global GRRR} and \eqref{LocalGRRR}, respectively. The time to remain infected is set to $T_I=30$, while the recovery time is $T_R=30$ in (a)-(b) and $T_R=45$ in (c). Each plot corresponds to a different infectivity radii parameter $\rho_0$ with contact tolerance $\kappa=0.6$ in (a) and $\kappa=0.8$ in (b)-(c). Note that for the globally homogeneous case (labeled as GH GRR), the infectivity neighborhood has fixed radius of $\rho_0$, whereas the locally homogeneous case (labeled as LH GRR) has a variable radii $\zeta_t$ at each iteration $t$ as given in \eqref{Bkeq}, which is a function of $\rho_0$.}
\label{SIRCA_Simulations}
\end{figure}
and locally homogeneous GRRs are from solving \eqref{Global GRRR} and \eqref{LocalGRRR}, respectively.
In the Figure, we observe agreement of long term behavior of the simulations with both the globally and locally homogeneous GRRs. For example, in Figure \ref{SIRCA_Simulations}, the left hand column of each row corresponds to the case where $\rho_0 = 0.02$. In the left column of (c), the average of the E-AB simulations for the fixed points or long term behavior is 3877.1 infected agents and 5790.9 recovered agents. Upon calculation, the relative error between the simulated fixed points and the GRR fixed points is $\mathcal{O}\left(10^{-4}\right)$. Further, the early time dynamics of the infected and recovered populations with the GRR estimates have behavior similar to that of the E-AB simulations.
However, the early infection dynamics of the GRRs do not exactly match the simulations for all cases. In Figure \ref{SIRCA_Simulations}(a)-(b), for a contact tolerance $\kappa=0.6$ and $\kappa=0.8$ (characterizing how easily an agent becomes infected), we observe that as the infectivity radius increases, the GRRs are able to more accurately capture the early time dynamics of the E-AB simulations. For an infectivity neighborhood of radius $\rho_0=0.02$, it is likely that there is not a sufficient number of agents in the region to accurately capture the early time dynamics of infectivity. We do observe that the locally homogeneous GRR provides a better approximation to the E-AB simulations in comparison to the globally homogeneous GRR. Similar trends are observed in Figure \ref{SIRCA_Simulations}(c), where the recovery time $T_R$ is increased.
To explicitly define how much ``better'' the locally homogeneous GRR is relative to the globally homogeneous GRR at capturing the E-AB dynamics for a particular parameter set, we need to develop a metric.
We have a sequence of points, $(1,U_1),(2,U_2),...,(M,U_M)$, from the simulation, where $U_t$, as previously defined, is the number of agents in state $\mathcal U$ at iteration $t$ for $t=1,\ldots,M$.
By linear spline interpolation of these points we will construct a function $g(t)$.
We also have a sequence of points, $(1, \hat{U}_1),(2, \hat{U}_2),\dots,(M,\hat{U}_M)$, from the GRR.
Given the fact that some of the error is due to translation and that the number of agents is much larger than the number of iterations, we need to normalize the data.
We scale the $t$-values so that $t_i\leftarrow t_i/M$ and $\hat{t}_i \leftarrow \hat{t}_i/M$.
Additionally, we let $\gamma = \max\{U_1,U_2,...,U_M\}$ and scale the $U$-values so that $U_i\leftarrow U_i/\gamma$ and $\hat{U}_i \leftarrow \hat{U}_i/\gamma$.
Our error metric $\nu$ in Equation \eqref{metric2} is a normalized least square, evaluating the average distance of each scaled GRR estimation to the scaled E-AB simulated curve. Details of the derivation for the error metric can be found in Appendix \ref{MetricDeriv}.
\begin{table}[H]
\begin{center}
\caption{Error between GRR and 1000 E-AB simulations using metric $\nu$ from Equation \eqref{metric2} for infectivity radii $\rho_0$, contact tolerances $\kappa$, time in infected state $T_I$, and time in recovered state $T_R$.}
{\tiny
\subfloat[Error for number of infected agents.]{
\begin{tabular}{|c||c|c|c|c||c|c|c|c|} \hline
\multirow{2}{*}{$N=10000$}
& \multicolumn{4}{c||}{$T_I=T_R=30$}
& \multicolumn{4}{c|}{$T_I=30$, $T_R=45$} \\
\cline{2-9}
& \multicolumn{2}{c|}{$\kappa=0.6$}
& \multicolumn{2}{c||}{$\kappa=0.8$}
& \multicolumn{2}{c|}{$\kappa=0.6$}
& \multicolumn{2}{c|}{$\kappa=0.8$} \\
\cline{2-9}
& Global & Local & Global & Local & Global & Local & Global & Local \\
\hline
$\rho_0=0.02$ & 0.036606 & 0.009573 & 0.052660 & 0.020494 &
0.035438 & 0.010552 & 0.059061 & 0.021560 \\
\hline
$\rho_0=0.04$ & 0.008092 & 0.001398 & 0.007514 & 0.001237 &
0.008595 & 0.001687 & 0.008653 & 0.001444 \\
\hline
$\rho_0=0.08$ & 0.004019 & 0.000652 & 0.003252 & 0.000360 &
0.004686 & 0.000692 & 0.003657 & 0.000391 \\
\hline
$\rho_0=0.16$ & 0.002030 & 0.000798 & 0.001589 & 0.000545 &
0.002352 & 0.000859 & 0.001827 & 0.000608 \\
\hline
\end{tabular} \label{dataTableI}} \\
\subfloat[Error for number of recovered agents.]{
\begin{tabular}{|c||c|c|c|c||c|c|c|c|}
\hline
\multirow{2}{*}{$N=10000$}
& \multicolumn{4}{c||}{$T_I=T_R=30$}
& \multicolumn{4}{c|}{$T_I=30$, $T_R=45$} \\
\cline{2-9}
& \multicolumn{2}{c|}{$\kappa=0.6$}
& \multicolumn{2}{c||}{$\kappa=0.8$}
& \multicolumn{2}{c|}{$\kappa=0.6$}
& \multicolumn{2}{c|}{$\kappa=0.8$} \\
\cline{2-9}
& Global & Local & Global & Local & Global & Local & Global & Local \\
\hline
$\rho_0=0.02$ & 0.022055 & 0.009537 & 0.027853 & 0.015889 &
0.028273 & 0.012485 & 0.036781 & 0.021142 \\
\hline
$\rho_0=0.04$ & 0.008436 & 0.000831 & 0.008046 & 0.000817 &
0.010412 & 0.000895 & 0.010120 & 0.000937 \\
\hline
$\rho_0=0.08$ & 0.003581 & 0.000126 & 0.003084 & 0.000104 &
0.004247 & 0.000138 & 0.003694 & 0.000103 \\
\hline
$\rho_0=0.16$ & 0.001247 & 0.000176 & 0.001117 & 0.000154 &
0.001350 & 0.000189 & 0.001260 & 0.000184 \\
\hline
\end{tabular} \label{dataTableR} }
}
\end{center}
\end{table}
We see from Figure \ref{SIRCA_Simulations}, as well as Table \ref{dataTableI} and Table \ref{dataTableR}, that the locally homogeneous GRR approaches the E-AB with less error than the Global GRR.
Despite the scaling and translation differences, the general behavioral trends of the Global GRR and locally homogeneous GRR emulate the E-AB agent state densities.
The surface plot in Figure \ref{fig: ErrorSurf} shows the mean error between the locally homogeneous GRR and the E-AB simulations with respect to the number of infected individuals.
The horizontal axis represents the expected number of susceptible agents in the initial infected agents neighborhood and the vertical axis represents the contact tolerance $\kappa$ for the mean error calculated from \eqref{metric2} using 150 iterations of data.
We fixed the number of agents, $N$, and varied the infectivity radius, $\rho_0$, to generate the error surface plot in Figure \ref{fig: ErrorSurf}; however, one can generate similar error surface plots by fixing $N$ and varying $\rho_0$.
Regardless of the contact tolerance $\kappa$, the GRR approaches the mean simulation's fixed point with only two
expected susceptible agents in the initial infected agent's neighborhood---the lower left subplot \textit{does} approach the fixed point when the simulation runs for more iterations. However, the early infection front is better captured as the density of agents in the initial infectivity radius increases, as shown in the dark blue bands in the surface plot in Figure \ref{fig: ErrorSurf}.
In the yellow and orange bands of the surface plot (approximately fewer than four susceptible agents in the initial infectivity radius), we find that in some simulations the infected agents transition to recovered before any susceptible agents become infected, reaching a point early in the epidemic where there are no infected agents. These cases, in which the epidemic ``dies out,'' skews the expected number of infected, leading to higher error.
The number of iterations required to match depends inversely on the expected number of susceptible agents in the initial infectivity radius. For very low density simulations, where fewer than two susceptible agents are expected to be in the initial infectivity radius, the epidemic has a greater chance of ``dying out,'' which makes our current analysis unreliable.
\begin{figure}
\centering
\includegraphics[width=4in]{MY_ComboR}
\caption{The surface plot in the center displays the error between the locally homogeneous GRR and the E-AB simulations with respect to the number of infected agents as a function of the contact tolerance and the expected number of susceptible agents located in the initial infected agent's neighborhood at $t=0$. The 6 outside plots show the locally homogeneous GRR infected population solution (blue) and the simulated solution (black) with a bound of $\pm$ one standard deviation (yellow).}
\label{fig: ErrorSurf}
\end{figure}
\section{Discussion and conclusions}
In this paper we have introduced a Global Recurrence Rule (GRR) for estimating the state densities for each iteration of a Markovian off-lattice AB model. We demonstrated its utility with a three state Epidemiological Agent-Based (E-AB) model. {However, this analysis can be used for any multi-state model of moving individuals that influence the states of other individuals within a neighborhood.} For the E-AB,
we were able to perform stability analysis using the GRR, as well as determine bounds for the efficacy of the GRR in early stages of the epidemic.
We note that other analytical techniques, besides stability analysis, can be explored with the GRR, including parameter sensitivity analysis and parameter estimation. Performing these calculations directly with an AB model would prove to be very computationally expensive. But, with the computational efficiency of the GRR, one could take existing epidemic data and find the E-AB parameters that best-fit the data. It is well known that the ease of quick and simple calibration of an AB model is critical \cite{Prieto2016}, and this methodology provides a framework to easily handle parameter sweeps.
We identified two classes of E-AB GRR: a globally homogeneous GRR, which assumes that the infected agents are uniformly distributed throughout the domain, and a locally homogeneous GRR, which assumes that there is an infectivity front expanding outward from the initially infected agent.
With relaxed assumptions, the locally homogeneous GRR performs better than the globally homogeneous GRR with respect to early epidemic prediction. However, we demonstrated and proved that the much simplified globally homogeneous GRR can predict long-term behavior just as well as the locally homogeneous GRR.
Further, we demonstrated that the GRR is a generalized model, but is not unique in its application---certain choices must be made.
The generalized GRR definition lends itself to be used as a framework when adapting similar models. For example, if the E-AB were three-dimensional or if the neighborhoods were a different geometry, then we could use our previously derived GRR equations \ref{Global GRRI}, \ref{Global GRRR}, and \ref{LocalGRRR} while only simply having to derive new expressions for $\mu(\mathcal N)$ and $\zeta_t$. Further, we assumed a constant number of agents, $N$, but we could derive a GRR to calculate $S_t$, $I_t$, and $R_t$ that incorporates a dynamically varying number of agents in much the same way as we did in Section \ref{eca-results}. The analysis would be similar, only in three-dimensional phase space instead of two-dimensional.
Previous analytical techniques, such as mean-field game theory, assumed the density of agents approaches infinity in order to calculate end behavior \cite{Lasry}. Other approaches take continuum limits to approximate the dynamics of AB state distributions as a system of PDEs \cite{Chaturapruek,Devitt-Lee,Othmer}, which often corresponds to reducing the scales to infinitesimal time or spatial steps. In contrast, the GRR analysis allows for and takes into account a finite number of agents in a discrete spatial and temporal domain, which in some cases might more closely reflect the outcome of interest for a particular application.
Moreover, the GRR incorporates the notion that the state changes are incurred through spatially defined neighborhoods. Traditional differential equation formulations of SIR models do not incorporate this feature since there is an assumption of a well-mixed population, but we saw that these neighborhoods affected our steady state solutions. Since our GRR analysis incorporates movements of individuals, this causes the contacts between individuals to be dynamic. We note that it is not feasible to do a direct comparison with a differential equation SIR model but that previous CA models with individuals at fixed locations and fixed neighborhoods have done some comparisons for specific cases \cite{Fuentes99}.
Our explicit GRR formulation for the E-AB model ultimately fails when the density of the infected population is zero. In general, the expansion of the wave of infectivity is caused by the infection spread, rather than the agent movement.
However, as can be seen in Figure \ref{fig: ErrorSurf}, when agent density is low, the early infectivity front growth relies on agent movement.
For the infection to not ``die out'' in these
cases, we require an increase in the ratio of the movement size to the neighborhood area to increase the probability that a susceptible agent encounters the infectivity region. {Recent work has studied disease dynamics on dynamic networks \cite{Bansal10,Danon11,Enright18,Rocha16}; the threshold at which a disease can become an epidemic is dependent on reshaping of the contact network \cite{Schwarzkopf10,Segbroeck10}. Volz and Meyers also derived analytical results that have shown epidemics are possible for a range of a mixing parameter that controls the heterogeneity of contacts \cite{Volz09}. In the future, it will be interesting to understand epidemic thresholds and continuum approximations of state changes in order to determine the probability that an infection will ``die out" using the GRR analysis for different movement rules for the agents, which could extensively change mixing and contacts of individuals.} This future analysis will establish density and parameter bounds for when the E-AB GRR formulation is reliable. Early predictions of disease dynamics are necessary \cite{Pellis2015,Roberts2015}, and the proposed framework can be extended to determine accuracy of these estimates for given parameter regimes.
\newpage
|
\section{Introduction}
Our quest to understand the fundamental laws of nature and to design ever more advanced technologies requires precise measurements with outstanding performance even in challenging environmental conditions.
Realizing breakthrough discoveries and revolutionary technologies, such as gravitational wave detection and self-driving cars,
often implies measuring extremely weak signals, demanding a continuous improvement of our measurement tools. Two figures of merit, sensitivity and stability, are crucial for these tasks: while quantum sensors have achieved sensitivities beyond any other technology, they are often prone to instability and decoherence due to external influences. One such quantum sensor is the Nitrogen Vacancy (NV) center in diamond, which is one of the most promising platforms for quantum sensing and many other applications of quantum mechanics~\cite{RN2, RN14}.
Increasing the coherence of NV centers via better controlled growth of diamonds~\cite{Balasubramanian2009}, implementation of dynamical decoupling sequences~\cite{Bar-Gill2013,deLange2010,Knowles2014}, and quantum memories~\cite{RN21} are few of the many advances that have led to an improved magnetic field sensitivity, able to probe nanoscale weak phenomena in condensed matter~\cite{RN12, RN17, RN18} and biology~\cite{RN19}. Measuring weak signals is however not just a matter of using a sensitive device, but also being able to extract signals out of environmental noise via long averaging. In turns, this requires using stable sensors as well as implementing protocols to suppress the effects of different noise sources.
As NV centers comprise an electronic and a nuclear spin within a single lattice site~\cite{RN3}, they enable a broader range of potential applications. Here we report the design of a compact combinatorial device containing two different large ensembles of sensors in the same footprint, taking advantage of the very high densities ($\sim\!10^{17}$~cm$^{-3}$) of solid-state systems. We demonstrate a cross-sensing application of the NV electronic and nuclear spin, where the nuclear spin is used as the primary sensor, while the electronic spin, by sensing the exact same fluctuations of the environment, is used to stabilize it. Specifically, we implement Ramsey interferometry with a nuclear spin ensemble, a protocol that could be exploited to make a rotation sensor. The operating principle of such a quantum spin gyroscope is based on the detection of the dynamic phase accumulated due to the rotation of a spin around its symmetry axis~\cite{RN4}. Alternative NV spin gyroscope designs are based on the measurement of the Berry phase~\cite{RN5, RN22, RN26}, the shift of the Larmor frequency of $^{13}$C nuclear spins due to pseudo-fields~\cite{Wood17}, or an effective AC magnetic field when the spins are rotating in a non-coaxial static magnetic field~\cite{RN6}. The two latter techniques rely on the ability of the NV centers to measure their magnetic field environment, which requires to use the highly sensitive NV electronic spins that suffer from shorter coherence time compared to the $^{14}$N nuclear spins. On the other hand, gyroscopes probing the geometric phase due to the adiabatic evolution of the Hamiltonian during a rotation will have similar performance as devices that measure the dynamic phase. Both can take advantage of using a large number of $^{14}$N nuclear spins with long dephasing time in an isotopically purified $^{12}$C diamond to promise rotation sensitivities of the order of 10$^{-1}$~$\mathrm{\ deg\ s^{-1}/\sqrt{Hz}}$~\cite{RN5, RN22, RN26}.
Indeed, because of its small gyromagnetic ratio ($\gamma_N\!=\!0.3$~kHz/G), the $^{14}$N nuclear spin is a poor magnetic field sensor, so it is less perturbed by any magnetic environmental noise and hence exhibits a coherence time of $\sim\!1$ ms in large ensembles.
To improve on that, we will pair our rotation sensing protocol with a feedback loop to reach long-term stability. Typically, feedback protocols as for instance quantum error correction (QEC) codes are implemented via ancilla qubits either for their isolation to the environment or to use redundant degrees of freedom. QEC codes have been proposed with NV centers for quantum metrology \cite{Kessler14,Unden16} where the NV electronic spin is used as main sensor and nearby nuclear spins as ancilla qubits. Here our approach is the opposite: we rely on the advantages of the nuclear spins as presented above and exploit the fact that the NV spin is instead very sensitive to its environment.
We quantify the stability of our quantum sensor in the presence of a controlled magnetic perturbation, in both a free running and a corrected regime, by computing the Allan deviation, which highlights characteristic features related to the different types of noise affecting our sensors. In particular, we demonstrate using this figure of merit that we can stabilize the output signal of the nuclear spins via an active feedback scheme using the NV electronic spin as a local magnetic field sensor to monitor the common environment of both spins. We thus recover a square root behavior of the Allan deviation, enabling an efficient averaging of the nuclear Ramsey signal over a period longer than a day.
\section{Results}
Our device is based on an ensemble of NV$^-$ defects in diamond (see Appendix~\ref{sec:fptESR}), providing a hybrid electron-nuclear spin system with optical addressability.
We designed optical and microwave control apparatus, as well as control protocols, in order to demonstrate the capabilities of this sensor.
By applying a green laser beam focused on a 10 $\mu$m spot during 30 $\mu$s, we initialize $10^9$ NV centers in the $m_S=0$ electronic spin ground state manifold.
Two permanent magnets in a Helmholtz configuration apply a static field of 420~G, aligned along the NV-center's $\langle111\rangle$ axis.
Coherent control between the $m_S=0$ and $m_S = -1$ spin states is performed via pulsed resonant microwave at 1.7 GHz. The longitudinal component of the NV spin state can be determined optically by monitoring the fluorescence intensity with 3~PIN photodiodes placed in contact with the edges of the diamond and thus collecting 6\% of the total fluorescence ~\cite{RN8}.
\begin{figure*}[tbh!]\centering
\includegraphics[width=0.8\textwidth]{GyroFig1.png}
\caption{\textbf{Control and readout of spin ensembles. }
(a) Schematics of the experimental device (See main text for a description). (b) Pulsed electronic spin resonance (ESR) spectrum of an ensemble of NV centers. The fluorescence signal is normalized to this measured prior to driving the NV spin with microwave. We added an offset to the blue curve for better visualization. In blue, NV centers are aligned with an external magnetic field of $\sim 26$~G and show an equal population in each nuclear spin state. On the other hand, for a magnetic field of $\sim 420$~G (red), the $^{14}$N nuclear spin is initialized in a particular spin state ($\ket{m_S=0, m_I=1}$) via a transfer of polarization that occurs close the excited state level anti-crossing (ESLAC).
(c) Coherence decay of the electronic spin. We use two subsequent Ramsey sequences, each one composed of two $\pi/2$ pulses detuned from the resonance frequency $\nu_e$~=~1.704 GHz by $\Delta \nu_e$~=~10~MHz. The phase of the second $\pi/2$ pulses are shifted by $\pi$ to measure both spin projections. Similarly to equation \eqref{eq:sigF}, we plot the signal difference f. This signal oscillates at the detuning frequency $\Delta \nu_e$ and decays with a coherence time T$_{2e}^*$ = 403(22)~ns.
(d) Coherence decay of the nuclear spin. The resonance frequency $\nu_n$ is 4.68 MHz and the detuning $\Delta \nu_n$~=~5.5~kHz. The nuclear spin is prepared in a superposition of state $(\ket{0,0} +\ket{0,1} )/2$. We measure a coherence time T$_{2n}^*$ = 840(79) $\mu$s via a Ramsey pulse sequence. A microwave $\pi$ pulse tuned at $\nu_e$ is used a selective mapping ($\ket{0,0} \rightarrow \ket{-1,0}$, $\ket{0,1} \rightarrow \ket{0,1}$) between the nuclear spin state and the NV electronic spin state to obtain higher contrast during the spin-dependent fluorescence readout. (e) Contrast response of the nuclear spin sensor to a linear change of the phase of the last $\pi$/2 pulse of the Ramsey sequence with (without) selective mapping in red (blue, Appendix~\ref{sec:raw}). This is the same response as if the NV sensor would be accumulated a quantum phase $\Phi$ during a time t = 600~$\mu$s due to physical rotation about the NV axis. The nuclear spin readout can be done without a mapping releasing the constraint of a narrowband, selective pulse but with lower contrast.
}
\label{fig:fig1}
\end{figure*}
Electron spin resonances are optically detected by sweeping the carrier frequency of a 5$\mu$s-long microwave pulse. We show in Fig.~\ref{fig:fig1}(b) the spectrum recorded from the ensemble of NV centers aligned with a magnetic field of 420~G (red) and of 26~G (blue). The hyperfine coupling between the NV center and the Nitrogen nuclear spin splits each electronic manifold into three non-degenerate states. The relative amplitude of each of the three Lorentzians of the fit provides an estimate of the degree of polarization of the nuclear spin state, which is clearly unpolarized in case of the second spectrum.
At a magnetic field of 420~G (corresponding to a Zeeman energy $\gamma_eB\approx1.18$~GHz), a level anticrossing in the excited state (ESLAC) (zero-field splitting $\approx1.5$~GHz) allows for polarization transfer from the electronic spin onto the nuclear spin~\cite{RN9, RN25}. This results in initializing both NV center spins into the $\ket{m_S=0, m_I=1}$
state with a polarization of $95(4)\%$ as is visible in the red spectrum (Fig.~\ref{fig:fig1}(b)).
We then choose a pair of hyperfine states ($\ket{0,1}$ and $\ket{0,0}$) as a basis for our sensing qubit that is coherently controlled with a radio-frequency radiation at 4.68 MHz. Reading out the difference of populations between these states is done via a selective mapping between the state $m_I = 0$ onto $m_S = -1$. Experimentally we apply a microwave pulse tuned at 1.704 GHz to be resonant on the transition $\ket{0,0}\rightarrow \ket{-1,0}$ before the fluorescence measurement. The spins that were initially in the state $m_I = 0$ will therefore fluoresce with a lower intensity ~\cite{RN11}. Our sensing qubit benefits from the longer coherence time of the $^{14}$N nuclear spin ~\cite{RN10}, which be measured by plotting the change of NV fluorescence signal F given by equation \eqref{eq:sigF} after applying a series of Ramsey pulse sequences (described in \ref{sec:gyro}). In Figure~\ref{fig:fig1}(d), we observe a decaying signal with a nuclear dephasing time of $T_{2n}^{*} = 840(79)\ \mu$s.
\subsection{Emulated nuclear spin gyroscope}
\label{sec:gyro}
Using all these steps together we can implement quantum sensing protocols to measure various physical quantities. The Ramsey sequence
is one simple example of such protocols where we drive the nuclear spin sensor to a superposition of state, after which it evolves freely during a precession time $t$ that is usually set on the order of the coherence time $T_{2n}^{*}$. It will thus acquire a relative phase dependent on the strength of the measured quantity that is transferred into a difference of population for optical readout.
In the case of a rotating spin at the rate $\Omega$ and in the presence of an external magnetic field $b$, the phase is given by $\Phi = (\gamma b + \Omega) t$. In other words, a physical rotation would be coupled into this dynamic phase and mapped out through a population difference. Equivalently, and more intuitively, we can consider the two rf $\pi/2$-pulses as being applied along two different axes. The first pulse is along a reference axis (x-axis by convention) while the second one is about an arbitrary axis (x$'$-axis) rotated by an angle $\theta=\Omega t$. In our experimental proof-of-principle of rotation sensing, we emulate such a phase accumulation by cycling the phase of the last pulse of the Ramsey sequence as shown in Fig.~\ref{fig:fig1}(e), followed by a spin readout that includes a mapping pulse (red). Setting the accumulation time $t = 600\ \mu$s close $T_{2n}^*$, this sequence simulates a rotation at the rate $\Omega_s = \theta/t$. From a statistical analysis of the data of Figure~\ref{fig:fig1}(e), we determine the rotation rate sensitivity as the signal amplitude equivalent to the amount of noise $\delta \Omega$ after averaging N$_\mathrm{seq}$ subsequent sequences during a total acquisition time of one second (i. e. the averaged signal for a signal-to-noise ratio of 1). Experimentally, it is measured as $\eta = \sigma_f(T) \sqrt{T}/dS_\Omega$ where $\sigma_f$ is the standard error in a set of fluorescence signal measurements, T the measurement time and $dS_\Omega$ is the slope of the fluorescence signal as a function of the rotation rate (here $\Omega = \Phi/t$ from Figure~\ref{fig:fig1}(e)). We obtained a sensitivity of $3000 \mathrm{\ deg\ s^{-1}/ \sqrt{\mathrm{Hz}}}$ for nuclear spins. A sensitivity comparison with electronic spins (sensitivity of $0.5 \times 10^6 \mathrm{\ deg\ s^{-1}/ \sqrt{\mathrm{Hz}}}$) indeed shows that the nuclear spin sensor benefits from longer coherence time.
We believe that our sensitivity is reduced by technical limitations that include an excess of electrical noise from the photodetectors as well as an excess of background light from the microwave circuit that reduces the contrast. We estimated that technical improvements could lead to sensitivities in the order of 10$\mathrm{\ deg\ s^{-1}/ \sqrt{\mathrm{Hz}}}$ (See Appendix \ref{sec:setup}). Further improvement can be obtained with dynamical decoupling techniques and spin-bath driving that would allow for an extension of the coherence time T$_2^*$~\cite{Chen18, Bauch18}.
Close to the ESLAC, the mechanism of nuclear spin repolarization provides a mean to read out the nuclear spin state without using a narrowband, selective microwave pulse. Indeed, the polarization process is enabled by flip-flops of the electronic and nuclear spins in the excited state. The ensuing state swapping also modifies the measured fluorescence intensity depending on the initial nuclear state.
Such a method has the main advantage of being intrinsic to the NV center and consequently not relying on any coherent control that requires calibration and stability. It faces however the drawback of a lower contrast, here by a factor 3 (blue, Fig.~\ref{fig:fig1}(e)), as the state swapping is stochastic in nature, due to the 12ns short excited state lifetime, and not coherent as for the selective pulse
Our device is built in a very compact design as nanoscale, combinatorial sensors are embedded in the same footprint and coherent control can be delivered by the same loop antenna. Thus, the nuclear and electronics spins allows for measurements of two independent quantities like temperature, magnetic and electric fields~\cite{RN7} as well as rotations, probed at the same lattice site. Besides, the two sensors are then sensitive to the same local environment that could cause errors and drifts of their output signal at different timescales.
In particular, in our setup, magnetic field amplitude drifts due to a change of magnetization of the permanent magnets will affect both sensors similarly: a change of magnetic field strength will induce a phase shift during the Ramsey sequence and will cause systematics in the rotation measurement. In the following we analyze the stability of our combinatorial quantum sensor and describe schemes to mitigate drifts.
\begin{figure*}[thb]\centering
\includegraphics[width=0.85\textwidth]{GyroFig2}
\caption{\textbf{Nuclear spin sensor with stabilized readout.} (a) The sequence consisting in alternating measurements of the quantum phase via a Ramsey sequence with the $^{14}$N nuclear spins ($t=600 \mu$s) and measurements of the energy shift via a set of ESR measurements with the NV spins. The full sequence is symmetrized to reject common noise. (b) Nuclear Ramsey contrast F as defined by equation \eqref{eq:sigF} with an uncorrected (blue) and corrected (red) selective mapping. A slowly varying magnetic field shifts the energy levels and perturb the mapping. (c) Allan deviations of the Ramsey signals of (b). The maximum averaging time $\tau$ used to calculate the Allan deviation is about a third of the total acquisition time of 1.15 day (see Appendix~\ref{sec:AllanApp}). The right axis is rescaled using a factor s = $4.4\times 10^6\ (\mathrm{deg\ s^{-1}})^{-1}$ calculated at the steepest point of curve F in Figure~\ref{fig:fig1}(e).}
\label{fig:fig2}
\end{figure*}
Here we design and implement an adaptive protocol ~\cite{RN12, RN13,RN13b, RN20} to locally probe magnetic field changes and feed this information back on both sensors to stabilize their combined output signal.
The control protocol is depicted in Figure~\ref{fig:fig2}(a) and consists in six interleaved measurements.
We apply a Ramsey sequence to the nuclear spins. The relative phase between the two $\pi/2$ pulses is chosen to $\theta_r=90^{\circ}$ so the nuclear signal is 0 when no phase is accumulated (see Figure~\ref{fig:fig1}(e)). As described above, a phase shift of the driving field around this ideal bias point mimics a physical rotation that we want to detect.
In a regime of small signals, the response of this first sensor will be a linear change of the fluorescence output signal $S_{90}$, measured with the spin readout which includes a mapping step as described above.
This sensing module is followed by two spin resonance measurements at the frequencies $\nu_4 = \nu+700$~kHz and $\nu_3= \nu+350$~kHz. The second half of the sequence consists in repeating a similar set of measurements with a relative phase $\theta_r = -90 ^\circ$ and the ESR frequencies $\nu_1 = \nu-700$~kHz, $\nu_2= \nu-350$~kHz. All these frequencies are graphically represented in Figure~\ref{fig:fig1}(b). In a regime where the measured physical quantities are slowly varying with respect to the total sequence length ($\sim 1$ ms), noise that have a similar signature on both output signals $S_{\pm90}$, as for example laser intensity noise, can be suppressed by using the effective signal
\begin{equation}
F = \frac{S_{90} - S_{-90}}{S_{90} + S_{-90}}.%
\label{eq:sigF}
\end{equation}
However, such a common-noise rejection scheme is still inefficient in case of sources of noise, such as magnetic fields, that act similarly to a signal. To suppress them, we exploited the four ESR measurements~\cite{RN1} to probe line shifts caused by these sources of noise, interleaved with the sensing protocol of the first sensor. The relative fluorescence intensity at four different microwave frequency allows for recovering the transition frequency $\nu$ and determining the strength of the field causing this line shift (see Appendix~\ref{sec:fptESR}).
In the following, we test our stabilizing schemes against an engineered, slowly drifting perturbation generated by applying an oscillating magnetic field created by a coil placed at 1~cm of the diamond sample. Its period is set to 1000~s and its strength along the NV axis is measured to be 0.14 G peak to peak via the four-point ESR measurements described above. In Figure~\ref{fig:fig2}(b), a clear oscillation of a period of 1000~s is visible in the signal of the nuclear Ramsey measurements (blue data points) as well as a contribution from a slower environmental noise, on which we have no control, on a timescale of a few hours. While nuclear spins are little sensitive to our applied magnetic perturbation, NV spins are far more affected by it, which highly disturbs the mapping step. Indeed, the transition frequency of selective pulses must be finely calibrated to maximize the readout fidelity and be stable along the full measurement dataset in order to limit readout errors. The feedback protocols we implemented however succeed in stabilizing both the nuclear and electronic spin transition frequency fluctuations. In what follows, we present two scenarios in which we isolate the effect of the perturbation to a single parameter that will be stabilized. First, we compensate the readout mapping pulse frequency to prevent a loss of contrast due to an off-resonance selective pulse. Then, while applying a stronger perturbation but no mapping, we use the signal of the electronic spin to adjust the nuclear spin driving frequency.
\subsection{Cross-sensor feedback stabilization}
We demonstrate here that we can use measurements of the local NV electronic spin to feedback on the nuclear Ramsey measurements to stabilize their results. To do that, we (i) repeat the sequence of Figure~\ref{fig:fig2}(a) N$_r$ = 2000 times, (ii) transfer the measurements on the control computer to compute the ESR shift and (iii) update the experimental parameters to compensate the measured magnetic drifts. The two last steps takes about 50~s, i.~e. for a total duration of 1~min, which would optimally set a lower bound for N$_r$ as one would like to maximize the duty cycle of the sensors. On the other hand, the characteristic timescale and amplitude of the noise limits the number N$_r$ of repetitions after which the correction has to be made as the frequency drift becomes significant. In the case of the engineered perturbation, we choose N$_r$ = 2000 as it corresponds to the maximal drift equivalent to a tenth of the Rabi frequency of the mapping pulse and it is smaller than the bandwidth of the ESR measurements ($\sim$ 500 kHz). We show that feedback helps making the measurement more stable over the full data acquisition of more than 1 day (Figure~\ref{fig:fig2}(b), red).
More quantitatively, we characterize the stability of our dual spin sensor by computing the Allan deviation of the data traces of Figure~\ref{fig:fig2}(b) (See Appendix~\ref{sec:AllanApp} and Figure~\ref{fig:fig2}(c)). We observe that the uncorrected signal (shown in blue) displays an overall decaying behavior with three features. The first two at T=50~s and T=1000~s are the signature of periodic noises at frequencies 1/$T$. They correspond to perturbations associated with (I) the episodically interrupted recordings to update the experimental parameters and (II) the magnetic perturbation that we apply to our sensor. The third noticeable feature (III) is due to the environmental noise that prevents sensors to operate accurately over long runs. We believe this is due mainly temperature changes that affect the magnetization of the permanent magnets. Shown in red is the Allan deviation for the corrected data set which displays a varying stability improvement at different timescale. The perturbation (II) is only partially corrected, mainly because of the comparable timescale of the data acquisition (about 1~min) and the magnetic perturbation period (1000~s), so that the measured field has already considerably changed by the time the correction is applied during the next acquisition.
This is not the case for the uncontrolled environment noise (III), as its variations are much slower: then the feedback protocol based on monitoring a second spin sensor
allows for improving the first sensor readout stability by an order of magnitude. While in this experiment we limited the feedback correction to the electronic spin driving frequency, we show next that we can obtain additional gain by directly correcting the nuclear spin control.
\begin{figure}[htb]\centering
\includegraphics[width=\columnwidth]{GyroFig3}
\caption{\textbf{Stabilized nuclear spin sensor.} (a) Dual sensor scheme. The ESR shift $\nu_\mathrm{ESR}$ collected from the NV spins serves also to feedback on the nuclear spin control parameters $\delta \nu_\mathrm{N}$ to stabilize the $^{14}$N nuclear Ramsey contrast G with respect to a noise common to both spins. (b) Allan deviation of the nuclear spin signal for different correction strengths. We directly collect the nuclear spin-state dependent fluorescence without mapping pulse to isolate the perturbation effect on the nuclear spins signal. The right axis is rescaled using a factor s = $1.4\times 10^6 \ (\mathrm{deg\ s^{-1}})^{-1}$ calculated at the steepest point of curve G in Figure~\ref{fig:fig1}(e). }\label{fig:fig3}
\end{figure}
As both sensors are spins, they are sensitive to magnetic field fluctuations through Zeeman coupling. Due to a small gyromagnetic ratio (10000 times smaller than the one of an electron), the $^{14}$N nuclear spin's response to magnetic fluctuations is weaker. Thus, to be able to see the effect of the magnetic perturbation, we increase its strength to a peak-to-peak value of $\sim$3~G. At the same time, we lengthen its period to 3000~s to stay within the limit of the previously presented four-point ESR bandwidth. Also, to isolate the effect of the perturbation on the nuclear spins, we extract the signal G directly from the bare fluorescence without any selective mapping (similarly defined as in eq. \eqref{eq:sigF}, see Appendix \ref{sec:raw}). We first plot in blue the Allan deviation of the uncorrected signal (Figure~\ref{fig:fig3}(b)). One can distinguish a small deviation from the expected square root behavior confirming that a relative phase due to the magnetic perturbation is imprinted during the free evolution of the nuclear Ramsey sequence.
To correct this, we exploit the fact that the NV spin can probe the strength of this perturbation with a good accuracy, at exactly the same location, to cross feedback between the two sensors (Figure~\ref{fig:fig3}(a)). We probe the Zeeman shift $\delta \nu_\mathrm{NV}$ with the four-point ESR scheme and update the driving frequency of the nuclear Ramsey pulses with a frequency shift $\delta \nu_\mathrm{N} = -\gamma_N/\gamma_e \delta \nu_\mathrm{NV}$. In addition to the free running Ramsey sequence (blue), two other data acquisitions were interleaved at the same time with different corrections factors: with a the correct shift given above (red) and with its opposite (black), thus doubling the error. We can see that we recover a stable data averaging for the good feedback correction factor, whereas the opposite correction leads to an amplification of the perturbation, thus proving that the source of noise is indeed the same for the two spins.
\subsection{History-based feedback protocols}
So far, our feedback protocol consisted only in updating the experimental parameters with averaged values recorded during the previous dataset. However, as we keep records of every dataset, it is in principle possible to use all this knowledge to correct for slower frequency drifts with more advanced protocols. In particular, schemes relying on machine learning techniques~\cite{RN13,RN13b} or the Bayesian rule~\cite{RN12} are potential candidates to extract the most important features of the noise and be able to apply efficient corrections. Here, we would like to assess the question of the efficiency of using the previous records in the presence of stochastic noise to correct the control parameters of an ensemble of sensors. We simulate a signal constantly equal to zero (similar to the signal F at the most sensitive operating point) on top of which is added a sinusoidal perturbation and a stochastic noise (Figure~\ref{fig:fig4}(a)). An intuitive approach to guess the best transition frequency at the $i+1$ step is to fit the $N$ previous points with a model function, and to extrapolate the result to the future point. In the case where the stochastic noise is absent, like for example with a perfect readout, a polynomial fit allows for perfectly suppressing the perturbation, as long as one suitably increases the degree of the polynomial (Figure~\ref{fig:fig4}(b)). This is not necessary true in the scenario of a non-zero stochastic noise. As we can notice that for noise amplitude of the same order as the sinusoidal perturbation, a linear regression between the two last datasets provides a better correction than higher order polynomials (Figure~\ref{fig:fig4}(c)). Experimentally, we have measured a ratio between the magnetic perturbation and our readout noise of 0.1 (Figure~\ref{fig:fig4}(d)), indicating that we are in the regime where taking into account the past evolution of the transition frequency does not provide any help to stabilize any further our sensors.
\begin{figure}[t]\centering
\includegraphics[width=0.48\textwidth]{GyroFig4}
\caption{\textbf{Simulation of history-based feedback protocols.}
(a) We simulate a sinusoidal perturbation with a stochastic noise of varying amplitude (red). A polynomial (here linear, blue dash) fit is used on certain number N of points [i,i+N] to guess the value of the point i+N+1, generating the guessed signal (orange). In blue, we plot the error between red and orange. (b) In the case of the absence of noise, we plot the error as previously defined for different degree d of the fitting polynomial. The number of points used is set N=d+1. (c) We plot the average error, defined as the mean value of the error (blue curve in (a)) averaged over 20 realizations, as a function of the noise amplitude and the degree of the polynomial. We noticed that the trend is different depending on the noise amplitude. (d) From experimental ESR data, we estimated a ratio between the magnetic perturbation and our experimental noise of 0.1. This is a regime where no improvement can be achieved using the history of the measurements as one can see in the red region of (c). }
\label{fig:fig4}
\end{figure}
\section{Discussion}
We used a large ensemble of NV centers in diamond to realize a combinatorial dual spin sensor, providing the capabilities to measure two physical quantities on the same micrometer scale footprint and to stabilize one sensor with local information collected in realtime via the second sensor. Both sensors were coherently controlled in microsecond timescales with microwave and radio-frequency radiations and read out after laser excitation via an efficient fluorescence collection scheme from the side of the diamond ~\cite{RN8}. We used electron spin resonance measurements to probe the magnetic field fluctuations and stabilize the output signal of interleaved Ramsey sequences performed on the $^{14}$N nuclear spin. Moreover, due to the strong interaction between the two spins that composed the NV center, one can use the electronic spin to increase the nuclear spin readout contrast by a factor 3. In turn, this would increase the rotation rate sensitivity by the same factor since the step of mapping extends the length of the sequence by only 5$\mu$s and doesn't affect the duty cycle. On the other hand, this mapping affects the stability of the sensor and prevent averaging beyond a certain number of repetitions. We show that our feedback scheme can improve the stability of the nuclear spin readout and the accuracy of their measurements. In figure~\ref{fig:fig2}(c), we see that the precision of the measurement tends to degrade after a total time of acquisition of $\sim 3 \tau_\mathrm{opt.}$ = 30000 s. At this optimal point, correcting the mapping pulse frequency allows to improve the precision on the averaged signal by a factor 2.5, down to a contrast error of $4 \times 10^{-6}$ (equivalent to a minimum rotation rate detectable of $\sim 1 \mathrm{\ deg\ s^{-1}}$), and almost reaches the minimum error given by a perfect average of independent measurements, which would have followed a square-root law. Given the experimental parameters, i. e. the total ESR measurement time ($2 \times 50 \mu$s) and the time to compute update the frequency (which can be reduced to less than 5 seconds), the stabilization stage does not extend significantly the sequence neither and consequently does not affect the sensitivity. Hence we believe that there is a benefit to use both mapping pulses and a stabilization procedure.
We anticipate that such a device can potentially find application as a very stable gyroscope, allowing navigation with limited or even no need of remote localization. Existing technologies like micro electro-mechanical systems~\cite{RN101} (MEMS) or spin comagnetometers~\cite{RN23, Limes18} are already successful in making sensitive gyroscopes that have thus gained ubiquitous usage in everyday life, from navigation and inertial sensing, to rotation ion sensors in hand-held devices and automobiles. Detailed comparisons in terms of sensitivity and stability between different technologies can be find in~\cite{takase08, Passaro17, RN4}. In particular, while commercial gyroscopes achieve typical sensitivities of $0.1 \mathrm{\ deg\ s^{-1}}/\sqrt{\mathrm{Hz}}$ in hundreds of micron size footprint, their accuracy is strongly affected by drifts after few minutes of operation, making them unattractive for geodetic applications ~\cite{RN100}. On the other hand, our results show sensing capabilities for over many hours, confirming the potential of NV centers in diamond as competitive modality for such applications. Furthermore, long-term stability is a key figure of merit is the search of discrepancies in the current theories in fundamental physics and long averaging is almost always required in current tests of Lorentz and CPT symmetries or search of clues to understand dark matter~\cite{Budker14,Garcon18,Rajendran17}.
|
\section{Introduction}
\IEEEPARstart{I}{nternet} of Things (IoT) is a key application paradigm for the next generation wireless communication systems.
Due to the energy, cost and complexity constraints, energy- and spectrum efficient communication technologies are desirable for the IoT devices \cite{Stankovic2014DirectionIoT}.
Cognitive backscatter communication is an emerging technology for green IoT to fulfill such demand \cite{LIU2013AMBC}, which is an integration of the well-known cognitive radio concept and the \emph{backscatter communication} (BackCom) technology. To be specific, the backscatter (secondary) system shares not only the same spectrum, but also the same RF source with the legacy (primary) system (e.g., cellular base stations, \emph{digital television} (DTV) transmitters, \emph{wireless fidelity} (Wi-Fi) access points, etc).
A typical \emph{cognitive backscatter system} (CBS) is illustrated in {\figurename~\ref{Cog_AmBC model}}, in which the \emph{backscatter device} (BD) transmits its own information by reflecting the \emph{radio-frequency} (RF) carriers from the legacy transmitter.
Hence, no power-hungry RF components (e.g., up-converters and power amplifiers) are required, and the power consumption of the BD can be ultra low \cite{Welbourne2009RFID,Boyer2014RFIDBC,ZhangR2013WIPT,Larsson2013SIPT}.
In addition, thanks to the spectrum sharing natural, the CBS also achieves desirable spectrum utilization efficiency \cite{KangX2017RidingPrimary}.
Cognitive backscattering is still in its infant stage, and there are various technical challenges arising from the data transmission and networking perspectives \cite{Niyato2018survey,Niyato2018AmBCMagazine,LiuW2017survey}.
In \cite{LIU2013AMBC}, the first prototype about this conception is introduced which is referred to as the \emph{ambient backscatter communications} (AmBC) for ultra short range communication between two passive tags.
In \cite{Darsena2017ModelandPerformance,darsena2016performance}, early attempts on the information theory and performance analyses are carried out for the \emph{orthogonal frequency division multiplexing} (OFDM) modulated RF source and Gaussian distributed backscatter signals.
In \cite{WangG2018AmBCrate}, a numerical method is presented to calculate the maximum achievable rate of the on-off modulated backscatter signals.
In \cite{Dinesh2015BackFi,Ensworth2017BLE,Talla2017LoRabackscatter}, enormous works are devoted to new prototyping for practical implementation, e.g., ``BakcFi'', ``BLE-Backscatter'', and ``Lora-Backscatter''.
In \cite{lyu2018optimal}, the optimal tradeoff between the energy harvesting and backscatter transmission is investigated by taking the finite battery capacity into account.
In \cite{Niyato2017ImproveNetPerform,Niyato2018TimeSchedulingHTB,wang2018stackelberg,lu2018wireless}, hybrid backscatter network is investigated where the CBS is employed to assist the conventional wireless powered radio networks. In these works, time scheduling protocols are investigated for the tradeoff between the cognitive backscattering and the conventional Backcom techniques.
\begin{figure}
\centering
\includegraphics[width=.7\columnwidth]{CognitiveAmBC_v3-eps-converted-to.pdf}
\caption{A basic CBS illustration.}
\label{Cog_AmBC model}
\vspace{-1.5em}
\end{figure}
In this paper, we mainly focus on the receiver design for CBS.
From {\figurename~\ref{Cog_AmBC model}}, the receiver of the backscatter system (i.e., reader) simultaneously receives the backscattered signal from the BD and the \emph{direct-link interference} (DLI) from the legacy system.
Different with the conventional Backcom where the DLI is ``unmodulated'' and can easily be eliminated, the RF carries from the legacy transmitter is ``modulated'' and unknown for the reader in CBS.
The randomness of the unknown strong DLI makes the backscatter symbols very hard to be distinguished.
In \cite{LIU2013AMBC}, noncoherent \emph{energy detector} (ED) is utilized to recover the backscatter symbols.
The performance of the ED and its modified versions are analyzed in \cite{TaoQ2018Manchester,GaoFF2017EDAmBC,wang2016ambient,GaoFF2017NDAmBC}.
In \cite{GaoFF2017NDAmBC,Guo2018IoT}, an interesting error floor problem of ED is pointed out: the \emph{bit error rate} (BER) for backscatter symbol detection converges to a non-zero floor with the increase of the transmitted power at the legacy RF source.
That means, the ED based backscatter system could not benefit from the increase of the RF source transmit power due to the existence of DLI.
As a result, to achieve reliable detection, the transmission rate of the backscatter system becomes quite limited.
To realize the high-speed transmission, it is critical to suppress the DLI at the reader.
Instead of the tag-to-tag communication demonstrated in \cite{LIU2013AMBC}, some works consider a battery-powered reader (employed as a IoT \emph{access point} (AP)), which allows more complicated receiver to remove the DLI effect and to increase the throughput of the backscatter system.
In the literature, some studies suggest to integrate the reader into the legacy transmitter, and then the self-interference cancellation techniques used in full-duplex communications can be exploited to suppress DLI \cite{darsena2016performance,Dinesh2015BackFi}.
Similarly, the reader can also be integrated into the legacy receiver, so that the source signal is jointly decoded together with the backscatter signal \cite{YangG2017CooperativeBC,RZLong2017beamform,DuanRF2017TVTRbistatic}.
Besides the above fully cooperative scenarios, in \cite{YangG2017OFDM}, the OFDM based legacy source is adopted, and the repeating structure of the cyclic prefix is exploited to design the DLI free backscatter waveforms based on some coordination between the legacy transmitter and the BD.
Nevertheless, it is worth noting that the validity of above methods depends on special requirements in terms of transceiver design, RF source modulation, and synchronization between the legacy system and backscatter system; otherwise, these aforementioned methods will fail to work.
We are interested in developing a general DLI-free detector for various RF sources and application scenarios.
For this purpose, it is assumed that there is no cooperation between the legacy system and the backscatter system.
One early attempt work is presented in \cite{Guo2018IoT}, where only one RF source is considered, and multiple receive antennas are exploited to suppress the DLI.
In this paper, a multiple-RF-source scenario is considered which is illustrated in {\figurename~\ref{backscatter model}}. This assumption is more general and practical. For example, in the smart-home applications, the reader may receive two or even three dominated RF source signals from the neighboring WiFi APs.
In addition, We assume that there is no pilots transmitted from the legacy transmitter or the BD to assist channel estimation at the reader.
The main differences between the CBS shown in {\figurename~\ref{backscatter model}} and the MIMO interference channel are twofold.
First, the backscattering is a multiplicative operation on the incident signals from RF sources in the analog domain.
Therefore, the signal from backscatter link contains a mixture version of the direct-link signals from all the RF sources.
To mitigate the direct link signal straightforwardly may not achieve good performance, which has already been pointed out in \cite{Guo2018IoT}.
Second, the unknown source signals act as unknown channel fading coefficients on the backscatter-link signals.
As a result, the backscatter signal detection is a challenge work, especially without any pilots to assist channel estimation.
In this case, we suggest BD employ on-off differential modulation for noncoherent transmission \cite{LIU2013AMBC,GaoFF2017NDAmBC}.
Novel detection algorithm is required to suppress the DLI from multiple RF sources.
\begin{figure}
[!t]
\centering
\includegraphics[width=.6\columnwidth]{system_v3-eps-converted-to.pdf}
\caption{The CBS consisting of $K$ RF sources, one BD and one multi-antenna reader. The reader receives signals from both direct link and backscatter link.
}
\vspace{-1.5em}
\label{backscatter model}
\end{figure}
The main works and contributions of this paper are summarized as follows:
\begin{itemize}
\item Firstly, the impacts of multiple RF sources is addressed for the CBS. To be specific, the impacts are twofold: providing carrier emitters for transmission, and causing DLI at the receiver.
We derive the \emph{Chernoff Information} (CI) \cite{cover2012elements} to analyze the relationship between the detection performance and the system parameters, such as the number of the RF sources and receive antennas.
Numerical results shown that, the increase of RF sources number may not necessarily lead to performance degradation, when multiple antennas are equipped at the receiver.
\item Secondly, we provide an insight into the error-floor phenomenon through the information-theoretic aspects by deriving the upper bound of the channel capacity.
It is pointed out that, high-throughput backscatter transmission can be realized by the multi-antenna receiver, as long as the antenna number is larger than the amount of RF sources.
\item Thirdly, the optimal (soft decision) and suboptimal (hard decision) detectors are derived to recover the BD original generated symbols based on the \emph{maximum likelihood} (ML) criterion. Numerical results verify that, the suboptimal detector achieves almost the same performance as that of the optimal one with much lower complexity.
\item Finally, the \emph{channel state information} (CSI) learning problem is addressed for practical implementation, bearing in mind that there is no pilots to be exploited. A novel blind channel estimation method is proposed based on clustering algorithm, thanks to the fact that, the received signals may fall into two clusters corresponding to different backscatter states. A modified version of the relative entropy is designed as a new distance metric to assist clustering. Then, channel estimation is realized by exploiting all the received samples in each cluster.
\end{itemize}
The rest of the paper is organized as follows. Section \ref{system model} outlines the CBS model.
In Section \ref{problem}, we provide some theoretical discussion on the impact of RF Sources, such as the maximum achievable rate and the CI.
Next, in Section \ref{coordinated}, the optimal and suboptimal detectors are derived to recover the original information bits of BD.
Then, for piratical application, blind channel estimation method is proposed in Section \ref{semi-blind}.
Numerical results are provided in Section \ref{simulation} and section \ref{conclusion} concludes the paper.
The notations used in this paper are listed as follows. ${\mathbb E}[\cdot]$ denotes statistical expectation, ${\cal Q} (\cdot)$ is the Q-function, $\mathbbm{1} (\cdot)$ is the indicator function, and ${\rm{sgn}}(\cdot)$ is the signum function. ${\rm Pr}(A)$ denotes the probability of event $A$ happens.
${\rm I}(X;Y)$ denotes the mutual information of random variables $X$ and $Y$, ${\rm H}(X)$ denotes the entropy of $X$, and ${\rm H}(X|Y)$ is the conditional entropy.
${\cal{CN}}(\mu, \sigma^2)$ denotes the \emph{circularly symmetric complex Gaussian} (CSCG) distribution with mean $\mu$ and variance $\sigma^2$.
For any general matrix ${\bf G}$, ${\bf G}^T$ and ${\bf G}^H$ denote the transpose and conjugate transpose, respectively.
${\bf{I}}_{M}$ denotes the $M \times M$ identity matrix.
${\rm{tr}}({\bf S})$ is the trace of a square matrix ${\bf S}$, $|{\bf S}|$ denotes is determinant, ${\rm{rank}}({\bf S})$ denotes its rank, and $\|{\bf S}\|_{\rm F}=\sqrt{{\rm{tr}}({\bf S}{\bf S}^H)}$ denotes its Frobenius norm.
$\|{\bf w}\|$ denotes the Euclidean norm of a vector ${\bf w}$.
The quantity $\max(a,b)$ denotes the maximum between two real numbers $a$ and $b$.
$a \oplus b$ represents the addition modulo $2$.
$|x|$ denotes the absolute value of a complex number $x$, and ${\rm{Re}}(x)$ and ${\rm{Im}}(x)$ denote its real part and imaginary part, respectively.
\section{System Model}\label{system model}
{\figurename~\ref{backscatter model}} depicts the CBS model considered in this paper, which consists of $K$ dominated legacy RF sources, a single-antenna BD and a reader equipped with $M$ antennas.
We only consider one BD transmission in every time slot. When the system contains multiple BDs, they can be scheduled by multiple access control protocols such as \emph{time division multiple access} (TDMA).
The BD communicates to a neighbouring reader by reflecting the RF signals with different antenna impedances.
The RF carrier wave is dominated by several surrounding RF sources.
For simplicity, flat block fading assumption is adopted, wherein the channel remains constant over consecutive symbol intervals (i.e., a block).
\subsection{RF Source Signals}
Denote $s_{k,n}$ as the $k$-th RF source signal transmitted at time instant $n$. We assume that $s_{k,n}$ is \emph{independent and identically distributed} (i.i.d.) at different time instants, and it follows the standard CSCG distribution, i.e., $s_{k,n} \sim{\cal{CN}}(0,1)$.
The RF source signal received at the reader consists of two components.
One is the signal from the direct link, $r_{m,n}^{\rm{d}}$, which is transmitted directly from the RF sources to the reader.
The other is the signal from the backscatter link, $r_{m,n}^{\rm{b}}$, which is backscattered from the BD to the reader.
Supposing the BD-reader distance is relatively short, the time delay between the receptions of $r_{m,n}^{\rm{d}}$ and $r_{m,n}^{\rm{b}}$ at the reader is negligible \cite{LIU2013AMBC,GaoFF2017NDAmBC,YangG2017CooperativeBC}. The received signals at the $m$-th antenna of the reader can be expressed as
\begin{equation} \label{equ:2_signalx}
y_{m,n}=r_{m,n}^{\rm{d}}+r_{m,n}^{\rm{b}}+u_{m,n},
\end{equation}
where $n=1,2,\ldots,N$, and $u_{m,n}$ is the CSCG noise with zero mean and unit power, i.e., $u_{m,n} \sim{\cal{CN}}(0,1)$.
\subsection{Direct Link}
At time instant $n$, the direct link signal from the $k$-th RF source received at the $m$-th antenna of the reader can be expressed as
\begin{equation} \label{equ:2dl_signalk}
r_{k,m,n}^{\rm{d}}=f_{k,m} \sqrt{P_{{\rm{s}},k}} s_{k,n},
\end{equation}
where $f_{k,m}$ denotes the small-scale fading from the $k$-th RF source to the reader with ${\mathbb E}[|f_{k,m}|^2]=1$ for all $k$ and $m$, and $P_{{\rm{s}},k}$ is the average received power from the $k$-th direct link.
The average received power $P_{{\rm{s}},k}$ is determined by the transmit power $P_{{\rm{t}},k}$ of the $k$-th RF source and the path loss \cite{Rappaport2002WCPP}:
\begin{equation} \label{equ:2PL_direct}
P_{{\rm{s}},k} = \frac{P_{{\rm{t}},k} G_{{\rm{t}},k} G_{\rm{r}} \lambda^2 }{ (4 \pi)^2 L_{{\rm{d}},k}^{\nu_1}},
\end{equation}
where $L_{{\rm{d}},k}$ is the distance between the $k$-th RF source and the reader in meter, $\nu_1$ is the path loss exponent, $G_{{\rm{t}},k}$ and $G_{\rm{r}}$ are the antenna gain of the $k$-th RF source and the reader, respectively.
Let $\kappa=\frac{{\lambda}^2}{(4 \pi)^2}$, and then \eqref{equ:2PL_direct} becomes
\begin{equation}
P_{{\rm{s}},k} = \frac{\kappa P_{{\rm{t}},k} G_{{\rm{t}},k} G_{\rm{r}} }{ L_{{\rm{d}},k}^{\nu_1}},
\end{equation}
Then the summation signal $r_{m,n}^{\rm{d}}=\sum_{k=1}^K r_{k,m,n}^{\rm{d}}$ is:
\begin{equation} \label{equ:d_signal}
r_{m,n}^{\rm{d}}=\sum_{k=1}^K f_{k,m} \sqrt{P_{{\rm{s}},k}} s_{k,n}.
\end{equation}
\subsection{Backscatter Link}
The RF source signals received at the BD can be expressed as
\begin{equation} \label{equ:2_signal_st}
s_{n}^{\rm{r}}= \sum_{k=1}^K l_k \sqrt{P_{{\rm{b}},k}} s_{k,n},
\end{equation}
where $l_k$ denotes the small-scale fading from the $k$-th RF source to the BD with ${\mathbb E}[|l_k|^2]$=1, and $P_{{\rm{b}},k}$ is the average available power from the $k$-th RF source before backscattering.
Assuming the same path loss exponent, we have
\begin{equation} \label{equ:2PL_BD}
P_{{\rm{b}},k} = \frac{\kappa P_{{\rm{t}},k} G_{{\rm{t}},k} G_{\rm{b}} }{ L_{{\rm{b}},k}^{\nu_1}},
\end{equation}
where $L_{{\rm{b}},k}$ is the distance between the $k$-th RF source and the BD, and $G_{\rm{b}}$ is the antenna gain of the BD.
From the antenna scatterer theorem \cite{Fuschini2008BackscatterAnalyze,Griffin2009LinkBudget,Boyer2014RFIDBC}, we assume that BD only has two backscattering states, which are denoted by $c=0$ and $c=1$ (i.e., \emph{on-off keying} (OOK)), respectively.
When $c=0$, only structural mode scattering exists, and when $c=1$, both the structural mode component and antenna mode component are exist.
Since the structural mode scattering is independent to the antenna load impedance and always exists, it has been already contained in the direct link signal. Define $\alpha$ as the reflection coefficient of the antenna mode scattering, and we further suppose that the BD backscattering state $c$ remains unchanged for $N$ consecutive source symbols. Then, during one BD symbol period, the signal backscattered from the BD to the reader is given by
\begin{equation} \label{equ:2_signal_yt}
s_{n}^{\rm{b}}=\alpha s_{n}^{\rm{r}} c,
\end{equation}
where $n=1,2,\ldots,N$, and $0<|\alpha|^2<1$.
The backscatter link signal received at the $m$-th antenna of the reader is expressed as
\begin{equation} \label{equ:2bl_signal}
r_{m,n}^{\rm{b}}= g_m \sqrt{ \frac{ G_{\rm{b}} G_{\rm{r}}\kappa }{ L_{\rm{c}}^{\nu_2} }} s_{n}^{\rm{b}}
,
\end{equation}
where $g_m$ is the small-scale fading from the BD to the reader with ${\mathbb E}[|g_m|^2]$=1,
$L_{\rm{c}}$ is the BD-reader distance, and ${\nu_2}$ is the path loss exponent\footnote{Generally, as $L_{\rm{c}}$ is relatively short, it is a line-of-sight path between the BD and the reader, and we have ${\nu_2}=2$.}.
Substituting \eqref{equ:2PL_direct}, \eqref{equ:2_signal_st} and \eqref{equ:2_signal_yt} into \eqref{equ:2bl_signal}, we have
\begin{equation} \label{equ:2bl_signaln}
\begin{aligned} [b]
r_{m,n}^{\rm{b}}
&=g_m \alpha c \sum_{k=1}^K l_k \sqrt{ \frac{ \kappa^2 P_{{\rm{t}},k} G_{{\rm{t}},k} G_{\rm{r}} G^2_{\rm{b}} } { L_{\rm{b}}^{\nu_1} L_{\rm{c}}^{\nu_2}}
} s_{k,n}\\
&= g_m \alpha c \sum_{k=1}^K l_k \sqrt{\frac { \kappa P_{{\rm{s}},k} G^2_{\rm{b}} {L_{{\rm{d}},k}^{\nu_1}}} {{L_{{\rm{b}},k}^{\nu_1}} L_{\rm{c}}^{\nu_2}}} s_{k,n}.
\end{aligned}
\end{equation}
Denote the total backscattering power loss of the $k$-th BD as
\begin{equation}
\widetilde \gamma_k=\frac{\kappa |\alpha|^2 G^2_{\rm{b}}{L_{{\rm{d}},k}^{\nu_1}} }{ {L_{{\rm{b}},k}^{\nu_1}} L_{\rm{c}}^{\nu_2}}.
\end{equation}
Let $\bar{\alpha}=\frac{\alpha}{|\alpha|}$ denote the phase shift due to backscattering.
Then we finally have
\begin{equation}
r_{m,n}^{\rm{b}}= g_m \bar{\alpha} c \sum_{k=1}^K l_k \sqrt{ \widetilde \gamma_k { P_{{\rm{s}},k} } } s_{k,n}.
\end{equation}
\subsection{Received Signal at The Reader}
Substituting $r_{m,n}^{\rm{d}}$ and $r_{m,n}^{\rm{b}}$ into \eqref{equ:2_signalx}, the received signal at the $m$-th antenna of the reader is expressed as
\begin{equation} \label{equ:2_signal}
y_{m,n}
=\sum_{k=1}^K \sqrt{P_{{\rm{s}},k}} s_{k,n} \left(f_{k,m}+\bar{\alpha} g_m l_k \sqrt{ \widetilde \gamma_k } c \right)+u_{m,n}.
\end{equation}
We further denote the average \emph{signal-to-noise ratio} (SNR) of the $k$-th direct link as $\gamma_{{\rm{d}},k} \triangleq P_{{\rm{s}},k}$, and the average SNR of the $k$-th backscatter link as $\gamma_{{\rm{b}},k} \triangleq \widetilde \gamma_k P_{{\rm{s}},k}$.
We assume that the direct link channel response $f_{k,m}$ and the backscatter link channel response $l_k g_m$ are mutually independent.
Then, when the transmit power from one of the RF source increases, or when the distance between the RF sources and the backscatter system ($L_{{\rm{b}},k}$ or $L_{{\rm{d}},k}$) becomes shorter, both $\gamma_{{\rm{d}},k}$ and $\gamma_{{\rm{b}},k}$ increase. Also, when the BD-reader distances become shorter, a larger $\widetilde \gamma_k$ is obtained, resulting in a stronger backscatter link.
With $M$ receive antennas at the reader, we denote the channel response vectors as
\begin{align}
{\bf{f}}_k&=[f_{k,1},f_{k,2},\ldots,f_{k,M}]^T,\\
{\bf{g}}&=[g_1,g_2,\ldots,g_M]^T.
\end{align}
By letting ${\bf{h}}_{k,1}=\sqrt{\gamma_{{\rm{d}},k}} {\bf{f}}_k$ and ${\bf{h}}_{k,2}=\bar{\alpha} l_k \sqrt{ \widetilde \gamma \gamma_{{\rm{d}},k}}{\bf{g}}$, the received signals collected by the $M$ antennas can further be expressed as
\begin{equation} \label{equ:2_signalv}
\begin{aligned} [b]
{\bf{y}}_n
&=\sum_{k=1}^K {{\bf{f}}_k} \sqrt{{\gamma_{{\rm{d}},k}}} s_{k,n}+
{\bf{g}} \sum_{k=1}^K l_k \bar{\alpha} \sqrt{\widetilde \gamma_k {\gamma_{{\rm{d}},k}}} s_{k,n} c
+{\bf{u}}_n\\
&= \sum_{k=1}^K {{{\bf{h}}_{k,1}}} s_{k,n} +\sum_{k=1}^K {{{\bf{h}}_{k,2}}} s_{k,n} c + {\bf{u}}_n
,
\end{aligned}
\end{equation}
where ${\bf{y}}_n=[y_{1,n},\ldots,y_{M,n}]^T$ and ${\bf{u}}_n=[u_{1,n},\ldots,u_{M,n}]^T$.
\subsection{Frame Structures and Differential Modulation}
Finally, we adopt a backscatter symbol frame as depicted in {\figurename~\ref{c_clusterframe}}, where each frame consists of $I$ original BD's information bits ${\bf{b}}=[b^{(1)}, b^{(2)},\cdots ,b^{(I)}]$, and $b^{(i)} \in \left\{0,1\right\}$ for all $i=1,2,\cdots,I$. Then the original bits are differentially encoded at the BD as follows
\begin{equation}\label{equ:diffmodulation}
c^{(i)}=c^{(i-1)} \oplus b^{(i)},
\end{equation}
where ${\bf{c}}=[c^{(1)}, c^{(2)},\cdots ,c^{(I)}]$ are the modulated symbols to be transmitted with the reference symbol $c^{(0)}=1$.
According to the block fading assumption, the channels remain invariant during one frame period.
Each BD symbol contains $N$ RF source symbols.
The $n$-th received sample in the $i$-th BD symbol period is
\begin{equation} \label{equ:2_signalvi}
\begin{aligned} [b]
{\bf{y}}_n^{(i)}=\sum_{k=1}^K {\bf{h}}_{k,1} s_{k,n}^{(i)}+ c^{(i)} \sum_{k=1}^K {\bf{h}}_{k,2} s_{k,n}^{(i)} +{\bf{u}}_n^{(i)},
\end{aligned}
\end{equation}
where $n=1,2,\ldots,N$ and $i=0,1,2,\ldots,I$.
Let ${\bf{Y}}^{(i)}=[{\bf{y}}_1^{{(i)}^T} ,{\bf{y}}_2^{{(i)}^T} ,\ldots,{\bf{y}}_N^{{(i)}^T}]^T$ denote the received signal sequence in the $i$-th symbol period.
Our goal is to recover all the $b^{(i)}$ from the observed ${\bf{Y}}^{(i)}$.
\begin{figure}
[!t]
\centering
\includegraphics[width=.99\columnwidth]{packet_structure-eps-converted-to.pdf}
\caption{Frame structures of involving symbols.
}
\vspace{-1.5em}
\label{c_clusterframe}
\end{figure}
\section{The Impact Of the DLI}\label{problem}
In CBS, the relationship between the transmit powers of the RF sources and the achievable rate of the backscatter system is complicated.
On the one hand, the RF sources provide carrier emitters for the BD transmission. Thus, like the traditional BackCom, a strong RF source signal will improve the backscatter transmission.
On the other hand, due to the spectrum sharing, the RF sources simultaneously cause interference to the backscatter system. This is a little like the multiuser MIMO system, where the strong interferences are harmful for the backscatter transmission.
As a result, the CBS model is much different with the conventional communication scenarios.
It is worth to process some analyses to provide an insight about the impact of RF sources on the backscatter transmission.
\subsection{Exploiting Multi-antennas for High-speed Transmission}
In CBS detection, a well-known phenomenon for the single-antenna receiver is the ``error floor'': the BER for backscatter symbol detection converges to a non-zero floor even though the transmit power of the RF sources increases to infinite \cite{GaoFF2017NDAmBC,Guo2018IoT}.
Due to that, given a BER requirement, the transmission rate is limited.
In this subsection, we try to provide an explanation on the rate limitation of single-antenna receiver from the information theoretic perspective.
Then, we will show that to exploit multiple receive antennas can realize high-speed backscatter transmission.
For simplicity, we set $N=1$, and the efficiency loss of the differential modulation is also ignored as well.
\subsubsection{Single-antenna Receiver}
When there is only one RF source and the receiver is equipped with single antenna, the received signal at the reader is
\begin{equation} \label{equ:2_oneantenna}
{{y}} =l g \sqrt{\widetilde \gamma {\gamma_{\rm{d}}}} s c +f \sqrt{{\gamma_{\rm{d}}}} s+u.
\end{equation}
The maximum achievable rate is $R_{\{M=1\}}={\max_{p(c)}}{\rm I}(c;{{y}})$.
As shown in \eqref{equ:2_oneantenna}, the backscatter symbol $c$ is corrupted by the unknown source signal $s$.
Due to the unknown $s$, ${R}_{\{M=1\}}$ is hard to be derived.
We consider a scenario, where the BD is allowed to exploit $s$ to design the transmit codewords. Then \eqref{equ:2_oneantenna} becomes
\begin{equation} \label{equ:2_oneanew}
y =l g \sqrt{\widetilde \gamma {\gamma_{\rm{d}}}} \widetilde{{c}} +f \sqrt{{\gamma_{\rm{d}}}} s+u,
\end{equation}
where $\widetilde{{c}}$ is the new transmitted symbol by BD which drops the interference of $s$ to achieve the optimal rate.
Denote the achievable rate in this scenario as ${\widetilde{R}}_{\{M=1\}}$, and we have
\begin{equation} \label{equ:R_oneantenna}
{\widetilde{R}}_{\{M=1\}}
= \log_2 (1+\frac{\widetilde \gamma \gamma_{\rm{d}}|l g|^2}{1+\gamma_{\rm{d}}|f|^2})
.
\end{equation}
Obviously, ${\widetilde{R}}_{\{M=1\}}$ is an upper bound of ${R}_{\{M=1\}}$ (i.e., ${R}_{\{M=1\}}\leq {\widetilde{R}}_{\{M=1\}}$).
Then we have following proposition:
\begin{mypro}\label{pro1}
When there is only one RF source and the receiver is equipped with single antenna, the achievable rate converges to a finite value with the increase of $\gamma_{\rm{d}}$.
\end{mypro}
\begin{IEEEproof}
Let ${\widetilde{R}}_{\{M=1\}}^{\infty}={\lim_{\gamma_{\rm{d}} \rightarrow \infty}} {\widetilde{R}}_{\{M=1\}}$ denote the extreme rate.
Then according to \eqref{equ:R_oneantenna}, we obtain
\begin{equation} \label{equ:Rup_oneantenna}
\begin{aligned} [b]
{\widetilde{R}}_{\{M=1\}}^{\infty}&= \lim_{\gamma_{\rm{d}}\rightarrow +\infty} \log_2 (1+\frac{\widetilde \gamma \gamma_{\rm{d}}|l g|^2}{1+\gamma_{\rm{d}}|f|^2})\\
&= \log_2 (1+\frac{\widetilde \gamma |l g|^2}{|f|^2})
.
\end{aligned}
\end{equation}
Therefore, ${\widetilde{R}}_{\{M=1\}}^{\infty}$ is dominated by the relative SNR $\widetilde \gamma$, and obviously, ${\widetilde{R}}_{\{M=1\}}^{\infty}<+\infty$.
\end{IEEEproof}
In CBS, $\widetilde \gamma$ is very small (less than $-20$ dB), and ${\widetilde{R}}_{\{M=1\}}^{\infty}$ is quiet limited due to the existence of the DLI. That is the key reason of the error floor phenomenon.
As a result, to suppress the DLI is the critial task to realize high-speed backscatter transmission.
\subsubsection{Multi-antenna Receiver}
To realize high-speed backscatter transmission, we exploit the receiving diversity in the spatial domain with the multi-antenna receiver ($M\geq 2$). Suppose that there are $K$ RF sources, according to \eqref{equ:2_signalv}, the received signal at the receiver is
\begin{equation} \label{equ:Mantenna}
{\bf{y}}
={\bf{g}} \sum_{k=1}^K l_k \sqrt{\widetilde \gamma_k {\gamma_{{\rm{d}},k}}} s_k c +\sum_{k=1}^K {{\bf{f}}_k} \sqrt{{\gamma_{{\rm{d}},k}}} s_k+{\bf{u}}.
\end{equation}
The theoretical maximum achievable rate is denoted as $R_{\{M,K\}}={\max_{p(c)}}{\rm I}(c;{\bf{y}})$.
As the $K$ RF sources are mutually independent, we assume ${{\bf{f}}_1}$, ${{\bf{f}}_2}$, $\cdots$, ${{\bf{f}}_K}$ and ${\bf{g}}$ are linearly independent, i.e., the equation $a_1 {{\bf{f}}_1}+a_2 {{\bf{f}}_2}+\cdots+a_K {{\bf{f}}_K}+a_{K+1} {\bf{g}}={\bf{0}}$
is only satisfied by $a_1=a_2=\cdots=a_{K+1}=0$.
Then we have following proposition:
\begin{mypro}\label{pro2}
When the receive antenna number is larger than $K$, i.e., $M>K$, infinite $R_{\{M,K\}}$ can be achieved as long as the transmit power of one RF source (${\gamma_{{\rm{d}},k_0}}$) approaches infinity.
\end{mypro}
\begin{IEEEproof}
See Appendix \ref{M_larger_K}.
\end{IEEEproof}
\subsection{Chernoff Information for the On-off Modulation}
To achieve low implementation cost, the BD usually only has two backscatter states, i.e., $c$ employs the on-off modulation.
However, the close-form expression of the maximum achievable rate in this scenario is unavailable (A numerical method is shown in Appendix \ref{A_onoff_rate}).
In this subsection, we resort to the CI as a tractable metric to predict the optimal detection performance, which is the maximum achievable error exponent \cite[eq 11.230]{cover2012elements}:
\begin{equation} \label{equ:chernoff}
D= \max \lim_{N\rightarrow \infty} -\frac{1}{N} \ln {P_{\rm e}},
\end{equation}
where $P_{\rm e}$ is the decision BER.
According to \cite[eq 11.239]{cover2012elements}, the standard definition of CI is
\begin{equation} \label{equ:Dchernoff}
D= - \min_{0\leq u \leq 1} \ln\left( {p^{u}({\bf{y}} | c=1 )} {p^{1-u}({\bf{y}} | c=0 )} \right),
\end{equation}
where ${p({\bf{y}} | c=1 )}$ and ${p({\bf{y}} | c=0 )}$ are the conditional \emph{probability density functions} (PDFs) given $c$.
Since both the RF source signals $s_k$ and the noise $\bf u$ follow CSCG distribution, we have:
\begin{align}
{p({\bf{y}} | c=1 )}
&= \frac{1}{\pi^M |{\bf{C}}_1|} e^{
- {{\bf{y}}}^H {\bf{C}}_1^{-1} {\bf{y}}
} \label{equ:Gpdfc1}
,\\
{p({\bf{y}} | c=0 )}
&= \frac{1}{\pi^M |{\bf{C}}_0|} e^{
- {{\bf{y}}}^H {\bf{C}}_0^{-1} {\bf{y}}
}, \label{equ:Gpdfc0}
\end{align}
where ${\bf{C}}_1$ and ${\bf{C}}_0$ are the channel statistical covariance matrices:
\begin{align}
{\bf{C}}_1&=\sum_{k=1}^K ({{\bf{h}}_{1,k}+{\bf{h}}_{2,k}}) ({{\bf{h}}_{1,k}+{\bf{h}}_{2,k}})^H+ {\bf{I}}_{M}. \label{equ:C1}\\
{\bf{C}}_0&=\sum_{k=1}^K {{\bf{h}}_{1,k}} {{\bf{h}}_{1,k}^H}+ {\bf{I}}_{M}, \label{equ:C0}
\end{align}
Let $G(u)={p^{u}({\bf{y}} | c=1 )} {p^{1-u}({\bf{y}} | c=0 )}$, where $u \in [0,1]$.
According to \eqref{equ:Dchernoff}, the CI is $D= - \min_{0\leq u \leq 1} \ln\left( G(u) \right)$.
Substituting \eqref{equ:Gpdfc1} and \eqref{equ:Gpdfc0} into $G(u)$, we have
\begin{equation} \label{equ:funcG}
\ln G(u)= \ln\left(\left|{\bf{K}}(u)\right|\right)-u \ln (|{\bf{C}}_1|)-(1-u) \ln (|{\bf{C}}_0|),
\end{equation}
where
\begin{equation}
{\bf{K}}^{-1}(u)=u {\bf{C}}_1^{-1}+(1-u) {\bf{C}}_0^{-1}.
\end{equation}
Based on the matrix differentiation identities:
\begin{align}
\frac{\rm d}{{\rm d}u} \ln\left(\left|{\bf{K}}(u)\right|\right)&={\rm tr} \left[{\bf{K}}^{-1}(u) \frac{{\rm d}{\bf{K}}}{{\rm d}u} (u)\right],
\\
\frac{\rm d}{{\rm d}u} {\bf{K}}^{-1}(u)&=-{\bf{K}}^{-1}(u) \frac{{\rm d}{\bf{K}}}{{\rm d}u} (u) {\bf{K}}^{-1}(u),
\end{align}
we have:
\begin{equation}
\frac{\rm d}{{\rm d}u} \ln\left(G(u)\right)=-{\rm tr} \left({\bf{K}}(u)\left({\bf{C}}_1^{-1}-{\bf{C}}_0^{-1}\right)\right)-\ln{\frac{|{\bf{C}}_1|}{|{\bf{C}}_0|}}.
\end{equation}
Solving $\frac{\rm d}{{\rm d}u} \ln\left(G(u^\ast)\right)=0$, CI is obtained
\begin{equation}
D= - \ln\left( G(u^\ast) \right).
\end{equation}
Further discussions will be presented in Section \ref{OOK_rate_sim}, which illustrates the impact of $K$ and $M$ on the CI.
\section{Maximum Likelihood Detectors}\label{coordinated}
In this section, we investigate on the backscatter signal detection for the multi-antenna receiver based on the ML criterion.
Based on \emph{\textbf{Proposition}} \ref{pro2}, we assume that $M>K$ to achieve reliable performance.
The backscatter system spanning $N$ consecutive source symbols to make sure that the BD transmission causes no negative impact on the legacy system \cite{YangG2017CooperativeBC,RZLong2017beamform}.
\subsection{Soft Messages of ${\bf{c}}$}
From the frame structures shown in {\figurename~\ref{c_clusterframe}}, the symbol vector ${\bf{c}}$ is the sufficient statistic of the original signal vector ${\bf{b}}$.
Hence, recovering ${\bf{c}}$ is the first step to decode ${\bf{b}}$.
During one symbol period, $c^{(i)}$ is equal to ``$0$'' or ``$1$''. As a result, to recover $c^{(i)}$ is equivalent to distinguish whether the backscatter-link signal is present or absent:
\begin{equation} \label{equ:3_sensingh}
{\bf{y}}_n^{(i)}=
\begin{cases}
\sum_{k=1}^K {\bf{h}}_{k,1} s_{k,n}^{(i)}+{\bf{u}}_n^{(i)}, & {\rm{if}} ~~ c^{(i)} = 0,
\\
\sum_{k=1}^K ({\bf{h}}_{k,1}+{\bf{h}}_{k,2}) s_{k,n}^{(i)}+{\bf{u}}_n^{(i)}, & {\rm{if}} ~~ c^{(i)} = 1.
\end{cases}
\end{equation}
where $n=1,2,\ldots,N$, and $i=1,2,\ldots,I$.
As we know, ${\bf{y}}_n^{(i)}$ is a CSCG distributed vector with conditional PDFs given $c^{(i)}$:
\begin{align}
{p({\bf{y}}_n^{(i)} | c^{(i)}=0 )}
&= \frac{1}{\pi^M |{\bf{C}}_0|} e^{
- {{\bf{y}}_n^{(i)}}^H {\bf{C}}_0^{-1} {\bf{y}}_n^{(i)}
}, \label{equ:3_Gpdfc0}\\
{p({\bf{y}}_n^{(i)} | c^{(i)}=1 )}
&= \frac{1}{\pi^M |{\bf{C}}_1|} e^{
- {{\bf{y}}_n^{(i)}}^H {\bf{C}}_1^{-1} {\bf{y}}_n^{(i)}
} \label{equ:3_Gpdfc1}
.
\end{align}
Then the likelihood function of the received signal sequence ${\bf{Y}}^{(i)}=[{\bf{y}}_1^{{(i)}^T} ,{\bf{y}}_2^{{(i)}^T} ,\ldots,{\bf{y}}_N^{{(i)}^T}]^T$ is
\begin{equation} \label{equ:LFY}
{\mathcal{L}}{({\bf{Y}}^{(i)} | c^{(i)} )}= \prod_{n=1}^N {p({\bf{y}}_n^{(i)} | c^{(i)} )}.
\end{equation}
Substituting \eqref{equ:3_Gpdfc0} and \eqref{equ:3_Gpdfc1} into \eqref{equ:LFY}, we have
\begin{align}
{\mathcal{L}}{({\bf{Y}}^{(i)} | c^{(i)}=0 )}
&= \prod_{n=1}^N {
\frac{1}{\pi^M |{\bf{C}}_0|} e^{
- {{\bf{y}}_n^{(i)}}^H {\bf{C}}_0^{-1} {\bf{y}}_n^{(i)}
}} \label{equ:YLH0}
,\\
{\mathcal{L}}{({\bf{Y}}^{(i)} | c^{(i)}=1 )}
&= \prod_{n=1}^N{
\frac{1}{\pi^M |{\bf{C}}_1|} e^{
- {{\bf{y}}_n^{(i)}}^H {\bf{C}}_1^{-1} {\bf{y}}_n^{(i)}
}} \label{equ:YLH1}
.
\end{align}
The likelihood functions \eqref{equ:YLH0} and \eqref{equ:YLH1} contain the whole information of $c^{(i)}$ inferred from ${\bf{Y}}^{(i)}$, which is referred to as the ``soft message'' of $c^{(i)}$.
\subsection{ML Detector for the Original Symbols ${\bf{b}}$}\label{ML_detection}
From \eqref{equ:diffmodulation}, the BD encodes the original symbol $b^{(i)}$ via differential modulation.
Therefore, $c^{(i)}$ and $c^{(i-1)}$ are the sufficient statistics of $b^{(i)}$. In the following, we will design the ML detectors for $b^{(i)}$ based on the soft messages of $c^{(i)}$ and $c^{(i-1)}$, respectively.
\subsubsection{Soft Decision}
Define ${\bf{z}}^{(i)}=[{\bf{Y}}^{(i-1)};{\bf{Y}}^{(i)}]$.
According to \eqref{equ:diffmodulation}, the likelihood functions of ${\bf{z}}^{(i)}$ conditional on $b^{(i)}$ are expressed as follows based on the soft messages of $c^{(i)}$ and $c^{(i-1)}$:
\begin{equation} \label{equ:LFYb0}
\begin{aligned} [b]
{\mathcal{L}}{({\bf{z}}^{(i)} | b^{(i)}=0 )}&= {\mathcal{L}}{({\bf{z}}^{(i)} | c^{(i-1)}=c^{(i)} )}
\\
&={\mathcal{L}}{({\bf{Y}}^{(i-1)} | c^{(i-1)}=0 )} {\mathcal{L}}{({\bf{Y}}^{(i)} | c^{(i)}=0 )}\\
&\quad +{\mathcal{L}}{({\bf{Y}}^{(i-1)} | c^{(i-1)}=1 )} {\mathcal{L}}{({\bf{Y}}^{(i)} | c^{(i)}=1 )}
,
\end{aligned}
\end{equation}
\begin{equation} \label{equ:LFYb1}
\begin{aligned} [b]
{\mathcal{L}}{({\bf{z}}^{(i)} | b^{(i)}=1 )}&= {\mathcal{L}}{({\bf{z}}^{(i)} | c^{(i-1)}\neq c^{(i)} )}
\\
&={\mathcal{L}}{({\bf{Y}}^{(i-1)} | c^{(i-1)}=0 )} {\mathcal{L}}{({\bf{Y}}^{(i)} | c^{(i)}=1 )}\\
&\quad +{\mathcal{L}}{({\bf{Y}}^{(i-1)} | c^{(i-1)}=1 )} {\mathcal{L}}{({\bf{Y}}^{(i)} | c^{(i)}=0 )}
.
\end{aligned}
\end{equation}
Then, the ML detector for the original symbols $b^{(i)}$ is expressed as follows:
\begin{equation}\label{equ:soft}
{\mathcal{L}}{({\bf{z}}^{(i)} | b^{(i)}=0 )}
\mathop{\gtrless}_{{\hat b}^{(i)}=1}^{{\hat b}^{(i)}=0}
{\mathcal{L}}{({\bf{z}}^{(i)} | b^{(i)}=1 )}
,
\end{equation}
where ${\hat b}^{(i)}$ is the decision result.
The ML detector in \eqref{equ:soft} is usually referred to as ``\emph{soft decision}'' (SD) method.
\subsubsection{Hard Decision}
In some scenarios, a two-step ML decision method are applied instead of the SD method in \eqref{equ:soft} for simplicity.
In the first step, the ``\emph{hard decision}'' (HD) of $c^{(i)}$ is obtained based on the ML criterion:
\begin{equation}\label{equ:hardmc}
{\mathcal{L}}{({\bf{Y}}^{(i)} | c^{(i)}=0 )}
\mathop{\gtrless}_{{\hat c}^{(i)}=1}^{{\hat c}^{(i)}=0}
{\mathcal{L}}{({\bf{Y}}^{(i)} | c^{(i)}=1 )}
,
\end{equation}
Substituting \eqref{equ:YLH0} and \eqref{equ:YLH1} into \eqref{equ:hardmc}, we obtain the ML detector of the transmitted symbol $c^{(i)}$ as follows:
\begin{equation}\label{equ:3_opt_CSCGv}
\sum_{n=1}^N
{{\bf{y}}_n^{(i)}}^H \left({\bf{C}}_0^{-1}-{\bf{C}}_1^{-1}\right) {\bf{y}}_n^{(i)}
\mathop{\gtrless}_{{\hat c}^{(i)}=0}^{{\hat c}^{(i)}=1}
N \ln \frac {|{\bf{C}}_1|} {|{\bf{C}}_0|}
,
\end{equation}
where $[{\hat c}^{(1)}, {\hat c}^{(2)},\cdots ,{\hat c}^{(I)}]$ are the decision results of $\bf c$.
In the second step, the decision result ${\hat b}^{(i)}$ is derived based on the differential modulation relationship shown in \eqref{equ:diffmodulation}:
\begin{equation} \label{equ:hard}
{\hat b}^{(i)}={\hat c}^{(i-1)} \oplus {\hat c}^{(i)}.
\end{equation}
When $M=1$, the ML detector for $c^{(i)}$ in \eqref{equ:3_opt_CSCGv} is equivalent to the ED, and thus the above HD method is also known as the joint-ED in \cite{GaoFF2017NDAmBC}.
\section{Clustering: Blind Channel Estimation for Practical Implementation}\label{semi-blind}
In section \ref{coordinated}, we have designed the optimal SD and suboptimal HD detectors with the knowledge of CSI (i.e., ${\bf{C}}_0$ and ${\bf{C}}_1$).
In conventional communication system, the CSI is usually estimated with the aid of pilot symbols.
Nonetheless, in CBS, the backscatter link is very weak ($\widetilde \gamma$ is generally smaller than $-20$ dB), so that a long pilot sequence is need which may incur severer efficiency loss \cite{GaoFF2017NDAmBC}.
Thus, channel estimation is another critical challenge for practical implementation, especially without the cooperation with the legacy system.
In this section, we try to propose a fully blind channel estimation method, which do no need any pilot symbols to achieve high transmission efficiency.
Based on that the BD transmits binary modulated signals, we suggest a fully blind channel estimation method via clustering \cite{Hastie2009slearn}.
Our intuition is summarized below:
\begin{itemize}
\item According to \eqref{equ:3_sensingh}, the received signals $\left\{{\bf{Y}}^{(i)}\right\}^I_{i=1}$ can be grouped into two clusters since they come from two different distributions (see equations \eqref{equ:YLH0} and \eqref{equ:YLH1}).
\item Thanks to the differential modulation, we actually do not need to map the two groups to the transmission states ``$c=0$'' and ``$c=1$''.
The CSI for each group can be carried out by combining all the received samples belonging to the same cluster.
\end{itemize}
The details of the clustering method will be presented in the next three subsections.
\subsection{Feature Extraction}\label{semi-blind-feature}
For a single antenna receiver, the energy of received symbol is a good feature to group all the $\left\{{\bf{Y}}^{(i)}\right\}^I_{i=1}$ into two clusters. Based on that, clustering is realized through a simple sorting algorithm, in which the first half and the second half of energy levels are classified into two groups, respectively \cite{{GaoFF2017NDAmBC}}.
However, when the reader has multiple antennas, this is a challenge task due to the high dimension of ${\bf{Y}}^{(i)}$.
To extract a proper feature from ${\bf{Y}}^{(i)}$ is the first key task for clustering.
Denote the channel statistical covariance matrix corresponding to ${\bf{Y}}^{(i)}$ as ${\bf{C}}^{(i)}$.
Then we have
\begin{equation}
{\bf{C}}^{(i)}=
\begin{cases}
{\bf{C}}_0, & {\rm{if}} ~~ c^{(i)} = 0,
\\
{\bf{C}}_1, & {\rm{if}} ~~ c^{(i)} = 1.
\end{cases}
\end{equation}
The ML estimation of ${\bf{C}}^{(i)}$ is the sample covariance matrix:
\begin{equation} \label{equ:samplecm}
{{\bf{R}}}^{(i)}=\frac{1}{N} \sum_{n=1}^{N}
{
{\bf{y}}_n^{(i)} {{\bf{y}}_n^{(i)}}^H
}
,
\end{equation}
which contains all the useful information of ${\bf{Y}}^{(i)}$.
In most cases, ${{\bf{R}}}^{(i)}$ is accurate enough.
Nevertheless, the estimation accuracy can be further improved by exploiting the spatial sparsity of ${\bf{C}}^{(i)}$, which will be significant when the antenna number $M$ is much larger than $K$.
Let us divide ${\bf{C}}^{(i)}$ into two components: the signal component ${{\bf{C}}}_{s}$ and the noise component ${{\bf{C}}}_u$. Since the noise follows the standard CSCG distribution, ${{\bf{C}}}_u={\bf{I}}_{M}$. Then we have
\begin{equation}\label{equ:estC}
{\bf{C}}^{(i)}={{\bf{C}}}_{s}+{\bf{I}}_{M}
.
\end{equation}
For different transmit symbols ($c=0$ or $c=1$), ${{\bf{C}}}_{s}$ is expressed as follows
\begin{align}
{\bf{C}}_{s,0}&=\sum_{k=1}^K {{\bf{h}}_{1,k}} {{\bf{h}}_{1,k}^H}, \label{equ:Cs0} \\
{\bf{C}}_{s,1}&=\sum_{k=1}^K ({{\bf{h}}_{1,k}+{\bf{h}}_{2,k}}) ({{\bf{h}}_{1,k}+{\bf{h}}_{2,k}})^H. \label{equ:Cs1}
\end{align}
Thus ${\rm{rank}}({{\bf{C}}}_{s})=K$, and there are some redundant elements in ${{\bf{C}}}_{s}$ when the the number of the dominated RF sources is smaller than $M-1$.
Consequently, a better estimation of ${{\bf{C}}}_{s}$ is:
\begin{equation}\label{equ:clusterRF}
{{\bf{R}}}^{(i)}_{s}= \sum_{k=1}^K ({\rm \lambda}_{k}-1) {\bf{w}}_{k} {\bf{w}}_{k}^H
,
\end{equation}
where ${\rm \lambda}_{k}$ is the $k$-th largest eigenvalue of ${{\bf{R}}}^{(i)}$, and ${\bf{w}}_{k}$ is the corresponding eigenvector.
Combining \eqref{equ:estC} and \eqref{equ:clusterRF}, we have a better estimation of the channel covariance matrix, which will be employed as the features for clustering:
\begin{equation}\label{equ:improvedesR}
{\hat{\bf{R}}}^{(i)}= {{\bf{R}}}^{(i)}_{s}+{\bf{I}}_{M}.
\end{equation}
\subsection{Distance Design}
After feature extraction, the second key task for clustering is to measure the difference between features ${\hat{\bf{R}}}^{(i)}$ and ${\hat{\bf{R}}}^{(j)}$, which is referred to as the ``distance'' in the statistical machine learning.
Based on ${\hat{\bf{R}}}^{(i)}$, we define the pseudo-PDF as follows
\begin{equation}\label{equ:pPDF}
{\hat{p}}{({\bf{Y}} | {\hat{\bf{R}}}^{(i)} )}
= \prod_{n=1}^N {
\frac{1}{\pi^M |{\hat{\bf{R}}}^{(i)}|} e^{
- {\bf{y}}_n^H ({{\hat{\bf{R}}}^{(i)}})^{-1} {\bf{y}}_n
}},
\end{equation}
where ${\bf{Y}}=[{\bf{y}}_1^T,{\bf{y}}_2^T,\ldots,{\bf{y}}_n^T]^T$ denotes the received signal sequence.
It is known that, the relative entropy between two $M$-dimension zero-mean multivariate CSCG distributions with covariance matrices ${\bf \Sigma}_0$ and ${\bf \Sigma}_1$ is
\begin{equation}\label{equ:KL}
D_{\rm KL} ({\bf \Sigma}_0 \parallel {\bf \Sigma}_1)=\frac{1}{2}\left(
{\rm{tr}}\left({\bf \Sigma}_1^{-1} {\bf \Sigma}_0\right)-M+\ln \frac {|{\bf \Sigma}_1|} {|{\bf \Sigma}_0|}
\right)
.
\end{equation}
Based on \eqref{equ:KL}, we propose a new distance for clustering, which is defined by
\begin{equation}\label{equ:distance}
\begin{aligned} [b]
d_{ij}&= \left(D_{\rm KL} ({\hat{\bf{R}}}^{(i)} \parallel {\hat{\bf{R}}}^{(j)})+D_{\rm KL} ({\hat{\bf{R}}}^{(j)} \parallel {\hat{\bf{R}}}^{(i)})\right)\\
&=\frac{1}{2}{\rm{tr}}\left(({\hat{\bf{R}}}^{(i)})^{-1} {\hat{\bf{R}}}^{(j)}+({\hat{\bf{R}}}^{(j)})^{-1} {\hat{\bf{R}}}^{(i)}\right)-M
.
\end{aligned}
\end{equation}
Obviously, the distance in \eqref{equ:distance} satisfies $d_{ii}=0$ and $d_{ij}=d_{ji}$.
\subsection{Clustering}
Based on the feature ${\hat{\bf{R}}}^{(i)}$ and the distance metric $d_{ij}$, we present the clustering algorithm.
For clustering, the key task is to find the cluster centroids according to a given distance metric.
We suggest a two-step centroid search method based on the distance defined in \eqref{equ:distance}:
\begin{itemize}
\item Firstly, we randomly choose one element ${\hat{\bf{R}}}^{(i)}$, then the first cluster centroid is ${\hat{\bf{R}}}^{(a_0)}$ where $a_0= \arg \max_{j} d_{ij}$.
\item Similarly, the second cluster centroid is ${\hat{\bf{R}}}^{(a_1)}$ where $a_1= \arg \max_{j} d_{a_0 j}$.
\end{itemize}
Using ${\hat c}^{(i)}=0$ to indicate that ${\hat{\bf{R}}}^{(i)}$ belongs to the first cluster, and ${\hat c}^{(i)}=1$ to indicate that ${\hat{\bf{R}}}^{(i)}$ belongs to the second cluster, the clustering is implemented as follow:
\begin{equation}\label{equ:clustering}
d_{a_0 i}
\mathop{\gtrless}_{{\hat c}^{(i)}=0}^{{\hat c}^{(i)}=1}
d_{a_1 i}
.
\end{equation}
Based on the clustering results, all the received symbols in the same cluster are combined to realize the estimation of the channel statistical covariance matrices:
\begin{align}
\hat{\bf{C}}_0 &=\frac{1}{\sum_{i=1}^I (1-{\hat c}^{(i)})} \sum_{i=1}^I (1-{\hat c}^{(i)}) \hat{\bf{R}}^{(i)}, \label{equ:clusterR0}\\
\hat{\bf{C}}_1 &=\frac{1}{\sum_{i=1}^I {\hat c}^{(i)}} \sum_{i=1}^I {\hat c}^{(i)} \hat{\bf{R}}^{(i)}. \label{equ:clusterR1}
\end{align}
Subsequently, signal detection can be executed based on the SD or HD in Section \ref{coordinated}.
\section{Numerical Results}\label{simulation}
In this section, we resort to numerical examples to evaluate the multi-antenna receiver schemes.
All the channels are assumed experience independent Rayleigh fading, and for simplicity, a homogenous scenario is assumed in which all the $\widetilde \gamma_k$ and $\gamma_{{\rm{d}},k}$ corresponding to different RF sources are the same.
Denote the summation of the incident direct-link power by $\gamma_{{\rm{d}}}=\sum \gamma_{{\rm{d}},k}$, and the relative SNR by $\widetilde \gamma=\widetilde \gamma_k$.
We first continue the discussion in Section \ref{problem} on the impact of multiple RF sources and multiple receive antennas.
Then, the BER performance of all the proposed detectors will be illustrated.
Finally, the effectiveness of the proposed channel estimation method will be evaluated.
\subsection{Achievable Rate and CI for Different $K$ and $M$}\label{OOK_rate_sim}
{\figurename~\ref{ergodic_rate}} illustrates the relationship between the average throughput $\overline{R}_{\rm o}$ and the system parameters, such as the prior probability of transmit symbol ``0'' (i.e., $\phi_0$), antenna number $M$, and amount of RF sources $K$.
The average throughput $\overline{R}_{\rm o}$ is obtained through Monte Carlo integration according to Appendix \ref{A_onoff_rate}.
It is noticed that, for all the combination cases of $M$ and $K$ considered in {\figurename~\ref{backscatter model}}, $\phi_0=0.5$ always achieves the maximum average throughput. Thus to transmit equiprobable symbols is optimal.
In addition, one can also see the significant improvement of $\overline{R}_{\rm o}$ when the condition $M>K$ is satisfied.
For the $M=K$ cases, the average throughput still increases slightly when $M$ becomes larger.
In the rest numerical discussions, we set BD transmits ``0'' and ``1'' with equal probability.
{\figurename~\ref{CI_vs_rd}} plots the CI (i.e., $D$ in the $Y$-axis) for different combinations of $M$ and $K$ with respect to the direct link SNR $\gamma_{\rm{d}}$ when $\widetilde \gamma=-25$ dB, when BD employs on-off modulation.
It is seen that, if $M>K$ is satisfied, $D$ increases as the $\gamma_{\rm{d}}$ increases.
However, when $M=K$, $D$ approaches to an finite upper bound with the increase of $\gamma_{\rm{d}}$.
Therefore, when $\gamma_{\rm{d}}$ is already high enough, to increase $\gamma_{\rm{d}}$ contributes little to the detection performance for the $M=K$ cases.
In addition, we observes an interesting phenomenon that, the CI of $\{M=4, K=2\}$ is high than that of $\{M=4, K=1\}$ when $\gamma_{\rm{d}}>35$ dB.
Hence, a larger $K$ does not necessarily implies poorer performance.
This is because that, although the multiple RF sources consume the degree of freedom for DLI cancelation, they also provides more carrier emitter energy which benefits the backscatter transmission.
Furthermore, it is shown that, the CI of $\{M=4, K=3\}$ is slightly better than that of $\{M=2, K=1\}$, although the rate $\frac{K}{M}$ of the $\{M=4, K=3\}$ cases is larger.
\begin{figure}
[!t]
\centering
\includegraphics[width=.99\columnwidth]{bits_diff_p-eps-converted-to.pdf}
\caption{The average throughput versus $\phi_0$ (the prior probability that $c=0$) when the BD adopts OOK, $\gamma_{\rm d}=40$ dB, and $\widetilde \gamma=-25$ dB.
}
\label{ergodic_rate}
\end{figure}
\begin{figure}
[!t]
\centering
\includegraphics[width=.99\columnwidth]{CI-eps-converted-to.pdf}
\caption{The CI versus $\gamma_{\rm d}$ when the BD transmits ``0'' and ``1'' with equal probability, and $\widetilde \gamma=-25$ dB.
}
\label{CI_vs_rd}
\end{figure}
\subsection{BER Performance}
In this subsection, we evaluate the effectiveness of the proposed detectors by simulation.
The backscatter frame length is fixed to $I=100$. In addition, $10^8$ Monte-Carlo runs are carried out to achieve reliable results.
\begin{figure}
[!t]
\centering
\includegraphics[width=.99\columnwidth]{BER_rd_M_2-eps-converted-to.pdf}
\caption{The BER versus $\gamma_d$, when $N=50$, $\widetilde \gamma = -25$ dB and $I=100$, for $M=2$ and $M=1$, respectively.
}
\label{BER_M2}
\end{figure}
\begin{figure}
[!t]
\centering
\includegraphics[width=.99\columnwidth]{BER_rd_M_4-eps-converted-to.pdf}
\caption{The BER versus $\gamma_d$, when $N=50$, $\widetilde \gamma = -25$ dB, $I=100$, and $M=4$, for different $K$.
}
\label{BER_M4}
\end{figure}
{\figurename~\ref{BER_M2}} illustrates the BER performance versus the direct link SNR $\gamma_{\rm{d}}$ for different detectors with $N=50$ and $\Delta \gamma = -25$ dB.
Specifically, the BERs of the optimal detector (i.e., SD) and suboptimal detector (i.e., HD) with perfect and estimated CSI are displayed, respectively.
It is seen that the HD achieves the same BER performance as that of the SD.
Besides, the detectors with estimated CSI suffer only a very small performance loss compared to the perfect CSI cases.
In addition, the BER meet the error floor when $M=K$.
However, when $M=2$ and $K=1$, the BER decreases significantly with the increase of $\gamma_{\rm{d}}$. Therefore, the error floor problem is efficiently eliminated.
{\figurename~\ref{BER_M4}} illustrates the BER performance versus $\gamma_{\rm{d}}$ for different $K$ when $M=4$ and $N=50$, $\Delta \gamma = -25$ dB.
The HD and SD still achieve almost the same performance.
Therefore, the performance loss caused by channel estimation is very small.
It is noticed that the error floor problem is existed when $K=4$, and it is overcome when $K\leq 3$.
Furthermore, when $K=1$ and $K=2$, the detectors achieve similar performance.
When $\gamma_{\rm{d}}$ increases, the BER performance of $K=2$ becomes better than that of $K=1$ cases, which is consistent with the prediction from {\figurename~\ref{CI_vs_rd}}.
{\figurename~\ref{BER_N}} plots the BER performance of the proposed detectors for different $N$, when $\gamma_{\rm{d}}=25$ dB and $\Delta \gamma = -20$ dB.
It is noticed that the performance of all detectors improves with the increase of $N$ at the cost of lower transmission rate.
For a target BER of $10^{-2}$, when $M=3$ and $K=2$, each backscatter symbol needs to span less than $N=20$ RF source symbols.
When $M=2$ and $K=1$, the backscatter symbol needs to span about $N=50$ RF source symbols.
However, when $M=1$ and $K=1$, the BER is unacceptable even when $N=500$.
Therefore, the transmission rate can be improved at least $10$ times, when multiple antennas is exploited and $M>K$ is satisfied.
In addition, we observe that the BER performance of $\{M=3,K=2\}$ is better than that of $\{M=2,K=1\}$, although the rate $\frac{K}{M}$ is increased.
\begin{figure}
[!t]
\centering
\includegraphics[width=.99\columnwidth]{BER_N-eps-converted-to.pdf}
\caption{The BER versus $N$, when $\gamma_{\rm{d}}=25$ dB, $\widetilde \gamma = -20$ dB, and $I=100$ for different $K$ and $M$.
}
\label{BER_N}
\end{figure}
\subsection{Channel Estimation Error}
Finally, the quality of the channel estimation by the proposed clustering-based method are assessed in this subsection.
The \emph{normalized mean squared error} (NMSE) is chosen as the evaluation metric, which is defined as follows\footnote{It should be noted that, the proposed clustering method only needs to classify the received samples into two groups, and it cannot (actually does not need to) map each group to $c=0$ and $c=1$, respectively. Thus, to apply \eqref{equ:NMSE}, we need to check the transmitted $c^{(a_0)}$ and $c^{(a_1)}$ in simulations to realize mapping.}:
\begin{equation}\label{equ:NMSE}
J_{\rm{NMSE}}= {\mathbb E} \left[ \frac{1}{2} \sum_{i=0}^{1} \frac {\| {\bf{C}}_i - \hat{\bf{C}}_i \|_{\rm F}^2 }
{\| {\bf{C}}_i \|_{\rm F}^2 }
\right]
.
\end{equation}
We suppose that the receiver equips $M=4$ antennas, and the backscatter frame consists of $I=100$ backscatter symbols.
{\figurename~\ref{NMSE_rd}} shows the NMSE performance over a range of $\gamma_d$ when $K$ increases from $1$ to $4$, in which $\widetilde \gamma = -20$ dB and $N=50$.
It is seen that the NMSE converges to a constant with the increase of $\gamma_d$.
This implies that, although the final detection BER can be improved as $\gamma_d$ increases when $M>K$, the channel estimation error always exists given a finite $\widetilde \gamma$ and $N$.
In addition, it is noticed that, when $K$ increases, the NMSE increases, since the number of directions of arrival required to be estimated increases in proportion to $K$.
\begin{figure}
[!t]
\centering
\includegraphics[width=.99\columnwidth]{NMSE_rd-eps-converted-to.pdf}
\caption{The NMSE versus $\gamma_d$, when $\widetilde \gamma = -20$ dB, $N=50$, $I=100$, and $M=4$, for different $K$.
}
\label{NMSE_rd}
\end{figure}
{\figurename~\ref{NMSE_N}} illustrate the NMSE with regard to the spreading factor $N$.
The parameters are set as $\gamma_d = 30$ dB and $\widetilde \gamma=-20$ dB.
It is seen that the NMSE for all the chosen $K$ significantly decreases with the increase of $N$.
The reason is that, the quality of the sample covariance matrix ${{\bf{R}}}^{(i)}$ in \eqref{equ:samplecm} becomes higher as $N$ increases, which results in better estimation performance.
\begin{figure}
[!t]
\centering
\includegraphics[width=.99\columnwidth]{NMSE_N-eps-converted-to.pdf}
\caption{The NMSE versus $N$, when $\gamma_d = 30$ dB, $\widetilde \gamma=-20$ dB, $I=100$, and $M=4$, for different $K$.
}
\label{NMSE_N}
\end{figure}
\section{Conclusion}\label{conclusion}
In this paper, we have investigated on the cognitive backscatter signal transmission, where multiple RF sources are considered, and there is no cooperative between the legacy system and the backscatter system.
Our study has demonstrated that to exploit multiple antennas at the receiver is a feasible way to achieve high-throughput backscatter signal transmission.
The optimal SD and suboptimal HD detectors have been designed based on the ML criterion.
In addition, a fully blind channel estimation method has been proposed based on the statistical clustering.
Simulation results have been provided to evaluated the performance of the proposed detectors.
It is verified that, the error-floor phenomenon can be avoided when $M>K$, the HD based suboptimal detector can achieve almost the optimal performance with lower complexity, and the detector using the estimated CSI can perform comparably as their counterparts with perfect CSI.
\appendices
\section{Proof of Proposition \ref{pro2}}\label{M_larger_K}
One simple method to suppress the DLIs from $K$ RF sources is the the \emph{zero-forcing} (ZF) beamforming receiver \cite{Tse2005Wirelesscom}, which projects the received signal ${\bf{y}}$ onto the subspace orthogonal to the one spanned by the vectors ${{\bf{f}}_1},{{\bf{f}}_2},\cdots,{{\bf{f}}_K}$\footnote{It is noted that ZF beamforming receiver is a suboptimal but simple linear receiver. We choose it in this proof to provide a performance lower bound.
In practice, the \emph{minimum mean square error} (MMSE) receiver or the ML receiver can be implemented to achieve better performance.}.
Letting ${\bf H}=[{{\bf{f}}_1},{{\bf{f}}_2},\cdots,{{\bf{f}}_K}, {\bf{g}}]$, the ZF beamforming vector ${\bf w}$ is the last row of the pseudoinverse ${\bf H}^\dagger$ of the matrix ${\bf H}$, defined by:
\begin{equation}
{\bf H}^\dagger := ({\bf H}^H {\bf H})^{-1} {\bf H}^H.
\end{equation}
Since $K<M$ is satisfied, we have ${\bf{w}} {{\bf{f}}_k}=0$ for $k=1,\cdots,K$ \cite{Tse2005Wirelesscom}.
Using ${\bf{w}}_{\rm{D}}=\frac{{\bf{w}}}{\|{\bf{w}}\|}$, the interference in \eqref{equ:Mantenna} is removed:
\begin{equation}\label{equ:ZF_cancel}
\begin{aligned} [b]
x&= {\bf{w}}_{\rm{D}} {\bf{y}}\\
&={\bf{w}}_{\rm{D}} {\bf{g}} \sum_{k=1}^K l_k \sqrt{\widetilde \gamma_k {\gamma_{{\rm{d}},k}}} s_k c+{\bf{w}}_{\rm{D}} {\bf{u}}
.
\end{aligned}
\end{equation}
Let ${\bf s}=[s_1,s_2,\cdots,s_K]$. From \eqref{equ:ZF_cancel}, ${\bf s}$ can be combined into the channel response. Then the achievable rate of the backscatter signal using ZF beamforming receiver is
\begin{equation}\label{equ:R_ZF}
\begin{aligned} [b]
{\widetilde{R}}_{\{M,K\}} &= {\mathbb E}_{\{ {\bf{s}} \}} \left[ \log_2 (1+
\frac {\left|{\bf{w}}_{\rm{D}} {\bf{g}} \sum_{k=1}^K l_k \sqrt{\widetilde \gamma_k {\gamma_{{\rm{d}},k}}} s_k\right|^2}
{ \| {\bf{w}}_{\rm{D}} \|^2}
) \right]\\
&= {\mathbb E}_{\{ {\bf{s}} \}} \left[ \log_2 (1+
{\left|{\bf{w}}_{\rm{D}} {\bf{g}}\right|^2 \sum_{k=1}^K {\widetilde \gamma_k} {\gamma_{{\rm{d}},k}} |l_k s_k|^2}
) \right]
.
\end{aligned}
\end{equation}
Obviously, we have ${{R}}_{\{M,K\}}\geq {\widetilde{R}}_{\{M,K\}}$.
Assuming ${\gamma_{{\rm{d}},k_0}}$ is infinite, we have
\begin{equation}\label{equ:R_ZF_k0}
\begin{aligned} [b]
{\widetilde{R}}_{\{M,K\}} &\geq {\mathbb E}_{\bf s} \left[ \log_2 (1+
{{\gamma_{{\rm{d}},k_0}} {\widetilde \gamma_{k_0}} \left|{\bf{w}}_{\rm{D}} {\bf{g}}\right|^2 |l_{k_0}|^2 |s_{k_0}|^2}
) \right]
.
\end{aligned}
\end{equation}
Since $s_{k_0} \sim{\cal{CN}}(0,1)$, $|s_{k_0}|^2$ follows exponential distribution.
By letting $\delta={{\gamma_{{\rm{d}},k_0}} {\widetilde \gamma_{k_0}} \left|{\bf{w}}_{\rm{D}} {\bf{g}}\right|^2 |l_{k_0}|^2}$, \eqref{equ:R_ZF_k0} becomes
\begin{equation}\label{equ:R_ZF_k0a}
\begin{aligned} [b]
{\widetilde{R}}_{\{M,K\}} &\geq {\mathbb E}_{\{ {\bf{s}} \}} \left[ \log_2 (1+
{\delta |s_{k_0}|^2}
) \right]\\
&=\int_0^{+\infty} { \log_2(1+\delta x) e^{-x}
}{\rm d}x\\
&=\frac{e^{\frac{1}{\delta}}}{\ln 2} \int_{\frac{1}{\delta}}^{+\infty} { \frac {e^{-x}}{x}
}{\rm d}x
.
\end{aligned}
\end{equation}
Finally, we have
\begin{equation} \label{equ:Rdown_Mantenna}
\begin{aligned} [b]
R_{\{M,K\}}^{\infty}&= \lim_{{\gamma_{{\rm{d}},k_0}} \rightarrow +\infty} R_{\{M,K\}}\\
&\geq \lim_{{\gamma_{{\rm{d}},k_0}} \rightarrow +\infty} {\widetilde{R}}_{\{M,K\}}\\
&\geq \lim_{\delta \rightarrow +\infty} \frac{e^{\frac{1}{\delta}}}{\ln 2} \int_{\frac{1}{\delta}}^{+\infty} { \frac {e^{-x}}{x} }{\rm d}x\\
&=\frac{1}{\ln 2} \int_0^{+\infty} { \frac {e^{-x}}{x} }{\rm d}x\\
&=+\infty
.
\end{aligned}
\end{equation}
\section{The Maximum Rate of the On-off Modulated $c$}\label{A_onoff_rate}
Let $R_{\rm o}$ denote the achievable rate for the on-off modulated $c$, which is the mutual information between the OOK modulated $c$ and received signal ${\bf{y}}$, i.e., $R_{\rm o} = {\rm I}(c;{\bf{y}})$.
Then the average maximum achievable rate $\overline{R}_{\rm o}$ is:
\begin{equation}\label{equ:ER_ook}
\overline{R}_{\rm o} = {\mathbb E}_{\{ {\bf{g}}, {\bf{f}},l \}} \left[ {\rm I}(c;{\bf{y}})\right]
.
\end{equation}
It is known that, the close-form expression of $R_{\rm o}$ is unavailable even for the $M=1$ cases \cite{WangG2018AmBCrate}.
Thus in this section, we try to present a numerical method to obtain $\overline{R}_{\rm o}$.
The mutual information ${\rm I}(c;{\bf{y}})$ is expressed as follows
\begin{equation}\label{equ:R_ookn}
\begin{aligned} [b]
{\rm I}(c;{\bf{y}})
&={\rm H}(c)-{\rm H}(c|{\bf{y}})\\
&={\rm H_b}(\phi_0)+{\mathbb E}_{\left\{{\bf{y}}_0\right\}} \left[ {\rm H}(c|{\bf{y}}_0) \right]
.
\end{aligned}
\end{equation}
where ${\bf{y}}_0$ is once realization of ${\bf{y}}$, $\phi_0$ and $\phi_1$ are the prior probabilities when ``$c=0$'' and ``$c=1$'', respectively, which satisfies $\phi_0+\phi_1=1$, ${\rm H_b}(\phi_0)\triangleq -\phi_0 \log_2 \phi_0-\phi_1 \log_2 \phi_1$ is the binary entropy function, and $p( c | {\bf{y}}_0)$ is the posterior probability of $c$ while receiving ${\bf{y}}_0$.
Since ${\rm H_b}(\phi_0)$ is independent of all the channel coefficients, the average achievable rate has the same expression as \eqref{equ:R_ookn}:
\begin{equation}\label{equ:ER_ookn}
\begin{aligned} [b]
\overline{R}_{\rm o} &= {\mathbb E}_{\{ {\bf{g}}, {\bf{f}},l \}} {\rm I}(c;{\bf{y}})\\
&={\rm H_b}(\phi_0)+{\mathbb E}_{\left\{ {\bf{y}}_0\right\}} \left[ {\rm H}(c|{\bf{y}}_0)\right]
.
\end{aligned}
\end{equation}
Note that, in \eqref{equ:R_ookn} and \eqref{equ:ER_ookn}, ${\rm H}(c|{\bf{y}}_0)$ is averaged with respect to ${\bf{y}}_0$ according to different prior distributions.
Using the PDFs in \eqref{equ:Gpdfc1} and \eqref{equ:Gpdfc0}, we have the posterior probability as follows:
\begin{equation}\label{equ:post_y}
\begin{aligned} [b]
p( c=j | {\bf{y}}_0) &= \frac{\phi_j p( {\bf{y}}_0| c=j) } {p({\bf{y}}_0)}
,
\end{aligned}
\end{equation}
where $p({\bf{y}}_0) = \phi_0 {p( {\bf{y}}_0| c=0) } +\phi_1 {p( {\bf{y}}_0| c=1) }$.
Let $\varepsilon_j=p( c=j | {\bf{y}}_0)$, for $j=0,1$.
Then we have the conditional entropy
\begin{equation}\label{equ:conditionalE}
\begin{aligned} [b]
{\rm H}(c|{\bf{y}}_0)&=- \sum_{j=0}^1 \varepsilon_j \log_2 \varepsilon_j \\
&={\rm H_b}(\varepsilon_0).
\end{aligned}
\end{equation}
Substituting \eqref{equ:conditionalE} into \eqref{equ:ER_ookn}, the average achievable rate is
\begin{equation}\label{equ:R_ookF}
\begin{aligned} [b]
\overline{R}_{\rm o}
&= {\rm H_b}(\phi_0)-{\mathbb E}_{\left\{ {\bf{y}}_0\right\}} \left[{\rm H_b}(\varepsilon)\right]\\
&= {\rm H_b}(\phi_0)-\int_{{\bf{y}}_0} {p({\bf{y}}_0) {\rm H_b}(\varepsilon) } {\rm d}{{\bf{y}}_0}.
\end{aligned}
\end{equation}
The integral in \eqref{equ:R_ookF} can be calculated by Monte Carlo integration. Firstly, the channel coefficients ${\bf{g}}$, ${\bf{f}}_1,\cdots,{\bf{f}}_K$, and $l_1,\cdots,l_K$ are randomly generalized, and then ${\bf{C}}_0$ and ${\bf{C}}_1$ are obtained for each realization based on \eqref{equ:C0} and \eqref{equ:C1}.
Secondly, $c$ and ${\bf{y}}_0$ are randomly generated for sufficient times given a group of $\phi_0$, $\phi_1$, ${\bf{C}}_0$ and ${\bf{C}}_1$,
and then an instantaneous result $R_t={\rm H_b}(\phi_0)-{ {\rm H_b}(\varepsilon) }$ is obtained.
Finally, the mean value of $R_t$ will well approximate the $\overline{R}_{\rm o}$ in \eqref{equ:R_ookF} based on the \emph{Law of Large Numbers}.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
|
\section{\noindent ~Introduction}
All electrochemical intercalation based batteries require not only ample lattice space and robust structure of an electrode, but also a multiple valance ion to maintain charge neutrality \cite{Rev1}. In this regard transition metal based oxides are widely studied as cathode as well as anode in Li/Na/K ion batteries \cite{Rev1, JameshJPS18, Rev TMO, revNMO2017, XiaoAEM18}. Among transition metal oxides, Na$_{0.44}$MnO$_2$ with tunnel type structure and multivalent Mn ion have been studied as cathode material (in pristine form) \cite{NM44synRev, ChenACSAMI18} as well as anode material (when doped with other appropriate transition metal ion) in a Na-ion battery \cite{Ti NMO}. The 3D lattice structure of Na$_{0.44}$MnO$_2$ comprise of MnO$_6$ octahedra and MnO$_5$ square based pyramids with 0.44 Mn in 3+ valence state and 0.56 Mn in 4+ valence state. Among these, all the Mn$^{4+}$ and 50\% Mn$^{3+}$ occupy the octahedral position whereas rest of the 50\% of Mn$^{3+}$ occupy pyramidal position resulting into three different sites for Na ions in two different tunnels, one large S-shaped and other smaller one \cite{SauvageInChem2007}. These tunnels in Na$_{0.44}$MnO$_2$ allow an easy insertion and reinsertion of the Na--ions during the oxidation and reduction of the cathode material.
The electrochemical properties and particularly the cycling capacity of Na$_{0.44}$MnO$_2$ critically depend on the synthesis method \cite{SauvageInChem2007,tempAssSolGel2016,polypyrolysisNanowire}, sintering temperature \cite{SauvageInChem2007,NM44synRev} and morphology \cite{Nanorods2016,DemirelMT,ZhaoRSC13Raman} of the samples. Despite of the potential, two major challenges need to be overcome while using these materials as cathode in Na-ion batteries. First is the poor retention of capacity upon cycling and low capacity at higher current rates. The poor cyclic capacity is owing to the structural instability upon insertion/extraction of the Na--ions \cite{NM44synRev,SauvageInChem2007}. Such deficiency can be resolved by altering the morphology (surface to volume ratio) by optimizing the growth technique. There have been various modified techniques employed for the synthesis of Na$_{0.44}$MnO$_2$ \cite{ZhangEA17, MaEA16, AstaACSAMI17, JuJPC18}, such as reverse microemulsion method \cite{reverseMicroMethod,NM44synRev}, polymer-pyrolysis method \cite{polypyrolysisNanowire}, modified pechini method \cite{modPechini,ZhaoRSC13Raman} and optimized solid state reaction method \cite{DemirelMT}. Apart from the synthesis, introduction of some disorder in the electrode material, especially in the form of doping, has been found to be increasing the capacity and performance of the electrode material \cite{Rev TMO, Rev1,Ti NMO,Cr dope Nat}.
Another major issue with Mn based transition metal oxides is the presence of the strong Jahn-Teller (JT) active Mn$^{3+}$ ion. The JT distortion makes the compound vulnerable to the structural changes during the Na insertion/extraction resulting into the reduction of capacity after cycling. One approach to overcome this limitation is to substitute the transition metal site with a diavalent ion \cite{Zn doping ref, Na content}, which eventually converts some of the Mn$^{3+}$ to Mn$^{4+}$, therefore reducing the number of JT ions. One such example is substitution of Ni at Mn site where the most stable state of Ni (which is Ni$^{2+}$) is utilized \cite{Na content}. Similarly, the most stable state of Zn is 2+ and therefore, can be a potential candidate for the replacement of Mn. Moreover, the ionic size of Zn is larger than that of Mn, so one can expect an increase in the unit cell volume with Zn substitution and hence easy insertion/reinsertion of Na ions upon charging/discharging. The larger ion in the lattice or larger unit cell volume upon substitution has been found to improve electrochemical properties in some oxides \cite{SarohaAppSurSci2017, LeeSciReP2017}.
Therefore, we report the synthesis of Zn-substituted tunnel type Na$_{0.44}$MnO$_2$ by sol-gel technique and investigation of their structural, morphological, transport and electrochemical properties. These samples stabilize in the orthorhombic structure and the activation energy decreases with increasing Zn concentration. We observed the specific capacity of about 100~mAh/g at current density of 4~mA/g. The capacity does not increase with Zn substitution, but we observed significant improvement in the cycling life of a Na-ion battery.
\section{\noindent ~Experimental}
We have synthesized Na$_{0.44}$Mn$_{1-x}$Zn$_x$O$_2$ ($x=$ 0, 0.005, 0.02) using sol gel method. The sodium nitrate (Merck, 99\%), manganese nitrate (Merck, 99\%), and zinc nitrate (Merck, 99.9\%) were added in a stoichiometric ratio in deionized water followed by homogeneouse mixing via stirring for 2~hrs. Then, citric acid (Sigma, 99.9\%) in molar ratio of 3:1 was added to the mixture as a complexing agent followed by overnight stirring at 90$^o$C, which results in the formation of the gel. The formed gel was dried at 120$^o$C for 24~hrs. The obtained powder then ground to get fine particles. The powder was first heated to 450$^o$C for 12~hrs to remove organic components before final sintering at 900$^o$C for 15~hrs.
The room temperature powder x-ray diffraction data were recorded with CuK$\alpha$ radiation (1.5406~\AA ) from Panalytical x-ray diffractometer in the 2$\theta$ range of 10 -- 60$^o$. The surface morphology of the prepared materials has been investigated using a scanning electron microscope (SEM) at 20~keV electron energy. Raman spectra of the prepared pellets were recorded with Renishaw inVia confocal Raman microscope at wavelength of 785~nm and grating of 1200~lines/mm with 1~mW laser power. Temperature dependent resistivity measurements were done using physical property measurement system (PPMS) from Quantum Design, USA.
For the electrochemical measurements, first the slurry for cathodes were prepared by mixing Na$_{0.44}$Mn$_{1-x}$Zn$_x$O$_2$ ($x=$ 0 -- 0.02) active material, PVDF (Polyvinylidene difluoride) as binder and carbon black as conductive additive in a weight ratio of 80:10:10 in NMP (N-methyl-2-pyrrolidinone) solvent followed by overnight stirring. The obtained slurry was coated on a current collector (Al foil in present case). In order to evaporate the solvent, the coated material was dried in vacuum oven at 135$^o$C for 10~hrs. After that, the circular disks of 12~mm diameter were prepared. The CR2016 coin half-cells were assembled in an nitrogen-filled glove box (Jacomex, $\leq$0.5 ppm of O$_2$ and H$_2$O level). The sodium disks prepared from sodium cubes (Sigma Aldrich, 99.9\%) with 16 mm diameter were used as counter as well as reference electrode. The glass fiber filter (Advantec, GB-100R) was used as separator and in house prepared NaClO$_4$ dissolved in ethylene carbonate (EC)/dimethyl carbonate (DMC) in a volume ratio of 1:1 was used as an electrolyte \cite{maheshMRB}. All the electrochemical characterizations were performed using VMP3 (Biologic) instrument and BTS-400 (Neware) at room temperature. The cyclic voltmmetry (CV) measurements were performed in the potential window of 1.8 -- 4.0~V vs. Na/Na$^+$ at a scan rate of 0.1/0.05~mVs$^{-1}$. The charging/discharging of the coin cell were performed in galvanostatic mode at different current densities.
\begin{figure}[h]
\includegraphics[width=3.3in]{Fig1_XRD}
\caption{The XRD patterns for the Na$_{0.44}$Mn$_{1-x}$Zn$_x$O$_2$ ($x=$ 0 -- 0.02) samples with peak indexing using JCPDS\# 27-0750.}
\label{fig:Fig1_XRD}
\end{figure}
\section{\noindent ~Results and Discussion}
The x-ray diffraction (XRD) patterns for the as synthesized Na$_{0.44}$Mn$_{1-x}$Zn$_x$O$_2$ ($x=$ 0 -- 0.02) are shown in Fig.~1. The structure for $x=$ 0 sample is orthorhombic with Pbam space group (JCPDS No: 27-0750), which is in agreement with earlier reports \cite{SauvageInChem2007, Nanorods2016, NM44synRev}. The Zn substitution does not affect the structure and it remain orthorhombic for all the samples. The compositions of all the samples have been confirmed by energy dispersive x-ray spectroscopy (not shown). We have estimated the strain and crystallite size using Williamson Hall plots \cite{WHallActa1953} from the XRD data. We found that with increasing Zn concentration, the strain increases from (2.13$\pm$0.5$)\times$10$^{-3}$ for $x$=0 to (2.6$\pm$0.2$)\times$10$^{-3}$ for $x$=0.005 and (3.2$\pm$0.3$)\times$10$^{-3}$ for $x$=0.02 smaples. This observation indicates that the substitution of Mn with the larger ion Zn might be causing this increase in internal strain. However, the crystallite size decreases from 100~nm to 50~nm with increasing Zn concentration.
\begin{figure*}
\includegraphics[width=7.0in]{Fig2_SEM}
\caption{The SEM images for Na$_{0.44}$Mn$_{1-x}$Zn$_x$O$_2$: (a) $x=$ 0, (b) $x=$ 0.005, and (c) $x=$ 0.02 samples.}
\label{fig:Fig2_SEM}
\end{figure*}
Fig.~2 shows the surface morphology of Na$_{0.44}$Mn$_{1-x}$Zn$_x$O$_2$ ($x=$ 0 -- 0.02) samples, investigated using SEM technique. In the SEM images [see Figs.~2(a-c)], we observed slabs of a few $nm$ width and $\mu$m length for Na$_{0.44}$MnO$_2$ sample. With Zn substitution, the morphology does not change much and slab structure prevails for $x=$ 0.005 and 0.02 samples. However, for $x=$ 0.02 sample there are some random shape particles also observed along with the slabs [Fig.~2(c)]. In Fig.~3, we have shown the Raman spectroscopy measurements for $x=$ 0 and 0.02 samples using a 785~nm wavelength laser. There are two major peaks observed in the Raman spectra for $x=$ 0 sample , i.e. at 654 and 369 cm$^{-1}$. These two peaks correspond to the stretching vibrations of Mn-O bonds and bending vibrations of Mn-O-Mn bonds, respectively. These peaks are slightly shifted and lower in intensity as observed at 649 and 364 cm$^{-1}$ for $x=$ 0.2 sample. Considering the diatomic approximation which assumes each metal oxygen bond as a separate and non-interacting oscillator and which has been used to explain the relation between Raman shift and bond length in many transition metal oxides \cite{WachsRamanSolidIon1991, WachsRamanJChemSoc.Faraday1996}, the shift in the Raman peaks towards lower wave number with increasing Zn concentration indicates a decrease in the force constant and thus increase in the metal oxygen bond length. Apart from these major peaks there are few minor peaks observed in both the samples. These results are consistent with earlier reports \cite{ZhaoRSC13Raman,Raman0.21}.
\begin{figure}
\includegraphics[width=3.3in]{Fig3_Raman1}
\caption {Raman spectra of Na$_{0.44}$Mn$_{1-x}$ZnO$_2$ ($x=$ 0 and 0.02) samples, measured at room temperature using 785~nm laser. The dashed lines are fitted with Lorentzian line shape.}
\label{fig:Raman}
\end{figure}
Now let us discuss the electronic properties which also play an important role in the electrochemical properties. The electronic band structure of 3$d$ transition metal oxides is mainly governed by the overlapping and interaction between 3$d$ orbitals of the neighboring transition metal ions. The substitution of the foreign ion alters the overall metal-metal distance, which manifest itself into the modification in the band structure. The intercalation of Na into the cathode material is correlated to the density of states near the Fermi level \cite{Molenda1 SSIon1986, XiaChemSci2018, KunduraciChemMat2006, Molenda2 RSC2017}. The high density of states (i.e. higher conductivity) results in to high rate capability, whereas the lower density of states (lower conductivity) in the cathode material leads to poor cycle life as described in the ref.~\cite{Molenda1 SSIon1986}. In a way, lattice constant, electronic properties and electrochemical properties are correlated. Therefore, in order to get an idea about how Zn substitution affects the electronic properties, we have performed the temperature dependent resistivity measurements on the three samples by four probe method. All the samples show highly insulating nature due to which we could not measure the resistivity below 280~K, as it increases beyond the measurement limit of the instrument. When compared with $x=$ 0 sample, the resistivity first decreases for $x=$ 0.005 sample and then increases for $x=$ 0.2 (Fig.~4). However, the fitting of the temperature dependent resistivity data using Arrhenius equation: $\rho=\rho_0e^{E_a/k_B T}$, where, E$_a$ is the activation energy, shows that the activation energy systematically decreases with increasing Zn substitution. For example, the activation energy is 420~meV, 360~meV and 330~meV for $x=$ 0, 0.005 and 0.2 samples, respectively. These values of the activation energies are well in agreement with the previous reports on Mn based transition metal oxides \cite{bandgap, maheshMRB}. The plausible explanation for reduction in the activation energy can be the defect levels in the band gap due to the Zn substitution. The systematic decrease in the activation energy with Zn substitution indicates narrowing the band gap and increase in density of states with doping. As mentioned above, this increase in the density of states near Fermi level in $x=$ 0.2 sample can be one of possible reason for better rate capability despite of lower specific capacity which is described in the following section.
\begin{figure}
\includegraphics[width=3.3in]{Fig4_RT}
\caption {The temperature dependent resistivity data of Na$_{0.44}$Mn$_{1-x}$ZnO$_2$ ($x=$ 0 -- 0.02). The inset shows a plot between ln $\rho$ and 1/T with fitting by the Arrhenius equation.}
\label{fig:Fig3_RT}
\end{figure}
The electrochemical behavior of Na$_{0.44}$Mn$_{1-x}$ZnO$_2$ ($x=$ 0 -- 0.02) as the electrode materials in Na-ion batteries has been investigated by cyclic voltametry (CV) and charging discharging measurements at different current densities on the prepared coin cells. The CV data gives information about the potentials at which the oxidation/reduction takes place, as shown in Figs.~5(a--c), the scan rate is kept as 0.1 or 0.05~mV/sec for all the CV scans. For $x=$ 0 sample, there are six distinct peaks in the oxidation, which repeats in the reduction as well as upon cycling indicating reversibility of Na ion in extraction/insertion. The multi peaks in the CV indicate a complex phase evolution where each peak corresponds to the extraction of Na from different crystallographic sites in the structure \cite{DemirelMT}. The difference between oxidation and reduction peak values is a measure of the number of free electrons (or Na-ion) involved in the charging/discharging and is given by $\Delta$E=59/n~meV (at room temperature) \cite{CV formula}, where, $\Delta$E is the difference between the oxidation and corresponding reduction peaks and n is the number of electrons involved. For n=0.44 (to extract all the Na-ion), the peak separation should be 0.13~mV. For our sample the peak separation is 0.11, 0.17, 0.17, 0.18, 0.21~mV for peaks from higher voltage to lower voltage side, which correspond to n=0.53, 0.34, 0.34, 0.32, 0.28. These values are close to the expected n = 0.44 in ideal case. The CV data for all three samples ($x=$ 0, 0.005, 0.02) are shown in Figs.~5(a--c). Now, if we look at the CV for $x=$ 0.005 and 0.02 samples, it is clear that $x=$ 0.005 sample contains all the oxidation and reduction peaks similar to that in $x=$ 0 sample with a small shift in the peak position. Moreover, the similar oxidation peaks have been observed for $x=$ 0.02 sample, but there are fewer reduction peaks when compared with the $x=$ 0 sample. In order to resolve the redox peaks for $x$=0.02 cathode we performed the CV at slower scan rate of 0.05~mV/s [Fig.~5(c)], but no significance change has been observed. The separation between the oxidation and reduction peaks has also increased indicating decrease in the number of available free electrons and irreversibility of Na ion for some sites, which affect the discharge capacity with higher Zn concentration at the Mn site.
\begin{figure}
\includegraphics[width=3.4in]{Fig5_CV}
\caption {(a--c) The cycling voltammogram and, (d--f) the capacity degradation upon cycling for Na$_{0.44}$Mn$_{1-x}$Zn$_x$O$_2$ with $x=$ 0, 0.005, 0.02, measured at different current densities.}
\label{Fig5_new}
\end{figure}
\begin{figure}
\includegraphics[width=3.5in]{Fig6_CD}
\caption {The charging discharging cycles measured at different current densities for $x=$ 0 sample (upper panel), 0.5$\%$ (middle panel) and 2$\%$ Zn substitutions (lower panel).}
\label{fig:Figure6_new}
\end{figure}
The galvanostatic charging-discharging curves at different current densities for Na$_{0.44}$Mn$_{1-x}$Zn$_x$O$_2$ ($x=$ 0, 0.005, 0.02) are shown in the Fig.~6 and the capacity fading with cycling is summarized in Figs.~5(d--f). For $x=$ 0 sample, the capacity is nearly 100~mAh/g (which is close to the theoretical capacity 121~mAh/g \cite{TheorycapRSC 2014}) for slow charging rates that is with current densities 4~mA/g and 6~mA/g. It can be seen that the capacity is almost constant after 10 cycles. However, for higher current density (14~mA/g) the capacity is low and degrade rapidly with cycling (nearly 25$\%$ degradation after 10 cycles), see Fig.~5(d). The possible reason for such behavior lie in the layered structure and presence of Jahn-Teller Mn$^{3+}$ ions in these samples. Similar capacity fading is also observed in other Mn based oxides where Mn has 3+ oxidation sate \cite{JTMn, maheshMRB}. As mentioned earlier the Zn substitution would decrease the concentration of Jahn-teller active Mn$^{3+}$ ions and one can expect improvement in the capacity degradation upon cycling. When comparing the capacity of $x=$ 0 and 0.005 samples after 10 cycles at 6~mA/g current density, we find that the capacity loss is nearly 4 $\%$ for the $x=$ 0 sample, whereas it is only about 1$\%$ for $x=$ 0.005 sample. In addition, the capacity is higher with improved cycling capacity at 14~mA/g current density. This clearly indicate the improvement in the cycle life with small Zn concentration (0.5\%). With further increasing Zn concentration (2\%), the overall capacity has decreases to about 60~mAh/g at slower current density of 4~mA/g. This decrease in the capacity is a manifestation of the absence of reduction peaks in the CV for 2$\%$ substituted sample, as discussed earlier. Despite of the low capacity, the durability is improved and the capacity remain unchanged after 10 cycles at 4~mA/g current density. Upon increasing the charging/discharging current density to 14~mA/g, the capacity decreases to nearly 25~mAh/g and remains constant for more than 30 measured cycles. In this way, we reveal that Zn substitution increases the cycle life possibly due to the reduction in Jahn-Teller Mn$^{3+}$ ions. Now, we discuss the possible explanation of the lower capacity in galvanostatic charging discharging for $x=$ 0.02 sample in the view of the CV measurements. The substitution of Zn$^{2+}$ can have the following effects in the parent sample: (1) it replaces the Mn$^{3+}$ (not Mn$^{4+} $, due to larger difference in ionic radii and charge) and simultaneously converts another Mn$^{3+}$ to Mn$^{4+}$ in order to maintain the charge neutrality, (2) Zn$^{2+}$ is reluctant in changing its valance state from 2+ and Mn$^{4+}$/Mn$^{5+}$ redox couple is rare and not accessible at ordinary potentials. Therefore, the presence of Zn$^{2+}$ hinders the extraction of Na from the cathode in these two ways. This hindrance clearly reflect in the CV of $x=$ 0.02 sample in terms of smaller anodic/cathodic peak current as well as less number of redox peaks, which may manifest in the low specific capacity.
\begin{figure}
\includegraphics[width=3.5in]{Fig7_XRD_SEM_1}
\caption {The comparison of XRD patterns for the as prepared sample and cycled electrode of $x=$ 0 and $x=$ 0.02 samples. The insets show SEM images of the cycled electrodes.}
\label{fig:Fig7_XRD}
\end{figure}
In order to understand how the cycling affect the structure and morphology, we recorded the XRD patterns and SEM images of the cycled electrodes. We compare the XRD data in Fig.~7 and SEM images in the insets. It is evident from the comparison of XRD data that there are no new peaks observed after cycling and the peaks in the as prepared materials matches with those in the the cycled electrode. This shows that there is no structural or phase change taking place during/after cycling process. However, in the case of $x=$ 0 sample, the peak broadening and intensity of peaks found to different indicating a change in particle size and morphology. The SEM images of $x=$ 0 and $x=$ 0.02 electrodes show that the electrode surface for $x=$ 0 is porous with large voids as compared to that for $x=$ 0.02 electrode, which might be responsible for rapid capacity degradation upon cycling.
\section{\noindent ~Conclusions}
In conclusion, we have synthesized nano-size rod shaped Na$_{0.44}$Mn$_{1-x}$Zn$_x$O$_2$ ($x=$ 0 -- 0.02) via sol gel method and studied the physical properties and electrochemical performance for Na-ion batteries. The structure is orthorhombic (Pbam space group) and it does not change with Zn substitution. All the samples are highly insulating and the activation energy, deduced from Arrhenius fitting, decreases with higher Zn concentration. The specific capacity for Na$_{0.44}$MnO$_2$ is nearly 100~mAh/g, which is close to the theoretically predicted value. The capacity does not increase with Zn substitution, but we observed significant improvement in cycling life. The capacity reduction is only 1$\%$ for the 0.5\% Zn sample after 10 cycles as compared to the 4\% of $x=$ 0 sample, measured at 6~mA/g current density. Our study reveal that Zn substitution changes Mn ion from 3+ to 4+ valence state and improve the battery life. We observed changes in the morphology of electrodes after cycling.
\section*{\noindent ~Acknowledgments}
RSD acknowledges the financial support from SERB-DST [Early Career Research (ECR) Award, project reference no. ECR/2015/000159], BRNS (DAE Young Scientist Research Award, project sanction no. 34/20/12/2015/BRNS), and IIT Delhi (FIRP project no. MI01418). Mahesh and Rishabh thank SERB-DST (NPDF, no PDF/2016/003565), and DST-inspire, respectively, for the fellowship. Rakesh thanks IIT Delhi for postdoctoral fellowship through FIRP project. The authors acknowledge central research facility (CRF) and the physics department of IIT Delhi for providing research facilities: XRD, SEM, Raman, and PPMS EVERCOOL--II.
|
\section{Introduction}\label{sec:intro}
Image segmentation refers to the process of uniquely identifying a
given pixel or set of pixels in a digital image as a member of a
region. In astronomy, this notion led to the development of
segmentation images (or maps) that define astrophysical sources in
pixelated images. Such images are often used used to define apertures
through which various measurements are made (for example brightness or
center-of-mass). Therefore, it is imperative to ensure that the
segmentation maps are free of pathological defects and false positives
are judiciously removed.
One common tool for algorithmically creating segmentation maps is
\texttt{Source Extractor} \citep[\texttt{SExtractor};][]{sex}, which has a host of
parameters that govern the properties of the deblending of nearby
sources. Although \texttt{SExtractor}\ is extremely efficient, it can often segment
an image in a way far different than a human might expect or desire
(notably regions near extended sources or bright point sources). But
also, it is challenging to force a segmentation region without
inadvertently creating many false sources. To remedy these issues,
several schemes have proposed to remove extraneous sources \citep[such
as hot/cold running;][]{bard12}. Here, I propose an alternative
paradigm: a graphical-user interface (GUI) that allows the user to
directly interact with the segmentation image pixels. Although there
are distinct advantages with an algorithmic approach (such as
repeatability), a properly cleaned segmentation map avoids erroneous
sources corrupting the aggregate properties of the sample and/or
biasing measurements of other sources in the field. With these
tradeoffs in mind, I expect the ideal approach is to begin with a
segmentation map derived by some algorithmic means then sensibly
polishing the map for clear mistakes, but \textit{primum non nocere}.
This document is organized as follows: I introduce the package,
including installation and functionality in \Sect{sec:editor}, I
present several examples in \Sect{sec:examples}, I describe the User
preferences in \Sect{sec:pref}, I present additional improvements
and limitations of the current implementation in \Sect{sec:add}, and
close with a few closing remarks and waivers in \Sect{sec:coda}.
\section{Segmentation Editor}\label{sec:editor}
\subsection{Installation, Dependencies, and Calling}
The GUI is entirely written in
IDL\footnote{\href{http://www.harrisgeospatial.com/SoftwareTechnology/IDL.aspx}{http://www.harrisgeospatial.com/SoftwareTechnology/IDL.aspx}},
and is installed by simply placing the directory tree in the
\texttt{IDL\_PATH} of the user. The code requires the astronomy
library \citep[\texttt{AstroLib};][]{astrolib}, which can be
similarly installed. The main component of \texttt{SegEditor} is
written as a class that is instantiated from the IDL command line as:
\begin{verbatim}
IDL> obj=obj_new('segeditor',SEGFILE,IMGFILE)
\end{verbatim}
\vspace*{1ex}
\noindent where \texttt{SEGFILE} and \texttt{IMGFILE} are variables
referring to the full path of the segmentation and direct image,
respectively. I also provide a procedural-oriented wrapper that
handles the object-oriented aspects internally. Since older versions
of IDL do not explicitly destroy objects, it may be imperative to
explicitly destroy the resultant object (\texttt{obj} in the above
example). However, if the GUI is properly closed, then the object
reference will also be destroyed. Therefore if \texttt{SegEditor} is
called from an external code-base, it may be necessary to verify the
object still exists (such as with \texttt{obj\_valid()}).
\subsection{Basic Functionality}
The initial state of the GUI is shown in \Fig{fig:segeditor} with data
from \citet{clash}. There are several quick keys to execute common
functions, often for the left side of the keyboard (see
\Tab{tab:keyboard}). However, the main controls are driven by the
state of the mouse, which is dictated by toolbar along the top. These
buttons are briefly described in \Tab{tab:buttons}.
\input{letters}
\begin{figure}
\includegraphics[width=3.3in]{segeditor.pdf}
\caption{Basic state of the \texttt{SegEditor} GUI. At the top row
is the primary toolbar in the default state. In the middle, is
the main image display with the segmentation map as a
semi-transparent three-color and the direct image as a gray-scale
image, respectively. The opacity of the segmentation map is
controlled by the right-mouse button described in
\Tab{tab:buttons}. In the four corners of the graphics window,
the $(x,y)$ positions, image pixel values, min/max display values
for the direct image, and the RA/Dec positions (from top left
going clockwise). Finally, at the very bottom is a brief
instruction of the current left-mouse state.
\label{fig:segeditor}}
\end{figure}
\section{Example Actions}\label{sec:examples}
Although it may be readily straightforward to manipulate the GUI, it may
be useful to describe a few actions to begin.
\begin{figure*}
\begin{center}
\includegraphics[width=1.62in]{step1.pdf}
\includegraphics[width=1.62in]{step2.pdf}
\includegraphics[width=1.62in]{step3.pdf}
\includegraphics[width=1.62in]{step4.pdf}
\caption{Example to separate two regions and regroup. On the far
left, I show the original image (a zoom from \Fig{fig:segeditor}).
The first step is to use the eraser tool \capline{button.pdf}
to separate the light blue region from the main pink region. Next
is to {\it ungroup} the separated blue region with the
disassociate tool \capline{drawing.pdf}. Finally, the
remaining light blue region can be grouped with the larger pink
region with merge tool
\capline{group.pdf}.\label{fig:example}}
\end{center}
\end{figure*}
In \Fig{fig:example}, I show an example in using \texttt{SegEditor}.
Here the light-blue object has been associated as a small object to
the lower right, but \texttt{SExtractor} identified several pixels
that probably should belong to the main galaxy to the upper left (pink
source). To reassign those pixels to the main galaxy, one can simple
erase the collection of pixels that connect the two regions with
\inline{button.pdf}. Once the smaller object is isolated from the
erroneous pixels to the upper left, then it can be ungrouped from the
pair with \inline{drawing.pdf} and finally two bits of the main
galaxy can be merged with \inline{group.pdf}.
But it is perhaps still the case that one may wish to increase the
outer edge of the region of a source, such as that as seen for the
pink galaxy. To execute such an operation, one selects the paint tool
\inline{paint.pdf} then will click and drag to paint the pixels.
However it is important to begin the click/drag motion on the region
to be expanded (the paint tool does not paint a new region from
scratch). To draw a new region, where one was not there before, the
draw tool \inline{segpoly.pdf} is used (see \Fig{fig:draw}).
Here, one clicks and drags to encompass an set of pixels (the region
is automatically closed), which will be given a unique segmentation
value (the current maximum segmentation value plus one).
\begin{figure}
\begin{center}
\includegraphics[width=1.62in]{draw1.pdf}
\includegraphics[width=1.62in]{draw2.pdf}
\caption{Example to draw a region. The first image is the
original, again a zoom from \Fig{fig:segeditor}. But since the
source in the center was unidentified by \texttt{SExtractor}, it
is necessary to create a region with the draw tool
\capline{segpoly.pdf}.\label{fig:draw}}
\end{center}
\end{figure}
Sometimes it is useful to identify a particular known source or scan
the segmentation for pathological issues (such as the smallest/largest
sources). Such operations are easily achievable from the tabulation
view, which is opened with \inline{dm.pdf}. With this interface,
one can click on a column heading to sort the table by this column,
and once the data is sorted a second click will sort in reverse order.
Also, by clicking on a row header, the primary GUI will pan to the
object of interest. However it is important to note, the tabulation
GUI will not update as operations are performed in the main window.
Therefore it is imperative to refresh the table with
\inline{dm.pdf}.
\begin{figure}
\begin{center}
\includegraphics[width=3.25in]{table.pdf}
\caption{Table view. The columns can be sorted either ascending
or descending by clicking on the column headers. By clicking on
a row header, the main GUI (\Fig{fig:segeditor}) will center on
that source.\label{fig:table}}
\end{center}
\end{figure}
\section{Preferences} \label{sec:pref}
There are several parameters that control the fundamental operations
of the GUI and the text displays. These can be edited with
\inline{propsheet.pdf} and \inline{text.pdf}, respectively.
In \Tab{tab:prefs}, I describe the meaning of most of these keywords.
\input{prefs}
\begin{figure}
\begin{center}
\includegraphics[width=2.25in]{operations.pdf}
\caption{Operations tab. The sub-GUI for various image processing
options. The {\it Compress} button will compress the
segmentation map to the lowest set of positive integers, still
reserving zero for the sky pixels. The {\it Dilate} button will
apply the dilation operation with size given by the adjacent
slider. The {\it Erode} button will apply the erosion operation
with the size given by the adjacent slider. Please see the
\texttt{IDL} manual for more details on these latter two
operations.\label{fig:operations}}
\end{center}
\end{figure}
\section{Additional Remarks and Functionality}\label{sec:add}
I have identified a few places where the current implementation does
not facilitate certain actions that a User may like. I may address
these in future versions, but am willing to work with Users wishing to
modify the codebase to enact these or other improvements.
\begin{description}
\item[Fixed window size] The size of the main graphics window is fixed
to facilitate performance rendering and manipulating the RGB$\alpha$
displays, which necessitated fixing the text sizes and margins.
Either of these maybe relaxed in future, but Users wishing to force
a different window size can easily adjust the \texttt{self.winsize}
in the \texttt{init} method of the file
\texttt{segeditor\_\_define.pro}. This variable is in units of
screen pixels.
\item[Segmentation map must exist] I have assumed that the
segmentation map exists and will be modified. However, one may
prefer to start with a blank segmentation map and simply draw the
entire collection of regions. If such functionality is required,
then a simple workaround is to create a blank segmentation map {\it
before} working with \texttt{SegEditor} and ensure that the
world-coordinate system variables match those of the direct image.
Of course, such a created image should likely have the
\texttt{BITPIX} keyword set to an integer-like value.
\item[More interaction in sub-GUIs] There are several sub-GUIs that
can be spawned from within \texttt{SegEditor}, however their
relative interaction with the primary graphics window is quite
limited. One such example of these limitation is related to the
tabulation GUI, where it might be desirable to have the table update
as regions are modified in the primary graphics window. Conversely,
it is often advantageous to interact with segmentation map via the
table (such as deleting regions in the table view). Until such
changes are implemented, it is necessary to refresh the sub-GUIs
and/or the primary graphics window as changes are made.
\end{description}
\section{Coda}\label{sec:coda}
I hope others find this software useful, and if so, then I would
gratefully appreciate a reference to this report. I encourage any
Users to report bugs and/or suggestions. I distribute the software
according to The MIT License.
\acknowledgments
I am very grateful to Duho Kim, Teresa Ashcraft, Andrea Bellini, Seth
Cohen, and Joe Hunkeler for their help in debugging and useful
suggestions.
\software{IDL}
\input{bib}
\hfill
\newpage
\input{buttons}
\end{document}
|
\section{Future work}
\label{sec:future-work}
It is well-known that the tile complexity of a 2D $N \times N$ square at temperature-2 is $O\left( \frac{\log N}{\log \log N} \right)$.
More formally, for an $N \times N$ square $S^2_N = S_N = \{0,1,\ldots, N-1\}\times\{0,1,\ldots,N-1\}$, $K^2_{USA}\left(S_N\right) = O\left( \frac{\log N}{\log \log N}\right)$ \cite{AdlemanCGH01}.
However, it is conjectured \cite{ManuchSS10} that $K^1_{USA}\left(S_N\right)=2N-1$, meaning that a 2D $N \times N$ square does not self-assemble efficiently at temperature-1.
Yet, the tile complexity of a just-barely 3D $N \times N$ square at temperature-1 is $O\left(\frac{\log N}{\log \log N}\right)$.
That is, $K^1_{USA}\left(S^3_N\right) = O\left(\frac{\log N}{\log \log N} \right)$, where $S^3_N$ is a just-barely 3D $N \times N$ square, satisfying $\{0,1,\ldots,N-1\}\times\{0,1,\ldots,N-1\}\times\{0\} \subseteq S^3_N \subseteq \{0,1,\ldots,N-1\}\times\{0,1,\ldots,N-1\}\times\{0,1\}$ \cite{jFurcyMickaSummers}.
So, a 2D $N \times N$ square has the same asymptotic tile complexity at temperature-2 as its just-barely 3D counterpart does at temperature-1.
Regarding thin rectangles, we know that $K^2_{USA}\left(R^2_{k,N}\right) = O\left(N^{\frac{1}{k}}\right)$ \cite{AGKS05g} and we speculate whether a similar upper bound holds for a just-barely 3D $k \times N$ thin rectangle at temperature-1.
In other words, is it the case that either $K^1_{SA}\left(R^3_{k,N}\right)$ or $K^1_{USA}\left(R^3_{k,N}\right)$ is equal to $O\left(N^{\frac{1}{k}}\right)$?
If not, then what are tight bounds for $K^1_{SA}\left(R^3_{k,N}\right)$ and $K^1_{USA}\left(R^3_{k,N}\right)$?
We conjecture that $K^1_{USA}\left( R^3_{k,N} \right) = o\left( N^{ \frac{1}{\left \lfloor \frac{k}{3} \right \rfloor} }\right)$.
\section{Upper bound}
\label{sec:thin-rectangle-construction}
In this section, we give a construction for a singly-seeded TAS in which a sufficiently large just-barely 3D rectangle uniquely self-assembles. Going forward, we say that $R^3_{k,N} \subseteq \mathbb{Z}^3$ is a 3D $k \times N$ \emph{rectangle} if $\{0,1, \ldots,N-1\} \times \{0,1,\ldots,k-1\} \times \{0\} \subseteq R^3_{k,N} \subseteq \{0,1\ldots,N-1\} \times \{0,1\ldots,k
-1\} \times \{0,1\}$. For the sake of clarity of presentation, we represent $R^3_{k,N}$ vertically. Here is our main positive result.
\addtocounter{theorem}{0}
\begin{theorem}
$K^1_{USA}\left( R^3_{k,N} \right) = O\left( N^{\frac{1}{\left \lfloor \frac{k}{3} \right \rfloor}} + \log N \right)$.
\end{theorem}
\begin{figure}[htp]
\begin{center}
\begin{subfigure}[t]{.48\textwidth}
\begin{center}
\includegraphics[width=60px]{super_high_level_overview}
\caption{\label{fig:super_high_level_overview} High-level: values of the counter. }
\end{center}
\end{subfigure}
~
\begin{subfigure}[t]{.48\textwidth}
\begin{center}
\includegraphics[width=90px]{Example_Full}
\caption{\label{fig:Example_Full} Low-level: full example. }
\end{center}
\end{subfigure}
\caption{\label{fig:example_low_high} A full example of a $11\times 56$ construction. The counter begins at 10-01-10 and is counting in ternary, so the initial value is $23 = 2\cdot 3^2 + 1\cdot 3^1 + 2 \cdot 3^0$. Note that the least significant bit for each digit, which is the lowest bit in each digit column, is actually a ``left edge'' indicator, where ``1'' means ``leftmost'' and ``0'' means ``not leftmost''. To help distinguish overlapping tiles, ``write gadgets'' are drawn in gray.}
\end{center}
\end{figure}
The basic idea of our construction for Theorem~\ref{thm:two} is to use a counter, the base of which is a function of $k$ and $N$. Then, we initialize the counter for our construction with a certain starting value and have it increment until its maximum value, at which point it rolls over to all 0's and the assembly goes terminal. Our construction is inspired by, but substantially different from, a similar construction by Aggarwal, Cheng, Goldwasser, Kao, Moisset de Espan\'{e}s and Schweller \cite{AGKS05g} for the self-assembly of two-dimensional $k \times N$ thin rectangles at temperature-2. Like theirs, our construction uses a counter whose base depends on $k$ and $N$. But unlike theirs, we represent each digit of the counter in our construction geometrically, using a one-bit-per-bump representation in an assembly that is three tiles wide and whose height is proportional to the binary representation of the base of the counter. See Figure~\ref{fig:example_low_high} for an example of the counter in our construction at different levels of granularity. The size of the tile set produced by the construction is $O\left(m + \log N\right) = O\left( N^{\frac{1}{\left \lfloor \frac{k}{3} \right \rfloor}} + \log N\right)$ and unique self-assembly follows from conditional determinism. The full construction of the tile set that proves Theorem~\ref{thm:two} is given in the remainder of this section.
\subsection{Notation for gadgets and figures}
In the context of the construction of a tile set, a \emph{gadget}, referred to by a name like {\tt Gadget}, is a group of tiles that perform a specific task as they self-assemble. All gadgets are depicted in a corresponding figure, where the input glue is explicitly specified by an arrow, output glues are inferred and glues internal to the gadget are configured to ensure unique self-assembly within the gadget. We say that a gadget is \emph{general} if its input and output glues are undefined. If {\tt Gadget} is a general gadget, then we use the notation ${\tt Gadget}({\tt a}, {\tt b})$ to represent the creation of the \emph{specific gadget}, or simply \emph{gadget}, referred to as {\tt Gadget}, with input glue label {\tt a} and output glue label {\tt b} (all positive glue strengths are $1$). If a gadget has two possible output glues, then we will use the notation ${\tt Gadget}({\tt a}, {\tt b}, {\tt c})$ to denote the specific version of {\tt Gadget}, where {\tt a} is the input glue and {\tt b} and {\tt c} are the two possible output glues, listed in the order north, east, south and west, with all of the $z=0$ output glues listed before the $z=1$ output glues. If a gadget has only one output glue (and no input glue), like a gadget that contains the seed, or if a gadget has only one input glue (and no output glue), then we will use the notation ${\tt Gadget}({\tt a})$. We use the notation $\langle \cdot \rangle$ to denote some standard encoding of the concatenation of a list of symbols.
Following standard presentation conventions for ``just-barely'' 3D tile self-assembly, we use big squares to represent tiles placed in the $z=0$ plane and small squares to represent tiles placed in the $z=1$ plane. A glue between a $z=0$ tile and a $z=1$ tile is denoted as a small black disk. Glues between $z=0$ tiles are denoted as thick lines. Glues between $z=1$ tiles are denoted as thin lines. The following is our main positive result.
\subsection{Parameters for the counter}
Since the height (number of tile rows) of each logical row in the counter depends on $k$ and $N$, we must choose its starting value carefully. Therefore, let $d=\left \lfloor \frac{k}{3} \right \rfloor$, $m=\left\lceil\left(\frac{N}{5}\right)^{\frac{1}{d}}\right\rceil$, $l=\left\lceil\log m\right\rceil+1$, $s=m^d-\left\lfloor\frac{N-3l-1}{3l+2}\right\rfloor$, $c=k\mod3$, and $r=N+1\mod 3l+2$, where $d$ is the number of digits in the counter, $m$ is the numerical base of the counter, $l$ is the number of bits needed to represent each digit in the counter's value plus one for the ``left edge'', $s$ is the numerical start of the counter, and $c$ and $r$ are the number of tile columns and tile rows, respectively, that must be filled in after and outside of the counter. Each digit of the counter requires a width of 3 tiles, which has a direct relation with the tile complexity of the construction. The values of $m$ and $s$ are chosen such that the counter stops just before reaching a height of $N$ tiles, at which point, the assembly is given a flat ``roof''. For example, in Figure~\ref{fig:example_low_high}, we have $d=3$, $m=3$, $l=3$, $s=23$, $c=2$, and $r=2$. We now informally justify each of the previously-defined parameters.
We define $d$ as the number of digit columns in the counter. The number of digits in the counter is limited by the width, $k$. Each digit column requires $3$ tiles. We maximize the range of the counter by maximizing the number of digit columns. Given those requirements and restrictions, we end up with $d=\left\lfloor\frac{k}{3}\right\rfloor$ digit columns.
Given at least one digit column, we have the ability to count to any number of counter rows so long as we choose an appropriate base. In that sense, choosing $m$, the base of the counter, is somewhat arbitrary. So long as the maximum height of our construction, where we count $m^d$ times, is greater than $N$, and the minimum height of our construction, where we count once, is less than $N$, then there exists a corresponding value of $s$ such that the counter stops just before reaching a height of $N$. We choose $m=\left\lceil\left(\frac{N}{5}\right)^{\frac{1}{d}}\right\rceil$.
The value of $l$, which is the number of binary bits needed to encode each digit of base-$m$, plus one for the ``left edge'', is easily verified by inspection.
Let $h$ be the height of the construction without any additional ``roof'' tiles. By inspection of the example construction in Figure~\ref{fig:Example_Full}, as well as the construction of the gadget units in subsequent sections, we see that the height of the construction without additional tiles is $3l+1$ for the \texttt{Seed} unit (see Section~\ref{sec:seed_unit}), plus $3l+2$ for each incrementation of the counter, adding a Counter unit (see Section~\ref{sec:counter_unit}) row each time. If we define $n$ as the number of Counter unit rows, then $h=n(3l+2)+3l+1$. So then the maximum height of the counter is $m^d(3l+2)+3l+1$.
\begin{lemma}
\label{lem:max-height-counter}
$N \leq m^d(3l+2)+3l+1$.
\end{lemma}
\begin{proof}
We have
\begin{eqnarray*}
N & = & 5\left(\frac{N}{5}\right) \
= \ 5\left( \left(\frac{N}{5}\right)^{\frac{1}{d}} \right)^d \
\leq \ 5 \left \lceil \left( \frac{N}{5} \right)^{\frac{1}{d}} \right \rceil^d \\
& = & 5m^d \
\leq \ 3lm^d+2m^d \
\leq \ 3lm^d+2m^d+3l+1 \\
& = & m^d(3l+2)+3l+1.
\end{eqnarray*}
\end{proof}
And the minimum height is $6l+3$.
\begin{lemma}
\label{lem:min-height-counter}
$6l+3 \leq N$, for $k\geq 3$ and sufficiently large values of $N$.
\end{lemma}
\begin{proof}
We have
\begin{eqnarray*}
6l+3 & = & 6 \left \lceil \log m \right \rceil + 9 \
\ = \ 6 \left \lceil \log \left \lceil \left( \frac{N}{5} \right)^{\frac{1}{d}} \right \rceil \right \rceil + 9 \\
& = & 6 \left \lceil \log \left( \left( \frac{N}{5} \right)^{\frac{1}{d}} \right) \right \rceil + 9 \quad\quad\quad\quad\quad \textmd{See \cite{Graham:1994} for a proof of this equality} \\
& \leq & 6 \log \left( \left( \frac{N}{5} \right)^{\frac{1}{d}} \right) + 15
\ = \ \frac{6}{d} \log \left( \frac{N}{5} \right) + 15 \\
& \leq & \frac{6}{d} \log N + 15
\ = \ \frac{6}{\left \lfloor \frac{k}{3} \right \rfloor } \log N+ 15 \\
& \leq & 6 \log N+ 15\quad\quad\quad\quad\quad\quad\textmd{Because } k \geq 3 \\
& \leq & N \quad\quad\quad\quad\quad\quad\quad\quad\quad\quad \textmd{For } N \geq 49.
\end{eqnarray*}
\end{proof}
By Lemma~\ref{lem:min-height-counter}, one row of the counter might result in a final assembly that will not be tall enough but Lemma~\ref{lem:max-height-counter} says that having all possible rows of the counter might result in a final assembly that is too tall. Therefore, we must start the counter at an appropriate value to get the correct height of the final assembly.
The counter can start at any whole number less than $m^d$ and ends when it reaches $0$ by rolling over $m^d-1$. This means that the number of Counter unit rows $n$, is $m^d-s$, where we have defined $s$ as the starting value of the counter. To choose the best starting value, we find the value for $n$ that gets $h$ close to $N$ without exceeding $N$. It follows from the equation $h=n(3l+2)+3l+1$, that $n=\left\lfloor\frac{N-3l-1}{3l+2}\right\rfloor$. Thus, $s=m^d-\left\lfloor\frac{N-3l-1}{3l+2}\right\rfloor$.
Recall that we must use $3d$ tile columns to encode the digits of the counter. Since the remaining tile columns must still be filled in to ensure a width of $k$, we must sometimes construct additional tile columns outside of the counter. We denote the number of additional tile columns by $c$ and conclude that its value is $k\mod3$.
Similarly, we must also account for the number of additional tile rows that must be filled in after the counter has finished counting. We denote the number of filler rows at the top of the construction with $r$, and conclude that it is the remainder of this quotient expression $\frac{N-3l-1}{3l+2}$, which is to say that $r=N-3l-1\mod 3l+2$. This modular expression simplifies to $r=N+1\mod 3l+2$.
\subsection{Tile set}
For the purposes of this construction, we assume the function $bin(a,b)$ gives the binary representation of $a$ that is truncated (or prepended with 0's) to a length of $b$. We will use the notation $\left(a\right)_b\left[i\right]$ to denote the digit in the $i$-th position from the right of $a$ in base-$b$. We define a \emph{gadget unit} as a collection of gadgets with a singular purpose. Gadgets belonging to the same gadget unit will have their figures grouped together. The set of all gadgets created in this subsection forms the tile set $T_{k,N}$.
\subsubsection{{\tt Vertical\_Column} tiles}
Since \texttt{Vertical\_Column} tiles are present in a majority of our gadget units, they will only be shown in Figure~\ref{fig:Vertical_Column}.
\begin{figure}[htp]
\begin{center}
\begin{subfigure}[t]{.4\textwidth}
\begin{center}
\includegraphics[width=12px]{Up_Column_Tile}
\caption{\label{fig:Up_Column_Tile}\texttt{Up\_Column\_Tile}}
\end{center}
\end{subfigure}
\begin{subfigure}[t]{.4\textwidth}
\begin{center}
\includegraphics[width=12px]{Down_Column_Tile}
\caption{\label{fig:Down_Column_Tile}\texttt{Down\_Column\_Tile}}
\end{center}
\end{subfigure}
\caption{\label{fig:Vertical_Column}\texttt{Vertical\_Column} tiles are used throughout the construction to adjust the height of gadget units.}
\end{center}
\end{figure}
\subsubsection{Seed unit}
\label{sec:seed_unit}
We begin by encoding the initial value of the counter with the Seed unit. It has $d$ columns, where each 3-wide column represents a digit of $s$ in base-$m$. A collection of bit-bumps on the columns' east sides encodes the digits into binary. A small ``lip'' is added on the west side of the Seed unit to increase the width of the assembly by $c$, which catches any vertical filler tiles at the end of the construction. The \texttt{Guess} tile on the east side of the unit initiates the first Counter unit. See Figure~\ref{fig:Seed}.
We define the Seed unit by creating the following gadgets:
\begin{itemize}
\item The first gadget depends on the value of $c$:
\begin{itemize}
\item If $c=0$, create $\texttt{Seed\_Start}\left(\left<\texttt{seed},\texttt{col},d,1\right>\right)$ from the general gadget in Figure~\ref{fig:Seed_Start_0}.
\item If $c=1$, create $\texttt{Seed\_Start}\left(\left<\texttt{seed},\texttt{col},d,1\right>\right)$ from the general gadget in Figure~\ref{fig:Seed_Start_1}.
\item If $c=2$, create $\texttt{Seed\_Start}\left(\left<\texttt{seed},\texttt{col},d,1\right>\right)$ from the general gadget in Figure~\ref{fig:Seed_Start_2}.
\end{itemize}
One gadget was created in this step.
\item For each $i=1,\ldots,d$:
\begin{itemize}
\item For each $j=1,\ldots,3l-3$, create\\$\texttt{Up\_Column}\left(\left<\texttt{seed},\texttt{col},i,j\right> \! ,\left<\texttt{seed},\texttt{col},i,j+1\right>\right)$ from the general gadget in Figure~\ref{fig:Up_Column_Tile}.
\item Create $\texttt{Seed\_Msb}\left(\left<\texttt{seed},\texttt{col},i,3l-2\right> \! ,\left<\texttt{seed},\texttt{bit},i,l-1\right>\right)$ from the general gadget in Figure~\ref{fig:Seed_Msb_0} if $((s)_m[i])_2[l]=0$ or Figure~\ref{fig:Seed_Msb_1} if $((s)_m[i])_2[l]=1$.
\item For each $j=2,\ldots,l-1$, create\\$\texttt{Seed\_Bit}\left(\left<\texttt{seed},\texttt{bit},i,j\right> \! ,\left<\texttt{seed},\texttt{bit},i,j-1\right>\right)$ from the general gadget in Figure~\ref{fig:Seed_Bit_0} if $((s)_m[i])_2[j]=0$ or Figure~\ref{fig:Seed_Bit_1} if $((s)_m[i])_2[j]=1$.
\end{itemize}
In this step, $\sum_{i=1}^d{\left(1 + \sum_{j=1}^{3l-3}{1} + \sum_{j=2}^{l-1}{1}\right)} = O(dl)$ gadgets were created. This means that
\begin{eqnarray*}
dl & = & O\left(d \log m\right) \\
& = & O\left( d \log \left \lceil \left(\frac{N}{5} \right)^{\frac{1}{d}} \right \rceil \right) \\
& = & O\left( d \log \left( 2\left( \frac{N}{5} \right)^{\frac{1}{d}} \right) \right) \\
& = & O\left( d \log \left( \frac{N}{5} \right)^{\frac{1}{d}} + d \log 2\right) \\
& = & O\left( \log N + \left \lfloor \frac{k}{3} \right \rfloor \right) \\
& = & O\left( \log N \right)
\end{eqnarray*}
gadgets were created in this step.
\item Create $\texttt{Seed\_Bit}\left(\left<\texttt{seed},\texttt{bit},d,1\right> \! ,\left<\texttt{seed},\texttt{bit},d,0\right>\right)$ from the general gadget in Figure~\ref{fig:Seed_Bit_1}. One gadget was created in this step.
\item For each $i=1,\ldots,d-1$:
\begin{itemize}
\item Create $\texttt{Seed\_Bit}\left(\left<\texttt{seed},\texttt{bit},i,1\right> \! ,\left<\texttt{seed},\texttt{bit},i,0\right>\right)$ from the general gadget in Figure~\ref{fig:Seed_Bit_0}.
\item Create $\texttt{Seed\_Spacer}\left(\left<\texttt{seed},\texttt{bit},i+1,0\right> \! ,\left<\texttt{seed},\texttt{col},i,1\right>\right)$ from the general gadget in Figure~\ref{fig:Seed_Spacer}.
\end{itemize}
In this step, $2(d-1) = O(k) = O(\log N)$ gadgets were created.
\item Create
\begin{align*}
\texttt{Seed\_End}( & \! \left<\texttt{seed},\texttt{bit},1,0\right> \!, \\
& \! \left<\texttt{inc},\texttt{read},1\right> \!, \\
& \! \left<\texttt{inc},\texttt{read},0\right> )
\end{align*}
One gadget was created in this step.
\end{itemize}
\begin{figure}[htp]
\begin{center}
\begin{subfigure}[b]{.2\textwidth}
\begin{center}
\includegraphics[width=12px]{Seed_Start_0}
\caption{\label{fig:Seed_Start_0}\texttt{Seed\_Start\_0}}
\vspace{.56\linewidth}
\includegraphics[width=27px]{Seed_Start_1}
\caption{\label{fig:Seed_Start_1}\texttt{Seed\_Start\_1}}
\vspace{.3\linewidth}
\includegraphics[width=42px]{Seed_Start_2}
\caption{\label{fig:Seed_Start_2}\texttt{Seed\_Start\_2}}
\end{center}
\end{subfigure}
\begin{subfigure}[b]{.2\textwidth}
\begin{center}
\includegraphics[width=42px]{Seed_Msb_0}
\caption{\label{fig:Seed_Msb_0}\texttt{Seed\_Msb\_0}}
\vspace{.17\linewidth}
\includegraphics[width=39px]{Seed_Msb_1}
\caption{\label{fig:Seed_Msb_1}\texttt{Seed\_Msb\_1}}
\vspace{.22\linewidth}
\includegraphics[width=42px]{Seed_Spacer}
\caption{\label{fig:Seed_Spacer}\texttt{Seed\_Spacer}}
\end{center}
\end{subfigure}
\begin{subfigure}[b]{.2\textwidth}
\begin{center}
\includegraphics[width=27px]{Seed_Bit_0}
\caption{\label{fig:Seed_Bit_0}\texttt{Seed\_Bit\_0}}
\vspace{.15\linewidth}
\includegraphics[width=24px]{Seed_Bit_1}
\caption{\label{fig:Seed_Bit_1}\texttt{Seed\_Bit\_1}}
\vspace{.15\linewidth}
\includegraphics[width=27px]{Seed_End}
\caption{\label{fig:Seed_End}\texttt{Seed\_End}}
\end{center}
\end{subfigure}
\vspace{20pt}
\begin{subfigure}[b]{.38\textwidth}
\begin{center}
\includegraphics[width=162px]{Example_Seed}
\caption{\label{fig:Example_Seed} The seed is the leftmost tile in the bottom row and self-assembly of the black tiles proceeds in a left-to-right fashion.}
\end{center}
\end{subfigure}
\caption{\label{fig:Seed}The Seed gadget unit. The actual seed tile is at the far-left of any of the \texttt{Seed\_Start} gadgets.}
\end{center}
\end{figure}
Since the number of tiles in each gadget in the Seed unit is $O(1)$, then based on the above computations, the number of gadgets created in this subsection is $O(\log N)$.
\subsubsection{Counter unit}
\label{sec:counter_unit}
For this construction, we require a set of $4m-1$ Counter gadget units to encode the digits of the counter, of which $m$ units will increment a digit of the counter, $m$ units will copy a digit of the counter, $m$ units will copy the most significant digit of the counter, and $m-1$ units will increment the most significant digit of the counter. Each row of the counter will have a total of $d$ Counter units. Each Counter unit reads over a series of bit-bumps protruding into their row from the preceding Seed unit or counter row. After a Counter unit reads its bit pattern with \texttt{Guess} tiles, the unit produces a new bit pattern in the row above it that encodes a copy or increment of the current digit. Of the ``less significant digit'' units, the one that increments $m-1$ to $0$ is unique because it initiates another increment unit, that is, a carry is passed to the digit to its left. Other increment units, as well as the copy units, will only initiate copy units (no carry is propagated to the left). The first bit read is always the ``left edge'' marker, which tells the unit if it represents the most significant digit of the counter value and needs to start a new row instead of another Counter unit. The counter terminates when the most significant digit follows an increment unit and reads $m-1$ in its column. At that point, the counter will have rolled over $m^d-1$ and the Roof unit takes over. The gadgets belonging to the Counter units are shown in Figure~\ref{fig:Counter}.
We define the Counter units by creating the following gadgets:
\begin{itemize}
\item For each $i = 0, \ldots, l - 2$ and each $u \in \{0,1\}^i$:
\begin{itemize}
\item Create
\begin{align*}
\texttt{Counter\_Read}( & \! \left<\texttt{inc},\texttt{read},0u\right> \!, \\
& \! \left<\texttt{inc},\texttt{read},10u\right> \!, \\
& \! \left<\texttt{inc},\texttt{read},00u\right> )
\end{align*}
from the general gadget in Figure~\ref{fig:Counter_Read_0}.
\item Create
\begin{align*}
\texttt{Counter\_Read}( & \! \left<\texttt{inc},\texttt{read},1u\right> \!, \\
& \! \left<\texttt{inc},\texttt{read},11u\right> \!, \\
& \! \left<\texttt{inc},\texttt{read},01u\right> )
\end{align*}
from the general gadget in Figure~\ref{fig:Counter_Read_1}.
\item Create
\begin{align*}
\texttt{Counter\_Read}( & \! \left<\texttt{copy},\texttt{read},0u\right> \!, \\
& \! \left<\texttt{copy},\texttt{read},10u\right> \!, \\
& \! \left<\texttt{copy},\texttt{read},00u\right> )
\end{align*}
from the general gadget in Figure~\ref{fig:Counter_Read_0}.
\item Create
\begin{align*}
\texttt{Counter\_Read}( & \! \left<\texttt{copy},\texttt{read},1u\right> \!, \\
& \! \left<\texttt{copy},\texttt{read},11u\right> \!, \\
& \! \left<\texttt{copy},\texttt{read},01u\right> )
\end{align*}
from the general gadget in Figure~\ref{fig:Counter_Read_1}.
\end{itemize}
These are read gadgets for all digit positions and all bits (except the most significant bit) for both copy and increment columns. In this step, $\sum_{i=0}^{l-2}{4\cdot 2^i} = 4 \left( 2^{l-1} - 1 \right) = 4 \left( 2^{\left \lceil \log m \right \rceil } - 1\right) \leq 4 \left( 2\cdot 2^{\log m}\right) = O(m) = O\left(N^{\frac{1}{\left \lfloor \frac{k}{3} \right \rfloor}}\right)$ gadgets were created.
\end{itemize}
Recall that the least significant bit of our digit columns represents whether the digit is for the most significant digit's place or not. It follows then, that $bin(2m-2,l)$ is the greatest digit of our base-$m$ counter but with an extra $0$ at the end, which is to say that it is m-1 but not for the most significant digit's place. To get $m-1$ for the most significant digit's place, we simply add a $1$ to that expression and get $bin(2m-1,l)$. Hence, we use the range $0$ to $2m-1$ to refer to all digits of base-$m$ in all positions / places, we use the range $0$ to $2m-3$ to refer to all digits of base-$m$ except $m-1$ in all positions / places, we use the index $2m-2$ to refer to the counter value $m-1$ when not in the most significant digit's place, and we use the index $2m-1$ to refer to the counter value $m-1$ when in the most significant digit's place.
\begin{itemize}
\item For each $i=0,\ldots,2m-1$, create
\begin{align*}
{\tt Counter\_Read\_Msb}( & \! \left \langle \texttt{copy},\texttt{read},bin(i,l) \right \rangle \! , \\
& \! \left \langle \texttt{copy},\texttt{write},bin(i,l) \right \rangle \! , \\
& \! \left \langle \texttt{d\_fill} \right \rangle)
\end{align*}
from the general gadget in Figure~\ref{fig:Counter_Read_Msb_0} if $(i)_2[l]=0$ or Figure~\ref{fig:Counter_Read_Msb_1} if $(i)_2[l]=1$. These are all read gadgets for the most significant bit in copy columns only. In this step, $2m = O\left(N^{\frac{1}{\left \lfloor \frac{k}{3} \right \rfloor}}\right)$ gadgets were created.
\item For each $i=0,\ldots,2m-3$, create
\begin{align*}
{\tt Counter\_Read\_Msb}( & \! \left \langle \texttt{inc},\texttt{read},bin(i,l) \right \rangle \! , \\
& \! \left \langle \texttt{copy},\texttt{write},bin(i+2,l) \right \rangle \! , \\
& \! \left \langle \texttt{d\_fill} \right \rangle )
\end{align*}
from the general gadget in Figure~\ref{fig:Counter_Read_Msb_0} if $(i)_2[l]=0$ or Figure~\ref{fig:Counter_Read_Msb_1} if $(i)_2[l]=1$. These are all read gadgets for the most significant bit in increment columns but only when the digit value is between $0$ and $m - 2$. In this step, $2m-2 = O\left(N^{\frac{1}{\left \lfloor \frac{k}{3} \right \rfloor}}\right)$ gadgets were created.
\item Create
\begin{align*}
\texttt{Counter\_Read\_Msb}( & \! \left \langle \texttt{inc},\texttt{read},bin(2m-2,l)\right \rangle \!, \\
& \! \left \langle \texttt{inc},\texttt{write\_all\_0s},1 \right \rangle \!, \\
& \! \left \langle \texttt{d\_fill} \right \rangle )
\end{align*}
from the general gadget in Figure~\ref{fig:Counter_Read_Msb_1}. This is the read gadget for the most significant bit in all increment columns (except the most significant digit) but only when the digit value is $m - 1$. One gadget was created in this step.
\item For each $i = 1, \ldots, l - 1$, create\\${\tt Counter\_Write}(\langle {\tt inc}, {\tt write\_all\_0s}, i\rangle, \langle {\tt inc}, {\tt write\_all\_0s}, i+1\rangle)$ from the general gadget in Figure~\ref{fig:Counter_Write_0}. These are the write gadgets for a digit value of all 0s due to incrementation of $m -1$. In this step, $l-1 = O(\log m) = O\left(\frac{\log N}{k}\right) = O(\log N)$ gadgets were created. This group of gadgets writes a series of 0 bits, because $m-1$ was incremented to $0$.
\item For each $u \in \{0,1\}^{l-1}$:
\begin{itemize}
\item Create $\texttt{Counter\_Write}\left(\left<\texttt{copy},\texttt{write},u0\right> \!,\left<\texttt{copy},\texttt{write},u\right>\right)$ from the general gadget in Figure~\ref{fig:Counter_Write_0}.
\item Create $\texttt{Counter\_Write}\left(\left<\texttt{copy},\texttt{write},u1\right> \! ,\left<\texttt{msd},\texttt{write},u\right>\right)$ from the general gadget in Figure~\ref{fig:Counter_Write_1}. Note that ``{\tt msd}'' stands for ``most significant digit'' (since the ``left edge'' bit is 1 in this case).
\end{itemize}
These are all gadgets to write the ``left edge'' marker in all copy columns. In this step, $2\cdot 2^{l-1} = 2\cdot 2^{\left \lceil \log m \right \rceil} \leq 2 \left( 2\cdot 2^{ \log m}\right) = O(m) = O\left(N^{\frac{1}{\left \lfloor \frac{k}{3} \right \rfloor}}\right)$ gadgets were created.
\item For each $i = 1, \ldots, l - 2$ and each $u \in \{0,1\}^i$:
\begin{itemize}
\item Create $\texttt{Counter\_Write}\left(\left<\texttt{copy},\texttt{write},u0\right> \!,\left<\texttt{copy},\texttt{write},u\right>\right)$ from the general gadget in Figure~\ref{fig:Counter_Write_0}.
\item Create $\texttt{Counter\_Write}\left(\left<\texttt{copy},\texttt{write},u1\right> \!,\left<\texttt{copy},\texttt{write},u\right>\right)$ from the general gadget in Figure~\ref{fig:Counter_Write_1}.
\item Create $\texttt{Counter\_Write}\left(\left<\texttt{msd},\texttt{write},u0\right> \!,\left<\texttt{msd},\texttt{write},u\right>\right)$ from the general gadget in Figure~\ref{fig:Counter_Write_0}.
\item Create $\texttt{Counter\_Write}\left(\left<\texttt{msd},\texttt{write},u1\right> \!,\left<\texttt{msd},\texttt{write},u\right>\right)$ from the general gadget in Figure~\ref{fig:Counter_Write_1}.
\end{itemize}
These gadgets write all digits (except the ``left edge'' marker) in all copy columns. In this step,
\begin{eqnarray*}
\sum_{i=1}^{l-2}{4\cdot 2^i} & = & 4 \left( 2^{l-1} - 2 \right) \\
& = & 4 \left( 2^{\left \lceil \log m \right \rceil } - 2\right) \\
& \leq & 4 \left( 2\cdot 2^{\log m}\right) \\
& = & O(m) = O\left(N^{\frac{1}{\left \lfloor \frac{k}{3} \right \rfloor}}\right)
\end{eqnarray*}
gadgets were created.
\item Create $\texttt{Counter\_Write\_Msb}\left(\left<\texttt{inc},\texttt{write},l\right> \!,\left<\texttt{inc},\texttt{down\_z\_0},1\right>\right)$ from the general gadget in Figure~\ref{fig:Counter_Write_Msb_0}. This gadget writes 0 for the most significant bit when digit value is all 0s after incrementing $m - 1$. One gadget was created in this step.
\item Create $\texttt{Counter\_Write\_Msb}\left(\left<\texttt{copy},\texttt{write},0\right> \!,\left<\texttt{copy},\texttt{down\_z\_0},1\right>\right)$ from the general gadget in Figure~\ref{fig:Counter_Write_Msb_0}. This gadget writes 0 for the most significant bit in any copy column except the most significant digit column. One gadget was created in this step.
\item Create $\texttt{Counter\_Write\_Msb}\left(\left<\texttt{copy},\texttt{write},1\right> \!,\left<\texttt{copy},\texttt{down\_z\_0},1\right>\right)$ from the general gadget in Figure~\ref{fig:Counter_Write_Msb_1}. This gadget writes 1 for the most significant bit in any copy column except the most significant digit column. One gadget was created in this step.
\item Create $\texttt{Counter\_Write\_Msb}\left(\left<\texttt{msd},\texttt{write},0\right> \!,\left<\texttt{msd},\texttt{down\_z\_0},1\right>\right)$ from the general gadget in Figure~\ref{fig:Counter_Write_Msb_0}. This gadget writes 0 for the most significant bit in the most significant digit (i.e., copy) column. One gadget was created in this step.
\item Create $\texttt{Counter\_Write\_Msb}\left(\left<\texttt{msd},\texttt{write},1\right> \!,\left<\texttt{msd},\texttt{down\_z\_0},1\right>\right)$ from the general gadget in Figure~\ref{fig:Counter_Write_Msb_1}. This gadget writes 1 for the most significant bit in the most significant digit (i.e., copy) column. One gadget was created in this step.
\item For each $i=1,\ldots,3l-2$:
\begin{itemize}
\item Create $\texttt{Down\_Column}\left(\left<\texttt{inc},\texttt{down\_z\_0},i\right> \!,\left<\texttt{inc},\texttt{down\_z\_0},i+1\right>\right)$ from the general gadget in Figure~\ref{fig:Down_Column_Tile}.
\item Create $\texttt{Down\_Column}\left(\left<\texttt{copy},\texttt{down\_z\_0},i\right> \!,\left<\texttt{copy},\texttt{down\_z\_0},i+1\right>\right)$ from the general gadget in Figure~\ref{fig:Down_Column_Tile}.
\end{itemize}
These gadgets go back down to the bottom of the new counter row in all columns (except the most significant digit column) while remembering an increment or copy is taking place. In this step, $2(3l-2) = O(l) = O(\log m) = O\left( \frac{\log N}{k} \right) = O(\log N)$ gadgets were created.
\item For each $i=1,\ldots,3l-3$, create\\$\texttt{Down\_Column}\left(\left<\texttt{msd},\texttt{down\_z\_0},i\right> \!,\left<\texttt{msd},\texttt{down\_z\_0},i+1\right>\right)$ from the general gadget in Figure~\ref{fig:Down_Column_Tile}. These gadgets go back down to the bottom of the new counter row in the most significant digit column while remembering an increment or copy is taking place. In this step, $3l-3 = O(l)=O(\log m) = O\left(\frac{\log N}{k}\right) = O(\log N)$ gadgets were created.
\item Create\\$\texttt{Counter\_Return\_Column\_Start}\left(\left<\texttt{inc},\texttt{down\_z\_0},3l-1\right> \!,\left<\texttt{inc},\texttt{down\_z\_1},1\right>\right)$ from the general gadget in Figure~\ref{fig:Counter_Return_Column_Start}. This is the first gadget to start going down to the bottom of the previous counter row in increment columns only. One gadget was created in this step.
\item Create\\$\texttt{Counter\_Return\_Column\_Start}\left(\left<\texttt{copy},\texttt{down\_z\_0},3l-1\right> \!,\left<\texttt{copy},\texttt{down\_z\_1},1\right>\right)$ from the general gadget in Figure~\ref{fig:Counter_Return_Column_Start}. This is the first gadget to start going down to the bottom of the previous counter row in copy columns only. One gadget was created in this step.
\item For each $i=1,\ldots,l-1$:
\begin{itemize}
\item Create\\$\texttt{Counter\_Return\_Column}\left(\left<\texttt{inc},\texttt{down\_z\_1},i\right> \!,\left<\texttt{inc},\texttt{down\_z\_1},i+1\right>\right)$ from the general gadget in Figure~\ref{fig:Counter_Return_Column}.
\item Create\\$\texttt{Counter\_Return\_Column}\left(\left<\texttt{copy},\texttt{down\_z\_1},i\right> \!,\left<\texttt{copy},\texttt{down\_z\_1},i+1\right>\right)$ from the general gadget in Figure~\ref{fig:Counter_Return_Column}.
\end{itemize}
These gadgets go back down to the bottom of the previous counter row in all columns (except the most significant digit column) while remembering an increment or copy is taking place. In this step, $2(l-1) = O(l) = O(\log m) = O\left( \frac{\log N}{k} \right) = O(\log N)$ gadgets were created.
\item Create
\begin{align*}
\texttt{Counter\_Return\_Column\_End}( & \! \left<\texttt{inc},\texttt{down\_z\_1},l\right> \!, \\
& \! \left<\texttt{inc},\texttt{read},1\right> \!, \\
& \! \left<\texttt{inc},\texttt{read},0\right> )
\end{align*}
from the general gadget in Figure~\ref{fig:Counter_Return_Column_End}. This gadget gets ready to read the ``left edge'' marker in the next digit (to the left) in order to increment it. One gadget was created in this step.
\item Create
\begin{align*}
\texttt{Counter\_Return\_Column\_End}( & \! \left<\texttt{copy},\texttt{down\_z\_1},l\right> \!, \\
& \! \left<\texttt{copy},\texttt{read},1\right> \!, \\
& \! \left<\texttt{copy},\texttt{read},0\right> )
\end{align*}
from the general gadget in Figure~\ref{fig:Counter_Return_Column_End}. This gadget gets ready to read the ``left edge'' marker in the next digit (to the left) in order to copy it. One gadget was created in this step.
\end{itemize}
Since the number of tiles in each gadget in the Counter unit is $O(1)$, then based on the above computations, the number of gadgets (and therefore the number of tile types) created in this subsection is $O\left(N^{\frac{1}{\left \lfloor \frac{k}{3} \right \rfloor}} + \log N\right)$.
\begin{figure}[htp]
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=12px]{Counter_Read_0}
\caption{\label{fig:Counter_Read_0}\texttt{Counter\_Read\_0}}
\end{subfigure}
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=12px]{Counter_Read_1}
\caption{\label{fig:Counter_Read_1}\texttt{Counter\_Read\_1}}
\end{subfigure}
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=27px]{Counter_Read_Msb_0}
\caption{\label{fig:Counter_Read_Msb_0}\texttt{Counter\_Read\_Msb\_0}}
\end{subfigure}
\vspace{10pt}
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=27px]{Counter_Read_Msb_1}
\caption{\label{fig:Counter_Read_Msb_1}\texttt{Counter\_Read\_Msb\_1}}
\end{subfigure}
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=27px]{Counter_Write_0}
\caption{\label{fig:Counter_Write_0}\texttt{Counter\_Write\_0}}
\end{subfigure}
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=24px]{Counter_Write_1}
\caption{\label{fig:Counter_Write_1}\texttt{Counter\_Write\_1}}
\end{subfigure}
\vspace{10pt}
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=42px]{Counter_Write_Msb_0}
\caption{\label{fig:Counter_Write_Msb_0}\texttt{Counter\_Write\_Msb\_0}}
\end{subfigure}
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=39px]{Counter_Write_Msb_1}
\caption{\label{fig:Counter_Write_Msb_1}\texttt{Counter\_Write\_Msb\_1}}
\end{subfigure}
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=12px]{Counter_Return_Column_Start}
\caption{\label{fig:Counter_Return_Column_Start}\texttt{Counter\_Return\_Column\_Start}}
\end{subfigure}
\vspace{10pt}
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=6px]{Counter_Return_Column}
\caption{\label{fig:Counter_Return_Column}\texttt{Counter\_Return\_Column}}
\end{subfigure}
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=24px]{Counter_Return_Column_End}
\caption{\label{fig:Counter_Return_Column_End}\texttt{Counter\_Return\_Column\_End}}
\end{subfigure}
\caption{\label{fig:Counter}The Counter gadget unit. A row of units shares half of its space with the row above and the other half of its space with the row below.}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=120px]{Example_Counter}
\caption{\label{fig:Example_Counter} An example with a row of Counter units colored black. The unit is incrementing 10-10-00 to 10-10-01.}
\end{figure}
\subsubsection{Return Row unit}
\label{sec:return_row_unit}
To begin a new row of Counter units following the completion of the current row, a Return Row gadget unit must return the frontier of the assembly to the east side of the construction. After returning the frontier to the east side, it initiates an incrementing Counter unit for the least significant digit. In cases where there is only one digit column in use, a single special gadget is used instead of the multi-piece unit. See Figure~\ref{fig:Return}
We define the Return Row unit by creating the following gadgets:
\begin{itemize}
\item If $d=1$, create
\begin{align*}
{\tt Return\_Row\_Single}( & \! \left \langle \texttt{return},\texttt{start}\right \rangle \! , \\
& \! \left \langle \texttt{inc},\texttt{read},1\right \rangle \!, \\
& \! \left \langle \texttt{d\_fill} \right \rangle \! , \\
& \! \left \langle \texttt{inc},\texttt{read},0\right \rangle )
\end{align*}
from the general gadget in Figure~\ref{fig:Return_Row_Single}. One gadget was created in this step.
\item Otherwise:
\begin{itemize}
\item Create
\begin{align*}
\texttt{Return\_Row\_Start}( & \! \left<\texttt{msd\_down\_z\_0},3l - 2\right> \!, \\
& \! \left \langle \texttt{d\_fill} \right \rangle \!, \\
& \! \left<\texttt{return},1\right> )
\end{align*}
from the general gadget in Figure~\ref{fig:Return_Row_Start}. One gadget was created in this step.
\item For each $i=1,\ldots,d-2$, create $\texttt{Return\_Row}\left(\left<\texttt{return},i\right> \! ,\left<\texttt{return},i+1\right>\right)$ from the general gadget in Figure~\ref{fig:Return_Row}. In this step, $d-2 = O(k) = O(\log N)$ gadgets were created.
\item Create
\begin{align*}
\texttt{Return\_Row\_End}( & \! \left<\texttt{return},d-1\right> \!, \\
& \! \left<\texttt{inc},\texttt{read},1\right> \!, \\
& \! \left<\texttt{inc},\texttt{read},0\right> )
\end{align*}
from the general gadget in Figure~\ref{fig:Return_Row_End}. One gadget was created in this step.
\end{itemize}
\end{itemize}
Since the number of tiles in each gadget in the Return Row unit is $O(1)$, then based on the above computations, the number of gadgets (and therefore the number of tile types) created in this subsection is $O\left( \log N \right)$.
\begin{figure}[htp]
\begin{center}
\begin{subfigure}[b]{.49\textwidth}
\begin{center}
\includegraphics[width=43.5px]{Return_Row_Start}
\caption{\label{fig:Return_Row_Start} \texttt{Return\_Row\_Start}}
\vspace{.15\linewidth}
\includegraphics[width=52.5px]{Return_Row}
\caption{\label{fig:Return_Row} \texttt{Return\_Row} }
\vspace{.15\linewidth}
\includegraphics[width=51px]{Return_Row_End}
\caption{\label{fig:Return_Row_End} \texttt{Return\_Row\_End} }
\vspace{.15\linewidth}
\includegraphics[width=42px]{Return_Row_Single}
\caption{\label{fig:Return_Row_Single} \texttt{Return\_Row\_Single} }
\end{center}
\end{subfigure}
\begin{subfigure}[b]{.49\textwidth}
\begin{center}
\includegraphics[width=132px]{Example_Return_Row}
\caption{\label{fig:Example_Return_Row} An example with the Return Row unit colored black.}
\end{center}
\end{subfigure}
\caption{\label{fig:Return}The Return Row gadget unit.}
\end{center}
\end{figure}
\subsubsection{Roof unit}
\label{sec:roof_unit}
The Roof gadget unit consists of a vertical tile column that reaches above the tiles from the last counter row, then extends the assembly to a height of $N$. Along the column are glues placed periodically that accept filler extensions. These extensions are designed to patch up holes that are left in the last counter row. The highest tile in the column has west-facing and east-facing glues. These glues accept shingle tiles which extend the roof westward and eastward so that the entire construction is ``covered''. (The eastward expansion is blocked if $r=0$.) Each shingle tile has a south-facing glue that binds to a filler tile, which will cover up any remaining gaps in the $k\times N$ shape. The Roof unit is shown in Figure~\ref{fig:Roof}.
We define the Roof unit by creating the following gadgets:
\begin{itemize}
\item Create $\texttt{Up\_Column}\left(\left<\texttt{inc},\texttt{read},bin(2m-1,l)\right> \! ,\left<\texttt{roof},\texttt{col},2\right>\right)$ from the general gadget in Figure~\ref{fig:Up_Column_Tile}. One gadget was created in this step.
\item Create $\texttt{Up\_Column}\left(\left<\texttt{roof},\texttt{col},2\right> \! ,\left<\texttt{roof},\texttt{col},3\right>\right)$ from the general gadget in Figure~\ref{fig:Up_Column_Tile}. One gadget was created in this step.
\item For each $i=3,\ldots,l+2$, create
\begin{align*}
\texttt{Roof\_Chimney}( & \! \left<\texttt{roof},\texttt{col},i\right> \!, \\
& \! \left<\texttt{roof},\texttt{col},i+1\right> \!, \\
& \! \left<\texttt{roof},\texttt{filler},1\right> )
\end{align*}
from the general gadget in Figure~\ref{fig:Roof_Chimney}. In this step, $(l+2)-3+1 = O(l) = O(\log m) = O\left( \frac{\log N}{k} \right) = O(\log N)$ gadgets were created.
\item For each $i=1,\ldots,d-1$, create\\$\texttt{Roof\_Filler}\left(\left<\texttt{roof},\texttt{filler},i\right> \! ,\left<\texttt{roof},\texttt{filler},i+1\right>\right)$ from the general gadget in Figure~\ref{fig:Roof_Filler}. In this step, $d - 1 = O(k) = O(\log N)$ gadgets were created.
\item For each $i=l+3,\ldots,l+r+4$,\\create $\texttt{Up\_Column}\left(\left<\texttt{roof},\texttt{col},i\right> \! ,\left<\texttt{roof},\texttt{col},i+1\right>\right)$ from the general gadget in Figure~\ref{fig:Up_Column_Tile}. In this step, $r+2 = O(l) = O(\log m) = O\left( \frac{\log N}{k} \right) = O(\log N)$ gadgets were created.
\item Create
\begin{align*}
\texttt{Roof\_Cap}( & \! \left<\texttt{roof},\texttt{col},l+r+5\right> \!, \\
& \! \left<\texttt{roof},\texttt{r\_shingle},1\right> \!, \\
& \! \left<\texttt{roof},\texttt{l\_shingle},1\right> )
\end{align*}
from the general gadget in Figure~\ref{fig:Roof_Cap}. One gadget was created in this step.
\item For each $i=1,\ldots,c+2$, create
\begin{align*}
{\tt Roof\_Left\_Shingle}( & \! \left \langle \texttt{roof},\texttt{l\_shingle},i \right \rangle \! , \\
& \! \left \langle \texttt{d\_fill} \right \rangle \! , \\
& \! \left \langle \texttt{roof},\texttt{l\_shingle},i+1 \right \rangle )
\end{align*}
from the general gadget in Figure~\ref{fig:Roof_Left_Shingle}. In this step, $c+2 = O(1)$ gadgets were created.
\item If $r>0$, then for each $i=1,\ldots,3d-3$, create
\begin{align*}
{\tt Roof\_Right\_Shingle}( &\! \left \langle \texttt{roof},\texttt{r\_shingle},i \right \rangle \! , \\
& \! \left \langle \texttt{roof},\texttt{r\_shingle},i+1 \right \rangle \!, \\\
& \! \left \langle \texttt{d\_fill} \right \rangle )
\end{align*}
from the general gadget in Figure~\ref{fig:Roof_Right_Shingle}. In this step, if $r > 0$, then $3d-3 = O(d) = O(k) = O(\log N)$ gadgets were created.
\end{itemize}
Since the number of tiles in each gadget in the Roof is $O(1)$, then based on the above computations, the number of gadgets (and therefore the number of tile types) created in this subsection is $O(\log N)$.
\begin{figure}[htp]
\begin{center}
\begin{subfigure}[b]{.49\textwidth}
\begin{center}
\includegraphics[width=13.5px]{Roof_Chimney}
\caption{\label{fig:Roof_Chimney}\texttt{Roof\_Chimney}}
\vspace{.1\linewidth}
\includegraphics[width=52.5px]{Roof_Filler}
\caption{\label{fig:Roof_Filler}\texttt{Roof\_Filler}}
\vspace{.1\linewidth}
\includegraphics[width=15px]{Roof_Cap}
\caption{\label{fig:Roof_Cap}\texttt{Roof\_Cap}}
\vspace{.1\linewidth}
\includegraphics[width=22.5px]{Roof_Left_Shingle}
\caption{\label{fig:Roof_Left_Shingle}\texttt{Roof\_Left\_Shingle}}
\vspace{.1\linewidth}
\includegraphics[width=22.5px]{Roof_Right_Shingle}
\caption{\label{fig:Roof_Right_Shingle}\texttt{Roof\_Right\_Shingle}}
\end{center}
\end{subfigure}
\begin{subfigure}[b]{.49\textwidth}
\begin{center}
\includegraphics[width=162px]{Example_Roof}
\caption{\label{fig:Example_Roof} An example with the Roof unit colored black.}
\end{center}
\end{subfigure}
\caption{\label{fig:Roof}The Roof gadget unit.}
\end{center}
\end{figure}
\subsection{Proof of Theorem~\ref{thm:two}}
We use the \emph{conditional determinism} method of Lutz and Shutters \cite{LutzShutters12} to prove that our construction is directed. Conditional determinism is weaker than the local determinism method of Soloveichik and Winfree \cite{SolWin07} but strong enough to imply unique production. Although Lutz and Shutters define conditional determinism for a 2D TAS, their result holds in 3D as well, namely, if $\mathcal{T}$ is a conditionally deterministic TAS, then $\mathcal{T}$ is directed.
We merely give a brief, intuitive review of conditional determinism (see \cite{LutzShutters12} for a thorough discussion). Let $\mathcal{T} = (T,\sigma,\tau)$ be a TAS and $\vec{\alpha}$ be an assembly sequence of $\mathcal{T}$ with result $\alpha$. Conditional determinism relies on the notion of \emph{dependence} (of one position on another position in an assembly). For a position $\vec{m} \in \textmd{dom}\left({\alpha}\right)$ and some unit vector\\$\vec{u} \in \{(0,1,0),(1,0,0),(0,-1,0),(-1,0,0),(0,0,1),(0,0,-1)\}$, we say that $\vec{m}+\vec{u}$ \emph{depends} on $\vec{m}$, or $\vec{m}+\vec{u}$ is a dependent of $\vec{m}$, if, in every assembly sequence in $\mathcal{T}$, a tile is placed at $\vec{m}$ before a tile is placed at $\vec{m}+\vec{u}$. Intuitively, $\vec{\alpha}$ is conditionally deterministic if the following three properties are satisfied: (1) every tile placed by $\vec{\alpha}$ binds with exactly temperature $\tau$, (2) for every position $\vec{m} \in \textmd{dom}\;{\alpha}$, if $\alpha(\vec{m})$ and the union of its immediate output and dependent neighbors (adjacent positions that depend on $\vec{m}$) are removed from $\alpha$ to get $\alpha'$, then $\alpha(\vec{m})$ is the only tile that may bind at position $\vec{m}$ in $\alpha'$, and (3) $\alpha$ is terminal. A TAS is conditionally deterministic if it exhibits a conditionally deterministic assembly sequence. We are now ready to prove Theorem~\ref{thm:two}.
Assume that $\sigma$ is the seed assembly that consists of the single seed tile that was previously specified and let $\mathcal{T}_{k,N} = \left(T_{k,N},\sigma,1\right)$. We are now ready to prove Theorem~\ref{thm:two}.
\begin{proof}(of Theorem~\ref{thm:two}) \\
First, note that $\left|T_{k,N}\right| = O\left( N^{\frac{1}{\left \lfloor \frac{k}{3} \right \rfloor}} + \log N\right)$ because there are two\\{\tt Vertical\_Column} tile types, a constant number of gadget units (Seed, Counter, Return Row, and Roof), the Counter unit contributes $O\left( N^{\frac{1}{\left \lfloor \frac{k}{3} \right \rfloor}} + \log N \right)$ tile types and the Seed, Return Row and Roof units each contribute $O\left( \log N \right)$ tile types.
Now, let $\vec{\alpha}$ be the assembly sequence in $\mathcal{T}_{k,N}$, with result $\alpha$, such that the order in which $\vec{\alpha}$ places tiles can be inferred from Figure~\ref{fig:Example_Full}, and with the additional constraint that, when building the assembly sequence in a depth-first manner, $\vec{\alpha}$ always tries to place as many tiles as possible in the $z=0$ plane before placing a tile in the $z=1$ plane, breaking ties on which direction in which to proceed based on the ordering: north, east, south and finally west. Thus, $R^3_{k,N}$ self-assembles in $\mathcal{T}_{k,N}$. To show that $\mathcal{T}_{k,N}$ is conditionally deterministic, and therefore directed, it suffices to show that $\vec{\alpha}$ is conditionally deterministic.
To prove that $\vec{\alpha}$ is conditionally deterministic, first note that, by the way we construct all the gadgets, all tiles placed by $\vec{\alpha}$ initially bind deterministically with exactly strength $\tau = 1$. Now let $\vec{m} \in \textmd{dom}\;{\alpha}$ and\\$\vec{u} \in \{(0,1,0),(1,0,0),(0,-1,0),(-1,0,0),(0,0,1),(0,0,-1)\}$ such that $\alpha(\vec{m}+\vec{u})$ is defined. The following statements can be verified by inspection of how the gadgets defined in the previous subsection attach to each other:
\begin{enumerate}
\item If $\alpha(\vec{m})$ has a strength-$1$ glue in direction $\vec{u}$ and $\alpha(\vec{m}+\vec{u})$ has a strength-$0$ glue in direction $-\vec{u}$, then $\vec{m}$ depends on $\vec{m} + \vec{u}$ in every assembly sequence in $\mathcal{T}_{k,N}$.
\item If $\alpha(\vec{m})$ has a strength-$1$ glue in direction $\vec{u}$ and $\alpha(\vec{m}+\vec{u})$ has a strength-$1$ glue in direction $-\vec{u}$, then there are two sub-cases to consider, depending on whether the glues interact:
\begin{enumerate}
\item The case where the glues are not equal and therefore do not interact does not happen in our construction. So, ignore this case.
\item If the glues are equal, then, by the way we construct the gadgets, in every assembly sequence in $\mathcal{T}_{k,N}$, $\vec{m}+\vec{u}$ is an output side of $\alpha\left(\vec{m}\right)$ or $\vec{m}+\vec{u}$ is the unique input side of $\alpha\left(\vec{m}\right)$.
\end{enumerate}
\end{enumerate}
We need not consider the cases where $\alpha(\vec{m})$ has a strength-$0$ glue in direction $\vec{u}$ and $\alpha(\vec{m}+\vec{u})$ has a strength-$0$ glue in direction $-\vec{u}$ and either one depends on the other or not. Thus, for every position $\vec{m} \in \textmd{dom}\;{\alpha}$, if $\alpha(\vec{m})$ and the union of its immediate output and dependent neighbors are removed from $\alpha$ to get $\alpha'$, then $\alpha(\vec{m})$ is the only tile that may bind (via its unique input side) at position $\vec{m}$ in $\alpha'$. Finally, since $\alpha$ is terminal, we may conclude that $\vec{\alpha}$ is conditionally deterministic and therefore $\mathcal{T}_{k,N}$ is directed.
\end{proof}
\begin{corollary}
If $R^3_{k,N}$ is a thin rectangle, then $K^1_{USA}\left( R^3_{k,N} \right) = O\left( N^{\frac{1}{\left \lfloor \frac{k}{3} \right \rfloor}} \right)$.
\end{corollary}
\begin{proof}
By the definition of thin rectangle, we have
$$
k < \frac{\log N}{\log \log N - \log \log \log N} < \frac{\log N}{\log \log N - \frac{1}{2}\log \log N} = \frac{2\log N}{\log \log N},
$$
where the second inequality holds for $N > 2^{2^{4}}$.
Then we have
\begin{eqnarray*}
\log N & = & 2^{\log \log N}
\ = \ \left( N^{\frac{1}{\log N}} \right)^{\log \log N}
\ = \ \left( N^{\frac{1}{\frac{\log N}{\log \log N}}} \right)^{\frac{2}{2}}
\ = \ N^{\frac{2}{\frac{2 \log N}{\log \log N}}} \\
& = & O\left( N^{\frac{2}{k}} \right) \ = \ O\left( N^{\frac{1}{\frac{k}{2}}} \right) \ = \ O\left( N^{\frac{1}{\frac{k}{3}}} \right) \ = \ O\left( N^{\frac{1}{\left \lfloor \frac{k}{3} \right \rfloor}} \right). \\
\end{eqnarray*}
\end{proof}
\section{Preliminaries}
\label{sec:prelims}
\newcommand{{\rm label}}{{\rm label}}
\newcommand{\colorf}[1]{\color(#1)}
\newcommand{{\rm str}}{{\rm str}}
\newcommand{\strengthf}[1]{{\rm str}(#1)}
\newcommand{\bbval}[1]{\left[\!\left[ #1 \right] \! \right]}
\newcommand{{\rm dom} \;}{{\rm dom} \;}
\newcommand{\mathcal{A}}{\mathcal{A}}
\newcommand{\asmbt}[2]{\mathcal{A}^{#1}_{#2}}
\newcommand{\mathcal{A}^\tau_T}{\mathcal{A}^\tau_T}
\newcommand{\ste}[2]{#1 \mapsto #2}
\newcommand{\frontier}[3]{{\partial}^{#1}_{#2}{#3}}
\newcommand{\frontiert}[1]{\partial^{\tau}{#1}}
\newcommand{\frontiertt}[1]{\frontier{\tau}{t}{#1}}
\newcommand{\frontiertx}[2]{{\partial}^{\tau}_{#1}{#2}}
\newcommand{\frontiertau}[1]{{\partial}^{\tau}{#1}}
\newcommand{\arrowstett}[2]{#1 \xrightarrow[\tau,T]{1} #2}
\newcommand{\arrowste}[2]{#1 \stackrel{1}{\To} #2}
\newcommand{\arrowtett}[2]{#1 \xrightarrow[\tau,T]{} #2}
\newcommand{\res}[1]{\textrm{res}(#1)}
\newcommand{\termasm}[1]{\mathcal{A}_{\Box}[\mathcal{#1}]}
\newcommand{\prodasm}[1]{\mathcal{A}[\mathcal{#1}]}
\newcommand{\fgg}[1]{fgg^\#_{#1}}
\newcommand{\ftdepth}[2]{\textrm{ft-depth}_{#1}\left(#2\right)}
\newcommand{\str}[1][*]{\textrm{str}_{#1}}
\newcommand{\col}[1]{\textrm{col}_{#1}}
\newcommand{\llcorner \negthinspace\lrcorner}{\llcorner \negthinspace\lrcorner}
\newcommand{G^\mathrm{f}}\newcommand{\bindinggraph}{G^\mathrm{b}}{G^\mathrm{f}}\newcommand{\bindinggraph}{G^\mathrm{b}}
\newcommand{\setr}[2]{\left\{\ #1 \ \left|\ #2 \right. \ \right\}}
\newcommand{\setl}[2]{\left\{\ \left. #1 \ \right|\ #2 \ \right\}}
In this section, we briefly sketch a 3D version of Winfree's abstract Tile Assembly Model. Going forward, all logarithms in this paper are base-$2$.
Fix an alphabet $\Sigma$.
$\Sigma^*$ is the set of finite strings over $\Sigma$. Let $\mathbb{Z}$, $\mathbb{Z}^+$, and $\mathbb{N}$ denote the set of integers, positive integers, and non-negative integers, respectively. Let $d \in \{2, 3\}$.
A \emph{grid graph} is an undirected graph $G=(V,E)$, where $V \subset \mathbb{Z}^d$, such that, for all $\left\{\vec{a},\vec{b}\right\} \in E$, $\vec{a} - \vec{b}$ is a $d$-dimensional unit vector. The \emph{full grid graph} of $V$ is the undirected graph $G^\mathrm{f}}\newcommand{\bindinggraph}{G^\mathrm{b}_V=(V,E)$,
such that, for all $\vec{x}, \vec{y}\in V$, $\left\{\vec{x},\vec{y}\right\} \in E \iff \| \vec{x} - \vec{y}\| = 1$, i.e., if and only if $\vec{x}$ and $\vec{y}$ are adjacent in the $d$-dimensional integer Cartesian space.
A $d$-dimensional \emph{tile type} is a tuple $t \in (\Sigma^* \times \mathbb{N})^{2d}$, e.g., a unit square (cube), with four (six) sides, listed in some standardized order, and each side having a \emph{glue} $g \in \Sigma^* \times \mathbb{N}$ consisting of a finite string \emph{label} and a non-negative integer \emph{strength}. We call a $d$-dimensional tile type merely a {\em tile type} when $d$ is clear from the context.
We assume a finite set of tile types, but an infinite number of copies of each tile type, each copy referred to as a \emph{tile}. A \emph{tile set} is a set of tile types and is usually denoted as $T$.
A {\em configuration} is a (possibly empty) arrangement of tiles on
the integer lattice $\mathbb{Z}^d$, i.e., a partial function $\alpha:\mathbb{Z}^d
\dashrightarrow T$. Two adjacent tiles in a configuration \emph{bind},
\emph{interact}, or are \emph{attached}, if the glues on their
abutting sides are equal (in both label and strength) and have
positive strength. Each configuration $\alpha$ induces a
\emph{binding graph} $\bindinggraph_\alpha$, a grid graph whose
vertices are positions occupied by tiles, according to $\alpha$, with
an edge between two vertices if the tiles at those vertices
bind. For two
non-overlapping configurations $\alpha$ and $\beta$, $\alpha \cup \beta$
is defined as the unique configuration $\gamma$ satisfying, for all
$\vec{x} \in {\rm dom} \;{\alpha}$, $\gamma(\vec{x}) = \alpha(\vec{x})$, for
all $\vec{x} \in {\rm dom} \;{\beta}$, $\gamma(\vec{x}) = \beta(\vec{x})$, and
$\gamma(\vec{x})$ is undefined at any point $\vec{x} \in \mathbb{Z}^d
\backslash \left( {\rm dom} \;{\alpha} \cup {\rm dom} \;{\beta} \right)$.
An \emph{assembly} is a connected, non-empty configuration,
i.e., a partial function $\alpha:\mathbb{Z}^d \dashrightarrow T$ such that
$\bindinggraph_{{\rm dom} \; \alpha}$ is connected and ${\rm dom} \; \alpha \neq
\emptyset$. Given $\tau\in\mathbb{Z}^+$, $\alpha$ is \emph{$\tau$-stable} if every cut-set
of~$\bindinggraph_\alpha$ has weight at least $\tau$, where the weight
of an edge is the strength of the glue it represents.\footnote{A
\emph{cut-set} is a subset of edges in a graph which, when removed
from the graph, produces two or more disconnected subgraphs. The
\emph{weight} of a cut-set is the sum of the weights of all of the
edges in the cut-set.} When $\tau$ is clear from context, we say
$\alpha$ is \emph{stable}. Given two assemblies $\alpha,\beta$, we
say $\alpha$ is a \emph{subassembly} of $\beta$, and we write $\alpha
\sqsubseteq \beta$, if ${\rm dom} \;\alpha \subseteq {\rm dom} \;\beta$ and, for all
points $\vec{p} \in {\rm dom} \;\alpha$, $\alpha(\vec{p}) = \beta(\vec{p})$.
A $d$-dimensional \emph{tile assembly system} (TAS) is a triple $\mathcal{T} =
(T,\sigma,\tau)$, where $T$ is a tile set, $\sigma:\mathbb{Z}^d
\dashrightarrow T$ is the finite, $\tau$-stable, \emph{seed assembly},
and $\tau\in\mathbb{Z}^+$ is the \emph{temperature}.
Given two $\tau$-stable assemblies $\alpha,\beta$, we write $\alpha
\to_1^{\mathcal{T}} \beta$ if $\alpha \sqsubseteq \beta$ and
$|{\rm dom} \;\beta \setminus {\rm dom} \;\alpha| = 1$. In this case we say $\alpha$
\emph{$\mathcal{T}$-produces $\beta$ in one step}. If $\alpha
\to_1^{\mathcal{T}} \beta$, $ {\rm dom} \;\beta \setminus
{\rm dom} \;\alpha=\{\vec{p}\}$, and $t=\beta(\vec{p})$, we write $\beta =
\alpha + (\vec{p} \mapsto t)$. The \emph{$\mathcal{T}$-frontier} of
$\alpha$ is the set $\partial^\mathcal{T} \alpha = \bigcup_{\alpha
\to_1^\mathcal{T} \beta} ({\rm dom} \;\beta \setminus {\rm dom} \;\alpha$), i.e.,
the set of empty locations at which a tile could stably attach to
$\alpha$. The \emph{$t$-frontier} of $\alpha$, denoted
$\partial^\mathcal{T}_t \alpha$, is the subset of
$\partial^\mathcal{T} \alpha$ defined as
$\setr{\vec{p}\in\partial^\mathcal{T} \alpha}{\alpha \to_1^\mathcal{T}
\beta \text{ and } \beta(\vec{p})=t}.$
Let $\mathcal{A}^T$ denote the set of all assemblies of tiles from $T$, and let $\mathcal{A}^T_{< \infty}$ denote the set of finite assemblies of tiles from $T$.
A sequence of $k\in\mathbb{Z}^+ \cup \{\infty\}$ assemblies $\vec{\alpha} = \left (\alpha_0,\alpha_1,\ldots \right)$ over $\mathcal{A}^T$ is a \emph{$\mathcal{T}$-assembly sequence} if, for all $1 \leq i < k$, $\alpha_{i-1} \to_1^\mathcal{T} \alpha_{i}$.
The {\em result} of an assembly sequence $\vec{\alpha}$, denoted as $\textmd{res}(\vec{\alpha})$, is the unique limiting assembly (for a finite sequence, this is the final assembly in the sequence).
We write $\alpha \to^\mathcal{T} \beta$, and we say $\alpha$ \emph{$\mathcal{T}$-produces} $\beta$ (in 0 or more steps), if there is a $\mathcal{T}$-assembly sequence $\alpha_0,\alpha_1,\ldots$ of length $k = |{\rm dom} \;\beta \setminus {\rm dom} \;\alpha| + 1$ such that
(1) $\alpha = \alpha_0$,
(2) ${\rm dom} \;\beta = \bigcup_{0 \leq i < k} {\rm dom} \;{\alpha_i}$, and
(3) for all $0 \leq i < k$, $\alpha_{i} \sqsubseteq \beta$.
If $k$ is finite then it is routine to verify that $\beta = \alpha_{k-1}$.
We say $\alpha$ is \emph{$\mathcal{T}$-producible} if $\sigma \to^\mathcal{T} \alpha$, and we write $\prodasm{\mathcal{T}}$ to denote the set of $\mathcal{T}$-producible assemblies.
An assembly $\alpha$ is \emph{$\mathcal{T}$-terminal} if $\alpha$ is $\tau$-stable and $\partial^\mathcal{T} \alpha=\emptyset$.
We write $\termasm{\mathcal{T}} \subseteq \prodasm{\mathcal{T}}$ to denote the set of $\mathcal{T}$-producible, $\mathcal{T}$-terminal assemblies. If $|\termasm{\mathcal{T}}| = 1$ then $\mathcal{T}$ is said to be {\em directed}.
In general, a $d$-dimensional shape is a set $X \subseteq \mathbb{Z}^d$. We say that a TAS $\mathcal{T}$ \emph{self-assembles} $X$ if, for all $\alpha \in
\termasm{\mathcal{T}}$, ${\rm dom} \;\alpha = X$, i.e., if every terminal
assembly produced by $\mathcal{T}$ places a tile on every point in $X$
and does not place any tiles on points in $\mathbb{Z}^d \backslash\, X$. We say that a TAS $\mathcal{T}$ \emph{uniquely
self-assembles} $X$ if $\termasm{\mathcal{T}} = \{ \alpha \}$ and ${\rm dom} \;\alpha = X$.
In the spirit of \cite{RotWin00}, we define the \emph{tile complexity} of a shape $X$ at temperature $\tau$, denoted by $K^\tau_{SA}(X)$, as the minimum number of distinct tile types of any TAS in which it self-assembles, i.e., \\$K^\tau_{SA}(X) = \min \left\{ n \; \left| \; \mathcal{T} = \left(T,\sigma,\tau\right), \left|T\right|=n \textmd{ and } X \textmd{ self-assembles in } \mathcal{T} \right.\right\}$. The \emph{directed tile complexity} of a shape $X$ at temperature $\tau$, denoted by $K^\tau_{USA}(X)$, is the minimum number of distinct tile types of any TAS in which it uniquely self-assembles, i.e., \\$K^\tau_{USA}(X) = \min \left\{ n \; \left| \; \mathcal{T} = \left(T,\sigma,\tau\right), \left|T\right|=n \textmd{ and } X \textmd{ uniquely self-assembles in } \mathcal{T} \right.\right\}$.
\section{Lower bound}
\label{sec:impossibility}
In this section, we prove Theorem~\ref{thm:one}, which is a lower bound on the tile complexity of 2D rectangles. So, going forward, let $k,N \in \mathbb{N}$. We say that $R^2_{k,N}$ is a 2D $k \times N$ \emph{rectangle} if $R^2_{k,N} = \{0,1, \ldots, N-1\} \times \{0,1,\ldots, k-1\}$. Throughout this section, we will denote $R^2_{k,N}$ as simply $R_{k,N}$. Our lower bound relies on the following observation regarding temperature-1 self-assembly.
\begin{observation}
\label{obs:simple}
If $\mathcal{T} = (T,\sigma,1)$ is a singly-seeded TAS in which some shape $X$ self-assembles and $\alpha \in
\mathcal{A}[\mathcal{T}]$ such that $\mathrm{dom}\ \alpha = X$, then $G^{\textmd{b}}_{\alpha}$ contains a simple path $s$
from the location of $\sigma$ to any location of $X$ and there is a corresponding (simple) assembly sequence $\vec{\alpha}$ that follows $s$ by placing tiles on and only on locations in $s$.
\end{observation}
Since, in Observation~\ref{obs:simple}, we do not necessarily assume that $\mathcal{T}$ uniquely produces $X$, there could be more than one assembly sequence for a given $s$. Throughout the rest of this section, unless stated otherwise, let $\mathcal{T} = (T,\sigma,1)$ be a singly-seeded TAS in which $R_{k,N}$ self-assembles.
\subsection{Window Movie Lemmas}
To prove our lower bound, we will use a variation of the Window Movie Lemma (WML) by Meunier, Patitz, Summers, Theyssier, Winslow and Woods and a corollary thereof. In this subsection, we review standard notation \cite{WindowMovieLemma} for and give the statements of the variation that we use in our lower bound proof.
A \emph{window} $w$ is a set of edges forming a cut-set of the full grid
graph of $\mathbb{Z}^d$. Given a window $w$ and an assembly $\alpha$, a window that {\em
intersects} $\alpha$ is a partitioning of $\alpha$ into two
configurations (i.e., after being split into two parts, each part may
or may not be disconnected). In this case we say that the window $w$
cuts the assembly $\alpha$ into two configurations $\alpha_L$
and~$\alpha_R$, where $\alpha = \alpha_L \cup \alpha_R$. Given a window $w$, its translation by a vector $\vec{\Delta}$, written $ w
+ \vec{\Delta}$ is simply the translation of each one of $w$'s elements (edges)
by~$\vec{\Delta}$.
For a window $w$ and an assembly sequence $\vec{\alpha}$, we define a \emph{glue window movie}~$M$ to be the order of placement, position and glue type for each glue that appears along the window $w$ in $\vec{\alpha}$. Given an assembly sequence $\vec{\alpha}$ and a window $w$, the associated glue window movie is the maximal sequence $M_{\vec{\alpha},w} = \left(\vec{v}_{1}, g_{1}\right), \left(\vec{v}_{2}, g_{2}\right), \ldots$ of pairs of grid graph vertices $\vec{v}_i$ and glues $g_i$, given by the order of the appearance of the glues along window $w$ in the assembly sequence $\vec{\alpha}$.
Furthermore, if $m$ glues appear along $w$ at the same instant (this happens upon placement of a tile which has multiple sides touching $w$) then these $m$ glues appear contiguously and are listed in lexicographical order of the unit vectors describing their orientation in $M_{\vec{\alpha},w}$. We use the notation $\mathcal{B}\left(M_{\vec{\alpha}, w}\right)$ to denote the \emph{bond-forming submovie} of $M_{\vec{\alpha},w}$, which consists of only those steps of $M_{\vec{\alpha},w}$ that place glues that eventually form positive-strength bonds in the assembly $\textmd{res}\left(\vec{\alpha}\right)$. We write $M_{\vec{\alpha}, w} + \vec{\Delta}$ to denote the translation of the glue window movie, that is $\left(\vec{v}_{1}+\vec{\Delta}, g_{1}\right), \left(\vec{v}_{2}+\vec{\Delta}, g_{2}\right), \ldots$.
Let $w$ be a window that partitions~$\alpha$ into two configurations~$\alpha_L$ and $\alpha_R$, and assume $w + \vec{\Delta}$ partitions~$\beta$ into two configurations $\beta_L$ and $\beta_R$.
Assume that $\alpha_L$, $\beta_L$ are the sub-configurations of $\alpha$ and $\beta$ containing the seed tile of $\alpha$ and $\beta$, respectively.
\begin{lemma}[Standard Window Movie Lemma]\label{lem:wml}
If $\mathcal{B}\left(M_{\vec{\alpha},w}\right) = \mathcal{B}\left(M_{\vec{\beta},w+\vec{\Delta}}\right) - \vec{\Delta}$, then the following two assemblies are producible: $\alpha_L \left(\beta_R - \vec{\Delta}\right) = \alpha_L \cup \left(\beta_R - \vec{\Delta}\right)$ and $\beta_L\left(\alpha_R + \vec{\Delta}\right) = \beta_L \cup \left(\alpha_R + \vec{\Delta}\right)$.
\end{lemma}
By Observation~\ref{obs:simple}, there exist simple assembly sequences $\vec{\alpha} = \left(\alpha_i \mid 0 \leq i < l\right)$ and $\vec{\beta} = \left(\beta_i \mid 0 \leq i < m\right)$, with $l,m\in\mathbb{Z}^+ $ that place tiles along simple paths $s$ and $s'$, respectively, leading to results $\alpha$ and $\beta$, respectively. The notation $M_{\vec{\alpha}, w} \upharpoonright s$ represents the \emph{restricted} glue window submovie (\emph{restricted to} $s$), which consists of only those steps of $M$ that place glues that eventually form positive-strength bonds along $s$.
\begin{corollary}[Restricted Window Movie Lemma]\label{lem:rwml}
If\\$M_{\vec{\alpha},w} \upharpoonright s = \left( M_{\vec{\beta},w+\vec{\Delta}} \upharpoonright s' \right) - \vec{\Delta}$, then the following two assemblies are producible: $\alpha_L \left(\beta_R - \vec{\Delta}\right) = \alpha_L \cup \left(\beta_R - \vec{\Delta}\right)$ and $\beta_L\left(\alpha_R + \vec{\Delta}\right) = \beta_L \cup \left(\alpha_R + \vec{\Delta}\right)$.
\end{corollary}
The proof of Corollary~\ref{lem:rwml} is identical to that of Lemma~\ref{lem:wml} and therefore is omitted. The proof is identical because $\vec{\alpha}$ $\left(\vec{\beta}\right)$ follows $s$ ($s'$) and $\tau=1$, so glues that do not follow $s$ ($s'$) are not necessary for the self-assembly of $\alpha$ ($\beta$). Note that Lemma~\ref{lem:wml} and Corollary~\ref{lem:rwml} both generalize to 3D. See Figure~\ref{fig:glue_movie_window_examples} for examples of $M_{\vec{\alpha},w}$, $\mathcal{B}\left( M_{\vec{\alpha},w} \right)$ and $M_{\vec{\alpha}, w} \upharpoonright s$. We now turn our attention to counting the number of restricted glue window submovies.
\begin{figure}
\centering
\begin{subfigure}[t]{0.15\textwidth}
\centering
\includegraphics[width=.75in]{glue_movie_window_example_assembly}
\caption{\label{fig:glue_movie_window_example_assembly} A subassembly of $\alpha$ and a window $w$ induced by a translation of the $y$-axis.}
\end{subfigure}%
~ \hspace{15pt}
\begin{subfigure}[t]{0.15\textwidth}
\centering
\includegraphics[width=.75in]{glue_movie_window_example_just_path}
\caption{\label{fig:glue_movie_window_example_just_path} A portion of the simple path $s$ through $G^b_{\alpha}$.}
\end{subfigure}%
~ \hspace{15pt}
\begin{subfigure}[t]{0.15\textwidth}
\centering
\includegraphics[width=.75in]{glue_movie_window_example_full}
\caption{\label{fig:glue_movie_window_example_full} The glue window movie $M_{\vec{\alpha},w}$ }
\end{subfigure}%
~ \hspace{15pt}
\begin{subfigure}[t]{0.15\textwidth}
\centering
\includegraphics[width=.75in]{glue_movie_window_example_bond_forming_submovie}
\caption{\label{fig:glue_movie_window_example_bond_forming_submovie} The bond-forming submovie $\mathcal{B}\left( M_{\vec{\alpha},w} \right)$ }
\end{subfigure}%
~ \hspace{15pt}
\begin{subfigure}[t]{0.15\textwidth}
\centering
\includegraphics[width=.75in]{glue_movie_window_example_restricted}
\caption{\label{fig:glue_movie_window_example_restricted} The restricted glue window submovie $M_{\vec{\alpha},w} \upharpoonright s$ }
\end{subfigure}
\caption{\label{fig:glue_movie_window_examples} An assembly, a simple path and the various types of glue window movies. }
\end{figure}
\subsection{Counting procedure for undirected self-assembly in 2D}
\label{sec:counting_procedure}
In this subsection, we develop a counting procedure that we will use to obtain an upper bound on the number of distinct restricted glue window submovies.
Let $\rho \in \mathcal{A}_\Box\left[ \mathcal{T} \right]$. By Observation~\ref{obs:simple}, there exists a simple path $s$ in $G^{\textmd{b}}_{\rho}$
from the location of $\sigma$ to any location in an extreme (i.e., leftmost or rightmost) column
of $R_{k,N}$. For the remainder of this section, unless stated otherwise, let $\vec{\alpha}$ denote any (simple) assembly sequence that follows a simple path from the location of $\sigma$ to some location in the furthest extreme column of $R_{k,N}$. So, if $\sigma$ is in the left half of $R_{k,N}$, then $s$ will go from $\sigma$ to some location in the rightmost column of $R_{k,N}$. If $\sigma$ is in the right half of $R_{k,N}$, then $s$ will go from $\sigma$ to some location in the leftmost column of $R_{k,N}$. If $\sigma$ is in the middle column of $R_{k,N}$, then $s$ can go to either the leftmost or rightmost column of $R_{k,N}$.
Index the columns of $R_{k,N}$ from 1 (left) to $N$ (right). Assume $c_\sigma$ is the index of the column in which the seed is contained.
Consider two consecutive columns, with indices $c_0$ and $c_1$, satisfying $\left| c_\sigma - c_0 \right| < \left| c_\sigma - c_1 \right|$, and such that either $c_0$ or $c_1$ (or both) are in between $c_{\sigma}$ and the furthest (from $\sigma$) extreme column.
Since $s$ is a simple path from the location of $\sigma$ to some location in the extreme column of $R_{k,N}$ that is furthest from $\sigma$, $s$ crosses between $c_0$ and $c_1$ through $e$ \emph{crossing edges} in $G^{\textmd{b}}_{\rho}$, where $1 \leq e \leq k$ and $e$ is odd, visiting a total of $2e$ endpoints.
The endpoint of a crossing edge in column $c_0$ ($c_1$) is its \emph{near} (far) endpoint. A crossing edge points \emph{away from} (towards) the seed if its near endpoint is visited first (second).
Observe that the first and last crossing edges visited by $s$ must point away from the seed but each crossing edge that points away from the seed (except the last crossing edge) is immediately followed by a corresponding crossing edge that points towards the seed, when skipping the part of $s$ that connects them without going through another crossing edge in between them. Assume that the rows of $R_{k,N}$ are assigned an index from $1$ (top) to $k$ (bottom). Let $E \subseteq \{1,\ldots, k\}$ be such that $|E| = e$ and $f \in E$ and $l \in E$ be the row indices of the first and last crossing edges visited, respectively. We define a \emph{near} (far) \emph{crossing pairing over} $E$ \emph{starting at} $f \in E$ (ending at $l \in E$) as a set of $p = \frac{e-1}{2}$ non-overlapping pairs of elements in $E\backslash \{f\}$ ($E\backslash\{l\}$), where each pair contains (the row indices of) one crossing edge pointing away from the seed and its corresponding crossing edge pointing towards the seed. See Figure~\ref{fig:counting_proc_good} for examples of the previous definitions. For counting purposes pertaining to a forthcoming argument, we establish an injective mapping
from near crossing pairings over $E$ to strings of balanced parentheses of length $e - 1$.
\begin{lemma}
\label{lem:near-crossing-pairing-to-parens}
There exists an injective function from the set of all near crossing pairings over $E$ starting at $f$ into the set of all strings of $2p$ balanced parentheses.
\end{lemma}
\begin{proof}
Given a near crossing pairing $P$ with $p$ pairs, build a string $x$ with $2p$
characters indexed from 1 to $2p$ going from left to right, as
follows. For each element $\{a,b\}$ of $P$, with $a<b$ and assuming $a$ is the $i$-th lowest row index and $b$ is the $j$-th lowest row index in $E \backslash \{f\}$, place a left parenthesis at index $i$ and a right parenthesis at index $j$.
The resulting string $x$ contains
exactly $p$ pairs of parentheses. Furthermore, all of the parentheses
in $x$ are balanced because each opening parenthesis appears to the
left of its closing parenthesis (thanks to the indexing used in the
string construction algorithm just described) and, for any two pairs
of parentheses in the string, it must be the case that either 1) they do not
overlap (i.e., the closing parenthesis of the leftmost pair is
positioned to the left of the opening parenthesis of the rightmost
pair) or 2) one pair is nested inside the other (i.e., the interval
defined by the indices of the nested pair is included in the interval
defined by the indices of the outer pair). The other case is
impossible, that is, when the two pairs $\{a,b\}$ and $\{c,d\}$ are
such that $a < c < b < d$ because it would be impossible for any path
crossing consecutive columns according to $P$ to be simple. Consider any simple
path $\pi_{\{a,b\}}$ that links crossing edges $a$ and $b$ without going through another crossing edge in between them. Since $P$ is a near crossing pairing, $\pi_{\{a,b\}}$ is fully
contained in the half-plane $H_0$ on the near side of $c_0$ toward the seed. This
path partitions $H_0$ into two spaces. Since the crossing edges $c$ and $d$ belong to different components of this
partition, any simple path linking these two crossing edges must
cross $\pi_{\{a,b\}}$. Therefore, $s$, crossing between $c_0$ and $c_1$, could not have been simple.
Finally, note that two different near crossing pairings $P_1$ and $P_2$ over $E$ starting at $f$ will map to two different strings of balanced parentheses, so the mapping is injective.
\qed
\end{proof}
\begin{corollary}
\label{lem:far-crossing-pairing-to-parens}
There exists an injective function from the set of all far crossing pairings over $E$ ending at $l$ into the set of all strings of $2p$ balanced parentheses.
\end{corollary}
\begin{figure}
\centering
\begin{subfigure}[t]{0.48\textwidth}
\centering
\includegraphics[width=2.2in]{counting_procedure_0}
\caption{\label{fig:counting_proc_0}Two consecutive columns, for ``height'' $k=12$, rotated 90 degrees counter-clockwise. We show the columns this way for ease of examining the corresponding strings of balanced parentheses.}
\end{subfigure}%
~
\begin{subfigure}[t]{0.48\textwidth}
\centering
\includegraphics[width=2.2in]{counting_procedure_1_2_3}
\caption{\label{fig:counting_proc_steps_1_2_3}Steps 1, 2 and 3. Here, $f=12$, $l=2$ and $E = \{2, 4, 5, 6, 8, 10, 12\}$, with chosen glues indicated by the little black rectangles.}
\end{subfigure}%
\\
\vspace{10pt}
\begin{subfigure}[t]{0.48\textwidth}
\centering
\includegraphics[width=2.2in]{counting_procedure_4_5}
\caption{\label{fig:counting_proc_steps_4_5}Steps 4 and 5. Here, $x_0 = ()(())$ and $x_1 = (())()$.}
\end{subfigure}%
~
\begin{subfigure}[t]{0.48\textwidth}
\centering
\includegraphics[width=2.2in]{counting_procedure_6}
\caption{\label{fig:counting_proc_step_6}The $\pi$ with $|\pi|=14=2e$ that is found by the algorithm in Step 6. }
\end{subfigure}
\caption{\label{fig:counting_proc_good}A ``good'' sample run of the counting procedure.}
\vspace{-20pt}
\end{figure}
The following is a systematic
procedure for upper bounding the number of ways to select and order the $2e$ endpoints of the $e$ crossing edges between an arbitrary pair of consecutive columns $c_0$ and $c_1$ (see Figure~\ref{fig:counting_proc_0}). Obviously, the number of ways to do this is less than or equal to $\binom{k}{e} (2e)!$. We will use crossing pairings and Catalan numbers to reduce this upper bound.
\begin{enumerate}
\item Choose the set $E$ of row indices of the $e$ crossing edges, out of $k$ possible edges between consecutive columns. There are $\binom{k}{e}$ ways to do this.
\item One of the crossing edges must be first, so, choose the first crossing edge $f$. There are $e$ ways to do this.
\item One of the crossing edges must be last, so, choose the last crossing edge $l$. There are $e$ ways to do this.
\end{enumerate}
The three previous steps are depicted in Figure~\ref{fig:counting_proc_steps_1_2_3}. We purposely allow choosing $l=f$ because our intention is to upper bound the number of ways to select and order the endpoints of the crossing edges. Moreover, $l=f$ when $e=1$. Denote as $\mathcal{C}_p = \frac{1}{p+1} \binom{2p}{p}$ the $p^{th}$ Catalan number (indexed starting at 0).
\begin{enumerate}
\setcounter{enumi}{3}
\item For a given pair of consecutive columns, in which $f$ is visited first, $s$ induces a near crossing pairing over $E$ starting at $f$, where the elements of each pair are row indices of near endpoints of crossing edges, including $l$, but not $f$. By Lemma~\ref{lem:near-crossing-pairing-to-parens}, it suffices to count the number of ways to choose a string of $2p$ balanced parentheses. Therefore, choose a string $x_0$ of $2p$ balanced parentheses, where $x_0[i]$ corresponds to the crossing edge in $E$ with the $i$-th lowest row index, excluding $f$. There are $\mathcal{C}_p$ ways to do this.
\item For a given pair of consecutive columns in which $l$ is visited last, $s$ induces a far crossing pairing over $E$ ending at $l$, where the elements of each pair are row indices of far endpoints of crossing edges, including $f$, but not $l$. By Corollary~\ref{lem:far-crossing-pairing-to-parens}, it suffices to count the number of ways to choose a string of $2p$ balanced parentheses. Therefore, choose a string $x_1$ of $2p$ balanced parentheses, where $x_1[i]$ corresponds to the crossing edge in $E$ with the $i$-th lowest row index, excluding $l$. There are $\mathcal{C}_p$ ways to do this.
\end{enumerate}
The two previous steps are depicted in Figure~\ref{fig:counting_proc_steps_4_5}. At this point, we have chosen the locations and connectivity pattern of all the crossing edges. We now show that there is at most one way in which both endpoints of every crossing edge may be visited by $s$, subject to the constraints imposed by the previous steps.
\begin{enumerate}
\setcounter{enumi}{5}
\item Let $I_j(r)$ be the index $i$, such that $x_j[i]$ corresponds to the crossing edge with row index $r$. The following greedy algorithm attempts to build a path $\pi$ of locations that (1) starts at the near endpoint of the first crossing edge, (2) ends at the far endpoint of the last crossing edge and (3) visits only the endpoints of the crossing edges while following the balanced parenthesis pairings of both $x_0$ and $x_1$.
\begin{algorithm}[H]
\SetAlgoLined
Initialize $\pi = \left( \left(c_0, f\right), \left(c_1, f\right) \right)$ to be a sequence of locations, $j = 1$ and $r = f$, where $r$ stands for ``row number'' and $j$ stands for the current side, near (0) or far (1).
\While{$r\ne l$}{
Let $r'$ be the unique row index of the crossing edge, such that, $I_j(r)$ and $I_j\left(r'\right)$ are paired.
Set $r = r'$.
Append $\left(c_j, r\right)$ to $\pi$.
Set $j = (j+1) \mod 2$.
Append $\left(c_j, r\right)$ to $\pi$.
}
\end{algorithm}
First, note that no endpoint can be visited more than once, so the algorithm always terminates. Second, note that, when the algorithm terminates, either $|\pi| = 2e$ (see Figure~\ref{fig:counting_proc_step_6}) or $|\pi|<2e$ (see Figure~\ref{fig:counting_proc_bad}). However, regardless of its length, $\pi$ is the unique sequence of endpoints of crossing edges that can be visited by any simple path starting at $\left(c_0,f\right)$, ending at $\left(c_1,l\right)$ and following the balanced parenthesis pairings of both $x_0$ and $x_1$. The uniqueness of $\pi$ follows from the uniqueness of $r'$, given $r$, based on the balanced parenthesis pairings of both $x_0$ and $x_1$.
\begin{figure}[h]
\centering \includegraphics[width=2.2in]{counting_procedure_bad}
\caption{\label{fig:counting_proc_bad}A ``bad'' sample run of the counting procedure. Here, $x_0=(())()$ and $x_1=()()()$. With $f=12$, $l=2$ and $E = \{2, 4, 5, 6, 8, 10, 12\}$, there is no valid $\pi$ (crossing edges 4 and 5 are skipped, assuming the pairs of parentheses are faithfully followed), so the algorithm terminates with $|\pi|=10<2e=14$.}
\vspace{-20pt}
\end{figure}
\end{enumerate}
By the above counting procedure, there are at most $\binom{k}{e}\left( e\frac{1}{p+1}\binom{2p}{p}\right)^2\cdot 1$ ways to select and order the endpoints of the crossing edges between an arbitrary pair of consecutive columns $c_0$ and $c_1$ as they are visited by a simple path.
\subsection{Lower bound for undirected self-assembly in 2D: Theorem~\ref{thm:one}}
To prove a lower bound on $K^1_{SA}\left(R_{k,N}\right)$, we turn our attention to upper bounding the number of restricted glue window submovies of the form $M_{\vec{\alpha}, w} \upharpoonright s$. For the remainder of this subsection, assume $w$ is always some window induced by (a translation of) the $y$-axis that cuts $R_{k,N}$ between some pair of consecutive columns.
\begin{lemma}
\label{lem:count_restricted_bond_forming_submovies}
The number of restricted glue window submovies of the form $M_{\vec{\alpha},w}~\upharpoonright~s$ is less than or equal to $|G|^k \cdot 2^{3k+2} \cdot k$, where $G$ is the set of all glues of (the tile types in) $T$.
\end{lemma}
\begin{proof}
Let $e$ be an odd number such that $1 \leq e \leq k$ and\\$M_{\vec{\alpha},w} \upharpoonright s = \left(\vec{v}_1,g_1\right), \ldots, \left(\vec{v}_{2e}, g_{2e}\right)$ be a restricted glue window submovie. Since $\vec{\alpha}$ follows a simple path, $g_{2i-1}=g_{2i}$ for $i=1,\ldots, e$. This means that we only need to assign $e$ glues, with $|G|$ choices for each glue. So, the number of ways to assign glues in $M_{\vec{\alpha},w} \upharpoonright s$ is less than or equal to $|G|^e$. Since $\vec{\alpha}$ follows a simple path, each location in $M_{\vec{\alpha},w} \upharpoonright s$ corresponds to an endpoint of a crossing edge that crosses $w$. So, the number of ways to assign locations in $M_{\vec{\alpha},w} \upharpoonright s$ is less than or equal to the number of ways to select and order the endpoints of $e$ crossing edges that cross $w$ via $\vec{\alpha}$. By the above counting procedure, the number of ways to select and order the endpoints of $e$ crossing edges that cross $w$ via $\vec{\alpha}$ is less than or equal to $\binom{k}{e}\left( e\frac{1}{p+1}\binom{2p}{p}\right)^2$. Thus, if $m$ is the total number of restricted glue window submovies of the form $M_{\vec{\alpha},w} \upharpoonright s$, then we have:
\begin{eqnarray*}
m & \leq & \sum_{\substack{1\leq e \leq k \\ e\ \mathrm{odd}}}\left( \binom{k}{e}\left( e\frac{1}{p+1}\binom{2p}{p}\right)^2 \ |G|^e \right) \\
& = & \sum_{\substack{1\leq e \leq k \\ e\ \mathrm{odd}}}\left( \binom{k}{e}\left( e\frac{2}{e+1}\binom{e-1}{(e-1)/2}\right)^2 \ |G|^e \right) \\
&\leq & 2^2 \ \sum_{\substack{1\leq e \leq k \\ e\ \mathrm{odd}}}\left( \binom{k}{e}\left( \frac{e}{e+1}\binom{e-1}{(e-1)/2}\right)^2 \ |G|^k \right) \\
& = & |G|^k \cdot 2^2 \ \sum_{\substack{1\leq e \leq k \\ e\ \mathrm{odd}}}\left( \binom{k}{e}\left( \frac{e}{e+1}\binom{e-1}{(e-1)/2}\right)^2 \right) \\
&\leq & |G|^k \cdot 2^2 \ \sum_{\substack{1\leq e \leq k\\ e\ \mathrm{odd}}}\left( 2^k\left( 1 \cdot \binom{e-1}{(e-1)/2}\right)^2 \right) \\
& = & |G|^k \cdot 2^{k+2} \ \sum_{\substack{1\leq e \leq k\\ e\ \mathrm{odd}}} \binom{e-1}{(e-1)/2}^2 \\
& = & |G|^k \cdot 2^{k+2} \ \sum_{\substack{0\leq e \leq k-1\\ e\ \mathrm{even}}} \binom{e}{e/2}^2 \\
& \leq & |G|^k \cdot 2^{k+2} \ \sum_{\substack{0\leq e \leq k-1\\ e\ \mathrm{even}}} \binom{k}{e/2}^2 \\
& \leq & |G|^k \cdot 2^{k+2}\ \ \sum_{\substack{0\leq e \leq k-1\\ e\ \mathrm{even}}}2^{2k} \\
& \leq & |G|^k \cdot 2^{3k+2} \cdot k. \\
\end{eqnarray*}
\qed
\end{proof}
We will use Lemma~\ref{lem:count_restricted_bond_forming_submovies} to prove our impossibility result.
\addtocounter{theorem}{-2}
\begin{theorem}
$K^1_{SA}\left(R_{k,N}\right)=\Omega\left( N^{\frac{1}{k}}\right)$.
\end{theorem}
\begin{proof}
Let $G$ be the set of glues of (the tile types in) $T$. It suffices to show that $|T| = \Omega\left( N^{\frac{1}{k}}\right)$. Since $s$ is the longest path in $G^{\textmd{b}}_{\rho}$, from the location of $\sigma$ to some location in an extreme column
of $R_{k,N}$, by Lemma~\ref{lem:count_restricted_bond_forming_submovies}, if $N > 2 \cdot |G|^k \cdot 2^{3k+2} \cdot k$, then there exists a window $w$, along with vectors $\vec{\Delta}_1$ and $\vec{\Delta}_2$, such that, $\vec{\Delta}_1 \ne \vec{0}$, $\vec{\Delta}_2 \ne \vec{0}$ and $\vec{\Delta}_1 \ne \vec{\Delta}_2$ and satisfying $M_{\vec{\alpha},w} \upharpoonright s = \left(M_{\vec{\alpha},w+\vec{\Delta}_1} \upharpoonright s \right) - \vec{\Delta}_1$ and $M_{\vec{\alpha},w+\vec{\Delta}_1} \upharpoonright s = \left(M_{\vec{\alpha},w+\vec{\Delta}_2} \upharpoonright s \right) - \vec{\Delta}_2$. Without loss of generality, assume that $w$ and $w+\vec{\Delta}_1$ are on the same side of $\sigma$ and $w+\vec{\Delta}_1$ is to the left of $w$. Assume that $w$ partitions $\alpha$ into $\alpha_L$ and $\alpha_R$ and $w+\vec{\Delta}_1$ partitions $\beta=\alpha$ into $\beta_L$ and $\beta_R$. Then, by Corollary~\ref{lem:rwml}, $\beta_L\left(\alpha_R+\vec{\Delta}_1\right) \in \mathcal{A}[\mathcal{T}]$. Since $\vec{\Delta}_1\ne \vec{0}$, $\textmd{dom}\left(\beta_L\left(\alpha_R + \vec{\Delta}_1\right)\right)\backslash R_{k,N} \ne \emptyset$. In other words, $\mathcal{T}$ produces some assembly that places at least one tile at a location that is not an element of $R_{k,N}$. Therefore, it must hold that $N \leq 2 \cdot |G|^k \cdot 2^{3k+2} \cdot k$, which implies that $|G|^k \geq \frac{N}{ 2^{3k+3} \cdot k}$, and thus $|G| \geq \left( \frac{N}{2^{3k+3} \cdot k} \right)^{\frac{1}{k}} \geq \frac{N^{\frac{1}{k}}}{\left( 2^{6k} \cdot 2^k \right)^{\frac{1}{k}}} = \frac{N^{\frac{1}{k}}}{128}$. Finally, note that $|T| \geq \frac{|G|}{4}$ and it follows that $|T| = \Omega\left( N^{\frac{1}{k}} \right)$.
\qed
\end{proof}
Note that the main technique for the proof of Theorem~\ref{thm:one} does not hold in 3D because Lemma~\ref{lem:near-crossing-pairing-to-parens} requires planarity.
\section{Introduction}
\label{sec:intro}
Intuitively, self-assembly is the process through which
simple, unorganized components spontaneously combine, according to
local interaction rules, to form some kind of organized final
structure.
While nature exhibits numerous examples of self-assembly, researchers have been investigating the extent to which the power of nano-scale self-assembly can be harnessed for the systematic nano-fabrication of atomically-precise computational, biomedical and mechanical devices. For example, in the early 1980s, Ned Seeman \cite{Seem82} exhibited an experimental technique for controlling nano-scale self-assembly known as ``DNA tile self-assembly''.
Erik Winfree's abstract Tile Assembly Model (aTAM) is a simple, discrete mathematical model of DNA tile self-assembly. In the aTAM, a DNA tile is represented as an un-rotatable unit square \emph{tile}. Each side of a tile may have a \emph{glue} that consists of an integer strength, usually 0, 1 or 2, and an alpha-numeric label. The idea is that, if two tiles abut with matching kinds of glues, then they bind with the strength of the glue. In the aTAM, a tile set may consist of a finite number of tiles, because individual DNA tiles are expensive to manufacture. However, an infinite number of copies of each tile are assumed to be available during the self-assembly process in the aTAM. Self-assembly starts by designating a \emph{seed} tile and placing it at the origin. Then, a tile can come in and bind to the seed-containing \emph{assembly} if it binds with total strength at least a certain experimenter-chosen, integer value called the \emph{temperature} that is usually 1 or 2. Self-assembly proceeds as tiles come in and bind one-at-a-time in an asynchronous and non-deterministic fashion.
Tile sets are designed to work at a certain temperature value. For instance, a tile set that self-assembles correctly at temperature-2 will probably not self-assemble correctly at temperature-1. However, if a tile set works correctly at temperature-1, then it can be easily modified to work correctly at temperature-2 (or higher). In what follows, we will refer to a ``temperature-2'' tile set as a tile set that self-assembles correctly only if the temperature is 2 and a ``temperature-1'' tile set as a tile set that self-assembles correctly if the temperature is 1.
Temperature-2 tile sets give the tile set designer more control over the order in which tiles bind to the seed-containing assembly.
For example, in a temperature-2 tile set, unlike in a temperature-1 tile set, the placement of a tile at a certain location can be prevented until after the placements of at least two other tiles at two respective adjacent locations.
This is known as \emph{cooperative binding}.
Cooperative binding in temperature-2 tile sets leads to the self-assembly of computationally and geometrically interesting structures, in the sense of Turing universality \cite{Winf98}, the efficient self-assembly of $N \times N$ squares \cite{AdlemanCGH01,RotWin00} and algorithmically-specified shapes \cite{SolWin07}.
While it is not known whether the results cited in the previous paragraph hold for temperature-1 tile sets, the general problem of characterizing the power of non-cooperative tile self-assembly is important from both a theoretical and practical standpoint. This is because when cooperative self-assembly is implemented in the laboratory \cite{RoPaWi04,BarSchRotWin09,SchWin07,MaoLabReiSee00,WinLiuWenSee98}, erroneous non-cooperative binding events may occur, leading to the production of invalid final structures. Of course, the obvious way to minimize such erroneous non-cooperative binding events is for experimenters to always implement systems that work in non-cooperative self-assembly because temperature-1 tile sets will work at temperature-1 or temperature-2. Yet, how capable is non-cooperative self-assembly, in general, or even in certain cases? At the time of this writing, no such general characterization of the power of non-cooperative self-assembly exists, but there are numerous results that show the apparent weakness of specific classes of temperature-1 tile self-assembly \cite{jLSAT1,ManuchSS10,WindowMovieLemma,MeunierW17}.
Although these results highlight the weakness of certain types of temperature-1 tile self-assembly, if 3D (unit cube) tiles are allowed to be placed in just-barely-three-dimensional Cartesian space (where tiles may be placed in just the $z=0$ and $z=1$ planes), then temperature-1 self-assembly is nearly as powerful as its two-dimensional cooperative counterpart. For example, like 2D temperature-2 tile self-assembly, just-barely 3D temperature-1 tile self-assembly is capable of simulating Turing machines \cite{CookFuSch11} and the efficient self-assembly of squares \cite{CookFuSch11,jFurcyMickaSummers} and algorithmically-specified shapes \cite{FurcyS18}.
Furthermore, Aggarwal, Cheng, Goldwasser, Kao, Moisset de Espan\'{e}s and Schweller \cite{AGKS05g} studied the efficient self-assembly of $k \times N$ rectangles, where $k < \frac{\log N}{\log \log N - \log \log \log N}$ at temperature-2 in 2D (the upper bound on $k$ makes such a rectangle \emph{thin}). They proved that the size of the smallest set of tiles that uniquely self-assemble into (i.e., the \emph{tile complexity} of) a thin $k \times N$ rectangle is $O\left(N^{\frac{1}{k}}+k\right)$ and $\Omega\left(\frac{N^{\frac{1}{k}}}{k}\right)$ at temperature-2. Their lower bound actually applies to all tile sets (temperature-1, temperature-2, etc.) but their upper bound construction requires temperature-2 and does not work correctly at temperature-1.
In this paper, we continue the line of research into the tile complexity of thin rectangles, initiated by Aggarwal, Cheng, Goldwasser, Kao, Moisset de Espan\'{e}s and Schweller, but exclusively for temperature-1 tile self-assembly.
\subsection{Main results of this paper}
The main results of this paper are bounds on the tile complexity of thin rectangles at temperature-1. We give an improved lower bound for the tile complexity of 2D thin rectangles as well as a non-trivial upper bound for the tile complexity of just-barely 3D thin rectangles. Intuitively, a just-barely 3D thin rectangle is like having at most two 2D thin rectangles stacked up one on top of the other. We prove two main results: one negative (lower bound) and one positive (upper bound). Our main negative result gives a new and improved asymptotic lower bound on the tile complexity of a 2D thin rectangle at temperature-1, without assuming unique production of the terminal assembly (unique self-assembly).
\begin{theorem}
\label{thm:one}
The tile complexity of a $k \times N$ rectangle for temperature-1 tile sets is $\Omega\left( N^{\frac{1}{k}}\right)$.
\end{theorem}
Currently, the best upper bound for the tile complexity of a $k \times N$ rectangle for temperature-1 tile sets is $N + k - 1$, and is obtained via a straightforward generalization of the ``Comb construction'' by Rothemund and Winfree (see Figure 2b of \cite{RotWin00}). So, while Theorem~\ref{thm:one} currently does not give a tight bound, its proof technique showcases a novel application of Catalan numbers to proving lower bounds for temperature-1 self-assembly in 2D, and could be of independent interest.
Our main positive result is the first non-trivial upper bound on the tile complexity of just-barely 3D thin rectangles.
\begin{theorem}
\label{thm:two}
The tile complexity of a just-barely 3D $k \times N$ thin rectangle for temperature-1 tile sets is $O\left(N^{\frac{1}{\left \lfloor \frac{k}{3} \right \rfloor}} + \log N \right)$. Moreover, our construction produces a tile set that self-assembles into a unique final assembly.
\end{theorem}
We say that our main positive result is the first non-trivial upper bound because a straightforward generalization of the aforementioned Comb construction would give an upper bound of $O\left(N + k\right)$ on the tile complexity of a just-barely 3D $k \times N$ thin rectangle.
\subsection{Comparison with related work}
Aggarwal, Cheng, Goldwasser, Kao, Moisset de Espan\'{e}s and Schweller \cite{AGKS05g} give a general lower bound of $\Omega\left( \frac{N^{\frac{1}{k}}}{k}\right)$ for the tile complexity of a 2D $k \times N$ rectangle for temperature-$\tau$ tile sets. Our main negative result, Theorem~\ref{thm:one}, is an asymptotic improvement of this result for the special case of temperature-1 self-assembly.
Aggarwal, Cheng, Goldwasser, Kao, Moisset de Espan\'{e}s and Schweller \cite{AGKS05g} also prove that the tile complexity of a 2D $k \times N$ thin rectangle for general positive temperature tile sets is $O\left( N^{\frac{1}{k}}+k\right)$. Our main positive result, Theorem~\ref{thm:two}, is inspired by but requires a substantially different proof technique from theirs. Our construction, like theirs, uses a just-barely 3D counter, the base of which depends on the dimensions of the target rectangle, but unlike theirs, ours self-assembles in a zig-zag manner and the digits of the counter are encoded geometrically, vertically-oriented and in binary.
\section*{Acknowledgement}
We thank Ryan Breuer (University of Wisconsin Oshkosh Computer Science Major, class of 2019) for writing a computer program that implements the construction for our main positive result. Through his implementation, Ryan uncovered a couple technical flaws with an earlier version of the construction, which have since been corrected in version that was presented here.
\bibliographystyle{amsplain}
|
\section{Introduction}
\label{sec:intro}
Since the discovery of graphene\cite{novoselov2004electric}, two-dimensional (2D) materials have been shown to possess remarkable electronic, mechanical, thermal, and optical properties with great potential for nanotechnology applications, such as semiconductors, ultrasensitive sensors, and medical devices\cite{neto2009electronic,hendry2010coherent,sevik2014assessment,lee2008measurement}.
Stacked 2D materials (or ``heterostructures'') have even more unusual and novel properties that their monolayer and 3D counterparts do not possess.\cite{geim2013van,novoselov20162d}
For example, the electronic band gap of a graphene bilayer can be tuned
by applying a variable external electric field, which allows great flexibility in the design and optimization of semiconductor devices such as p-n junctions and transistors.\cite{zhang2009direct}
A different manifestation of interesting behavior not found in the bulk is the
recently reported superconductivity in intentionally misaligned
(by a relative twist of $\sim 1.1^\circ$) graphene bilayers\cite{graphene_sc_2018}.
As a prototype of a stacked 2D material, multilayer graphene (``graphitic structure'' hereafter) exhibits strong $sp^2$ covalent bonds within layers and weak van der Waals (vdW) and orbital repulsion interactions between layers.
Although weak, it is the interlayer interaction that defines the function of nanodevices such as nanobearings, nanomotors and nanoresonators.\cite{kolmogorov2005registry}
To simulate the mechanical behavior of graphitic structures
it is necessary to model the interactions between the electrons and the ions, which
produce the forces governing atomic motion and deformation.
First-principles approaches that involve solving the Schr\"{o}dinger equation are most accurate, but due to hardware and algorithmic limitations,
this approach is typically limited to studying small molecular systems and crystalline materials characterized by compact unit cells with
an upper limit on the number of atoms in the range of $\sim 10^3$.
Empirical interatomic potentials are computationally far less costly than
first-principles methods and can therefore be used to compute static
and dynamic properties that are inaccessible to quantum calculations,
such as dynamical tribological properties of large-scale graphene
interfaces.\cite{wen2015interpolation, leven2014ilp, leven2016interlayer}
There have been many efforts to produce an interatomic potential that would
adequately describe the properties of graphitic structures, in particular the interactions between layers.
However, as we argue in detail in this paper, the existing potentials fall short
of capturing key elements of the graphitic structures of interest.
Therefore, there is a pressing need to construct an accurate interlayer potential that will
elucidate many of the important structural properties of these structures.
The paper is structured as follows.
In \sref{sec:need} we briefly review the nature of existing
interatomic potentials that might be applied to graphitic structures,
we explain their shortcomings, and
elaborate on the need for constructing a new potential.
In \sref{sec:model}, the functional form of the new model is
presented, together with
a description of the fitting process that determines the values of
all the parameters that appear in it.
In \sref{sec:test}, the predictions of the new model for several canonical
properties of interest are compared with other potentials and
results from \emph{ab initio} total-energy calculations based
on density functional theory (DFT).
Large-scale applications of the new model are discussed in \sref{sec:app}.
The paper is summarized in \sref{sec:sum}.
\section{Need for new graphitic potential
\label{sec:need}
}
\begin{figure*}
\begin{subfigure}{0.65\columnwidth}
\includegraphics[width=\columnwidth]{high_sym_configs.pdf}
\caption{}
\label{fig:high:sym:confs}
\end{subfigure}
\begin{subfigure}{0.65\columnwidth}
\includegraphics[width=\columnwidth]{energy_vs_translation.pdf}
\caption{}
\label{fig:sliding:energy}
\end{subfigure}
\begin{subfigure}{0.65\columnwidth}
\includegraphics[width=\columnwidth]{force_rotation_center_sep3_4.pdf}
\caption{}
\label{fig:rotation:force}
\end{subfigure}
\caption{Energy and force variations when sliding and twisting a graphene bilayer.
(a) Schematic representation of high-symmetry graphene bilayer configurations: AA, AB, and saddle point (SP) stacking.
(b) Energy variation of sliding one layer relative to the other along the armchair direction.
(c) Out-of-plane component of the force on the atom at the rotation center
(blue circle labeled 1 in the bottom layer in panel (a)).
Rotation by $0^\circ$ corresponds to AB stacking,
and rotation by $\pm60^\circ$ corresponds to AA stacking.
In both sliding and twisting, periodic boundary conditions are applied and the layer spacing is fixed at $3.4~\text{\AA}$. Details are provided in \sref{sec:test}.}
\end{figure*}
A large number of interatomic potentials have been developed to model the strong covalent bonds in carbon systems. Among these are bond-order potentials, such as the Tersoff\cite{tersoff1988empirical,tersoff1989modeling} and REBO\cite{brenner1990physical,brenner2002rebo} potentials, which allow for bond breaking and formation depending on the local atomic environments. Such models have been shown to be accurate for many problems and are widely used,
but are not suitable for layered 2D materials since
they do not include long-range weak interactions.
To address this, the AIREBO\cite{stuart2000airebo} potential (based on REBO) added a \mcr{6-12 form of the} Lennard-Jones (LJ) potential\cite{lennardjones1931} to model vdW interactions.
For graphitic structures, the LJ potential works well in describing the overall binding characteristics between graphene layers.
For example, the LJ parameterization used in AIREBO predicts an equilibrium layer spacing of $3.357~\text{\AA}$ and a $c$-axis elastic modulus of $37.78~\text{GPa}$ for graphite,
in good agreement with first-principles and experimental results.
The isotropic nature of LJ, that is,
the fact that it depends only on distance between atoms and not orientation,
makes it too smooth to distinguish energy variations for different relative alignments of layers.\cite{zhang2017energy}
\fref{fig:sliding:energy} shows the energy variation obtained by sliding one layer relative to the other along the armchair direction of a graphene bilayer.
The energy remains nearly constant with a maximal difference of $0.41~\text{meV/atom}$ between the AA and AB stackings,
a small fraction (6.6\%) of the DFT result (also shown in the figure).
The reason that the LJ potential fails to capture the energy variations
due to interlayer sliding
is that in addition to vdW, the interlayer interactions include
short-range Pauli repulsion between overlapping $\pi$ orbitals of adjacent layers.
These repulsive interactions are not well described by a simple pair potential like LJ.\cite{kolmogorov2005registry,leven2014ilp,leven2016interlayer} To account for this registry effect (relative alignment of layers), Kolmogorov and Crespi (KC) developed a registry-dependent interlayer potential for graphitic structures.\cite{kolmogorov2005registry}
In the KC potential, the dispersive (vdW) attraction between layers is described
using the same theoretically-motivated $r^{-6}$ term as in LJ,
and $\pi$ orbital overlap is modeled by a Morse\cite{morse1929diatomic} type exponential multiplied by a registry-dependent modifier that depends on the transverse distance between atom pairs.
\mc{The KC potential has been modified and reparameterized to better fit the energy
variations between different stacking states predicted by DFT-D (DFT with
dispersion corrections)\cite{lebedeva2011interlayer}.}
It has also been adapted for other 2D materials such as h-BN\cite{leven2014ilp}
and graphene/h-BN\cite{leven2016interlayer, maaravi2017interlayer} heterostructures.
\mc{The energy corrugation obtained by the KC potential is in agreement with
DFT as shown in \fref{fig:sliding:energy}.}
However, the forces obtained from the KC potential deviate significantly from the DFT results. This implies that equilibrium structures associated with
energy minima will differ as well.
To illustrate this point, consider a graphene bilayer where one layer is
rigidly rotated relative to the other. \fref{fig:rotation:force} shows the force
in the $z$-direction (perpendicular to the layers) acting on the bottom atom on
the rotation axis (atom~1 in the bottom layer in \fref{fig:high:sym:confs}) as
a function of rotation angle.
The force predicted by the KC potential decreases and then increases from
AA ($\pm60^\circ$) to AB ($0^\circ$), whereas DFT predicts a monotonic increase from AA to AB\@.
In particular, the KC potential yields the same $z$-force for the AA and AB stackings\footnote{Note that the $x$ and $y$ components of the force are zero at AA and AB due to symmetry}, which indicates that the KC potential cannot distinguish the overlapping atoms at the rotation center in these states.
This is intrinsic to the KC potential. The force on the central atom in the AA and AB states is identical, regardless of the choice of KC parameters.
The modified KC potential\cite{lebedeva2011interlayer} has the same problem.
The LJ potential does even worse (\fref{fig:rotation:force}) predicting a constant
force on the central atom that is independent of the rotation angle.
In the present paper, a new registry-dependent interlayer potential for graphitic structures is developed that addresses the limitations of the KC potential described above.
A dihedral-angle-dependent term is introduced into the registry modifier of the repulsive part that makes it possible to distinguish forces in AA and AB states.
We refer to this potential as the \ipfullname (\ipshortname).
\ipshortname is validated by showing that it
correctly reproduces the DFT energy and forces for different sliding and rotated states as well as structural and elastic properties.
It is then applied to study structural relaxation in twisted graphene bilayers
and exfoliation of graphene from graphite; these representative example are
large-scale applications that cannot be studied using DFT.
The potential has been implemented as a \ipshortname Model Driver\cite{drip_driver} and the parameterization in this paper has been implemented as a Model\cite{drip_model} at OpenKIM\cite{tadmor2011kim,tadmor2013nsf}.
(See details in
Appendix~\ref{appendix:kim}.)
\section{Definition of new model}
\label{sec:model}
The \ipshortname functional form is
\begin{equation}\label{eq:tot:eng}
\mathcal{V} = \frac{1}{2} \sum_{i \in \text{layer 1}} \sum_{j \in \text{layer 2}} (\phi_{ij} + \phi_{ji}) ,
\end{equation}
where the pairwise interaction is based on the KC form with dihedral modifications:
\begin{widetext}
\begin{equation}
\phi_{ij} = f_\text{c}(x_r) \left[ e^{-\lambda(r_{ij} - z_0 )} \left[C+f(\rho_{ij})
+ g(\rho_{ij}, \{\alpha_{ij}^{(m)}\}) \right]
- A\left (\frac{z_0}{r_{ij}} \right)^6 \right], \; \;
m=1,2,3.
\label{eq:phi}
\end{equation}
\end{widetext}
The cutoff function $f_\text{c}(x)$ is same as that used in the ReaxFF potential\cite{duin2001reaxff} and the interlayer potential for h-BN\cite{leven2014ilp, leven2016interlayer}:
\begin{equation} \label{eq:cutoff}
f_\text{c}(x)=
20x^7 - 70x^6 + 84x^5 - 35x^4 +1,
\end{equation}
for $0\leq x \leq 1$ and vanishes for $x>1$,
while it has zero first and second derivatives at $x=1$;
in the expressions where
this function appears its argument is always non-negative.
The variable $x_r$ in \eqn{eq:phi} is the scaled pair distance
$x_r = r_{ij}/r_\text{cut}$.
The use of $f_\text{c}(x_r)$ ensures that \ipshortname is smooth at the cutoff $r_\text{cut}$ (set to 12~\AA), a feature that the KC model does not possess.
The term with $r_{ij}^{-6}$ dependence in
\eqn{eq:phi} models attractive vdW interactions (as in LJ),
while the repulsive interactions due to orbital overlap are modeled by
the exponential term multiplied by a registry-dependent modifier.
The transverse distance function $f(\rho)$ has the same form as in KC:
\begin{equation}
f(\rho) = e^{-y^2} \left[ C_0 + C_2 y^2 + C_4 y^4 \right],
\; \; y=\frac{\rho}{\delta}
\end{equation}
with its argument in \eqn{eq:phi} given by the expression
\begin{equation}
\rho_{ij}^2 = r_{ij}^2 - (\bm n_i \cdot \bm r_{ij})^2,
\end{equation}
in which $\bm r_{ij}$ is the vector connecting atoms $i$ and $j$, $r_{ij}$ is the corresponding pair distance, and $\bm n_i$ is the layer normal at atom $i$.
For example, as shown in \fref{fig:normal:dihedral}, $\bm n_i$ can be defined as the normal to the plane determined by the three nearest-neighbors of atom $i$: $k_1, k_2$ and $k_3$:
\begin{equation}
\bm n_i =
\frac{\bm r_{k_1k_2}\times \bm r_{k_1k_3}}{\| \bm r_{k_1k_2}\times \bm r_{k_1k_3} \|}.
\end{equation}
Note that in general $\rho_{ij} \neq \rho_{ji}$ because the normals $\bm n_i$ and $\bm n_j$ depend on their local environments.
\begin{figure}
\includegraphics[width=0.60\columnwidth]{normal_dihedral.pdf}
\caption{Schematic representation of an atomic geometry that
defines the normal vectors $\bm n_i$
and $\bm n_j$ and the dihedral angle $\Omega_{k_1 ijl_2}$.}
\label{fig:normal:dihedral}
\end{figure}
The dihedral angle function is given by
\begin{equation} \label{eq:g}
g(\rho, \{\alpha_{ij}^{(m)}\}) = B f_\text{c}(x_\rho) \sum_{m=1}^3 e^{-\eta \alpha_{ij}^{(m)}},
\end{equation}
where $\alpha_{ij}^{(m)}$ is the product of the three cosines of the dihedral angles formed by atom $i$ (in layer~1), its $m$th nearest-neighbor $k_m$, atom $j$ (in layer~2), and its three nearest-neighbors $l_1, l_2$ and $l_3$:
\begin{equation} \label{eq:alpha}
\alpha_{ij}^{(m)} = \cos\Omega_{k_mijl_1} \cos\Omega_{k_mijl_2} \cos\Omega_{k_mijl_3}
\end{equation}
\begin{equation} \label{eq:cos:dihedral}
\cos\Omega_{kijl} = \bm e_{jik} \cdot \bm e_{ijl}
\end{equation}
\begin{equation} \label{eq:normal:to:plane}
\bm e_{jik} = \frac{\bm r_{ik} \times \bm r_{ji}} {\|\bm r_{ik} \times \bm r_{ji}\|}, \qquad
\bm e_{ijl} = \frac{\bm r_{jl} \times \bm r_{ij}} {\|\bm r_{jl} \times \bm r_{ij}\|}.
\end{equation}
To understand the physical origin of the terms defined in
\eqns{eq:alpha}--\eqref{eq:normal:to:plane}, recall that
a dihedral angle $\Omega$ is the angle between two planes defined by four points that intersect at a line defined by two of them as shown in \fref{fig:normal:dihedral}.
Here, the intersection line is defined by atoms $i$ and $j$.
The two planes are then defined by atoms $(j, i, k_1)$ and $(i, j, l_2)$.
The normals to these planes are $\bm e_{jik_1}$ and $\bm e_{ijl_2}$, respectively,
defined in \eqn{eq:normal:to:plane}, with
the corresponding dihedral angle given by \eqn{eq:cos:dihedral}.
The dihedral product $\alpha_{ij}^{(m)}$ monotonically decreases when twisting a graphene bilayer from AB to AA stacking, and consequently can be utilized to construct a potential function that distinguishes AB and AA stacking and the intermediate stacking states.
The cutoff function $f_\text{c}(x_\rho)$ in
\eqn{eq:g} is the same as that in \eqn{eq:cutoff}, and
$x_\rho = \rho/\rho_\text{cut}$,
where we set $\rho_\text{cut} = 1.562~\text{\AA}$ to include only a few
of the computationally expensive 4-body dihedral angle interactions.
The potential has a total of
ten parameters, $C_0$, $C_2$, $C_4$, $C$, $\delta$, $\lambda$, $B$, $\eta$, $A$, and $z_0$, and two cutoffs $r_\text{cut}$ and $\rho_\text{cut}$.
To determine the values of all the parameters that appear in the
\ipshortname potential, we constructed
a training set of energies and forces for graphene bilayers
at different separation, sliding, and twisting states.
The training set is generated from DFT calculations using the Vienna Ab initio Simulation Package (VASP)\cite{vasp1,vasp2}.
The exchange-correlation energy of the electrons is treated within the generalized gradient approximated (GGA) functional of Perdew, Burke and Ernzerhof (PBE)\cite{pbe}.
Standard density functionals such as the local density approximation (LDA) and GGA accurately represent Pauli repulsion in interlayer interactions,
but fail to capture vdW forces that result from dynamical correlations between fluctuating charge distributions.\footnote{GGA predicts no binding at all at physically meaningful spacings for graphite. LDA gives the correct interlayer spacing for AB stacking, however, it underestimates the exfoliation energy by a factor of two and overestimates the compressibility.\cite{kolmogorov2005registry}}
To address this limitation,
various approximate corrections have been proposed including the D2 method\cite{grimme2006semiempirical}, the D3 method\cite{grimme2010consistent},
the Tkatchenko and Scheffler (TS) method\cite{tkatchenko2009accurate},
the TS method with iterative Hirshfeld partitioning (TSIHP) method\cite{bucko2013improved}, the many-body dispersion (MBD) method\cite{tkatchenko2012accurate},
and the dDsC dispersion correction method\cite{steinmann2011generalized}.
To select a correction for the \ipshortname training set, we used these dispersion
correction methods to compute a number of structural, energetic, and elastic properties.
\mc{The results are shown in \tref{tab:properties} along with experimental
values and more accurate adiabatic-connection fluctuation-dissipation theory
based random-phase-approximation (ACFDT-RPA) computations that have been shown
to provide a very accurate description of vdW interactions\cite{zhou2015van,lebegue2010cohesive}.
The conclusion from these comparisons is that
D2 and D3 provide inaccurate estimates for the layer spacing of AB graphene and graphite
($d_\text{AB}$ and $d_\text{graphite}$), and TS, TSIHP, and dDsC significantly
overestimate the graphite binding energy $E_\text{graphite}$.
MBD provides the best overall accuracy for all considered properties and is therefore
the vdW correction used in this work together with the PBE functional.}
\begin{table*}
\caption{Properties obtained from various DFT vdW corrections compared with
ACFDT-RPA and experimental results.
\mc{
Also included are results from various empirical potentials.
The properties include:
equilibrium layer spacings of bilayer graphene in AB stacking, $d_\text{AB}$,
bilayer graphene in AA stacking, $d_\text{AA}$,
and graphite, $d_\text{graphite}$;
optimal interlayer binding energies for bilayer graphene (binding energy at the
equilibrium spacing in AB stacking), $E_\text{AB}$,
and graphite, $E_\text{graphite}$;
energy differences between AA-stacked and AB-stacked bilayers, $\Delta E_\text{AA-AB}$,
and SP and AB stackings, $\Delta E_\text{SP-AB}$, at a layer spacing of $d=3.4~\text{\AA}$;
and the elastic modulus along the $c$-axis for graphite, $C_{33}$.
All properties are computed using the in-plane lattice constant $a=2.46~\text{\AA}$.
}
}
\label{tab:properties}
\begin{ruledtabular}
\begin{tabular}{ccccccccc}
&$d_\text{AB}$ &$d_\text{AA}$ &$d_\text{graphite}$ &$E_\text{AB}$
&$E_\text{graphite}$ &$\Delta E_\text{AA-AB}$ &$\Delta E_\text{SP-AB}$ &$C_{33}$\\
& (\AA) &(\AA) &(\AA) &(meV/atom)
&(meV/atom) &(meV/atom) &(meV/atom) &(GPa) \\
\hline
PBE+D2 &3.248 &3.527 &3.218 &24.84 &55.20 &10.35 &1.16 &39.12 \\
PBE+D3 &3.527 &3.713 &3.483 &21.40 &47.09 &3.80 &0.42 &35.04 \\
PBE+TS &3.357 &3.511 &3.329 &36.36 &82.33 &7.97 &1.01 &68.31 \\
PBE+TSIHP &3.379 &3.529 &3.350 &35.73 &80.42 &7.48 &1.22 &64.73 \\
PBE+MBD &3.423 &3.638 &3.398 &22.63 &48.96 &6.17 &0.69 &31.64 \\
PBE+dDsC &3.447 &3.639 &3.410 &28.04 &63.00 &5.53 &0.74 &38.43 \\
ACFDT-RPA &3.39\footnotemark[1] & &3.34\footnotemark[2] & &48\footnotemark[2] & & &36\footnotemark[2] \\
Experiment & & &3.34\footnotemark[3] & &\mcr{$43\pm5$}\footnotemark[4],
$35\pm10$\footnotemark[5], $52\pm5$\footnotemark[6] &\mcr{7.7\footnotemark[7]}
&\mcr{0.86\footnotemark[7]} &36.5\footnotemark[8], 38.7\footnotemark[9] \\
\hline
AIREBO &3.391 &3.418 &3.357 &22.85 &48.86 &0.41 &0.04 &37.78 \\
LCBOP &3.346 &3.365 &3.346 &12.51 &25.03 &0.47 &0.01 &29.77 \\
KC &3.374 &3.602 &3.337 &21.60 &47.44 &6.07 &0.44 &34.45 \\
\ipshortname &3.439 &3.612 &3.415 &\mc{23.05} &\mc{47.38} &6.02 &0.71 &\mc{32.00} \\
\end{tabular}
\end{ruledtabular}
\footnotetext[1]{Ref.~\onlinecite{zhou2015van}.}
\footnotetext[2]{Ref.~\onlinecite{lebegue2010cohesive}.}
\footnotetext[3]{Ref.~\onlinecite{baskin1955lattice}.}
\footnotetext[4]{Ref.~\onlinecite{girifalco1956energy}.}
\footnotetext[5]{Ref.~\onlinecite{benedict1998microscopic}.}
\footnotetext[6]{Ref.~\onlinecite{zacharia2004interlayer}.}
\footnotetext[7]{Ref.~\onlinecite{popov2012barriers}. \mcr{Values inferred from experimental data on shear mode frequencies.}}
\footnotetext[8]{Ref.~\onlinecite{blakslee1970elastic}.}
\footnotetext[9]{Ref.~\onlinecite{bosak2007elasticity}.}
\end{table*}
Each monolayer of the graphene bilayer is modeled as a slab with in-plane lattice constant $a=2.46~\text{\AA}$, and the supercell size in the direction perpendicular to the slab is set to $30~\text{\AA}$ to minimize the interaction between periodic images.
The sampling grid in reciprocal space is $20 \times 20 \times 1$, with an energy cutoff of $500~\text{eV}$.
A primitive unit cell of a graphene bilayer consists of four basis atoms.
To generate a graphene bilayer with different translational registry, the two atoms in the bottom layer are fixed at fractional positions $\bm b_1 = (0,0,0)$ and $\bm b_2 = (\frac{1}{3},\frac{1}{3},0)$ relative to the graphene lattice vectors $\bm a_1, \bm a_2$, and $\bm c$, where $\bm c$ is perpendicular to the plane defined by $\bm a_1$ and $\bm a_2$ with length equal to the interlayer distance $d$.
The other two atoms are located at $\bm r_1 = (p,q,1)$ and $\bm r_2 = (p+\frac{1}{3}, q+\frac{1}{3}, 1)$.
The two parameters $p\in[0,1]$ and $q\in[0,1]$ determine the translational registry.
For example, the graphene bilayer is in AA stacking (\fref{fig:aa}) when $p=0$ and $q=0$, and in AB stacking (\fref{fig:ab}) when $p=\frac{1}{3}$ and $q=\frac{1}{3}$.
Due to the symmetry of the honeycomb lattice, only 1/12 of the area defined by $\bm a_1$ and $\bm a_2$ needs to be sampled to fully explore all translational
registry states (see the shaded region in \fref{fig:sample:grid}).
The \ipshortname training set comprised the seven states indicated in the
shaded region of \fref{fig:sample:grid}, specifically
$(p,q)=(0,0)$, $(0,\frac{1}{6})$, $(0,\frac{2}{6})$, $(0,\frac{3}{6})$,
$(\frac{1}{6}, \frac{1}{6})$, $(\frac{1}{6}, \frac{2}{6})$,
$(\frac{2}{6}, \frac{2}{6})$.
\mc{These states include all the high-symmetry states of interest, including AA, AB, and the saddle point (SP) stacking ($p=0, q=\frac{3}{6}$).
The seven translational registry states are sampled at different layer spacings
$d$, varying from $2.7~\text{\AA}$ to $4.5~\text{\AA}$ with a step size of
$0.1~\text{\AA}$.
For layer spacings larger than $4.5~\text{\AA}$ but smaller than the cutoff
$r_\text{cut} = 12~\text{\AA}$, only bilayer graphene in AB stacking is included
since the difference between the stacking states in this range is negligible
(see discussion in \sref{sec:test}).
Thus $7\times 19 + 75 = 208$ translation configurations are included in the training set.
}
\begin{figure}
\begin{subfigure}{0.32\columnwidth}
\includegraphics[width=\columnwidth]{AA_stacking.pdf}
\caption{}
\label{fig:aa}
\end{subfigure}
\begin{subfigure}{0.32\columnwidth}
\includegraphics[width=\columnwidth]{AB_stacking.pdf}
\caption{}
\label{fig:ab}
\end{subfigure}
\begin{subfigure}{0.32\columnwidth}
\includegraphics[width=\columnwidth]{sample_grid.pdf}
\caption{}
\label{fig:sample:grid}
\end{subfigure}
\caption{Primitive unit cell of a graphene bilayer: (a) AA stacking, (b) AB stacking, and (c) unique sampling region and sampling points.}
\end{figure}
In addition to translation configurations, a set of
twisted bilayer configurations are included in the training set.
It is possible to construct a commensurate supercell arbitrarily close to any twisting angle according to the commensuration condition\cite{shallcross2010electronic,zhang2017energy,tritsaris2016}
\begin{equation}\label{eq:commensuration}
\theta = \cos^{-1}\left(\frac{3n^2 - m^2}{3n^2 + m^2}\right) ,
\end{equation}
where $m$ and $n$ are any two integers satisfying $0 < m < n$.
As an example, considering the AB-stacked bilayer in \fref{fig:twisted:bilayer:before}, a commensurate bilayer can be obtained by rotating one of the layers by $\theta = 27.8^\circ$ ($m=3,n=7$) with the supercell shown in \fref{fig:twisted:bilayer:after}.
Four types of twisted bilayers with rotation angles $9.43^\circ$, $21.79^\circ$, $32.30^\circ$ and $42.10^\circ$ (corresponding to $(m,n) = (1,7)$, $(1,3)$, $(1,2)$ and $(2,3)$) are included in the training set.
\mc{The twisted configurations were evaluated at layer spacings from 3.0~\AA\xspace
to 4.0~\AA\xspace with a step size of 0.1~\AA.
Thus $4 \times 11 = 44$ twisted configurations are included in the training set.
}
This does not include rotations for
$\theta = 0^\circ$ and $\theta = \pm60^\circ$ corresponding
to the AB and AA stacking states, respectively,
which are already included in the training set.
\begin{figure}
\begin{subfigure}{0.6\columnwidth}
\includegraphics[width=\columnwidth]{twisted_bilayer_angle_27_8.pdf}
\caption{}
\label{fig:twisted:bilayer:before}
\end{subfigure}
\begin{subfigure}{0.35\columnwidth}
\includegraphics[width=\columnwidth]{twisted_bilayer_angle_27_8_rotated.pdf}
\caption{}
\label{fig:twisted:bilayer:after}
\end{subfigure}
\caption{Example of commensuration of a graphene bilayer. (a) The two layers are commensurate when rotated relative to each other by $\cos^{-1} (\frac{23}{26})=27.8^\circ$, which corresponds to $m=3, n=7$ according to the condition in
\eqn{eq:commensuration}. (b) The resulting supercell after rotation, with 26 atoms in each layer.}
\label{fig:twisted:bilayer}
\end{figure}
The parameters of the potential are optimized by minimizing a loss function that quantifies the difference between the interatomic potential predictions and the training set. The training set includes $M$ configurations with
concatenated coordinates $\bm r_m$ for $m\in[1,M]$, such that
$\bm r_m\in\mathbb{R}^{3N_m}$ where $N_m$ is the number of atoms
in configuration $m$. The loss function is
\begin{align} \label{eq:loss}
L(\bm{\xi})
=& \sum_{m=1}^M \frac{1}{2} w^\text{e}_m \left[E_m(\bm r_m; \bm{\xi}) - E_m^\text{DFT} \right]^2 \nonumber \\
+& \sum_{m=1}^M \frac{1}{2} w^\text{f}_m \|\bm f(\bm r_m; \bm{\xi}) - \bm f_m^\text{DFT} \|^2,
\end{align}
where $\bm{\xi}$ is the set of potential parameters,
$E_m$ and
$\bm{f}(\bm r_m; \bm{\xi}) = -\left.(\partial\mathcal{V}/\partial \bm{r})\right|_{\bm{r}_m} \in \mathbb{R}^{3N_m}$
are the \ipshortname potential energy and concatenated forces
in configuration $m$,
and $w^\text{e}_m$ and $w^\text{f}_m$ are the weights associated with
the energy and forces of configuration $m$. For energy in units of eV
and forces in units of eV/\AA, these weights
have units of eV$^{-2}$ and (eV/\AA)$^{-2}$, respectively.
The DFT energy and forces used in the loss function, \eqn{eq:loss},
$E_m^\text{DFT}$ and $\bm f_m^\text{DFT}$, require explanation.
Since DFT provides only the total energy and forces on atoms
due to both intralayer and interlayer interactions it is necessary
to separate out the interlayer contributions when constructing the training set.
This is accomplished as follows. For configuration $m$,
first the total energy and forces of the bilayer are
obtained from DFT: $E_m^\text{DFT, bilayer}$, $\bm f_m^\text{DFT, bilayer}$.
Then each monolayer is computed separately by
removing all atoms from the other monolayer. Thus, there will be two
energies, $E_m^\text{DFT, layer 1}$ and $E_m^\text{DFT, layer 2}$,
and two forces, $\bm f_m^\text{DFT, layer 1}$ and $\bm f_m^\text{DFT, layer 2}$
(although each force vector will only contain nonzero components for the
atoms belonging to its monolayer).
The DFT interlayer energy and forces appearing in \eqn{eq:loss}
are then defined as:
\begin{align}
E_m^\text{DFT} &= E_m^\text{DFT, bilayer}
- E_m^\text{DFT, layer 1} - E_m^\text{DFT, layer 2},
\label{eq:E:DFT}
\\
\bm f_m^\text{DFT} &=
\bm f_m^\text{DFT, bilayer}
- \bm f_m^\text{DFT, layer 1} - \bm f_m^\text{DFT, layer 2}.
\label{eq:f:DFT}
\end{align}
In the present case, the training set includes $M=252$ configurations.
Both the energy weight $w^\text{e}_m$ and force weight $w^\text{f}_m$
($m=1,\dots,252$) are set to 1.
The optimization was carried out using the KIM-based Learning-Integrated
Fitting Framework (KLIFF)\cite{kliff}
with a geodesic Levenberg-Marquardt minimization algorithm\cite{wen2016potfit, transtrum2010nonlinear,transtrum2011geometry}.
The objective is to find the set of parameters $\bm{\xi}$ that minimizes $L(\bm{\xi})$.
The optimal parameter set identified by this process and preset cutoffs
are listed in \tref{tab:params}.
\begin{table}
\caption{\mc{\ipshortname parameters obtained by minimizing the loss function
$L(\bm\xi)$ defined in \eref{eq:loss} and preset cutoffs.}}
\label{tab:params}
\begin{ruledtabular}
\begin{tabular}{cccc}
Parameter &Value &Parameter &Value \\
\hline
$C_0$~(meV) &11.598 &$B$~(meV) &7.6799 \\
$C_2$~(meV) &12.981 &$\eta$~(1/\AA) &1.1432 \\
$C_4$~(meV) &32.515 &A~(meV) &22.216 \\
$C$~(meV) &7.8151 &$z_0$~(\AA) &3.3400 \\
$\delta$~(\AA) &0.83679 &$r_\text{cut}$~(\AA) &\mc{12} \\
$\lambda$~(1/\AA) &2.7158 &$\rho_\text{cut}$~(\AA) &1.562 \\
\end{tabular}
\end{ruledtabular}
\end{table}
\section{Testing of the new potential}
\label{sec:test}
We performed an extensive set of calculations to test
the ability of \ipshortname to reproduce its training set
(described in \sref{sec:model}),
and test its transferability to configurations outside the training set.
The calculations using the potential were performed with LAMMPS\cite{plimpton1995fast,lammps} and DFT calculations with VASP\cite{vasp1,vasp2}.
Periodic boundary conditions are applied in both in-plane directions, and the in-plane lattice constant is fixed at $a=2.46~\text{\AA}$.
The setup for the DFT computations is the same as that used for generating the training set in \sref{sec:model}.
\begin{figure*}
\begin{subfigure}{0.65\columnwidth}
\includegraphics[width=\columnwidth]{fz_sep3_4_p3q7_lj.pdf}
\caption{}
\label{fig:z:forces:lj}
\end{subfigure}
\begin{subfigure}{0.65\columnwidth}
\includegraphics[width=\columnwidth]{fz_sep3_4_p3q7_kc.pdf}
\caption{}
\end{subfigure}
\begin{subfigure}{0.65\columnwidth}
\includegraphics[width=\columnwidth]{fz_sep3_4_p3q7_wen.pdf}
\caption{}
\end{subfigure}
\caption{Out-of-plane component of the forces on the 26 atoms in the bottom layer of the twisted bilayer shown in \fref{fig:twisted:bilayer} (each represented as a bar) computed from DFT and the (a) LJ potential, (b) KC potential, and (c) \ipshortname model. The layer spacing is 3.4~\AA.}
\label{fig:z:forces}
\end{figure*}
\fref{fig:z:forces} shows the unrelaxed forces on the atoms in the bottom layer of the twisted bilayer shown in \fref{fig:twisted:bilayer} with a layer spacing of $d=3.4~\text{\AA}$.
There are 26 atoms in the bottom layer. For each, the out-of-plane force ($z$-component)
is displayed as a bar.
The plot compares the results of LJ, KC and \ipshortname with DFT.
For the LJ potential, the parameterization in the AIREBO potential is used.
The \ipshortname forces are in very good agreement with DFT, whereas the LJ
potential yields almost zero forces, and the KC potential greatly overestimates
the forces. (Note that the force ranges in the three panels are different).
The force on the central atom when twisting a bilayer obtained from \ipshortname (denoted as 1 in \fref{fig:high:sym:confs}) is
displayed in \fref{fig:rotation:force} as a function of rotation.
The results are in agreement with DFT, indicating that the
dihedral modification in \ipshortname successfully addresses the deficiency
of the KC potential discussed in \sref{sec:need}.
To investigate the accuracy of the potentials in a dynamical setting,
trajectories are generated at a temperature of 300~K
using \emph{ab initio} molecular dynamics (AIMD)
for bilayers in AA and AB stackings, and the twisted bilayer shown in
\fref{fig:twisted:bilayer}.
For each configuration along the trajectories, the DFT forces due to interlayer
interactions are computed using the procedure defined in \eqn{eq:f:DFT}
and explained above. Next, LAMMPS is used to compute the LJ, KC and \ipshortname
interlayer forces for the AIMD configurations.
The error in the potential forces is shown in \fref{fig:force:error}.
Each dot in the plot represents one atom pulled from one of the configurations
along the AIMD trajectories.
\mc{The horizontal coordinate in the plot is the magnitude of the in-plane component (left panels)
and out-of-plane component (right panels) of the DFT interlayer force acting on the atom. The force
is separated in this way because the in-plane component is significantly smaller than the out-of-plane
component. (Note that this is only the force due to interlayer interactions. The force due
to intralayer bonding is not included.)
The vertical coordinate is the magnitude of the
difference between the potential and DFT force vectors for that atom.
We see that the in-plane force error for LJ aligns with the diagonal,
i.e.\ the error equals the DFT force, which means that LJ predicts an
in-plane force close to zero.
This is because LJ provides a poor model for the
anisotropic overlap of electronic orbitals between adjacent layers and thus
has almost no barrier for relative sliding.
The KC model performs better in the sense that it predicts resistance
to sliding, however the overall accuracy in forces is poor
(see \sref{sec:need} for a discussion of the limitations of the KC model).
In contrast, \ipshortname provides consistently accurate in-plane forces
across the range of DFT forces with errors less than 20~meV/\AA.
For the out-of-plane component both LJ and \ipshortname perform comparably
providing good accuracy across the range of DFT forces, whereas
the KC model again shows poor accuracy with very large errors
in some cases.}
\begin{figure}
\includegraphics[width=1\columnwidth]{force_error_vs_DFT_force.pdf}
\caption{\mc{Deviation of potential forces from DFT results due to interlayer
interactions. The configurations are taken from three AIMD trajectories at 300~K.}}
\label{fig:force:error}
\end{figure}
Next, we consider energetics.
The interlayer binding energy $E_\text{b}$ of a graphene bilayer as a function of layer spacing $d$ is shown in \fref{fig:energy:layer:dist} for AB and AA stackings and the twisted configuration shown in \fref{fig:twisted:bilayer}.
The LJ potential (\fref{fig:energy:layer:dist:lj})
cannot distinguish these states and gives nearly identical binding energy versus layer spacing curves for all three.
Both KC (\fref{fig:energy:layer:dist:kc}) and \ipshortname (\fref{fig:energy:layer:dist:present}) correctly capture the energy differences between the three stacking states.
For all three potentials, the twisted bilayer curve lies between the other two,
which is expected since the AB and AA stackings are minimum and maximum energy
states. Also notable is that at large layer spacing, the curves for all three
stacking states merge since registry effects due to $\pi$-orbital overlap
become negligible and interactions are dominated by vdW attraction, which are the
same for all three states and captured equally well by all three potentials.
\begin{figure*}
\begin{subfigure}{0.65\columnwidth}
\includegraphics[width=\columnwidth]{energy_vs_layer_dist_lj.pdf}
\caption{}
\label{fig:energy:layer:dist:lj}
\end{subfigure}
\begin{subfigure}{0.65\columnwidth}
\includegraphics[width=\columnwidth]{energy_vs_layer_dist_kc.pdf}
\caption{}
\label{fig:energy:layer:dist:kc}
\end{subfigure}
\begin{subfigure}{0.65\columnwidth}
\includegraphics[width=\columnwidth]{energy_vs_layer_dist_wen.pdf}
\caption{}
\label{fig:energy:layer:dist:present}
\end{subfigure}
\caption{Interlayer binding energy $E_\text{b}$ of a graphene bilayer versus layer spacing $d$ for AA stacking, AB stacking, and a twisted bilayer with rotation angle $\theta=27.8^\circ$ (see \fref{fig:twisted:bilayer}) using (a) LJ potential, (b) KC potential, and (c) \ipshortname model, compared to DFT results.}
\label{fig:energy:layer:dist}
\end{figure*}
A more complete view of the interlayer energetics is obtained by considering
the generalized stacking fault energy (GSFE) surface obtained by sliding one layer
relative to the other while keeping the layer spacing fixed.
\fref{fig:gsfe} shows the results for a layer spacing of $d=3.4~\text{\AA}$
calculated using \ipshortname and DFT.
\ipshortname is in quantitative agreement with DFT results.
The KC GSFE has a similar appearance and the LJ GSFE is nearly flat.
The KC and LJ results are not included for
brevity, but the energies of the three potentials along the dashed line in the
left panel of \fref{fig:gsfe} are displayed in \fref{fig:sliding:energy}.
\begin{figure}
\includegraphics[width=\columnwidth]{bilayer_gsfe.pdf}
\caption{
\mc{The GSFE obtained by sliding one layer relative to the other at a fixed
layer spacing of $d=3.4~\text{\AA}$.
The energy is relative to the AB state, which is $-22.98~\text{meV/atom}$
for \ipshortname (on the left) and $-22.33~\text{meV/atom}$ for DFT (on the
right).
The sliding parameters $\Delta \bm a_1$ and $\Delta \bm a_2$ are in units of
in-plane lattice constant $a=2.46~\text{\AA}$.
}
}
\label{fig:gsfe}
\end{figure}
As a final test, \tref{tab:properties} shows the predictions of \ipshortname for
a number of structural, energetic, and elastic properties. The table also
includes results for the LCBOP\cite{los2003intrinsic} and
AIREBO\cite{stuart2000airebo} potentials, as well as DFT and experimental
results as described in \sref{sec:model}.
The LCBOP potential uses two Morse\cite{morse1929diatomic} type terms to
model long-range interactions, and LJ\cite{lennardjones1931}
is used in the AIREBO potential as discussed in \sref{sec:intro}.
The properties of the \ipshortname model are in good agreement with the PBE+MBD
DFT computations with which the training set was generated.
\section{Applications}
\label{sec:app}
To further compare the predictions of the KC potential and \ipshortname,
we carried out
two large-scale simulations, beyond the capability of DFT:
(1) structural relaxation in a twisted graphene bilayer, and (2) exfoliation of a graphene layer off graphite. In these simulations, the interlayer interactions are modeled using either KC or \ipshortname, and the REBO\cite{brenner2002rebo} potential is used to model the intralayer interactions.
\subsection{Structural relaxation of a twisted graphene bilayer}
\label{sec:bilayer:relax}
The electronic properties of stacked 2D materials can be manipulated by controlling
the relative rotation between the layers, which in turn leads to different
structural relaxation. A prototypical problem is the twisting of a graphene bilayer.
The bilayer is created by rotating one layer relative to the other by $\theta=0.82^\circ$,
setting $(m,n) = (1,81)$ as discussed in \sref{sec:model}.
The out-of-plane relaxation
$\delta$ of an atom is obtained by subtracting the mean out-of-plane coordinates of all atoms in the top layer from the out-of-plane coordinate of that atom:
\begin{equation}
\delta_i = z_i - \frac{1}{N} \sum_{j=1}^N z_j
\end{equation}
where $z_i$ is the out-of-plane coordinate of atom $i$ in the top layer and $N=9842$ is the number of atoms in the top layer\footnote{Using the atoms in the bottom layer will yield the same results because the relaxed structure of the bottom layer and the top layer are identical.}.
\begin{figure}
\begin{subfigure}{1.0\columnwidth}
\includegraphics[width=\columnwidth]{relax_twsited_bilayer_z_corrugation.pdf}
\caption{}
\label{fig:twisted:bilayer:relax:a}
\end{subfigure}
\begin{subfigure}{0.9\columnwidth}
\includegraphics[width=\columnwidth]{relax_twsited_bilayer_z_corrugation_along_diagonal.pdf}
\caption{}
\label{fig:twisted:bilayer:relax:b}
\end{subfigure}
\caption{Out-of-plane relaxation in a twisted bilayer with a relative rotation of $\theta=0.82^\circ$.
(a) Contour plot obtained from the \ipshortname model and the KC potential, and (b) relaxation along the diagonal indicated by the dashed line in panel (a).
\mc{The bilayers shown in the figure corresponds to $3\times3$ supercells used in the computation,
i.e.\ $a_1$ and $a_2$ are in units of in-plane lattice constant $a=2.46~\text{\AA}$.} }
\label{fig:twisted:bilayer:relax}
\end{figure}
The out-of-plane relaxation of the twisted bilayer is plotted in \fref{fig:twisted:bilayer:relax}.
The results of the \ipshortname and KC models are qualitatively similar.
The bright spots correspond to high-energy AA stacking, the long
narrow ribbons correspond to SP stacking, and the triangular regions
correspond to alternating AB and BA stacking. It has been shown that the
formation of this structure is due to local rotation at AA domains.\cite{zhang:tadmor:2018}
Quantitatively, however, the two potentials give different out-of-plane relaxation, especially at the peaks as seen in \fref{fig:twisted:bilayer:relax:b}.
The peak value predicted by \ipshortname is $0.076~\text{\AA}$, which is $26\%$ smaller than the KC potential value of $0.103~\text{\AA}$.
This difference at the peaks could lead to significant differences in electronic properties because twisted graphene bilayers develop highly-localized states around AA-stacked regions for small twist angles
\cite{gonzalez2017electrically}.
\subsection{Exfoliation of graphene from graphite}
\begin{figure}
\begin{subfigure}{0.9\columnwidth}
\includegraphics[width=\columnwidth]{peel_one_layer_schematics.pdf}
\caption{}
\label{fig:peel:one:layer:a}
\end{subfigure}
\begin{subfigure}{0.9\columnwidth}
\includegraphics[width=\columnwidth]{peel_one_layer_force_vs_disp.pdf}
\caption{}
\label{fig:peel:one:layer:b}
\end{subfigure}
\caption{(a) Schematic demonstrating the process of peeling a graphene layer off graphite, and (b) the normal force, $f_z$, needed to peel the top layer as a function of the displacement at the left end of the top layer, $d - d_0$. The armchair direction of
graphite is aligned with the $x$-axis. The initial layer spacing is $d_0=3.35~\text{\AA}$.}
\label{fig:peel:one:layer}
\end{figure}
Graphene can be prepared by exfoliating graphite. In this process, the vdW
attraction between layers is overcome by peeling a single layer off a
graphite crystal. A method as simple as sticking scotch tape to graphite and
applying an upward force can be used.\cite{novoselov2004electric}
To simulate this process, one edge of the top layer of a graphite crystal
is pulled up under displacement control conditions as illustrated in
\fref{fig:peel:one:layer:a}.
The atoms at the left end of the top layer are displaced in the $z$-direction according to $d = d_0 + 0.2k$, where $d_0=3.35~\text{\AA}$ is the initial layer spacing, and $k = 0, 1, \dots, 99$ is the step number.
At each step $k$, once the displacement is applied to the left atoms,
the remaining atoms in the top layer are relaxed.
The substrate (bottom three layers) is kept rigid during this process.
The system contains 600 atoms in each layer of size $105.83~\text{\AA}$ and $14.76~\text{\AA}$ in the $x$ and $y$ directions, respectively.
The system is periodic in the $y$ direction, and non-periodic the other two directions.
The normal force, $f_z$, needed to pull the left end of the top layer is plotted in \fref{fig:peel:one:layer:b}.
Both the KC and \ipshortname models give qualitatively similar results.
The force first increases as the left end is pulled up and then exhibits a sudden
drop at about 3~\AA. The normal force has two contributions:
(1) interlayer interactions with atoms in the substrate;
and (2) covalent-bonded interactions with other atoms in the top layer.
The former is almost unchanged before and after the load drop, therefore the drop is mainly due to the in-plane interactions in the top layer.
Before the load drop, the right-end of the top layer is trapped in a local minimum created by the substrate (similar to the one denoted as AB in \fref{fig:gsfe}, although there we only consider a graphene bilayer), and consequently as the left end is pulled up, the top layer experiences an increasing axial strain.
At about $3~\text{\AA}$, the right-end of the top layer snaps into an adjacent local minimum by moving in the negative $x$ direction (see Supplemental Material for a movie showing the snap-throughs associated with the load drop).
As a result, the axial strain in the top layer is released and the load is
reduced.
The same explanation applies to the load drop at a displacement of about $15~\text{\AA}$, and it is expected to continue to occur periodically
with continued pulling.
As for the results in \sref{sec:bilayer:relax}, KC and \ipshortname are in
qualitative agreement, but there are quantitative differences.
The KC potential predicts an initial peeling load of about 0.65~eV/\AA,
which is about 75\% of the 0.87~eV/\AA\xspace value predicted by \ipshortname.
The second snap-through occurs at a displacement of $16.6~\text{\AA}$ for \ipshortname,
and at $15.0~\text{\AA}$ for KC.
\section{Summary}
\label{sec:sum}
The interlayer interactions in stacked 2D materials play an important role in determining the functionality of many nanodevices.
For graphitic structures, the two-body pairwise LJ potential is too smooth to model the energy corrugation in different stacking states.
The registry-dependent KC potential improves on this and correctly captures the energy variation,
but fails to yield reasonable forces. In particular, the KC model does not distinguish forces
on atoms in the AA and AB stacking states that are different in DFT calculations.
The KC model is also discontinuous at the cutoff, which can lead to difficulties
in energy minimization and loss of energy conservation in dynamic applications.
To address these limitations, we developed a new potential for graphitic structures based on the
KC model.
The \ipfullname (\ipshortname) has a smooth cutoff
and includes a dihedral-angle-dependent term to distinguish different stacking
states and obtain accurate forces.
The potential parameters were determined by training on a set of energies
and forces for a graphene bilayer at different layer spacing, sliding and
twisting, computed using GGA-DFT calculations, augmented with the MBD dispersion
correction to account for the long-range vdW interactions.
To test the quality of the potential,
we employed it to compute energetics, forces, and structural and elastic properties for a graphene bilayer in different states and graphite.
The validation tests show that compared with first-principles results:
\begin{enumerate}
\item \ipshortname correctly predicts the equilibrium layer spacing, interlayer binding energy, and generalized stacking fault energy of a graphene bilayer, as well as the equilibrium layer spacing of graphite.
\item \mc{\ipshortname underestimates the $c$-axis elastic modulus $C_{33}$ of graphite by about 10\%
relative to ACFDT-RPA and experiments, but this result
is in good agreement with PBE+MBD to which \ipshortname was fit.}
\item \ipshortname provides more accurate forces than the KC model across the entire range of bilayer rotations and in particular distinguishes the forces in the AA and AB states that the KC potential cannot.
\end{enumerate}
In two large-scale applications, not amenable to DFT calculations,
we showed that
\ipshortname and the KC potential agree qualitatively, but differ quantitatively
by 26\% in the out-of-plane relaxation of a twisted graphene bilayer,
and by 23\% in the normal force required to peel one graphene layer off graphite.
The added four-body dihedral-angle-dependent correction in \ipshortname
is very short-ranged ($\rho_\text{cut}=1.562$~\AA) and therefore the
computational overhead relative to KC is small. In fact, for the large-scale
applications (bilayer relaxation and peeling) described in \sref{sec:app},
\ipshortname was actually faster than the KC potential in terms of the
overall computation time due to improved convergence.
Although \ipshortname was parameterized against a training set consisting of graphene bilayers, it can be used to describe interlayer interactions for other systems such as graphite and multi-walled carbon nanotubes where the carbon atoms are arranged in layers.
This potential only provides a description of the interlayer interactions, and therefore must be used together with a companion model that provides the intralayer interactions, such as the Tersoff\cite{tersoff1988empirical,tersoff1989modeling} or REBO\cite{brenner1990physical,brenner2002rebo} potentials.
The \ipshortname functional form and associated carbon parameterization
are archived in the
OpenKIM repository\cite{drip_driver,drip_model,tadmor2011kim} at \url{https://openkim.org}.
They can be used with any KIM-compliant molecular simulation code,
see Appendix~\ref{appendix:kim} for details.
\begin{acknowledgments}
This research was partly supported by the Army Research Office (W911NF-14-1-0247) under the MURI program, and the National Science Foundation (NSF) under grants No.~DMR-1408211 and DMR-1408717.
The authors wish to acknowledge the Minnesota Supercomputing Institute (MSI) at the University of Minnesota for providing resources that contributed to the results reported in this paper.
M.~W.~thanks the University of Minnesota Doctoral Dissertation Fellowship for supporting his research.
The authors thank Alexey Kolmogorov for reading the manuscript and for his valuable comments.
\end{acknowledgments}
|
\section{Introduction}
\setcounter{equation}{0}
In this paper we analyze the massive gravity theory constructed in 1965 by
Ogievetsky and Polubarinov (OP) \cite{OP}\footnote{V. I. Ogievetsky and
I. V. Polubarinov worked in Dubna in the Soviet Union times.}.
To our knowledge, this had been the
first serious work on massive gravity after Fierz and Pauli \cite{Fierz:1939ix}, and
the first ever systematic study of interacting massive gravitons.
Among other things, OP obtained one of the ``ghost-free" massive gravity models
rediscovered again only in 2010 \cite{deRham:2010kj}. However, their work
is almost unknown in the modern massive gravity community,
presumably because their strategy was quite different
from what is generally adopted at present.
Therefore,
in what follows we shall present our analysis of the OP theory and of
some of its applications.
To understand the OP's motivations, consider free massive gravitons in Minkowski space described
by a symmetric tensor ${ \Psi }_{\mu\nu}$ subject to \cite{Fierz:1939ix}
\begin{eqnarray}
(\partial^\sigma\partial_\sigma -m^2){ \Psi }_{\mu\nu}&=&0\,, \label{1a} \\
\partial^\sigma { \Psi }_{\sigma\mu}&=&0\,, \label{2a} \\
{ \Psi }^\sigma_{~\sigma}&=&0\,. \label{3a}
\end{eqnarray}
The tensor ${ \Psi }_{\mu\nu}$ has 10 independent components, but
the five conditions in \eqref{2a} and \eqref{3a}
eliminate the spin-1 and two spin-0 representations
and there remains only the massive spin-2
carrying $5=10-5$ degrees of freedom (DoF). As emphasized already by Fierz and Pauli
(FP) \cite{Fierz:1939ix}, the
Lorentz condition \eqref{2a} is absolutely essential, because if it
were replaced by some other four conditions to keep the same number of DoF,
the canonical energy of ${ \Psi }_{\mu\nu}$ would be non-positive.
The OP's goal was to find a non-linear completion for Eqs.\eqref{1a}--\eqref{3a}.
They adopted the field-theoretical approach initiated by Papapetrou \cite{Papapetrou:1948jw},
Gupta \cite{Gupta}, and Feynman \cite{Feynman:1963ax}, and considered gravitons
as interacting fields in {\it flat space}. Therefore,
they kept
${ \Psi }_{\mu\nu}$ (or rather ${ \Psi }^{\mu\nu}$) as the principle variables
and were looking for non-linear
terms to be added to \eqref{1a}
to describe the graviton interactions.
To illustrate the idea, let us consider the massless case -- the General Relativity.
It is well known that the Einstein equations can be represented in the Papapetrou form
(see, e.g. \cite{poisson}) as
\begin{equation} \label{Papa}
\partial_{\mu}\partial_\nu \left(
{\mathfrak g}^{\alpha\beta}{\mathfrak g}^{\mu\nu}-{\mathfrak g}^{\alpha\nu}{\mathfrak g}^{\beta\mu}\right)
=16\pi G\,(-g)\,
t^{\alpha\beta}_{\rm LL}\,,
\end{equation}
where ${\mathfrak g}^{\mu\nu}=\sqrt{-g}\,g^{\mu\nu}$ and
the Landau-Lifshitz pseudo-tensor
$t^{\alpha\beta}_{\rm LL}$ does not contain
second derivatives. It is always possible to impose the harmonic gauge condition
$\partial_\mu {\mathfrak g}^{\mu\nu}=0$. Introducing the tensor ${ \Psi }^{\mu\nu}$ via
\begin{equation} \label{g-g}
\sqrt{-g}\,g^{\mu\nu}
=\eta^{\mu\nu}+{ \Psi }^{\mu\nu},
\end{equation}
Eqs.\eqref{Papa} assume the form
\begin{eqnarray} \label{5}
\partial^\sigma\partial_\sigma\,
{ \Psi }^{\mu\nu}&=&(\mbox{terms non-linear in ${ \Psi }^{\mu\nu}$}), \\
\partial_\sigma { \Psi }^{\sigma\nu}&=&0. \label{5a}
\end{eqnarray}
These equations can be viewed as describing gravitons
in Minkowski space. The non-linear terms on the right in
\eqref{5} describe graviton interactions.
In the linear approximation one neglects the interaction terms
and the equations describe free gravitons,
\begin{eqnarray} \label{7}
\partial^\sigma\partial_\sigma\, { \Psi }^{\mu\nu}&=&0, \\
\partial_\sigma { \Psi }^{\sigma\nu}&=&0. \label{7a}
\end{eqnarray}
One can then wonder if it is possible to go back from these linear equations
to the non-linear ones \eqref{5},\eqref{5a} and
apply the field theory methods to
reconstruct the interaction terms~? In other words, can one
obtain the General Relativity as the
non-linear completion for the theory of free gravitons, without relying on methods of
differential geometry~? Today we know that this is indeed possible
\cite{Deser:1969wk}, but in 1965
this fact was not known.
Therefore, the OP's aim was to apply the field theory methods to
construct non-linear terms to be added to the right hand side of \eqref{1a} to obtain
a consistent self-interacting theory.
Remarkably, they achieved the goal
and, starting from the very first principles, constructed a fully interacting theory
whose action contains the Einstein-Hilbert kinetic term and has also
a graviton mass term. Sending the graviton mass to zero they recovered the General Relativity.
Therefore, OP have been the first to obtain the General Relativity by applying only field-theory
methods, without using the differential geometry\footnote{
The well-known paper \cite{Deser:1969wk} of
Deser on a similar subject (considering only the massless case)
appeared a few years after the OP's work.
Deser used the bootstrap method, quite different from the OP's approach. }.
The central role in their construction is played by the
subsidiary conditions \eqref{2a} and \eqref{3a}.
However, OP had realized that it would have been technically too difficult
to keep both of them. Therefore they
imposed only one combined Hilbert-Lorentz condition,
\begin{equation} \label{4a}
\partial_\mu( { \Psi }^{\mu\nu}+q\,\eta^{\mu\nu} { \Psi }^\alpha_{~\alpha})=0,
\end{equation}
with constant $q$. This is necessary, although not sufficient, for exclusion of negative energies.
They called this condition ``spin limitation principle".
It excludes the spin-1 and a spin-0, but not the
second spin-0, hence there remain altogether 6 DoF.
OP required the formula \eqref{4a} to be
exactly the same also in the presence of interactions: it should always contain
partial and not covariant derivatives\footnote{This is indeed
possible in a bimetric theory.}.
Therefore, the spin limitation condition always remains ``clean"
and removes precisely the spin-1 and spin-0 in the strict representation theory sense.
This is probably the most important moment: OP keep control over the spin content
of their theory.
\textcolor{black}{
At this point, it is worth posing to compare the OP's strategy with the
logic commonly adopted at present, according to which
the ``healthy" massive gravity theory has to have 5 DoF
to avoid the Boulware-Deser (BD)
ghost \cite{Boulware:1973my}, hence it should contain 5 constraints \cite{deRham:2010kj}.
For a flat background\footnote{\textcolor{black}{More generally, for Einstein space backgrounds.}}
these constraints have the structure similar to that in Eqs.\eqref{2a},\eqref{3a} and they eliminate
precisely the spin-1 and two spin-0 states. However, for arbitrary backgrounds the fifth
constraint has a rather complex structure and it is not obvious what spin states it
eliminates\footnote{\textcolor{black}{Already the linearized version
of the 5-th constraint is very complex; see Appendix B in \cite{Mazuet:2018ysa}.
}}.
This suggests that for generic backgrounds the theory may propagate
superpositions of states of different spins,
even though the total number of DoF is always 5.
Therefore, the theory controls the number of DoF but does not seem to always control
their spin contents, which
might explain why it shows pathologies
for some backgrounds \cite{DeFelice:2012mx,Fasiello:2013woa,Chamseddine:2013lid}.
}
The OP's strategy was quite different. They constructed a theory with 6 DoF and did not care
about the BD mode (the ghost problem was not known at the time).
Instead, they preferred to have
control over the spin contents of their theory -- it
contains only the spin-2 and a spin-0, while spin-1 states are definitely excluded.
Whether or not this makes sense is to be understood.
Getting back to their construction,
OP required the condition \eqref{4a} to be
a differential consequence of the second order field equations.
This requirement lead to certain identity relations for the Lagrangian,
implying the existence of a local internal symmetry. By analyzing the structure of the
symmetry generators, OP concluded that the symmetry must formally coincide with the
spacetime diffeomorphism symmetry, viewed in their approach
as the {\it internal symmetry }
acting on gravitons in flat space. OP were then able to construct
the interaction terms order by order by requiring that the symmetry algebra closes.
They ended up with a theory whose
kinetic term coincides with the standard Einstein-Hilbert term for the
``effective" metric $g_{\mu\nu}$ whose inverse $g^{\mu\nu}$ is
related to the graviton field ${ \Psi }^{\mu\nu}$ via the relation similar to \eqref{g-g},
\begin{equation} \label{g-g1}
\left(\frac{\sqrt{-g}}{\sqrt{-\eta}}\right)^{s+1} ((\hat{g}^{-1})^n)^{\mu\nu}=\eta^{\mu\nu}+{ \Psi }^{\mu\nu}.
\end{equation}
Here the parameters $s,n$ are {\it real} and the precise meaning of the
matrix power will be specified below.
Therefore, the spacetime metric
$g_{\mu\nu}$ arises in their approach as a secondary object related to
the primary graviton field ${ \Psi }^{\mu\nu}$ in a very
non-linear way, via \eqref{g-g1}. Notice that this transformation is invertible
and can be resolved with respect to $g_{\mu\nu}$.
The OP action contains also a mass term constructed
from $g^{\mu\nu}$ and $\eta_{\mu\nu}$.
Once the OP theory is obtained, it can be formulated
entirely in terms of $g_{\mu\nu}$ and $\eta_{\mu\nu}$, and then it can be viewed
simply as a bimetric theory. It implies certain on-shell conservation conditions which,
when expressed in terms of the variables ${ \Psi }^{\mu\nu}$ defined by Eq.\eqref{g-g1},
assume the form of the linear
``spin limitation principle".
All of this will be explained below.
Summarizing, there are two aspects of the OP's work.
First, it presents the first systematic derivation of the Einstein-Hilbert
kinetic term starting from the free theory and applying
only the field theory principles. This is, of course,
a remarkable achievement for which OP should be fully credited, in our opinion.
Secondly, their procedure gives also a particular mass term, but the status of this
is less clear, since it gives rise to
6 DoF -- a property considered today
as unacceptable.
At the same time, the OP mass term is a part of the very carefully designed
derivation procedure.
For a one-dimensional subset of the parameter space it shows
the FP property and propagates only 5 DoF
around flat space. For one particular point of this FP subset the
theory propagates 5 DoF even at the non-linear level and
coincides with one of the ghost-free dRGT\footnote{dRGT
-- after the names of authors of
\cite{deRham:2010kj}.}
models. Therefore, the OP procedure gave in 1965
the result that was rediscovered again only in 2010 !
All of this suggests that the OP massive gravity deserves studying, even though
it propagates in general 6 DoF. Therefore, we present in what follows our analysis
of this theory and of its solutions. Skipping its derivation indicated above and described
in the OP's paper \cite{OP}, we come directly to the theory itself.
In modern terms, this is a bimetric theory\footnote{\textcolor{black}{A bimetric theory is any theory
with two metrics. It can be a massive gravity if only one of the
metrics is dynamical, or a bigravity if both metrics are dynamical. }
} containing the dynamical metric
$g_{\mu\nu}$ and a non-dynamical reference metric $f_{\mu\nu}$, with a specially designed
interaction potential constructed from these metrics.
In Section II, we rewrite the theory in modern notation and explain
how the linear ``spin limitation condition" follows from the non-linear field equations.
In brief, this is simply the condition for the tensor obtained by varying
the action with respect to $f_{\mu\nu}$. This tensor is conserved on-shell, which is true in any
bimetric theory, but only in the OP theory the conservation condition can be made linear
by the non-linear field redefinition \eqref{g-g1}.
We then study in Section III the simplest
solutions, such as the de Sitter or Minkowski,
and explicitly show that there are 6 propagating DoF,
unless for a one-dimensional subset of the parameter space for which there are
only 5 DoF. In Section IV we obtain the effective
action for fluctuations and, surprisingly, find that the de Sitter space
is completely free of ghosts and gradient instabilities for a large region of the parameter
space, although the flat space always
shows ghost away from the FP limit.
This is, perhaps, our most interesting finding -- the fact that the 6-th
polarization can be totally harmless.
We then proceed to study in Section V other
homogeneous and isotropic cosmologies in the theory. We find many different types
of such solutions, but unfortunately most of
them are unstable. At the same time, it turns out that the Milne space --
a sector of Minkowski space -- is stable in the UV limit. This suggests that the ghost
instability of the flat space develops only for long waves, similarly
to the classical Jeans instability. Our conclusions are formulated in Section VI,
while the Appendix contains the derivation of the stability conditions for the
homogeneous and isotropic cosmologies.
Few words about the impact of the OP's work. In the older days it
was mentioned in the massive gravity context \cite{vanDam:1970vg,Vainshtein:1972sx,Zeldovich:1986ft}.
The important special case in which the theory propagates only 5 DoF and coincides
with the dRGT theory was studied by Maheshwari in 1972 \cite{Maheshwari:1972mb}
(see \cite{Pitts:2015aqa} for an interesting historical account).
Nowadays it is cited by experts in various field theory domains (see for example
\cite{Ivanov:1981wn,Cutler:1986dv,Tomboulis:1996cy,Boulanger:2000rq,%
Zinoviev:2006im,Morris:2018mhd}), but it is almost totally unknown
to the modern massive gravity community. This has given us the motivation for writing this text.
\section{The OP theory}
\setcounter{equation}{0}
The OP theory is a particular case of
bimetric massive gravity.
Any such theory is described by the dynamical metric
$g_{\mu\nu}$ and a non-dynamical reference metric $f_{\mu\nu}$. There is no
general rule for choosing the latter, for example one can set it to be the Minkowski metric,
$f_{\mu\nu}=\eta_{\mu\nu}$,
but one can just as well leave it unspecified for the time being.
The action of the theory is
\begin{equation} \label{act}
S=M_{\rm Pl}^2\int\left(\frac12\, R(g)-m^2 U +{\cal L}_{\rm matter} \right)\sqrt{-g}\, d^4x\,,
\end{equation}
where $m$ is a mass parameter and
$U$ is a scalar function constructed from products of $f_{\mu\nu}$ with
the inverse of the physical metric, $g^{\mu\nu}$.
Introducing the matrix $\hat{S}$ with components
\begin{equation} \label{Sh}
\textcolor{black}{ (\hat{S})^\mu_{~\nu}\equiv S^\mu_{~\nu}=g^{\mu\sigma}f_{\sigma\nu} }
\end{equation}
and using brackets to denote trace, $[\hat{S}]=S^\sigma_{~\sigma}$, the potential
can be any function of traces of powers of $\hat{S}$,
\begin{equation}
U=U\left([\hat{S}],[\hat{S}^2],[\hat{S}^3],\det(\hat{S})\right).
\end{equation}
For example, it can be given by the series
\begin{equation}
U=a_0+a_1\,[\hat{S}]+a_2\,[\hat{S}^2]+a_3\,[\hat{S}]^2+\ldots,
\end{equation}
where $a_k$ are constant coefficients.
This theory generically shows
6 DoF in the gravity
sector. If $a_2+a_3=0$ then
the theory is said to fulfill the FP property and
shows only 5 DoF around flat space.
However, even then an extra 6-th polarization
emerges when one deviates from flat space,
\textcolor{black}{unless the potential is further fine-tuned}.
This extra mode carries a negative
kinetic energy and is called BD ghost \cite{Boulware:1973my}.
As a result, there exist infinitely many massive gravity theories corresponding
to infinitely many possibilities to choose the potential $U$. One therefore needs
a guiding principle to select one particular theory. An example of this is provided
by the dRGT potential selected by the requirement that the theory should always propagate 5 DoF,
for any backgrounds \cite{deRham:2010kj}. This potential is expressed in terms
of fractional powers of $\hat{S}$,
\begin{equation} \label{dRGT}
{U}=\beta_0+\sum_{n=1}^3 \beta_k\,{U}_k\,,
\end{equation}
where $\beta_0,\beta_k$ are real parameters and
\begin{eqnarray} \label{4}
{U}_1&=&[\hat{\gamma}],~~~~~
{U}_2
=\frac{1}{2!}([\hat{\gamma}]^2-[\hat{\gamma}^2]),~~~~~
{U}_3=
\frac{1}{3!}([\hat{\gamma}]^3-3[\hat{\gamma}][\hat{\gamma}^2]+2[\hat{\gamma}^3]),
\end{eqnarray}
the matrix $\hat{\gamma}$ being determined by the condition $\hat{\gamma}^2=\hat{S}$
hence
$\hat{\gamma}=\hat{S}^{1/2}$.
The OP potential is selected by a different requirement: the theory should imply the linear
Hilbert-Lorentz condition as a consequence of the field equations.
This potential contains arbitrary real powers of $\hat{S}$.
Splitting the inverse metric as
\begin{equation} \label{split}
g^{\mu\nu}=f^{\mu\nu}+\chi^{\mu\nu}\,,
\end{equation}
one has
\begin{equation} \label{Shh}
S^\mu_{~\nu}=g^{\mu\sigma}f_{\sigma\nu}=\delta^\mu_\nu+\chi^{\mu\sigma}f_{\sigma\nu}\equiv
\delta^\mu_\nu+\chi^\mu_{~\nu}\,,
\end{equation}
which can be written as\footnote{
\textcolor{black}{We use the hat is used to denote matrices, e.g. $\hat{S}$, the matrix
components being denoted either without hat, $S^\mu_{~\nu}$, or as $(\hat{S})^\mu_{~\nu}$.}
}
\begin{equation}
\hat{S}=\hat{g}^{-1}\hat{f}\equiv \hat{1}+\hat{\chi}\,.
\end{equation}
An arbitrary real power of $\hat{S}$ is defined via the series,
\begin{equation} \label{Sigma}
\hat{\Sigma}\equiv
\hat{S}^n=\hat{1}+n\hat{\chi}+\frac{n(n-1)}{2}\,\hat{\chi}^2+\ldots
\end{equation}
Introducing the scalar
\begin{equation} \label{1}
\phi=\frac{1}{\sqrt{{\det} ({\hat S})}}=
\left(
1+[\hat{\chi}]+\frac12( [\hat{\chi}]^2-[\hat{\chi^2}])+\ldots
\right)^{-1/2}\,,
\end{equation}
the OP mass term is given by
\begin{eqnarray} \label{pot}
U&=&{\cal U} +\lambda_g \, ,
\end{eqnarray}
where the $\hat{S}$-dependent part is
\begin{equation} \label{U}
{\cal U}=\frac{1}{4n^2}\,\phi^s\, [\hat{S}^n]=\frac{1}{4n^2}\left(\det({ \hat S})\right)^{-s/2}[\hat{S}^n],
\end{equation}
while the constant part is
\begin{equation}
\lambda_g=\frac{n-2s-2}{2n^2}.
\end{equation}
The potential depends on two real parameters
$n,s$ (OP use instead $n,p=-(s+1)/n$).
The OP theory propagates 6 DoF for generic values of $n$ and $s$.
If the parameters belong to the ellipse shown in Fig.\ref{Fig1} below,
then the theory has the FP property and propagates only 5 DoF around flat space.
For the particular point at the ellipse corresponding to
\begin{equation}
s=0,~~~n=\frac12,
\end{equation}
the OP potential \eqref{pot} coincides with
the dRGT potential \eqref{dRGT} for
\begin{equation}
\beta_0=-3, ~~~\beta_1=1,~~~ \beta_2=\beta_3=0.
\end{equation}
The OP theory becomes ``ghost-free" in this case, in the sense
that it propagates only 5 DoF for generic backgrounds.
The matter term in the action \eqref{act} can be arbitrary, but we shall be considering
just the (boldfaced) cosmological constant,
\begin{equation} \label{mat}
{\cal L}_{\rm matter}=-\bm{\Lambda}.
\end{equation}
\subsection{Field equations}
Let us vary the two metrics, $g_{\mu\nu}\to g_{\mu\nu}+\delta g_{\mu\nu}$ and
$f_{\mu\nu}\to f_{\mu\nu}+\delta f_{\mu\nu}$ (the metric $f_{\mu\nu}$
can be varied even though it is non-dynamical,
\textcolor{black}{in order to obtain identities similar to the Bianchi identity}).
The potential \eqref{pot} then receives the variation
\begin{equation} \label{dU}
\delta\,\left({U}\sqrt{-g}\right)=\frac12\left(\sqrt{-g}\, {X}^{\mu\nu}\,
\delta g_{\mu\nu}-\sqrt{-f}\, {Y}^{\mu\nu}\, \delta f_{\mu\nu}\right),
\end{equation}
where ${X}^{\mu\nu}={X}^\mu_{~\alpha} g^{\alpha\nu}$ and ${Y}^{\mu\nu}
={Y}_{~\alpha}^{\mu} f^{\alpha\nu}$
with
\begin{eqnarray} \label{XY}
X^\mu_{~\nu}&=&\frac{1}{2n^2}\,\phi^s\left(
n\, \Sigma^\mu_{~\nu}
-\frac{s+1}{2}\,\Sigma^\alpha_{~\alpha}\,\delta^\mu_{~\nu}
\right), \nonumber \\
Y^\mu_{~\nu}&=&\frac{1}{2n^2}\,\phi^{s+1}\left(
n\, \Sigma^\mu_{~\nu}
-\frac{s}{2}\,\Sigma^\alpha_{~\alpha}\,\delta^\mu_{~\nu}
\right),
\end{eqnarray}
where
$
\Sigma^\mu_{~\nu}=(\hat{S}^n)^\mu_{~\nu}\,.
$
Consider an infinitesimal diffeomorphism generated by a vector field $\xi^\mu$.
It induces the variations of both metrics,
\begin{equation}
\delta g_{\mu\nu}=\overset{(g)}{\nabla}_{(\mu} \overset{(g)}{\xi}_{\nu)},~~~~~~~~~~~
\delta f_{\mu\nu}=\overset{(f)}{\nabla}_{(\mu} \overset{(f)}{\xi}_{\nu)},
\end{equation}
where $\overset{(g)}{\xi}_\mu=g_{\mu\sigma}\,\xi^\sigma$ and
$\overset{(f)}{\xi}_\mu=f_{\mu\sigma}\,\xi^\sigma$ while $\overset{(g)}{\nabla}$ and
$\overset{(f)}{\nabla}$ are the covariant derivatives with respect to the g-metric and
f-metric, respectively.
Inserting this to \eqref{dU}, integrating over the manifold, dropping the boundary term and using
the fact that $U$ is a scalar and hence its integral
does not change under diffeomorphisms, gives the identity
\begin{equation} \label{id}
\sqrt{-g}\, \overset{(g)}{\nabla}_\mu X^\mu_{~\nu}=\sqrt{-f}\, \overset{(f)}{\nabla}_\mu Y^\mu_{~\nu}\,.
\end{equation}
Let us now vary the whole action \textcolor{black}{only} with respect to $g_{\mu\nu}$.
Setting the variation to zero, yields the equations
\begin{equation} \label{eq}
G^\mu_{~\nu}+m^2\,\lambda_g\,\delta^\mu_{~\nu}=m^2 X^\mu_{~\nu}+
T^{\rm (m)\mu}_{~~~~~\nu}\,,
\end{equation}
where
$T^{\rm (m)}_{~\mu\nu}$ is obtained by varying ${\cal L}_{\rm matter}$.
If the latter is given by \eqref{mat} then
\begin{equation}
T^{\rm (m)\mu}_{~~~~~\nu}=-\bm{\Lambda}\,\delta^\mu_{~\nu}.
\end{equation}
\subsection{Subsidiary conditions}
In view of the Bianchi identities
$\overset{(g)}{\nabla}_\mu G^{\mu}_{~\nu}=0$
and owing to the matter conservation condition
$\overset{(g)}{\nabla}_\mu T^{{\rm (m)}\mu}_{~~~~~\nu}=0$,
equations \eqref{eq} imply that
\begin{equation} \label{consX}
\overset{(g)}{\nabla}_\mu X^\mu_{~\nu}=0.
\end{equation}
This in turn implies, in view of the identity \eqref{id}, that
\begin{equation} \label{consY}
\overset{(f)}{\nabla}_\mu Y^\mu_{~\nu}=0.
\end{equation}
Now, according to \eqref{Shh}, one has $S^\mu_{~\nu}=\delta^\mu_{~\nu}+\chi^\mu_{~\nu}$,
where $\chi^\mu_{~\nu}$ vanishes if $g_{\mu\nu}=f_{\mu\nu}$.
The weighted power of this matrix can be represented similarly,
\begin{equation} \label{prim}
\phi^{s+1}\,(S^n)^\mu_{~\nu}=
\delta^\mu_{~\nu}+\Psi^\mu_{~\nu},
\end{equation}
where $\Psi^\mu_{~\nu}$ vanishes when the two metrics coincide.
This relation is in fact equivalent to the one in \eqref{g-g1},
assuming that the indices are moved with the f-metric so that
\begin{equation}
((\hat{g}^{-1})^n)^{\mu\nu}=(\hat{S}^n)^\mu_{~\sigma}\,f^{\sigma\nu},~~~~~~
{ \Psi }^{\mu\nu}={ \Psi }^\mu_{~\sigma}\,f^{\sigma\nu}.
\end{equation}
Using \eqref{prim} reduces \eqref{consY} to
\begin{equation} \label{psi0}
\overset{(f)}{\nabla}_\mu\left( \Psi^\mu_{~\nu}
-\frac{s}{2n}\,(\Psi^\alpha_{~\alpha})\,\delta^\mu_{~\nu}\right)=0.
\end{equation}
If the reference metric is chosen to be flat Minkowski,
$f_{\mu\nu}=\eta_{\mu\nu}$, then the derivatives $\overset{(f)}{\nabla}_\mu$ become
ordinary partial derivatives and \eqref{psi0} reduces to the Lorentz-Hilbert condition
for the $\Psi$-field,
\begin{equation} \label{psi}
\partial_\mu \Psi^\mu_{~\nu}+q\,\partial_\nu \Psi^\alpha_{~\alpha}=0,
\end{equation}
with $q=-s/(2n)$. This explains the OP's trick -- the linear in
$ \Psi^\mu_{~\nu}$ subsidiary condition \eqref{psi}
indeed follows from the non-linear field equations \eqref{eq}.
This explains also why this condition is not manifestly covariant -- the field equations
\eqref{eq} are covariant if only both metrics are allowed to simultaneously transform,
but the covariance is lost as soon as the metric $f_{\mu\nu}$ is fixed (unitary gauge).
Even though the OP theory is formulated entirely in terms of the metrics $g_{\mu\nu}$
and $f_{\mu\nu}$, according to the OP's philosophy, the metric $g_{\mu\nu}$
is only a secondary object. The primary object is supposed to be the graviton field
$ \Psi^\mu_{~\nu}$ determining the metric $g_{\mu\nu}$ via \eqref{prim}.
Setting the philosophy aside, the mathematical statement is that in the OP theory
there exists the {invertible} non-linear transformation \eqref{prim}
expressing the metric $g_{\mu\nu}$
in terms of ${ \Psi }^{\mu\nu}$
such that the condition $\partial_\mu Y^\mu_{~\nu}=0$ becomes {\it linear} in
${ \Psi }^{\mu\nu}$.
One should stress at the same time that the tensor $Y^\mu_{~\nu}$ can be defined
via \eqref{dU} in any bimetric theory,
for any choice of the mass term $U$.
The condition \eqref{consY} will always hold on-shell, and
setting
$f_{\mu\nu}=\eta_{\mu\nu}$ one always obtains $\partial_\mu Y^\mu_{~\nu}=0$.
However, $Y^\mu_{~\nu}$ will in general contain non-linear terms that cannot
be absorbed by redefining the variables \textcolor{black}{via \eqref{prim}}, hence conditions
$\partial_\mu Y^\mu_{~\nu}=0$ will not
have the Lorentz-Hilbert form needed for the ``spin limitation".
For example, in
the dRGT theory with the potential \eqref{dRGT}
one has
\begin{equation}
{Y}^\mu_{~\nu}=-\frac{1}{\det(\hat{\gamma } ) }\left\{
\left(\beta_1+\beta_2\, U_1+\beta_3\,U_2 \right)
{{\gamma}}^\mu_{~\nu}-(\beta_2+\beta_3\, U_1)\,(\hat{\gamma}^2)^\mu_{~\nu}
+\beta_3\, (\hat{\gamma}^3)^\mu_{~\nu}
\right\}.
\end{equation}
If $\beta_2=\beta_3=0$ then the conservation condition $\partial_\mu Y^\mu_{~\nu}=0$
reduces to
\begin{equation}
\partial_\mu\left(\frac{1}{\det(\hat{\gamma } )}\, {{\gamma}}^\mu_{~\nu}\right)=0,
\end{equation}
\textcolor{black}{which has the form \eqref{prim} and
can be linearized by setting}
\begin{equation}
\frac{1}{\det(\hat{\gamma } )}\, {{\gamma}}^\mu_{~\nu}={\delta}^\mu_{~\nu}+{{ \Psi }}^\mu_{~\nu}\,,
\end{equation}
which yields $\partial_\mu{{ \Psi }}^\mu_{~\nu}=0$. \textcolor{black}{However, the same trick does not work
for generic values of $\beta_k$. }
For example, if $\beta_1=\beta_3=0$ and $\beta_2\neq 0$ then one obtains
\begin{equation}
\partial_\mu\left(\frac{1}{\det(\hat{\gamma } )}\,\left([\hat{\gamma}] {{\gamma}}^\mu_{~\nu}
-(\hat{\gamma}^2)^\mu_{~\nu}\right)\right)=0,
\end{equation}
\textcolor{black}{
which cannot be linearized by applying \eqref{prim}. Of course, this can be linearized
by a different transformation. However, within the OP approach, the transformation should be
the same as the one OP used to derive the Einstein-Hilbert kinetic term starting from the
linear theory, hence it must have the form \eqref{prim}, which is equivalent to \eqref{g-g1}.
}
Summarizing, the OP potential is adjusted in such a way that the
tensor $Y^\mu_{~\nu}$ has the structure
$Y^\mu_{~\nu}=D^\mu_{~\nu}+const.\times [\hat{D}]\,\delta^\mu_{~\nu}$
where $D^\mu_{~\nu}$ is the weighted power of $g_{\mu\nu}$.
Changing the variables via
$D^\mu_{~\nu}=\delta ^\mu_{~\nu}+{ \Psi }^\mu_{~\nu}$, the one-shell condition
$\partial_\mu Y^\mu_{~\nu}=0$ assumes the linear form
$\partial_\mu \Psi^\mu_{~\nu}+q\,\partial_\nu \Psi^\alpha_{~\alpha}=0$.
According to OP, this property is very important, since the linear
condition for ${ \Psi }^{\mu\nu}$ restricts the spin of states
in the theory. \textcolor{black}{To understand what this property gives in practical terms,
we shall now study the phenomenology of the theory. }
\section{Solutions with proportional metrics}
\setcounter{equation}{0}
We shall first consider the simplest solutions for which the reference metric $f_{\mu\nu}$
is not fixed once and forever
but related to the physical metric via
\begin{equation} \label{prop}
f_{\mu\nu}=\xi^2\, g_{\mu\nu}
\end{equation}
with constant $\xi$. This implies that
\begin{equation} \label{XX}
S^\mu_{~\nu}=\xi^2 \delta^\mu_{~\nu}~~~~~\Rightarrow~~~~~~
X^\mu_{~\nu}=\lambda_g\,\xi^{2(n-2s)} \delta^\mu_{~\nu}\,,
\end{equation}
and the field equations \eqref{eq} reduce to
\begin{equation} \label{ee}
G_{\mu\nu}+\Lambda\,g_{\mu\nu}=0
\end{equation}
with
\begin{equation} \label{Lam}
\Lambda=\bm{\Lambda}+m^2\lambda_g\left[1-\xi^{2(n-2s) } \right].
\end{equation}
It follows that the solution is an Einstein space with $R_{\mu\nu}=\Lambda\, g_{\mu\nu}$,
for example the de Sitter space.
If the parameters are adjusted such that $\Lambda=0$,
then the flat space will be a solution. We emphasize once again that to
different solutions there correspond different reference metrics defined via \eqref{prop}.
Let us now analyze the stability of such solutions. To this end, we consider
small perturbations of the g-metric without changing the reference metric,
\begin{equation}
g_{\mu\nu}\to g_{\mu\nu}+{ h }_{\mu\nu},~~~~~~~~f_{\mu\nu}\to f_{\mu\nu},
\end{equation}
hence
\begin{equation}
g^{\mu\nu}\to g^{\mu\nu}-{ h }^{\mu\nu}+\ldots \,,
\end{equation}
with
${ h }^{\mu\nu}=g^{\mu\alpha}g^{\nu\beta} { h }_{\alpha\beta}$.
Let us consider the linear in perturbations part of the field equations,
\begin{equation} \label{pert}
\delta G_{\mu\nu}+(\bm{\Lambda}+m^2\lambda_g)\,\delta g_{\mu\nu}=m^2 \delta X_{\mu\nu}\,,
\end{equation}
where
\begin{equation}
\delta X_{\mu\nu}=\delta\left( g_{\mu\sigma}X^\sigma_{~\nu} \right)=h_{\mu\sigma}\,X^\sigma_{~\nu}
+g_{\mu\sigma}\,\delta X^\sigma_{~\nu}=\lambda_g\,\xi^{2(n-2s)}\,h_{\mu\nu}+g_{\mu\sigma}\,\delta X^\sigma_{~\nu}.
\end{equation}
To calculate the variation $\delta X^\sigma_{~\nu}$, we notice that the tensor $X^\mu_{~\nu}$ defined by \eqref{XY}
contains powers of the tensor $S^\mu_{~\nu}=g^{\mu\sigma}f_{\sigma\nu}$, whose variation is
\begin{equation}
\delta S^\mu_{~\nu}=\delta g^{\mu\sigma} f_{\sigma\nu}=-h^{\mu\sigma}f_{\sigma\nu}
=-{ h }^\mu_{~\sigma}S^\sigma_{~\nu}=
-\xi^2 { h }^\mu_{~\nu}\,.
\end{equation}
The background tensor $S^\mu_{~\nu}=\xi^2\delta^\mu_\nu$ is proportional to the unit tensor,
hence it
commutes with $\delta S^\mu_{~\nu}$, which implies that
\begin{equation}
\delta{\Sigma}^\mu_{~\nu}=\delta ({S}^n)^\mu_{~\nu}
=n\,\delta {S}^\mu_{~\sigma}\, (S^{n-1})^\sigma_{~\nu}=-n{h}^\mu_{~\sigma}\,(S^n)^\sigma_{~\nu}
=-n\xi^{2n}\,h^\mu_{~\nu}.
\end{equation}
It follows also that
\begin{equation}
\delta(\det\hat{S})=\delta (\det{\hat{g}}^{-1})\det\hat{f}=-\det(\hat{S})\, h
\end{equation}
with $h=h^\mu_{~\mu}$
and hence
\begin{equation}
\delta(\phi^s)=\delta\left(\det \hat{S})^{-s/2}\right)=\frac{s}{2}\,\phi^s\,h=\frac{s}{2}\,\xi^{-4s}h.
\end{equation}
As a result,
\begin{eqnarray}
\delta X^\mu_{~\nu}
&=&\frac{s}{2}\,h\,X^\mu_{~\nu}-\frac{1}{2}\,\xi^{2n-4s}
\left(n h^\mu_{~\nu}-\frac{s+1}{2}\,h\delta^\mu_{~\nu}\right) \nonumber \\
&=&\xi^{2n-4s}\left(
-\frac12 \,h^\mu_{~\nu}+\left[\frac{s}{2}\,\lambda_g+\frac{s+1}{4n}\right]h\delta^\mu_{~\nu}
\right),
\end{eqnarray}
and the perturbation equations \eqref{pert} assume the form
\begin{equation} \label{HIG}
E_{\mu\nu}\equiv \delta G_{\mu\nu}+\Lambda\,h_{\mu\nu}+\frac{M^2}{2}\,
(h_{\mu\nu}-\lambda h\, g_{\mu\nu})=0
\end{equation}
with $\Lambda$ defined by \eqref{Lam} and with
\begin{equation}
M^2=m^2 \xi^{2n-4s}
\end{equation}
while
\begin{equation} \label{zeta}
\lambda=1+\frac{1-(n-2s-1)^2-3n^2}{4n^2}\equiv 1+\zeta.
\end{equation}
For $\lambda=1$ Eqs.\eqref{HIG} reduce to those studied by Higuchi
to describe massive gravitons in de Sitter space \cite{Higuchi},
the parameter $M$ then determines the
graviton mass. The equations show the FP property in this case --
they propagate only 5 DoF.
However, for $\lambda\neq 1$ the number
of DoF is 6. Let us remind the corresponding counting argument.
There are 10 equations in \eqref{HIG}, where one has
\begin{eqnarray}
\delta G_{\mu\nu}=&&\left.\frac12\left(
\nabla^\sigma \nabla_\mu { h }_{\sigma\nu}
+\nabla^\sigma \nabla_\nu { h }_{\sigma\mu}-\Box { h }_{\mu\nu}-\nabla_\mu \nabla_\nu h-R\,{ h }_{\mu\nu}\right)\right.
\nonumber \\
&&+\left.\frac12 \,
g_{\mu\nu}\left( \Box h-\nabla^\alpha\nabla^\beta h_{\alpha\beta} +R^{\alpha\beta}h_{\alpha\beta} \right)\right..
\end{eqnarray}
For $R_{\mu\nu}=\Lambda\,g_{\mu\nu}$ and $R=4\Lambda$ there is the
identity relation
$
\nabla^\mu (\delta G_{\mu\nu}+\Lambda h_{\mu\nu})=0,
$
hence taking the divergence of \eqref{HIG} yields four constraints
\begin{equation} \label{FP1}
\frac{M^2}{2}\left(\nabla^\mu { h }_{\mu\nu}-\lambda \nabla_\nu h\right)=0
\end{equation}
from \textcolor{black}{which} $\nabla^\mu { h }_{\mu\nu}=\lambda \nabla_\nu h$. Using these relations,
Eqs.\eqref{HIG} reduce to
\begin{eqnarray}
-\Box h_{\mu\nu}+(2\lambda-1)\nabla_{\mu\nu}h
-2R_{\mu\alpha\nu\beta}\,h^{\alpha\beta}
&+&\left[(1-\lambda)\Box h+\Lambda h \right] g_{\mu\nu} \nonumber \\
&+&M^2\left[h_{\mu\nu}-\lambda h\, g_{\mu\nu} \right]=0,
\end{eqnarray}
and taking the trace one obtains
\begin{equation} \label{FP2}
2(1-\lambda)\Box h+\left[2\Lambda+(1-4\lambda)M^2 \right]h=0.
\end{equation}
If $\lambda=1$ then this yields
\begin{equation}
\left(2\Lambda-3M^2 \right)h=0,
\end{equation}
implying the fifth constraint, $h=0$ (unless in the partially massless limit $2\Lambda=3M^2$).
Therefore, the number of DoF is the number of components of $h_{\mu\nu}$
minus the number of constraints in \eqref{FP1},\eqref{FP2}, which gives
$10-5=5$. This corresponds to the Fierz-Pauli theory.
If $\lambda\neq 1$ then the FP property is lost, because
there is the non-trivial kinetic term in \eqref{FP2},
hence the trace $h$ becomes a dynamical mode, so that
there are 6 DoF.
\begin{figure}[th]
\hbox to \linewidth{ \hss
\resizebox{12cm}{9cm}{\includegraphics{fg_OP1.pdf}}
\hss}
\caption{The FP subset of the OP theory.}
\label{Fig1}
\end{figure}
Therefore, the OP theory respects
the FP property if the parameter $\lambda$
defined by \eqref{zeta} is equal to one, hence if
\begin{equation}
(n-2s-1)^2+3n^2=1.
\end{equation}
This defines an ellipse in the $ns$-plane shown in Fig.\ref{Fig1}.
Points of this ellipse correspond to the special case of the OP theory in which the FP property is respected
and there are only 5 DoF
around the de Sitter (or flat for $\Lambda=0$) space.
Points not belonging to the ellipse correspond to the generic OP
theory with 6 DoF.
If $\Lambda=0$ and the background geometry is flat, then
among the 6 DoF there is ghost --
a mode with a negative kinetic energy.
One might think that all of the OP theories with 6 DoF are unphysical, since flat space is unstable
in such theories.
However, as we shall now see,
the whole interior of the ellipse corresponds to theories in which the
de Sitter space is stable \textcolor{black}{at the level of liner perturbations}.
\section{Stability conditions}
\setcounter{equation}{0}
Let us assume the background metric $g_{\mu\nu}$
to be of the Friedmann-Lema$\hat{\i}$tre-Robertson-Walker (FLRW) type,
\begin{equation}
ds_g^2=-N^2(t)\,dt^2+a^2(t)\, \delta_{ij}\, dx^i dx^j=\frac{1}{\xi^2}\, ds_f^2,
\end{equation}
which fulfills background equations \eqref{ee} with $\Lambda\equiv 3H^2>0$,
\begin{equation}
\frac{\dot{a}^2}{N^2 a^2}=H^2.
\end{equation}
This describes the de Sitter space expressed in the spatially flat slicing.
Perturbing the solution,
\begin{equation} \label{pert0}
g_{\mu\nu}\to g_{\mu\nu}+h_{\mu\nu},~~~~~~~f_{\mu\nu}\to f_{\mu\nu},
\end{equation}
the perturbations can be decomposed into the
scalar, vector, and tensor parts via
\begin{eqnarray} \label{h}
h_{00}&=&-N^2 {\bf S}_3, \nonumber \\
h_{0i}&=&Na \left( \partial_i {\bf S}_4+{\bf W}_i \right), \nonumber \\
h_{ik}&=&a^2\left(
{\bf S}_1\,\delta_{ik}+\partial^2_{ik}{\bf S}_2+\partial_i {\bf V}_k+\partial_k {\bf V}_i+{\bf D} _{ik}
\right),
\end{eqnarray}
where
\begin{equation}
\sum_k \partial_k {\bf V}_k=\sum_k \partial_k {\bf W}_k=0,~~~~~
\sum_k \partial_k {\bf D}_{ki}=0,~~~~\sum_k {\bf D}_{kk}=0.
\end{equation}
The spatial dependence of the modes is given by the plane waves
$\exp(i\,{\bf px})=\exp(ipz)$, where the wave vector
can be oriented along the 3-rd (z) axis.
The amplitudes ${\bf S}_4$ and ${\bf V}_k$ have dimension of length,
while ${\bf S}_2$ has dimension of length squared. To pass to dimensionless quantities,
we introduce a mass scale $\mu$ and set
\begin{equation}
{\bf S}_1=S_1(t) e^{ipz},~~~~~
{\bf S}_2=\frac{1}{\mu^2}\,S_2(t) e^{ipz},~~~~~
{\bf S}_3=S_3(t) e^{ipz},~~~~~
{\bf S}_4=\frac{1}{\mu}\,S_4(t) e^{ipz},~~~~~
\end{equation}
the vector amplitudes are chosen as
\begin{equation}
{\bf V}_k=\frac{1}{\mu}\,[V_{1}(t),V_{2}(t),0]\,e^{ipz},~~~~~~~
{\bf W}_k=[W_{1}(t),W_{2}(t),0]\,e^{ipz},
\end{equation}
while for the tensor modes the only non-trivial components of ${\bf D}_{ik}$ are
\begin{equation}
{\bf D}_{11}=-{\bf D}_{22}=D_{1}(t)\,e^{ipz},~~~~~
{\bf D}_{12}={\bf D}_{21}=D_{2}(t)\,e^{ipz}.
\end{equation}
The mass scale $\mu$ can be, for example, the Planck mass $M_{\rm Pl}$, or
the Hubble rate $H$, or
the graviton mass $M$.
However, we prefer not to specify it to be able to consider the limits such as
$H\to 0$
or $M\to 0$.
Inserting everything into the perturbation equations $E_{ik}=0$ \eqref{HIG}, they
split into three independent groups for the scalar, vector, and tensor modes. These equations
determine the effective action, which is the sum of three independent terms,
\begin{equation} \label{act0}
I=I_{\rm T}+I_{\rm V}+I_{\rm S}=\frac{M_{\rm Pl}^2 }{4} \int Na^3\, \bar{h}^{\mu\nu}E_{\mu\nu}\, dt\,d^3x,
\end{equation}
where the bar denotes complex conjugation.
One obtains in the tensor sector
\begin{equation} \label{IT}
I_{\rm T}=\frac{M_{\rm Pl}^2 }{4}\, \int Na^3\left(
\frac{1}{N^2}\left(\dot{D}_{1}^2+\dot{D}_{2}^2\right)-\left[M^2+\frac{p^2}{a^2} \right](D_{1}^2+D_{2}^2)
\right)dt\,d^3x.
\end{equation}
Inspecting the equations in the vector sector one finds that the two amplitudes $W_1$ and $W_2$ can be
expressed in therms of $V_1$ and $V_2$, the latter being governed by the action
\begin{equation} \label{IV}
I_{\rm V}=\frac{M_{\rm Pl}^2 }{4\mu^2}\,M^2 \int Na^3\left(
\frac{p^2}{N^2(p^2/a^2+M^2)}\left(\dot{V}_{1}^2+\dot{V}_{2}^2\right)-p^2(V_{1}^2+V_{2}^2)
\right)dt\,d^3x.
\end{equation}
Neither tensor nor vector modes are sensitive to the value of the parameter $\zeta$
describing the deviation from the FP limit. The vector modes become non-dynamical when the mass $M$
tends to zero.
In the scalar sector, the amplitudes $S_3$ and $S_4$ can be expressed in terms of $S_1$ and $S_2$,
the effective action for the latter being
\begin{equation} \label{IS}
I_{\rm S}^{(\zeta)}=\frac{M_{\rm Pl}^2 }{4}\, M^2 \int Na^3\sum_{a,b=1,2} \left(
\frac{1}{N^2}\,K_{ab}\,\dot{S}_a\dot{S}_b
+\frac{\cal Q}{N}\,\epsilon_{ab}\,\dot{S}_a S_b - U_{ab}\,S_a S_b
\right)dt\, d^3 x.
\end{equation}
Here the kinetic matrix has components
\begin{equation} \label{KK}
K_{11}=\zeta\,\frac{3a^2M^2+4p^2}{X},~~~
K_{22}=\frac{2a^2H^2p^4}{\mu^4\,X},~~~~
K_{12}=K_{21}=-\zeta\,\frac{a^2 M^2 p^2}{\mu^2 X},
\end{equation}
with
\begin{equation} \label{X}
X=8H^2p^2+M^2a^2(6H^2-\zeta M^2),
\end{equation}
and one has
\begin{equation} \label{QQ}
{\cal Q}=-2(1+2\zeta)\,\frac{H p^4}{\mu^2 X},~~~~~~~\epsilon_{ab}=-\epsilon_{ba},~~~\epsilon_{12}=1.
\end{equation}
Components of the potential matrix have more complex structure,
\begin{eqnarray}
U_{11}=\frac{1}{2a^2X^2}\{-32H^2\, p^6 +4a^2[48H^4-6M^2(2\zeta+5)H^2+\zeta M^4]\,p^4 \nonumber \\
+2M^2a^4 [144H^4+6(4\xi^2-11\xi-12)M^2H^2+\zeta(5\zeta+6)M^4]\,p^2 \nonumber \\
+3M^4a^6(6H^2-\zeta M^2)(6H^2-(3+4\zeta)M^2)\}, \nonumber \\
U_{22}=\frac{p^4}{2\mu^4X^2}\{-64\zeta H^4\, p^4
-8a^2H^2M^2[(2\zeta-10)H^2+(1+2\zeta-\zeta^2)M^2]\,p^2 \nonumber \\
+M^4a^4(6H^2-(1+2\xi)M^2)(6H^2-\zeta M^2)
\}, \nonumber \\
U_{12}=U_{21}=\frac{p^2}{2\mu^2X^2}\{16H^2[ (6\zeta-1)H^2+(1+\zeta)M^2]\,p^4 \nonumber \\
+2a^2M^2[6(2\zeta-7)H^4-(10\zeta^2-17\zeta-18)M^2H^2-\zeta(1+\zeta)M^4 ]\,p^2 \nonumber \\
-M^4a^4(6H^2-\zeta M^2)(6H^2-(3+4\zeta)M^2)
\}.
\end{eqnarray}
We notice that the scalars become non-dynamical in the $M\to 0$ limit.
Let us analyze the positivity of the kinetic matrix $K_{ab}$. This eigenvalues
$\lambda_1$ and $\lambda_2$ of this
matrix will be positive definite if the trace ${\rm tr}( K_{ab})=\lambda_1+\lambda_2$
and the determinant $\det (K_{ab})=\lambda_1\lambda_2$ are both positive. One has
\begin{eqnarray} \label{XK}
\mbox{tr}(K_{ab})&=&\frac{1}{\mu^4\,X}\,(2H^2a^2 p^4+4\zeta \mu^4\,p^2+3\zeta a^2 \mu^4 M^2)
~~~\stackrel{p\to\infty}{\longrightarrow}
~~~\frac{a^2 p^2}{4\mu^4} +{\cal O}(1), \nonumber \\
\det(K_{ab})&=&\frac{1}{\mu^4 X}\,\zeta a^2 p^4
~~~\stackrel{p\to\infty}{\longrightarrow}
~~~\frac{\zeta a^2}{8\mu^4H^2} \,p^2+{\cal O}(1).
\end{eqnarray}
Therefore, as long as $\zeta>0$,
the kinetic matrix is always positive-definite in the UV limit
$p\to\infty$, and it will be positive definite for {\it any} momenta if
\begin{equation}
6H^2>\zeta M^2
\end{equation}
since $X>0$ in this case.
This conclusion applies only when the Hubble rate of the background
metric is non-zero, $H\neq 0$.
If $H=0$ then the background metric
is flat and one can set $a=1$, hence \eqref{X} yields $X=-\zeta M^4$.
Eq.\eqref{XK} then gives
\begin{equation} \label{Mink}
\mbox{tr}(K_{ab})=-\frac{4p^2+3M^2 }{M^4}<0,~~~~~~
\det(K_{ab})=-\frac{p^4}{\mu^4 M^4}<0,
\end{equation}
hence one of the two eigenvalues $\lambda_1$ and $\lambda_2$ is always
negative.
As a result, there is ghost around flat space if $\zeta\neq 0$. This fact is of course well-known.
What is new is that the ghost becomes a benign mode in the de Sitter space
if $\zeta>0$.
One may wonder if there are \textcolor{black}{gradient instabilities} in the system.
The sound speed $C_{\rm S}$ is determined by the algebraic equation
\begin{equation} \label{speed}
\det\left(\frac{p^2C_{\rm S}^2}{a^2}\,K_{ab}+i\,\frac{p\,C_{\rm S}}{a}\,{\cal Q}\,\epsilon_{ab}-U_{ab} \right)=0.
\end{equation}
This gives two different values for $C_{\rm S}^2$ determining the speed of the
scalar component of the massive graviton and that of the 6-th mode.
Both have the same UV limit,
\begin{equation} \label{speed1}
\lim_{p\to\infty}C_{\rm S}^2=1.
\end{equation}
Since $C_{\rm S}^2>0$, there are no \textcolor{black}{gradient instabilities} in this limit.
This conclusion applies
both for the de Sitter space ($H\neq 0$) and for Minkowski space ($H=0$).
\subsection{The FP limit}
Let us finally see what happens when $\zeta\to 0$. The only
non-vanishing component of the kinetic matrix $K_{ab}$ in this limit is $K_{22}$,
hence the amplitude $S_1$ becomes non-dynamical and can be algebraically
expressed in terms of $S_2$. Injecting this expression back to the action yields
\begin{equation}
I_{\rm S}^{(\zeta=0)}=\frac{3M_{\rm Pl}^2}{4\mu^4}\,M^2(M^2-2H^2)\int Na^3 \left(
\frac{1}{N^2}\,K\dot{S}_2^2-US_2^2
\right)\,dt\, d^3x
\end{equation}
with
\begin{equation}
K=\frac{a^4p^4}{Y},~~~~U=a^2p^4\,
\frac{16H^2a^2p^4+(p^2+a^2M^2)Y}{Y^2},
\end{equation}
where $Y=4p^4+3a^2(M^2-2H^2)(4p^2+3M^2a^2)$. We see that the scalar mode becomes
non-dynamical either for $M=0$ (massless limit) or for $M^2=2H^2$ (partially
massless limit). The kinetic therm is always positive if $M^2>2H^2$ but for $M^2<2H^2$
it becomes negative and the scalar mode becomes the
(Higuchi) ghost. The speed of sound is equal to one in the UV limit.
Sending $H\to 0$, one can see that the flat space is stable in this case.
\subsection{Generic massive gravity }
The above results are actually quite general and
apply not only in the OP theory but also in the
generic bimetric theory \eqref{act}.
Specifically, introducing
$
H^\mu_{~\nu}=\delta^\mu_{~\nu}-S^\mu_{~\nu}
$
with $S^\mu_{~\nu}=g^{\mu\sigma}f_{\sigma\nu}$,
the potential $U$ in \eqref{act} can be expanded as
\begin{equation}
U=a_0+a_1[\hat{H}]+a_2[\hat{H}^2]+a_3[\hat{H}]^2+\ldots
\end{equation}
If the matter
term in \eqref{act} is given by \eqref{mat}, then the field equations read
\begin{equation} \label{dSE}
G_{\mu\nu}+{\bm\Lambda}g_{\mu\nu}+m^2(2a_1 f_{\mu\nu}+4a_2 f_{\mu\sigma}H^\sigma_{~\nu}
+4a_3[\hat{H}]f_{\mu\nu}+(a_0+a_1[\hat{H}]) g_{\mu\nu}+\ldots)=0.
\end{equation}
If $f_{\mu\nu}=g_{\mu\nu}$ then $H^\mu_{~\nu}=0$ and the equations reduce to
\begin{equation}
G_{\mu\nu}+\Lambda g_{\mu\nu}=0~~~~~~{\rm with}~~~~~~
\Lambda={\bf \Lambda}+m^2(a_0+2 a_1).
\end{equation}
Therefore, if the reference metric is chosen to be de Sitter,
then the theory admits a solution for which the physical metric is also de Sitter.
Consider perturbations
$
g_{\mu\nu}=f_{\mu\nu}\to f_{\mu\nu}+h_{\mu\nu}
$
with fixed $f_{\mu\nu}$. Linearizing Eqs.\eqref{dSE} with respect to
$h_{\mu\nu}$ then yields precisely the Higuchi equations \eqref{HIG} with
\begin{equation}
M^2=8 m^2 a_2,~~~~~~\lambda=1+\zeta=-\frac{a_3}{a_2}.
\end{equation}
Therefore, the above analysis directly applies and one can say at once that the
flat space (obtained if $\Lambda=0$) is stable
if $a_2+a_3=0$,
while the de Sitter space is stable if $\lambda>1$.
\section{More general cosmologies}
\setcounter{equation}{0}
Let us now study more general solutions of the OP theory.
We shall be considering FLRW cosmologies described by
\begin{eqnarray} \label{gf}
ds_g^2&=&-N^2(t)\,dt^2+a^2(t)\,\Omega_{ij}dx^i dx^j, \nonumber \\
ds_f^2&=&-N_f^2(t)\,dt^2+a_f^2(t)\, \Omega_{ij}dx^i dx^j,
\end{eqnarray}
where $\Omega_{ij}dx^i dx^j$ is the metric of the maximally symmetric 3-space
with constant curvature $K=0,\pm 1$. We can assume without loss of generality that the
functions $N,N_f,a,a_f$ are positive. The time reparameterization freedom
can be used to impose one gauge condition, for example, one can set $N=1$, or $N_f=1$, or $a=t$.
Denoting
\begin{equation} \label{xi}
\frac{a_f}{a}\equiv \xi,~~~~~~~
\frac{N_f}{N}\equiv c\,\xi\,,
\end{equation}
and injecting everything to \eqref{Sh} and \eqref{1} yields
\begin{equation}
S^\alpha_{~\beta}=\xi^2\,{\rm diag}[\,c^2,1,1,1],~~~~~~~~~
\phi=(c\xi^4)^{-1}.
\end{equation}
The function $c$ is the speed of light
measured with respect to the reference metric.
\subsection{Stability conditions }
Before we study solutions of the form \eqref{gf},
let us describe their stability conditions.
These conditions are derived in the Appendix\footnote{\textcolor{black}{Eqs. \eqref{zzz} and \eqref{ccc0} are
derived in the Appendix for the spatially flat $K=0$ background, but they are
valid also for the spatially open $K=1$ and closed $K=-1$ backgrounds.} },
without imposing the background field equations
and only assuming that $\dot{a}\neq 0$.
It turns out that backgrounds
\eqref{gf} may in general accommodate
ghost and \textcolor{black}{gradient instabilities}. Ghosts (excitations with a negative kinetic energy)
will be absent in the UV limit, for momenta
much larger than the Hubble parameter,
\begin{equation} \label{UV}
p\gg H=\frac{\dot{a}}{Na},
\end{equation}
if the following condition holds:
\begin{equation} \label{zzz}
\zeta(c)\equiv -\frac{(2n-s-1)(2n-s)c^{2n}+3s(s+1) }{4n^2}>0.
\end{equation}
However, soft ghosts with a wavelength
of the order or larger than the cosmological horizon may still be present.
Notice that if
$c\to 1$ then $\zeta(c)$ reduces to
$\zeta$ defined by \eqref{zeta}.
\textcolor{black}{The gradient instability} is
characterized by an imaginary
sound speed. The sound speed $C_{\rm S}$ is determined by the
algebraic equation (again assuming the UV limit \eqref{UV})
\begin{equation} \label{ccc0}
2C^2_{\rm V}\,\zeta(c)\,C_{\rm S}^4
+\left[C^4_{\rm V}\left(4Z\,\zeta(c)-\omega^2 \right)+c^2\right] C_{\rm S}^2
+2c^2C^2_{\rm V}Z=0,
\end{equation}
where
\begin{eqnarray}
\omega&=&-\frac{1+c^2}{2C^2_{\rm V}}+\frac{2s+1}{2n}\,(1+c^{2n})
-\frac{s(s+1)}{2n^2}\,(3+c^{2n}), \nonumber \\
Z&=&-\frac{1}{4n^2}\left[2n(2n-2s-1)+s(s+1)(3+c^{2n}) \right],
\nonumber \\
C^2_{\rm V}&=&n\,\frac{c^2-1}{c^{2n}-1}.
\end{eqnarray}
There are in general two different values of $C_{\rm S}^2$ which fulfill \eqref{ccc0},
hence two different sound speeds.
\textcolor{black}{The gradient instability} will be absent if
\begin{equation}
C_{\rm S}^2>0.
\end{equation}
In the $c\to 1$ limit Eq.\eqref{ccc0} reduces to
$
2\zeta(C_{\rm S}^2-1)^2=0,
$
in agreement with the previous result \eqref{speed1}.
\subsection{Simplest solutions}
Let us now consider the field equations for metrics \eqref{gf}.
Eq.\eqref{XY} yields the following non-zero components for the tensor $X^\mu_{~\nu}$:
\begin{eqnarray} \label{eee}
X^0_{~0}&=&\frac{1}{2n^2}\, c^{-s} \xi^{2n-4s}\left(n\, c^{2n}-\frac{s+1}{2}\left( c^{2n}+3\right)\right), \nonumber \\
X^1_{~1}=X^2_{~2}=X^3_{~3}&=&
\frac{1}{2n^2}\, c^{-s} \xi^{2n-4s}\left(n-\frac{s+1}{2}\left(c^{2n}+3\right)
\right).
\end{eqnarray}
The field equations \eqref{eq} then reduce to
\begin{eqnarray} \label{eqn}
\frac{3\dot{a}^2 }{N^2 a^2}+\frac{3K}{a^2}&=&m^2\lambda_g+\bm{\Lambda}-m^2 X^0_{~0}\,, \\
\frac{1}{N}\,\left(\frac{\dot{a}}{Na} \right)^{\bm\cdot}-\frac{K}{a^2}&=&\frac{m^2}{2}(X^0_0-X^1_1). \label{eqna}
\end{eqnarray}
These two equations plus an additional gauge condition
do not determine all four functions $N,N_f,a,a_f$ and one of them
remains free. This is the consequence of the fact
that the system is undetermined because the reference metric is not yet completely
specified. An extra assumption is needed
in order to determine all four functions.
As the simplest option, let us assume the physical metric to be flat.
There are two possibilities for this, for the flat
metric of the form \eqref{gf} can be either Minkowski,
\begin{equation}
\dot{a}=0,~~~~~K=0,
\end{equation}
\textcolor{black}{or} Milne,
\begin{equation}
\dot{a}=N\neq 0,~~~~~K=-1.
\end{equation}
In both cases Eq.\eqref{eqna} requires that $X^0_0=X^1_1$ which implies that
$c=1$. Eq.\eqref{eqn} then reduces to
\begin{equation}
0=m^2\lambda_g+\bm{\Lambda}-\frac{m^2}{2n^2}\, \left(n-s-1\right)\xi^{2n-4s}\,,
\end{equation}
which determines a constant value for $\xi$, hence the reference metric is also
either Minkowski or Milne, respectively.
The Minkowski space has already been
considered above. Its stability is determined by the arguments
given around Eq.\eqref{Mink} -- it is stable only in the FP limit corresponding to the
ellipse in Fig.\ref{Fig1}.
For the Milne solution one has $\dot{a}\neq 0$
and one can apply Eq.\eqref{zzz}.
This gives the same result as for the de Sitter space
considered above: it is stable everywhere in the region inside the ellipse
in Fig.\ref{Fig1}, where the Minkowski space is unstable. This sounds odd, since the Milne space
is merely a sector of Minkowski space expressed in different coordinates.
If $T,R$ are the Minkowski time and radial coordinate, then one has
\begin{equation}
T=a(t)\cosh(r),~~~~R=a(t)\sinh(r),
\end{equation}
hence the Milne coordinates $t,r$ cover the interior of the future light cone.
How can it be that the unstable Minkowski
space
becomes stable when expressed in different
coordinates~? The answer is that the stability condition \eqref{zzz}
guarantees the absence of ghosts in the UV limit \eqref{UV}, but there could still be soft ghosts
with momenta
\begin{equation}
p\leq H=\frac{\dot{a}}{Na}=\frac{1}{a}=\frac{1}{\sqrt{T^2-R^2}}.
\end{equation}
It follows that, although the Milne space does not have UV ghosts, it must contain
soft ghosts with wavelengths of the order or larger than the Milne horizon.
One may then argue that the Minkowski space too is actually unstable only with respect
to long wave ghosts. At first glance, this
contradicts the fact that the ghost modes in Minkowski space exist for any momenta.
However, Eq.\eqref{XK}
shows that the leading UV contributions come from interactions of
perturbations with the background curvature, which
suggests that the kinetic energy of
perturbations around Minkowski should be dominated by nonlinear
interactions. Therefore, the linear ghost instability of Minkowski space in
the UV limit could presumably be cured by the nonlinear terms.
\subsection{\textcolor{black}{Solutions with constant $\xi$}}
To study general solutions of Eqs.\eqref{eqn}, \eqref{eqna},
it is convenient to consider their
consequence: the conservation condition
\begin{equation}
\left(X^0_{~0}\right)^{\mbox{.}}=3\,\frac{\dot{a}}{a}\left(X^1_{~1}-X^0_{~0}\right). \label{eqnn}
\end{equation}
If $\dot{a}\neq 0$ then this replaces the second order
equation \eqref{eqna}. This condition
can be represented in the form
\begin{equation} \label{eqn1}
\frac{\dot{c}}{c}+F_1(c)\,\frac{\dot{a}}{a}+F_2(c)\,\frac{\dot{\xi}}{\xi}=0,
\end{equation}
where
\begin{equation}
F_1(c)=\frac{6n(c^{2n}-1)}{W},~~~~F_2(c)=\frac{2(n-2s)((2n-s-1)c^{2n}-3(s+1) ) }{W},
\end{equation}
with $W=(2n-s-1)(2n-s)\,c^{2n}+3s(s+1)$.
As a result, the solutions are obtained
by solving Eqs.\eqref{eqn} and \eqref{eqn1}, supplemented by
an extra condition to totally specify the system.
For example, one can assume
$\xi$ to be a given function of $a$, then Eq.\eqref{eqn1} becomes
\begin{equation} \label{eqn2}
\frac{1}{c}\,\frac{dc}{da}+F_1(c)\,\frac{1}{a}+F_2(c)\,\frac{{\xi^\prime(a)}}{\xi(a)}=0,
\end{equation}
integrating which yields $c=c(a)$. Injecting this to \eqref{eee} will give $X^0_{~0}(a)$,
which will allow one to integrate Eq. \eqref{eqn}.
As the simplest option, one can set
$\xi(a)=\xi_0=const$, in which case the spatial parts of the two
metrics are proportional with the constant factor $\xi_0$.
Eq.\eqref{eqn2} then
reduces to
\begin{equation} \label{eqn22}
\textcolor{black}{\frac{1}{c}\frac{dc}{da} + F_1(c)\frac{1}{a}=0.}
\end{equation}
One can fulfill this by setting
$
c=1$, then the full 4-metrics are conformally related by the factor
$\xi_0$. This corresponds to the solutions with proportional
metrics already discussed above.
The general solution of \eqref{eqn22} for $c\neq const.$ is
\begin{equation} \label{a0}
a=a_0\,\frac{ c^{\alpha}}{(c^{2n}-1)^\beta }~~~~\mbox{with}~~~~
\alpha=\frac{s(s+1)}{2n},~~~~\beta=\frac{n(2n-2s-1)+2s(s+1)}{6n^2},
\end{equation}
where $a_0$ is the integration constant.
Inverting this to obtain $c=c(a)$ and
injecting to \eqref{eqn} yields
\begin{equation} \label{eqn1a}
\frac{3}{N^2}\frac{\dot{a}^2}{a^2}
+\frac{3K}{a^2}=m^2\lambda_g+\bm{\Lambda}
-\frac{m^2}{4n^2}\, \xi_0^{2n-4s}
\left[(2n-s-1)\, c^{2n-s}(a)-3(s+1)c^{-s}(a)\right].
\end{equation}
Imposing the $N=1$ gauge, this equation determines
$a(t)$, which specifies all four
metric amplitudes $N,N_f,a,a_f$.
\subsection{\textcolor{black}{Solutions with Minkowski fiducial metric}}
Let us apply the \textcolor{black}{procedure outlined above} to construct all solutions
in the case where the reference metric is flat Minkowski.
Therefore, the extra assumption is $\dot{a}_f=K=0$.
Using the definition of $\xi$ in \eqref{xi},
one can represent \eqref{eqn1} as
\begin{equation}
\frac{\dot{c}}{c} + [F_1(c)-F_2(c)]\,
\frac{\dot{a}}{a}
+F_2(c)\,\frac{\dot{a}_f}{a_f} =0, \label{eqn222a}\,
\end{equation}
and if $\dot{a}_f=0$ this reduces to
\begin{eqnarray} \label{eqn3}
a\,\frac{dc}{da}= \frac{2c (n-2s-2)\left[(2n-s)c^{2n}-3s\right]}{(2n-s-1)(2n-s)c^{2n}+3s(s+1)}.
\end{eqnarray}
It is clear that
\begin{equation}
c=c\left(\frac{a}{a_0}\right)\equiv c(\alpha),
\end{equation}
where $a_0$ is an integration constants, hence we set
\begin{equation}
a(t)=a_0 \,\alpha(t)
\end{equation}
and we also set the lapse function to the constant value,
\begin{equation}
N^2=\frac{12n^2}{m^2}\left(\frac{a_f}{a_0}\right)^{4s-2n}.
\end{equation}
Eq.\eqref{eqn} then assumes the form of the energy conservation,
\begin{equation} \label{cons}
\frac{\dot{\alpha}^2}{\alpha^2}
+{\rm U}(\alpha)={\cal E},
\end{equation}
with the ``potential energy"
\begin{equation} \label{V}
{\rm U}(\alpha)=\alpha^{4s-2n}\left[
(2n-s-1)c(\alpha)^{2n-s}-3(s+1)c(\alpha)^{-s}
\right],~~~~~
\end{equation}
and the ``total energy"
\begin{equation}
{\cal E}=\frac13\,N^2 (m^2\lambda_g+\bm{\Lambda}).
\end{equation}
The sign of the ``total energy" is determined by that of $m^2\lambda_g+\bm{\Lambda}$
while its absolute value depends on $a_0$ and hence can be
arbitrary.
Solving \eqref{eqn3} yields $c(\alpha)$, injecting which to \eqref{V} determines ${\rm U}(\alpha)$, and then
\eqref{cons} determines $\alpha(t)$. It is clear that $\alpha$ should be confined to the
region where ${\rm U}(\alpha)\leq {\cal E}$.
The solution of the problem contains several subcases, depending
on values of the parameters $n,s$.
Let us first consider cases where the amplitude $c$ is constant.
\underline{\underline{I. $n=2(s+1)$.}} The solution of \eqref{eqn3} is an arbitrary constant
that can be assumed to be positive, $c=c_0>0$. The potential becomes
\begin{equation} \label{V0}
{\rm U}=\frac{3(s+1)c_0^{-s}(c_0^{4(s+1)}-1)}{\alpha^4},
\end{equation}
which can be positive or negative, depending on values of $s$ and $c_0$.
If ${\rm U}$ is negative, then it describes an effective radiation
mimicked by massive gravitons. The scale factor then evolves as in the universe containing
a radiation and an effective cosmological term mimicked by ${\cal E}$.
If ${\rm U}>0$ then the massive gravitons mimic a ``phantom radiation".
One has in this case ${\rm U}(\alpha)\to +\infty$ as $\alpha\to 0$ hence there is an infinite potential barrier near singularity so that
the solution is a {\it bounce}: the universe first shrinks up to a minimal non-zero size,
then hits the potential barrier and expands again.
It is worth noting that the very existence of bounces indicates that the
Null Energy Condition is violated.
At the same time, the solutions can be free of ghosts and \textcolor{black}{gradient instabilities}. Indeed, choosing
$c_0$ to be close to unity, the no-ghost condition $\zeta(c_0)>0$
expressed by \eqref{zzz} will have approximately the same solutions as for $c_0=1$,
corresponding to the interior of the ellipse in Fig.\ref{Fig1}.
The two sound speeds
will then be close to unity.
\underline{\underline{II. $n\neq 2(s+1)$, $c=const$.}} The solution of \eqref{eqn3}
\textcolor{black}{in this case} is
\begin{equation} \label{cc}
c^{2n}=\frac{3s}{2n-s}\equiv c_\ast^{2n},
\end{equation}
where one should assume that either $2n>s>0$ or $2n<s<0$ for $c_\ast$ to be positive.
One has
\begin{equation}
{\rm U}=-\frac{6n c_\ast^{-s}}{2n-s}\, \alpha^{4s-2n}<0.
\end{equation}
If $n=2s$ then ${\rm U}=\textcolor{black}{-4}$ and the universe expands with a constant Hubble \textcolor{black}{expansion} rate.
If $n>2s$ then for small $\alpha$ the universe is dominated by an effective \textcolor{black}{``fluid"}
mimicked by ${\rm U}$ while for large $\alpha$ the potential approaches zero and the universe expands
with a constant Hubble \textcolor{black}{expansion} rate $H^2={\cal E}\textcolor{black}{/N^2}$ (which should be positive).
If $n<2s$ then the \textcolor{black}{``fluid term"} ${\rm U}$
grows without bound as $\alpha \to\infty$.
The no-ghost condition \eqref{UV} reduces to
\begin{equation}
\zeta(c_\ast)=-\frac{3s}{2n}>0,
\end{equation}
which is impossible to fulfill
since $n$ and $s$ should be either both positive or both negative \textcolor{black}{for $c_\ast$ to be positive}.
Therefore, type II solutions always have ghost.
\textcolor{black}{\underline{\underline{III. $n\neq 2(s+1)$, $c\neq const$.}}}
Let us now study solutions for which the amplitude $c$ is not constant.
The general solution of \eqref{eqn3} for $n\neq 2(s+1)$ is
\begin{equation} \label{gen}
\frac{(2n-s)\,c^{2n}-3s }{c^{s+1}}=\left(\frac{a}{a_0} \right)^{2(n-2s-2) }.
\end{equation}
Several subcases are to be considered. The left hand side of this expression
is positive in one of the following cases,
\begin{eqnarray} \label{cases}
&&{\rm (a)}~s=2n<0;~~~~
{\rm (b)}~s=0, n>0;~~~~
{\rm (c)}~n>s/2>0, c^{2n}>3s/(2n-s);~~ \nonumber \\
&&{\rm (d)}~n<s/2<0, c^{2n}<|3s/(2n-s)|;~~~~
{\rm (e)}~n>s/2,~s<0.
\end{eqnarray}
Each of these cases further splits into subcases depending on the signs of $n$,
of $s+1$,
of $2n-s-1$, and of $n-2s-2$. This renders the
classification of solutions a bit tedious, but still manageable.
\textcolor{black}{\underline{IIIa.}} Let us first consider case (a) in \eqref{cases}, where $s=2n<0$.
Then, properly redefining the constant $a_0$ in \eqref{gen},
one obtains
\begin{equation} \label{c}
c^{2n+1}=\left(\frac{a}{a_0} \right)^{2(3n+2)}\equiv \alpha ^{2(3n+2)},
\end{equation}
hence, \textcolor{black}{as $\alpha$ increases from zero to infinity,} $c$ either increases from zero to infinity or decreases from
infinity to zero. The potential is
\begin{equation} \label{VV}
{\rm U}(\alpha)=-\alpha^{6n}-3(2n+1)\alpha^{-\frac{2n}{2n+1}}.
\end{equation}
Depending on value of $n$ there are several subcases.
\textcolor{black}{IIIa-1: $-1/2<n<0$, $s=2n$.} $c(\alpha)$ increases \textcolor{black}{as $\alpha$ increases}.
The potential ${\rm U}(\alpha)$ is always negative and \textcolor{black}{its absolute value}
becomes large for $\alpha\to 0$ and for $\alpha\to\infty$.
\textcolor{black}{IIIa-2: $-2/3<n<-1/2$, $s=2n$.} $c(\alpha)$ decreases \textcolor{black}{as $\alpha$ increases}.
The potential ${\rm U}(\alpha)$ is qualitatively similar to the one
shown in panel B in Fig.\ref{Fig2} below.
It
is large and positive at small $\alpha$ while
for $\alpha\to\infty$ it approaches zero from below, hence there is a minimal value
${\rm U}_{\rm min}={\rm U}(\alpha_{\rm min})<0$ \textcolor{black}{with $\alpha_{\rm min}=1$}.
If
${\cal E}={\rm U}_{\rm min}$ then the system always rests at the minimum
and the geometry is flat \textcolor{black}{with $c(\alpha_{\rm min})=1$}.
If ${\rm U}_{\rm min} <{\cal E}<0$ then $\alpha(t)$
oscillates around the value $\alpha_{\rm min}$ while $c(t)$ oscillates
around $c(\alpha_{\rm min})=1$.
For ${\cal E}\geq 0$ the solution is a bounce.
\textcolor{black}{IIIa-3: $n<-2/3$, $s=2n$.} $c(\alpha)$ increases
\textcolor{black}{as $\alpha$ increases}.
The potential ${\rm U}(\alpha)$
is qualitatively similar to the one
shown in panel A in Fig.\ref{Fig2} below.
It is large and negative for small $\alpha$, then passes through a maximal value
${\rm U}_{\rm max}={\rm U}(\alpha_{\rm max})>0$ \textcolor{black}{with $\alpha_{\rm max}=1$},
then approaches
zero from above as $\alpha\to\infty$.
Depending on value of ${\cal E}$,
the motions in this potential correspond either to cosmologies with an
initial
singularity or to non-singular bounces. If
${\rm U}_{\rm max}={\cal E}$ then
there is a solution for which $\alpha(t)$ grows
from the constant value $\alpha_{\rm max}$ in the past to infinity
in the future,
hence the universe interpolates between the flat space
and de Sitter space.
There are also two boundary cases, $n=-1/2$ and $n=-2/3$.
For $n=-1/2$ the scale factor $a$ should be constant, as seen from Eq.\eqref{eqn3},
hence this is a particular case of the Minkowski solutions.
The $n=-2/3$ case corresponds to the intersection with family I since one has then
$n=2(s+1)$.
For all type \textcolor{black}{IIIa} solutions the no-ghost condition {\eqref{zzz}}
reduces to
\begin{equation}
\zeta(c)=-3\,\frac{2n+1}{2n}>0\,,
\end{equation}
which is independent of $c$. This condition is fulfilled for solutions
of type \textcolor{black}{IIIa-1}, hence they are ghost-free. However, one finds then that the sound speeds are imaginary,
hence there are \textcolor{black}{gradient instabilities}.
The no-ghost condition is violated for
solutions of types \textcolor{black}{IIIa-2 and IIIa-3}.
\textcolor{black}{\underline{IIIb.}} In case (b) in \eqref{cases}, for $s=0$ and $n>0$, \textcolor{black}{the no-ghost condition \eqref{zzz} becomes}
\begin{equation}
\zeta(c)=\frac{1-2n}{2n}\,c^{2n} \textcolor{black}{>0}.
\end{equation}
This is \textcolor{black}{satisfied} for $n<1/2$, however, there are \textcolor{black}{gradient instabilities}
in this case \textcolor{black}{as the coefficient of $C_S^2$ in \eqref{ccc0} is positive}.
\textcolor{black}{\underline{IIIc.}} Let us now consider case (c) in \eqref{cases}, with $n>s/2>0$.
Then, redefining the integration constant $a_0$, Eq.\eqref{gen} can be rewritten as
\begin{equation}
\frac{c^{2n}-c_\ast^{2n}}{c^{s+1}}=\left(\frac{a}{a_0} \right)^{2(n-2s-2) }\equiv \alpha^{2(n-2s-2) },
\end{equation}
where $c_\ast$ is the same as in \eqref{cc}, hence
one should have $c\in[c_\ast,\infty)$ for the left hand side
to be positive.
There are again several cases to study
and one finds the following possibilities.
\textcolor{black}{IIIc-1: $2n>s>0$, $n>2(s+1)$.} $c(\alpha)$ increases from $c_\ast$
to infinity; the potential
${\rm U}(\alpha)$ has a maximum as shown in panel A in Fig.\ref{Fig2}.
\textcolor{black}{IIIc-2: $2n>s>0$, $\max\{(s+1)/2,~2s\}<n<2(s+1)$.} $c(\alpha)$ decreases
from infinity to $c_\ast$; the potential
${\rm U}(\alpha)$ has a minimum as shown in panel B in Fig.\ref{Fig2}.
\begin{figure}[th]
\hbox to \linewidth{ \hss
\resizebox{8cm}{6cm}{\includegraphics{fg1.pdf}}
\resizebox{8cm}{6cm}{\includegraphics{fg3.pdf}}
\hss}
\hbox to \linewidth{ \hss
\resizebox{8cm}{6cm}{\includegraphics{fg2.pdf}}
\resizebox{8cm}{6cm}{\includegraphics{fg4.pdf}}
\hss}
\caption{The potential ${\rm U}(\alpha)$ and $c(\alpha)$ for generic $2n>s>0$.}
\label{Fig2}
\end{figure}
\textcolor{black}{IIIc-3: $2n>s>0$, $(s+1)/2<n<2s$.} Both $c(\alpha)$ and ${\rm U}(\alpha)$
decrease as shown in panel C in Fig.\ref{Fig2}.
\textcolor{black}{IIIc-4: $2n>s>0$, $s/2<n<(s+1)/2$.} Both $c(\alpha)$ and ${\rm U}(\alpha)$
are two-valued functions defined only for $\alpha\geq \alpha_{\rm min}>0$
as shown in panel D in Fig.\ref{Fig2}. For ${\cal E}<{\rm U}_{\rm max}$ there are two different bounce solutions
corresponding to reflections from either the lower or upper branch of $U$.
For solutions of types \textcolor{black}{IIIc-1, IIIc-2, and IIIc-3}
the amplitude $c(\alpha)$ approaches the value
$c_\ast$ either for small or for large $\alpha$, which insures that there is ghost
\textcolor{black}{since $\zeta(c_\ast)=-(3s)/(2n)<0$}
in these cases.
The situation is more complex in case \textcolor{black}{IIIc-4}.
If one plots $a$, ${\rm U}$, and $\zeta(c)$ against $c$ in this case, one finds that
$a(c)$ attains a minimal non-zero value at some $c_m$,
the potential ${\rm U}(c)$ attains
a maximum at the same time, while $\zeta(c)$ changes sign. As a result, both $c(\alpha)$
and ${\rm U}(\alpha)$ are double-valued functions as shown in panel D in Fig.\ref{Fig2}.
Their two branches determine two different solutions
which should be considered independently. The upper branch of ${\rm U}(\alpha)$ corresponds
to the lower branch of $c(\alpha)$ where one has $\zeta(c)<0$,
hence there is ghost. The lower branch of ${\rm U}(\alpha)$ corresponds to the upper
branch of $c(\alpha)$ where $\zeta(c)>0$, hence this branch is ghost-free.
However, one finds
\textcolor{black}{gradient instabilities} there.
\textcolor{black}{\underline{IIId and IIIe.}}
Nothing qualitatively new is found in cases (d) and (e) in \eqref{cases}.
There are again several subcases to study, but each time one finds the potential to
be either of one of the types show in Fig.\ref{Fig2}, or of type \textcolor{black}{IIIa-1}.
All of these solutions contain ghosts and/or \textcolor{black}{gradient instabilities}.
This gives all solutions with the Minkowski reference metric.
Only type I solutions can be stable.
\subsection{\textcolor{black}{Solutions with Milne fiducial metric}}
\textcolor{black}{If the fiducial metric is Milne, then $K=-1$, $N_f=\dot{a}_f$, and
Eq.\eqref{eqn222a} reduces to }
\begin{equation}
\frac{\dot{c}}{c} + [F_1(c)-F_2(c)]\,
\frac{\dot{a}}{a}
+F_2(c)\,c\,\frac{N}{a} =0. \label{eqn222}
\end{equation}
Setting $c$ to a constant value yields
\begin{equation} \label{XXy}
\frac{\dot{a}}{N}=\frac{c F_2(c)}{F_2(c)-F_1(c)},
\end{equation}
injecting which to \eqref{eqn} \textcolor{black}{with $K=-1$} one obtains
\begin{equation}
3\left(\left[\frac{c F_2(c)}{F_2(c)-F_1(c)} \right]^2-1\right)\frac{1}{a^2}
=m^2\lambda_g+\bm{\Lambda}
-\frac{m^2}{4n^2}\, c^{-s} \left[(2n-s-1)\, c^{2n}-3\left(s+1\right)\right]\xi^{2n-4s}.
\end{equation}
This equation determines $\xi=\xi(a)$, while \eqref{XXy},
imposing the gauge where $N=1$, insures that $a$ is a linear function of time.
This specifies all functions in the problem.
\textcolor{black}{If $c\neq 1$ then
the physical metric is not flat.}
If $c$ is close to unity then one can adjust $s,n$ such that the solution will be free
of ghosts and tachyons (in the UV limit).
More general solutions with $\dot{c}\neq 0$ can be obtained by setting
\textcolor{black}{$N=a$ in \eqref{eqn222}}, which gives the equation containing only $c$ and $a$.
However, solutions of this equations are not immediately obvious.
At the same time, as we learned above, solutions with a non-trivial $c$ usually
show ghosts and/or \textcolor{black}{gradient instabilities}.
At this point, we terminate our analysis of the OP theory and come to the conclusions.
\section{Summary and concluding remarks}
We presented above our analysis of the theory constructed by
Ogievetsky and Polubarinov in \cite{OP}.
Taking apart the remarkable way it was obtained,
it is just a bimetric massive gravity with a specially designed mass term.
As any other bimetric theory, it
has two tensors, $X^\mu_{~\nu}$ and $Y^\mu_{~\nu}$, obtained
by varying with respect to the two metrics, respectively.
If the reference metric is flat, then one has on-shell $\partial_\mu Y^\mu_{~\nu}=0$.
The specialty of the OP theory is that, by the non-linear field redefinition \eqref{g-g1}, its physical metric
can be algebraically expressed in terms of the ``graviton field" ${ \Psi }^{\mu\nu}$
in such a way that $Y^\mu_{~\nu}$ becomes {\it linear} in ${ \Psi }^{\mu\nu}$.
Therefore, the field equations imply as a consequence
the linear ``spin limitation condition"
$\partial_\mu \Psi^\mu_{~\nu}+q\,\partial_\nu \Psi^\alpha_{~\alpha}=0$.
In the OP's view, this property is very important as it allows one to carry
out a ``clean" classification of the spin states present in the theory.
The linear subsidiary condition arising on-shell is the
key property that distinguishes the OP theory
among other bimetric models.
It is, however, not immediately obvious what this property gives in practical terms.
The most striking point is that the OP theory has a non-zero intersection
with the dRGT ghost-free massive gravity. However, for generic parameter
values it propagates 6 degrees of freedom and shows ghost on flat background.
According to the currently adopted view, these properties are unacceptable.
Nevertheless, we studied the phenomenology of the theory to see if
something interesting may emerge.
We found many different types of homogeneous and isotropic cosmological solutions,
including self-accelerating cosmologies, bounces, oscillating solutions, etc.
Unfortunately, most of them show ghost and/or \textcolor{black}{gradient instabilities}.
Surprizingly, however, we find that the de Sitter space is stable
in a large region of the parameter space, in spite of the presence of the
6-th degree of freedom, hence the Boulware-Deser mode becomes benign.
Moreover, even the instability of the flat space does not seem
so dramatic, as it turns out that the Milne universe, which is a sector of
Minkowski space, is actually UV stable in the theory. Therefore, the instability
can only be due to the soft modes with wavelengths of the order or larger than the
Milne horizon. This suggests that the flat space ghost instability is similar
to the classical Jeans instability
with respect to long-wave perturbations. Therefore, the BD mode
could probably
be viewed as some kind of non-relativistic fluid \cite{Gumrukcuoglu:2016jbh}.
Our conclusion is that the OP theory does show interesting features, despite the presence of the
6-th polarization. This may be due to the
``spin limitation" encoded in the theory, even though there are other
massive gravities with a healthy 6-th mode \cite{Celoria:2017bbh}.
It would be interesting to study also other solutions in the theory, as for example black holes.
It seems also that the particular OP theory with $n=1/2$ and $s=0$, which
shows both the ``spin limitation" and the ``freedom of the ghost", should be further
studied.
\section*{Acknowledgements}
The work of S.M. was supported by Japan Society for the Promotion of Science (JSPS) Grants-in-Aid for Scientific Research (KAKENHI) No. 17H02890, No. 17H06359, and by
World Premier International Research Center Initiative (WPI), MEXT, Japan.
M.S.V. thanks for hospitality the YITP in Kyoto, where a part of this work was completed.
Discussions during the workshop YITP-T-17-02 ``Gravity and Cosmology 2018" and the
YKIS2018a symposium ``General Relativity -- The Next Generation" were useful.
His work was also partly supported by the Russian Government Program of Competitive Growth
of the Kazan Federal University.
|
\section{Introduction}
One of the most fundamental and studied problems in enumerative geometry is the following: what is $N_d^{\delta}$, the number of
degree $d$ curves in $\mathbb{P}^2$ that have $\delta$ distinct nodes and pass through $\frac{d(d+3)}{2} - \delta$ generic points?
The question was studied more than a hundred years ago by Zeuthen (\cite{Zeuthen}) and has been studied extensively in the last thirty
years
by Ran (\cite{Ran1}, \cite{Ran2}), Vainsencher (\cite{Van}), Caporaso-Harris (\cite{CH}), Kazarian (\cite{Kaz}), Kleiman and Piene (\cite{KP1}),
Florian Block (\cite{FB}),
Tzeng and Li (\cite{Tz}, \cite{Tzeng_Li}), Kool, Shende and Thomas (\cite{KST}), Berczi (\cite{Berczi}) and Fomin and Mikhalkin (\cite{FoMi}),
amongst others. This question has been investigated from several perspectives and is very well understood. \\
\hf \hf The problem motivates a natural generalization considered by Kleiman and Piene in \cite{KP2}, where they study the
enumerative geometry of singular curves in a moving family of
surfaces. More recently, this question has been studied further by T.~Laarakker in \cite{TL}, where he obtains
a formula for the following number: how many degree $d$ curves are
there in $\mathbb{P}^3$ whose image lies in a $\mathbb{P}^2$, that pass through $\frac{d(d+3)}{2} + 3 - \delta$ generic lines and have $\delta$-nodes
(provided $d\geq \delta$). This can be viewed as a family version of the classical problem of computing $N_d^{\delta}$. \\
\hf \hf Motivated by the papers of Kleiman and Piene (\cite{KP2}) and T.~Laarakker (\cite{TL}),
we have studied the parallel question of counting stable
rational maps into a family of moving target spaces. This can be viewed as a family version of the famous question of enumerating rational curves
in $\mathbb{P}^2$, that was studied by Kontsevich-Manin (\cite{KoMa}) and Ruan-Tian (\cite{RT}). The main result of our paper is as follows:
\begin{mresult}
Let $N_{d}^{\mathbb{P}^3, \mathrm{Planar}}(r,s)$ be the number of genus zero, degree $d$ curves in $\mathbb{P}^3$, whose image lies in a $\mathbb{P}^2$, intersecting
$r$ generic lines and $s$ generic points (where $r + 2s = 3d+2$). We have
a recursive formula to compute $N_{d}^{\mathbb{P}^3, \mathrm{Planar}}(r,s)$ for all $d \geq 2$.
\end{mresult}
\begin{rem}
Note that for $d=1$, the corresponding question is classical Schubert calculus and there $r+2s = 4$ as opposed to $5$.
\end{rem}
\begin{rem}
We note that when $s=3$ and $d \geq 2$, the number $N_{d}^{\mathbb{P}^3, \mathrm{Planar}}(3d-4,3)$ is the number of rational curves in $\mathbb{P}^2$
through $3d-1$ points; this is because $3$ generic points in $\mathbb{P}^3$ determine a unique $\mathbb{P}^2$. We also note that when $s>3$,
$N_d^{\mathbb{P}^3, \mathrm{Planar}}(r,s)$ is zero, since $4$ or more generic points do not lie in a plane.
\end{rem}
We have written a mathematica program to implement our formula;
the program is available on our web page
\[ \textnormal{\url{https://www.sites.google.com/site/ritwik371/home}}. \]
In section \ref{low_deg_check}, we verify that the numbers we compute are logically consistent with those obtained by T.~Laarakker in \cite{TL} till $d=6$.
This gives strong evidence to support the conjecture that his formulas for $\delta$-nodal planar degree $d$ curves in $\mathbb{P}^3$
are expected to be enumerative when $d \geq 1 + [\frac{\delta}{2}]$
(as opposed to $d \geq \delta$ which is proved in \cite{TL}).
Starting from $d=7$, we can not use the result of \cite{TL} to make any consistency check, since the corresponding nodal polynomial is not expected to
be enumerative (due to the presence of double lines); this is explained in section \ref{low_deg_check}.
\section{Notation }
Let us define a \textbf{planar} curve in $\mathbb{P}^3$ to be a curve, whose image lies inside a $\mathbb{P}^2$.
We will now develop some notation to describe the space of planar curves of a given degree $d$. \\
\hf \hf Let us denote the dual of $\mathbb{P}^3$ by
$\hat{\mathbb{P}}^3$; this is the space of $\mathbb{P}^2$ inside $\mathbb{P}^3$.
An element of $\hat{\mathbb{P}}^3$ can be thought of as
a non zero linear functional $\eta: \mathbb{C}^4 \longrightarrow \mathbb{C}$ upto scaling (i.e., it is the projectivization of the dual of $\mathbb{C}^4$).
Given such an $\eta$, we define the projectivization of its zero set as $\mathbb{P}^2_{\eta}$. In other words,
\begin{align*}
\mathbb{P}^2_{\eta} &:= \mathbb{P}(\eta^{-1}(0)).
\end{align*}
Note that this $\mathbb{P}^2_{\eta}$ is a subset of $\mathbb{P}^3$. Next, when $d \geq 2$, we
define the moduli space of planar degree $d$ curves into $\mathbb{P}^3$
as a fibre bundle over $\hat{\mathbb{P}}^3$. More precisely,
we define
\[ \pi: \overline{\mathcal{M}}_{0,k}^{\mathrm{Planar}}(\mathbb{P}^3,d) \longrightarrow \hat{\mathbb{P}}^3 \]
to be the fiber bundle, such that
\[ \pi^{-1}([\eta]) := \overline{\mathcal{M}}_{0,k}(\mathbb{P}^2_{\eta}, d). \]
Here we are using the standard notation to denote $\mathcal{M}_{0,k}(X, \beta)$ to be the moduli space of genus zero stable maps,
representing the class $\beta \in H_2(X, \mathbb{Z})$ and $\overline{\mathcal{M}}_{0,k}(X, \beta)$ to be its stable map compactification.
Since the dimension of a fiber bundle is dimension of the base, plus the dimension of the fiber, we conclude that the dimension of
$\overline{\mathcal{M}}_{0,k}^{\mathrm{Planar}}(\mathbb{P}^3,d)$ is $3d+2+k$.\\
\hf\hf Next, we note that the space of planes in $\mathbb{P}^3$ can also be thought of as the Grassmanian $\mathbb{G}(3,4)$. Let
$\gamma_{3,4}$ denote the tautological three plane bundle over the Grassmanian. Since $\mathbb{G}(3,4)$ can be identified with $\hat{\mathbb{P}}^3$,
we can think of $\gamma_{3,4}$ as a bundle over $\hat{\mathbb{P}}^3$. \\
\hf \hf When $d=1$, we define $\overline{\mathcal{M}}_{0,0}^{\mathrm{Planar}}(\mathbb{P}^3,1)$ to be
\[ \overline{\mathcal{M}}_{0,0}^{\mathrm{Planar}}(\mathbb{P}^3,1) := \mathbb{P}(\gamma_{3,4}^*) \longrightarrow \hat{\mathbb{P}}^3.\]
We note that an element of $\overline{\mathcal{M}}_{0,0}^{\mathrm{Planar}}(\mathbb{P}^3,1)$
is of the form $(L, H)$, where $L$ is a line in $\mathbb{P}^3$ and $H$ is a plane containing $L$.
Since a line is not contained in a unique plane, $\overline{\mathcal{M}}_{0,0}^{\mathrm{Planar}}(\mathbb{P}^3,1)$ is not the
same as the space of lines; infact we note that the space of lines is $4$
dimensional, while the dimension of $\overline{\mathcal{M}}_{0,0}^{\mathrm{Planar}}(\mathbb{P}^3,1)$ is $5$.\\
\hf \hf We will now define a few numbers by intersecting cycles on $\overline{\mathcal{M}}_{0,0}^{\mathrm{Planar}}(\mathbb{P}^3,d)$, the moduli space with zero marked points
(this includes the case $d=1$; unless otherwise stated we always include the case $d=1$ in any of our statements).
Let
$\mathcal{H}_L$ and $\mathcal{H}_p$ denote the classes of the cycles in $\overline{\mathcal{M}}_{0,0}^{\mathrm{Planar}}(\mathbb{P}^3,d)$ that corresponds
to the subspace of curves passing through a generic line and a point respectively.
We also denote $a$
to be the standard generator of $H^*(\hat{\mathbb{P}}^3; \mathbb{Z})$.
Let us now define
\begin{align}
N_d^{\mathbb{P}^3,\mathrm{Planar}}(r,s,\theta) &:= \langle a^{\theta}, ~~\overline{\mathcal{M}}_{0,0}^{\mathrm{Planar}}(\mathbb{P}^3,d) \cap \mathcal{H}_L^r \cap \mathcal{H}_p^s \rangle.
\label{nd_rs_theta}
\end{align}
We formally define the the left hand side of $\eqref{nd_rs_theta}$ to be zero if $r+2s + \theta \neq 3d+2$
(since the right hand side of \eqref{nd_rs_theta} doesn't make sense unless $r+2s + \theta = 3d+2$).
Note that when $\theta =0$, $r + 2s =3d+2$ and $d \geq 2$,
$N_d^{\mathbb{P}^3,\mathrm{Planar}}(r,s,0)$ is precisely equal to the number of rational planar degree $d$-curves in $\mathbb{P}^3$ intersecting $r$ generic lines and
$s$ generic points. \\
\hf \hf Next, we will define a number $B_{d_1,d_2}(r_1,s_1,r_2,s_2,\theta)$ by intersecting it on the product
of two moduli spaces as follows:
\begin{eqnarray}
B_{d_1,d_2}(r_1,s_1,r_2,s_2,\theta) &:= \Big\langle \pi_1^*( \mathcal{H}_L^{r_1} \mathcal{H}_p^{s_1})\cdot \pi_2^*( \mathcal{H}_L^{r_2} \mathcal{H}_p^{s_2}) (\pi_2^*a)^{\theta}\cdot
(\pi^*\Delta), \nonumber \\
& \qquad [\overline{\mathcal{M}}_{0,0}^{\mathrm{Planar}}(\mathbb{P}^3,d_1) \times \overline{\mathcal{M}}_{0,0}^{\mathrm{Planar}}(\mathbb{P}^3,d_2)] \Big\rangle. \label{bd_delta}
\end{eqnarray}
Here $\Delta$ denotes the class of the diagonal in $\hat{\mathbb{P}}^3 \times \hat{\mathbb{P}}^3$ and $\pi_1$ and $\pi_2$ are the obvious projection
maps.
Again, we formally define the left hand
side of \eqref{bd_delta} to be zero, unless $r_1 + 2s_1 + r_2 + 2 s_2 + \theta = 3d_1 + 3d_2 + 4$ (since in that case,
the right hand side doesn't make sense). Note that when $\theta =0$ and $r_1 + 2s_1 + r_2 + 2 s_2 = 3d_1 + 3d_2 + 4$, the number $B_{d_1,d_2}(r_1,s_1,r_2,s_2,0)$
denotes the number of two component rational curves in $\mathbb{P}^3$ of degree $(d_1, d_2)$, such that the two components intersect and lie in a common plane
and the $d_i$ component intersects $r_i$ lines and $s_i$ points, for $i=1$ and $2$.
\section{Recursive Formula and its Proof}
We are now ready to state our recursion formula
\begin{lmm}
\label{base_case_recursion}
If $d=1 $, then the number $N_d^{\mathbb{P}^3, \mathrm{Planar}}(r,s,\theta)$ is given by
\begin{align}
\label{base_formula}
N_1^{\mathbb{P}^3,\mathrm{Planar}}(r,s,\theta) & = \begin{cases} 0 & \mbox{if} ~~(r,s,\theta) =(1,2,0), \\
0 & \mbox{if } ~~(r,s,\theta) =(3,1,0),\\
0 & \mbox{if } ~~(r,s,\theta) =(5,0,0),\\
1 & \mbox{if} ~~(r,s,\theta) =(0,2,1),\\
1 & \mbox{if} ~~(r,s,\theta) =(2,1,1),\\
2 & \mbox{if} ~~(r,s,\theta) =(4,0,1), \\
1 & \mbox{if} ~~(r,s,\theta) =(1,1,2), \\
2 & \mbox{if} ~~(r,s,\theta) =(3,0,2), \\
0 & \mbox{if} ~~(r,s,\theta) =(0,1,3), \\
1 & \mbox{if} ~~(r,s,\theta) =(2,0,3), \\
0 & \mbox{\textnormal{otherwise}}.\end{cases}
\end{align}
\end{lmm}
\begin{lmm}
\label{second_case_recursion}
If $d=2 $, then the number $N_d^{\mathbb{P}^3, \mathrm{Planar}}(r,s,\theta)$ is given by
\begin{align}
\label{second_case_formula}
N_2^{\mathbb{P}^3,\mathrm{Planar}}(r,s,\theta) & = \begin{cases} 92 & \mbox{if} ~~(r,s,\theta) =(8,0,0), \\
18 & \mbox{if } ~~(r,s,\theta) =(6,1,0),\\
4 & \mbox{if } ~~(r,s,\theta) =(4,2,0),\\
1 & \mbox{if} ~~(r,s,\theta) =(2,3,0),\\
34 & \mbox{if} ~~(r,s,\theta) =(7,0,1),\\
6 & \mbox{if} ~~(r,s,\theta) =(5,1,1), \\
1 & \mbox{if} ~~(r,s,\theta) =(3,2,1), \\
0 & \mbox{if} ~~(r,s,\theta) =(1,3,1), \\
8 & \mbox{if} ~~(r,s,\theta) =(6,0,2), \\
1 & \mbox{if} ~~(r,s,\theta) =(4,1,2), \\
0 & \mbox{if} ~~(r,s,\theta) =(2,2,2), \\
0 & \mbox{if} ~~(r,s,\theta) =(0,3,2), \\
1 & \mbox{if} ~~(r,s,\theta) =(5,0,3), \\
0 & \mbox{if} ~~(r,s,\theta) =(3,1,3), \\
0 & \mbox{if} ~~(r,s,\theta) =(1,2,3), \\
0 & \mbox{\textnormal{otherwise}}.\end{cases}
\end{align}
\end{lmm}
\begin{thm}
\label{rec_formula_mthm}
If $d\geq 3$, then
\begin{align}
\label{triv_eqn}
N_d^{\mathbb{P}^3, \mathrm{Planar}}(r,s,\theta) &= \begin{cases} 0 & \mbox{if} ~~r + 2s + \theta \neq 3d+2, \\
0 & \mbox{if } ~~s>3,\\
0 & \mbox{if } ~~\theta>3.\end{cases}
\end{align}
In the remaining case when $r+2s + \theta = 3d+2$, $s \leq 3$ and $\theta \leq 3$, we have
\begin{align}\label{recurfor}
N_d^{\mathbb{P}^3, \mathrm{Planar}}(r,s,\theta)& =2d N_d^{\mathbb{P}^3, \mathrm{Planar}}(r-2,s+1,\theta) \nonumber \\
& +\sum_{r_1 =0}^{r-3}\sum_{s_1 =0}^{s}\sum_{d_1=1}^{d-1}\binom{r-3}{r_1}\binom{s}{s_1}d_1^2 d_2\times \nonumber \\
& \Big(d_2 B_{d_1,d_2}(r_1+1,s_1,r_2-2,s_2,\theta)
-d_1B_{d_1,d_2}(r_1,s_1,r_2-1,s_2,\theta)\Big),
\end{align}
where in the above expression, $d_2:= d-d_1$, $r_2:= r-r_1$ and $s_2:= s-s_1$.
Furthermore,
$\forall d_1, d_2 \geq 1$, we have \\
\begin{equation}\label{b1}
B_{d_1,d_2}(r_1,s_1,r_2,s_2,\theta) =\sum_{i=0}^3 N_{d_1}^{\mathbb{P}^3, \mathrm{Planar}}(r_1,s_1,i)\times N_{d_2}^{\mathbb{P}^3, \mathrm{Planar}}(r_2,s_2,\theta+3-i).
\end{equation}
\end{thm}
\begin{rem}
We note that equations \eqref{base_formula}, \eqref{second_case_formula}, \eqref{triv_eqn},
\eqref{recurfor} and \eqref{b1} allow us to compute $N_d^{\mathbb{P}^3, \mathrm{Planar}}(r,s,\theta)$
for all $d \geq 1$ and all $r, s, \theta \geq 0$.
\end{rem}
\noindent \textbf{Proof of Theorem \ref{rec_formula_mthm}:}
We will start by proving equation \eqref{triv_eqn}. The first equation is true simply because we are pairing a cohomology class with a homology
class of different dimensions (see the remark after equation \eqref{nd_rs_theta}). Next, when $s>3$, we note that there can not be any planar curves,
because $4$ generic points do not lie on a plane. Finally, we note that $a^{\theta} =0 \in H^*(\hat{\mathbb{P}}^3;\mathbb{Z})$ if $\theta >3$, which proves the last equality. \\
\hf \hf We now justify the main thing, which is equation \eqref{recurfor}.
The idea is very similar to the idea used to compute
the number of rational degree $d$ curves in $\mathbb{P}^2$ (and also $\mathbb{P}^3$) that is given in
\cite{KoMa}, \cite{McSa} and \cite{FuPa}.
As in the case of counting curves in $\mathbb{P}^2$,
let us first consider $\overline{\mathcal{M}}_{0,4}$. This space is isomorphic to $\mathbb{P}^1$; hence we have equivalence of the
following divisors
\begin{align}
(x_1x_2|x_3 x_4) \equiv (x_1x_3|x_2 x_4). \label{wdvv_p1}
\end{align}
Here $(x_1x_2|x_3 x_4)$ denotes the domain which is a wedge of two spheres with the marked points $x_1$ and $x_2$ on the first
sphere and $x_3$ and $x_4$ on the second sphere. The domain $(x_1x_3|x_2 x_4)$ is defined similarly. Since $\overline{\mathcal{M}}_{0,4} \approx \mathbb{P}^1$
is path connected, any two points determine the same divisor. \\
\hf \hf Let us now consider the projection map
\[\pi: \overline{\mathcal{M}}_{0,4}^{\textrm{Planar}}(\mathbb{P}^3, d) \longrightarrow \overline{\mathcal{M}}_{0,4}. \]
We define a cycle $\mathcal{Z}$ in $\overline{\mathcal{M}}_{0,4}^{\textrm{Planar}}(\mathbb{P}^3, d)$, given by
\begin{align*}
\mathcal{Z} &:= \textnormal{ev}_1^*([H]) \cdot \textnormal{ev}_2^*([H]) \cdot \textnormal{ev}_3^*([L]) \cdot \textnormal{ev}_4^*([L]) \cdot \mathcal{H}_L^{r-3}
\cdot \mathcal{H}_p^s \cdot a^{\theta},
\end{align*}
where $[H]$ and $[L]$ denote the class of a hyperplane and a line in $\mathbb{P}^3$, $\textnormal{ev}_i$ denotes the evaluation map at the $i^{\textnormal{th}}$
marked point and $a$ denotes the generator of $H^*(\hat{\mathbb{P}}^3;\mathbb{Z})$.
Let us now intersect the cycle $\mathcal{Z}$ by pulling back the left hand side of \eqref{wdvv_p1}, via $\pi$.
We now note that
\begin{align}
\pi^*(x_1 x_2|x_3 x_3) \cdot \mathcal{Z} &=
N_d^{\mathbb{P}^3,\mathrm{Planar}}(r,s,\theta)\; \nonumber \\
& +\sum_{r_1=0}^{r-3}\sum_{s_1=0}^s\sum_{d_1=1}^{d-1}\binom{r-3}{r_1}\binom{s}{s_1}d_1^3 d_2\times B_{d_1,d_2}(r_1,s_1,r_2-1,s_2,\theta),\label{wdvv1}
\end{align}
where in the above expression, $d_2:= d-d_1$, $r_2:= r-r_1$ and $s_2:= s-s_1$. \\
\hf \hf Next, we note that
\begin{align}
\pi^*(x_1 x_3|x_2 x_4) \cdot \mathcal{Z} &= 2d N_d^{\mathbb{P}^3,\mathrm{Planar}}(r-2,s+1,\theta)\; \nonumber \\
& +\sum_{r_1=0}^{r-3}\sum_{s_1=0}^s\sum_{d_1=1}^{d-1}\binom{r-3}{r_1}\binom{s}{s_1}d_1^2 d_2^2 \times B_{d_1,d_2}(r_1+1,s_1,r_2-2,s_2,\theta),
\label{wdvv2}
\end{align}
where as before $d_2:= d-d_1$, $r_2:= r-r_1$ and $s_2:= s-s_1$. \\
\hf \hf We now note that equations \eqref{wdvv_p1}, \eqref{wdvv1} and \eqref{wdvv2}, imply our desired recursive formula \eqref{recurfor}. \\
\hf \hf Next, we will justify the formula for $B_{d_1, d_2}(r_1, s_1, r_2, s_2,\theta)$ (equation \eqref{b1}).
This follows immediately from the fact that the class of the diagonal
is given by
\begin{align*}
\Delta_{\hat{\mathbb{P}}^3 \times \hat{\mathbb{P}}^3} &= \pi_1^* a^3 + \pi_1^*a^2 \cdot \pi_2^*a + \pi_1^*a \cdot \pi_2^*a^2 + \pi_2^* a^3,
\end{align*}
where $a$ denotes the generator of $H^*(\hat{\mathbb{P}}^3;\mathbb{Z})$ and $\pi_1, \pi_2$ denote the respective projection maps. The formula now follows immediately
from the definition of $$B_{d_1, d_2}(r_1, s_1, r_2, s_2,\theta)$$ (namely, equation \eqref{bd_delta}).\\
\hf \hf It remains to prove the two base cases of the recursion, namely Lemma \ref{base_formula} and Lemma \ref{second_case_formula}. \qed \\
\noindent \textbf{Proof of Lemma \ref{base_case_recursion}:} We recall that $\overline{\mathcal{M}}_{0,0}^{ \mathrm{Planar}}(\mathbb{P}^3, 1)$
is defined to be the projective bundle $\mathbb{P}(\gamma_{3,4}^*) \longrightarrow \mathbb{\hat{\mathbb{P}}}^3$. Now, we note that the
Chern classes of the rank three vector bundle $\gamma_{3,4}^* \longrightarrow \mathbb{\hat{\mathbb{P}}}^3$ are given by
\begin{align*}
c_i(\gamma_{3,4}^*) &= a^i \in H^{2i} (\mathbb{\hat{\mathbb{P}}}^3; \mathbb{Z}).
\end{align*}
The reason is explained in \cite[Page 18]{Zing_Notes}. Here $a$ is the standard generator of
$H^{*} (\mathbb{\hat{\mathbb{P}}}^3; \mathbb{Z})$.\\
\hf \hf Next, we note that $a^i =0$ if $i >3$. Hence, the cohomology ring of $H^*(\mathbb{P}(\gamma_{3,4}^*))$ is given by
\begin{align}
H^*(\mathbb{P}(\gamma_{3,4}^*)) &\approx \frac{\mathbb{Z}[a, \lambda]}{\langle \lambda^{3} + \lambda ^2 a + \lambda a^2 + a^3\rangle }
\end{align}
where $\tilde{\gamma}\longrightarrow \mathbb{P}(\gamma_{3,4}^*)$
is the tautological line bundle over the projectivized bundle $\mathbb{P}(\gamma_{3,4}^*)$
and $\lambda := c_1(\tilde{\gamma}^*) \in H^*(\mathbb{P}(\gamma_{3,4}^*))$. This follows from \cite[Page 270]{BoTu}.\\
\hf \hf We now note the following two important facts:
intersecting a generic line, corresponds to the cycle
\[ \mathcal{H}_l = \lambda + a.\]
Furthermore, passing through a generic point, corresponds to the cycle
\[ \mathcal{H}_p = \lambda a. \]
The reason for this can again be found in \cite[Pages 18 and 19]{Zing_Notes}. Hence,
to compute $N_1^{\mathbb{P}^3, \mathrm{Planar}}(r, s, \theta)$ we have to compute the expression
\[ (\lambda+a)^r (\lambda a)^s a^{\theta}, \]
use the relationship
\[ \lambda^3 = -(\lambda ^2 a + \lambda a^2 + a^3) \]
and extract the coefficient of $\lambda^2 a^3$. This gives us all the numbers for various values of $r, s$ and $\theta$. \qed \\
\noindent \textbf{Proof of Lemma \ref{second_case_recursion}:} First we note that every conic in $\mathbb{P}^3$ lies inside a unique plane
(except a double line).
Hence, let us consider the projective bundle
\[\mathbb{P}(\textnormal{Sym}^2(\gamma_{3,4}^*)) \longrightarrow \mathbb{\hat{\mathbb{P}}}^3.\]
This space $\mathbb{P}(\textnormal{Sym}^2(\gamma_{3,4}^*))$ is the space of conics in $\mathbb{P}^3$ and a plane
that contains the conic. The space of all smooth conics form an open dense subspace of $\mathbb{P}(\textnormal{Sym}^2(\gamma_{3,4}^*))$.
Hence, to compute the numbers $N_2^{\mathbb{P}^3, \mathrm{Planar}}(r,s, \theta)$ (which is defined as an intersection number on
$\overline{\mathcal{M}}^{\mathrm{Planar}}_{0,0}(\mathbb{P}^3,2)$), we can instead compute the relevant intersection number on
$\mathbb{P}(\textnormal{Sym}^2(\gamma_{3,4}^*))$.\\
\hf \hf Next, we note that $\mathbb{P}(\textnormal{Sym}^2(\gamma_{3,4}^*))$
is a $\mathbb{P}^5$ bundle over $\mathbb{\hat{\mathbb{P}}}^3$.
The cohomology ring structure of
the total space is given by
\begin{align}
\label{ring_conic}
H^*(\mathbb{P}(\textnormal{Sym}^2(\gamma_{3,4}^*)) &\approx \frac{\mathbb{Z}[a, \lambda]}{\langle \lambda^6 + 4 \lambda^5 a + 10 \lambda^4 a^2 + 20\lambda^3 a^3 \rangle}.
\end{align}
This follows from \textit{splitting principle} (see page $275$ in \cite{BoTu}).
We now note the following two important facts:
intersecting a generic line, corresponds to the cycle
\[ \mathcal{Z}_l = \lambda + 2a.\]
Furthermore, passing through a generic point, corresponds to the cycle
\[ \mathcal{Z}_p = \lambda a. \]
The reason for this can again be found in \cite[Pages 18 and 19]{Zing_Notes}.
Hence, to compute $N_2^{\mathbb{P}^3, \mathrm{Planar}}(r, s, \theta)$ we have to compute the expression
\[ (\lambda+2a)^r (\lambda a)^s a^{\theta}, \]
use the relationship given by the cohomology ring structure in \eqref{ring_conic}, i.e.
\[ \lambda^6 = -(4 \lambda^5 a + 10 \lambda^4 a^2 + 20\lambda^3 a^3) \]
and extract the coefficient of $\lambda^5 a^3$. This gives us all the numbers for various values of $r, s$ and $\theta$. \qed \\
\begin{comment}
\newpage
We explain briefly how we got these numbers as shown in table. Since $ d=1 $, we count the number of planar lines in $ \mathbb{P}^3 $ passing
through $ r $ number of lines and $ s(\leq 3) $ number of points such that $ r+2s+\theta=5 $. Let $ \overline{E}_{0,k}^1 = \mathbb{P}(\gamma_{3,4}^{*}) $ denotes
the space of planar lines in $ \mathbb{P}^3 $, this can also be thought of as a fiber
bundle over $ \hat{\mathbb{P}}^3$ ($ \pi:\overline{E}_{0,k}^1 \rightarrow \hat{\mathbb{P}}^3 $). Let $ \gamma_{\hat{\mathbb{P}}^3}^*\rightarrow \hat{\mathbb{P}}^3 $ denotes
the dual tautological line bundle over $ \hat{\mathbb{P}^3} $ and $ a:= c_1(\gamma_{\hat{\mathbb{P}}^3}^*) $ then $ \pi^{*}\hat{\gamma} $ is a tautological line bundle over $ \overline{E}_{0,k}^1 $ and $ c_1(\pi^{*}{\hat{\gamma}^*})=\lambda $. Then $ \lambda+a $ (see page $19$ of \cite{Zin1}) represents the condition that the line passes through a line and $ \lambda a $ represents the condition that the line passes through a point. So to compute the number $ N_1^{\mathcal{P}lanar}(r,s,\theta)$ we extract the coefficient of $ a^3\lambda^2 $ from the expression $ (\lambda+a)^r(\lambda a)^sa^{\theta} $.
\vspace{1cm}
Suppose $ d\geq 2 $. Let $u:\mathbb{P}^{1}\rightarrow \mathbb{P}^{3}$ be a rational map such that the image of $u$ lie inside some $\mathbb{P}_{\eta}^2\subset \mathbb{P}^3$. We think of a plane in $ \mathbb{P}^3 $ (which also corresponds to a point in $ \hat{\mathbb{P}}^3 $ ) as a non-zero linear functional $ \eta:\mathbb{C}^4\rightarrow \mathbb{C} $ upto scaling (i.e. we consider $ [\eta] $). Also, $ s:\mathbb{P}_{\alpha}^2\rightarrow \mathcal O(d) $ is a section, then $ s^{-1}(0) $ denotes a degree $ d $ curve in $\mathbb{P}_{\eta}^2$, which can be represented by $ X : \{([\eta],[s],p)\in \mathbb{P}(V)\times \mathbb{P}^3: \eta(p)=0,s(p)=0 \}$. We can express the space of genus zero planar maps in $ \mathbb{P}^3 $ as the projective fiber bundle over $\hat{\mathbb{P}}^3$ such that the fiber over each point in $\hat{\mathbb{P}}^3$ is $\overline{\mathcal{M}}_{0,n}(\mathbb{P}^2_{\eta}, d)$. ${\mathcal{M}}_{0,n}(\mathbb{P}^2_{\eta}, d)$ is a coarse moduli space which parametrizes (isomorphism classes of) Kontsevich stable $ n $-pointed degree $ d $ maps to $ \mathbb{P}^2_{\eta} $. For our case we take $n=3d+2$ where the required intersection will take place with the boundary divisors of $\overline{\mathcal{M}}_{0,{3d+2}}^{\mathcal{P}lanar}(\mathbb{P}^3,d)$, this is a moduli space which parametrizes the space of planar degree $ d $ rational curves with $ 3d+2 $ marked points upto an isomorphism and the dimension of this space is $ 6d+1 $.\footnote{see page 74 kock notes}\\
Next we concentrate for $d\geq 3$ case\\
Let $x_1, x_2, y_1, y_2\cdots y_{3d+2-2=3d}$ denotes the marked point on the curve $ \mathbb{P}^1 $. Next we fix two hyperplanes $H_1$ and $H_2$ and, $ r-1 $ lines $L_1,L_2,l_3, \cdots l_{r-1}$ and, $ s $ points $p_1,p_2\cdots p_s$ all lying in general position in $ \mathbb{P}^3 $ such that $ r+2s=3d+2 $. Consider the variety
$\overline{\mathcal{M}}_{0,{3d+2}}^{\mathcal{P}lanar}(\mathbb{P}^3,d)$. Let
\begin{equation}
S = ev_1^{*}{H_1} \cap ev_2^{*}{H_2} \cap ev_1^{*}{L_1} \cap ev_2^{*}{L_2} \cap \mathcal{H}_L^{r-3} \cap \mathcal{H}_p^{s} \cap a^{\theta}
\end{equation}
be the subvariety of maps from $ \mathbb{P}^1 $ which sends $ x_i $ to a point in $ H_i $ and $ y_i $ to a point in $ L_i , i=1,2$ and rest lie in the image. By counting the dimension we can show that $ S $ is one dimensional. Hence, the intersection of $ S $ with each pull-back divisor in $\overline{\mathcal{M}}_{0,{3d+2}}^{\mathcal{P}lanar}(\mathbb{P}^3,d)$ is a finite collection of points. We will count these points.\\
Let's Consider out of $r+2s=3d+2$ if we concentrate on four marked point and suppose the evaluation map $ev: \overline{\mathcal{M}}_{0,4}^{\mathcal{P}lanar}(\mathbb{P}^3,d) \rightarrow \overline{\mathcal{M}}_{0,4}.$ Here rest marked points will lie in the image. \\
Since $\overline{\mathbb{M}}_{0,4}$ is path connected so $D(x_1,x_2 \mid x_3,x_4) \sim D(x_1,x_3 \mid x_2,x_4)$ as divisors. Then $\pi^{*}{D(x_1,x_2 \mid x_3,x_4)}$ is equivalent to $\pi^{*}{D(x_1,x_3 \mid x_2,x_4)}$ as pull-divisors in $\overline{\mathcal{M}}_{0,4}^{\mathcal{P}lanar}(\mathbb{P}^3,d)$.\\
Then the above equivalence gives equality as numbers as
\begin{equation}
\begin{split}
\pi^{*}{D(x_1,x_2 \mid x_3,x_4)} \cap ev_1^{*}(H_1) \cap ev_1^{*}(H_2) \cap ev_1^{*}(L_1) \cap ev_1^{*}(L_2) \cap \mathcal{H}_L^r \cap \mathcal{H}_p^s \cap a^{\theta} \\
= \pi^{*}{D(x_1,x_3 \mid x_2,x_4)} \cap ev_1^{*}(H_1) \cap ev_1^{*}(H_2) \cap ev_1^{*}(L_1) \cap ev_1^{*}(L_2) \cap \mathcal{H}_L^r \cap \mathcal{H}_p^s \cap a^{\theta}.
\end{split}
\end{equation}
i.e, $S\cap \pi^{*}{D(x_1,x_2 \mid x_3,x_4)}= S\cap \pi^{*}{D(x_1,x_3 \mid x_2,x_4)}$.
Now we compute the intersection of $S$ with the boundary divisor on the left-hand side of the equation 3.2,
$$\pi^*D(x_1,x_2 |y_1,y_2) = \sum{D(1,2;d_1,d_2)}$$ such that $d= d_1 +d_2$. Let us compute the intersection of $S$ with each
one of the component of the summation, it has three types of breakup, i.e., when either of $d_1=0$ or $d_2 =0$ and $d_1=d$ or $d_2=d$ respectively and the non-trivial case when $d_1>0$ and $d_2>0$. We will compute these numbers using the non symmetric configuration as shown below:
\begin{tikzpicture}
\draw (0,2) circle (2cm);
\node[above left=50pt of {(3,1.5)}, outer sep=2pt,fill=white] {$x_1$};
\node[above right=15pt of {(.5,1)}, outer sep=2pt,fill=white] {$x_2$};
\node[above right=2pt of {(-5,1)}, outer sep=2pt,fill=white] {$H_1$};
\node[above left=5pt of {(-2,.3)}, outer sep=2pt,fill=white] {$H_2$};
\draw[<->] (-4,0)--(1,1.5);
\draw[<->] (-4,0)--(1,1.5);
\draw[<->] (-7,0)--(.5,2.5);
\draw (0,-2) circle (2cm);
\draw[<->] (3,-3)--(-1.5,-2);
\draw[<->] (4,0)--(-1,-1.5);
\node[below left=70pt of {(0,0)}, outer sep=2pt,fill=white] {$y_2$};
\node[below left=45pt of {(0,0)}, outer sep=2pt,fill=white] {$y_1$};
\node[above right=42pt of {(2,-1)}, outer sep=2pt,fill=white] {$l_1$};
\node[below right=95pt of {(1.5,0)}, outer sep=2pt,fill=white] {$l_2$};
\end{tikzpicture}
From the above picture we can see that our domain looks like a wedge of two spheres and we represent the upper sphere as degree $d_1$ component and
lower as degree $d_2$ component. Here $x_1$ hits the plane $H_1$, $x_2$ hits the plane $H_2$ and $y_1,y_2$ hit the lines $L_1, L_2$ respectively.
Observe that the term $d_2 = 0, d_1 = d$ does not contribute anything to the total sum. When $d_2=d,d_1 = 0$ we
get $ N_d^{\mathcal{P}lanar}(r,s,\theta)$. Now when $ d_1>0, d_2>0 $ such that $d=d_1 + d_2$, this contribute to
the sum when $3d_1-1$ generic lines considered as effect from first component. There are $r-3\choose r_1 $ $ s \choose s_1$ such
components having contribution $ B_{d_1,d_2}(r_1,s_1,r_2-1,s_2,\theta) $, this is given by \eqref{b1}. Now observe that $x_1$ has to
fall on the intersection of image of the $ d_1 $ component and the plane $H_1$ then by Bezout theorem there are $d_1$ such points.
Similarly, image of $x_2$ will intersect the plane $ H_2 $ at $d_1$ many points. Therefore, the first component must
go through $d_1^2$ points. Finally, the intersection point of images of $ d_1 $ component and $ d_2 $ component must
go through $d_1d_2$ points of intersection of the two (Bezout's theorem again). So we got one side of recursion formula by computing L.H.S of equation 3.2. It is written as
\begin{equation}\label{nonsym}
\begin{split}
N_d^{\mathcal{P}lanar}(r,s,\theta)\;+\sum_{r_1=0}^r\sum_{s_1=0}^s\sum_{d_1+d_2=d}\binom{r-3}{r_1}\binom{s}{s_1}d_1^3 d_2\\
\times B_{d_1,d_2}(r_1,s_1,r_2-1,s_2,\theta)
\end{split}
\end{equation}
Now let us compute the intersection of $ S $ with the boundary divisor on R.H.S of equation 4.2 using the symmetrical picture as shown below :
\begin{tikzpicture}
\draw (0,2) circle (2cm);
\node[above left=50pt of {(3,1.5)}, outer sep=2pt,fill=white] {$x_1$};
\node[above right=15pt of {(.5,1)}, outer sep=2pt,fill=white] {$y_1$};
\node[above right=2pt of {(-5,1)}, outer sep=2pt,fill=white] {$l_1$};
\node[above left=5pt of {(-2,.3)}, outer sep=2pt,fill=white] {$H_1$};
\draw[<->] (-4,0)--(1,1.5);
\draw[<->] (-4,0)--(1,1.5);
\draw[<->] (-7,0)--(.5,2.5);
\draw (0,-2) circle (2cm);
\draw[<->] (3,-3)--(-1.5,-2);
\draw[<->] (4,0)--(-1,-1.5);
\node[below left=70pt of {(0,0)}, outer sep=2pt,fill=white] {$x_2$};
\node[below left=45pt of {(0,0)}, outer sep=2pt,fill=white] {$y_2$};
\node[above right=42pt of {(2,-1)}, outer sep=2pt,fill=white] {$H_2$};
\node[below right=95pt of {(1.5,0)}, outer sep=2pt,fill=white] {$l_2$};
\end{tikzpicture}\\
Observe that the contribution to the sum when either $d_1=0,d_2=d$ or $d_1=d,d_2=0$ is counted two times, i.e., we have $2N_d^{\mathcal{P}lanar}(r-2,s+1,\theta)$ with
the multiplication of $d$. This is due to the fact that the image of each component intersect a generic hyperplane in $d$ many points (Bezout's theorem).
Now when $ d_1>0,d_2>0 $ satisfying $d=d_1+d_2$, this configuration contributes to the sum only when the image curve of first
component passes through $3d_1+2$ generic lines and there are $r-3 \choose r_1$ $s \choose s_1$ such
components in $\pi^*D(x_1,y_1 |x_2,y_2)$. Now for the above choice there are $B_{d_1,d_2}(r_1+1,s_1,r_2-1,s_2,\theta)$ ways to put
this curve, which is given by \eqref{b1}. Also, the marked point $y_1$ must map to the intersection of image of first component
with $H_1$ giving us $d_1$ many choices. Similarly, from the second component we get $d_2$ many choices and the marked point belonging
to the intersection of images of these two components gives us $d_1d_2$ many choices (Bezout's theorem). Hence, the total contribution of R.H.S of equation 3.2 is given by
\begin{equation}
\begin{split}
2d N_d^{\mathcal{P}lanar}(r-2,s+1,\theta)\;+\sum_{r_1=0}^r\sum_{s_1=0}^s\sum_{d_1+d_2=d}\binom{r-3}{r_1}\binom{s}{s_1}d_1^2 d_2^2\\
\times B_{d_1,d_2}(r_1+1,s_1,r_2-2,s_2,\theta)
\end{split}
\end{equation}
Now the fundamental linear equivalence of divisors equation 3.2 gives us the required recursion formula.
\end{comment}
\section{Low degree checks}
\label{low_deg_check}
We now describe concrete low degree checks that we have performed. Using our recursive formula, we have obtained the following numbers:
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
$(d,r,s)$ & $(3,11,0)$ & $(4,14,0)$ & $(5,17,0)$ & $(6,20,0)$ \\
\hline
$N_d^{\mathbb{P}^3, \mathrm{Planar}}(r,s)$ & $12960$ & $3727920$ & $1979329280$ & $1763519463360$ \\
\hline
\end{tabular}
\end{center}
Next, let $N^{\textnormal{Node}, \delta}_d(r,s)$ denote the number of planar degree $d$ curves in $\mathbb{P}^3$ with $\delta$ (unordered) nodes
intersecting $r$ generic lines and $s$ generic points. These numbers are computed in \cite{TL}. Using that, we get the following table
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
$(d,r,s, \delta)$ & $(3,11,0,1)$ & $(4,14,0,3)$ & $(5,17,0,6)$ & $(6,20,0,10)$ \\
\hline
$N_d^{\mathrm{Node}, \delta}(r,s)$ & $12960$ & $4057340$ & $2487128120$ & $2681467886460$ \\
\hline
\end{tabular}
\end{center}
Finally, let us denote the by $\textnormal{Red}^{\textnormal{Node}, \delta}_d(r,s)$ to be the number of reducible
planar degree $d$ curves in $\mathbb{P}^3$ with $\delta$ (unordered) nodes
intersecting $r$ generic lines and $s$ generic points.
This number can be computed using \cite[Proposition $8.4$]{TL}. Using that, we get
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
$(d,r,s, \delta)$ & $(3,11,0, 1)$ & $(4,14,0,3)$ & $(5,17,0,6)$ & $(6,20,0,10)$ \\
\hline
$\textnormal{Red}_d^{\mathrm{Node}, \delta}(r,s)$ & $0$ & $329420$ & $507798840$ & $917948423100$ \\
\hline
\end{tabular}
\end{center}
We now note that in all the cases we have tabulated,
\[N_d^{\mathbb{P}^3, \mathrm{Planar}}(r,s) = N_d^{\mathrm{Node}, \frac{(d-1)(d-2)}{2}}(r,s) - \textnormal{Red}_d^{\mathrm{Node}, \frac{(d-1)(d-2)}{2}}(r,s).\]
This is positive evidence for the fact that T.~Laarakker's formula for $N^{\textnormal{Node}, \delta}_d(r,s)$
is actually enumerative when $d \geq 1 + [\frac{\delta}{2}]$ (as opposed to $d \geq \delta$, which is proved in \cite{TL}).
We also note that when $d=7$ and $\delta = \frac{(d-1)(d-2)}{2} = 15$, the formula for $N_d^{\mathrm{Node}, \frac{(d-1)(d-2)}{2}}(r,s)$
is not expected to be enumerative because of an obvious geometric reasons. To see why, suppose $s=0$ and $r = 23$.
Through the required $r=23$ lines, we can place a double line
through $4$ lines and through the remaining $19$ lines we can place a quintic that intersects the line so that the double line and the quintic
lie in a plane. This is a degenerate configuration, and hence $N_7^{\mathrm{Node}, 15}(23, 0)$ is not expected to be enumerative.\\
\hf \hf This is analogous to the case of counting $\delta$-nodal degree $d$ curves in $\mathbb{P}^2$; let $N^{\delta}_d$ denote that number.
A formula for this number can be
explicitly found in \cite{FB} for instance. On the
other hand, let $N_d^{\mathbb{P}^2}$ denote the number of rational degree $d$ curves in $\mathbb{P}^2$ through $3d-1$ generic points.
Till $d=6$, we can verify that $N_d^{\mathbb{P}^2}$ is logically consistent with the corresponding value of $N^{\frac{(d-1)(d-2)}{2}}_d$ (after subtracting
the number of irreducible curves). From $d=7$, we can not make any such consistency check, because $N^{15}_7$ is not enumerative; there are double
lines that can pass through two of the $20$ points and a quintic through the remaining $18$ points. We also note that this fact is consistent with the
{G}\"ottsche threshold of when the number $N^{\delta}_{d}$ is supposed to be enumerative; this is proved in \cite{KlSh}.
Our computations give evidence
to show that a similar bound is likely to be true for the case of planar curves in $\mathbb{P}^3$. \\
\section{Acknowledgement}
The first author is grateful to Martijn Kool and Ties Laarakker for a very fruitful exchange of ideas and discussions
related to the parallel question of counting $\delta$-nodal planar curves in $\mathbb{P}^3$, that is investigated in
\cite{TL}. We are also very grateful to ICTS for their hospitality and conducive atmosphere for doing mathematics research.
We would like to specially acknowledge two programs conducted by ICTS where a significant part of the project was carried out:
Analytic and Algebraic Geometry (Code: ICTS/Prog-AAG2018/03) and
Integrable Systems in Mathematics, Condensed Matter and Statistical Physics (Code: ICTS/integrability2018/07). The first author
would also like to thank Center for Quantum Geometry of Moduli Space at Aarhus, Denmark (DNRF95) for giving him a chance to spend $6$ weeks
there, when the author got the initial idea for this project; the visit was
was mainly paid by the grant ``EU-IRSES Fellowship within FP7/2007-2013 under grant
agreement number 612534, project MODULI - Indo European Collaboration on Moduli Spaces."
Finally, the first author would like to acknowledge the External Grant
he has obtained, namely
MATRICS (File number: MTR/2017/000439) that has been sanctioned by the Science and Research Board (SERB).
|
\section{Introduction} \label{introduction}
The relationship between super-massive black holes (SMBHs) and their host galaxies is an active area of research \citep[see, e.g.][]{Kormendy2013, Hickox2014, Delvecchio2015, Bongiorno2016, Yang2018}. Several empirical correlations between the mass of a black hole (BH) and the physical properties of its host galaxy have been reported \citep[e.g.][]{Magorrian1998, Merloni2003, Merritt2006, Kormendy2009}. Of these correlations the BH mass-stellar velocity dispersion (the $\mathrm{M_{\bullet}} - \sigma$ relation) \citep[e.g][]{Ferrarese2000, Gebhardt2000, Mcconnell2011} historically gave the first clues on a feedback driven co-evolution of BHs and their host galaxies \citep{Silk1998, King2003}.
Theoretical and observational studies suggest that major mergers play a fundamental role in establishing feedback and feeding cycles \citep[e.g.][]{DiMatteo2005, Hopkins2006, Kormendy2011}. Further support for the importance of mergers comes from the increased scatter in the BH-host correlations at larger redshifts which is a natural consequence of the central-limit theorem and an increasing number of BH mergers for SMBH toward low redshifts \citep{Schawinski2006, Peng2006, Hirschmann2010}. However, mergers are not the only physical processes involved; offsets in the $\mathrm{M_{\bullet}}-\mathrm{M_{Bulge}}$ relation corresponding to disc galaxies can be explained through the co-evolution of SMBHs with their disc-galaxy hosts through secular processes \citep{Volonteri2016, Simmons2017, Martin2018}.
SMBHs with masses of $M_{\bullet}\sim 10^{9}\, \mathrm{M_{\odot}}$ have been observed in galaxies at high redshifts ($z\sim$ 6 -- 7) \citep{Fan2006a, Mortlock2011, Banados2018} corresponding to less than a gigayear after The Big Bang. This population of BHs must have had a rapid formation process to reach the masses observed at this early epoch. Indeed, if the growth rate of SMBHs is limited by the Eddington accretion rate \citep[see however,][]{Natarajan2012,Pacucci2017}, they must be seeded by some massive progenitor at an early epoch $z \geq 10$, prior to the onset of reionization and the shutdown of Population-III stars \citep{Paardekooper2015,Johnson2013a}.
Given the e-folding nature of an Eddington limited BH-growth rate on a Salpeter time scale of $t_{\mathrm{Sal}} = 0.45 \, \eta/(1 - \eta)\; \mathrm{Gyr}$ (where $\eta \sim 0.1$ is the radiative efficiency \citep[see, e.g.][]{King2008}), varying the initial seed mass by factors of ten can have a strong impact on relaxing the constraints on the formation time in the early Universe. Consequently, various seed formation processes are discussed \citep[see, e.g. for a review][]{Volonteri2010}, including population III stellar remnants \citep{Madau2001} and the collapse of dense stellar clusters \citep{Clark2008,Yajima2016} or the direct collapse of gas through the state of a super-massive star \citep{Bromm2003,Lodato2006,Begelman2006,Begelman2010,Agarwal2012b}. The latter channel has received heightened attention due to the massive seeds it produces and the ability to grow to super-massive scales with less stringent constraints on the average accretion rate \citep{Agarwal2012b, Latif2013, Latif2014, Pacucci2015, Agarwal2016a}. \footnote{\citet{Johnson2011} simulated the radiative feedback from such a seed BH showing that the average accretion rate is very low, indicating that the feedback might off-set the advantage you gain of having a higher initial mass.}
Direct collapse BHs (DCBHs) form during the collapse of pristine gas in haloes with virial temperatures of $T_{\rm v}\gtrsim 10^{4}$ K \citep{Bromm2003}. Provided a halo remains pristine and the local intensity of the Lyman-Werner radiation field is greater than the critical intensity required to dissociate any H$_{2}$ gas \citep{Agarwal2016}, cooling within the halo will only take place via atomic hydrogen. The gas temperature in such a halo will be kept at $T_{\rm g}\sim 10^{4}$ K during collapse with a Jeans Mass of $M_{\rm J}\sim 10^{6} \, \mathrm{M_{\odot}}$; preventing the fragmentation into gas clumps and stars, and leading to the isothermal collapse of a massive gas cloud into a single BH \citep{Bromm2003}, possibly via an intermediate stage of a super-massive star \citep{Begelman2010}. This process results in the formation of massive seed BHs with masses of $M_{\bullet}\sim 10^{4}-10^{6}\, \mathrm{M_{\odot}}$ at $z\sim 10 - 20$, prior to the formation of the host galaxy in the halo \citep{Agarwal2012b}. If SMBHs are truly seeded by DCBHs it would affect the early stages of galaxy evolution. Gas build up around the gravitational potential well of a DCBH through cosmological accretion and halo merging, would not only lead to further BH growth but also potentially to the gradual growth of a proto-galaxy around the BH. Besides feedback from the BH affecting the proto-galaxy, initially such a proto-galaxy would be gravitationally dominated by the mass of the BH as well. However, it is not clear how this would affect the processes of galaxy evolution, such as star formation, and the cycle of baryons in the galaxy.
Recently, a first potential candidate for an observed DCBH system has been proposed (\citealt{Sobral2015}; but see \citealt{Bowler2017}). The system, called CR7, is a very bright Ly $\alpha$ emitter at $z=6.6$ with $L_{\mathrm{Ly}\alpha} \sim 10^{44} \mathrm{erg \,s^{-1}}$ \citep{Matthee2015, Sobral2015, Bowler2017}. \citet{Sobral2015} have identified CR7 as a combination of three components: Two clumps which appear to be evolved galaxies in close proximity to a third clump, which provides the vast majority of the Ly $\alpha$ flux. This third clump has been successfully modelled by \citet{Agarwal2016a} as a $M_{\bullet} \sim 4.4 \times 10^6 \, \mathrm{M}_{\odot}$ BH formed through direct collapse around $z\sim20$. Recent work has shown either an active galactic nucleus (AGN) or a young starburst population are also likely explanations for the observed characteristics of CR7 \citep{Bowler2017}. With the former being potentially seeded via the stage of a DCBH, and the latter not requiring a DCBH at all.
The evolution during the initial stages of a potential DCBH-systems such as CR7 is unknown and yet likely consists of a constant interplay between star formation and BH growth. The formation of stars in proto-galaxies is driven by the accretion of gas and subsequent gravitational collapse. The star formation law relating the star formation rate (SFR) surface density in a disc galaxy to its gas surface density \citep{Schmidt1959}, once confirmed by observations of local galaxies \citep{Kennicutt1998}, has more recently been shown to extend to $z\sim1.5$ \citep[see, e.g.][]{Carilli2013}. At higher redshifts, the higher densities imply shorter cooling times (as $t_{\rm cool} \propto \rho^{-1}$). Rapid cooling means that the SFR is only limited by the total gas accretion rate and the growth rate of gravitational instabilities in a galaxy \citep{Dekel2009,Dekel2013}. Previous studies on galaxy evolution and star formation have related the global SFR to disc properties via the growth rate of instabilities \citep{Fall1980, Lacey1983, Wang1994, Schaye2004, Elmegreen2010}. For example, \citet{Elmegreen2010} modeled the growth in gas mass and turbulence driven by gas accretion onto a galaxy and found the SFR was mainly a function of the gas accretion rate. However, star formation is a local process. Star formation can only take place where the gas is unstable to gravitational collapse \citep{Wang1994}. In this context the BH may also play an important role in the stability of the disc \citep{Lodato2012}. The gas properties will change throughout the galaxy with some regions being more unstable than others. Indeed, \citet{Schaye2004} found that if disc galaxies are rotational supported against collapse this will be particularly at large radii, limiting the radial extent of star formation to within some truncation radius.
A further complicating factor for the growth of proto-galaxies around DCBHs is that the hosting halo is generally in the vicinity of a more massive halo it is likely to merge with at a later stage during its evolution \citep{Agarwal2014a}. During the satellite-phase the provision of fuel for star formation will cease due to stripping processes in the environment \citep{VanDenBosch2008}. The growth of the host galaxy will thus be affected and in turn the path to the locally observed BH-galaxy correlations.
The aim of this paper is two-fold, we want to model the stabilising effect of DCBHs on the gaseous disc in proto-galaxies and their impact on the onset of star formation, and based on these models present arguments for the evolution of DCBHs toward locally observed correlations with host galaxies. First, we lay out the star formation model we use which relates star formation rate to disc instabilities (section~\ref{starform}). The model is first introduced by discussing a non-evolving case in section~\ref{Non-evolving} before being fully explored in section~\ref{Evolving} with evolving the halo and stellar mass. Finally, we discuss the implications of the model for massive seed hosting galaxies and the onset of star formation within them (section~\ref{conclusions}). Throughout the paper a $\Lambda$CDM Universe is assumed with $\rm H_{0} = 70 km\, s^{-1}\, Mpc^{-1}$, $\Omega_{\mathrm{m, \, 0}} = 0.27$ and $\Omega_{\mathrm{\Lambda, \, 0}} = 0.73$.
\section{Star Formation} \label{starform}
The empirical star formation law derived from local observation \citep{Kennicutt1998},
\begin{equation}
\dot{\Sigma}_{\star}^{\rm KS}(t) = 1.515 \times 10^{-4} \, \mathrm{M_{\odot} \, yr^{-1} \,kpc^{-2}} \left(\Sigma_{\mathrm{g}}/1 \, \mathrm{M_{\odot} \, pc^{-2}}\right)^{1.4}
\label{eq:sf10}
\end{equation}
\noindent where the amplitude has been adjusted to fit with a \citet{Chabrier2003} IMF \citep{Schaye2010}, has been seen to hold to high redshift \citep[see, e.g.][and references therein]{Carilli2013}. Furthermore, the relation appears to hold for both local surface density values and those integrated over an aperture \citep{Kennicutt1998}. One can understand this relation between star formation rate surface density and gas density using a star formation timescale \citep[see, e.g.][]{Wang1994,Kennicutt1998,Krumholz2007,Elmegreen2010}. The local dynamical or free-fall time within a star forming region is often used to relate the timescale to the gas density while the different mechanisms that would work against gravitational collapse, such as thermal and rotational support, are factored in either explicitly \citep{Wang1994,Elmegreen2010} or as part of an efficiency parameter \citep{Krumholz2007}.
In contrast to global models of star formation in proto-galactic discs \citep[e.g.][]{Elmegreen2010} we are here focusing on the radial star formation profile, which depends on local gravitational instabilities in the disc and allows the investigation of the impact of massive seed BHs.
The Ansatz for the star formation model we use is: (1) Star formation can only take place above a minimum threshold $\Sigma_{\mathrm{g}} > \Sigma_{\mathrm{th}} = 10.0 \, \mathrm{M_{\odot} \, pc^{-2}}$ \citep{Schaye2004}. (2) No star formation will take place if the disc is locally stabilised against gravitational collapse $Q_{\rm Toomre} \geq 1$ or (3) if the local density is too low to overcome tidal forces $Q_{\rm tidal} \geq 1$ (see section~\ref{modelsetup}).
If these conditions for star formation are met, the SFR surface density is calculated by relating the timescale for star formation to the maximal growth rate of axisymmetric perturbations, $\omega_{\rm WS}$ \citep{Wang1994}:
\begin{equation}
t_{\rm SF} = \frac{1}{\omega_{\rm WS}} = \frac{Q_{\rm Toomre}}{\kappa \sqrt{1 - Q_{\rm Toomre}^{2}}}
\label{eq:sf8}
\end{equation}
\noindent where $\kappa$ is the epicyclic frequency and $Q_{\rm Toomre}$ is the Toomre disc instability parameter (see section~\ref{Non-evolving}). The SFR surface density $\dot{\Sigma}_{\star}$ can then be written as a function of this timescale, the gas surface density profile $\Sigma_{\mathrm{g}}$, and the star formation efficiency parameter $\epsilon$, to obtain the following \citep{Wang1994}:
\begin{equation}
\begin{aligned}
\dot{\Sigma}_{\star}^{\rm WS}(t) = \epsilon \frac{\Sigma_{\mathrm{g}}}{t_{\rm SF}} &= \epsilon \frac{ \kappa}{Q_{\rm Toomre}} \Sigma_{\mathrm{g}} \left(1 - Q_{\rm Toomre}^{2}\right)^{0.5} \\
&= \epsilon \frac{ \pi G \Sigma_{\mathrm{d}}}{\sigma} \Sigma_{\mathrm{g}} \left(1 - Q_{\rm Toomre}^{2}\right)^{0.5}
\label{eq:sf9}
\end{aligned}
\end{equation}
\noindent where $\sigma$ is the velocity dispersion (from here on taken to be the sound speed, $c_s$) and the $\Sigma_{\rm d} =\Sigma_{\mathrm{g}} + \Sigma_{\star}$ is the total surface density of the disc. This formulation includes explicitly the effects of rotation on the growth rate of instabilities in the disc. The rotation of the disc provides support against gravitational instabilities, preventing the collapse of gas to form stars \citep{Fall1980, Lacey1983}. This then relates the SFR to the growth rate of gravitational instabilities rather than simply the free-fall timescale and allows one to take the structure of the disc in to account. \footnote{Equation~\ref{eq:sf9} will result in a steeper $\dot{\Sigma}_{\star} - \Sigma_{\mathrm{g}}$ relation ($\dot{\Sigma}_{\star} \propto \Sigma_{\mathrm{g}}^{2}$) than the \citet{Kennicutt1998} law (equation~\ref{eq:sf10}) and is therefore likely to overestimate the total SFR, providing a conservative, upper-limit estimate for our model which we are seeking for in this study.}
\section{Non-evolving, non-star forming case} \label{Non-evolving}
\subsection{Model Setup} \label{modelsetup}
This first case looks at the effect of the BH in a non-evolving gaseous disc (The fully evolving case is addressed in section~\ref{Evolving}). The mass of the halo is kept constant, $\dot{M}_{\rm 200} = 0$, and the stellar mass is zero throughout, $M_{\star}=0$.
To investigate the effects of a seed BH on a galaxy forming in its host halo it is necessary to probe the inner region of the proto-galactic disc. For this reason we take the radial dependencies of system's properties into account. For the non-evolving case we have a purely gaseous disc embedded in a halo we model simply as an isothermal sphere\footnote{Whether the inner density profile of a dark matter halo in this scenario should be less steep is not clear due to the effects of baryons on the halo \citep[see, e.g.][]{Davis2014}.} (see Section~\ref{stability} for a discussion of the implications of this). We assume the gas disc has an exponential profile centred on the BH such that the surface density of gas goes with radius as,
\begin{equation}
\Sigma_{\mathrm{g}}(R) = \Sigma_{g,\,0} \exp(-R/R_{\mathrm{d}})
\label{eq:sigprofile}
\end{equation}
\noindent $R_{\rm d}$ is the disc scale radius which is set by the halo parameters and $\Sigma_{g,\,0} = M_{\rm g} / (2 \pi R_{\rm d}^2 )$. As we assume the halo is an isothermal sphere, the disc scale radius is calculated using the following \citep{Mo1998}:
\begin{equation}
R_{\mathrm{d}} = \frac{1}{\sqrt{2}} \left (\frac{j_{\rm d}}{m_{\rm d}} \right) \,\lambda\, r_{\rm 200}
\label{eq:Rd}
\end{equation}
\noindent Throughout this study $j_{\rm d}/m_{\rm d}$ is assumed to be unity and the spin parameter is taken from the log-normal distribution used by \citet{Mo1998} with the first and second moments: $\bar{\lambda}=0.05$ and $\sigma_{\lambda}$ respectively. For simplicity the first moment of the distribution is used as our fiducial value for $\lambda$ unless otherwise stated, though it is important to note changing $\lambda$ will have an effect on the model. For example, taking $\lambda$ at the 10\% point of the distribution roughly halves the disc scale radius which, for the same disc mass, doubles the surface density, $\Sigma_{g,\,0}$. For further discussion in the context of the model see Appendix~\ref{app:rd}.
The local stability of the disc against gravitational collapse is parametrised by the Toomre stability parameter \citep{Toomre1964}. If we have a disc which has both a stellar and gas component and assume the velocity dispersion of each component is such that $\sigma \equiv \sigma_{\mathrm{g}} = \sigma_{\mathrm{s}}$, the Toomre parameter becomes:
\begin{equation}
Q_{\rm Toomre} = \frac{\kappa\,\sigma}{\pi\,G\,\Sigma_{\mathrm{d}}}
\label{eq:toomre}{}
\end{equation}
\noindent where $\kappa$ is the epicyclic frequency and $\sigma$ is the velocity dispersion \citep{Toomre1964, Wang1994, Romeo2011}. If $Q_{\rm Toomre}>1$ the disc is stable to gravitational instabilities; $Q_{\rm Toomre}<1$ the disc is unstable and $Q_{\rm Toomre}\sim1$ the disc is partially stable \citep[see, e.g.][]{Lodato2007}. The velocity dispersion is taken as the sound speed of the gas, $c_{\rm s}$. This acts as a lower limit as we do not take turbulent motions into account, however at $T_{\rm g}= 8000$ K, the sound speed should provide a significant fraction of the gas velocity dispersion\footnote{With $\sigma = c_{\rm s}$, Equation~\ref{eq:toomre} is similar to $Q_{\rm Thermal}$ as described recently by \citet{Stark2018}.}. The epicyclic frequency describes the rotational support of the system due to the gravitational potential. It can be expressed as a function of the angular velocity, $\Omega = V_{\rm c} / R$ (where $V_c$ is the circular velocity which is calculated from the radial derivative of the gravitational potential):
\begin{equation}
\kappa^{2} = \frac{2 \,\Omega}{R}\frac{d}{dR}(R^{2}\Omega)
\label{eq:kappa}
\end{equation}
\noindent This is therefore a function of the three components of the system: the halo, disc and BH. As the potential due to these components can be combined to find the full potential, $\kappa$ can be split into the corresponding parts. The relative importance of the component of $\kappa$ due to the BH will increase with proximity to the BH.
Similarly to the Toomre parameter, the critical tidal density, $\rho_{\mathrm{tidal}}$, defines the limit to the local density of the disc below which the local self gravity of the disc is weak compared to the tidal forces on the disc \citep{Hunter2001,Martig2009}.
\begin{equation}
\rho_{\mathrm{tidal}}=\frac{3 \, \Omega \, R}{2 \pi G}\abs*{\frac{d\Omega}{dR}}
\label{eq:rhotidal}
\end{equation}
Dominant tidal forces inhibit the growth of density perturbations, preventing stars from forming. To compare the tidal and Toomre stability of the disc we need to make a comparison between the critical tidal density and the Toomre parameter. We define the critical tidal surface density as $\Sigma_{\mathrm{tidal}} = 2 H \rho_{\mathrm{tidal}}$ and using the scale height of the disc, $H = \sigma^2/(\pi\,G\,\Sigma_{\mathrm{d}})$, we obtain:
\begin{equation}
\Sigma_{\mathrm{tidal}}=3 \Omega R \frac{\sigma^2}{\pi^2 G^2 \Sigma_{\mathrm{d}}}\abs*{\frac{d\Omega}{dR}}.
\label{eq:sdtidal}
\end{equation}
\noindent With some rearranging we can see the square root of the ratio of the critical tidal surface density to the local surface density is of a similar form to the Toomre parameter.
\begin{equation}
Q_{\rm tidal} \equiv \sqrt{\frac{\Sigma_{\mathrm{tidal}}}{\Sigma_d}}=\frac{\sigma}{\pi G \Sigma_{\mathrm{d}}}\sqrt{3 \Omega R\abs*{\frac{d\Omega}{dR}}}. = Q_{\rm Toomre} \frac{\nu}{\kappa}
\label{eq:sdratio}
\end{equation}
\noindent where we define the tidal frequency $\nu$,
\begin{equation}
\nu^2 = 3 \Omega R\abs*{\frac{d\Omega}{dR}}
\label{eq:tidalfreq}
\end{equation}
\noindent It follows that $Q_{\rm tidal}$ will behave similarly to the Toomre parameter; if $Q_{\rm tidal} < 1$ it implies $\rho > \rho_{\mathrm{tidal}}$ and the disc's local self gravity dominates but if $Q_{\rm tidal} > 1$, $\rho < \rho_{\mathrm{tidal}}$ and the disc is locally unstable to tidal forces, and the growth of gravitational instabilities locally in the disc is inhibited.
We combine the two stability parameters by defining $Q_*$ as the maximum of the two:
\begin{equation}
Q_*=\mathrm{max}[Q_{\rm tidal},\;Q_{\rm Toomre}]
\label{eq:Qmax}
\end{equation}
\noindent this reduces our conditions for star formation down to two: that the surface density is above the threshold ($\Sigma_{\mathrm{g}} > \Sigma_{\mathrm{th}} = 10.0 \, \mathrm{M_{\odot} \, pc^{-2}}$) and that $Q_*>1$.
\begin{table}
\centering
\caption{Table of non-evolving, non-star forming model parameters.}
\label{tab:nonevopara}
\begin{tabular}[h]{l|l|l}
\hline
Parameter & Definition & Fiducial\\
\hline
$M_{200}$ & dark matter halo mass ($\mathrm{M_{\odot}}$) & $5\times10^{8}$ \\
$\lambda$ & halo spin parameter& 0.05\\
$j_{\rm d}/m_{\rm d}$ & disc and halo specific momenta ratio& 1.0\\
$f_{\rm b}$ & baryon fraction & 0.17\\
$T_{\rm g}$ & gas temperature (K)& 8000\\
$z$ & redshift& 10.0\\
\end{tabular}
\end{table}
\subsection{Stability Parameters Profiles} \label{stability}
For our analysis here we use a fiducial model of an atomic cooling halo at $z\sim10$ (see Table~\ref{tab:nonevopara}). The total mass of the system was calculated for an atomic cooling halo $M_{\rm total}\sim M(T_{\rm vir}=8000\,K)$ \citep{Mo2010} and the disc mass was calculated by taking the baryonic mass of the halo, $M_{\rm d} = f_{\rm b} M_{\rm total}$ where $f_{\rm b}=0.17$ is the universal baryon fraction, unless otherwise stated. One would expect only a fraction of the baryonic mass of the halo, $p\,f_{\rm b}$ (where $p<1$), to reach the disc \citep[see, e.g.][]{Dekel2013a, Dekel2013} and therefore, $f_{\rm b} M_{\rm total}$ is an upper limit on the disc mass. Taking this upper limit allows us to look at the most unstable case as lowering the disc mass would increase the disc stability. For our model, decreasing $p$ would have the same effect as decreasing the overall baryon fraction. Section~\ref{minQ} discusses how lowering the baryon fraction would change the stability profile of the disc for different BH and disc masses.
Figure~\ref{fig:Qprofile} shows the radial profiles for the Toomre and tidal stability parameters respectively for our fiducial set-up as summarized in table \ref{tab:nonevopara}. If we first look at the upper panel of Figure~\ref{fig:Qprofile}, the increase in $\kappa$ at small radii due to the presence of the BH stabilises the inner-most region of the disc, shown by the increase in $Q_{\rm Toomre}$ at small radii. Increasing the BH mass increases this effect, narrowing the region of the disc where star formation can take place. For a constant disc mass an accreting BH thus would be able to prevent a larger fraction of the disc from forming stars as it grows in mass. At larger radii the influence of the BH diminishes and the disc determines the shape of the stability profiles except for cases with the largest BH masses. After reaching a minimum both of the stability parameters increase as the disc surface density decreases with radius.
The tidal stability parameter profile (the lower panel of Figure~\ref{fig:Qprofile}) shows how the BH also has a strong tidal effect on the disc at small radii. The tidal and Toomre parameter profiles are similar in shape however, the tidal parameter appears to be below the critical value of 1 over a narrower range in radius. This suggests $Q_{\mathrm{tidal}} \leq 1$ is a stricter condition for star formation in the disc within the model than simply $Q_{\rm Toomre} \leq 1$.
\begin{figure}
\includegraphics[width=\columnwidth]{qtoti_plot}
\caption{The stability parameters profiles for the same disc without a BH and for four different mass BHs. The top panel shows the Toomre stability parameter profile and the bottom shows the tidal stability parameter profile. The disc is the same in all cases with $M_{\rm d} = 1.02\times10^8 \, \mathrm{M_{\odot}}$ and $R_{\rm d} = 86.5 \, \mathrm{pc}$.}
\label{fig:Qprofile}
\end{figure}
For a constant disc mass the inner critical radius, the inner-most radius where star formation can occur (where both $Q_{\rm Toomre}$ and $Q_{\rm tidal}$ are $\leq 1$), increases with BH mass (Figure~\ref{fig:RvsMbh}). The inner critical radius, $R_{\rm c, \, in}$, is smallest when we have no BH at $R_{\rm c, \, in} = 27.3 \, \mathrm{pc}$ and is $R_{\rm c, \, in} = 32.3 \, \mathrm{pc}$ for $M_{\bullet} = 10^6 \, \mathrm{M_{\odot}}$. If we increase the BH mass enough we reach a point where the whole disc becomes stabilised (For example, the yellow $M_{\bullet} = 10^8 \, \mathrm{M_{\odot}}$ BH case in Figure~\ref{fig:Qprofile}). The point where the disc becomes completely stabilised is represented in Figure~\ref{fig:RvsMbh} by the point where the $R_{\rm c, \, in}$ lines end around $M_{\bullet} \sim 10^7 \, \mathrm{M_{\odot}}$ which is less than a tenth of the disc mass ($M_{\rm d} = 1.02\times10^8 \, \mathrm{M_{\odot}}$).
It can be useful to compare this inner critical radius to the radius of the sphere of influence of the BH, $R_{\bullet}$. We calculate this from $M_{\bullet}$ and the circular velocity profile of the system, $V_c(R)$, using:
\begin{equation}
R_{\bullet} = \frac{G M_{\bullet}}{V_c^2}
\label{eq:rbullet}
\end{equation}
\noindent Due to the radial dependence of $V_c(R)$ it is necessary to solve Equation~\ref{eq:rbullet} iteratively such that $V_c=V_c(R_{\bullet})$.
As we increase the BH mass towards this disc-stabilising value, $R_{\rm c, \, in}$ and $R_{\rm c, \, out}$ are of the same order as $R_{\rm d}$, with $R_{\rm c, \, in}$ roughly 2.5 times $R_{\bullet}$.
The radius at which either stability parameter is minimized is always of the order of the disc scale radius (see Figure~\ref{fig:Rqmin} and the following section). Starting at BH mass fractions of $10\%$ the radius at which the tidal stability parameter is minimized quickly catches up with the scale radius of the disc. This helps to explain why $R_{\rm c, \, in}$ approaches $R_{\rm d}$ as the disc becomes stabilised as we increase the BH mass.
As the disc approaches stability the minimum value of $Q_{\rm tidal}$ approaches 1 until the BH mass is sufficiently massive to fully stabilise the disc ($M_{\bullet}/M_{\rm d}\sim 10\%$) and $R_{\rm c, \, in} = R_{\rm c, \, out} = R_{\rm Q_{min}}$. At lower BH masses, the inner critical radius is generally less than the disc scale radius but it can be significantly greater than $R_{\bullet}$ depending on the masses of the BH and disc.
It is important to note that the halo profile also plays a role in determining the shape of the $Q$ profiles. In the case of a less centrally dominated halo profile, such as the NFW profile \citep{Navarro1996}, the inner disc would be more unstable in the absence of a massive BH, decreasing $R_{\rm c, \, in}$. However, the BH would provide a more signifcant fraction of the total mass in the inner region of the system and the relative stabilising effect of the BH would therefore be enhanced in a NFW profile halo.
\begin{figure}
\includegraphics[width=\columnwidth]{r_vs_mbh_plot}
\caption{The relationship between characteristic radii versus BH mass for a disc with $M_{\rm d} = 1.02\times10^8 \, \mathrm{M_{\odot}}$ (shown as the vertical, dashed line) and a scale radius of $R_{\rm d} = 86.5 \, \mathrm{pc}$. The disc scale radius is shown in pink, the sphere of influence radius of the BH ($R_{\rm \bullet}$) is shown in purple, the inner and outer critical radii ($R_{\rm c, \, in}$ and $R_{\rm c, \, out}$) are shown in green and orange respectively. The green and orange lines stop at $M_{\bullet} \sim 10^7 \, M_{\odot}$, the mass at which the BH completely stabilises the disc.}
\label{fig:RvsMbh}
\end{figure}
\subsection{Minima of the Stability Parameter Profiles} \label{minQ}
\begin{figure}
\includegraphics[width=\columnwidth]{rqminrdvsmbhmd}
\caption{The top panel shows radius where the Toomre stability parameter is minimized as a fraction of the disc scale radius as a function of the mass ratio of the BH and disc. The bottom panel shows the same for the tidal stability parameter. The black line was found for the fiducial case of $f_{\rm b} = 0.17$, $ \lambda=\bar{\lambda}\equiv0.05$. The remaining lines represent cases where either $f_{\rm b}$ or $\lambda$ are changed from fiducial case to the values indicated in the legend. Unlike the curves, the position of the markers are dependent on the total mass of the system. The $M_{\rm total}=10^9 \mathrm{M}_{\odot}$ case is shown as an example. The dots correspond to the point where $Q_{\rm tidal,\, min}=Q_{\rm Toomre, \, min}$ i.e. to the left of these dots the Toomre parameter is a stricter criterion for star formation and to the right the tidal parameter is more the strict of the two. The stars correspond to the point where $Q_{*,\,\rm min}=1$. There are no star symbols shown for the low baryon fraction and high spin parameter cases (green and pink lines respectively) as the disc is fully stable for both, even in the absence of a BH.}
\label{fig:Rqmin}
\end{figure}
The radii of the minimum of the two stability parameter profiles as a fraction of the disc scale radius are only a function of the BH-to-disc mass ratio, the baryon fraction and the spin parameter of the halo (Assuming $j_{\rm d}/m_{\rm d} = 1$). Each of the lines in Figure~\ref{fig:Rqmin} represent a different combination of baryon fraction and spin parameter. While changing the total mass of the system will change the absolute value of $Q_*$, it does not change the positions (as a fraction of $R_{\rm d}$) of the minimum values of $Q_{\rm Toomre}$ and $Q_{\rm tidal}$. That is the ratios $f_{\rm b}$ and $M_{\bullet}/M_{\rm d}$ define the relative importance of the different components of each $Q$ and therefore shape of the $Q$ profiles. The spin parameter $\lambda$ defines $R_{\rm d}$ and therefore changes the surface density of the disc. Hence, the $R_{Q_{\rm min}}/R_{\rm d}$ -- $M_{\bullet}/M_{\rm d}$ relationships are influenced by $\lambda$ as the disc surface density and velocity profile (and therefore the disc stability) are dependent on it. Increasing the baryon fraction or decreasing the spin parameter leads to a similar change in the $R_{Q_{\rm min}}/R_{\rm d}$ -- $M_{\bullet}/M_{\rm d}$ curves. The curves shown with the higher and lower spin parameters correspond to the upper and lower limits of the 80\% confidence interval of the $\lambda$ probability distribution \citep{Mo1998}. The upper limit to the baryon fraction is unlikely to be much greater than our fiducial value \citep{Qin2017}. Therefore, $f_{\rm b}=0.2$ would be an extreme case. Though a lower baryon fraction than $f_{\rm b}=0.085$ is possible \citep{Qin2017}, such a system would struggle to have an unstable disc in our model. Over this range of values, the lower spin parameter limit case, $\lambda = 0.026344$, has the largest range in $R_{Q_{\rm min}}/R_{\rm d}$, with a factor of $<2.5$ change. This indicates $R_{Q_{\rm min}}\sim R_{\rm d}$ over the relevant parameter space.
The curves in Figure~\ref{fig:Rqmin} were found to follow the functional form:
\begin{equation}
\frac{R_{Q_{\rm min}}}{R_{\rm d}} = A-B \tan ^{-1}\left[C \left(\frac{M_{\bullet}}{M_{\rm d}}\right)^{-D}\right]
\end{equation}\label{eq:rqminfunc}
See Table~\ref{tab:fitdata} for the values corresponding to each curve in Figure~\ref{fig:Rqmin}.
For all curves, $R_{Q_{\rm tidal,\, min}}/R_{\rm d}$ -- $M_{\bullet}/M_{\rm d}$ is steepest between $M_{\bullet}/M_{\rm d}\sim0.1$ and $M_{\bullet}/M_{\rm d}\sim2$ and the points found for $Q_{*,\,\rm min}=1$ all lie in that range. Note, for $M_{\rm total}=10^9\, \mathrm{M}_{\odot}$, the tidal parameter becomes the more strict criterion at $M_{\bullet}/M_{\rm d} < 0.1$ in each case shown.
\begin{table}
\centering
\caption{Table of the $R_{Q_{\rm min}}/R_{\rm d}$ -- $M_{\bullet}/M_{\rm d}$ fit parameters for each curve in Figure~\ref{fig:Rqmin}. The fiducial case is shown in the top row.}
\label{tab:fitdata}
\begin{tabular}[h]{c|c|c|c|c|c}
\hline
& &
\multicolumn{4}{c}{$Q_{\rm Toomre}$} \\
$f_{\rm b}$ & $\lambda$ & A & B & C & D\\
\hline
0.17 & 0.05 & 1.43 & 0.4596 & 0.3904 & 0.6898 \\
\hline
0.085 & 0.05 & 1.371 & 0.3571 & 0.4572 & 0.7338\\
0.2 & 0.05 & 1.441 & 0.4842 & 0.3773 & 0.6782\\
0.17 & 0.026344 & 1.463 & 0.5414 & 0.3505 & 0.6496 \\
0.17 & 0.094898 & 1.381 & 0.3731 & 0.4454 & 0.7275 \\
\hline
& & \multicolumn{4}{c}{$Q_{\rm tidal}$} \\
$f_{\rm b}$ & $\lambda$ & A & B & C & D\\
\hline
0.17 & 0.05 & 1.338 & 0.3776 & 0.6429 & 0.7381\\
\hline
0.085 & 0.05 & 1.254 & 0.2669 & 0.7043 & 0.7739\\
0.2 & 0.05 & 1.357 & 0.4046 & 0.6302 & 0.7281\\
0.17 & 0.026344 & 1.398 & 0.4665 & 0.6028 & 0.7031\\
0.17 & 0.094898 & 1.267 & 0.2836 & 0.6938 & 0.7691 \\
\end{tabular}
\end{table}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{vsmbh_plot}
\caption{Each of the panels show a property of a disc relating to its stability as a function of BH mass for different disc masses. The halo mass, $M_{200}=5 \times 10^8\,\mathrm{M}_{\odot}$, and the disc scale radius, $R_{\rm d}=86.5\,\mathrm{pc}$, are the same in all cases. The top left panel shows the inner critical radius. The curves tend to the no BH case at the low mass BH end. The disc scale radius is shown as a black dashed line. The top right panel shows the relative difference in the inner critical radius with and without a BH $f_{R_{\rm c, \; in}}$ (see Equation~\ref{eq:reldiff}). The bottom left panel shows the stable fraction of the disc $\gamma$ (see Equation~\ref{eq:gamma}). The bottom right panel shows the star formation rate (SFR) for both with and without a BH (the solid and dashed lines respectively).}
\label{fig:RcvsMbh}
\end{figure*}
\subsection{Change in Disc Stability with Black Hole and Disc Mass}
The inner critical radius is shown as a function of BH mass for different disc masses in the top left panel of Figure~\ref{fig:RcvsMbh}. The purple line represents the same disc mass as used in Figures~\ref{fig:RvsMbh}. Increasing the disc mass decreases the critical radius as the disc becomes more unstable. For each case in the top left panel of Figure~\ref{fig:RvsMbh} the curve tends to the no BH case at the low mass BH end. As said above, as the BH mass increases the critical radius does increase, however, this increase behaves slightly differently for the higher disc masses.
Since the higher disc mass decreases $R_{\rm c, \, in}$, the critical radius is closer to the BH. This in turn means small changes to the BH mass at the low mass end has a greater influence on the value of $R_{\rm c, \, in}$; the relative difference in $R_{\rm c, \, in}$ between the cases with and without a BH is defined as:
\begin{equation}
\begin{aligned}
f_{R_{\rm c, \, in}}\equiv(R_{\rm c, \, in} - R_{\rm c, \, in}(M_{\bullet}=0))/R_{\rm c, \, in}(M_{\bullet}=0)
\label{eq:reldiff}
\end{aligned}
\end{equation}
This is shown to increase with disc mass in the top right panel of Figure~\ref{fig:RcvsMbh} which shows the relative difference as a function of BH mass for different disc masses. At higher BH masses the lower mass disc curves are steeper as the BH mass is increasingly comparable with the disc until the disc is fully stabilised. This is not seen in the higher disc mass cases as the BH masses investigated do not reach the range required to stabilise these discs. Note the BH mass required to fully stabilise the $M_{\rm d} = 3.24\times10^8 \, \mathrm{M_{\odot}}$ disc (orange curve) is $M_{\bullet} \gtrsim M_{\rm d}$ whereas the lowest disc mass case needs only $M_{\bullet} \lesssim 0.1 \, M_{\rm d}$. These numbers are in line with the range where the dependence of $R_{Q_{\rm tidal,\, min}}/R_{\rm d}$ on $M_{\bullet}/M_{\rm d}$ is strongest ($0.1 \lesssim M_{\bullet}/M_{\rm d} \lesssim 2$; see previous section).
We define the stable fraction of the disc as fraction of the disc mass outside the unstable region between $R_{\rm c, \; in}$ and $R_{\rm c, \; out}$,
\begin{equation}
\begin{aligned}
\gamma &\equiv 1 - \frac{M_{\rm g}(<R_{\rm c, \; out}) - M_{\rm g}(<R_{\rm c, \; in})}{M_{\rm g}}
\label{eq:gamma}
\end{aligned}
\end{equation}
\noindent The bottom left panel of Figure~\ref{fig:RcvsMbh} shows the stable fraction of the disc mass $\gamma$ as a function of the BH mass for the same cases as the other panels in the figure. In the model this is the fraction of the disc that is stabilised against gravitational collapse, i.e. the fraction of the disc that is outside the region where star formation can take place. That the most massive disc case is almost completely unstable is to be expected and the BH has no effect on the stability fraction for this case. Such a disc could not form as in this case as it greatly outweighs its host halo. Looking at the two lowest mass cases, there is a sharp change in the stable fraction of the disc as the BH mass becomes more comparable with the disc mass ($\sim10\%$), in line with trend seen in $R_{\rm c, \, in}$ panel as the disc reaches stability and $R_{\rm c, \, in} \sim R_{\rm d}$.
The bottom right panel of Figure~\ref{fig:RcvsMbh} shows the how the total SFR changes as a function of BH mass for different disc masses. The two cases with the lowest disc masses are where the difference between the cases with and without a black can be most significant. In these cases, the steep drop in SFR we see as the BH mass increases appears to simply reflect the stable gas fraction increase on the adjacent panel. In the lowest disc mass the drop-off occurs with $M_{\bullet} \lesssim 0.1\,M_{\rm d}$ (green line) while with the next higher mass disc the drop-off is at $M_{\bullet}\sim M_{\rm d}$, following the trend in stable fraction.
\subsection{Star Formation Timescale Profile}
To see how star formation is affected by the mass of the BH in more detail we need to look at the star formation timescale. Not only is the region where star formation can take place constrained by the BH but also the star formation timescale in the model can, in principle, be affected by the presence of the BH. This is because the timescale, $t_{\rm SF}$, is dependent on $Q_{\rm Toomre}$ and therefore $\kappa$ (see Equation~\ref{eq:sf8}), which depends on the BH mass as outlined above. The top panel of Figure~\ref{fig:tsfvsx} shows how the star formation timescale varies as a function of radius for different BH masses and a constant disc mass ($M_{\rm d} = 1.02\times10^8 \, \mathrm{M_{\odot}}$). At a given radius close to the BH, increasing the BH mass increases the timescale until the disc becomes locally stable ($Q_*(R) \geq 1$). There is little variation between the profiles except for the largest BH mass case with $M_{\bullet} = 10^7 \, \mathrm{M_{\odot}}$, where the BH mass is 10\% of the disc mass. Between this case and the lowest BH mass case, the width and area of the star forming region of the disc are around 2/3 smaller while the value of the timescale increases at a given radius by $\sim$15\%. The reduction in the fraction of the gas capable of forming stars at higher BH masses provides a more significant decrease in the star formation rate than the increase in the value of star formation timescale.
A comparison is made with the no BH case in the bottom panel of Figure~\ref{fig:tsfvsx} through finding the ratio of star formation surface density profiles for with and without a BH. The star formation rate in the presence of a BH is less than the no-BH case at all radii and is $\sim$15\% lower close to the BH at $R_{\rm c, \, in}$ for the two most massive BH cases. The presences of the BH changes the SFR surface density profile, and therefore the total SFR, due to the change in the Toomre parameter profile. By limiting the range in radius where stars can form the BH limits the total SFR and would confine the stellar mass to a narrow ring in the disc, ignoring any following redistribution of stars (e.g. through stellar or tidal interactions).
\begin{figure}
\includegraphics[width=\columnwidth]{sfprofiles_plot2}
\caption{The two panels show the following for a disc of mass $M_{\rm d} = 1.02\times10^8 \, \mathrm{M_{\odot}}$ for different BH masses: The top panel shows the star formation timescale versus radius for different BH masses; the bottom panel shows the ratio of the star formation surface density profiles for with and without a BH for different BH masses.}
\label{fig:tsfvsx}
\end{figure}
\section{Evolving Halo Model} \label{Evolving}
\subsection{DCBH Hosting Haloes} \label{DCBH haloes}
The formation of a seed BH through direct collapse is expected to take place in haloes within regions of high-intensity local LW radiation \citep[e.g][]{Agarwal2012b}. Recent studies have shown a local source of H$_2$-dissociating radiation from nearby quasars or PopII or PopIII stars is required \citep{Agarwal2016a} to provide the critical LW intensity. For this reason the properties of our model proto-galaxy are chosen to reflect those expected in proximity to a larger galaxy which formed at an earlier time. We assume the LW radiation field is sufficient to entirely dissociate molecular hydrogen in the proto-galaxy and the gas temperature is set to $T_{\rm g}= 8000$ K.
Due to this assumed proximity there is a strong likelihood of a DCBH hosting halo to undergo a merger in its evolution \citep{Agarwal2014a}. To account for the variation in the growth of seed BH hosting dark matter haloes we will focus on two evolutionary paths reported in simulations \citep{Agarwal2014a} (see Figure~\ref{fig:halodiagram}): (1) an isolated, growing halo and (2) a halo that forms a BH before it becomes a satellite at some later in-fall redshift, $z_{\rm infall}$. The main difference between these paths is the rate at which new baryons enter the proto-galaxy.
\begin{figure}
\includegraphics[width=\columnwidth]{diagram2}
\caption{A schematic diagram illustrating the cases investigated here for the evolution of DCBH hosting haloes. The blue case on the left is the isolated, growing halo case and the pink shows an in-falling halo case. The star symbol represents the DCBH hosting proto-galaxy/galaxy while the dots represent a separate, central galaxy.}
\label{fig:halodiagram}
\end{figure}
As mentioned above, the cooling in DCBH hosting haloes must be limited to occur via atomic hydrogen \citep{Agarwal2012b}. This constrains the metallicity, virial temperature, and therefore the mass of the haloes. All modelled haloes therefore have $T_{\rm v}\gtrsim 10^{4}$ K and the gas is comprised of atomic hydrogen and helium. For simplicity we assume that the cooling timescale is very short and cooling occurs on a dynamical time of the halo, meaning gas accreted by the halo reaches the proto-galaxy on a halo dynamical time\citep{Dekel2009, Khochfar2011}. An upper limit on the mass of such a proto-galaxy is therefore the baryonic mass fraction of the halo.
\subsection{Halo Growth} \label{haloevo}
We here model the time evolution of the system composed of a gaseous disc, a stellar disc, a BH and a dark matter halo. The evolving model follows the growth of a galactic disc within an isolated growing host halo after the formation of a massive BH seed at some initial redshift, $z_{\rm i}$, down to redshifts where SMBHs have been observed $z_{\rm end} \sim 6.0$ \citep{Fan2006a, Mortlock2011}. The seed formation redshift for our fiducial model is $z_{\rm i}\sim 10$ \citep{Agarwal2012b} though a range of seed formation redshifts are possible \citep[e.g.][]{Begelman2006} which we will investigate below as well. We follow the evolution of isolated and in-falling DCBH hosting haloes (see Figure~\ref{fig:halodiagram}).
The total mass of the system, $M_{\mathrm{total}}$, is made up of the dark matter (DM) halo and the baryons that make up the massive BH and the galaxy disc. To fit with the conditions expected for the formation of a DCBH, the total initial mass, $M_{\mathrm{total, \;i}}$, is calculated by estimating the mass of an atomic hydrogen cooling halo at $z_{\rm i}$ \citep{Mo2010}. The baryon fraction is assumed to follow the universal value of $f_{\mathrm{b}} = 0.17$. Initially, the BH seed is given a mass in the range $M_{\mathrm{\bullet, \; i}} = 10^{4 - 6} \, \mathrm{M_{\odot}}$, and the remaining baryons make up the disc mass.
The total mass increases through cosmological accretion of mass onto the system. The following equation taken from \citet{Dekel2013} is used to calculate the growth of the system:
\begin{equation}
M_{\rm total}(z) = M_{\mathrm{total, \;i}} \, \mathrm{e}^{-\alpha(z - z_{\mathrm{i}})}
\label{eq:1}
\end{equation}
Two methods are used to calculate the halo growth parameter, $\alpha$. In general, $\alpha=3/2 \, s \, t_1 = 0.806$ where $s = 0.030 \, \mathrm{Gyr^{-1}}$\citep{Dekel2013} and $t_1 = 2/3 \, \Omega_{m}^{-1/2} \, H_{0}^{-1} \sim 17.9 \, \mathrm{Gyr}$. As an alternative growth rate, $\alpha = 0.586$ was found by fitting an exponential to the median of the model growth histories for the host halo of the CR7 DCBH from \citet[][see their Figure 4]{Agarwal2016a}.
The total mass is split into the dark matter halo $M_{\mathrm{200}}(z) = M_{\mathrm{total}}(z) \, (1 - f_{\mathrm{b}})$ and the baryons. The model assumes the disc and central BH comprise all the baryons in the system and that accreted baryons go directly onto the disc, conserving mass. The equilibrium solutions are used here rather than allowing the accreted material to reach the disc over a dynamical time as the latter required an extra step in the calculation while having little baring on the disc mass at later times and, therefore, the redshift at which the disc became unstable. Hence, the mass of the disc at any given time is the difference between the total baryonic mass and the mass of the BH.
When sufficient gas is available, the BH is assumed to grow at a constant Eddington fraction, $f_{\mathrm{Edd}}$, leading to the following equation for the growth of the BH:
\begin{equation}
M_{\mathrm{\bullet}}(t) = M_{\mathrm{\bullet, \;i}} \, \exp\left(f_{\mathrm{Edd}} \frac{t - t_{\mathrm{i}}}{t_{\mathrm{Sal}}}\right)
\label{eq:5}
\end{equation}
\noindent where $t_{\mathrm{Sal}} = 0.45 \, \eta/(1 - \eta)\; \mathrm{Gyr}$ is the Salpeter timescale and $\eta$ is the radiative efficiency which we assume throughout as $\eta=0.1$ \citep{King2008}. The BH is assumed to only accrete gas. With a high Eddington fraction the BH accretion rate can exceed the baryonic growth rate of the halo at late times. This results in a decrease in the gas mass of the disc. If the gas mass drops to zero, the BH accretion rate will be limited to the baryonic growth rate of the halo.
\subsection{Star Formation and Stellar and Gaseous Disc} \label{evodisc}
Initially, we assume a gaseous disc which is fed by the net accretion of gas resulting from the difference in the gas accreted onto the disc and the BH. The growth of the disc translates into a growth in the gas surface density such that the central surface density becomes a function of time.
\begin{equation}
\Sigma_{\rm g,\,0}(t) = \frac{M_{\rm g}(t)}{2\pi R_{\rm d}(t)^2}
\label{eq:sigg0}
\end{equation}
After the onset of star formation in the proto-galaxy, some of the gas is converted into stars. The SFR is calculated using the method discussed in section~\ref{starform}. For simplicity, stars are assumed to remain on circular orbits where they form in the disc. The stars therefore follow a different surface density profile to the gas and the total disc surface density is simply the sum of the stellar and gas surface densities.
\begin{equation}
\Sigma_{\mathrm{d}}(R,t) = \Sigma_{\mathrm{g}}(R,t) + \Sigma_{\mathrm{\star}}(R,t)
\label{eq:7}
\end{equation}
For the purpose of our model we neglect feedback from stars and the accreting BH and note that the stellar mass is an upper limit on what could be expected. The velocity dispersion is assumed to be dominated by the sound speed of the gas ($c_{\rm s}\sim 10\, \mathrm{km/s}$ with $T_{\rm g}=8000$ K). However, including feedback effects should lead to an increase in the velocity dispersion of $\sim 10 \, \mathrm{km/s}$ \citep[see, e.g.][]{Wada2002, Dib2006, Agertz2009} due to supernovae after a few Myr \citep{Schaerer2002}, driving outflows and suppressing star formation.
\begin{table}
\centering
\caption{Table of model parameters.}
\label{tab:para_table}
\begin{tabular}[h]{l|l|l|l}
\hline
Parameter & Definition & Fiducial (Range)\\
\hline
$M_{\rm \bullet, \,i}$ & BH seed mass ($\mathrm{M_{\odot}}$) & $10^{6}$ ($10^{4 - 6}$) \\
$f_{\rm Edd}$ & Eddington fraction & 0.25 ($0.0 - 1.0$)\\
$\alpha$ & Halo growth parameter & 0.806 ($0.806, 0.586$)\\
$z_{\rm i}$ & Seed formation redshift & 10.0 ($20.0 - 10.0$)\\
$z_{\rm infall}$ & In-fall redshift & 0.0 ($10.0, 8.5, 7.0, 0.0$)\\
\end{tabular}
\end{table}
\subsection{Fiducial Case} \label{fiducialevo}
Table~\ref{tab:para_table} summarises the parameters discussed above with their fiducial values and the relevant ranges used. For our fiducial evolving model we make comparisons between cases both with and without a BH and with and without BH accretion. For the growth of the halo and the disc we assume an accretion rate in line with \citet{Dekel2013}. Using the lower accretion rate and a $T_{\rm v}\sim 10^{4}$ K halo at $z_{\rm i}=10$ would result in a system where the disc was never massive enough to be unstable prior to $z = 6$, independent of the BH mass\footnote{At higher formation redshifts, $z_{\rm i}\sim20$, a $T_{\rm v}\sim 10^{4}$ K halo can form a disc capable of becoming unstable at later times even for the assumed lower accretion rates.}. Our fiducial value for the seed mass is $M_{\bullet}=10^6 \, M_{\odot}$ and we assume an Eddington fraction of $f_{\rm Edd}=0.25$ for the accreting BH case.
\begin{figure}
\includegraphics[width=\columnwidth]{plot_m002run}
\caption{The mass evolution of each of the system components for the fiducial cases. The solid line is the case with no BH, the dashed line is for the case of a BH with a constant mass of $M_{\bullet} = 10^6 \, M_{\odot}$, and the dotted line is for the case with a BH growing from an initial mass of $M_{\bullet} = 10^6 \, M_{\odot}$ with $f_{\rm Edd} =0.25$.}
\label{fig:fid_mevo}
\end{figure}
Figure~\ref{fig:fid_mevo} shows the evolution with redshift of the mass of each component of the model for our three fiducial cases (no BH, non-accreting BH, accreting BH). The stellar mass evolution varies between the different models. The case without the BH has the largest stellar mass at all redshifts after the onset of star formation while the accreting BH has the lowest. As the accreting case has the most massive BH it will have the longest star formation timescales and the highest stable disc fraction, leading to lower star formation rates and hence lower stellar masses. Furthermore, the higher BH mass leads to a delay in the onset of star formation. The higher BH changes the $Q_*$ profile such that $Q_{*,\,\rm min}$ is higher for a given disc mass and therefore the critical mass of the disc required for it to become unstable is higher. In the constant BH mass case, this higher critical mass requirement delays the time at which the disc is first unstable as each model has the same cosmological accretion rate. In fact this delay is further enhanced in the accreting BH case as the net growth rate of the disc will be reduced.
\subsection{Star Formation Rate Surface Density Profile}
\begin{figure}
\includegraphics[width=\columnwidth]{sigsfrprofile}
\caption{The radial profiles of the SFR surface density and the stellar surface density at three snapshots during the evolution of two models of disc galaxies. The solid lines represent the fiducial model with a growing BH and the dashed line is the case with no BH. The first point in time ($z=6.45$) is taken immediately prior to the onset of star formation in the BH case and the final snapshot ($z=6.0$) is at the end of the calculation. In most cases, all star formation is taking place within the scale radius of the disc. Where this is not the case a marker of the corresponding colour indicates the disc scale length. The disc scale radius at each point in time is as follows: $R_{\rm d}=143.2$ pc for $z=6.45$, $R_{\rm d}=155.6$ pc for $z=6.25$, and $R_{\rm d}=167.6$ pc for $z=6$. Exponential surface density profiles, $\Sigma_{\star} \propto \exp\left (-a\, R/R_{\rm d} \right )$, were fitted to the stellar surface density profiles at $z=6$ (shown in black). At this redshift the fit parameter was found to be $a=0.51$ (with a turn over starting at $r=84$ pc) and $a=15.1$ (with a turn over at $r=124$ pc) in the BH and no BH cases respectively.}
\label{fig:sigsfrprofile}
\end{figure}
The change in the Toomre and tidal parameter radial profiles due to the presence of a BH has an affect on the SFR in the disc. Figure~\ref{fig:sigsfrprofile} shows how the SFR surface density evolves in the model with and without a BH. The region where star formation takes place in the model is shifted outward in the cases with a BH compared to the one without. Over time the SFR increases throughout the unstable region. This is expected in our model as the formation of stars in a region increases the stellar surface density while the corresponding decrease in gas density is spread out throughout the disc. Meanwhile more gas is accreted through cosmological accretion and $R_{\rm d}$ increases as the halo grows. This means even as the gas density profile is stretch out there is an overall increase in the total surface density in a region undergoing star formation and this increases the SFR in that region (see Equation~\ref{eq:sf9}). The lack of feedback effects or any momentum and mass transfer in the stellar disc results in runaway star formation. As the BH is allowed to accrete gas there is further increase in the difference in the SFR at late times as the gas surface density is reduced. The differences in the stability profiles and gas density means the SFR and stellar mass surface densities are higher at each point in time and at each radius in the case without the BH. Looking at the lowest redshift, the presence of the BH has resulted in a decrease in the width of the annulus of the disc where stars can form by $\sim 1/3$. The resulting ring of stars occupies this same smaller region and the inner region is void of stars, effectively creating a hole in the galaxy stellar disc by enlarging the central region void of stars from $\sim64$ pc to $\sim110$ pc. The change in the $Q_*$ profiles does result in a fractional increase ($9\%$) in the outer radius of the star formation region but this only has a minor effect on the total SFR of the system.
\subsection{Evolution of the Star Formation Rate}
\begin{figure}
\includegraphics[width=\columnwidth]{sfr_and_dsfr_plot}
\caption{The evolution of the star formation rate for the fiducial cases and the difference between the cases with and without a BH. $\Delta \mathrm{SFR} = \mathrm{SFR}_{\rm no\,BH} - \mathrm{SFR}$.}
\label{fig:fid_sfr}
\end{figure}
The radially integrated SFR in Figure~\ref{fig:fid_sfr} shows the difference in the total SFR over time. The SFR of the no BH case is highest at all redshifts after the onset of star formation and the accreting BH case results in the lowest. As we go forward in time we see that the difference between the SFRs increases. As the stellar density increases in the unstable region of the disc, the gas mass will continue to be spread across the total disc profile and the total density will increase within the unstable region, resulting in a local increase in the SFR (see Equation~\ref{eq:sf9}). As the model does not follow stellar migration, the stellar mass is not redistributed and the SFR simply increases the surface density and so on, leading to a run away effect until the gas density reaches the star formation threshold value. This means, within the model, once a galaxy has a higher SFR and stellar mass it becomes hard for another model to catch up, unless the gas is used up less efficiently due to the presence of a BH.
Note the inclusion of feedback from accretion onto the BH would have the potential to regulate the star formation in the disc further. The resulting heating and ejection of gas could further stabilise the disc against star formation but also regulate the growth of the BH \citep{Latif2018}. If the BH were to maintain the assumed growth rate, the star formation rate would be expected to decrease due to the decrease in the gas surface density and the increase in the gas temperature. The onset of star formation would be delayed and once the disc did reach instability it would have a lower unstable fraction, lowering the total SFR.
\begin{figure}
\includegraphics[width=\columnwidth]{ssfr_and_dssfr_plot}
\caption{The evolution of the specific star formation rate for the fiducial cases and the difference between the cases with and without a BH. $\Delta \mathrm{sSFR} = \mathrm{sSFR}_{\rm no\,BH} - \mathrm{sSFR}$.}
\label{fig:fid_ssfr}
\end{figure}
The evolution of the specific star formation (sSFR) rate of each of the models (Figure~\ref{fig:fid_ssfr}) shows interestingly that the sSFR is higher for the higher BH mass cases. The rise in $Q_{\rm Toomre}$ and $Q_{\rm tidal}$ due to the BH decreases the SFR and therefore a significant decrease in the stellar mass over time, resulting in an increase in the sSFR. Indeed, without the BH the sSFR is lower at early times as the stellar surface density will be significantly larger due to the difference in the time at which star formation can first occur in the disc. As the system progresses the stellar masses become more comparable and the difference in the sSFR decreases.
When compared to observations \citep{Stark2013}, we find our model sSFR is greater by a factor of $~10$ at $z=6.8$, though the lower mass BH and no BH cases appear to be following a trend which would agree with the $z=5.9$ data point. However, the relationship between SFR and stellar mass has a large scatter and the slope varies with stellar mass \citep{Whitaker2014}, meaning large deviations from this median value in sSFR for individual galaxies is to be expected, particularly at low masses. Indeed, our findings suggest DCBH hosting galaxies should generally have a higher sSFR, providing a possible tool for identifying candidate DCBH hosts.
\subsection{Onset of Star Formation}
\begin{figure*}
\begin{center}$
\begin{array}{cc}
\includegraphics[width=\columnwidth]{zun_zi10_mbhi_fg1} &
\includegraphics[width=\columnwidth]{zun_zi10_mbht_fg1}\\
\includegraphics[width=\columnwidth]{zun_zi10_mbhi_fg15} &
\includegraphics[width=\columnwidth]{zun_zi10_mbht_fg15}\\
\includegraphics[width=\columnwidth]{zun_zi20_mbhi_CR7} &
\includegraphics[width=\columnwidth]{zun_zi20_mbht_CR7}
\end{array}$
\end{center}
\caption{The left panels show the redshift at which the modelled disc first becomes unstable (i.e. $Q_{*\,\rm min}=1$) and is able to form stars ($z_{\rm SF}$) as a function of the initial BH seed mass ($M_{\bullet, \, i}$) for different fractions of the Eddington limit accretion rate ($f_{\rm Edd}$). The right panels show the same except with the BH mass at $z_{\rm SF}$ on the x-axis. The time difference between the onset of star formation with and without a BH is also shown on the right hand side to indicate the delay caused by the stabilising effect of the BH. The upper limit of the initial BH mass range is the total baryonic mass of the atomic hydrogen cooling halo at the DCBH formation redshift of $z\sim10$ for the top two rows of panels and $z\sim20$ for the bottom. The growth rate of the halo follows Equation~\ref{eq:1} with $\alpha=0.806$ for the top panels, $\alpha=1.209$ for the middle and $\alpha=0.586$ for the bottom. The $f_{\rm Edd} = 0.25 - 1$ lines in the top panels each reach a maximum seed mass above which the disc will never become unstable and be able to form stars. In the bottom panels this is only seen for the $f_{\rm Edd} = 1$ line.}
\label{fig:zsf}
\end{figure*}
The top panels of Figure~\ref{fig:zsf} show how the redshift at which star formation first occurs in the model depends on the seed mass and the growth rate of the BH for an atomic hydrogen cooling halo that forms a DCBH at $z=10$ while the remainder of its baryonic mass goes into making a disc. For the case of Eddington limited accretion ($f_{\rm Edd} = 1$) even the lowest mass in the estimated range of the DCBH masses, $M_{\bullet,i}=10^4 \; M_{\odot}$, results in a disc that will never undergo star formation. Yet, high accretion rates close to the Eddington limit are required for even the most massive DCBHs at $z\sim10$ to reach the $M_{\bullet}\sim10^9\, \mathrm{M}_{\odot}$ by $z\gtrsim6-7$ as observed \citep{Fan2006a,Mortlock2011}. This indicates DCBH formed and grew into SMBHs in separate progenitors from their eventual host galaxies, in order for these massive quasars to be observed within massive galaxies at $z\sim6$. In fact, for the upper limit of the DCBH mass range, $M_{\bullet,i}=10^6 \; M_{\odot}$, star formation is inhibited for the $f_{\rm Edd} = 0.5$ case and is delayed by $\sim 100$ Myr with $f_{\rm Edd} = 0.25$. Most notably at this formation redshift ($z_{\rm i}\sim10$), any combination of seed mass and growth rate which leads to the growth of a $M_{\bullet}\sim 10^9 \, \mathrm{M}_{\odot}$ SMBH by $z\sim6$ inhibits star formation in the host.
In the cases of no BH growth and $f_{\rm Edd} = 0.1$, the disc will eventually undergo star formation, even when the BH seed mass is at its maximum i.e. at the seed's formation it takes up the entire baryonic mass of the halo ($M_{\bullet,i}= M_{\rm b} = 7.82 \times 10^6 \; M_{\odot}$) and the disc mass is initially zero. The onset of star formation is delayed somewhat in these $M_{\bullet,i}= M_{\rm b}$ seed cases, with $f_{\rm Edd} = 0.1$ leading to a delay by $\sim 200$ Myr which is significant as this is around a fifth of the age of the universe at this epoch. Note in all models the BH mass never exceeds the total baryonic mass in the redshift range we investigate in the models where $z_{\rm SF}$ is defined.
The growth rate of the halo and therefore the disc greatly influences this result. The lower the growth rate of the halo the more delayed star formation will be. As highlighted above, the growth rate modelled for the DCBH hosting halo of CR7 by \citet{Agarwal2016a} is sufficiently low that with $z_{\rm i} = 10$ the surface density of the disc is never high enough for stars to form over the redshift range we investigate. However, at earlier formation times the role of the BH decreases as the growth rate of the halo becomes higher at larger redshift. With a formation redshift of $z_{\rm i}=20$, a BH with an initial seed mass of $M_{\bullet, \rm i}=10^6 \, \mathrm{M}_{\odot}$ growing at the Eddington limit will be unable to prevent star formation, only delaying the onset by $\sim40$ Myr (see bottom panels of Figure~\ref{fig:zsf}).
Similarly to increasing the formation redshift, increasing the growth rate of the halo decreases the influence of the BH. \citet{Genel2008} find a scatter in the growth rate of DM haloes which they approximate as $\approx \left \langle \dot{M}_{\rm DM} \right \rangle \left (2.5/(1+z) \right )^{0.2} $ where $\left \langle \dot{M}_{\rm DM} \right \rangle$ is the mean halo growth rate. Assuming the growth rate is a linear function of the halo mass, in line with Equation~\ref{eq:1}, leads to $\sigma_\alpha \approx \left \langle \alpha \right \rangle \left (2.5/(1+z) \right )^{0.2}$. The middle panels of Figure~\ref{fig:zsf} shows the case for a formation redshift of $z_{\rm i} = 10$ with a growth rate of $\alpha = 1.209$, $\times 1.5$ the fiducial rate and within the 1$-\sigma_\alpha$ scatter at $z=6$. With this case the disc rapidly becomes more massive than the fiducial case and therefore becomes unstable much earlier. The BH mass required to keep the disc stable increases; it has to grow much faster to keep up with the disc and prevent star formation. This is illustrated by looking at a $M_{\bullet, \rm i}\sim 2\times 10^5 \, \mathrm{M}_{\odot}$ seed case. Growing at the Eddington limit, such a seed does prevent star formation, however, this is achievable at the same growth rate by a seed with a mass 20 times smaller at $M_{\bullet, \rm i}\sim 5\times 10^3 \, \mathrm{M}_{\odot}$ in the fiducial case. Furthermore, a $M_{\bullet, \rm i}\sim 2\times 10^5 \, \mathrm{M}_{\odot}$ seed in the fiducial case is capable of preventing the onset of star formation growing at half the BH accretion rate.
\begin{figure*}
\begin{center}$
\begin{array}{cc}
\includegraphics[width=\columnwidth]{alphafeddcontourzi10_long} &
\includegraphics[width=\columnwidth]{alphafeddcontourzi20_long}
\end{array}$
\end{center}
\caption{The variation of $z_{\rm SF}$, the redshift at which star formation can first occur, with $\alpha$ and $f_{\rm Edd}$, the accretion rate parameter of the halo and the BH Eddington fraction respectively. Where no value for $z_{\rm SF}$ is shown, the system will never undergo star formation. The initial seed mass is $M_{\bullet, \rm i}=10^6 \, \mathrm{M}_{\odot}$. The formation redshift is $z_{\rm i}=10$ in the left panel and $z_{\rm i}=20$ in the right panel. The range in $\alpha$ shown is from the lower 1$-\sigma_{\alpha}$ limit to the upper 2$-\sigma_{\alpha}$ limit at $z=6$. The average halo accretion rate parameter, and our fiducial value, $\alpha=0.809$ is shown as the grey, dashed line. The upper 1$-\sigma_{\alpha}$ limit of $\alpha=1.467$ is shown as the grey dot-dash line. The grey, dotted line represents the minimum Eddington fraction for which a $M_{\bullet, \rm i}=10^6 \, \mathrm{M}_{\odot}$ seed BH will reach $M_{\bullet}=10^9 \, \mathrm{M}_{\odot}$ by $z=6$. A best-fitting line for $f_{\rm Edd}>0.01$ is shown in both cases for the critical values of $\alpha$ and $f_{\rm Edd}$ where the model transfers from becoming unstable at some redshift to never being able to form stars.}
\label{fig:alpha_fedd_zsf}
\end{figure*}
The interplay of the halo and BH growth rates is well summarized in Figure~\ref{fig:alpha_fedd_zsf}. The figure shows how the onset of star formation varies with the growth rate of the halo and the BH and also depends on the formation redshift. $z_{\rm SF}$ is calculated as a function of the halo growth parameter, $\alpha$, and Eddington fraction for the same seed mass of $M_{\bullet, \rm i}=10^6\, \mathrm{M}_{\odot}$ at formation redshift $z_{\rm i}=10$ and $z_{\rm i}=20$. The range in $\alpha$ shown is from the lower 1$-\sigma_{\alpha}$ limit to the upper 2$-\sigma_{\alpha}$ limit at $z=6$, where $\sigma_{\alpha}$ was calculated using the approximation from \citet{Genel2008} as outlined above. In the $z_{\rm i}=10$ case, a significant fraction of the parameter space results in a model that is unable to ever form stars, particularly at higher Eddington fractions. Above the Eddington fraction at which the BH reaches $M_{\bullet}=10^9 \, \mathrm{M}_{\odot}$ at $z=6$ ($f_{\rm Edd}=0.728$), a higher than average halo growth parameter is required for star formation to occur. At even higher BH accretion rates, $f_{\rm Edd}\gtrsim0.9$, only haloes growing more than 1$-\sigma_{\alpha}$ faster than the average growth rate are sufficient. However, with $z_{\rm i}=20$, only the models with a slower than average halo growth rate, $\alpha \lesssim 0.8$, have a significantly delayed onset of star formation.
How strong an effect the BH has on the galaxy will depend on the growth rate of both the BH and the disc. Figure~\ref{fig:MbhMsevo} shows the evolution of the BH and stellar mass for different seed masses and accretion rates. The BH mass is initially significantly more massive but as the SFR is significantly larger than the BH accretion rate in these cases, the stellar mass quickly catches up with the BH mass. However, by the end of the calculation at $z=6$, only the lowest mass seed region reaches to the $M_{\bullet}\sim 10^{-3}\, M_{\star}$ line seen empirically at lower redshifts \citep{Haring2004, Kormendy2013, Reines2015}. This implies a BH cannot grow to lie on the BH-stellar mass relation at this point in cosmic time if the halo grows with an average growth rate. Indeed, earlier studies have indicated that systems reach the BH-stellar mass relation through periods of growth triggered by galaxy encounters \citep{Lamastra2010, Valiante2014}. The results of our model indicate a boost to the stellar mass is required possibly through mergers with evolved galaxies hosting only small or no BHs once the SMBH has grown. This suggests it is likely that these seeds are generated in satellites prior to falling into their host galaxies to lie on the relation.
It is also thought that the empirical scaling relation of BH and stellar mass is linked to the interaction of AGN and star formation; the feedback attributed to AGN helps regulate the SFR and vice-versa \citep[see, e.g.][]{Silk1998, Gabor2010, Silk2013}, producing this correlation. Here we do not model the feedback from either the BH or stars yet it is not clear that the inclusion of feedback would resolve the discrepancy of the results with the empirical relation. This model predicts BH and galaxy masses that would place the model galaxies above the relation, meaning the BHs are too massive relative to their hosts. Yet the inclusion of feedback from stars to regulate the growth of BHs would be insufficient as our BHs gain most of their mass prior to the onset of star formation. The inclusion of feedback from accretion onto the BH itself would be expected to regulate both the growth of the BH and the SFR. If, however, the BH were able to grow at the assumed rate in the model, the heating of the disc due the BH growth would decrease the fraction of the gas able to form stars. The resulting decrease in the SFR would lead to such systems existing further above the BH-stellar mass relation.
\begin{figure}
\includegraphics[width=\columnwidth]{fbplot0508-1100}
\caption{The evolution of the total stellar mass and BH mass was calculated for a range of growth rates ($f_{\rm Edd} = 0.0,\,0.1,\,0.2,\,0.5,\,1.0$; note $f_{\rm Edd}=1.0$ cases are not shown) and seed masses ($M_{\bullet,\rm \,i}=10^6 \, \mathrm{M_{\odot}},\,10^5 \, \mathrm{M_{\odot}},\,10^4 \, \mathrm{M_{\odot}}$). The coloured regions represent the spread in the evolution for different accretion rates given the same initial seed mass. Each model starts from the left hand side of the plot (with a stellar mass of zero) and once the onset of star formation is reached the evolution traces from left to right. The BH accretion rates shown here are limited to $f_{\rm Edd} \le 0.2$, $f_{\rm Edd} \le 0.5$, and $f_{\rm Edd} \le 0.5$ in the $M_{\bullet,\rm \,i}=10^6 \, \mathrm{M_{\odot}}$, $M_{\bullet,\rm \,i}=10^5 \, \mathrm{M_{\odot}}$, and $M_{\bullet,\rm \,i}=10^4 \, \mathrm{M_{\odot}}$ seed mass cases respectively as the higher accretion rates entirely prevent star formation. The dashed line represents $M_{\bullet}= 10^{-3}\, M_{\star}$ \citep[see, e.g.][]{Kormendy2013} and the coloured dotted lines connect points at the same redshift for each seed mass case. From left to right the lines correspond to $z=6.7,\,6.5,\,6.3,\,6.1$.}
\label{fig:MbhMsevo}
\end{figure}
\subsection{In-falling Host Halo}
The in-fall of the seed-BH-hosting halo to become a satellite of a more massive, central galaxy is modelled by cutting the growth of the halo. This reflects the starving of satellites as they are unable to accrete fresh material and existing material will either be used up or stripped. At a given redshift, $z_{\rm infall}$, the halo growth rate is set to zero and in turn accretion of fresh baryons stops. The disc mass only changes as the BH continues to accrete the remaining gas.
If the in-fall event happens prior to the onset of star formation the disc will never become unstable, due to the halt in the growth of the disc mass. This would result in a massive BH surrounded by a primordial gaseous disc (assuming the BH is unable to accrete all the gas), orbiting a central galaxy. For example, this would be true in our fiducial case if the in-fall occurs at a higher redshift than the onset of star formation at around $z\lesssim 7$. As the delay in onset of star formation increases with the mass of the seed BH, an in-fall event is more likely to occur prior to the onset of star formation in systems with more massive BHs.
In general, the SFR is maintained by the influx of gas to the disc. Cutting off this supply by having an in-fall after the onset of star formation leads to a rapid decrease in the gas density as the gas is converted into stars and accreted by the BH. The rate at which the SFR then decreases to zero will depend on the stellar mass and the BH accretion rate. Eventually, the stellar and BH masses will each reach a maximum and stop growing, starved by the lack of gas. For our fiducial model values for BH formation redshift and halo growth rate, only the models with lowest mass and slowest growing BH seeds can reach the $M_\bullet/M_\star\sim0.001$ relation by $z\sim6$ (Figure~\ref{fig:MbhMsevo}). Therefore, if the in-fall of the model galaxy were to take place at $z>6$, the resulting satellite would likely have an oversized BH relative to its stellar mass. This low mass satellite dominated by a massive BH could survive to lower redshifts due to the absence of further accretion of gas. This does however depend on the timing of the subsequent merger of the satellite with the central galaxy and, hence, the separation and relative masses of the merging galaxies. Indeed, the merging of massive BHs from the accretion of BH-dominated satellite galaxies could help form the SMBHs found in massive central galaxies at lower redshift \citep{Volonteri2003}, however, the time-scale for massive BHs to merge could be large \citep{Tremmel2015}.
A study by \citet{Agarwal2014a} identified haloes where DCBHs formed within a cosmological, hydrodynamical simulation. We highlight two of the cases they identified with a DCBH formation redshifts of $z_{\rm i}\sim 10$. One where the seed forms in a site close to one dominant galaxy and another where the formation in a clustered environment. In first case the DCBH host falls in to its largest neighbour at $z\sim8.5$, which is $\lesssim 200$ Myr after the formation of the seed. Within our model this would likely occur prior to the onset of star formation in the seed hosting halo. In the second case the seed hosting halo undergoes an in-fall at a later time, around $z\sim6$, well after the likely onset of star formation from our model. However, the environment of the cluster may play a stronger role in this scenario.
\section{Discussion \& Conclusions} \label{conclusions}
We use an analytical model to investigate the effect of a DCBH seed on the stability of proto-galaxy discs and the resulting suppression of star formation. We look at how the Toomre and tidal stability parameters profiles of an exponential disc change due to the presence of a BH in the centre of the system and link the stability of the disc to the star formation rate. We show how the BH has a gravitationally stabilising effect on the inner region of the disc which increases the star formation timescale locally and limits the region of the disc where star formation can occur, decreasing the modelled SFR. We also model the growth of a galaxy around a seed BH to investigate how the interplay of cosmological accretion, accretion onto the BH and the stabilizing effect of the BH can be important in determining the circumstances under which stars can form.
After the initial onset of star formation, we find that the radial extent of the star forming region remains relatively constant. Under the assumption of stars staying on circular orbits and not migrating in the disc, the process of forming stars increases the local surface density ($\Sigma_{\rm g} + \Sigma_{\star}$). This increases the self-gravity of the disc locally and decreases the effect of tidal forces on the gas. Removing the support from the tidal shear against gravitational collapse then leads to the further formation of stars in this same region. Following a short period beginning at the onset of star formation (while the stellar mass is still negligible), all subsequent star formation in the disc is largely confined to the region where stars have already formed. As stability increases in the presence of a massive BH, the radial extent of the region where stars can form narrows and the total SFR is reduced.
The radial extent of the region where stars can form in the model disc is small ($\sim 100$ pc) due to the disc properties at $z=6$, even in the absence of a BH. For the evolving model with a formation redshift at $z_{\rm i}=10$, we calculate the angular size of the stellar disc in the no BH case at $z=6$ to be $\theta< 0.02$ arcseconds and note this is less than the angular resolution of the James Webb Space Telescope (JWST), even at the shortest possible wavelengths. Resolved observations of such objects at this redshift would therefore by infeasible with current instruments.
The presence of a growing BH seed can greatly affect the star formation history of its host galaxy, even preventing the formation of stars entirely. Increasing the mass of the BH or the scale radius of the disc increases the stability of the disc, while increasing the disc mass decreases the stability. In the fiducial case, the disc becomes more unstable in the star forming region as the disc mass increases with the growth of the halo, resulting in SFR increasing with time. We find the sSFR in the model increases with higher BH mass and that the sSFR we calculate is higher than the observed median value at high redshift \citep{Stark2013}, particularly at times close to the onset of star formation. Our results suggest that systems hosting DCBHs should occupy the upper envelope of the sSFR distribution for any given stellar mass. Indeed, high sSFR galaxies could potential be used for the identification of DCBH hosts. As we evolve the model to lower redshifts, the discrepancy between the model sSFR and the observations decreases.
Increasing the accretion rate of the BH leads to an increase in the stability of the disc at a given time as the BH mass increases and the disc mass decreases. This can lead to a delay in the time where the disc first becomes unstable and forms stars. This delay in the onset of star formation is not only dependent on the BH growth rate and seed mass but also the growth of the disc and halo. As halo growth rates are higher at high redshift, the delay is also a function of the formation redshift of the BH. For a sufficiently high BH accretion rate and seed mass, the disc can be prevented from ever forming stars. At the lowest halo growth rates and high BH accretion rates, even models with early formation times have no stars forming. Such a low halo growth rate is typical of satellite galaxies \citep[see, e.g.][]{DeLucia2012}. This suggests the chance of a SMBH forming with no stellar disc counterpart is more likely in satellite galaxies. Indeed, this would also occur if an in-fall event were to occur prior to the onset of star formation.
We find that the halo in which a seed is born at $z=10$ is prevented from having significant star formation if the BH grows at the Eddington limit. If a seed BH is to grow at the rate required to increase in mass by $\gtrsim 3$ orders of magnitude between $z\sim10$ and $z\sim6$, star formation in its host is suppressed, placing such a system above the BH-stellar-mass relation. This suggests that DCBH galaxies will move towards the local BH-stellar mass relation via potential mergers with already evolved galaxies without massive BHs and not self-regulated co-evolution. Alternatively, this discrepancy can be resolved if either the formation of the DCBH is pushed to higher redshift ($z\sim20$) or if the evolution of the BH-galaxy system takes place in haloes with higher than average growth rates.
Though we do not model the feedback from the accreting BH we acknowledge that this would change the star formation and BH growth histories \citep{Schawinski2006, Latif2018}. BH feedback would heat and eject gas in the disc, acting to stabilise it, reducing the star formation rate in the model. The process of stabilising the disc through BH feedback would complement the gravitationally stabilising effect of the BH, delaying the onset of star formation further and decreasing the area of the disc able to form stars. This does not take into account the inclusion of ``positive feedback'' \citep{Gaibler2012}, where the inducing of star formation through jets leads to an increase in the SFR. However, this induced star formation would take place at large radii, meaning the inner region close to the BH would still be void of stars.
\section*{Acknowledgements}
\defcitealias{Astropy2013}{Astropy Collaboration, 2013}
DSE acknowledges the financial support of the Science
and Technology Facilities Council through a studentship
award. The Authors thank the reviewer for their helpful comments and suggestions. The Authors would also like to thank Tilman Hartwig, Bhaskar Agarwal and Marta Volonteri for their useful discussions and comments. This research made use of Astropy, a community-developed core Python package for Astronomy \citepalias{Astropy2013}.
\bibliographystyle{mnras}
|
\section{Introduction}
Assume one is interested in sampling from a target probability density on $\mathbb{R}^d$ which can be evaluated pointwise up to an intractable normalizing constant.
In this context one can use Markov chain Monte Carlo (MCMC) algorithms to sample from, and compute expectations with respect to the target measure.
Despite their great success, standard MCMC methods, such as the ubiquitous Metropolis--Hastings algorithm,
tend to perform poorly on high-dimensional targets. To address this issue,
several new methods have been proposed over the past few decades. Popular alternatives include the Metropolis-adjusted Langevin algorithm (MALA)
\cite{roberts1996exponential}, Hamiltonian, or Hybrid, Monte Carlo (HMC) \cite{duane1987hybrid} and slice sampling \cite{neal2003slice}.
Recently, a novel class of non-reversible, continuous-time MCMC algorithms based on piecewise-deterministic Markov processes (PDMP) has appeared in applied probability \cite{monmarche2016piecewise,bierkens2017piecewise}, automatic control \cite{mesquita2012jump},
physics \cite{peters2012rejection,michel2014generalized,nishikawa2016event} statistics and machine learning
\cite{bouncy2018,bierkens2016zig,RHMC,vanetti2017piecewise,B_BC_D_D_F_R_J_16,pakman2016stochastic,wu2017generalized}.
Most of the current literature revolves around two piecewise-deterministic MCMC (PDMCMC) schemes: the Bouncy Particle Sampler (BPS) \cite{peters2012rejection,bouncy2018} and the Zig-Zag sampler \cite{bierkens2016zig}.
A practical advantage of the BPS and Zig-Zag algorithms is that in many models it is possible to simulate their piecewise linear paths without time-discretization \cite{bouncy2018}. In contrast, methods based on either diffusions or Hamiltonian paths require time discretization. Moreover their performance is known to collapse if the discretization is too coarse.
Despite the increasing interest in these piecewise linear PDMCMC algorithms, our theoretical understanding of their properties remains limited, although
a fair amount of progress has been achieved recently in establishing geometric ergodicity, see \cite{deligiannidis2017exponential,durmus2018geometric} for BPS and
\cite{fetique2017long,bierkens2017ergodicity} for Zig-Zag. However, all of these results tend to provide convergence rates that deteriorate with the dimension and thus fail to capture the
empirical performance of these PDMCMC algorithms on high-dimensional targets.
Scaling limits have become a very popular tool for analysing and comparing MCMC algorithms in high-dimensional scenarios since their introduction in the seminal paper \cite{roberts1997weak}; see, e.g., \cite{roberts1998optimal,beskos2013optimal}. They have been used to establish the computational complexity of
the most popular MCMC algorithms, which is $O(d^2)$ for Random Walk Metropolis (RWM), $O(d^{4/3})$ for MALA and $O(d^{5/4})$ for HMC; here computational complexity is defined in terms of the expected squared jump distance.
In this direction, the recent work of \citet{bierkens2018high} has established scaling limits for both Zig-Zag and global BPS for high-dimensional standard Gaussian targets.
They obtain the scaling limits of several finite dimensional statistics, namely the angular velocity, the log-density and the first coordinate.
In this context, it is shown that Zig-Zag has algorithmic complexity $O(d)$ for all three types of statistics, whereas global BPS has complexity $O(d)$ for angular momentum and
$O(d^2)$ for the other two types of statistics. Benefits of Zig-Zag over global BPS are to be expected in this scenario. Indeed, when applied to a product target, the Zig-Zag sampler
factorizes into independent components and is closely related to Local-BPS (LBPS); see
\cite{peters2012rejection,bouncy2018}. The standard (global) BPS studied herein and \citet{bierkens2018high}, just like RWM, MALA and HMC, is an algorithm whose dynamics do not distinguish between product and non-product targets.
In the present paper, we also study scaling limits for BPS but concentrate on the first coordinate and its corresponding velocity in a regime which differs from the one considered in \cite{bierkens2018high}. It provides a different scaling limit,
which suggests that BPS has algorithmic complexity $O(d^{3/2})$, at least on weakly dependent targets. This is in agreement with the empirical results reported in \cite{bouncy2018}. The first difference is that \cite{bierkens2018high} considers BPS with the location evolving at unit speed, whereas in our scenario the velocity is Gaussian, therefore with speed scaling like $\sqrt{d}$ in the dimension.
The second difference is that \cite{bierkens2018high} considers scaling limits for the first coordinate of the location process only,
whereas we look at both location and velocity. Finally the third difference is that \cite{bierkens2018high} rescales time with a factor $d$, whereas we obtain our limiting process on the natural time scale. Given the different regimes and different objects studied in \cite{bierkens2018high} and the present paper, it is not surprising that the
two scaling limits differ significantly, with our bound being tighter and seemingly better at capturing the empirical behaviour of the process. In \cite{bierkens2018high} the first location coordinate converges to a Langevin diffusion, whereas in the present paper the process
tracking the first location and velocity components converges to a piecewise deterministic Markov process known as Randomized Hamiltonian Monte Carlo (RHMC).
Although the corresponding Fokker-Planck equation was studied in \citet{dolbeault2015hypocoercivity}, using a related approach to ours,
RHMC was first studied in a Monte Carlo context in \cite{RHMC}.
To the best of our knowledge, our result is the first in the literature establishing a direct link between BPS and Hamiltonian dynamics.
It is our understanding that the Langevin diffusion obtained in \cite{bierkens2018high} can be obtained from RHMC by a further limiting procedure similar to the \textit{overdamped} regime of the Langevin equation.
The second part of the paper is concerned with the convergence properties of RHMC. This process was studied in \cite{RHMC} where it was established that it is \textit{geometrically ergodic}. However, it is not clear whether such an approach can provide dimension independent convergence rates. The earlier work of \cite{dolbeault2015hypocoercivity} studies the corresponding Fokker-Planck equation, tracking the evolution of densities rather than conditional expectations.
In recent years, there has been great success in obtaining dimension-free convergence rates of MCMC schemes for strongly log-concave targets with bounded Hessians; see for example \cite{dalalyan2017theoretical,durmus2017nonasymptotic,mangoubi2017rapid,bou2018coupling,dwivedi2018log}.
In particular, in relation to HMC, the papers \cite{mangoubi2017rapid,bou2018coupling} use coupling techniques to obtain convergence
rates in terms of Wasserstein or total variation distances, but these usually leverage independent momentum refreshment to obtain a Markov process in the location components only.
We establish here these convergence rates in weighted Wasserstein distance using coupling ideas, and also
in $L^2$ using \textit{hypocoercivity}; see, e.g., \cite{dric2009hypocoercivity,pavliotis2014stochastic}.
The rates we provide may generally not be the optimal ones for specific scenarios. However, the optimal rates for a specific scenario can be obtained by solving a multivariate optimisation problem. \citet{dolbeault2015hypocoercivity} also uses hypocoercivity, albeit with a much different flavour, and does not seem to provide explicit rates. After the first version of the present paper appeared online, the approach of \cite{dolbeault2015hypocoercivity} was extended in \citet{andrieu2018hypercoercivity} to cover several PDMPs, including BPS, Zig-Zag and RHMC.
The approach in \cite{dolbeault2015hypocoercivity} and \cite{andrieu2018hypercoercivity} is quite distinct to ours.
In particular \cite{andrieu2018hypercoercivity} also obtain dimension-free bounds for RHMC under similar assumptions; their explicit rates depend on various additional parameters of the problem and therefore a detailed comparison with the explicit rates in our Theorem~\ref{thm:hypoco} was not performed in \cite{andrieu2018hypercoercivity}. Their approach is quite general but much less direct for RHMC than ours, as they rely on generic results by Dolbeault, Mouhot, and Schmeiser. Due to this additional layer of complexity, it does not seem possible to obtain as sharp bounds for RHMC as with our more direct approach.
In addition the bounds of \cite{andrieu2018hypercoercivity} for BPS suggest that its computational cost scales like $O(d^2)$. This seems to capture the worst case scenario and agrees for example with results \cite{bierkens2018high} for the log-density of the target, which recommends scaling the refreshment rate with the dimension. Our results suggest that when one is interested in low-dimensional projections, then it is computationally more efficient to not scale the refreshment rate with the dimension, achieving computational cost of order $O(d^{3/2})$. Empirical results in Section \ref{sec:empirical} seem to suggest that this may also be the case for certain classes of functions depending on all the coordinates, such as the sum of all coordinates. A common scenario where this type of scaling limit is extremely relevant is for example that of Bayesian inference where typically one may only be interested in estimating the posterior means, variances and covariances of the high-dimensional state components (this is a set of one and two dimensional marginals). Finally, it is intuitively clear that the log-density will not mix well in a high-dimensional target for the global BPS, see \cite{bierkens2018high} for a detailed study. We conjecture that the functions that exhibit this type of behaviour form a low-dimensional sub-space of $L^2(\pi)$. Recently \cite{bierkens2019spectral} has obtained very detailed results on the whole spectrum of the one-dimensional Zig-Zag process, it would be interesting if similar results could be obtained for BPS in high dimensional scenarios.
Apart from the intrinsic interest of the RHMC process, our motivation for studying its convergence rates is as follows. In the scaling literature for MCMC the limiting processes are usually Langevin diffusions. These have very well understood convergence rates which, at least under additional assumptions, are dimension-free. Therefore, in high-dimensions the cost of running the (time-rescaled) algorithm serves as a proxy for its computational complexity. In our case, the algorithm ran on its natural time scale converges to RHMC, which
as we establish here, also enjoys dimension-free convergence rates under appropriate assumptions. Therefore the cost of running BPS for a unit of process time serves as a proxy
for its algorithmic complexity.
The next section contains the statements of the main results of the paper along with necessary notation and definitions.
The remaining sections contain the proofs of the main results.
\section{Main results}
\subsection{Notation}
Let $k\geq 1$. For vectors $u,v\in \mathbb{R}^k$ we write $|v|$ and $u\cdot v$ for the Euclidean norm and inner product respectively, whilst for a function $f:\mathbb{R}^k \mapsto \mathbb{R}$ we write $\nabla f, \nabla^2 f$ for its (weak) gradient and Hessian respectively.
When considering functions $f=f(a,b)$, where $a,b\in\mathbb{R}^k$, that is $f:\mathbb{R}^{2k}\mapsto \mathbb{R}$, we will write $\nabla_a f$, $\nabla_b f$ to denote the gradient with respect to $a\in\mathbb{R}^{k}$ and $b\in\mathbb{R}^{k}$ variables respectively.
For vector valued functions $f:\mathbb{R}^{d}\to \mathbb{R}^k$, we will write $\nabla f$ for the Jacobian matrix of derivatives.
For a locally compact Hausdorff space $\mathcal{Z}$, let $C_0(\mathcal{Z})$ denote the space of continuous functions $f:\mathcal{Z}\mapsto \mathbb{R}$ that vanish at infinity, that is $f\in C_0(\mathcal{Z})$ if $f$ is continuous and for any $\epsilon$,
there exists a compact set $K_\epsilon\subset \mathcal{Z}$ such that $|f(z)|<\epsilon$ for all $z\notin K_\epsilon$.
Recall that $C_0(\mathcal{Z})$ is a Banach space with respect to the $\|\cdot\|_\infty$ norm, which is defined as usual through $\|f\|_\infty = \sup |f|$.
Also let $C_c^\infty (\mathcal{Z})$ be the space of infinitely differentiable functions $f:\mathcal{Z}\mapsto \mathbb{R}$ with compact support.
For $d\geq 1$, let $\mathcal{Z}:=\mathbb{R}^d\times \mathbb{R}^d$. For $n\geq1$, define the Borel probability measure $\pi_n(\mathrm{d} \bm{z})$ on $\mathcal{Z}^n$ with density w.r.t. Lebesgue measure given by
$$\pi_n(\bm{z}) = \pi_n(\bm{x},\bm{v})\propto \exp\left\{ -U_n(\bm{x}) - \bm{v}^T \bm{v}/2\right\},$$
where $U_n:\mathbb{R}^{n\times d}\mapsto \mathbb{R}_+$ is a potential.
For a measure $\pi$ on $\mathcal{Z}$, we will write $L^2(\pi)$ for the usual Hilbert space, and
$\langle \cdot,\cdot\rangle, \| \cdot\|$
to denote the inner product and norm in $L^2(\pi)$ respectively, whereas $L_0^2(\pi)$ will denote the orthogonal complement of the constant functions, i.e., functions with mean zero under the distribution $\pi$. Finally for $f:\mathcal{Z}\to \mathbb{R}^d$ and $g:\mathcal{Z}\to \mathbb{R}^d$ we will write
$$\langle f, g\rangle = \int \pi(\mathrm{d} z) \langle f(z), g(z) \rangle.$$
We also define
$$H^1:= H^1(\pi):= \left\{h\in L_0^2(\pi): \nabla_x h, \nabla_v h\in L^2(\pi) \right\},$$
the Sobolev space of functions in $L^2(\pi)$ with weak derivatives in $L^2(\pi)$ and for $f,g\in H^1(\pi)$
we will denote the inner product and norm on $H^1(\pi)$ with $\langle \cdot ,\cdot\rangle_{H^1(\pi)}$ and $\|\cdot \|_{H^1(\pi)}$ respectively, where
$$\langle f, g\rangle_{H^1(\pi)}= \langle \nabla_x f, \nabla_x g\rangle+ \langle \nabla_v f, \nabla_v g\rangle.$$
\subsection{The Bouncy Particle Sampler}
For $(\bm{x},\bm{v}) \in \mathcal{Z}^n$, define
\begin{equation}\label{eq:bounceoperator}
R_n(\bm{x})\bm{v} := \bm{v} - 2 \frac{\langle \nabla U_n(\bm{x}), \bm{v} \rangle}{|\nabla U_n(\bm{x})|^2}\nabla U_n(\bm{x}).
\end{equation}
The vector $R_n(\bm{x})\bm{v}$ can be interpreted as a Newtonian collision on the hyperplane orthogonal to the gradient of the potential $U_n$, hence the interpretation of $\bm{x}$ as a position, and $\bm{v}$, as a velocity.
The Bouncy Particle Sampler (BPS), first introduced in \cite{peters2012rejection} and in a statistical context in \cite{bouncy2018}, defines a ${\pi_n}$-invariant, non-reversible, piecewise deterministic Markov process
$\{\bm{Z}_n(t):t\geq 0\}=\{(\bm{X}_t, \bm{V}_t): t\geq 0\}$ taking values in $\mathcal{Z}^n$ whose generator
$\mathcal{A}_n$, for smooth enough functions $f:\mathcal{Z}^n \mapsto \mathbb{R}$, is given by
\begin{align*}
\mathcal{A}_n f(\bm{x},\bm{v})
&= \langle \nabla f(\bm{x}), \bm{v})\rangle + \max\{0, \langle \nabla U_n(\bm{x}), \bm{v}\rangle\}
\left[\mathfrak{R}_nf\left(\bm{x},\bm{v}\right) - f\left(\bm{x},\bm{v}\right) \right]\\
&\qquad
+ \lambda_{\textrm{ref}} \left[ Q_{\alpha,n} f \left(\bm{x},\bm{v}\right)-f\left(\bm{x},\bm{v}\right)\right],
\end{align*}
where
\begin{equation*}
\mathfrak{R}_n f\left(\bm{x},\bm{v}\right)
:= f\left(\bm{x},R_n(\bm{x})\bm{v}\right),
\quad Q_{\alpha,n} f\left(\bm{x},\bm{v}\right)
:= \frac{1}{(2\pi)^{nd/2}}\int_{\mathbb{R}^{nd}} \mathrm{e}^{-|\boldsymbol{\xi}|^2/2} f\left(\bm{x},\alpha\bm{v}+\sqrt{1-\alpha^2} \boldsymbol{\xi}\right) \mathrm{d} \boldsymbol{\xi},
\end{equation*}
for $0\leq\alpha<1$ and a positive refreshment rate $\lambda_{\textrm{ref}}>0$.
We also write $\bm{Z}_n(t)=\left(Z_n^{(1)}(t), \dots, Z_n^{(n)}(t)\right)$ where $Z_n^{(k)}(t)=(X_n^{(k)}(t), V_n^{(k)}(t)) \in \mathcal{Z}$ is the $k$-th component.
The original formulation of the BPS algorithm corresponds to $\alpha=0$, that is refreshment occurs independently. The generalization $\alpha > 0$ \cite{vanetti2017piecewise} consists in refreshments that are performed according to an auto-regressive process.
\subsection{Randomized Hamiltonian Monte Carlo}
We define here RHMC as this is the process we will obtain as the weak limit of $Z_n^{(1)}(t)=(X_n^{(1)}(t), V_n^{(1)}(t)) \in \mathcal{Z}$ as $n\to\infty$. Define the Hamiltonian
\begin{equation}\label{eq:Hamiltonian}
H(x,v)=U(x) + |v|^2/2,
\end{equation}
for $(x,v)\in \mathcal{Z}$ and the corresponding probability density on $\mathcal{Z}$
\begin{equation}\label{targetRHMC}
\pi(x,v)=\overline{\pi}(x) \cdot \psi(v) \propto \exp\{-U(x)-|v|^2/2\}.
\end{equation}
The Hamiltonian dynamics associated to (\ref{eq:Hamiltonian}) is an ordinary differential equation in $\mathcal{Z}$ of drift $(\nabla_v H,-\nabla_x H)=(v,-\nabla U)$.
For $f\in C^{\infty}_c (\mathcal{Z})$, the RHMC generator is then given by
\begin{equation}\label{eq:RHMC_generator}
\mathcal{A}f(x,v) = \langle \nabla_x f, v\rangle - \langle \nabla_v f, \nabla U\rangle
+ \lambda_{\textrm{ref}} \left[Q_\alpha f(x,v) - f(x,v)\right],
\end{equation}
where $\lambda_{\textrm{ref}}>0$ and
\begin{equation}
Q_\alpha f(x,v):= \frac{1}{(2\pi)^{d/2}}\int \mathrm{e}^{-|{\xi}|^{2}/2}
f\left(x,\alpha v+\sqrt{1-\alpha^2} \xi\right) \mathrm{d} \xi,
\end{equation} for some $0\leq\alpha<1$.
We will write $\{P^t: t\geq 0\}$ for the semi-group generated by $\mathcal{A}$, which is $\pi$-invariant, and
$\{Z_t:t \geq 0\}$ for the associated RHMC process.
The RHMC process thus corresponds to the Hamiltonian dynamics associated with $H$, with the velocity/momentum being refreshed at the arrival times of an independent homogeneous Poisson process of intensity $\lambda_{\textrm{ref}}$. The refreshment is done in an auto-regressive manner. From now on, we will restrict ourselves for BPS and RHMC to $0<\alpha<1$. The reason for using $\alpha>0$ is that it allows us to establish the Feller property which greatly simplifies the rest of the proofs.
Since the autoregressive process mixes exponentially fast there is no loss in terms of mixing potentially at the cost of more frequent refreshments, something which has also been observed empirically.
\subsection{Main results}
\subsubsection{RHMC as Scaling Limit of BPS}
Before stating our weak convergence result, we will make some assumptions.
\begin{assumption}\label{ass:iid}
We have $d=1$ and the potential $U_n:\mathbb{R}^n\mapsto \mathbb{R_+}$ takes the form
\begin{equation}
U_n(\bm{x}) = U_n(x_1, \dots, x_n) = \sum_{i=1}^n U(x_i),
\end{equation}
for $\bm{x}=(x_1,\dots, x_n) \in \mathbb{R}^n$ and a potential $U:\mathbb{R}\mapsto \mathbb{R_+}$. \end{assumption}
\begin{assumption}\label{ass:feller}
The potential $U:\mathbb{R}\mapsto \mathbb{R_+}$ is continuously differentiable.
\end{assumption}
For some of our results, we will actually need the following stronger assumption.
\begin{assumption}\label{ass:weakconv}
The potential $U$ belongs to $C^{\infty}(\mathbb{R})$ and $\|U''\|_\infty\leq M < \infty$.
\end{assumption}
\begin{assumption}\label{ass:gradintegral}
We have
$$\int \mathrm{e}^{-U(x)} |U'(x)|^2\mathrm{d} x <\infty.$$
\end{assumption}
The following theorem is our first main result.
\begin{theorem}\label{thm:weakconv}
Suppose Assumptions~\ref{ass:iid}, \ref{ass:weakconv} and \ref{ass:gradintegral} hold, that $0<\alpha<1$ and that the BPS process $\{\bm{Z}_n(t):t\geq 0\}$ is initialized
at stationarity, i.e., $\bm{Z}_n(0)\sim \pi_n$. Then the process $\{Z_n^{(1)}(t):t\geq 0\}$ corresponding to the first location and velocity components of the BPS process converges weakly to the RHMC process $\{Z_t:t\geq 0\}$ as $n\to\infty$.
\end{theorem}
\begin{remark}
To illustrate Theorem~\ref{thm:weakconv} in Figure \ref{fig:weakconv} we have plotted the paths of the BPS process and the equi-energy contours of the Hamiltonian corresponding to the deterministic dynamics of RHMC.
The target distribution has potential $U(x)=|x|^b/2$ and we have tested two values of $b$, $b=2$ (Gaussian) and $b=4$. These figures show the first coordinate of the position and velocity vectors. As we can see, as the dimension increases, the paths of BPS indeed appear more and more similar to the contours of the Hamiltonian.
\begin{figure}
\centering
\begin{subfigure}[t]{0.49\textwidth}
\raisebox{-\height}{\includegraphics[width=\textwidth]{beta2d10.pdf}}
\caption{$b=2$, $d=10$}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.49\textwidth}
\raisebox{-\height}{\includegraphics[width=\textwidth]{beta2d100.pdf}}
\caption{$b=2$, $d=100$}
\end{subfigure}
\begin{subfigure}[t]{0.49\textwidth}
\raisebox{-\height}{\includegraphics[width=\textwidth]{beta4d10.pdf}}
\caption{$b=4$, $d=10$}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.49\textwidth}
\raisebox{-\height}{\includegraphics[width=\textwidth]{beta4d100.pdf}}
\caption{$b=4$, $d=100$}
\end{subfigure}
\caption{Convergence of the BPS process to RHMC in high dimensions for $U(x)=|x|^b/2$.}
\label{fig:weakconv}
\end{figure}
\end{remark}
\begin{remark}
Theorem~\ref{thm:weakconv} can be straightforwardly extended to the scenario where $d>1$ and to any finite number of coordinates.
In addition, it will be clear from the proof that the result can also be extended to non
i.i.d.\ scenarios; roughly speaking it is enough to have the following conditional, self-normalised central limit theorem,
$$ \frac{\langle \nabla U_n (\bm{X}_n), \bm{V}_n\rangle}{\| \nabla U_n (\bm{X}_n)\|} \Big| X_1, V_1\Rightarrow \mathcal{N}(0, 1),$$
as $n\to \infty$ with $\bm{Z}_n = (\bm{X}_n, \bm{V}_n) \sim \pi_n$.
For example one could have a potential with local interactions of the form
$$U_n(x_1, \dots, x_n) = U_1(x_1) + \sum_{i=2}^n U (x_{i-1}, x_{i}),$$
under which, when $(X_1, \dots, X_n)\sim \pi_n$, $(X_i)_{i\geq 1}$ forms a Markov chain. In this scenario the above central limit theorem would hold under regularity assumptions.
\end{remark}
\subsubsection{Dimension-free Convergence Rates for RHMC}
We consider the RHMC process on the target (\ref{targetRHMC}) defined on $\mathcal{Z}:=\mathbb{R}^d\times \mathbb{R}^d$ for ${\pi}(x)$ a strongly log-concave target distribution on $\mathbb{R}^d$ having a potential with bounded Hessian. This is a standard assumption adopted in \cite{bou2018coupling,mangoubi2017rapid,dalalyan2017theoretical,dwivedi2018log,durmus2017nonasymptotic}.
\begin{assumption}\label{ass:hypoco}
Assume that $U\in C^2(\mathbb{R}^d)$ and that for some $0<m<M$, and all $x,v\in \mathbb{R}^d$
\begin{equation}
m \langle v, v\rangle \leq \langle v, \nabla^2 U(x) v\rangle \leq M \langle v, v\rangle.
\end{equation}
\end{assumption}
The following proposition shows that the expected number of bounces per unit time for BPS in stationary distribution is $O(\sqrt{d})$.
\begin{proposition}\label{proposition:BPSexpectednbofbounces}
Suppose that $\pi(x)=\exp(-U(x))$ is a distribution on $\mathbb{R}^d$. Then the BPS process on $\mathcal{Z}$ initiated in the stationary distribution has a expected number of bounces per unit time that equals
\[\Lambda_{b}:=\mathbb{E}_{X\sim \pi, Z\sim N(0,\mathbb{I}_d)}\l( \langle \nabla U(X),Z\rangle_+\r),\]
for any choice of refreshment rate $\lambda_{\textrm{ref}}$ and auto-regressive parameter $\alpha$. Moreover, if $\pi$ satisfies Assumption \ref{ass:hypoco}, then we have
\begin{equation}
\frac{\sqrt{m (d-1/2)}}{\sqrt{2\pi}}\le \Lambda_{b}\le \frac{\sqrt{M d}}{\sqrt{2\pi}}.
\end{equation}
\end{proposition}
\paragraph{Wasserstein distance.}
For $t\ge 0$, let $Z^{(1)}(t)=(X^{(1)}(t),V^{(1)}(t))$ denote a path of the RHMC process. We couple this with another path $Z^{(2)}(t)=(X^{(2)}(t), V^{(2)}(t))$ such that their refreshments happen simultaneously and the same multivariate normal random variables are used for updating their velocities. Then the coupled process $\left(Z^{(1)}(t),Z^{(2)}(t)\right)$ is Markov and we write $L_{1,2}$ for the corresponding generator.
Notice that the $2\times 2$ real valued matrix
\begin{equation}
A:=\l(\begin{matrix}a &b\\ b &c\end{matrix}\r), \label{eq:matrixA}
\end{equation}
is positive definite, denoted $A\succeq 0$, if and only if $a>0$, $c>0$ and $b^2<ac$. For such a matrix, let
\begin{align*}&d_A^2(Z_1(t),Z_2(t)):=\\
&a \|X^{(2)}(t)-X^{(1)}(t)\|^2 + 2 b \inner{X^{(2)}(t)-X^{(1)}(t)}{V^{(2)}(t)-V^{(1)}(t)}+c\|V^{(2)}(t)-V^{(1)}(t)\|^2
\end{align*}
denote a distance function called weighted Wasserstein distance. It is equivalent up to constant factors to the standard Wasserstein distance on $\mathcal{R}^{2d}$ and the standard Wasserstein distance corresponds to the special case $a=1$, $b=0$, $c=1$. However, due to the effect of the generator $L_{1,2}$ on $d_A^2(Z_1(t),Z_2(t))$, it will never be a contraction when $b=0$, and thus weighting the Wasserstein distance is essential for obtaining convergence rates.
By the Brascamp-Lieb inequality (\cite{BrascampLieb}), it follows that for mean zero Lipschitz-functions $f\in L_0^{2}(\pi)$ we have
\begin{equation}
\|f\|^2\le \max\l(\frac{1}{m}, 1\r) \|f\|_{\mathrm{Lip}}^2,
\end{equation}
where $\|f\|_{\mathrm{Lip}}$ denotes the Lipschitz coefficient of $f:\mathcal{R}^{2d}\to \mathcal{R}$ with respect to the Euclidean distance. Our main result in this section is the following.
\begin{theorem}\label{thm:Wasserstein}
Suppose that $0\le \alpha<1$, Assumption~\ref{ass:hypoco} holds and let
$$\lambda_{\textrm{ref}}=\frac{1}{1-\alpha^2}\l(2\sqrt{M+m}-\frac{(1-\alpha) m}{\sqrt{M+m}}\r), \qquad \mu=\frac{(1+\alpha)m}{\sqrt{M+m}}-\frac{\alpha m^{3/2}}{2(M+m)}.$$
Then there exist constants $a$, $b$ and $c$ depending on $m$, $M$ and $\alpha$, such that the corresponding matrix $A$ is positive definite, and for any $t \geq 0$ we have
\begin{equation}\label{eq:Wassersteincontraction1}
L_{1,2} \ d_A^2(Z_1(t),Z_2(t))\le -\mu\cdot d_A^2(Z_1(t),Z_2(t)).
\end{equation}
Moreover, for every $f\in L_0^{2}(\pi)$, we have
\begin{equation}\label{eqL2bndWass}
\|P^t f \|^2 \leq \min(C e^{-\mu t},1) \|f\|^2,
\end{equation}
where $C=\frac{ac+b^2+2\sqrt{ac b^2}}{ac-b^2}$.
\end{theorem}
\begin{remark}
Due to the non-reversibility of RHMC, the convergence rates in Wasserstein distance do not directly imply bounds on the asymptotic variance for every function in $L^2(\pi)$,
but only for Lipschitz functions. In the next section, we obtain convergence rates based on hypocoercivity. This approach will allow to obtain variance bounds
for a much larger class of functions.
\end{remark}
As we shall see in the next proposition, it is possible to obtain sharper convergence rates for Gaussian target distributions. For this result, we generalise the weighted Wasserstein distance and consider distances of the form
\begin{equation}
d_D^2(Z_1(t),Z_2(t)):=\inner{Z_2(t)-Z_1(t)}{D (Z_2(t)-Z_1(t))},
\end{equation}
where $D$ is a real valued $2d\times 2d$ positive definite matrix.
\begin{proposition}\label{prop:WassersteinGauss}
Suppose that $\overline{\pi}$ is Gaussian and its inverse covariance matrix $H$ satisfies $m I\preceq H\preceq MI$.
Let $$\lambda_{\textrm{ref}}=\frac{2\sqrt{m}}{1-\alpha}, \qquad \mu=\frac{\sqrt{m}}{3}.$$
Then there exists a $2d\times 2d$ real valued matrix $D$ such that for any $t \geq 0$ we have
\begin{equation}
L_{1,2} \ d_D^2(Z_1(t),Z_2(t))\le -\mu\cdot d_D^2(Z_1(t),Z_2(t)).
\end{equation}
Moreover, for every $f\in L_0^{2}(\pi)$, we have
\begin{equation}\label{eqL2bndGaussian}
\|P^t f \|^2 \leq \min(C e^{-\mu t},1) \|f\|^2,
\end{equation}
where $C=\frac{ac+b^2+2\sqrt{ac b^2}}{ac-b^2}$.
\end{proposition}
\paragraph{Hypocoercivity}
Our next convergence result is based on the hypocoercivity approach; see, e.g., \cite{monmarche2014hypocoercive,nier2005hypoelliptic,dric2009hypocoercivity,dolbeault2015hypocoercivity,roussel2017spectral}.
Our result will be stated in terms of the following modified Sobolev norm
\begin{equation}\label{eq:newnorm}
\langle\!\langle h, h\rangle\!\rangle:=a\|\nabla_v h\|^2 - 2b \langle \nabla_x h, \nabla_v h\rangle + c\|\nabla_x h\|^2,
\end{equation}
which again for $a,c>0$ and $b^2<ac$ defines a norm equivalent to the $H^1$ norm.
In particular following the calculations in \cite{dric2009hypocoercivity}, by Young's inequality we get
\begin{align*}
\left( 1+ \frac{|b|}{\sqrt{ac}}\right) \left[ a\|\nabla_v h\|^2+c\|\nabla_x h\|^2\right]\ge \langle\!\langle h, h\rangle\!\rangle
&\geq \left( 1- \frac{|b|}{\sqrt{ac}}\right) \left[ a\|\nabla_v h\|^2+c\|\nabla_x h\|^2\right].
\end{align*}
By the Efron-Stein inequality (\cite{EfronStein}) and the fact that $\pi(x,v)=\overline{\pi}(x)\psi(v)$ is the product of two independent distributions, we have
\begin{align*}
\|h\|^2=\mathrm{Var}_{\pi}(h)\le \mathrm{Var}_{\psi}(\mathbb{E}_{\overline{\pi}}(h))+\mathrm{Var}_{\overline{\pi}}(\mathbb{E}_{\psi}(h)),
\end{align*}
for any $h \in L_0^2(\pi)$.
Now by using Brascamp-Lieb inequality (\cite{BrascampLieb}) and the strong log-concavity of the distributions $\overline{\pi}$ and $\psi$, we obtain that
\[a\|\nabla_v h\|^2+c\|\nabla_x h\|^2\ge a\cdot 1 \cdot \mathrm{Var}_{\psi}(\mathbb{E}_{\overline{\pi}}(h))+c\cdot m\cdot \mathrm{Var}_{\overline{\pi}}(\mathbb{E}_{\psi}(h))\ge \min(a, cm) \|h\|^2.
\]
Therefore convergence in the $\langle\!\langle \cdot, \cdot \rangle\!\rangle$ norm implies convergence in $L_0^2(\pi)$.
\begin{theorem}\label{thm:hypoco}
Suppose that Assumptions~\ref{ass:feller}, \ref{ass:weakconv} and \ref{ass:hypoco} hold.
Let
$$\lambda_{\textrm{ref}}=\frac{1}{1-\alpha^2}\l(2\sqrt{M+m}-\frac{(1-\alpha) m}{\sqrt{M+m}}\r), \qquad
\mu=\frac{(1+\alpha)m}{\sqrt{M+m}}-\frac{\alpha m^{3/2}}{2(M+m)}.$$
Then there exist constants $a,b,c$ depending on $m$, $M$ and $\alpha$ such that $a>0, c>0, b^2<a c$, and for every $f \in \mathcal{D}({B})\subset H^1(\pi)\subset L_0^2(\pi)$(with $B$, $\mathcal{D}({B})$ as defined in \eqref{eq:definitionofB}),
\begin{equation}\frac{\mathrm{d} }{\mathrm{d} t} \langle\!\langle P^t f, P^t f\rangle\!\rangle\le -\mu \langle\!\langle P^t f, P^t f\rangle\!\rangle.
\label{eq:hypoco_rate}
\end{equation}
Moreover, for every $f\in L_0^{2}(\pi)$ and $t\ge 0$, we have
\begin{equation}
\|P^t f \|^2 \leq \min(C e^{-\mu t},1) \|f\|^2,
\end{equation}
where $C=\frac{ac+b^2+2\sqrt{ac b^2}}{ac-b^2}$.
\end{theorem}
\begin{remark}
Since the first-coordinate process of BPS converges to RHMC, whose mixing we established above, in the natural time-scale
the computational cost of running BPS for one time unit serves as a proxy for its algorithmic complexity. This cost is proportional to the number of total events per time unit, including bounces and refreshments. Proposition \ref{proposition:BPSexpectednbofbounces} shows that the expected number of bounces per unit time under Assumption \ref{ass:hypoco} is at least $\frac{\sqrt{m(d-1/2)}}{2\sqrt{\pi}}$, which is much larger than the expected number of refreshments ($\lambda_{\textrm{ref}}$) if the refreshment rate is chosen as recommended by Theorems \ref{thm:Wasserstein} and \ref{thm:hypoco} (as long as $M/m\ll d$ and $\alpha$ is not too close to 1). Therefore in these cases it is justified to choose $\lambda_{\textrm{ref}}$ in order to maximize the contraction rate $\mu$ of the limiting RHMC process.
Since each bounce has a computational cost of order $O(1)$ per gradient evaluation, our results suggests that BPS scales like $O(d^{1/2})$ in gradient evaluations under our assumptions. This is the scaling observed in the simulations presented in the next section. It should be noted that this scaling could be quite different for strongly dependent targets.
\end{remark}
\subsection{Empirical results for different functions}\label{sec:empirical}
In this section, we show some simulation results about the computational cost of the BPS for a $d$ dimensional standard normal target, and seven different test functions defined as follows,
\begin{align*}
f_1(x)&=x_1 \quad (\text{first coordinate}),\\
f_2(x)&=\sum_{i=1}^{d}x_i \quad (\text{sum of all coordinates}),\\
f_3(x)&=\sum_{i=1}^{d-1}\sin(x_i+x_{i+1}) \quad (\text{a sum of sines depending on two component each}),\\
f_4(x)&=\|x\| \quad (\text{radius}),\\
f_5(x)&=\frac{\|x\|^2}{2}=\sum_{i=1}^{d}\frac{x_i^2}{2} \quad (\text{log-density}),\\
f_6(x)&=x_1^2 \quad (\text{square of first coordinate}),\\
f_7(x)&=x_1 x_2 \quad (\text{product of first and second coordinates}).
\end{align*}
In order to estimate the effective sample sizes, we have run 100 parallel BPS simulations with $10^6$ events per simulation, starting from the Gaussian target distribution. The autoregressive parameter $\alpha$ was set as $\alpha=0$.
Figure \ref{fig:empirical} shows the number of events required for one effective sample for dimensions $d=10$, $100$, $1000$ and $10000$ for these 7 functions, with refreshment parameter choices $\lambda_{\textrm{ref}}=1$ (as suggested by Theorems \ref{thm:Wasserstein} and \ref{thm:hypoco}) and $\lambda_{\textrm{ref}}=\sqrt{d}$ (as suggested by \cite{bierkens2018high} and Table 1 of \cite{andrieu2018hypercoercivity}). The number of events is a correct proxy for the computational cost as each event requires one gradient evaluation (see Section 2.3 of \cite{bouncy2018} for the description of the implementation of BPS for Gaussian targets). As we can see, if the refreshment rate is chosen as $\lambda_{\textrm{ref}}=1$, these simulation results show $O(\sqrt{d})$ scaling in the number of events required for an effective sample for all of the functions except the radius and the log-density ($f_4$ and $f_5$). In contrast, the choice $\lambda_{\textrm{ref}}=\sqrt{d}$ seems to require significantly more events per effective sample, with $O(d)$ scaling observed empirically. In the cases of the radius and the log-density, the choice $\lambda_{\textrm{ref}}=\sqrt{d}$ still seems to require $O(d)$ events per effective sample, while $\lambda_{\textrm{ref}}=1$ is doing worse, approximately $O(d^{4/3})$ events per effective sample is required. The scaling limits for this function were studied in \cite{bierkens2018high}, who has recommended choosing $\lambda_{\textrm{ref}}=O(\sqrt{d})$ to obtain the best mixing for the log-density, consistently with our empirical results.
To sum up, we can see that if the goal of the simulation is to estimate the posterior mean or posterior covariance matrix, or other quantities only depending a small subset of the coordinates, then choosing $\lambda_{\textrm{ref}}$ as recommended by Theorems \ref{thm:Wasserstein} and \ref{thm:hypoco} yield good empirical performance ($O(\sqrt{d})$ scaling in the number of events required for an effective sample). For functions depending on all of the coordinates the situation is more complicated, and the best choice of $\lambda_{\textrm{ref}}$ is strongly function dependent in this case.
\begin{figure}
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=.8\linewidth]{f1.pdf}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=.8\linewidth]{f2.pdf}
\end{subfigure}\\
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=.8\linewidth]{f3.pdf}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=.8\linewidth]{f4.pdf}
\end{subfigure}\\
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=.8\linewidth]{f5.pdf}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=.8\linewidth]{f6.pdf}
\end{subfigure}\\
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=.8\linewidth]{f7.pdf}
\end{subfigure}%
\caption{Number of BPS events per effective sample for 7 different functions}
\label{fig:empirical}
\end{figure}
\section{Proof of Weak Convergence Result - Theorem~\ref{thm:weakconv}}
The proof will be based on a sequence of auxiliary results.
First we will show that the RHMC semigroup $\{P^t: t \geq 0\}$ is Feller, and that the space $C_c^\infty$ is a \textit{core} for its generator given in \ref{eq:RHMC_generator}, in the sense that
$C_c^\infty$ is dense in $\mathcal{D}(\mathcal{A})$ with respect to the norm $\vvvert \cdot \vvvert := \|f\|_\infty + \|\mathcal{A} f \|_\infty$.
This, and a sequence of auxiliary results, will allow us to
apply \cite[Corollary~8.6]{EK_05} to prove Theorem~\ref{thm:weakconv}.
\subsection{Feller property}
Recall that in the context of Theorem ~\ref{thm:weakconv}, we have $d=1$ and $\mathcal{Z}=\mathbb{R}^{2}$. A Markov process taking values in $\mathcal{Z}$,
with transition semigroup $\{P^{t}:t\geq0\}$, is called a Feller
process and $\{P^t:t\geq 0\}$ a Feller semigroup, if it satisfies the following two properties
\begin{description}
\item [{Feller property:}] for all $t\geq0$ and $f\in C_{0}(\mathcal{Z})$ we have
$P^{t}f\in C_{0}(\mathcal{Z})$, and
\item [{Strong continuity:}] $\|P^{t}f - f\|_\infty \to 0$ as $t\to0$ for $f\in C_{0}(\mathcal{Z})$.
\end{description}
\begin{proposition}\label{lem:feller}
Under Assumption~\ref{ass:feller} the RHMC process $\{Z_t\}_{t\geq 0}$ with generator $\mathcal{A}$ given by \ref{eq:RHMC_generator}, with $\alpha \in (0,1)$ and $\lambda_{\textrm{ref}}>0$ is a Feller process. If Assumption~\ref{ass:weakconv} also holds, then $C_c^\infty$ is a core for its generator.
\end{proposition}
Note that a more technical approach proposed recently in \citet{holderrieth2019core} requires weaker assumptions.
\subsubsection{Proof of Proposition~\ref{lem:feller}}
Before we proceed let us first define the \textit{resolvent operator} for $\lambda > 0$
$$\mathcal{R}_\lambda f(z) := (\lambda I-\mathcal{A})^{-1} f(z)= \int_0^\infty \mathrm{e}^{-\lambda s} P^s f(z) \mathrm{d} s
=\int_0^\infty \mathrm{e}^{-\lambda s} \mathbb{E}^z \left[ f(Z_s)\right] \mathrm{d} s.$$
The proof will proceed as follows. First we will first show that
$\mathcal{R}_\lambda: C_0(\mathcal{Z}) \to C_0(\mathcal{Z})$,
and then use \cite[Corollary~1.23]{Bottcher13} to establish that $\{P^t:t\geq 0\}$ has the Feller property, that is for all $t\geq 0$ $P^t: C_0(\mathcal{Z}) \to C_0(\mathcal{Z})$.
Once the Feller property is established, by \cite[Lemma~1.4]{Bottcher13}, to prove strong continuity it suffices to prove the weaker statement
$P^t f(z) \to f(z)$, for all $f\in C_0(\mathcal{Z})$ and $z \in \mathcal{Z}$.
We now establish this property.
Let $T_1, T_2, \dots$ be the arrivals times of the jumps. Then we have for $h> 0$
\begin{align*}
P^h f(z) - f(z)
&= \mathbb{E}^z \left[ f(Z_h)\right] - f(z)\\
&= \mathbb{E}^z \left[ f(Z_h) \mathbb{1} \{T_1 \geq h \}\right] - f(z)
+ \mathbb{E}^z \left[ f(Z_h) \mathbb{1} \{T_1 < h \}\right]\\
&= f\left( \Xi(h,z)\right)\mathrm{e}^{-\lambda_{\textrm{ref}} h} - f(z)
+ \mathcal{E},
\end{align*}
where we write $\Xi(z,t)$ for the solution of the Hamiltonian dynamics at time $t$ initialized at $z_0=z$. It is well-known that if $H:\mathbb{R}\times \mathbb{R} \to \mathbb{R}$ is continuously differentiable everywhere then $\Xi(z,s)$ is well defined for all $s>0$ (see for example \cite[Theorem~1.186]{Chi}), $H\big(\Xi(z,s)\big)=H(z)$ for all $s>0$ and
$\Xi(z,h) \to z$.
Since $f$ is bounded it easily follows that as $h\to 0$
$$|\mathcal{E}| \leq \|f\|_\infty (1-\mathrm{e}^{-\lambda_{\textrm{ref}} h}) \to 0.$$
Since $\Xi(z,h)\to z$ as $h\to 0$, the result follows.
\paragraph{Proof of the Feller property.} From \cite[Equation~2.6]{DuCo_08} we know that we can express the resolvent kernel as follows for a measurable set $A$
\begin{equation}\label{eq:resolvent_sum}
\mathcal{R}_\lambda(z,A) = \sum_{j=0}^\infty J_\lambda^j K_\lambda (z,A),
\end{equation}
where
\begin{align}
K_\lambda (z,A)
&:=\int_0^\infty \mathrm{e}^{-\lambda s -\lambda_{\textrm{ref}} s} \mathbb{1}_A \left( \Xi(z,s)\right) \mathrm{d} s,\\
J_\lambda (z,A)
&:=\int_0^\infty \lambda_{\textrm{ref}} \, \mathrm{e}^{-\lambda s -\lambda_{\textrm{ref}} s} Q\left(\Xi(z,s), A \right)
\mathrm{d} s,
\end{align}
with $\Xi(z,s)=\Xi\big((x,v),s\big)$ as defined above.
We will now show that $\mathcal{R}_\lambda f \in C_0(\mathcal{Z})$ for any $f\in C_0(\mathcal{Z})$.
This follows from the next result.
\begin{lemma}\label{lem:resolvent}
Assume that Assumption~\ref{ass:feller} holds and let $f\in C_0(\mathcal{Z})$. Then, for any $\lambda>0$, we have $J_\lambda f, K_\lambda f \in C_0(\mathcal{Z})$ and $\|J_\lambda f\|_\infty\leq \lambda_{\textrm{ref}}/(\lambda+ \lambda_{\textrm{ref}})\|f\|_\infty$. In particular
$$\mathcal{R}_\lambda f = \sum_{j=0}^\infty J_\lambda^j K_\lambda f \in C_0(\mathcal{Z}).$$
\end{lemma}
\begin{proof}[Proof of Lemma~\ref{lem:resolvent}]
Let $\lambda >0$ and let us first look at $K_\lambda$.
Suppose now that $f\in C_0(\mathcal{Z})$ and that $z_n\to z$. Then
\begin{align*}
|K_\lambda f(z) - K_\lambda f(z_n)|
&\leq \int_0^\infty \lambda_{\textrm{ref}} \, \mathrm{e}^{-\lambda s -\lambda_{\textrm{ref}} s} |f \left(\Xi(z,s)\right)-f \left(\Xi(z_n,s)\right)| \mathrm{d} s \to 0,
\end{align*}
by the bounded convergence theorem, since $f$ is bounded and the functions
$s\mapsto |f \left(\Xi(z,s)\right)-f \left(\Xi(z_n,s)\right)|$ vanish pointwise by the continuity of $f$ and the continuous dependence of the solution $\{\Xi(z,s):s\geq 0\}$ on the initial condition; see, e.g., \cite[Theorem~1.3]{Chi}.
This establishes that $K_\lambda f$ is continuous.
Next we prove that $K_\lambda f$ vanishes at infinity. Let $\epsilon>0$ be arbitrary.
By Assumption~\ref{ass:feller}, the level sets $\mathcal{H}_L:=\{z: H(z)\leq L\}$ are compact and $\mathcal{Z} = \cup_{L>0}\{z: H(z)\leq L\}$. Therefore we can find $L=L(\epsilon)$ such that
$|f(z)|<\epsilon(\lambda + \lambda_{\textrm{ref}})$ for $z \notin \mathcal{H}_L$. For all such $z$, since $H(\Xi(z,s))=H(z)$ for all $s>0$, we have that
\begin{align*}
|K_\lambda f(z)|
&\leq \int_0^\infty \mathrm{e}^{-\lambda s -\lambda_{\textrm{ref}} s} \left|f \left( \Xi(z,s)\right)\right| \mathrm{d} s\\
&< \epsilon(\lambda + \lambda_{\textrm{ref}}) \int_0^\infty \mathrm{e}^{-\lambda s -\lambda_{\textrm{ref}} s} \mathrm{d} s = \epsilon.
\end{align*}
Thus we conclude that for all $\lambda >0$ we have $K_\lambda :C_0(\mathcal{Z}) \to C_0(\mathcal{Z})$.
Now we move on to $J_\lambda$. First notice that for any $f\in C_0(\mathcal{Z})$ we have $Q_\alpha f$ is also continuous. To see why let $z_n=(x_n,v_n)\to z=(x,v)$
and notice that as $d=1$
\begin{multline*}
|Q_\alpha f \left(z_n\right)-Q_\alpha f \left(z\right)|
\\ \leq \frac{1}{\sqrt{2\pi}} \int_{-\infty}^\infty
\left| f\left(x_n, \alpha v_n + \sqrt{1-\alpha^2} \xi\right)
-f\left(x, \alpha v + \sqrt{1-\alpha^2} \xi\right) \right|\mathrm{e}^{-\xi^2/2} \mathrm{d} \xi \to 0,
\end{multline*}
by the bounded convergence theorem, since $f$ is continuous and bounded, and therefore $Q_\alpha f$ is continuous.
Next, for any $\delta>0$ we can choose a compact set $K_\delta$ such that $|f(z)|<\delta$ for $z\notin K_\delta$. In particular, since $K_\delta$ is compact, for any $\delta>0$ we can also find $M_\delta>0$ such that
$$K_\delta \subset \{(x,v): |x|, |v|\leq M_\delta\}.$$
Fix $\epsilon \in (0, 1/2)$ and choose $z_\epsilon$ such that $\Phi(z_{\epsilon})=1-\epsilon$, where $\Phi$ is the cumulative distribution function of the standard normal distribution. Then
\begin{align*}
|Q_\alpha f \left(z\right)|
&\leq \epsilon \|f\|_\infty
+\frac{1}{\sqrt{2\pi}} \int_{\xi=-z_{\epsilon}}^{z_\epsilon}
\left| f\left(x, \alpha v + \sqrt{1-\alpha^2} \xi\right)
\right|\mathrm{e}^{-\xi^2/2}\mathrm{d} \xi.
\end{align*}
Then for all $z=(x,v)$ and $\xi$ such that $|x|>M_\epsilon$, $|v|>(M_\epsilon+z_\epsilon)/\alpha$ and $|\xi|<z_\epsilon$ we have
\begin{align*}
\Big|\alpha v + \sqrt{1-\alpha^2}\xi\Big|
&\geq \alpha |v| - \sqrt{1-\alpha^2}|\xi|
\geq \alpha |v| -|\xi|\geq M_\epsilon - z_\epsilon > M_\epsilon.
\end{align*}
Therefore for such $z$ we have that
\begin{align*}
|Q_\alpha f \left(z\right)|
&\leq \epsilon \|f\|_\infty
+\frac{1}{\sqrt{2\pi}} \int_{\xi=-z_{\epsilon}}^{z_\epsilon}
\left| f\left(x, \alpha v + \sqrt{1-\alpha^2} \xi\right)
\right|\mathrm{e}^{-\xi^2/2}\mathrm{d} \xi\\
&< \epsilon \|f\|_\infty
+\frac{\epsilon}{\sqrt{2\pi}} \int_{\xi=-z_{\epsilon}}^{z_\epsilon}\mathrm{e}^{-\xi^2/2}\mathrm{d} \xi,
\end{align*}
and since $\epsilon>0$ is arbitrary it follows that $Q_\alpha f \in C_0$.
Observe that $J_\lambda f(z) =\lambda_{\textrm{ref}} K_\lambda Q_\alpha f(z)$. Therefore
if $f\in C_0$, since we have already shown that $Q_\alpha :C_0\to C_0$ and $K_\lambda :C_0\to C_0$, it follows that $J_\lambda f \in C_0$.
Finally, since clearly $\|Qf\left( \Xi(z,s)\right)\|_\infty \leq \|f\|_\infty$
\begin{align*}
\|J_\lambda f\|_\infty
&= \sup_{z} \Big|
\int_0^\infty \lambda_{\textrm{ref}} \, \mathrm{e}^{-\lambda s -\lambda_{\textrm{ref}} s} Qf \left(\Xi(z,s)\right) \mathrm{d} s
\Big|\\
&\leq \int_0^\infty \lambda_{\textrm{ref}} \, \mathrm{e}^{-\lambda s -\lambda_{\textrm{ref}} s} \|Q f\left(\Xi(z,s)\right)\|_\infty \mathrm{d} s\\
&\leq \int_0^\infty \lambda_{\textrm{ref}} \, \mathrm{e}^{-\lambda s -\lambda_{\textrm{ref}} s} \|f\|_\infty \mathrm{d} s
=\frac{\lambda_{\textrm{ref}}}{\lambda + \lambda_{\textrm{ref}}} \|f\|_\infty,
\end{align*}
and since $\lambda >0$ we can see that this is a strict contraction.
From this, it follows that the sequence
$$\sum_{j=0}^n J_\lambda^j K_\lambda f, $$
is Cauchy in the Banach space $\left(C_0(\mathcal{Z}), \| \cdot \|_\infty\right)$, whence the conclusion follows.
\end{proof}
\paragraph{$C_c^\infty$ is a core.}
Define the semigroup $\{Q_t:t \geq 0\}$, where for each $t\geq 0$
$Q^t: C_0(\mathcal{Z})\mapsto C_0(\mathcal{Z})$ is defined through
$Q^t f(z)=f\left(\Xi (z,t) \right)$, with $\xi(z,t)$ denoting as before the solution of the Hamiltonian dynamics started from $z$ at time $t$.
It can be easily shown that the generator of $Q^t$ is given for $f\in C_c^\infty(\mathcal{Z})$ by
$$Bf(x,v) = \langle \nabla_x f, v \rangle - \langle \nabla_v f , \nabla U(x)\rangle,$$
that is the first two terms of the generator $\mathcal{A}$ of RHMC.
Let $f$ be supported on a compact set $K$. By our assumptions on the Hamiltonian $H$, there exists $L>0$ such that $K\subseteq \mathcal{H}_L:=\{(x,v): H(x,v) \leq L\}$. Letting $z \notin \mathcal{H}_L$,
for all $t\geq 0$, we have by definition $H(\Xi(z,t))=H(z)$ and thus $\Xi(z,t) \notin \mathcal{H}_L$. Therefore $Q^tf$ will have compact support.
Notice next, that by Assumption~\ref{ass:weakconv}, for any $t\geq 0$ the mapping $z\mapsto \Xi(z,t)$ is infinitely differentiable, see e.g. \cite[Exercise~1.185]{Chi}. From this and the above discussion we conclude that for any $f\in C_c^\infty (\mathcal{Z})$ and $t\geq 0$ we have $Q^t f \in C_c^\infty$.
Therefore from \citet[Theorem~1.9]{davies1980one}, and since $C_c^\infty (\mathcal{Z})\subset C_0 (\mathcal{Z})$ is dense, we conclude that $C_c^\infty$ is a core for $B$, and in particular that for any $f\in \mathcal{D}(B)$, there exists a sequence $\{f_n:n\geq 0\} \subset C_c^\infty (\mathcal{Z})$ such that
$$\| f_n -f \|_\infty + \|Bf_n - B_f \|_\infty \to 0.$$
Since the operator $\lambda_{\textrm{ref}} [Q_\alpha - I]$ is clearly bounded on $C_0(\mathcal{Z})$ for any $\alpha$, it follows that $\mathcal{D}(\mathcal{A})= \mathcal{D}(B)$, and that for the sequence $\{f_n\}$ above we also have
$$\| f_n -f \|_\infty + \|\mathcal{A}f_n - \mathcal{A}_f \|_\infty \to 0,$$
proving that $C_c^\infty(\mathcal{Z})$ is a core for $\mathcal{A}$.
\subsection{Proof of Theorem~\ref{thm:weakconv}}
Recall that we write $\{\bm{Z}_n(s):s\geq 0\}$ for BPS initialized from $\pi_n$, the generator of which we denote with $\mathcal{A}_n$, and write $\{Z^{(1)}_n(s):s\geq 0\}$ for its first component.
In addition let
$$\mathcal{F}^n_t:=\sigma\{\bm{Z}_n(s): s\leq t\}, \quad \text{and} \quad \mathcal{G}^n_t:=\sigma\left\{Z^{(1)}_n(s): s\leq t\right\}.$$
Let $\epsilon_n\to 0$ be monotone and to be specified later on. All expectations will be with respect to the path measure of BPS started from $\pi_n$.
We proceed with the usual construction. For some function $f:\mathcal{Z} \to \mathcal{R}$, that is $f$ is a function only of $Z_n^{(1)}$, such that $f\in C^{\infty}_c$, smooth with compact support, we define
\begin{align}
\xi_n(t)&:= \epsilon_n^{-1} \int_0^{\epsilon_n}\mathbb{E}\left[ f\big( Z_n^{(1)}(t+s) \big) \big| \mathcal{G}_t^n \right] \mathrm{d} s,\label{eq:xidef}\\
\phi_n(t)&:=\epsilon_n^{-1} \mathbb{E}\left[\left. f\left( Z_n^{(1)}(t+\epsilon_n) \right) - f\left( Z_n^{(1)}(t) \right) \right| \mathcal{G}_t^n \right].\label{eq:phidef}
\end{align}
We have already established that $(\mathcal{A}, C_c^\infty)$ generates the strongly continuous semigroup $\left\{ P^{t}:t\geq0\right\}$ corresponding to RHMC.
To apply \citep[Corollary~8.6 of Chapter~4]{EK_05} we need to check
the following:
\begin{itemize}
\item \textbf{Strongly Separating algebra:} the closure of the linear span of $C_c^\infty$
contains an algebra that strongly separates points, see \cite[Section~3.4]{EK_05} for the definition. This is obvious since $C_c^\infty(\mathcal{Z})$ strongly separates points and its closure contains the algebra $C_0(\mathcal{Z})$.
\item \textbf{Generator~~convergence:} for each $f\in\mathcal{D}(\mathcal{A})$ and $T>0$, for $\xi_{n},\phi_{n}$ as defined in \eqref{eq:xidef},\eqref{eq:phidef}
\begin{align}
\sup_{n}\sup_{t\leq T}\mathbb{E}[|\xi^{(n)}(t)|] & <\infty\label{eq:firstcondition}\\
\sup_{n}\sup_{t\leq T}\mathbb{E}[|\phi^{(n)}(t)|] & <\infty\label{eq:secondcondition}\\
\lim_{n\to\infty}\mathbb{E}\left[\left|\xi^{(n)}(t)-f\left(Z_n^{(1)}(t)\right)\right|\right] & =0,\label{eq:thirdcondition}\\
\lim_{n\to\infty}\mathbb{E}\left[\left|\phi^{(n)}(t)-\mathcal{A}f\left(Z_n^{(1)}(t)\right)\right|\right] & =0,\label{eq:fourthcondition}
\end{align}
and in addition
\begin{equation}
\lim_{n\to\infty}\mathbb{E}\left\{ \sup_{t\in\mathbb{Q}\cap[0,T]}|\xi_{n}(t)-f(Z_n^{(1)}(t))|\right\} =0,\label{eq:833}
\end{equation}
and for some $p>1$
\begin{equation}
\sup_{n\to\infty}\mathbb{E}\left[\left(\int_{0}^{T}|\phi_{n}(s)|^{p}\mathrm{d}s\right)^{1/p}\right]<\infty.\label{eq:834}
\end{equation}
\end{itemize}
\subsubsection{Proof of Equations (\ref{eq:833}) and (\ref{eq:thirdcondition}).}
Since condition (\ref{eq:thirdcondition}) is implied by (\ref{eq:833}), we will establish \eqref{eq:833}.
Fix $T>0$. Then for each $n$, since BPS is non-explosive for every $n$ and $\delta>0$ we can find a $K_{n,\delta}>0$ such that
$$\mathbb{P}\left[ \sup_{t\leq T+1} \|\bm{Z}_n(t)\| \geq K_{n,\delta} \right] \leq \delta.$$
For $\delta_n\to 0$ and by a diagonal argument, we can find a sequence $K_{n,\delta_n}$ such that
$$\mathbb{P}\left[ \sup_{t\leq T+1}\|\bm{Z}_n(t)\| \geq K_{n,\delta_n} \right] \leq \delta_n \to 0.$$
We will write $G_n$ for the event
$$G_n := \left\{ \sup_{t\leq T+1}\|\bm{Z}_n(t)\| \leq K_{n,\delta_n} \right\}.$$
Then we have for $\epsilon_n \to 0$, to be specified later on,
\begin{align*}
\lefteqn{\mathbb{E}\left[ \sup_{t\in[0,T] \cap \mathbb{Q}} \left| \xi_n(t) - f\left( Z_n^{(1)}(t) \right)\right| \right]}\\
&=\mathbb{E}\left[\sup_{t\in[0,T] \cap \mathbb{Q}} \left| \epsilon_n^{-1}\int_0^{\epsilon_n}\mathbb{E}\left[\left. f\left( Z^{(1)}_n(t+r)\right)- f\left( Z^{(1)}_n(t)\right)\right| \mathcal{G}_t^n \right]\mathrm{d} r \right| \right]\\
&=\mathbb{E}\left[\sup_{t\in[0,T] \cap \mathbb{Q}} \left| \epsilon_n^{-1}\int_0^{\epsilon_n}\mathbb{E}\left[\left. \mathbb{E}\left\{\left. f\left( Z^{(1)}_n(t+r)\right)- f\left( Z^{(1)}_n(t)\right)
\right| \mathcal{F}_t^n \right\} \right| \mathcal{G}_t^n \right]\mathrm{d} r \right| \right]\\
&=\mathbb{E}\left[\sup_{t\in[0,T] \cap \mathbb{Q}} \left| \epsilon_n^{-1}\int_0^{\epsilon_n}\mathbb{E}\left[\left. \mathbb{E}\left\{\left. \left( f\left( Z^{(1)}_n(t+r)\right)- f\left( Z^{(1)}_n(t)\right)\right) \mathbb{1}_{G_n}
\right| \mathcal{F}_t^n \right\} \right| \mathcal{G}_t^n \right]\mathrm{d} r \right| \right]\\
&\qquad +\mathbb{E}\left[\sup_{t\in[0,T] \cap \mathbb{Q}} \left| \epsilon_n^{-1}\int_0^{\epsilon_n}\mathbb{E}\left[\left. \mathbb{E}\left\{\left. \left( f\left( Z^{(1)}_n(t+r)\right)- f\left( Z^{(1)}_n(t)\right)\right) \mathbb{1}_{G_n^\mathtt{c}}
\right| \mathcal{F}_t^n \right\} \right| \mathcal{G}_t^n \right]\mathrm{d} r \right| \right]\\
&:= J_1 + J_2.
\end{align*}
For the term $J_2$ we have for $p>1$
\begin{align}
J_2
&\leq 2\|f\|_\infty \mathbb{E}\left[\sup_{t\in[0,T] \cap \mathbb{Q}} \mathbb{E}\left[\left. \mathbb{1}_{G_n^\mathtt{c}}
\right| \mathcal{G}_t^n \right]\right]
\leq 2\|f\|_\infty\mathbb{E}\left[ \left(\sup_{t\in[0,T] \cap \mathbb{Q}} \mathbb{E}\left[\left. \mathbb{1}_{G_n^\mathtt{c}}
\right| \mathcal{G}_t^n \right] \right)^p\right]^{1/p}\notag\\
&\leq 2\|f\|_\infty\frac{p}{p-1} \mathbb{E}\left[ \mathbb{E}\left[\left. \mathbb{1}_{G_n^\mathtt{c}} \right| \mathcal{G}_T^n \right]^p\right]^{1/p}
\leq 2\|f\|_\infty\frac{p}{p-1} \mathbb{E}\left[ \mathbb{1}_{G_n^\mathtt{c}}^p\right]^{1/p} = 2\|f\|_\infty\frac{p}{p-1} \delta_n^{1/p},\label{eq:j2_doob}
\end{align}
where we used Jensen's inequality, the fact that $\mathbb{E}[\mathbb{1}_{G_n^\mathtt{c}}\mid \mathcal{G}_t^n]$ is a martingale and Doob's martingale inequality.
We proceed with the term $J_1$ as follows
\begin{align*}
J_1
&= \mathbb{E}\left[\sup_{t\in[0,T] \cap \mathbb{Q}} \left| \epsilon_n^{-1}\int_0^{\epsilon_n}\mathbb{E}\left[\left. \mathbb{E}\left\{\left. \left[ f\left( Z^{(1)}_n(t+r)\right)- f\left( Z^{(1)}_n(t)\right)\right] \mathbb{1}_{G_n}
\mathbb{1}\{\tau_1^\text{ref}(t)> \epsilon_n\} \right| \mathcal{F}_t^n \right\} \right| \mathcal{G}_t^n \right]\mathrm{d} r \right| \right]\\
&\qquad + \mathbb{E}\left[\sup_{t\in[0,T] \cap \mathbb{Q}} \left| \epsilon_n^{-1}\int_0^{\epsilon_n}\mathbb{E}\left[\left. \mathbb{E}\left\{\left. \left[ f\left( Z^{(1)}_n(t+r)\right)- f\left( Z^{(1)}_n(t)\right)\right] \mathbb{1}_{G_n}
\mathbb{1}\{\tau_1^\text{ref}(t)\leq \epsilon_n\} \right| \mathcal{F}_t^n \right\} \right| \mathcal{G}_t^n \right]\mathrm{d} r \right| \right]\\
&=: J_{1,1} + J_{1,2},
\end{align*}
where we denote by $\tau_1^\text{ref}(t)$ the first refreshment time after time $t$. Since refreshment happens independently we can bound $J_{1,2}$
\begin{align*}
J_{1,2}
&\leq 2\|f\|_\infty
\mathbb{E}\left[\sup_{t\in[0,T] \cap \mathbb{Q}} \left| \epsilon_n^{-1}\int_0^{\epsilon_n} (1-\mathrm{e}^{-\lambda_{\textrm{ref}} \epsilon_n})\mathrm{d} r \right| \right]\leq 2\|f\|_\infty \lambda_{\textrm{ref}}\, \epsilon_n \to 0.
\end{align*}
We control the term $J_{1,1}$ in two steps. To keep notation short we introduce the notation $G_n'(t):= \{ \tau_1^\text{ref}(t)> \epsilon_n\}$. Then
\begin{align*}
J_{1,1}
&\leq
\mathbb{E}\Bigg[\sup_{t\in[0,T] \cap \mathbb{Q}} \bigg| \epsilon_n^{-1}\int_0^{\epsilon_n}\mathbb{E}\Big[ \mathbb{E}\Big\{ \Big[ f\left( X^{(1)}_n(t+r), V^{(1)}_n(t+r)\right) \\
&\qquad\qquad - f\left( X^{(1)}_n(t), V^{(1)}_n(t+r)\right)\Big] \mathbb{1}_{G_n} \mathbb{1}_{G'_n(t)} \Big| \mathcal{F}_t^n \Big\} \Big| \mathcal{G}_t^n \Big]\mathrm{d} r \bigg| \Bigg]\\
&\qquad + \mathbb{E}\Bigg[\sup_{t\in[0,T] \cap \mathbb{Q}} \bigg| \epsilon_n^{-1}\int_0^{\epsilon_n}\mathbb{E}\Big[ \mathbb{E}\Big\{ \Big[ f\left( X^{(1)}_n(t), V^{(1)}_n(t+r)\right) \\
&\qquad\qquad\qquad - f\left( X^{(1)}_n(t), V^{(1)}_n(t)\right)\Big] \mathbb{1}_{G_n} \mathbb{1}_{G'_n(t)} \Big| \mathcal{F}_t^n \Big\} \Big| \mathcal{G}_t^n \Big]\mathrm{d} r \bigg| \Bigg]\\
&=: J_{1,1,1}+ J_{1,1,2}.
\end{align*}
For the first term, since only the location component changes we have
\begin{align*}
J_{1,1,1}
&\leq \|\partial_x f\|_\infty
\mathbb{E}\Bigg[\sup_{t\in[0,T] \cap \mathbb{Q}} \bigg| \epsilon_n^{-1}\int_0^{\epsilon_n}
\mathbb{E}\Big[ \mathbb{E}\Big\{ \Big| X^{(1)}_n(t+r) - X^{(1)}_n(t)\Big|\times \mathbb{1}_{G_n} \mathbb{1}_{G'_n(t)} \Big| \mathcal{F}_t^n \Big\} \Big| \mathcal{G}_t^n \Big]\mathrm{d} r \bigg| \Bigg]\\
&\leq \|\partial_x f\|_\infty
\mathbb{E}\Bigg[\sup_{t\in[0,T] \cap \mathbb{Q}} \bigg| \epsilon_n^{-1}\int_0^{\epsilon_n}
\mathbb{E}\Big[ \epsilon_n \Big| V^{(1)}_n(t)\Big|\times \mathbb{1}_{G_n} \mathbb{1}_{G'_n(t)} \Big| \mathcal{F}_t^n \Big\} \Big| \mathcal{G}_t^n \Big]\mathrm{d} r \bigg| \Bigg],
\end{align*}
where the second inequality follows from the linear dynamics of BPS, since on the event $G_n'(t)$ there is no refreshment event and therefore the norm of the velocity component does not change.
Finally, recalling the definition of the event $G_n$ we obtain
\begin{align*}
J_{1,1,1}
&\leq \|\partial_x f\|_\infty \epsilon_n K_{n,\delta_n},
\end{align*}
Next we have to control the term $J_{1,1,2}$ for which we point out that, since there is no refreshment event, the velocity will remain constant on the interval $[t, t+\epsilon_n]$ unless there is a bounce.
Writing $\sigma_1(t)$ for the arrival time of the first bounce after time $t$ we thus have
\begin{align*}
&J_{1,1,2}\\
&= \mathbb{E}\Bigg[\sup_{t\in[0,T] \cap \mathbb{Q}} \bigg| \epsilon_n^{-1}\int_0^{\epsilon_n}\mathbb{E}\Big[ \mathbb{E}\Big\{ \Big[ f\left( X^{(1)}_n(t), V^{(1)}_n(t+r)\right) \\
&\qquad\qquad\qquad - f\left( X^{(1)}_n(t), V^{(1)}_n(t)\right)\Big] \mathbb{1}\{\sigma_1(t)< \epsilon_n\}\mathbb{1}_{G_n} \mathbb{1}_{G'_n(t)} \Big| \mathcal{F}_t^n \Big\} \Big| \mathcal{G}_t^n \Big]\mathrm{d} r \bigg| \Bigg]\\
&\leq 2\|f\|_\infty \mathbb{E}\Bigg[\sup_{t\in[0,T] \cap \mathbb{Q}} \bigg| \epsilon_n^{-1}\int_0^{\epsilon_n}\mathbb{E}\Big[ \mathbb{E}\Big\{ \mathbb{1}\{\sigma_1(t)< \epsilon_n\}\mathbb{1}_{G_n} \mathbb{1}_{G'_n(t)} \Big| \mathcal{F}_t^n \Big\} \Big| \mathcal{G}_t^n \Big]\mathrm{d} r \bigg| \Bigg]\\
&\leq 2\|f\|_\infty \mathbb{E}\Bigg[\sup_{t\in[0,T] \cap \mathbb{Q}} \bigg| \epsilon_n^{-1}\int_0^{\epsilon_n}\mathbb{E}\Big[ \mathbb{E}\Big\{ \mathbb{1}\{\sigma_1(t)< \epsilon_n\} \Big| \mathcal{F}_t^n \Big\} \Big| \mathcal{G}_t^n \Big]\mathrm{d} r \bigg| \Bigg]\\
&\leq 2\|f\|_\infty \mathbb{E}\Bigg[\sup_{t\in[0,T] \cap \mathbb{Q}} \bigg| \epsilon_n^{-1}\int_0^{\epsilon_n}
\mathbb{E}\Big[\mathbb{E}\Big\{ \Big[ 1-\exp\Big(-\int_0^{\epsilon_n} \langle \nabla U_n(\bm{X}_n(t+s)), \bm{V}_n(t+s) \rangle_+
\mathrm{d} s \Big) \Big] \Big| \mathcal{F}_t^n \Big\} \Big| \mathcal{G}_t^n \Big]\mathrm{d} r \bigg| \Bigg],
\end{align*}
where we dropped the indicators in order to be able to compute the probability of no bounce. We again decompose according to the event $G_n$ in order to proceed
\begin{align*}
&J_{1,1,2}\\
&\leq 2\|f\|_\infty \mathbb{E}\Bigg[\sup_{t\in[0,T] \cap \mathbb{Q}} \bigg| \epsilon_n^{-1}\int_0^{\epsilon_n}
\mathbb{E}\Big[\mathbb{E}\Big\{ \Big[ 1-\exp\Big(-\int_0^{\epsilon_n} \langle \nabla U_n(\bm{X}_n(t+s)), \bm{V}_n(t+s) \rangle_+
\mathrm{d} s \Big) \Big]\mathbb{1}_{G_n} \Big| \mathcal{F}_t^n \Big\} \Big| \mathcal{G}_t^n \Big]\mathrm{d} r \bigg| \Bigg]\\
&\,+ 2\|f\|_\infty \mathbb{E}\Bigg[\sup_{t\in[0,T] \cap \mathbb{Q}} \bigg| \epsilon_n^{-1}\int_0^{\epsilon_n}
\mathbb{E}\Big[\mathbb{E}\Big\{ \Big[ 1-\exp\Big(-\int_0^{\epsilon_n} \langle \nabla U_n(\bm{X}_n(t+s)), \bm{V}_n(t+s) \rangle_+
\mathrm{d} s \Big) \Big]\mathbb{1}_{G_n^\mathtt{c}} \Big| \mathcal{F}_t^n \Big\} \Big| \mathcal{G}_t^n \Big]\mathrm{d} r \bigg| \Bigg].
\end{align*}
Since the integrand is bounded above by 1, a calculation similar to the one for the term $J_2$ in \eqref{eq:j2_doob} shows that the second term above vanishes as $n\to \infty$, and therefore using the inequality
$1-\exp(-x) \leq x$ for $x>0$ we have for $p>1$
\begin{align*}
&J_{1,1,2}\\
&\leq C \delta_n^{1/p}+
2\|f\|_\infty \mathbb{E}\Bigg[\sup_{t\in[0,T] \cap \mathbb{Q}} \bigg| \epsilon_n^{-1}\int_0^{\epsilon_n}
\mathbb{E}\Big[ \mathbb{E}\Big\{ \int_0^{\epsilon_n} \|\nabla U_n(\bm{X}_n(t+s))\| \|\bm{V}_n(t+s)\| \mathrm{d} s \mathbb{1}_{G_n}\Big| \mathcal{F}_t^n \Big\} \Big| \mathcal{G}_t^n \Big]\mathrm{d} r \bigg| \Bigg]\\
&\leq C \delta_n^{1/p}+\\
&\quad +2\|f\|_\infty \mathbb{E}\left[
\sup_{t\in[0,T] \cap \mathbb{Q}}
\left| \epsilon_n^{-1}\int_0^{\epsilon_n}\mathbb{E}\left[ \left. \mathbb{E} \left\{ \left.\int_0^{\epsilon_n}\left( \frac{1}{2}\|\nabla U_n(\bm{X}_n(t+s))\|^2 + \frac{1}{2} \|\bm{V}_n(t+s)\|^2 \right)\mathrm{d} s \times
\mathbb{1}_{G_n}\right| \mathcal{F}_t^n \right\} \right| \mathcal{G}_t^n \right]\mathrm{d} r \right| \right]\\
&\leq C \delta_n^{1/p}+\\
&\quad 2 C \|f\|_\infty \mathbb{E}\left[
\sup_{t\in[0,T] \cap \mathbb{Q}}
\left| \epsilon_n^{-1}\int_0^{\epsilon_n}\mathbb{E}\left[ \left. \mathbb{E} \left\{ \left.\int_0^{\epsilon_n}
\left(n+ \|\bm{X}_n(t+s)\|^2 + \|\bm{V}_n(t+s)\|^2 \right)\mathrm{d} s \times
\mathbb{1}_{G_n}\right| \mathcal{F}_t^n \right\} \right| \mathcal{G}_t^n \right]\mathrm{d} r \right| \right]\\
&\leq C \delta_n^{1/p}+\\
&\quad 2C\|f\|_\infty \mathbb{E}\Bigg[\sup_{t\in[0,T] \cap \mathbb{Q}} \bigg| \epsilon_n^{-1}\int_0^{\epsilon_n}
\mathbb{E}\Big[ C\epsilon_n \left( n+ \|\bm{Z}_n(t+s)\|^2\right) \mathbb{1}_{G_n}\Big| \mathcal{F}_t^n \Big\} \Big| \mathcal{G}_t^n \Big]\mathrm{d} r \bigg| \Bigg]
\leq 2C\|f\|_\infty \epsilon_n (n+K_{n,\delta_n}^2),
\end{align*}
where in the penultimate inequality we used the fact that $|U'(x)| \leq C (1+|x|)$, for some generic constant $C>0$ by Assumption \ref{ass:weakconv}.
We choose $\epsilon_n$ such that $\epsilon_n (n^2+K^2_{n,\delta_n}) \to 0$.
\subsubsection{Proof of \eqref{eq:fourthcondition}.}
Next we prove \eqref{eq:fourthcondition}. First, by stationarity notice that we can equivalently check
$$\mathbb{E}\left[ \left| \phi_n(0) - \mathcal{A}f\left( Z_n^{(1)}(0) \right)\right| \right] \to 0.$$
Notice first that $f\in \mathrm{Dom}\big(\widetilde{\mathcal{A}}_n\big)$, the domain of the extended generator, since $f$ is smooth and bounded (see \cite[Theorem~26.14]{D_93})
\begin{align*}
\phi_n(0)
&=\epsilon_n^{-1} \mathbb{E}\left[\left. f\left( Z_n^{(1)}(\epsilon_n) \right) - f\left( Z_n^{(1)}(0) \right) \right| \mathcal{G}_0^n \right]\\
&=\epsilon_n^{-1} \mathbb{E}\left[\left. \int_0^{\epsilon_n} \widetilde{\mathcal{A}}_n f \left(\bm{Z}_n(s) \right)\mathrm{d} s + \mathcal{R}_n(s) \right| \mathcal{G}_0^n \right]\\
&=\epsilon_n^{-1} \mathbb{E}\left[\left. \int_0^{\epsilon_n} \widetilde{\mathcal{A}}_n f \left(\bm{Z}_n(s) \right)\mathrm{d} s \right| \mathcal{G}_0^n \right],
\end{align*}
where we used the facts that $\mathcal{R}_n(t)$ is an $\mathcal{F}_t^n$-martingale and $\mathcal{F}_t^n \subseteq \mathcal{G}_t^n$, whence
$$\mathbb{E}\left[\left. \mathcal{R}_n(s)\right| \mathcal{G}_0^n \right]
=\mathbb{E}\left\{\left. \mathbb{E}\left[\left. \mathcal{R}_n(s)\right| \mathcal{F}_0^n \right]\right|\mathcal{G}_0^n \right\}=0.
$$
We also notice that $g_n:=\widetilde{\mathcal{A}}_n f\in \mathrm{Dom}\big(\widetilde{\mathcal{A}}_n\big)$ the domain of the extended generator.
Therefore
\begin{align*}
\phi_n(0)
&=\epsilon_n^{-1} \int_0^{\epsilon_n} \mathbb{E}\left[\left. g_n \left(\bm{Z}_n(s) \right) \right| \mathcal{G}_0^n \right]\mathrm{d} s\\
&=\epsilon_n^{-1} \int_0^{\epsilon_n} \mathbb{E}\left[\left. \widetilde{\mathcal{A}}_n f \left(\bm{Z}_n(0) \right) + \int_0^s \widetilde{\mathcal{A}}_n g_n \left(\bm{Z}_n(r) \right) + \mathcal{R}_n'(s) \mathrm{d} r \right| \mathcal{G}_0^n \right]\mathrm{d} s\\
&=\epsilon_n^{-1} \int_0^{\epsilon_n} \mathbb{E}\left[\left. \widetilde{\mathcal{A}}_n f \left(\bm{Z}_n(0) \right) + \int_0^s \widetilde{\mathcal{A}}_n g_n \left(\bm{Z}_n(r) \right)\mathrm{d} r \right| \mathcal{G}_0^n \right]\mathrm{d} s,
\end{align*}
where, from \cite[Theorem~26.12]{D_93}, it follows that the local martingale $\{\mathcal{R}_n'(s):s\geq 0\}$ is actually a proper martingale,
and therefore using the same arguments as before, for $s>0$,
$$\mathbb{E}\left[\left. \mathcal{R}_n'(s) \right| \mathcal{G}_0^n \right]=0.$$
Then we have
\begin{align}
\mathbb{E}\left[ \left| \phi_n(t) - \mathcal{A}f\left( Z_n^{(1)}(t) \right)\right| \right]
&\leq \mathbb{E}\left[ \left|\mathbb{E}\left[\left. \widetilde{\mathcal{A}}_n f \left(\bm{Z}_n(0) \right)\right| \mathcal{G}_0^n\right] - \mathcal{A}f\left( Z_n^{(1)}(0) \right)\right| \right] \notag\\
&\qquad + \mathbb{E}\left\{ \left| \epsilon_n^{-1} \int_0^{\epsilon_n}\mathbb{E}\left[ \left. \int_0^s \widetilde{\mathcal{A}}_n g_n \left(\bm{Z}_n(r) \right)\mathrm{d} r \right| \mathcal{G}_0^n \right]\mathrm{d} s\right|\right\}
\notag\\
&\leq \mathbb{E}\left[ \left|\mathbb{E}\left[\left. \widetilde{\mathcal{A}}_n f \left(\bm{Z}_n(0) \right)\right| \mathcal{G}_0^n\right] - \mathcal{A}f\left( Z_n^{(1)}(0) \right)\right| \right]
\notag\\
&\qquad + \epsilon_n^{-1} \int_0^{\epsilon_n}\int_0^s\mathbb{E}\left\{ \mathbb{E}\left[ \left. \big| \widetilde{\mathcal{A}}_n g_n \left(\bm{Z}_n(r) \right)\big| \right| \mathcal{G}_0^n \right]\right\}\mathrm{d} r \mathrm{d} s
\notag\\
&:=\mathbb{E}\left[ \left|\mathbb{E}\left[\left. \widetilde{\mathcal{A}}_n f \left(\bm{Z}_n(0) \right)\right| \mathcal{G}_0^n\right] - \mathcal{A}f\left( Z_n^{(1)}(0) \right)\right| \right]
+ \mathcal{R}_n,\label{eq:Rndef}
\end{align}
by conditional Jensen's inequality. Finally by the tower law and by stationarity of $\{\bm{Z}_n(t):t\geq 0\}$ when initialized from $\pi_n$
\begin{align*}
\mathcal{R}_n
&= \epsilon_n^{-1} \int_0^{\epsilon_n}\int_0^s\mathbb{E}\left\{ \mathbb{E}\left[ \left. \big| \widetilde{\mathcal{A}}_n g_n \left(\bm{Z}_n(r) \right)\big| \right| \mathcal{G}_0^n \right]\right\}\mathrm{d} r \mathrm{d} s\\
&=\epsilon_n^{-1} \int_0^{\epsilon_n}\int_0^s\mathbb{E}\left\{ \big| \widetilde{\mathcal{A}}_n g_n \left(\bm{Z}_n(r) \right)\big| \right\}\mathrm{d} r \mathrm{d} s =\frac{\epsilon_n}{2}\mathbb{E}\left\{ \big| \widetilde{\mathcal{A}}_n g_n \left(\bm{Z}_n(0) \right)\big| \right\}.
\end{align*}
\paragraph{Error term.}
We will now control this error term. Recall first that for $f \in C^\infty_c(\mathcal{Z})\subset \mathcal{D}(\mathcal{A}_n)$ we have
\begin{align*}
\mathcal{A}_n f(\bm{x},\bm{v})
&= \langle \nabla f(\bm{x}), \bm{v})\rangle + \max\{0, \langle \nabla U_n(\bm{x}), \bm{v}\rangle\}
\left[\mathfrak{R}f\left(\bm{x},\bm{v}\right) - f\left(\bm{x},\bm{v}\right) \right]
+ \lambda_{\textrm{ref}} \left[ Qf \left(\bm{x},\bm{v}\right)-f\left(\bm{x},\bm{v}\right)\right],\\
\mathfrak{R}f\left(\bm{x},\bm{v}\right)
&:= f\left(\bm{x},\bm{v}- 2 \frac{\langle \nabla U_n(\bm{x}), \bm{v})\rangle}{\| \nabla U_n(\bm{x})\|^2} \nabla U_n(\bm{x})\right),\\
Qf\left(\bm{x},\bm{v}\right)
&:= \frac{1}{(2\pi)^{n/2}}\int_{\mathbb{R}^d} \mathrm{e}^{-|\boldsymbol{\xi}|^2/2} f\left(\bm{x},\alpha\bm{v}+\sqrt{1-\alpha^2} \boldsymbol{\xi}\right) \mathrm{d} \boldsymbol{\xi}.
\end{align*}
Potentially abusing notation, for $n\geq 1$ and $\bm{x}\in \mathbb{R}^n$ we define a mapping $\mathfrak{R}_x^n: \mathbb{R}^n \mapsto \mathbb{R}^n$ through
$$\mathfrak{R}_x^n \bm{v}:= \bm{v}- 2 \frac{\langle \nabla U_n(\bm{x}), \bm{v})\rangle}{\| \nabla U_n(\bm{x})\|^2} \nabla U_n(\bm{x}).$$
We decompose the generator $\mathcal{A}_n$ into three parts
$$\mathcal{A}_n = \mathcal{A}_n^{(1)}+\mathcal{A}_n^{(2)}+\mathcal{A}_n^{(3)},$$
where
\begin{align*}
\mathcal{A}_n^{(1)} f (\bm{x},\bm{v})
&=\left.\frac{\mathrm{d}}{\mathrm{d} t}f\left(\bm{x}+t \bm{v}, \bm{v}\right) \right|_{t=0}=\langle \nabla f(\bm{x}), \bm{v}\rangle,\\
\mathcal{A}_n^{(2)} f (\bm{x},\bm{v})
&=\max\{0, \langle \nabla U_n(\bm{x}), \bm{v}\rangle\}
\left[\mathfrak{R}f\left(\bm{x},\bm{v}\right) - f\left(\bm{x},\bm{v}\right) \right],\\
\mathcal{A}_n^{(3)} f (\bm{x},\bm{v})
&=\lambda_{\textrm{ref}} \left[ Qf \left(\bm{x},\bm{v}\right)-f\left(\bm{x},\bm{v}\right)\right].
\end{align*}
Therefore when considering $\mathcal{A}_n g_n=\mathcal{A}_n\mathcal{A}_n f_n$ we will need to consider all possible combinations $\mathcal{A}_n^{(i)} \mathcal{A}_n^{(j)}$ since the operators do not necessarily commute.
\paragraph{Case $i=1$.}
Using the fact that $f(\bm{x},\bm{v})=f(x,v)$, where we write $(x,v)$ for the first location and velocity components of $(\bm{x},\bm{v})$, the first term reduces to
\begin{align*}
\mathcal{A}_n^{(1)} \mathcal{A}_n^{(1)}f (\bm{x},\bm{v})
&= \left.\frac{\mathrm{d} }{\mathrm{d} t} \langle \nabla f(\bm{x}), \bm{v}\rangle\right|_{t=0}= \left.\frac{\mathrm{d} }{\mathrm{d} t} \frac{\partial}{\partial x}f(x+tv, v)v\right|_{t=0}\\
&=\frac{\partial^2 f}{\partial x^2}(x,v) v^2.
\end{align*}
Since $f\in C^\infty_c (\mathbb{R}\times \mathbb{R})$, it follows that $\partial^2_x f(x,v)$ is also continuous and compactly supported and therefore bounded. Thus
\begin{align*}
\mathbb{E}\left|\frac{\partial^2 f}{\partial x^2}(X^{(1)},V^{(1)}) \left(V^{(1)}\right)^2 \right|
&\leq \left\|\frac{\partial^2 f}{\partial x^2}\right\|_\infty\mathbb{E}\left[\left(V^{(1)}\right)^2 \right]\leq \left\|\frac{\partial^2 f}{\partial x^2} \right\|_\infty,
\end{align*}
since under $\pi_n$, $V^{(1)}$ is centered Gaussian with unit variance.
The second term takes the form
\begin{align*}
\mathcal{A}_n^{(1)} \mathcal{A}_n^{(2)}f (\bm{x},\bm{v})
&= \left.\frac{\mathrm{d} }{\mathrm{d} t} \mathcal{A}_n^{(2)}f (\bm{x}+t\bm{v},\bm{v})\right|_{t=0}\\
&= \left.\frac{\mathrm{d} }{\mathrm{d} t} \max\{0, \langle \nabla U_n(\bm{x}+t\bm{v}), \bm{v})\rangle\} \right|_{t=0}
\left[\mathfrak{R}f\left(\bm{x},\bm{v}\right) - f\left(\bm{x},\bm{v}\right) \right]\\
&\qquad +\left.\frac{\mathrm{d} }{\mathrm{d} t} \left[\mathfrak{R}f\left(\bm{x}+t\bm{v},\bm{v}\right) - f\left(\bm{x}+t\bm{v},\bm{v}\right) \right] \right|_{t=0}
\max\{0, \langle \nabla U_n(\bm{x}), \bm{v})\rangle\}\\
&=: J_1 + J_2.
\end{align*}
For $J_1$, first notice that
\begin{align*}
\lefteqn{\left|\max\{0, \langle \nabla U_n(\bm{x}+h\bm{v}), \bm{v}\rangle\}-\max\{0, \langle \nabla U_n(\bm{x}), \bm{v}\rangle\}\right|}\\
&\leq \left|\langle \nabla U_n(\bm{x}+h\bm{v}), \bm{v}\rangle-\langle \nabla U_n(\bm{x}), \bm{v})\rangle\right|\\
&\leq \sum_{i=1}^n\left|U'\left(x_{i}+h v_i\right) -U'\left(x_i\right)\right| |v_i|\leq \sum_{i=1}^n \|U''\|_\infty h |v_i|^2.
\end{align*}
Therefore we have that, for $h\in (0,1)$
\begin{align*}
h^{-1} \left|\max\{0, \langle \nabla U_n(\bm{x}+h\bm{v}), \bm{v}\rangle\}-\max\{0, \langle \nabla U_n(\bm{x}), \bm{v}\rangle\}\right|
&\leq \sum_{i=1}^n \|U''\|_\infty |v_i|^2 \in L^1(\pi),
\end{align*}
since the $V_i$ are standard normal random variables.
In addition since $f$ is bounded it follows that
$\mathfrak{R}f\left(\bm{x},\bm{v}\right) \leq \|f\|_\infty$. Therefore by the dominated convergence theorem we have that
\begin{align*}
\pi \left[J_1\right]
&\leq 2\|f\|_\infty \mathbb{E} \left[ \left|\frac{\mathrm{d} }{\mathrm{d} t} \max\{0, \langle \nabla U_n(\bm{X}+t\bm{V}), \bm{V})\rangle\} \Big|_{t=0}\right| \right]\\
&\leq 2\|f\|_\infty \mathbb{E} \left[\lim_{h\to 0} h^{-1}\left| \langle \nabla U_n(\bm{X}+h\bm{V}), \bm{V})\rangle -\langle \nabla U_n(\bm{X}+t\bm{V}), \bm{V})\rangle\right| \right]\\
&\leq 2\|f\|_\infty \sum_{i=1}^n \mathbb{E} \left[ \left|U''\left(X_i \right) \right| |V_i|^{2} \right]
= 2\|f\|_\infty n \pi(|U''|) = O(n).
\end{align*}
For $J_2$ a lengthy but straightforward calculation shows that
\begin{align*}
\left.\frac{\mathrm{d} }{\mathrm{d} t} \mathfrak{R}f\left(\bm{x}+t\bm{v},\bm{v}\right)\right|_{t=0}
&=\left.\frac{\mathrm{d} }{\mathrm{d} t}
f\left(x_1+tv_1,v_1- 2 \frac{\langle \nabla U_n(\bm{x}+t\bm{v}), \bm{v})\rangle}{\| \nabla U_n(\bm{x}+t\bm{v})\|^2} U'(x_1+tv_1)\right)
\right|_{t=0}\\
&= (\mathcal{R}\partial_x f)(\bm{x},\bm{v})v_1 + (\mathcal{R}\partial_v f)(\bm{x},\bm{v})\times \mathfrak{U}(\bm{x},\bm{v}),
\end{align*}
where
\begin{multline*}
\mathfrak{U}(\bm{x},\bm{v})
=\frac{1}{\|\nabla U(\bm{x})\|^2}\left\{
U'(x_1) \sum_{k=1}^n U''(x_k) v_k^2 + U''(x_1) v_1^2 \langle \nabla U(\bm{x}), \bm{v}\rangle
\right\}\\
-
\frac{1}{\|\nabla U(\bm{x})\|^4}\left\{
2U'(x_1) \sum_{k=1}^n U'(x_k) U''(x_k) v_k \langle \nabla U(\bm{x}), \bm{v}\rangle
\right\},
\end{multline*}
and thus since $\|U''\|_\infty \leq C$ we have
\begin{align*}
\left|\mathfrak{U}(\bm{x},\bm{v}) \right|
&\leq \frac{1}{\|\nabla U(\bm{x})\|^2}\left\{
C\|\nabla U(\bm{x})\| \|\bm{v}\|^2+ C \|\bm{v}\| |v_1|^2 \|\nabla U(\bm{x})\|
\right\}\\
&\qquad +
\frac{1}{\|\nabla U(\bm{x})\|^4}\left\{
2 C \|\nabla U(\bm{x})\|^3 \|\bm{v}\|^2
\right\},
\end{align*}
whence
\begin{multline*}
\left|\mathfrak{U}(\bm{x},\bm{v}) \max\{0, \langle \nabla U_n(\bm{x}), \bm{v})\rangle\} \right|
\leq \frac{ \|\bm{v}\|}{\|\nabla U(\bm{x})\|}\left\{
C\|\nabla U(\bm{x})\| \|\bm{v}\|^2+ C \|\bm{v}\|^3 \|\nabla U(\bm{x})\|
\right\}\\
+
\frac{ \|\bm{v}\|}{\|\nabla U(\bm{x})\|^3}\left\{
2 C \|\nabla U(\bm{x})\|^3 \|\bm{v}\|^2
\right\}\leq C (\|\bm{v}\|^3+ \|\bm{v}\|^4).
\end{multline*}
Thus overall,
\begin{align*}
\left|\frac{\mathrm{d} }{\mathrm{d} t} \mathfrak{R}f\left(\bm{x}+t\bm{v},\bm{v}\right)\!\Big|_{t=0} \max\{0, \langle \nabla U_n(\bm{x}), \bm{v})\rangle\} \right|
&\leq \|\partial_x f\|_\infty \|\bm{v}\|^2 \|\nabla U(\bm{x})\|
+ C \|\partial_v f\|_\infty \left(\|v\|^4+ \|v\|^3\right),
\end{align*}
On the other hand
\begin{align*}
\left|\left.\frac{\mathrm{d} }{\mathrm{d} t} f\left(\bm{x}+t\bm{v},\bm{v}\right) \right|_{t=0}
\max\{0, \langle \nabla U_n(\bm{x}), \bm{v})\rangle\}\right|
&\leq \|\partial_x f\|_\infty \|\nabla U(\bm{x})\| \|\bm{v}\|^2.
\end{align*}
Thus overall we have that, using the fact that $(V_1, \dots, V_n)$ are i.i.d.\ standard Gaussian and Assumption~\ref{ass:gradintegral}
\begin{align*}
\pi[|J_2|]
&\leq C \mathbb{E}\left[\|\nabla U(\bm{X})\|\right]\mathbb{E}\left[\|\nabla \bm{V}^2\| \right]
+C \mathbb{E}\left[\| \bm{V}\|^4 \right]\\
&\leq Cn \mathbb{E}\left[\|\nabla U(\bm{X})|^2\right]^{1/2}
+C n^2 = O(n^2)
\end{align*}
and thus overall we have that $\pi \left[ \left| \mathcal{A}_n^{(1)} \mathcal{A}_n^{(2)}f \right|\right] = O(n^2)$.
For the final term, since $Qf(\bm{x},\bm{v}) = Qf(x_1, v_1)$ we have
\begin{align*}
\mathcal{A}_n^{(1)} \mathcal{A}_n^{(3)}f (\bm{x},\bm{v})
&= \left.\frac{\mathrm{d} }{\mathrm{d} t} \mathcal{A}_n^{(3)}f (\bm{x}+t\bm{v},\bm{v})\right|_{t=0}\\
&=\lambda_{\textrm{ref}} \left.\frac{\mathrm{d} }{\mathrm{d} t} \left[ Qf(x_1+tv_1,v_1)-f(x_1+t v_1,v_1)\right]\right|_{t=0}\\
&=\lambda_{\textrm{ref}} \left[ Q(\partial_x f)(x_1+tv_1,v_1)-\partial_x f(x_1,v_1)\right],
\end{align*}
by an application of dominated convergence. In addition, since $\partial_x f$, we can easily see from the above that
$\pi \big[ \mathcal{A}_n^{(1)} \mathcal{A}_n^{(3)}f \big] =O(1)$ as $n\to \infty$.
\paragraph{Case $i=2$.}
For the first term $\mathcal{A}_n^{(2)} \mathcal{A}_n^{(1)}f$, notice first that since $f(\bm{x},\bm{v}) = f(x_1, v_1)$ we have
$$\mathcal{A}_n^{(1)}f(\bm{x}, \bm{v}) = \partial_x f(x_1,v_1) v_1:=h(x_1, v_1).$$
Therefore
\begin{align*}
\mathfrak{R}_n h(\bm{x},\bm{v})
&= \partial_x f\left(x_1,v_1- 2 \frac{\langle \nabla U_n(\bm{x}), \bm{v})\rangle}{\| \nabla U_n(\bm{x})\|^2} U'(x_1)\right)
\left(v_1- 2 \frac{\langle \nabla U_n(\bm{x}), \bm{v})\rangle}{\| \nabla U_n(\bm{x})\|^2} U'(x_1)\right),
\end{align*}
whence
\begin{align*}
\mathfrak{R}_n h(\bm{x},\bm{v})- h(\bm{x},\bm{v})
&= v_1 \left[ \mathfrak{R}_n \partial_xf(\bm{x},\bm{v})- \partial_x f(\bm{x},\bm{v})\right] - 2 \mathfrak{R}_n \partial_x f(\bm{x},\bm{v}) \frac{\langle \nabla U_n(\bm{x}), \bm{v})\rangle}{\| \nabla U_n(\bm{x})\|^2} U'(x_1),
\end{align*}
and thus
\begin{align*}
\mathbb{E}\left|\mathcal{A}_n^{(2)} \mathcal{A}_n^{(1)}f (\bm{X},\bm{V})\right|
&\leq \mathbb{E}\left[ \left| \langle\nabla U_n(\bm{X}), \bm{V}\rangle\right|\times |V_1| \times \left|\mathfrak{R}_n \partial_xf(\bm{X},\bm{V})- \partial_x f(\bm{X},\bm{V}) \right| \right]\\
&\qquad + 2\mathbb{E}\left[ \left| \langle\nabla U_n(\bm{X}), \bm{V}\rangle\right| \times \left|\mathfrak{R}_n \partial_x f(\bm{X},\bm{V}) \frac{\langle \nabla U_n(\bm{X}), \bm{V})\rangle}{\| \nabla U_n(\bm{X})\|^2} U'(X_1) \right| \right]\\
&\leq\left( \|\mathfrak{R}_n \partial_x f\|_\infty + \|\partial_x f\|_\infty \right)
\sum_{i=1}^n \mathbb{E}\left[ |V_1| |U'(X_i)| |V_i| \right]\\
&\qquad + 2\|\mathfrak{R}_n \partial_x f\|_\infty \mathbb{E}\left[ \frac{\langle \nabla U_n(\bm{X}), \bm{V}\rangle ^2}{\| \nabla U_n(\bm{X})\|^2} \left|U'(X_1)\right| \right].
\end{align*}
Using the Cauchy-Schwartz inequality inside the second term we have
\begin{align*}
\mathbb{E}\left|\mathcal{A}_n^{(2)} \mathcal{A}_n^{(1)}f (\bm{X},\bm{V})\right|
&\leq\left( \|\mathfrak{R}_n \partial_x f\|_\infty + \|\partial_x f\|_\infty \right)
\sum_{i=1}^n \mathbb{E}\left[ |V_1| |U'(X_i)| |V_i| \right]\\
&\qquad + 2\|\mathfrak{R}_n \partial_x f\|_\infty \mathbb{E}\left[ \|\bm{V}\|^2 \left|U'(X_1)\right| \right] = O(n),
\end{align*}
as $n\to \infty$ by Assumption~\ref{ass:gradintegral}.
For the next term $\mathcal{A}_n^{(2)} \mathcal{A}_n^{(2)}f$ first we write
\begin{align*}
\mathcal{A}_n^{(2)} \mathcal{A}_n^{(2)} f (\bm{x},\bm{v})
&= \max\left\{ 0, \langle \nabla U_n(\bm{x}), \bm{v}\rangle \right\}
\left[\mathfrak{R}_n \mathcal{A}_n^{(2)} f (\bm{x},\bm{v}) - \mathcal{A}_n^{(2)} f (\bm{x},\bm{v}) \right].
\end{align*}
Then notice that
\begin{align*}
\mathfrak{R}_n \mathcal{A}_n^{(2)} f (\bm{x},\bm{v})
&= \max \left\{0, \left\langle \nabla U_n(\bm{x}), \bm{v}- 2 \frac{\langle \nabla U_n(\bm{x}), \bm{v})\rangle}{\| \nabla U_n(\bm{x})\|^2} \nabla U_n(\bm{x}) \right\rangle \right\}\\
&\qquad \times
\left[\mathfrak{R}_n f\left( x_1, v_1- 2 \frac{\langle \nabla U_n(\bm{x}), \bm{v})\rangle}{\| \nabla U_n(\bm{x})\|^2} U'(x_1)\right) - f\left( x_1, v_1- 2 \frac{\langle \nabla U_n(\bm{x}), \bm{v})\rangle}{\| \nabla U_n(\bm{x})\|^2} U'(x_1)\right) \right]
\\
&= \max \left\{0, \left\langle \nabla U_n(\bm{x}), -\bm{v}\right\rangle \right\}\\
&\qquad \times
\left[\mathfrak{R}_n f\left( x_1, v_1- 2 \frac{\langle \nabla U_n(\bm{x}), \bm{v})\rangle}{\| \nabla U_n(\bm{x})\|^2} U'(x_1)\right) - f\left( x_1, v_1- 2 \frac{\langle \nabla U_n(\bm{x}), \bm{v})\rangle}{\| \nabla U_n(\bm{x})\|^2} U'(x_1)\right) \right],
\end{align*}
and therefore that
\begin{align*}
\left|\mathfrak{R}_n \mathcal{A}_n^{(2)} f (\bm{x},\bm{v}) - \mathcal{A}_n^{(2)} f (\bm{x},\bm{v})\right|
&\leq 2 \|f\|_\infty \left| \left\langle \nabla U_n(\bm{x}), \bm{v}\right\rangle \right|\\
\left|\mathcal{A}_n^{(2)}\mathcal{A}_n^{(2)} f (\bm{x},\bm{v})\right|
&\leq 2 \|f\|_\infty \left\langle \nabla U_n(\bm{x}), \bm{v}\right\rangle ^2.
\end{align*}
Thus
\begin{align*}
\mathbb{E}\left|\mathcal{A}_n^{(2)} \mathcal{A}_n^{(2)} f (\bm{X},\bm{V})\right|
&\leq C \|f\|_\infty \mathbb{E}\left[ \left\langle \nabla U_n(\bm{X}), \bm{V}\right\rangle ^2 \right]\\
&\leq C \|f\|_\infty \mathbb{E}\left[ \left(\sum_{i=1}^n U'(X_i) V_i \right)^2 \right] = O(n),
\end{align*}
from Assumption \ref{ass:gradintegral} and using the fact that the terms are independent and mean zero.
Next we consider the term $\mathcal{A}_n^{(2)} \mathcal{A}_n^{(3)}f$. Since $f$ is bounded, it easily follows that $\mathcal{A}_n^{(3)}f$ is also bounded and therefore that
\begin{align*}
\left|\mathcal{A}_n^{(2)} \mathcal{A}_n^{(3)}f (\bm{x}, \bm{v})\right|
&= \max\left\{ 0, \langle \nabla U_n(\bm{x}), \bm{v}\rangle \right\}
\left|\mathfrak{R}_n \mathcal{A}_n^{(3)} f (\bm{x},\bm{v}) - \mathcal{A}_n^{(3)} f (\bm{x},\bm{v}) \right|\\
&\leq 2 \|f\|_\infty \max\left\{ 0, \langle \nabla U_n(\bm{x}), \bm{v}\rangle \right\}.
\end{align*}
Therefore
\begin{align*}
\mathbb{E} \left|\mathcal{A}_n^{(2)} \mathcal{A}_n^{(3)}f (\bm{X}, \bm{V}) \right|
&\leq C \mathbb{E}\left| \langle \nabla U_n(\bm{X}), \bm{V}\rangle \right|
\leq C \mathbb{E}\left[ \langle \nabla U_n(\bm{X}), \bm{V}\rangle^2 \right]^{1/2} \\
&=C \mathbb{E}\left[\left(\sum_{i=1}^n U'(X_i) V_i\right)^2 \right]^{1/2}
=O(n^{1/2}),
\end{align*}
from Assumption \ref{ass:gradintegral} and using the fact that the terms are independent and mean zero.
\paragraph{Case $i=3$.} The first term to consider is
\begin{align*}
\mathcal{A}_n^{(3)} \mathcal{A}_n^{(1)}f(\bm{x}, \bm{v})
&= \lambda_{\textrm{ref}} \left[Q \mathcal{A}_n^{(1)}f(\bm{x},\bm{v})-\mathcal{A}_n^{(1)}f(\bm{x},\bm{v}) \right] \\
&= \lambda_{\textrm{ref}} \int \left[\mathcal{A}_n^{(1)}f(x_1,\alpha v_1+\sqrt{1-\alpha^2}\xi)-\mathcal{A}_n^{(1)}f(\bm{x},\bm{v}) \right]\phi(\xi) \mathrm{d} \xi\\
&= \lambda_{\textrm{ref}} \int \left[\partial_x f(x_1,\alpha v_1+\sqrt{1-\alpha^2}\xi)\left(\alpha v_1+\sqrt{1-\alpha^2}\xi\right)-\partial_x f(x_1,v_1)v_1 \right]\phi(\xi) \mathrm{d} \xi,
\end{align*}
where $\phi$ denotes the standard normal density.
Since $\|\partial_x f\|_\infty <\infty$ we have
\begin{align*}
\mathbb{E}\left|\mathcal{A}_n^{(3)} \mathcal{A}_n^{(1)}f(\bm{X}, \bm{V})\right|
&\leq \lambda_{\textrm{ref}}\|\partial_x f\|_\infty \mathbb{E}\left[\left|\alpha V_1 + \sqrt{1-\alpha^2}\xi\right| + |V_1|\right] =O(1),
\end{align*}
as $n\to \infty$.
For the second term we have, using Jensen's inequality on the Markov kernel $Q$,
\begin{align*}
\mathbb{E}\left|\mathcal{A}_n^{(3)} \mathcal{A}_n^{(2)}f(\bm{X},\bm{V})\right|
&\leq \lambda_{\textrm{ref}} \mathbb{E}\left[\left| Q \mathcal{A}_n^{(2)}f(\bm{X},\bm{V}) \right|\right] + \lambda_{\textrm{ref}} \mathbb{E}\left[\left|\mathcal{A}_n^{(2)}f(\bm{X},\bm{V})\right| \right] \\
&\leq \lambda_{\textrm{ref}} \mathbb{E}\left[ Q \left(\big|\mathcal{A}_n^{(2)}f\big|\right)(\bm{X},\bm{V}) \right] + \lambda_{\textrm{ref}} \mathbb{E}\left[\left|\mathcal{A}_n^{(2)}f(\bm{X},\bm{V})\right| \right].
\end{align*}
At this point notice that $Q$ is $\pi_n$-invariant and therefore
$$\mathbb{E}\left[ Q \left(\big|\mathcal{A}_n^{(2)}f\big|\right)(\bm{X},\bm{V}) \right] = \mathbb{E}\left[ \big|\mathcal{A}_n^{(2)}f (\bm{X},\bm{V})\big| \right],$$
whence we conclude that
\begin{align*}
\mathbb{E}\left|\mathcal{A}_n^{(3)} \mathcal{A}_n^{(2)}f(\bm{X},\bm{V})\right|
&\leq 2\lambda_{\textrm{ref}} \mathbb{E}\left[\left|\mathcal{A}_n^{(2)}f(\bm{X},\bm{V})\right| \right]\\
&\leq 4 \lambda_{\textrm{ref}} \|f\|_\infty \mathbb{E}\left[\left|\langle \nabla U_n(\bm{X}), \bm{V}\rangle \right|\right]\\
&\leq 4 \lambda_{\textrm{ref}} \|f\|_\infty \mathbb{E}\left[\langle \nabla U_n(\bm{X}), \bm{V}\rangle ^2\right]^{1/2}\\
&\leq 4 \lambda_{\textrm{ref}} \|f\|_\infty \mathbb{E}\left[\left( \sum_{i=1}^n U'(X_i)V_i \right)^2\right]^{1/2} = O(n^{1/2}),
\end{align*}
from Assumption \ref{ass:gradintegral} and using the fact that the terms are independent and mean zero.
Finally, by similar arguments as above the last term is given by
\begin{align*}
\mathbb{E}\left|\mathcal{A}_n^{(3)} \mathcal{A}_n^{(3)}f(\bm{X},\bm{V})\right|
&\leq 2\lambda_{\textrm{ref}} \mathbb{E}\left[\left|\mathcal{A}_n^{(3)}f(\bm{X},\bm{V})\right| \right]\\
&\leq 4\lambda_{\textrm{ref}}^2 \|f\|_\infty = O(1).
\end{align*}
Overall we have shown that the error term defined in \eqref{eq:Rndef} satisfies
$$\mathcal{R}_n= \frac{\epsilon_n}{2} \mathbb{E}\left[ \left| \mathcal{A}_n \mathcal{A}_n f \left( \bm{Z}_n (0)\right) \right| \right]
= O(n \epsilon_n) = o(1),$$
since we have chosen $\epsilon_n$ such that $\epsilon_n n^2\to 0$, as $n\to \infty$.
\paragraph{Main term.}
Having controlled the error term, we now focus on the main term given by
$$\mathbb{E}\left[ \left|\mathbb{E}\left[\left. \widetilde{\mathcal{A}}_n f \left(\bm{Z}_n(0) \right)\right| \mathcal{G}_0^n\right]- \mathcal{A}f\left( Z_n^{(1)}(0) \right)\right| \right],$$
where we recall that $\widetilde{\mathcal{A}}_n$ is the extended generator.
Notice that for $f(\bm{x},\bm{v}) = f(x_1, v_1)$,
\begin{align*}
\mathcal{\mathcal{A}}_n f\left(\bm{x},\bm{v} \right)
&= \partial_x f(x_1, v_1) v_1 + \max\left\{ 0, \langle \nabla U_n(\bm{x}), \bm{v}\rangle \right\} \left[ \mathfrak{R}_n f(\bm{x}, \bm{v})-f(\bm{x}, \bm{v})\right] + \lambda_{\textrm{ref}}\left[ Q f(x_1, v_1) - f(x_1,v_1)\right]\\
\mathcal{\mathcal{A}} f\left(x_1,v_1 \right)
&= \partial_x f(x_1, v_1) v_1 - \partial_v f(x_1,v_1) U'(x_1) + \lambda_{\textrm{ref}}\left[ Q f(x_1, v_1) - f(x_1, v_1)\right],
\end{align*}
and thus the first and third terms are in fact identical and will cancel out. We thus only have to consider the difference of the second terms. We first notice that
\begin{align*}
\lefteqn{\mathbb{E}\left[ \left. \max\left\{ 0, \langle \nabla U_n(\bm{X}), \bm{V}\rangle \right\} \left[ \mathfrak{R}_n f(\bm{X}, \bm{V})-f(\bm{X}, \bm{V})\right] \right| \mathcal{G}_0^n \right] }\\
&= \mathbb{E}\Bigg[ \max\left\{ 0, U'(X_1) V_1 + \sum_{i=2}^n U'(X_i)V_i \right\}\\
&\qquad \times\left[ f\left(X_1, V_1 - 2 \frac{U'(X_1) V_1 + \sum_{i=2}^n U'(X_i)V_i}{U'(X_1)^2 + \sum_{i=2}^n U'(X_i)^2}U'(X_1)\right)
-f(X_1, V_1)\right] \Bigg| \mathcal{G}_0^n \Bigg] \\
&= \mathbb{E}\Bigg[ \max\left\{ 0, U'(X_1) V_1 + \sum_{i=2}^n U'(X_i)V_i \right\}\\
&\qquad \times
\partial_v f(X_1, V_1) \left\{-2 \frac{U'(X_1) V_1 + \sum_{i=2}^n U'(X_i)V_i}{U'(X_1)^2 + \sum_{i=2}^n U'(X_i)^2}U'(X_1)\right\}\Bigg| \mathcal{G}_0^n \Bigg] + \mathcal{E}_n,
\end{align*}
where, using the $C_r$-inequality,
\begin{align*}
|\mathcal{E}_n|
&\leq 2 U'(X_1)^2\|\partial^2_{vv} f\|_\infty
\mathbb{E}\left[ \left.
\frac{\left|\sum_{i=1}^n U'(X_i)V_i \right|^3}{\left(\sum_{i=1}^n U'(X_i)^2\right)^2} \right| X_1, V_1 \right]\\
&\leq C U'(X_1)^2\|\partial^2_{vv} f\|_\infty \mathbb{E}\left[ \left.
\frac{\left|U'(X_1)V_1 \right|^3}{\left(\sum_{i=2}^n U'(X_i)^2\right)^2} \right| X_1, V_1 \right]\\
&\quad + C U'(X_1)^2\|\partial^2_{vv} f\|_\infty
\mathbb{E}\left\{\left. \frac{1}{\left(\sum_{i=2}^n U'(X_i)^2\right)^2}
\mathbb{E}\left[ \left. \left|\sum_{i=2}^n U'(X_i)V_i \right|^3 \right| (X_j)_{j=1}^n, V_1 \right] \right| X_1, V_1 \right\}.
\end{align*}
The first term is bounded above by $C |U'(X_1)| |V_1|^3 \|\partial^2_{vv} f\|_\infty$.
To control the inner expectation we apply Holder's inequality and then Khintchine's inequality, see \cite{newman1975}, conditionally on $X_1, \dots, X_n$, to obtain
\begin{align*}
\mathbb{E}\left[ \left. \left|\sum_{i=2}^n U'(X_i)V_i \right|^3 \right| (X_j)_{j=1}^n, V_1 \right]
&\leq \mathbb{E}\left[ \left. \left|\sum_{i=2}^n U'(X_i)V_i \right|^4 \right| (X_j)_{j=1}^n, V_1 \right]^{3/4}\\
&\leq C \mathbb{E}\left[ \left. \left(\sum_{i=2}^n U'(X_i)V_i \right)^2 \right| (X_j)_{j=1}^n, V_1 \right]^{3/2}\\
&\leq C \mathbb{E}\left[ \left. \sum_{i=2}^n U'(X_i)^2V_i^2 \right| (X_j)_{j=1}^n, V_1 \right]^{3/2}
=C \left[\sum_{i=2}^n U'(X_i)^2 \right]^{3/2},
\end{align*}
since the $V_i$'s have all mean zero and unit variance.
It follows that
\begin{align*}
|\mathcal{E}_n|
&\leq C |U'(X_1)| |V_1|^3 \|\partial^2_{vv} f\|_\infty +
2 C U'(X_1)^2\|\partial^2_{vv} f\|_\infty
\mathbb{E}\left[ \left.
\frac{\left[\sum_{i=2}^n U'(X_i)^2 \right]^{3/2}}{\left(\sum_{i=1}^n U'(X_i)^2\right)^2} \right| X_1, V_1 \right]\\
&\leq C |U'(X_1)| |V_1|^3 \|\partial^2_{vv} f\|_\infty + C |U'(X_1)|\, \|\partial^2_{vv} f\|_\infty
\mathbb{E}\left[ \left.
\left[ \frac{\sum_{i=2}^n U'(X_i)^2}{\sum_{i=1}^n U'(X_i)^2} \right]^{3/2}
\frac{|U'(X_1)|}{\left(\sum_{i=1}^n U'(X_i)^2\right)^{1/2}}
\right| X_1, V_1 \right]\\
&\leq C |U'(X_1)| |V_1|^3 \|\partial^2_{vv} f\|_\infty + C |U'(X_1)|\,\|\partial^2_{vv} f\|_\infty\\
&= C |U'(X_1)|\,\|\partial^2_{vv} f\|_\infty
\left[ |V_1|^3 +1 \right],
\end{align*}
which is in $L^1(\pi)$ by Assumption~\ref{ass:gradintegral}.
Notice that we can also write
\begin{align*}
\mathbb{E} |\mathcal{E}_n|
&\leq 4C \|\partial^2_{vv} f\|_\infty\mathbb{E} \left\{
\frac{|U'(X_1)|^2 \left[ 1+|V_1|^3\right]}{\left(\sum_{i=1}^n U'(X_i)^2\right)^{1/2}}
\right\} \to 0,
\end{align*}
as $n\to \infty$, by the dominated convergence theorem, since the integrand vanishes a.s.\ by a simple application of the law of large numbers since the $X_i$ are i.i.d.,
where as in the earlier calculation we showed that
$$|\mathcal{E}_n| \leq 2 |U'(X_1)|\, \|\partial^2_{vv}f\|_\infty \times (|V_1|^3+1)\in L^1(\pi).$$
Finally, having controlled the error terms,
to complete the proof of \eqref{eq:fourthcondition}, it remains to show that the following term vanishes
\begin{align*}
&\mathbb{E}_{\pi} \bigg[\left|\partial_v f(X_1, V_1)\, U'(X_1)\right|\times \\
&\qquad \qquad \bigg| \mathbb{E}\bigg[ \max\bigg\{ 0, \sum_{i=1}^n U'(X_i)V_i \bigg\}
\left( \frac{-2 \sum_{i=1}^n U'(X_i)V_i}{\sum_{i=1}^n U'(X_i)^2}\right) \bigg| X_1, V_1 \bigg] + 1 \bigg| \bigg].
\end{align*}
First we compute
\begin{align*}
\lefteqn{\mathbb{E}\left[\left. \max\left\{ 0, \sum_{i=1}^n U'(X_i)V_i \right\}
\left( \frac{-2 \sum_{i=1}^n U'(X_i)V_i}{\sum_{i=1}^n U'(X_i)^2}\right) \right| X_1, V_1 \right]}\\
&\qquad = -2\mathbb{E}\left[\left. \frac{\left(\sum_{i=1}^n U'(X_i)V_i\right)^2}{\sum_{i=1}^n U'(X_i)^2} \mathbb{1}\left\{ \frac{\sum_{i=1}^n U'(X_i)V_i}{\left(\sum_{i=1}^n U'(X_i)^2\right)^{1/2}}>0\right\} \right| X_1, V_1 \right].
\end{align*}
For any fixed $x_1, v_1$, the random variable
$$Z_n:= \frac{\sum_{i=1}^n U'(X_i)V_i }{\left(\sum_{i=1}^n U'(X_i)^2\right)^{1/2}},$$
converges in distribution to a standard Gaussian by Slutsky's theorem, since $(X_i)_i$, $(V_i)_i$ are i.i.d.\ sequences and $X_i$ is independent of $V_i$
\begin{align*}
\mathbb{E}[ U'(X_i) V_i]&=\mathbb{E}[U'(X_i)] \mathbb{E}[V_i]=0,\\
\mathbb{E}\left[ \left( U'(X_i)^2 V_i\right)^2\right] &= \mathbb{E}\left[ U'(X_i)^2 \right]\mathbb{E}[V_i^2] = \mathbb{E}\left[ U'(X_i)^2 \right] =:\sigma^2,
\end{align*}
whence by the central limit theorem and the law of large numbers
$$\frac{1}{\sqrt{n}}\sum_{i=2}^n U'(X_i) V_i \Rightarrow \mathcal{N}(0,\sigma^2), \qquad \frac{1}{n} \sum_{i=1}^n \left[ U'(X_i)\right]^2 \to \sigma^2,$$
where $ \sigma^2<\infty$ because of Assumption~\ref{ass:gradintegral}.
Notice then that by using again Khintchine's inequality
\begin{align*}
\mathbb{E}\left[ \left. Z_n^4 \right| X_1, V_1 \right]
&\leq \mathbb{E}\left[ \left. \frac{\left(\sum_{i=1}^n U'(X_i)V_i\right)^4}{\left(\sum_{i=1}^n U'(X_i)^2\right)^2} \right| X_1, V_1 \right] \\
&\leq \mathbb{E}\left\{ \left. \mathbb{E} \left[ \left. \frac{\left(\sum_{i=1}^n U'(X_i)V_i\right)^4}{\left(\sum_{i=1}^n U'(X_i)^2\right)^2} \right| (X_j)_{j=0}^n, V_1\right] \right| X_1, V_1 \right\} \\
&\leq \mathbb{E}\left\{ \left. \frac{ \mathbb{E} \left[ \left. \left(\sum_{i=1}^n U'(X_i)V_i\right)^4\right| (X_j)_{j=0}^n, V_1\right]}{\left(\sum_{i=1}^n U'(X_i)^2\right)^2} \right| X_1, V_1 \right\} \\
&\leq \mathbb{E}\left\{ \left. \frac{ C\mathbb{E} \left[ \left. \left(\sum_{i=1}^n U'(X_i)V_i\right)^2\right| (X_j)_{j=0}^n, V_1\right]^2}{\left(\sum_{i=1}^n U'(X_i)^2\right)^2} \right| X_1, V_1 \right\} \\
&\leq \mathbb{E}\left\{ \left. \frac{ C\left[ \sum_{j=1}^n U'(X_j)^2\right]^2}{\left(\sum_{i=1}^n U'(X_i)^2\right)^2} \right| X_1, V_1 \right\} \leq C.
\end{align*}
Therefore $Z_n$ has bounded 4th moments, and therefore we conclude that $Z_n^2$ is uniformly integrable, and in particular we have for any $x_1, v_1$ that
$$-2\mathbb{E}[ Z_n^2 \mathbb{1}\{Z_n >0\} \mid X_1=x_1, V_1=v_1]\to -2\mathbb{E}[ \xi^2 \mathbb{1}\{\xi>0\}] = -1.$$
Thus
$$\mathbb{E}\bigg[ \max\bigg\{ 0, \sum_{i=1}^n U'(X_i)V_i \bigg\}
\left( \frac{-2 \sum_{i=1}^n U'(X_i)V_i}{\sum_{i=1}^n U'(X_i)^2}\right) \bigg| X_1, V_1 \bigg] + 1 \to 0,$$
pointwise. Notice also that
\begin{align*}
\lefteqn{\left|\mathbb{E}\bigg[ \max\bigg\{ 0, \sum_{i=1}^n U'(X_i)V_i \bigg\}
\left( \frac{-2 \sum_{i=1}^n U'(X_i)V_i}{\sum_{i=1}^n U'(X_i)^2}\right) \bigg| X_1, V_1 \bigg]\right|}\\
&\leq C \mathbb{E}\bigg[ \frac{\mathbb{E}\left[\left. \left(\sum_{i=1}^n U'(X_i)V_i\right)^2 \right| (X_j)_{j=0}^n, V_1 \right]}{\sum_{i=1}^n U'(X_i)^2} \bigg| X_1, V_1 \bigg]\\
&\leq C \mathbb{E}\bigg[ \frac{\sum_{i=1}^n U'(X_i)^2}{\sum_{i=1}^n U'(X_i)^2} \bigg| X_1, V_1 \bigg] \leq C,
\end{align*}
and therefore by the dominated convergence theorem, since $\|\partial_ v f\|_\infty <\infty$ and $\pi \left(|U'|\right)<\infty$ we conclude that
\begin{align*}
&\mathbb{E}_{\pi} \bigg[\left|\partial_v f(X_1, V_1)\, U'(X_1)\right|\times \bigg| \mathbb{E}\bigg[ \max\bigg\{ 0, \sum_{i=1}^n U'(X_i)V_i \bigg\}
\left( \frac{-2 \sum_{i=1}^n U'(X_i)V_i}{\sum_{i=1}^n U'(X_i)^2}\right) \bigg| X_1, V_1 \bigg] + 1 \bigg| \bigg]\to 0,
\end{align*}
as $n\to \infty$.
\subsubsection{Proof of \eqref{eq:834}.}
Next we need to verify \eqref{eq:834}for some $p>1$ for which we proceed as follows
\begin{align*}
\mathbb{E}\left[\left(\int_{0}^{T}|\phi_{n}(t)|^{p}\mathrm{d}t\right)^{1/p}\right]^p
&\leq\mathbb{E}\left[\int_{0}^{T}|\phi_{n}(t)|^{p}\mathrm{d}t\right]=\int_{0}^{T}\mathbb{E}\left[|\phi_{n}(t)|^{p}\right]\mathrm{d}t\\
&=\int_{0}^{T}\mathbb{E}\left[\left|
\epsilon_n^{-1} \mathbb{E}\left\{\left. f\left(Z_n^{(1)}(t+\epsilon_n)\right)- f\left(Z_n^{(1)}(t)\right)\right| \mathcal{G}_t^n \right\}
\right|^{p}\right]\mathrm{d}t\\
&=\int_{0}^{T}\mathbb{E}\left[\left|
\epsilon_n^{-1} \mathbb{E}\left\{\left.
\int_0^{\epsilon_n}\widetilde{\mathcal{A}}_n f\left(Z_n^{(1)}(t+s) \right) + R_{t+s}
\right| \mathcal{G}_t^n \right\}\mathrm{d} s
\right|^{p}\right]\mathrm{d}t\\
\intertext{and using the fact that $\mathbb{E}[ R_{t+s} \mid \mathcal{G}_t^n] = \mathbb{E}[ \mathbb{E}[ R_{t+s} \mid \mathcal{F}_t^n ] \mid \mathcal{G}_t^n]=0$}
&=\int_{0}^{T}\mathbb{E}\left[\left|
\epsilon_n^{-1} \int_0^{\epsilon_n}\mathbb{E}\left\{\left.
\widetilde{\mathcal{A}}_n f\left(Z_n^{(1)}(t+s) \right)
\right| \mathcal{G}_t^n \right\}\mathrm{d} s
\right|^{p}\right]\mathrm{d}t\\
\intertext{and by Jensen's inequality}
&\leq\int_{0}^{T}\mathbb{E}\left[
\epsilon_n^{-1} \int_0^{\epsilon_n}\mathbb{E}\left\{\left. \left|
\widetilde{\mathcal{A}}_n f\left(Z_n^{(1)}(t+s) \right) \right|^{p}
\right| \mathcal{G}_t^n \right\}\mathrm{d} s
\right]\mathrm{d}t\\
&= \int_{0}^{T}\epsilon_n^{-1} \int_0^{\epsilon_n}\mathbb{E}\left[
\mathbb{E}\left\{\left. \left|
\widetilde{\mathcal{A}}_n f\left(Z_n^{(1)}(t+s) \right) \right|^{p}
\right| \mathcal{G}_t^n \right\}
\right]\mathrm{d} s\mathrm{d}t\\
&= \int_{0}^{T}\epsilon_n^{-1} \int_0^{\epsilon_n}\mathbb{E}\left[
\left|
\widetilde{\mathcal{A}}_n f\left(Z_n^{(1)}(t+s) \right) \right|^{p}
\right]\mathrm{d} s\mathrm{d}t\\
&=T \mathbb{E}\left[
\left|
\widetilde{\mathcal{A}}_n f\left(Z_n^{(1)}(0) \right) \right|^{p}
\right],
\end{align*}
by stationarity.
Next recalling the decomposition of $\widetilde{\mathcal{A}_n}$ into $\mathcal{A}_n^{(i)}$, $i=1,2,3$ notice that
$$\sup_{x,v} \left|\mathcal{A}_n^{(1)}f(x,v)\right| = \sup_{x,v}\left|\partial_x f(x,v) v\right|<\infty,$$
since $(x,v)\mapsto \partial_x f (x,v) v$ is continuous and has compact support, since $f$ has compact support. Similarly it follows easily that $\|\mathcal{A}_n^{(3)}f\|_\infty <\infty$ and therefore the only term we have to control corresponds to $\mathcal{A}_n^{(2)}$. For this term notice that
\begin{align*}
\lefteqn{\mathbb{E}\left[
\left| \mathcal{A}_n^{(2)} f\left(Z_n^{(1)}(0) \right) \right|^{p}
\right]}\\
&= \mathbb{E}\left[
\left| \max\left\{0, \sum_{i=1}^n U'(X_i) V_i \right\}
\left[ f\left(X_1, V_1 - 2 \frac{\sum_{i=1}^n U'(X_i)V_i}{\sum_{i=1}^n U'(X_i)^2}U'(X_1)\right) -f(X_1, V_1)\right] \right|^{p}
\right]\\
&\leq 2\|\partial_v f\|
\mathbb{E}\left[
\frac{ \left| \sum_{i=1}^n U'(X_i) V_i \right|^{2p} |U'(X_1)|^p}
{ \left(\sum_{i=1}^n U'(X_i)^2 \right)^p }\right]\\
&\leq 2\|\partial_v f\|
\mathbb{E}\left[ \frac{ |U'(X_1)|^p}
{ \left(\sum_{i=1}^n U'(X_i)^2 \right)^p }
\mathbb{E}\left\{ \left. \left| \sum_{i=1}^n U'(X_i) V_i \right|^{2p} \right| X_1, \dots, X_n \right\}\right].
\end{align*}
Once again using Khintchine's inequality, with $p\leq 2$,
\begin{align*}
\mathbb{E}\left[
\left| \mathcal{A}_n^{(2)} f\left(Z_n^{(1)}(0) \right) \right|^{p}
\right]
&\leq 2\|\partial_v f\|
\mathbb{E}\left[ \frac{ |U'(X_1)|^p}
{ \left(\sum_{i=1}^n U'(X_i)^2 \right)^p }
\mathbb{E}\left\{ \left. \left| \sum_{i=1}^n U'(X_i) V_i \right|^{2p} \right| X_1, \dots, X_n \right\}\right]\\
&\leq 2 B_p \|\partial_v f\|
\mathbb{E}\left[ \frac{ |U'(X_1)|^p}
{ \left(\sum_{i=1}^n U'(X_i)^2 \right)^p }
\mathbb{E}\left\{ \left. \sum_{i=1}^n U'(X_i)^2 \right| X_1, \dots, X_n \right\}^p\right]\\
&= 2 B_p \|\partial_v f\|
\mathbb{E}\left[ \frac{ |U'(X_1)|^p}
{ \left(\sum_{i=1}^n U'(X_i)^2 \right)^p }
\left(\sum_{i=1}^n U'(X_i)^2 \right)^p\right]\\
&= 2 B_p\|\partial_v f\|
\mathbb{E}\left[ |U'(X_1)|^p \right]<\infty,
\end{align*}
from Assumption~\ref{ass:gradintegral} and this bound is independent of $n$.
\subsubsection{Proof of \eqref{eq:firstcondition} and \eqref{eq:secondcondition}}
Notice that \eqref{eq:firstcondition} follows immediately since $\|f\|_\infty < \infty$, whereas \eqref{eq:secondcondition} follows from calculations similar to the ones used to prove
\eqref{eq:834}.
\section{Proofs of Wasserstein rates}
\subsection{Proof of Theorem~\ref{thm:Wasserstein}}
Let $\widetilde{X}(t):=X^{(2)}(t)-X^{(1)}(t)$ and $\widetilde{V}(t):=V^{(2)}(t)-V^{(1)}(t)$ denote the differences between the two paths in position and momentum. Ignoring for the moment the refreshment events, $\left(\widetilde{X}(t), \widetilde{V}(t)\right)$ will evolve according to the Hamiltonian dynamics, that is
\begin{equation}\label{eq:Hamcouplingdynamicsforward}
\begin{aligned}
\widetilde{X}'(t)&=\widetilde{V}(t),\\
\widetilde{V}'(t)&=-(\nabla U( X^{(2)}(t) )-\nabla U( X^{(1)}(t) ))=-H(t) \widetilde{X}(t),\text{ where}\\
H(t)&:=\int_{s=0}^{1} \nabla^2 U(sX^{(1)}(t)+ (1-s)X^{(2)}(t)) \mathrm{d} s.
\end{aligned}
\end{equation}
By convexity, we can see that $H(t)$ satisfies that $m I\preceq H(t)\preceq MI$ where $I$ denotes the identity matrix, where we write
$A\preceq B$ to denote that $B-A$ is positive definite. The effect of the generator $L_{1,2}$ on $\|\widetilde{X}(t)\|^2$, $\inner{\widetilde{X}(t)}{\widetilde{V}(t)}$ and $\|\widetilde{V}(t)\|^2$ is given by
\begin{equation}\label{eq:L12forward}
\begin{aligned}
L_{1,2} \|\widetilde{X}(t)\|^2&=2\inner{\widetilde{X}(t)}{\widetilde{V}(t)},\\
L_{1,2}\inner{\widetilde{X}(t)}{\widetilde{V}(t)}&=\|\widetilde{V}(t)\|^2-\inner{\widetilde{X}(t)}{H(t)\widetilde{X}(t)}-\lambda_{\textrm{ref}} (1-\alpha)\inner{\widetilde{X}(t)}{\widetilde{V}(t)},\\
L_{1,2} \|\widetilde{V}(t)\|^2&=-2\inner{\widetilde{V}(t)}{H(t)\widetilde{X}(t)}-\lambda_{\textrm{ref}} (1-\alpha^2)\|\widetilde{V}(t)\|^2.
\end{aligned}
\end{equation}
The claim of Theorem~\ref{thm:Wasserstein} is equivalent to showing that $-\mu\cdot d_A^2(Z_1(t),Z_2(t))-L_{1,2} d_A^2(Z_1(t),Z_2(t))\ge 0$. This can be expressed as
\begin{align*}
&-\mu\cdot d_A^2(Z_1(t),Z_2(t))-L_{1,2} d_A^2(Z_1(t),Z_2(t))\\
&=
-\mu a \|\widetilde{X}(t)\|^2 + 2[-\mu b+\lambda_{\textrm{ref}} (1-\alpha)b -a]\inner{\widetilde{X}(t)}{\widetilde{V}(t)}+[-c \mu+\lambda (1-\alpha^2)c-2b]\|\widetilde{V}(t)\|^2\\
&+2b \inner{\widetilde{X}(t)}{H(t)\widetilde{X}(t)}+2c \inner{\widetilde{V}(t)}{H(t)\widetilde{X}(t)}.
\end{align*}
Let \begin{align*}
X&:=\l(\begin{matrix}\|\widetilde{X}(t)\|^2 &\inner{\widetilde{X}(t)}{\widetilde{V}(t)}\\ \inner{\widetilde{X}(t)}{\widetilde{V}(t)} & \|\widetilde{V}(t)\|^2\end{matrix}\r), \quad
P:=\l(\begin{matrix}\inner{\widetilde{X}(t)}{H(t)\widetilde{X}(t)} &\inner{\widetilde{V}(t)}{H(t)\widetilde{X}(t)}\\ \inner{\widetilde{V}(t)}{H(t)\widetilde{X}(t)} &\inner{\widetilde{V}(t)}{H(t)\widetilde{V}(t)} \end{matrix}\r),\\
V&:=\l(\begin{matrix}-\mu a & - a+b \lambda_{\textrm{ref}}(1-\alpha)-\mu b\\- a+b \lambda_{\textrm{ref}}(1-\alpha)-\mu b & -c \mu+c \lambda_{\textrm{ref}} (1-\alpha^2)-2b \end{matrix}\r),
\quad W:=\l(\begin{matrix}2 b & c\\c & 0 \end{matrix}\r).
\end{align*}
We have
\[-\mu\cdot d_A^2(Z_1(t),Z_2(t))-L_{1,2} d_A^2(Z_1(t),Z_2(t))=\mathrm{Tr}(VX+WP),\]
so our goal is to show that $\mathrm{Tr}(VX+WP)\ge 0$ for all the possible $X, P$.
Using the fact that $m I\preceq H(t)\preceq M I$, we have $0\preceq m X\preceq P\preceq M X$.
Let $Y:=P-mX$, and $Z:=M X-P$, then $Y\succeq 0$, $Z\succeq 0$, and for $M>m$, we have
\begin{align*}&X=\frac{Y+Z}{M-m}, \quad P=\frac{MY + m Z}{M-m},
\intertext{ and hence }
&\mathrm{Tr}(VX+WP)=\frac{1}{M-m}\l(\mathrm{Tr}((V+MW)Y+(V+mW)Z) \r).
\end{align*}
When $M=m$, we have $H(t)=M I$ and $P=MX$, hence
\[\mathrm{Tr}(VX+WP)=\mathrm{Tr}((V+MW) X).\]
Note that in both cases, $\mathrm{Tr}(VX+WP)\ge 0$ if both $V+MW\succeq 0$ and $V+mW\succeq 0$. This can be equivalently written as the following set of inequalities,
\begin{align}
\label{ineq1}-\mu a+2M b&\ge 0,\\
-\mu a+2m b&\ge 0,\\
-c \mu+c \lambda_{\textrm{ref}} (1-\alpha^2)-2b&\ge 0,\\
(-a+b \lambda_{\textrm{ref}}(1-\alpha)-\mu b +M c)^2 &\le (-\mu a+2M b)(-c \mu+c \lambda_{\textrm{ref}} (1-\alpha^2)-2b),\\
\label{ineq5}(-a+b \lambda_{\textrm{ref}}(1-\alpha)-\mu b +m c)^2 &\le (-\mu a+2m b)(-c \mu+c \lambda_{\textrm{ref}} (1-\alpha^2)-2b).
\end{align}
As we have stated, let $\lambda_{\textrm{ref}}=\frac{1}{1-\alpha^2}\l(2\sqrt{M+m}-\frac{(1-\alpha) m}{\sqrt{M+m}}\r)$, $\mu=\frac{(1+\alpha)m}{\sqrt{M+m}}-\frac{\alpha m^{3/2}}{2(M+m)}$. Moreover, let $a:=1$,
\begin{align}\label{eq:Wassbcdef}
b&:=\frac{1+\alpha-\alpha \l(\frac{m}{M+m}\r)^{3/4}+\frac{3}{4} \frac{\alpha m}{M+m}}{2\sqrt{M+m}},\\
c&:=\frac{1+\alpha- \frac{\alpha }{2}\l(\frac{m}{M+m}\r)^{1/2} }{M+m}.
\end{align}
For this choice of $a,b,c$, the five inequalities can be shown to hold for every possible $0\le \alpha<1$, $0<m\le M$ (by homogeneity,
they can be rearranged to only depend on $\alpha$ and $m/M$, and then proven using for example Mathematica). Hence the bound \eqref{eq:Wassersteincontraction1} follows.
To show our $L^2$ bounds, we are also going to study the adjoint process $(P^t)^*$. Using the exact same coupling as before, the dynamics \eqref{eq:Hamcouplingdynamicsforward} ran backwards in time becomes
\begin{equation}\label{eq:Hamcouplingdynamicsbackward}
\begin{aligned}
\widetilde{X}'(t)&=-\widetilde{V}(t),\\
\widetilde{V}'(t)&=H(t) \widetilde{X}(t),\\
\end{aligned}
\end{equation}
with $H(t)$ defined as in \eqref{eq:Hamcouplingdynamicsforward}. For the velocity updates, forward in time we had $v'=\alpha v+\sqrt{1-\alpha^2}Z$ where $Z\sim N(0,I_d)$. Since in stationary we have $v,v'\sim N(0,I_d)$ and $\mathbb{E}(v (v')^T)=\rho I_d$, one can see that the updates backward in time are still the same.
Hence the effect of the adjoint becomes
\begin{equation}\label{eq:L12backward}
\begin{aligned}
L_{1,2}^* \|\widetilde{X}(t)\|^2&=2\inner{\widetilde{X}(t)}{-\widetilde{V}(t)},\\
L_{1,2}^*\inner{\widetilde{X}(t)}{\widetilde{V}(t)}&=\|\widetilde{V}(t)\|^2-\inner{\widetilde{X}(t)}{H(t)\widetilde{X}(t)}-\lambda_{\textrm{ref}} (1-\alpha)\inner{\widetilde{X}(t)}{-\widetilde{V}(t)},\\
L_{1,2}^* \|\widetilde{V}(t)\|^2&=-2\inner{-\widetilde{V}(t)}{H(t)\widetilde{X}(t)}-\lambda_{\textrm{ref}} (1-\alpha^2)\|\widetilde{V}(t)\|^2.
\end{aligned}
\end{equation}
Notice that this is very similar to the forward case \eqref{eq:L12forward}, except that we need to replace $\widetilde{V}(t)$ by $-\widetilde{V}(t)$. Based on this, by repeating the previous argument for $A':=\l(\begin{matrix}a &-b\\ -b &c\end{matrix}\r)$, we have
\begin{equation}\label{eq:Wassersteincontraction2}
L_{1,2}^* \ d_{A'}^2(Z_1(t),Z_2(t))\le -\mu\cdot d_{A'}^2(Z_1(t),Z_2(t)),
\end{equation}
where $a$, $b$ and $c$ are defined as in \eqref{eq:Wassbcdef}.
Hence we have shown that the adjoint process is also a contraction with the same rate $\mu$, but with respect to a different metric $d_{A'}$ instead of $d_A$ used for the forward process. The difference of these metrics means that we are not able to show contraction for the semigroup $((P^t)*+P^t)/2$. Now we are going to show that $d_A^2$ and $d_{A'}^2$ are equivalent up to a constant factor $C:=\frac{ac+b^2+2\sqrt{ac b^2}}{ac-b^2}$. Notice that for any $z_1,z_2\in \mathcal{R}^{2d}$,
\begin{equation}\label{eqdAApequivalence}
d_A^2(z_1,z_2)/C\le d_{A'}^2(z_1,z_2)\le d_A^2(z_1,z_2)\cdot C,
\end{equation}
as long as $A' \preceq C A$ and $A\preceq C A'$, and by rearrangement, this is equivalent to
\[\l(\begin{matrix}a(C-1)& -b(1+C)\\-b(1+C)& c(C-1) \end{matrix}\r)\succeq 0 \text{ and } \l(\begin{matrix}a(C-1)& b(C+1)\\b(C+1)& c(C-1) \end{matrix}\r)\succeq 0,\]
which holds for $C$ defined as above.
For $f:\mathcal{R}^{2d}\to \mathcal{R}$, let
$$\|f\|_{\mathrm{Lip},d_A}:=\sup_{z_1,z_2\in \mathcal{R}^{2d}, z_1\ne z_2}\frac{|f(z_1)-f(z_2)|}{d_{A}(z_1,z_2)},$$
be its Lipschitz coefficient with respect to the $d_A$ distance.
Then based on \eqref{eq:Wassersteincontraction1},\eqref{eq:Wassersteincontraction2}, and \eqref{eqdAApequivalence}, for any $t\ge 0$, $f:\mathcal{R}^{2d}\to \mathcal{R}$, have
\begin{align*}
\|(P^t)^* P^t f \|_{\mathrm{Lip},d_A}&\le \sqrt{C}\|(P^t)^* P^t f \|_{\mathrm{Lip},d_{A'}}\le \sqrt{C}\exp\l(-\frac{\mu t}{2}\r) \|P^t f\|_{\mathrm{Lip},d_{A'}}\\
&\le C\exp\l(-\frac{\mu t}{2}\r) \|P^t f\|_{\mathrm{Lip},d_{A}}
\le C\exp\l(-\mu t\r) \|f\|_{\mathrm{Lip},d_{A}}.
\end{align*}
Based on Propositions 29 and 30 of \cite{OllivierJFA} with $\kappa=1-C\exp\l(-\mu t\r)$, it follows that for any $t>\frac{\log(C)}{\mu}$, the reversible kernel $(P^t)^* P^t$ has as spectral radius of at most $C\exp\l(-\mu t\r)$. Thus for every $f\in L_0^{2}(\pi)$, we have
\begin{equation}
\|P^t f \|^2=\inner{f}{(P^t)^* P^t f} \leq \|f\| \|(P^t)^* P^t f\|\le C e^{-\mu t} \|f\|^2,
\end{equation}
and the claim of the Theorem follows by noticing that $\|P^t f \|^2\le \|f\|^2$ for every $t\ge 0$.
\begin{remark}
We note that for any given $\lambda_{\textrm{ref}}>0, \mu>0$, the contraction rate of $d_A^2(Z_1(t),Z_2(t))$ is at least $\mu$ as long as there are constants $a,b,c$ such that $a>0$, $c>0$, $b^2<ac$ and inequalities \eqref{ineq1}-\eqref{ineq5} hold. Unfortunately due to the non-linearity of these inequalities we did not manage to find an analytical expression for the largest possible $\mu$ for a given $\lambda_{\textrm{ref}}$ (and then the largest possible $\mu$ for any $\lambda_{\textrm{ref}}$). The reader can possibly slightly improve these rates by numerical optimization for a given $\alpha$, $m$ and $M$. Note however that in our numerical experiments, it seems that the choices of $\lambda_{\textrm{ref}}$ as stated leads to $\mu$ that is close to optimal in most of the domain $0\le \alpha<1$, and $0<m\le M$ (i.e. if we increase $\mu$ by a few percent, typically there is no longer a $\lambda_{\textrm{ref}}>0$ and parameters $a,b,c$ satisfying all of the inequalities).
\end{remark}
\subsection{Proof of Proposition~\ref{prop:WassersteinGauss}}
Assume without loss of generality that $m=1$ (the general case can be obtained from this by rescaling). Let $D:=\l(\begin{matrix}aH & bI\\b I & c I \end{matrix}\r)$ be a block matrix.
Then
\[d_{D}^2(Z_1(t),Z_2(t))=a\inner{\widetilde{X}(t)}{H \widetilde{X}(t)}+2b\inner{\widetilde{X}(t)}{\widetilde{V}(t)}+ c\|\widetilde{V}(t)\|^2,\]
and the effect of the generator on these terms equal
\begin{align}
L_{1,2} \inner{\widetilde{X}(t)}{H \widetilde{X}(t)}&=2\inner{\widetilde{X}(t)}{H\widetilde{V}(t)},\\
L_{1,2}\inner{\widetilde{X}(t)}{\widetilde{V}(t)}&=\|\widetilde{V}(t)\|^2-\inner{\widetilde{X}(t)}{H\widetilde{X}(t)}-\lambda_{\textrm{ref}} (1-\alpha)\inner{\widetilde{X}(t)}{\widetilde{V}(t)},\\
L_{1,2} \|\widetilde{V}(t)\|^2&=-2\inner{\widetilde{V}(t)}{H\widetilde{X}(t)}-\lambda_{\textrm{ref}} (1-\alpha^2)\|\widetilde{V}(t)\|^2.
\end{align}
We have
\begin{align*}
&-\mu\cdot d_D^2(Z_1(t),Z_2(t))-L_{1,2} d_D^2(Z_1(t),Z_2(t))\\
&=
2[-\mu b+\lambda_{\textrm{ref}} (1-\alpha)b]\inner{\widetilde{X}(t)}{\widetilde{V}(t)}+[-c \mu+\lambda_{\textrm{ref}} (1-\alpha^2)c-2b]\|\widetilde{V}(t)\|^2\\
&+(2b-\mu a) \inner{\widetilde{X}(t)}{H\widetilde{X}(t)}+2(c-a) \inner{\widetilde{V}(t)}{H\widetilde{X}(t)}.
\end{align*}
Let $X$ and $P$ defined as in the proof of Theorem \ref{thm:Wasserstein}, and let
\begin{align*}
V&:=\l(\begin{matrix}0 & b \lambda_{\textrm{ref}}(1-\alpha)-\mu b\\ b \lambda_{\textrm{ref}}(1-\alpha)-\mu b & -c \mu+c \lambda_{\textrm{ref}} (1-\alpha^2)-2b \end{matrix}\r),
\quad W:=\l(\begin{matrix}2 b-\mu a & c-a\\c-a & 0 \end{matrix}\r).
\end{align*}
Then we have $-\mu\cdot d_D^2(Z_1(t),Z_2(t))-L_{1,2} d_D^2(Z_1(t),Z_2(t))=\mathrm{Tr}(VX+WP)$, and using the same argument as in the proof of Theorem \ref{thm:Wasserstein},
it follows that $\mathrm{Tr}(VX+WP)\ge 0$ if both $V+MW\succeq 0$ and $V+mW\succeq 0$. This can be verified (for example by Mathematica) for the choices
$\lambda_{\textrm{ref}}=2\sqrt{m}/(1-\alpha)$, $\mu=\frac{\sqrt{m}}{3}$, $a=1$, $b=\frac{1}{4}$, $c=1$.
The proof of \eqref{eqL2bndGaussian} is analogous to the proof of \eqref{eqL2bndWass}. First we show that for $D':=\l(\begin{matrix}aH & -bI\\-b I & c I \end{matrix}\r)$,
\begin{equation}
L_{1,2}^* \ d_{D'}^2(Z_1(t),Z_2(t))\le -\mu\cdot d_{D'}^2(Z_1(t),Z_2(t)),
\end{equation}
then use the same argument as previously.
\section{Proof of Theorem~\ref{thm:hypoco}}
The generator of the RHMC process will be denoted by $\mathcal{A}$ and it is given for smooth enough functions by
$$\mathcal{A}f(x,v) = \langle \nabla_x f, v\rangle - \langle \nabla U, \nabla_v f\rangle
+ \lambda_{\textrm{ref}} \left[Q_\alpha f(x,v) - f(x,v)\right],$$
where recall that $\alpha \in (0,1)$ and
$$Q_\alpha f(x,v):= \frac{1}{\sqrt{2\pi}^d}\int \mathrm{e}^{-\bm{\xi}'\bm{\xi}/2}
f\left(x,\alpha v+\sqrt{1-\alpha^2} \xi\right) \mathrm{d} \bm{\xi}.$$
\paragraph{Asymptotic Variance.}
In the context of MCMC one is interested in optimising the computational resources
needed to produce an estimate of a certain precision. For this reason we are also interested in understanding the asymptotic variance.
Geometric ergodicity is enough to show that a large class of functions, determined by the Lyapunov function, have finite asymptotic variance.
However, since the convergence rates are not explicit in the parameters of the process, geometric ergodicity often does not allow one to optimise the asymptotic variance.
Usually controlling the asymptotic variance for a large enough class of functions is closely related to establishing a spectral gap.
In the reversible case, it is well known that geometric ergodicity is equivalent to having a spectral gap,
but in the non-reversible case this is no longer true, see \cite{kontoyiannis2012} and references therein.
This fact is actually observed for piecewise deterministic Markov processes such as the BPS and Zig-Zag samplers,
see \cite{peters2012rejection,bouncy2018,bierkens2016zig}. This class of processes also includes RHMC.
Although geometric ergodicity has been established for BPS (\cite{deligiannidis2017exponential,durmus2018geometric}), Zig-Zag (see \cite{bierkens2017ergodicity,fetique2017long}) and RHMC (\cite{RHMC}) \cite{RHMC},
and easy calculations show that, writing $\mathcal{L}$ for the generator of any of the above processes, we have $\langle \mathcal{L}f, f\rangle = 0$ for any function $f\in L^2(\pi)$ such that $f(x,v)= f(x)$, that is functions of the location only. The reason for this is that the Dirichlet form $\mathcal{E}(f,f):=\langle \mathcal{L}f, f \rangle$ only captures the symmetric part of the generator $\mathcal{L}$, which in these processes only affects the velocity component, whereas the location component is only affected by the anti-symmetric part of the generator. This means that although BPS, Zig-Zag and RHMC are geometrically ergodic, there is no hope of obtaining a spectral gap in the classical $L^2$ sense.
In fact this situation arises very often in so called kinetic equations which include for example the underdamped Langevin processes. For such processes a range of methods have been developed recently that are widely termed as \textit{hypocoercivity},
see \cite{nier2005hypoelliptic,dric2009hypocoercivity,dolbeault2015hypocoercivity} and references therein. In fact such methods have already been applied to piecewise deterministic Markov processes, see \cite{monmarche2014hypocoercive}.
Although this approach is often quite deep and involved, the underlying principle is that of adjusting the norm, or metric, in which the convergence is studied. This principle has been extremely successful recently, for example in the convergence of HMC when log-concavity fails locally in \cite{bou2018coupling}. In the case of hypocoercive estimates, the principle is to move from the $L^2$ norm to a stronger norm, usually some form of Sobolev norm.
\subsection{Strong continuity in $H^1(\pi)$. }
We will establish that the abstract Cauchy problem
\begin{align*}
\frac{\partial u(t,z) }{\partial t} &= \mathcal{A} u,\\
u(0,z) &= f,
\end{align*}
admits a unique solution in $H^1(\pi)$ given by $u(t,z):= P^t f(z)$. This will justify computing the time derivatives of $\langle\!\langle P^t f, P^t f \rangle\!\rangle$.
Before we proceed we will need to introduce some additional notation. We decompose the generator $\mathcal{A}$ of RHMC into its symmetric and antisymmetric component as follows
$$\mathcal{A}f(x,v) = Bf(x,v) + \lambda_{\textrm{ref}} (-S) f,$$
where
\begin{equation}\label{eq:definitionofB}
Bf := \langle \nabla_x f, v\rangle - \langle \nabla_v f, \nabla U \rangle, \qquad Sf := [I-Q_\alpha]f.
\end{equation}
As before we write $\{P^t: t\geq 0\}$ for the semi-group of transition kernels of RHMC, but in this section we slightly change our point of view and consider it as a semigroup on $L^2(\pi)$, that is
$P^t: L^2(\pi)\to L^2(\pi)$. Its generator will be given by $\mathcal{A}$ for smooth enough functions.
In fact even more is true as we will next show that $P^t$ is also strongly continuous as a semi-group on $H^1(\pi)$. To see why, first recall that the anti-symmetric operator $B$ generates the Hamiltonian flow $z\mapsto \Xi(t,z)$ with respect to $H(\bm{x},\bm{v}) = U(\bm{x}) + \bm{v}'\bm{v}/2$. Let us write $\{Q^t:t\geq 0\}$ for the semigroup generated by $B$, that is
$Q^t f(z) = f\left( \Xi(t,z)\right)$ for $z\in \mathcal{Z}$. Then
given a smooth function $f\in H^1(\pi)$, from the chain rule we have
$$\nabla P^t f(z) = \nabla f\left( \Xi(t,z)\right) \nabla\Xi(t,z).$$
From the variational equations of the Hamiltonian dynamics (see Section 6.1.2 of \cite{IntroductiontoHamiltonian}) and the upper bounds $M$ and $1$ of the Hessians of $U(\bm{x})$ and $\frac{\|\bm{v}\|^2}{2}$ it follows that for $C=\max(1,M)$, we have
$\|\nabla\Xi(t,z)\|\leq \mathrm{e}^{Ct}$ for every $t\ge 0$. Using this, we conclude that
\begin{align*}
\|\nabla_x P^t f\|^2 + \|\nabla_v P^t f\|^2
&\leq \mathrm{e}^{2Ct} \iint \pi(\mathrm{d} z)\left[ \left\|\nabla_x f \left(\Xi(t,z)\right) \right\|^2
+\left\|\nabla_v f \left(\Xi(t,z)\right) \right\|^2\right]\\
&= \mathrm{e}^{2Ct} \iint \pi(\mathrm{d} z)\left[ \left\|\nabla_x f \left(z\right) \right\|^2
+\left\|\nabla_v f \left(z\right) \right\|^2\right],
\end{align*}
by stationarity of the flow. By an approximation argument we can further show that
$Q^t: H^1(\pi)\to H^1(\pi)$ for all $\geq 0$. Finally $\{Q^t:t\geq 0\}$ is strongly continuous on $H^1(\pi)$, since
\begin{align}
\| \nabla Q^h f - \nabla f\|^2
&= \int\| \nabla f\left( \Xi(h, z) \right)\nabla \Xi(t,z) - \nabla f\left(z \right)\nabla \Xi(t,z)\|^2 \pi(\mathrm{d} z)\notag\\
&\leq \int\| \nabla f\left( \Xi(h, z) \right) \left[ \nabla \Xi(t,z) - I\right] \|^2 \pi(\mathrm{d} z) \notag\\
&\qquad + \int\| \nabla f\left( \Xi(h, z) \right) - \nabla f\left(z \right)\|^2 \pi(\mathrm{d} z)\notag\\
&\leq \int\| \nabla f\left( \Xi(h, z) \right)\|^2 \|\nabla \Xi(t,z) - I \|^2 \pi(\mathrm{d} z) \notag\\
&\qquad + \int\| \nabla f\left( \Xi(h, z) \right) - \nabla f\left(z \right)\|^2 \pi(\mathrm{d} z)\notag\\
&\leq \int\| \nabla f\left( \Xi(h, z) \right)\|^2 \|\nabla \Xi(t,z) - I \|^2 \pi(\mathrm{d} z) \notag\\
&\qquad + 2\int\|Q^h \nabla f(z) - \nabla f(z)\|^2 \pi(\mathrm{d} z).\label{eq:strongcontinuityinH1}
\end{align}
Since $g:=\nabla f\in L^2(\pi)$, for every $\epsilon>0$ there is a smooth, compactly supported function $g_\epsilon$ such that $\|g-g_\epsilon\|_{L^2(\pi)}<\epsilon$. Then
\begin{align*}
\int\|Q^h g(z) - g(z)\|^2 \pi(\mathrm{d} z)
&= \int\|Q^h g(z) - Q^h g_\epsilon(z)+ Q^h g_\epsilon(z) - g_\epsilon(z) + g_\epsilon(z)- g(z)\|^2 \pi(\mathrm{d} z)\\
&\leq \int \pi(\mathrm{d} z) \left\| g\left(\Xi(h,z)\right)-g_\epsilon\left(\Xi(h,z)\right)\right\|^2
+ \int \pi(\mathrm{d} z) \left\| g\left(z\right)-g_\epsilon\left(z\right)\right\|^2\\
&\qquad + \int \pi(\mathrm{d} z) \left\| g_\epsilon\left(\Xi(h,z)\right)-g_\epsilon\left(z\right)\right\|^2\\
&= 2 \|g-g_\epsilon\|_{L^2(\pi)} + \int \pi(\mathrm{d} z) \left\| g_\epsilon\left(\Xi(h,z)\right)-g_\epsilon\left(z\right)\right\|^2\\
&\leq 2 \epsilon + \int \pi(\mathrm{d} z) \left\| g_\epsilon\left(\Xi(h,z)\right)-g_\epsilon\left(z\right)\right\|^2.
\end{align*}
For every fixed $\epsilon >0$, the second term vanishes by bounded convergence. Since $\epsilon>0$ is arbitrary this shows that $\|Q^h \nabla f - \nabla f\|^2 \pi(\mathrm{d} z) \to 0$ as $h \to 0$.
Going back to \eqref{eq:strongcontinuityinH1}, notice that the first term also vanishes by the dominated convergence theorem, since $\|\nabla \Xi(t,z) - I\| \leq 2\mathrm{e}^{Ct}$ uniformly in $z$, $\|\nabla \Xi(t,z) -I \| \to 0$ pointwise.
Thus $Q^t$ is strongly continuous and therefore it admits a densely defined generator, which we denote by $B$,
$$B: \mathcal{D}(B) \subseteq H^1(\pi) \to H^1(\pi).$$
Again it is straightforward to check that $B$ has the expression given earlier.
In addition notice that $S$ is a bounded operator on $H^1(\pi)$. To see why first notice that an easy calculation, which will be provided later on for completeness,
shows that $\nabla_x Q_\alpha = Q_\alpha \nabla_x$ and $\nabla_v Q_\alpha = \alpha Q_\alpha \nabla_v$ whence
$$\|\nabla_x Q_\alpha f\|^2+ \|\nabla_v Q_\alpha f\|^2 \leq
\| Q_\alpha \nabla_xf\|^2+ \alpha |Q_\alpha \nabla_v f\|^2 \leq C \left( \|\nabla_x f\|^2+ \|\nabla_v f\|^2\right),$$
since $Q_\alpha$ is a contraction on $L^2(\pi)$.
Therefore, applying \cite[Theorem~3.2]{phillips1953perturbation}, the operator
$\mathcal{A}:=B+\lambda_{\textrm{ref}}(-S)$ has domain $\mathcal{D}(B)$ and generates a strongly continuous on $H^1(\pi)$, which we will denote again by $\{P^t:t\geq 0\}$. This implies that for every $f\in \mathcal{D}(B)$, $P^t f \in \mathcal{D}(\mathcal{A})$ for all $t\geq 0$ and $\mathcal{A}P^t f= P^t \mathcal{A} f$.
This essentially shows that given $f \in \mathcal{D}(B)$ the abstract Cauchy problem
\begin{align*}
\frac{\partial u(t,z) }{\partial t} &= \mathcal{A} u,\\
u(0,z) &= f,
\end{align*}
admits a unique solution in $H^1(\pi)$ given by $u(t,z):= P^t f(z)$.
\subsection{Proof of Theorem~\ref{thm:hypoco}.}
We introduce some additional notation to keep the presentation concise. First recall the decomposition $\mathcal{A} = B +\lambda_{\textrm{ref}}(-S)$ where
$$
Bf = \langle \nabla_x f, v\rangle - \langle \nabla_v f, \nabla U \rangle, \qquad Sf = [I-Q_\alpha]f,$$
and let us define the Dirichlet form $\mathcal{E}(h,g) := \langle h, S g\rangle$.
We will also write $A:= \nabla_v$, $C:= \nabla_x$.
From \cite[p. 40]{dric2009hypocoercivity}, or an easy calculation,
we have
$$[A,B]=AB-BA=\nabla_x, \quad\text{ and} \qquad [B,C]= \nabla^2 U \cdot \nabla_v = \nabla^2 U \cdot A.$$
We need to compute the time derivative of $\langle\!\langle P^t f, P^t f\rangle\!\rangle$. To keep notation to a minimum we will write $h$ rather than $P^t f$.
Since $P^t = \exp (t \mathcal{A})$, where $\mathcal{A}$ is the generator of the RHMC process, an easy calculation shows that for all $h, g \in \mathcal{D}(\mathcal{B})$ we have
\begin{align*}
\frac{\mathrm{d}}{\mathrm{d} t }\langle P^t h, P^t g \rangle \Big|_{t=0}
&=\langle Lh, g\rangle +\langle h, Lg\rangle,
\end{align*}
This also implies that
$$\frac{\mathrm{d}}{\mathrm{d} t }\langle P^t h, P^t h \rangle \Big|_{t=0} = 2\langle Lh, h\rangle
=-2\lambda_{\textrm{ref}} \mathcal{E}(h,h).$$
since $B$ is antisymmetric, in the sense that
$\langle Bf, g\rangle=-\langle f, Bg\rangle$.
We want to compute $\mathrm{d} \langle\!\langle P^t h, P^t h\rangle\!\rangle/\mathrm{d} t|_{t=0}$. We proceed by computing the
derivative of each term individually,
\begin{align*}
\frac{\mathrm{d}}{\mathrm{d} t }\|A h\|^2
&= 2\langle A h, A \mathcal{L}h\rangle = -2\lambda_{\textrm{ref}}\langle A h, A Sh\rangle + 2\langle A h, A Bh\rangle, \\
\frac{\mathrm{d}}{\mathrm{d} t }\langle C f, A f\rangle
&= \langle C f, A (-\lambda_{\textrm{ref}} S+B) f\rangle + \langle C (-\lambda_{\textrm{ref}} S+B)f, A f\rangle,\\
\frac{\mathrm{d}}{\mathrm{d} t }\|C h\|^2
&= 2\langle C h, C \mathcal{L} h\rangle = -2\lambda_{\textrm{ref}}\langle C h, C S h\rangle + 2\langle C h, C B h\rangle,
\end{align*}
\paragraph{Term one. }
We now compute the first term which is given by
$$-2\lambda_{\textrm{ref}}\langle A h, A Sh\rangle + 2\langle A h, A Bh\rangle.$$
Notice that
\begin{align*}
\frac{\partial}{\partial v_i} Q_\alpha f(x,v)
&= \frac{\partial}{\partial v_i} \mathbb{E} \left[ f\left(x,\alpha v+ \sqrt{1-\alpha^2} \xi\right)\right]\\
&= \mathbb{E} \left[ \alpha f_{v_i}\left(x,\alpha v+ \sqrt{1-\alpha^2} \xi\right)\right]\\
&= \alpha \mathbb{E} \left[ f_{v_i}\left(x,\alpha v+ \sqrt{1-\alpha^2} \xi\right)\right],
\end{align*}
where to keep notation clear we write
$\partial G(x,v) /\partial v_i$ to denote the derivative of the expression $G(x,v)$ w.r.t.\ $v_i$, whereas we write $f_{v_i}\left(x,\alpha v+ \sqrt{1-\alpha^2} \xi\right)$ to denote the derivative
of $f$ w.r.t.\ $v_i$ evaluated at $\alpha v+ \sqrt{1-\alpha^2} \xi$.
The above calculation shows that $A Q_\alpha = \alpha Q_\alpha A$ and therefore
\begin{align*}
-\lambda_{\textrm{ref}} \langle A h, A S h\rangle
&= \lambda_{\textrm{ref}} \langle A h, A (Q_\alpha -I) h\rangle\\
&= \lambda_{\textrm{ref}} \langle A h, A Q_\alpha h\rangle - \lambda_{\textrm{ref}} \langle A h, A h\rangle\\
&= \lambda_{\textrm{ref}} \alpha \langle A h, Q_\alpha A h\rangle - \lambda_{\textrm{ref}} \langle A h, A h\rangle\\
&=\lambda_{\textrm{ref}}\langle A h, (\alpha Q_\alpha -I) A h\rangle\\
&=\lambda_{\textrm{ref}}\langle A h, \alpha( Q_\alpha -I) A h\rangle
-(1-\alpha)\lambda_{\textrm{ref}}\langle A h, A h\rangle\\
&=-\lambda_{\textrm{ref}}\alpha\langle A h, SA h\rangle
-(1-\alpha)\lambda_{\textrm{ref}}\langle A h, A h\rangle.
\end{align*}
Continuing we have
\begin{align*}
\langle A h, A B h\rangle
&= \langle A h, A B h\rangle\\
&= \langle A h, (A B - BA) h\rangle + \langle A h, B A h\rangle\\
&= \langle A h, [A,B] h\rangle + 0 = \langle A h, C h\rangle=\langle A h, C h\rangle.
\end{align*}
since by the anti-symmetry of $B$, it follows that $\langle g, Bg\rangle=0$ for any $g$.
\paragraph{Term two.}
We next compute the second term
$$\langle C f, A (-\lambda_{\textrm{ref}} S+B) f\rangle + \langle C (-\lambda_{\textrm{ref}} S+B)f, A f\rangle.$$
First we compute the derivative along $B$
\begin{align*}
\langle A B h, C h\rangle + \langle A h, C B h\rangle
&= \langle A B h, C h\rangle + \langle A h, C B h\rangle \\
&= \langle A B h, C h\rangle + \langle A h, BC h\rangle + \langle A h, [C,B] h\rangle \\
\intertext{and using that $B^\ast = -B$}
&= \langle A B h, C h\rangle - \langle BA h, C h\rangle + \langle A h, [C,B] h\rangle \\
&= \langle [A, B] h, C h\rangle + \langle A h, [C,B] h\rangle \\
&= \langle C h, C h\rangle + \langle A h, [C,B] h\rangle \\
&= \|Ch\|^2- \langle A h, \nabla^2 U A h\rangle.
\end{align*}
To compute the derivative along $S$ first notice that
$C Q_\alpha = Q_\alpha C$, where in the r.h.s.\ we tensorise $Q_\alpha$ allowing it to act on each component separately, in the sense that
$$\frac{\partial}{\partial x_i} \mathbb{E}\left[ f\left(x, \alpha v + \sqrt{1-\alpha^2} \xi\right)\right]
= \mathbb{E}\left[ \frac{\partial}{\partial x_i}f\left(x, \alpha v + \sqrt{1-\alpha^2} \xi\right)\right].$$
Therefore
\begin{align*}
\lefteqn{ -\lambda_{\textrm{ref}}\langle A S h, C h\rangle -\lambda_{\textrm{ref}}\langle A h, C S h\rangle }\\
&= \lambda_{\textrm{ref}} \langle A (Q_\alpha -I) h, C h\rangle + \lambda_{\textrm{ref}} \langle A h, C (Q_\alpha -I)h\rangle\\
&= \lambda_{\textrm{ref}}\langle (\alpha Q_\alpha -I) A h, C h\rangle + \lambda_{\textrm{ref}}\langle A h,
(Q_\alpha -I) C h\rangle\\
&= \alpha\lambda_{\textrm{ref}}\langle ( Q_\alpha -I) A h, C h\rangle + (\alpha-1) \lambda_{\textrm{ref}}\langle A h, C h\rangle+ \lambda_{\textrm{ref}}\langle A h, (Q_\alpha -I) C h\rangle\\
&= -(1+\alpha) \lambda_{\textrm{ref}} \langle SA h, C h\rangle -(1-\alpha) \langle A h, C h\rangle,
\end{align*}
where we used again the fact that $Q_\alpha$ is positive.
\paragraph{Term three.}
Using the same arguments as before we have
\begin{align*}
\langle C h, C Q_\alpha h\rangle
&= \sum_{i=1}^d \left\langle \frac{\partial}{\partial x_i}h, \frac{\partial}{\partial x_i} Q_\alpha h\right\rangle\\
&= \sum_{i=1}^d \left\langle\frac{\partial}{\partial x_i}h, Q_\alpha \frac{\partial}{\partial x_i} h\right\rangle
= \langle C h, Q_\alpha C h\rangle,
\end{align*}
where we are overloading the inner product by allowing it to take both vectors and scalars as arguments, in the case of scalars it integrates the product, in the case of vectors the vector inner product.
Therefore
\begin{align*}
-\lambda_{\textrm{ref}}\langle C h, C S h\rangle
&= \lambda_{\textrm{ref}} \langle C h, C (Q_\alpha -I) h\rangle
=-\lambda_{\textrm{ref}} \langle C h, SC h\rangle
\end{align*}
The next one is
\begin{align*}
\langle C h, C B h\rangle
&= \langle C h, C B h\rangle\\
&= \langle C h, B C h\rangle - \langle C h, [B,C] h\rangle = 0- \langle C h, \nabla^2 U\cdot A h\rangle.
\end{align*}
\paragraph{Combining all the terms.}
We now have the tools to compute the derivative of
$$\langle\!\langle h, h\rangle\!\rangle:=a\|A h\|^2 - 2b \inner{C h}{A h} + c\|C h\|^2,$$
which, after multiplying by $-1$, is given by
\begin{align*}
\lefteqn{-\frac{\mathrm{d} }{\mathrm{d} t} \langle\!\langle h, h\rangle\!\rangle}\\
&= - a\frac{\mathrm{d} }{\mathrm{d} t}\|A h\|^2 + 2b\frac{\mathrm{d} }{\mathrm{d} t} \inner{A h}{C h} - c\frac{\mathrm{d} }{\mathrm{d} t}\|C h\|^2\\
&=2 a \left[\lambda_{\textrm{ref}}(1-\alpha)\|A h\|^2 +\lambda_{\textrm{ref}}\alpha \inner{S Ah}{Ah}
-\inner{Ah}{Ch}\right]\\
&\qquad +2 b \left[ \|Ch\|^2 - \inner{\nabla^2 U Ah} {Ah}
- (1+\alpha) \lambda_{\textrm{ref}} \langle S^{1/2}Ah, S^{1/2}Ch\rangle
-(1-\alpha)\lambda_{\textrm{ref}}\inner{Ah}{Ch}\right]\\
&\qquad +2c \left[ \lambda \inner{S C h}{Ch} + \inner{\nabla^2 U Ah}{Ch}
\right]
\\
&= 2 a\lambda_{\textrm{ref}}(1-\alpha)\|A h\|^2 -2(a+(1-\alpha)b\lambda_{\textrm{ref}})\inner{Ah}{Ch} +2 b \|Ch\|^2
\\
&\qquad
-2b \inner{\nabla^2 U Ah} {Ah} + 2c\inner{\nabla^2 U Ah}{Ch}\\
&\qquad + 2 a\lambda_{\textrm{ref}}\alpha \inner{S Ah}{Ah}+ 2 c\lambda \inner{S C h}{Ch}- 2(1+\alpha) b\lambda_{\textrm{ref}} \inner{S Ah}{Ch}.
\end{align*}
\begin{remark}
At this stage we can rewrite the above inequality as
\begin{align}
-\frac{1}{2}\frac{\mathrm{d} }{\mathrm{d} t} \langle\!\langle h, h\rangle\!\rangle
&\geq
\left[ a(1-\alpha) \lambda_{\textrm{ref}} - bM\right]\|A h\|^2 +b \|Ch\|^2 - \left\|J Ah\right\| \|Ch\|
\notag\\
&\qquad +a\alpha \lambda_{\textrm{ref}} \|S^{1/2}Ah\|^2 + c\lambda_{\textrm{ref}} \|S^{1/2}Ch\|^2 - (1+\alpha)b\lambda_{\textrm{ref}}
\|S^{1/2}Ah\| \|S^{1/2}Ch\|,\label{eq:quadraticformtooptimize}
\end{align}
where $S^{1/2}$ is the positive, self-adjoint square root of $S$, and
$$Jf:= \left( aI+(1-\alpha)b \lambda_{\textrm{ref}} I -c \nabla^2 U\right) f,$$
which is also self-adjoint, since $\nabla^2 U$ is symmetric, whence its norm is given by
\begin{align*}
\sup_{\|f\|=1} | \langle Jf, f\rangle|
&=\sup_{\|f\|=1} \left| \left[a+b\lambda_{\textrm{ref}}(1-\alpha)\right] \langle f, f\rangle - c \langle \nabla^2 U f, f\rangle \right|\\
&=\sup_{\|f\|=1}
\max \left\{ \left[a+b\lambda_{\textrm{ref}}(1-\alpha)\right] \langle f, f\rangle - c \langle \nabla^2 U f, f\rangle, c \langle \nabla^2 U f, f\rangle-\left[a+b\lambda_{\textrm{ref}}(1-\alpha)\right] \langle f, f\rangle \right\}\\
&\leq\sup_{\|f\|=1} \max\left\{ \left(a+(1-\alpha)\lambda_{\textrm{ref}} b\right) \|f\| - c m \|f\|, cM \|f\| - \left(a+(1-\alpha)\lambda_{\textrm{ref}} b\right) \|f\| \right\}\\
&= \max\left\{ a+(1-\alpha)\lambda_{\textrm{ref}} b - c m, cM - a-(1-\alpha)\lambda_{\textrm{ref}} b \right\}.
\end{align*}
Therefore, if we can find $a,b,c>0$, such that $b<\sqrt{4 a \alpha c}/(1+\alpha)$ and
\begin{align*}
4\left[ a(1-\alpha) \lambda_{\textrm{ref}} - bM\right]b &> \max\{cM-a-(1-\alpha)\lambda_{\textrm{ref}} b, a+(1-\alpha)b\lambda_{\textrm{ref}} - cm\}^2,
\end{align*}
then the RHS of \eqref{eq:quadraticformtooptimize} is a positive definite quadratic form. In principle this can be used to optimise the convergence rates among
norms of the form \eqref{eq:newnorm}.
\end{remark}
We take a slightly different approach.
Our goal is to show that for every $h$, we have $\frac{\mathrm{d} }{\mathrm{d} t} \langle\!\langle h, h\rangle\!\rangle\le -\mu \langle\!\langle h, h\rangle\!\rangle$, or equivalently
\[-\frac{\mathrm{d} }{\mathrm{d} t} \langle\!\langle h, h\rangle\!\rangle-\mu \langle\!\langle h, h\rangle\!\rangle\ge 0,\]
After rearrangement, we obtain that
\begin{align}
\nonumber&-\frac{\mathrm{d} }{\mathrm{d} t} \langle\!\langle h, h\rangle\!\rangle-\mu \langle\!\langle h, h\rangle\!\rangle\\
\nonumber&= a(2\lambda_{\textrm{ref}}(1-\alpha)-\mu)\|A h\|^2 -2(a+(1-\alpha)b\lambda_{\textrm{ref}}-\mu b)\inner{Ah}{Ch} +(2 b-c \mu) \|Ch\|^2\\
\nonumber&\quad -2b \inner{\nabla^2 U Ah} {Ah} + 2c\inner{\nabla^2 U Ah}{Ch}\\
&\quad + 2 a\lambda_{\textrm{ref}}\alpha \inner{S Ah}{Ah}+ 2 c\lambda \inner{S C h}{Ch}- 2(1+\alpha) b\lambda_{\textrm{ref}} \inner{S Ah}{Ch}.\label{eq:dperdthh}
\end{align}
We will use the following two lemmas.
\begin{lemma}\label{lem:VXWPZQ}
If $V, W, Z, A \in \mathbb{R}^{2 \times 2}$ are symmetric matrices such that $0 \preceq A$, $-Z \preceq A$, $A \preceq V+mW$ and $A \preceq V+MW$, then $\mathrm{Tr}(VX + WP + ZQ) \ge 0$ for all symmetric matrices $X,P,Q$ such that $0 \preceq Q \preceq X$ and $mX \preceq P \preceq MX$.
\end{lemma}
\begin{proof}[Proof of Lemma~\ref{lem:VXWPZQ}]
First, suppose that $M=m$. By the assumptions we have $P=mX$, $A\succeq 0$, and $Z+A\succeq 0$. Note that if $S,T$ are symmetric positive semidefinite matrices, then $\mathrm{Tr}(ST) \ge 0$. Using this fact, it follows that
\begin{align*} \mathrm{Tr}(VX + WP + ZQ)&=\mathrm{Tr}((V+mW)X + ZQ)\ge \mathrm{Tr}(AX + (Z+A)Q-AQ)\\
&\ge \mathrm{Tr}(A(X-Q))\ge 0.
\end{align*}
Now suppose that $M>m$.
Let
$$A_1 = Z+A,\quad A_2 = A, \quad A_3 = \frac{1}{M-m}(V+MW-A), \quad A_4 = \frac{1}{M-m}(V+mW-A).$$
Then $A_1, A_2, A_3, A_4 \succeq 0$, and
$$V = A_2 - m A_3 + M A_4, \quad W = A_3 - A_4, \quad Z = A_1 - A_2.$$
So
\begin{align*}
VX + WP + ZQ &= (A_2 - m A_3 + M A_4) X + (A_3 - A_4) P + (A_1-A_2) Q \\
&= A_1 Q + A_2 (X-Q) + A_3 (P-mX) + A_4 (MX-P).
\end{align*}
Using positive definiteness of both terms in the matrix products, we have
$$\mathrm{Tr}(A_1 Q), \mathrm{Tr}(A_2(X-Q)), \mathrm{Tr}(A_3(P-mX)), \mathrm{Tr}(A_4(MX-P)) \ge 0,$$
and therefore
\begin{equation*}
\mathrm{Tr}(V X + W P + Z Q) \ge 0.\qedhere
\end{equation*}
\end{proof}
Now we are ready to complete the proof of Theorem~\ref{thm:hypoco}.
\begin{proof}[Proof of Theorem \ref{thm:hypoco}
Let $a:=1$, and
\begin{align*}
b&:=\frac{1+\alpha-\alpha \l(\frac{m}{M+m}\r)^{3/4}+\frac{3}{4} \frac{\alpha m}{M+m}}{2\sqrt{M+m}},\\
c&:=\frac{1+\alpha- \frac{\alpha }{2}\l(\frac{m}{M+m}\r)^{1/2} }{M+m},\\
X&:=\l(\begin{matrix}\|A h\|^2 & \inner{Ah}{Ch}\\\inner{Ah}{Ch} & \|C h\|^2 \end{matrix}\r),\\
P&:=\l(\begin{matrix}\inner{\nabla^2 U(x) A h}{A h} & \inner{\nabla^2 U(x) Ah}{Ch}\\\inner{\nabla^2 U(x) Ah}{Ch} & \inner{\nabla^2 U(x) C h}{Ch} \end{matrix}\r),\\
Q&:=\l(\begin{matrix}\inner{S A h}{A h} & \inner{S Ah}{Ch}\\\inner{S Ah}{Ch} & \inner{S C h}{Ch} \end{matrix}\r),\\
V&:=\l(\begin{matrix} 2a(1-\alpha)\lambda-a\mu & -a -(1-\alpha)b \lambda+b\mu \\
-a -(1-\alpha)b \lambda+b\mu & 2 b-c \mu \end{matrix}\r),\\
W&:=\l(\begin{matrix}-2b & c\\c & 0 \end{matrix}\r),\\
Z&:=\l(\begin{matrix}2a \alpha \lambda & -(1+\alpha)b \lambda \\ -(1+\alpha)b \lambda & 2 c\lambda \end{matrix}\r),\\
A&:= \l(\begin{matrix}\frac{4(-3+2m-2M)(-1+\alpha)}{3\sqrt{m+M}(1+\alpha)} & -\frac{(-3+2m-2M)(-1+\alpha)}{3(m+M)}\\-\frac{(-3+2m-2M)(-1+\alpha)}{3(m+M)} & -\frac{(-3+2m-2M)(-1+\alpha)(1+\alpha)}{3(m+M)^{3/2}} \end{matrix}\r).
\end{align*}
It is easy to check that $b^2<a c$. Using the assumption that $m I \preceq \nabla^2 U\preceq M I$, we have $m X\preceq P \preceq M X$. Moreover, using the fact that $0\preceq S \preceq I$, we have $0\preceq Q\preceq X$. Based on \eqref{eq:dperdthh} and the above definitions it follows that
\begin{equation}
-\frac{\mathrm{d} }{\mathrm{d} t} \langle\!\langle h, h\rangle\!\rangle-\mu \langle\!\langle h, h\rangle\!\rangle=\mathrm{Tr}(VX + WP + ZQ).
\end{equation}
One can check, for example using Mathematica, that for every $M\ge 1, 0\le \alpha<1$, the inequalities $0 \preceq A$, $-Z \preceq A$, $A \preceq V+mW$ and $A \preceq V+MW$ hold for $A$ defined as above. Therefore \eqref{eq:hypoco_rate} follows from Lemma \ref{lem:VXWPZQ}, and by Gr\"onwall's lemma, this implies that $\langle\!\langle P^t f, P^t f\rangle\!\rangle\le \exp(-\mu t)\langle\!\langle f,f\rangle\!\rangle$.
To show our $L^2$ bound, we study the reversed process. Denote the variant of the scalar product $\langle\!\langle\cdot,\cdot \rangle\!\rangle$ when $b$ is replaced by $-b$ by $\langle\!\langle \cdot,\cdot\rangle\!\rangle'$, i.e.
\begin{equation}\label{eq:newnormp}
\langle\!\langle h, h\rangle\!\rangle':=a\|\nabla_v h\|^2 + 2b \langle \nabla_x h, \nabla_v h\rangle + c\|\nabla_x h\|^2.
\end{equation}
Then by repeating the same arguments as above with $v$ replaced by $-v$ everywhere, one can show that we have
\begin{equation}\frac{\mathrm{d} }{\mathrm{d} t} \langle\!\langle (P^*)^t f, (P^*)^t f\rangle\!\rangle'\le -\mu \langle\!\langle (P^*)^t f, (P^*)^t f\rangle\!\rangle',
\label{eq:hypoco_rate}
\end{equation}
and hence $\langle\!\langle (P^*)^t f, (P^*)^t f\rangle\!\rangle'\le \exp(-\mu t)\langle\!\langle f,f\rangle\!\rangle'$. Similarly to the previous proofs, we can show that $\langle\!\langle\cdot, \cdot\rangle\!\rangle$ and $\langle\!\langle\cdot, \cdot\rangle\!\rangle'$ are equivalent up to the same constant factor $C$, and
\[\langle\!\langle (P^t)^* P^t f, (P^t)^* P^t f\rangle\!\rangle\le C^2\exp(-2\mu t) \langle\!\langle f,f\rangle\!\rangle.\]
In addition, there exist constants $C_1, C_2>0$ such that
$\langle\!\langle f, f\rangle\!\rangle \leq C_1 \|\nabla f\|^2$ and $\|f\|^2\leq C_2 \langle\!\langle f, f\rangle\!\rangle$.
Thus, letting $f$ be $k$-Lipschitz we have
\begin{align*}
\|(P^t)^* P^t f\|^2
&\leq C_2 \langle\!\langle (P^t)^* P^t f, (P^t)^* P^t f\rangle\!\rangle'\\
&\leq C_2\exp(-2\mu t)\langle\!\langle f,f\rangle\!\rangle'\\
&\leq C_1 C_2 \exp(-2\mu t) \|\nabla f\|^2 \leq C_1 C_2 k^2 \exp(-2\mu t).
\end{align*}
Choose $t$ such that $C_1C_2 k \mathrm{e}^{-2\mu t}=: 1-\kappa<1$ and define the self-adjoint operator $Q = (P^t)^* P^t$.
Iterating the above we have for $n\geq 1$ that
\begin{align*}
\|Q^n f\|^2
&\leq C_1 C_2 (1-\kappa)^{2n} k^2 =: C(f) (1-\kappa)^{2n}.
\end{align*}
The rest is similar to the proof of Proposition~2.8 from \citet{hairer2014spectral}.
Let $f$ be $k$-Lipschitz, and without loss of generality also assume that $\|f\|=1$.
Let $\nu_f$ be the spectral measure corresponding to the self-adjoint operator $Q$ applied to the function $f$. In particular, since $\|f\|=1$, $\nu_f$ is a probability measure.
Then
\begin{align*}
\|Q^n f\|^2
&= \int_{-1}^1 t^{2n} \nu_f (\mathrm{d} t)\\
&= \int_{-1}^1 t^{2n (2n+2m)/(2n+2m)} \nu_f (\mathrm{d} t)\\
&\leq \left(\int_{-1}^1 t^{2(n+m)}\nu_f (\mathrm{d} t)\right)^{\frac{2n}{2(n+m)}}\\
&= \left(\|Q^{n+m} f\|^2 \right)^{\frac{2n}{2(n+m)}}\\
&\leq \left[C(f) (1-\kappa)^{2(n+m)} \right]^{\frac{2n}{2(n+m)}}\\
&\leq C(f)^{\frac{2n}{2(n+m)}} (1-\kappa)^{2n},
\end{align*}
and letting $k\to \infty$ we get for any $k$-Lipschitz $f$
\begin{align*}
\|Q^n f\|^2
&\leq \|f\|^2(1-\kappa)^{2n},
\end{align*}
noticing that the upper bound is independent of the Lipschitz constant.
Since Lipschitz functions are dense we conclude.
\end{proof}
\begin{remark}
Given any $\lambda_{\textrm{ref}}>0,\mu>0$, the contraction $\frac{\mathrm{d} }{\mathrm{d} t} \langle\!\langle h, h\rangle\!\rangle\le -\mu \langle\!\langle h, h\rangle\!\rangle$ holds as long as there exists coefficients $a,b,c\in \mathcal{R}$ and a $2\times 2$ real valued symmetric matrix $A$ such that $a>0$, $c>0$, $b^2<a c$ and
$0 \preceq A$, $-Z \preceq A$, $A \preceq V+mW$ and $A \preceq V+MW$ (with $V$ and $W$ defined as above). Note that as in the proof of Theorem~\ref{thm:Wasserstein}, due to the non-linearity of the constraints we did not manage to find an analytical expression for the largest possible $\mu$ for a given $\lambda$, and the largest possible $\mu$ for any $\lambda$. However, we believe that the choice of $\lambda$ and $\mu$ as given here is close to optimal in most of the parameter range $0<m\le M<\infty$, $0\le \alpha<1$.
\end{remark}
\section*{Acknowledgements}
The authors would like to thank Peter Holderrieth for a careful reading of the manuscript and his invaluable suggestions and Philippe Gagnon for his insightful comments on the manuscript. G.D.\ would like to thank Gabriel Stoltz for many useful discussions.
This material is based upon work supported in part by the U.S. Army Research Laboratory and the U. S. Army Research Office, and by the U.K. Ministry of Defence (MoD) and the U.K. Engineering and Physical Research Council (EPSRC) under grant number EP/R013616/1 and by the EPSRC EP/R034710/1. A part of this research was done while A. Doucet, G. Deligiannidis and D. Paulin were hosted by the Institute for Mathematical Sciences in Singapore.
\nocite{*}
\bibliographystyle{plainnat}
|
\section{Introduction}
Conformal higher spin gauge theories attract considerable attention because despite being non-unitary they give tractable examples of interacting Lagrangian theories extending conformal gravity and involving higher spin fields. The simplest conformal higher spin (CHS) fields are totally symmetric tensor fields subject to first order gauge transformations. These are also known as Fradkin-Tseytlin fields and were originally proposed in~\cite{Fradkin:1985am} in 4 dimensions and generalized in~\cite{Segal:2002gd} to higher even dimensions. Interacting theory for these fields was proposed much later~\cite{Segal:2002gd,Tseytlin:2002gz} and elaborated further in~\cite{Bekaert:2010ky} (see also~\cite{Bonezzi:2017mwr} for a recent discussion).
CHS fields are intimately related to Fronsdal fields in anti de Sitter space (AdS) of one extra dimension in the context of AdS/CFT correspondence. More specifically, CHS fields in $n$-dimensions can be regarded as leading boundary values of the Fronsdal fields in AdS space of ${n+1}$ dimensions. In so doing the CHS Lagrangian arises as the holographic Weyl anomaly \cite{Metsaev:2008fs,Metsaev:2009ym} (see also~\cite{Bekaert:2012vt,Bekaert:2013zya} for the gauge-covariant analysis at the level of equations of motion).
Wave operators for CHS fields (CHS operators) are of order $d-4+2s$ and were conjectured~\cite{Tseytlin:2013jya} to factorize into a product of 2nd order operators when written over the constant curvature background (see also~\cite{Metsaev:2007fq,Metsaev:2007rw,Joung:2012qy} for the relevant earlier contributions). It turns out that in 4 dimensions the factors have the mass terms identical to those of partially-massless fields~\cite{Deser:1983mm,Deser:2001pe} whose order of gauge transformation (known as ``depth'') ranges from $1$ to $s$. In higher dimensions in addition to partially-massless-like wave operators one also finds among the factors the wave operators of certain massive fields~\cite{Metsaev:2014iwa}. The existence of the factorized form allows to express the partition function of CHS fields in terms of the known partition functions of the partially-massless fields~\cite{Tseytlin:2013jya} (see also~\cite{Beccaria:2015vaa,Beccaria:2016tqy}). The manifestly factorized form of the CHS operators was given in~\cite{Nutma:2014pua}.
For a scalar ($s=0$) conformal field the factorization amounts to the familiar factorization of higher-order conformal operators known in the context of conformal geometry as GJMS operators~\cite{Paneitz:1983,Fradkin:1981jc,GJMS}. These also have a natural generalization~\cite{Gover:2005mn} to tractor fields on conformally-Einstein manifolds (for an introduction to tractors see e.g.~\cite{Eastwood,BEG,Cap:2002aj}). The similarity with GJMS operators suggests that tractor technique can be useful in studying factorization of CHS wave operators as well.
As far as the structure of CHS operators on constant curvature background is concerned in addition to manifestly-factorized form~\cite{Nutma:2014pua} it is also worth mentioning a suggestive ordinary derivative Lagrangian formulations proposed in~\cite{Metsaev:2014iwa}. Furthermore, the manifestly-conformal formulation of CHS equations was proposed in~\cite{Bekaert:2012vt,Bekaert:2013zya} by employing a version of the ambient space technique. In our study of the CHS operators we use this formulation as a starting point.
In this work we are concerned with more general class of the CHS wave operators for totally symmetric fields, which includes those with gauge invariance of arbitrary order $t\leq s$. These operators are of order $d-2+2(s-t)$ and can be considered as those of the boundary values of depth-$t$ partially-massless fields in $AdS_{n+1}$. We relate these CHS wave operators to GJMS ones by constructing a certain embedding of tensor fields into tractors. More specifically, we realize tractors using the parent-formulation approach~\cite{Barnich:2006pc,Bekaert:2009fg}, where the usual ambient-space construction is employed to describe the tangent space rather than the spacetime, and demonstrate that CHS fields can be embedded in such a way that GJMS operators coincides with CHS wave operators. As a byproduct we clarify and elaborate in some details on the parent approach description of tractors originally put forward in~\cite{Grigoriev:2011gp}.
Constructing the factorized form requires introducing the so-called scale tractor~\cite{Gover:2005mn} and a new ingredient -- operators $B_k$, $k=1,2,\ldots$. For $k=1$ this was already employed in the literature in the context of tractor description of low spin~\cite{Gover:2008sw} and higher-spin~\cite{Grigoriev:2011gp} fields in constant curvature spaces. In the formulation developped in this paper CHS operator simply takes a manifestly factorized form $B_\ell \ldots B_1$, $\ell=\frac{d-4}{2}+s$. This representation turns out to be useful in analyzing gauge invariance. In particular, we show that $B_t$ is the wave operator of depth-$t$ partially-massless field in traceless gauge.
The paper is organized as follows: in Section~\bref{sec:ambient} we recall the ambient space formulation of tractors and GJMS operators. There we also introduce parent formulation technique which allows to work with fields defined on the conformal space rather than ambient one but still benefit from the manifest realizations of $o(n,2)$-symmetry. In section~\bref{sec:CHS}
we introduce CHS fields, propose a new manifestly $o(n,2)$-invariant formulation for them, and construct the manifestly factorized factorized form. Technical details are relegated to Appendices.
\section{Ambient space and tractors in the parent approach}
\label{sec:ambient}
\subsection{Ambient space}
In this work we are concerned with conformal gauge fields defined on the conformally-flat spaces. The conformal symmetry can be seen as originating from the conformal isometries which, at the infinitesimal level, are given by conformal Killing vector fields. These form $o(n,2)$ algebra, where $n$ is the space-time dimension.
Conformally invariant equations can be described~\cite{Dirac:1936fq} in a manifestly $o(n,2)$-invariant way by employing the ambient space construction (which in turn originates from that of Klein). An ambient space is a pseudo-Eucledean space $\mathbb{R}^{n,2}$ equipped with the metric $\eta_{AB}$ of signature $(-,+,+...,+,-)$. In what follows we use ambient coordinates $(X^+, X^a, X^-)$ so the metric has the form
\begin{equation}
\eta_{AB}dX^AdX^B = 2dX^+dX^- + \eta_{ab}dX^adX^b
\end{equation}
Where $\eta_{ab}$ is Minkowski metric $(n-1, 1)$. The cone is a zero locus $\lbrace X^2 = 0 \rbrace \backslash \lbrace 0\rbrace$. In this picture, the $n$-dimensional conformal space $M$ is the projectivization of the cone (projective cone in what follows), i.e. the quotient space of $X^2=0$ modulo the equivalence relation $X^A \sim \lambda X^A, \lambda \in \mathbb{R} \backslash \lbrace 0 \rbrace $. The quotient is equipped with the conformal structure and with a natural action of $o(n,2)$ as well as the entire conformal group (in what follows we restrict to infinitesimal analysis and hence concentrate on the conformal algebra). The action comes from the standard $o(n,2)$-action on the ambient space.
To pick a representative of the equivalence class of metrics on $M$ one can embed $M$ as a submanifold of $X^2=0$ such that each ray intersects $M$ once and only once. The metric (which is conformally flat by construction) is then obtained by pulling back the ambient metric to $M$. Scalar densities of conformal weight $w$ on $M$ can be described ambiently as:
\begin{subequations}
\label{amb-scalar}
\begin{gather}
(X\cdot\frac{\partial}{\partial X} - w)\Phi = 0\\
\Phi \sim \Phi + X^2\chi\,,
\end{gather}
\end{subequations}
in terms of the ambient space functions $\Phi=\Phi(X)$. Here and in what follows $\cdot$ denotes $o(n,2)$-invariant contraction of indices, e.g. $Z\cdot W=\eta_{AB}Z^A W^B$ and $X^2=X\cdot X$. Because both the constraint and the equivalence relation are manifestly $o(n,2)$-invariant the conformal algebra act on the space defined by~\eqref{amb-scalar} (the same applies to the conformal group).
In a similar fashion we can consider tensor fields on the ambient space satisfying an analog of~\eqref{amb-scalar}. If we restrict ourselves to totally symmetric fields it is convenient to work in terms of generating functions defined on the cotangent bundle over the ambient space
\begin{equation}
\Phi(X,P)=\sum_{i=0} \Phi^{A_1 \ldots A_i}P_{A_1}\ldots P_{A_i}\,.
\end{equation}
Here $P_A$ are coordinates on the fibers and we assume $\Phi$ to be polynomial in $P$. It follows that the space defined by
\begin{subequations}
\label{ambient_tractors}
\begin{gather}
(X\cdot \frac{\partial}{\partial X} - w)\Phi(X,P) = 0 \label{at1}\\
\Phi(X) \sim \Phi(X) + X^2\chi(X,P) \label{at2}
\end{gather}
\end{subequations}
is that of totally symmetric tractors of weight $w$, which we denote by $\mathcal{E}^{\bullet}[w]$. Indeed, equation \eqref{at2} implies that such tensors are actually defined on the cone $X^2=0$ while \eqref{at1} says that these are actually defined on the projective cone. It is clear that $\mathcal{E}^{\bullet}[w]$ is equipped with a natural action of $o(n,2)$ induced by that on the cotangent bundle over the ambient space.
\subsection{GJMS operators}
Representing tractor fields through~\eqref{ambient_tractors} is useful in studying conformally invariant differential operators defined on tractors. In particular, in these terms it is easy to define so-called GJMS-operators. These were originally proposed~\cite{GJMS} for scalar densities and later extended to generic tractors~\cite{Gover:2002ay,Gover:2005mn} using the Feffermann-Graham construction which reduces to the above ambient space approach in the conformally flat case.
If tractors are described through~\eqref{ambient_tractors} the GJMS-operators are simply powers of the ambient Laplacian
\begin{equation}
P^{2\ell} \coloneqq \Box_X^{\ell},\quad P^{2\ell}: \mathcal{E}^{\bullet}[\ell-\frac{n}{2}] \mapsto \mathcal{E}^{\bullet}[-\ell-\frac{n}{2}]\,, \qquad \Box_X\coloneqq \dl{X}\cdot\dl{X}
\end{equation}
This operator is well defined on equivalence classes~\eqref{at2} provided the weights are as above. To see this, it is instructive to exploit that the following 3 operators
\begin{equation}
H \coloneqq X\cdot\dl{X} +\frac{n+2}{2}\,, \qquad E\coloneqq -\half X^2\,, \qquad F \coloneqq \half \dl{X}\cdot \dl{X}\,.
\end{equation}
define a representation of $sl(2)$-algebra on the ambient space functions. For trivial $\Phi(X)=X^2 \chi$ with $\chi$ of weight $w = \ell-2-\frac{n}{2}$ one finds
\begin{equation}
\Box_X^{\ell}(X^2\chi) = X^2\Box_X^{\ell}\chi + 4\ell\Box_X^{\ell-1}(w_{\chi} + \frac{n}{2} - \ell +2)\chi = X^2\Box_X^{\ell}\chi
\end{equation}
because $w_{\chi} = w_{\Phi} - 2 = \ell - 2 - \frac{n}{2}$.
To write explicit formulas for GJMS operators one chooses local coordinates $x^\mu$ on $M$ and particular metric $g_{\mu\nu}$ in the conformal class. The first non-trivial example of GJMS operators is Yamabe operator $\bar\nabla^2 - \frac{(n- 2)}{4(n-1)}R$ defined on scalar densities of weight $1- \frac{n}{2}$, where $\bar\nabla$ is the Levi-Civita connection determined by $g$.
If one identifies $M$ with flat Minkowski space then GJMS operators are just $\bar\nabla^{2\ell}$ acting on scalar densities of weight $w = \ell - \frac{n}{2}$.
There is a dual description of GJMS operators
using the following system \cite{GJMS} (see also~\cite{Bekaert:2013zya}):
\begin{equation}
\label{dualGJMS}
\begin{gathered}
\Box_X \Phi(X) = 0\,,\\
(X\cdot\dl{X}-\ell+\frac{n}{2})\Phi(X)=0\,,\\
\Phi(X) \sim \Phi(X) + (X^2)^{\ell}\chi\,.
\end{gathered}
\end{equation}
Here $\chi$ also satisfies analogous constraints but with $-\ell+\frac{n}{2}$ replaced with $\ell+\frac{n}{2}$.
It follows this system is equivalent to~\eqref{ambient_tractors} supplemented by $\Box_X^{\ell} \Phi=0$ and with $w=\ell-\frac{n}{2}$. It is a remarkable property of~\eqref{dualGJMS} that the first two equations considered in the vicinity of the hyperboloid
$X^2=-1$ in the ambient space describe the scalar field of mass $w(w+n)$. Moreover, this gives a systematic way~\cite{Bekaert:2012vt,Bekaert:2013zya} to describe boundary values of the (A)dS field on the hyperboloid.
\subsection{Thomas-D operator}
An important object well-defined on the equivalence classes \ref{ambient_tractors} is Thomas D-operator $D^A:\mathcal{E}^{B..C}[w] \mapsto \mathcal{E}^{AB..C}[w-1]$. Here we denote by $\mathcal{E}^{B..C}[w]$ tensor fields of arbitrary symmetry and rank of weight $w$ satisfying:
\begin{subequations}
\label{ambient_tractors2}
\begin{gather}
(X\cdot \frac{\partial}{\partial X} - w)\Phi = 0\\
\Phi \sim \Phi + X^2\,\chi \
\end{gather}
\end{subequations}
Of course, \eqref{ambient_tractors2} contains \eqref{ambient_tractors}.
By slightly abusing notations, Thomas D-operator can be defined as follows
\begin{equation}
\label{Thomas}
D_A \Phi = (2(X \cdot \dl{X} + \frac{n}{2})\frac{\partial}{\partial X^A} - X_A\Box_X)\Phi\,,
\end{equation}
where $\Phi$ is subject to~\ref{ambient_tractors2}.
It has the following properties:
\begin{itemize}
\item $[D_A, D_B] = 0$
\item $D_A D^A = X^2\Box_X^2$\,.
\end{itemize}
Note that the above definition and properties are to be modified in the conformally non-flat case.
There is a useful relation between GJMS operators and Thomas-D observed in~\cite{Gover:2002ay}:
\begin{equation}
\label{D->P1}
D_{A_1}...D_{A_{\ell}}\Phi = (-1)^kX_{A_1}...X_{A_{\ell}}P^{2\ell}\Phi\,, \qquad \Phi \in \mathcal{E}^{\bullet}[\ell-\frac{n}{2}]
\end{equation}
A version of this relation is going to be very useful in what follows.
\subsection{Tractors in parent formulation}\label{sec:tractors-parent}
The naive ambient construction sketched above is very useful in describing fields on a conformally-flat background. However, it operates in terms of the equivalence classes of fields on the ambient space rather then fields explicitly defined on the conformal manifold. Moreover, ambient description is not directly applicable to a conformally-flat space which is equivalent to the projective cone only locally.
Although these issues can be resolved by employing a full-scale Fefferman-Graham construction~\cite{FG} there is a relatively simple and concise alternative. It is based on
reformulating the system~\eqref{ambient_tractors} in the so-called parent form~\cite{Barnich:2006pc,Bekaert:2009fg} in which the ambient construction is realized in the formal version of the ambient space rather than in the space-time. Moreover, this approach has proved useful in describing gauge fields including CHS fields and hence provides a framework to study the structure of the CHS wave operators.
The parent counterpart of the system~\eqref{ambient_tractors} is constructed by first introducing the formal version of the ambient space with coordinates $Y^A$, where one considers totally symmetric tensor fields. As before we work with tensors in terms of the generating function
\begin{equation}
\Phi=\Phi(Y,P)\,.
\end{equation}
The dependence on $Y$ is assumed formal, i.e. as functions on the formal ambient space one takes
polynomials in $P$ with coefficients in formal series in $Y$. Given a nonvanishing ambient vector $V_0^A$
one defines a ``twisted'' realization~\cite{Barnich:2006pc,Bekaert:2009fg} of $o(n,2)$ on the space of the above functions in $Y,P$:
\begin{equation}
\label{twisted}
\rho(\alpha)\Phi=\alpha^A_B\left[P_A\dl{P_B}-(Y^B+V_0^B)\dl{Y^A}\right]\,, \qquad \alpha \in o(n,2)\,.
\end{equation}
Then, the conformal structure on $M$ can be encoded in terms of the vector bundle $\mathcal{V}$ over $M$ whose fiber is a copy of the flat ambient space. More precisely, the bundle is equipped with the fiber-wise pseudo-Euclidean metric $\eta$, nonvanishing section $V\in \Gamma (\mathcal{V})$ such that $\eta(V,V)=0$, and an $o(d,2)$ connection $dx^\mu\omega_\mu$ compatible with $\eta$ and such that $\nabla V$ has maximal rank (i.e. seen as a fiber-wise map $TM\to\mathcal{V}(M)$ it has a vanishing kernel). As we are now interested in flat conformal structures we restrict ourselves to flat connections, i.e. $d\omega+\omega\omega=0$.
Given this data one can consider an associated bundle whose fibre is the above space of ``functions'' in $Y,P$, where $o(n,2)$ acts according to the twisted representation~\eqref{twisted}, where at a given point $V_0$ is just $V$ at this point. Moreover we assume that the local frame is chosen in such a way that $V^A$ is constant. For instance, the associated covariant derivative of a section $\Phi$ is given explicitly by
\begin{equation}
\pmb{\nabla}_\mu \Phi=\dl{x^\mu}\Phi+\omega^A_{\mu B}\left[P_A\dl{P_B}-(Y^B+V^B)\dl{Y^A}\right]\Phi\,.
\end{equation}
Here and in what follows we use the local identification of sections of this bundle with functions in $x,P,Y$. In particular, for $Y$-indpendent sections the usual covarinat derivative is reproduced. In the conformally flat case we are concerned with $\pmb{\nabla}$ is flat, i.e. $\pmb{\nabla}^2=0$.
The standard choice of the local frame of $\mathcal{V}(M)$ is such that:
\begin{equation}
\label{bulky_equation}
\begin{gathered}
\omega_{\mu\enspace B}^{\enspace A} =
\begin{pmatrix}
0 & -e_{\mu b} & 0 \\
J_{\mu}^{\enspace a} & \omega_{\mu \enspace b}^{\enspace a} & e_{\mu}^{\enspace a} \\
0 & -J_{\mu b} & 0
\end{pmatrix}
\,,
\qquad
\quad
V^A = \begin{pmatrix}
V^-\\
V^a\\
V^+
\end{pmatrix}
= \begin{pmatrix}
0\\
0\\
1
\end{pmatrix}\,,
\end{gathered}
\end{equation}
where $e^A=\nabla V^A=\omega^A{}_B V^B$ and $J_{\mu}^{\enspace a} = e^{a\nu}J_{\mu\nu}$ with $J_{\mu\nu}$ being the Schouten tensor of the metric $g_{\mu\nu}=e^A_\mu e^B_\nu \eta_{AB}$. In dimensions $n\geq 3$ the Schouten tensor is defined in terms of the Ricci tensor and the scalar curvature as $J_{\mu\nu} = \frac{1}{n-2}\left(R_{\mu\nu} - \frac{R}{2(n-1)}g_{\mu\nu} \right)$. We denote $J$ to be the trace of $J_{\mu\nu}$.
Although the construction is frame-independent we assume for simplicity that the frame is chosen as above.
Note that it's also possible to allow for non-constant $V^A$ at the price of extra terms in the covariant derivative,
see \cite{Barnich:2006pc, Bekaert:2009fg} for more details.
With the above prerequisites we are ready to give a local version of the ambient definition of tractors. More precisely, consider the following system:
\begin{equation}
\label{parentsystem}
\begin{gathered}
\pmb{\nabla}_{\mu}\Phi = 0\,,\\
((Y+V)\cdot\frac{\partial}{\partial Y} - w)\Phi = 0\,,\\
\Phi \sim \Phi + (Y+V)^2 \chi\,,
\end{gathered}
\end{equation}
where $\chi=\chi(x,P,Y)$ satisfies analogous system with $w$ replaced by $w-2$. The space of equivalence classes of sections determined by~\eqref{parentsystem} is precisely the space of tractors of weight $w$ that we keep denoting $\mathcal{E}^\bullet[w]$.
The easiest way to see this is to observe that any $Y$-independent section $\Phi_0(x,P)$ admits a unique (up to an equivalence) lift to $\Phi(x,P,Y)$ satisfying~\eqref{parentsystem}. Indeed, taking into account the explicit form
\eqref{bulky_equation} one finds that the 1-st and the 2-nd equations are first-order in $y^a,Y^+$ and hence solution exists and can be constructed recursively.~\footnote{More precisely, to give a rigorous argument it is useful to introducing Grassmann-odd ghost variables associated to all the constraints in~\eqref{parentsystem} (note that those associated to the components of the covariant derivative are precisely the basis differentials $dx^\mu$) and to employ the homological perturbation theory. In this way it is manifest that consistency conditions are fulfilled at each step.} The arbitrariness in the solution at each step is in adding a function in $Y^-$ but this arbitrariness is taken into account by the equivalence relation in the last line of~\eqref{parentsystem}.
The GJMS operators can also be defined in terms of~\eqref{parentsystem} as
\begin{equation}
P^{2\ell} \Phi(x,P,Y)= \Box^{\ell} \, \Phi(x,P,Y)\,, \qquad
\Phi\in \mathcal{E}^{\bullet}[\ell-\frac{n}{2}]\,,\qquad
\Box\coloneqq (\dl{Y}\cdot\dl{Y})\,.
\end{equation}
To recover the previous definition of GJMS operators in terms of $Y$-independent fields let $\Phi(x,P,Y)$
be a unique (up to equivalence) solution to~\eqref{parentsystem} with $w=\ell-\frac{n}{2}$ such that $\Phi(x,P,Y)|_{Y=0}=\Phi_0(x,P)$. Then
\begin{equation}
P^{2\ell} \Phi_0(x,P)= \left(\Box^{\ell} \, \Phi(x,P,Y)\right)\Big|_{Y=0}\,.
\end{equation}
The parent analog of the system \eqref{dualGJMS} describing a scalar of weight $\ell-\frac{n}{2}$ subject to GJMS equation of order $2\ell$ reads as
\begin{equation}
\label{dual_GJMS}
\begin{gathered}
\pmb{\nabla}_{\mu}\Phi = 0\,,\\
((Y+V)\dl{Y} - \ell + \frac{n}{2})\Phi = 0\,,\\
\Box\Phi = 0\\
\Phi \sim \Phi + (Y+V)^{2\ell} \chi\,,
\end{gathered}
\end{equation}
where $(Y+V)^{2k} \coloneqq ((Y+V)^2)^k$. In this system the GJMS equation arises at order $(Y^-)^{\ell-1}$ of $(\Box \Phi)\big|_{Y^+= Y^a = 0} = 0$.
There exists a partially gauge-fixed version of \eqref{parentsystem}. More precisely, for a given $\tilde\Phi$ satisfying \eqref{parentsystem} one can construct an equivalent representative $\Phi$ of the same equivalence class such that
\begin{equation}
\label{pgf-tract}
\Box \Phi= (Y+V)^{2(\ell-1)}\alpha
\end{equation}
for some $\alpha(x,P,Y)$. It is easy to see that the new representative is defined up to a restricted equivalence relation:
\begin{equation}
\label{new-equiv}
\Phi\sim \Phi+(Y+V)^{2\ell}\chi\,.
\end{equation}
More formally, \eqref{parentsystem} is equivalent to the partially gauge-fixed system consisting of the first two equations of
\eqref{parentsystem} supplemented by \eqref{pgf-tract} and the new equivalence relation~\eqref{new-equiv}. In terms of
\eqref{parentsystem},\eqref{pgf-tract} the GJMS equation can be written as $\alpha=(Y+V)^{2}\beta$ for some $\beta$.
This makes manifest that GJMS equation shows up at order $\ell-1$ in the expansion of $\Box\Phi$ in powers of $(Y+V)^2$.
In terms of tractors described through~\eqref{parentsystem}, Thomas D-operator can be defined as follows:
\begin{equation}
\label{TomasD}
\mathcal{D}_A \Phi(x,P,Y)\coloneqq
\left(2((Y+V)\cdot \frac{\partial}{\partial Y} + \frac{n}{2})\frac{\partial}{\partial Y^A} - (Y+V)_A\frac{}{}\Box\right) \Phi(x,P,Y)\,,
\end{equation}
where $\Phi\in \mathcal{E}^{\bullet}[w]$. It is easy to check that $P\cdot\mathcal{D}$ is a well-defined map $\mathcal{E}^{\bullet}[w]\to \mathcal{E}^{\bullet}[w-1]$. The explicit relation between $\mathcal{D}_A$ and conventional Thomas-D operator defined on tractors reads as
\begin{equation}
\label{ThomasD}
D_A \Phi_0(x,P)=\left(\mathcal{D}_A \Phi(x,P,Y)\right)|_{Y=0}\,,
\end{equation}
where as usually $\Phi$ denotes a lift of $\Phi_0$ satisfying~\eqref{parentsystem}. In particular, this gives an alternative systematic way to derive the explicit expression for Thomas-D operator. The details of the derivation as well as the explicit expression for $D_A$ are given in Appendix~\bref{sec:components}. Note that a version of this derivation was in~\cite{Grigoriev:2011gp}.
The relation~\eqref{D->P1} between Thomas-D and GJMS operators take the form:
\begin{equation}
\label{parentGJMS}
\mathcal{D}_{A_1}... \mathcal{D}_{A_{\ell}}\Phi(x, P, Y) = (-1)^{\ell}(Y+V)_{A_1}...(Y+V)_{A_{\ell}}\Box^\ell\Phi(x, P, Y)\,.
\end{equation}
\subsection{Scale tractor and factorization of GJMS operators}
It is known that the GJMS operator factorises on a conformally-Einstein (in particular, a conformally-flat) background~\cite{Gover:2005mn}. This factorization is easy to arrive at explicitly by making use of additional
important ingredient, the so-called scale tractor. By definition, a scale tractor is a nowhere vanishing weight $0$ and rank $1$ tractor tensor $I^A$ which is parallel,
i.e. satisfying covariant-constancy condition:
\begin{equation}
\label{parallel}
\nabla_\mu I^A=\partial_{\mu} I^A + \omega_{\mu\enspace B}^{\enspace A}I^{B}=0\,.
\end{equation}
In the conformally-Einstein case~\eqref{parallel} implies that there exists a scalar density $\sigma$ such that $I^A=\frac{1}{n}D^A\sigma$, where $D^A$ is a Thomas-D derivative determined by~\eqref{ThomasD}. In terms of components
\begin{equation}
I^A = \begin{pmatrix}
\sigma \\
\bar\nabla^a \sigma\\
-\frac{1}{n}(J + \bar\nabla^2)\sigma
\end{pmatrix}
\end{equation}
where $\bar\nabla_a\coloneqq e_a^\mu \bar\nabla_\mu$. Here $\bar\nabla_\mu$ denotes Levi-Civita covariant derivative determined by the metric $g_{\mu\nu}=\nabla_\mu V^A \nabla_\nu V^B \eta_{AB}$.
If $I^A=\frac{1}{n}D^A\sigma$ and $\nabla_\mu I^A=0$ then $\sigma$ determines a constant curvature representative of the conformal equivalence class of the metric. More precisely, $g^{c}_{\mu\nu}\coloneqq \sigma^{-2} g_{\mu\nu}$ is constant curvature \cite{LeBrun85}. In the case where $g_{\mu\nu}$ is constant curvature from the very beginning one may simply take $\sigma=1$ so that
\begin{equation}
I^A = \begin{pmatrix}
1\\
0\\
-\frac{J}{n}
\end{pmatrix}\,.
\end{equation}
Note that $V^AI_A =1$ with this choice. Although all the general constructions of this and the next section are valid for general parallel $I^A$ in all the explicit examples we always assume that the metric is constant curvature from the very beginning ($AdS$ for definiteness) and $\sigma=1$.
Because $I$ commutes with $\nabla_\mu$, it also commutes with Thomas D: $[I_A,D_B] = 0$. It is easy to see that $I^AP_A$ satisfies \eqref{parentsystem} with $w=0$ and hence $I_A$ also commutes with $\mathcal{D}_B$.
It means that the formula for GJMS operator \eqref{parentGJMS} can be written slightly differently~\cite{Gover:2005mn}:
\begin{equation}
\begin{gathered}
\label{useless}
P^{2\ell} \Phi(x,P) = (-1)^{\ell}I^{A_1}{D}_{A_1}...I^{A_{
\ell}}{D}_{A_{\ell}} \Phi(x,P)\,,\qquad
\end{gathered}
\end{equation}
Note that $\Phi(x,P)$ is to be understood as a tractor field of weight $w=\ell-\frac{n}{2}$.
Note also that $(I\cdot\mathcal{D})\Phi(x, P)$ has weight $w-1$.
Using~\eqref{useless} one can obtain explicit formulas for GJMS operators. Indeed, for a tractor field $\Phi_w(x,P)$ of generic weight $w$ one has
\begin{equation}
\label{ID}
(I\cdot \mathcal{D})\Phi_w(x,P)=-\lbrace \nabla^2 +\frac{2J}{n}(n+w-1)(w)\rbrace \Phi_w(x,P)\,.
\end{equation}
Here and in what follows $\nabla^2=g^{\mu\nu}\nabla_\mu\nabla_\nu$ where by slight abuse of notations we denote by $\nabla_\mu$ the covariant derivative $\hat{\nabla}_\mu$ extended to tensors with values in symmetric tractors (identified with polynomials in $P_A$). For instance, for $A_\mu=A_\mu(x,P)$ one has $\nabla_\mu A_\nu=\hat{\nabla}_\mu A_\nu-\Gamma_{\mu\nu}^\rho A_\rho$, where $\Gamma_{\mu\nu}^\rho$ are coefficients of the Levi-Civita connection.
Combining this with~\eqref{useless} one gets:
\begin{equation}
\label{GJMS2}
\prod_{i=0}^{\ell-1} \lbrace \nabla^2 +\frac{2J}{n}(\ell+\frac{n}{2}-i-1)(\ell - \frac{n}{2}-i)\rbrace \Phi(x,P) = P^{2\ell}\Phi(x,P)\,.
\end{equation}
According to \eqref{ID} the order of terms entering \eqref{GJMS2} is the following: the term with $i=0$ acts first, then $i=1$, and so on.
\section{Conformal higher spin fields}
\label{sec:CHS}
\
There exist conformally invariant equations for totally symmetric tensor fields proposed originally by Fradkin and Tseytlin~\cite{Fradkin:1985am} in 4 dimensions and extended to all even dimensions in~\cite{Segal:2002gd}. The equations are Lagrangian and possess gauge invariance. In Minskowski space the equations and gauge transformations have the following structure:
\begin{equation}
(\d^2)^{\frac{n-4}{2}+s}\phi_{a_1\ldots a_s}+\ldots=0\,, \qquad \delta\phi_{a_1\ldots a_s}=\d_{(a_1}\epsilon_{a_2\ldots a_s)}+\eta_{(a_1a_s}\omega_{a_3\ldots a_s)}\,,
\end{equation}
where $\ldots$ denote terms proportional to $\d^{a_1}\phi_{a_1\ldots a_s}$ and $\phi^{a_1}_{a_1\ldots a_s}$. The algebraic gauge symmetry with parameter $\omega$ can be employed to set $\phi^{a_1}_{a_1\ldots a_s}=0$. In what follows we always assume this gauge condition.
CHS fields can be seen as a linearization of the nonlinear CHS theory proposed in~\cite{Tseytlin:2002gz,Segal:2002gd} (see also~\cite{Bekaert:2010ky,Bonezzi:2017mwr}) about a Minkowski space vacuum. The respective action functional arises as an induced action for the scalar field in the higher-spin background. It is remarkable, that the consistency of the scalar in higher-spin background naturally determines nonlinear gauge transformations for the background higher-spin fields, giving the gauge symmetries of the induced action~\cite{Segal:2002gd}.
In addition to CHS fields whose gauge transformations are of first order in derivatives there are conformal gauge fields whose gauge transformations are of order $t\leq s$ in derivatives, which we refer to as depth-$t$ CHS fields. The law-spin fields of this type were already in~\cite{Drew:1980yk,Barut:1982nj,Deser:1983mm} while higher spin ones were described much later~\cite{Erdmenger:1997wy,Vasiliev:2009ck,Bekaert:2013zya,Beccaria:2015vaa}. In what follows we use the term ``CHS fields'' for the entire family including Fradkin-Tseytlin fields as well as their depth-$t$ generalizations. Note that higher-depth CHS fields in $n$-dimensions are somewhat similar to so-called partially-massless fields~\cite{Deser:1983mm,Deser:2001pe} and in fact can be understood as boundary values of the partially-massless fields on $AdS_{n+1}$~\cite{Bekaert:2013zya}. It is also worth mentioning that in addition to totally symmetric fields there exist mixed symmetry CHS fields~\cite{Vasiliev:2009ck,Chekmenev:2015kzf} which remain beyond the scope of the present work.
\subsection{Manifestly $o(n, 2)$-invariant description of CHS fields}
Our goal is to study the structure of CHS equations and, in particular, the factorization of the CHS wave operators. As a starting point of our analysis we use the ambient space formulation of the CHS equations, which is available in the literature. More precisely, CHS equations of motion can be encoded in the following system~\cite{Bekaert:2012vt,Bekaert:2013zya} (see also~\cite{Chekmenev:2015kzf}):
\begin{equation}
\label{FT}
\begin{gathered}
\pmb{\nabla} \Phi = 0 , \qquad ((Y+V)\cdot\frac{\partial}{\partial Y}+\frac{n}{2}-\ell)\Phi = 0, \qquad (Y+V)\cdot\frac{\partial}{\partial P}\Phi = 0\,,\\
\frac{\partial}{\partial Y}\cdot\frac{\partial}{\partial P}\Phi = 0\,,\qquad \frac{\partial}{\partial P}\cdot\frac{\partial}{\partial P}\Phi = 0,\qquad
P\cdot\frac{\partial}{\partial P}\Phi = s\Phi\,,\\
\end{gathered}
\end{equation}
\begin{equation}
\label{Box}
\Box\Phi = 0\,,
\end{equation}
\begin{equation}
\label{equiv-l}
\Phi \sim \Phi+(Y+V)^{2\ell}\chi\,,
\end{equation}
which is formulated in the setting of Section~\bref{sec:tractors-parent} and where $\ell=\frac{n}{2}+s-t-1$. The equivalence relation is to be understood as follows: two configurations are equivalent if their difference can be represented as $(Y+V)^{2\ell}\chi$ for some $\chi$. Note that it's not difficult to extract explicitly the conditions $\chi$ ought to satisfy: these also have the form~\eqref{FT},\eqref{Box} but with $\ell$ replaced with $-\ell$.
In terms of the above representation the CHS gauge transformations can be written as follows:
\begin{equation}
\label{gt-PdY}
\begin{gathered}
\delta \Phi = (P\cdot\frac{\partial}{\partial Y})^t \epsilon\,,
\end{gathered}
\end{equation}
where $\epsilon=\epsilon(x,P,Y)$ is subject to the analogous system with $\ell$ replaced by $\ell+t$
and $s$ with $s-t$.
For $t=1$ system~\eqref{FT}-\eqref{gt-PdY} describes the usual CHS fields while for $t=2,\ldots,s$ their higher-depth generalizations. The system is manifestly $o(n, 2)$ invariant.
Strictly speaking system \eqref{FT}-\eqref{gt-PdY} is not equivalent to CHS equations of motion and gauge symmetries. More precisely, in addition to CHS equations
it also encodes conformal gauge conditions which, however, can be consistently removed in one or another way, giving an equivalent formulation of CHS fields (see~\cite{Bekaert:2012vt,Bekaert:2013zya,Chekmenev:2015kzf} for more details).
To see how exactly CHS equations are encoded in the above system let us consider~\eqref{FT} supplemented with
\begin{equation}
\label{Box-lrel}
\Box \Phi=(Y+V)^{2(\ell-1)}\alpha
\end{equation}
in place of~\eqref{Box}. By rephrasing the analysis of~\cite{Bekaert:2012vt,Bekaert:2013zya,Chekmenev:2015kzf} in the present terms one finds that any traceless $\phi(x,p)$ can still be lifted to $\Phi(x,P,Y)$ satisfying not only \eqref{FT} but also \eqref{Box-lrel}. Moreover, the condition that
\begin{equation}
\alpha =(Y+V)^2\beta
\end{equation}
for some $\beta$ encodes CHS equation and conformal gauge conditions. Here $\alpha$ is understood as a function of $\phi$ and its $x^\mu$-derivatives obtained by solving \eqref{FT},\eqref{Box-lrel}.
It turns out that just CHS equations can be written as
\begin{equation}
\label{CHS}
\alpha\big|_{Y=P^\pm=0}=0\,.
\end{equation}
The equation $\alpha\big|_{Y=0}=0$ is conformal by construction.
To check the conformal invariance of \eqref{CHS} one observes that the equation sitting at $P^{\pm}=0$ is of order $2\ell$ while the equations in different components are of higher order and hence their conformal transformations can't compensate the transformations of \eqref{CHS} so that \eqref{CHS} should be conformally invariant~\cite{Chekmenev:2015kzf}. Their gauge invariance can also be shown on general grounds following~\cite{Bekaert:2009fg,Bekaert:2013zya,Chekmenev:2015kzf}. In any case in this work we give an independent proof of the gauge invariance.
It is also worth mentioning that if one drops the equivalence relation in~\eqref{FT}, takes $V$ such that $V^2=-1$, and takes as $\Phi$ a field on $AdS_{n+1}$ rather than $n$-dimensional conformal space, the above system
is precisely the one from~\cite{Alkalaev:2011zv} see also~\cite{Barnich:2006pc,Alkalaev:2009vm}, which describes partially-massless fields on $AdS_{n+1}$. In this form it is manifest that depth-$t$ FT fields in $n$-dimensions are boundary values of the partially-massless fields on $AdS_{n+1}$.
There exist a ``dual'' system that also describes CHS fields but where
the harmonicity condition is replaced by $(\frac{\partial}{\partial Y}\cdot\frac{\partial}{\partial Y})^{\ell}\Phi(x, P, Y) = 0$:
\begin{equation}
\label{FT1}
\begin{gathered}
\pmb{\nabla} \Phi = 0 , \qquad ((Y+V)\cdot\frac{\partial}{\partial Y}+\frac{n}{2}-\ell)\Phi = 0, \qquad (Y+V)\cdot\frac{\partial}{\partial P}\Phi = 0\,,\\
\frac{\partial}{\partial Y}\cdot\frac{\partial}{\partial P}\Phi = 0\,,\qquad
\frac{\partial}{\partial P}\cdot\frac{\partial}{\partial P}\Phi = 0\,,\qquad
P\cdot\frac{\partial}{\partial P}\Phi = s\Phi\,,
\end{gathered}
\end{equation}
\begin{equation}
\label{Box-l}
\Box^{\ell}\Phi = 0 \,,
\end{equation}
\begin{equation}
\label{equivalence-2}
\Phi \sim \Phi+(Y+V)^2\alpha\,.
\end{equation}
Note that \eqref{FT1} is identical to \eqref{FT}.
In what follow it is useful to introduce a natural map $L^{-1}$ that sends elements of $\mathcal{E}^\bullet[w]$ (in particular solutions to~\eqref{FT1}) to tensor fields on $M$. In terms of generating function $\Phi(x,P,Y)$ it is given by
\begin{equation}
L^{-1}\Phi=\Phi\big|_{Y=P^{\pm}=0}\,.
\end{equation}
We have the following:
\begin{prop}
\label{prop:lift}
For all $\ell>0$ or all non-integer $\ell$ any $\phi(x, p)$ satisfying $\dl{p}\cdot \dl{p}\phi=0$ and $p\cdot\dl{p}\phi=s\phi$ can be lifted to $\Phi(x,P,Y)$ satisfying~\eqref{FT1} and such that $\phi=L^{-1}\Phi$.
The lift is unique if one takes into account the equivalence relation~\eqref{equivalence-2}.
\end{prop}
The statement can be inferred from the analysis of~\cite{Bekaert:2012vt,Bekaert:2013zya,Chekmenev:2015kzf}. Some details of the proof are also given in Appendix~\bref{sec:lift-obst}.
\begin{prop}
Let $\Phi(x,P,Y)$ be a lift of $\phi(x,p)$ as described in Proposition~\bref{prop:lift} with $\ell=\frac{n}{2}+s-t-1$, then the operator defined by
\begin{equation}
\label{CHS-operator}
A_{s,t}\phi=L^{-1}\left(\Box^{\ell}\Phi\right)
\end{equation}
is well-defined on equivalence classes~\eqref{equivalence-2} and coincides with CHS wave operator.
\end{prop}
\begin{proof}
Operator $A_{s,t}$ is clearly well-defined on equivalence classes $\Phi \sim \Phi+(Y+V)^2\alpha$ and hence determines an operator of derivative order $2\ell$ on totally symmetric traceless tensor fields.
To explicitly relate $A_{s,t}$ to CHS operator we first note that for any $\Phi_0$ satisfying~\eqref{FT1} one can find an equivalent element $\Phi=\Phi_0+(Y+V)^2\ldots$ such that
\begin{equation}
\label{Box-ell-rel}
\Box \Phi=(Y+V)^{2(\ell-1)}\alpha\,.
\end{equation}
Indeed, this is achived by taking
\begin{equation}
\begin{gathered}
\Phi(x, P, Y) = \sum_{k=0}^{l-1}(Y+V)^{2k}\Phi_{k}\\
\Phi_{k} = -\frac{1}{4k(\ell - k)}\Box\Phi_{k-1}\,.
\end{gathered}
\end{equation}
Note that the residual equivalence relation is precisely~\eqref{equiv-l}. In this way we have found that the system~\eqref{FT},\eqref{equiv-l},\eqref{Box-lrel} results from~\eqref{FT1},\eqref{equivalence-2} by partially taking into account the equivalence relation~\eqref{equivalence-2} and hence these systems are equivalent.
Finally, applying $\Box^{\ell-1}$ to both sides of~\eqref{Box-ell-rel} and setting to zero $Y,P^+,P^-$
one finds
\begin{equation}
L^{-1}(\Box^\ell \Phi)=r\,L^{-1}(\alpha)
\end{equation}
where $r$ is a non-vanishing coefficient
and hence equation $L^{-1}(\Box^\ell \Phi)=0$
is equivalent to CHS equations~\eqref{CHS}.
\end{proof}
Let us comment on the relation between the above description of CHS fields and tractors.
In contrast to tractor fields, which can be seen as certain tensor fields on the $n+2$-dimensional ambient space restricted to $n$-dimensional submanifold, CHS fields (at the off-shell level, i.e. before imposing CHS equations of motion) are tensor fields (more precisely, tensor densities) in $n$-dimensions on which the action of gauge transformations and conformal transformations is defined. Equations~\eqref{FT1},\eqref{equivalence-2} can be seen as a mean to embedd off-shell CHS fields as a subspace of tractor fields in such a way that the GJMS operator produces the CHS equations of motion through~\eqref{CHS-operator}.
\subsection{Modified system and factorization of CHS operators}
It turns out that it is useful to employ a certain modification of the system~\eqref{FT1},\eqref{equivalence-2}. In particular, the gauge invariance of the CHS equations is conveniently analysed in the modified formulation. Consider the following system:
\begin{equation}
\label{FT2}
\begin{gathered}
\pmb{\nabla} \Phi = 0 , \qquad ((Y+V)\cdot\frac{\partial}{\partial Y} -w)\Phi = 0, \qquad (Y+V)\cdot\frac{\partial}{\partial P}\Phi = 0\\
\mathcal{D}\cdot\frac{\partial}{\partial P}\Phi = 0\,, \qquad
\frac{\partial}{\partial P}\cdot\frac{\partial}{\partial P}\Phi = 0,\qquad
P\cdot\frac{\partial}{\partial P}\Phi = s\Phi\,,
\end{gathered}
\end{equation}
\begin{equation}
\label{equiv-22}
\Phi \sim \Phi + (Y+V)^2\chi\,,
\end{equation}
where $\chi(x, P, Y)$ satisfies \eqref{FT2} with $w$ replaced by $w-2$.
We denote by $S[s, w]$ the space of equivalence classes determined by this system.
The gauge transformations can be now defined as $\delta\Phi = (P\cdot\mathcal{D})^{t}\epsilon$, where $\epsilon$ also satisfies \eqref{FT2},\eqref{equiv-22} with $w,s$ replaced with $w+t,s-t$
so that $(P\cdot\mathcal{D})^{t}$ determines a well-defined map $S[s-t, w+t]\to S[s, w], \quad w = s-t-1$
It is easy to check that for $w\neq -\frac{n}{2}$ equations \eqref{FT2} are equivalent to \eqref{FT1}. Indeed,
assuming all the other constraints in \eqref{FT2} but $\mathcal{D}\cdot\dl{P}$ satisfied one finds:
\begin{equation}
\mathcal{D}\cdot\dl{P}=(2w+n)\dl{P}\cdot\dl{Y}\,.
\end{equation}
In particular, for $w=\ell-\frac{n}{2}$ if $\Phi(x,P,Y)$ satisfying \eqref{FT2} denotes a lift of $\phi(x,p)$ ($\dl{p}\cdot \dl{p})\phi=0$ then \eqref{CHS-operator} defines CHS wave operator.
For the special value $w=-\frac{n}{2}$ any $\phi(x,p)$ satisfying $\dl{p}\cdot\dl{p}\phi=0$ can be lifted to $\Phi(x,P,Y)$ satisfying \eqref{FT2}. However,
in contrast to \eqref{FT1} the lift is not unique even if one takes into account the equivalence relation~\eqref{equiv-22}. The uniqueness can be restored by introducing the following additional equivalence relation:
\begin{equation}
\label{extra-equiv}
\Phi \sim \Phi+ ((Y+V)\cdot P)\beta\,.
\end{equation}
Assuming that definiton of $S[s,w]$ in the case of $w=-\frac{n}{2}$ also involves~\eqref{extra-equiv} we conclude that for $w \geq -\frac{n}{2}$, the space $S[s,w]$ is one-to-one with that of totally symmetric traceless tensor fields on $M$.
The apparent disadvantage of defining CHS operator through~\eqref{CHS-operator} is that it requires extracting particular components of $\Box^\ell\Phi$.
This can be cured as follows: pick a particular metric in the conformal class and consider the following operator defined on $S[s, s-k-1]$:
\begin{equation}
B_k \coloneqq I \cdot\mathcal{D} - \frac{1}{k}(P\cdot\mathcal{D})(I\cdot \dl{P})
\end{equation}
Indeed, it is well defined on the equivalence classes~\eqref{equiv-22} for these values of parameters. If $\Phi(x, P, Y) \in S[s, s-k-1]$, then $(B_k\Phi)(x, P, Y) \in S[s, s-k-2]$, so that $B_k$ determines a well defined map $S[s, s-k-1]\to S[s, s-k-2]$. Note that in contrast to the operators employed above $B_k$ is not $o(n,2)$-invariant because it contains
the scale tractor that breaks $o(n,2)$-symmetry. With our choice of $I^A$ the residual symmetry is just $(A)dS$-isometries. Note that for $k=1$ this operator was employed in~\cite{Grigoriev:2011gp}, while for $k=1$ and $s=1,2$ it was in~\cite{Gover:2008sw}.
One can also define powers of $B_k$ as follows: $B^{\ell}_{k}\coloneqq B_{k+l-1}\circ ... \circ B_{k}$ which act on $\Phi\in S[s,s-k-1]$ according to $B^{\ell}_{k}: \Phi \mapsto (B_k^{\ell}\Phi) \in S[s,s-k-1-{\ell}]$. We have the following:
\begin{prop}
\label{proposition}
Let $\Phi(x, P, Y)\in S[s, w]$ with $\quad w = \ell - \frac{n}{2}, \ell = \frac{n-2}{2}+s-t$ be a lift of $\phi(x,p)$, i.e. $L^{-1}\Phi=\phi$. Then equation
$B^{\ell}_{t}\Phi = 0$ is equivalent to $A_{s,t}\phi=0$ and hence is a CHS equation formulated in terms of $S[s, w]$.
\end{prop}
It is clear that $B_t^{\ell}$ is well defined on $S[s, w]$ in this case. The proof that it indeed determines CHS wave operator is relegated to Appendix~\bref{sec:Blt-proof}.
\def\frac{n}{2}{\frac{n}{2}}
It follows from the identification of $S[s,s-k-1]$ with totally symmetric tensor densities that $B_k$ defines an
operator on tensor densities. More precisely, if $L_{s,w}$ denote a map that sends $\phi(x,p)$ to $\Phi(x,P,Y)$ satisfying \eqref{FT2} and $L^{-1}$ the map defined by $L^{-1}\Phi=\Phi\big|_{Y^A,P^\pm=0}$ then for $\phi=\phi(x,p)$ of rank $s$ and weight $s-k-1$
\begin{equation}
\bar B_k\phi=L^{-1}B_k L_{s,s-k-1}\phi
\end{equation}
is a second order differential operator on tensor densities. Representing $A_{s,t}$ as
\begin{multline}
A_{s,t}\phi=(-1)^{\ell}L^{-1}B_k^\ell L_{s,s-t-1}
=\\=
(-1)^{\ell}(L^{-1}B_{t+\ell-1} L_{s,s-t-\ell})( L^{-1}\ldots L_{s,s-t-2})(L^{-1} B_t L_{s,s-t-1})
\end{multline}
one finds
\begin{equation}
\label{A-exp}
A_{s,t}\phi=(-1)^{\ell}\bar B_{t+\ell-1}\ldots \bar B_{t} \phi\,.
\end{equation}
In other words we have arrived at the manifestly factorized form of the CHS wave operator. Note that although all the above arguments apply to generic conformally-flat background metric $g_{\mu\nu}$ operators $\bar B_k$ in general depend on scale $\sigma$ so that only on constant curvature spaces where one can take $\sigma=1$ this gives a genuine factorization of CHS wave operator into natural second-order operators.
\subsection{Explicit form of the factors}
Now we are ready to give an explicit component expressions for the CHS operators and the operators $\bar B_k$ in terms of tensor densities on $M$. Leaving the detailed computations for the Appendix~\bref{sec:lift-obst} we get
\begin{multline}
\bar B_k \phi(x,p)=L^{-1}B_k L_{s,s-k-1}\phi(x, p) = \\ =-
\lbrace \bar\nabla^2 +\frac{2J}{n}(-s + (n+s-k-2)(s-k-1) ) -
\frac{n+2s-4}{k(n+2s -k-3)}(p\cdot\bar\nabla)(\frac{\partial}{\partial p}\cdot\bar\nabla)
+\\+
\frac{1}{k(n+2s-k-3)}p^2(\frac{\partial}{\partial p}\cdot\bar\nabla)^2 \rbrace \phi(x,p)\,.
\end{multline}
It follows from the structure of the mass-like term in the above operator that it coincides with the one of partially-massless field of spin $s$ and depth $k-1$. More precisely, $B_k$ explicitly coincides with the partially massless operator provided both are written in the gauge where $(\dl{p}\cdot\bar\nabla) \phi(x,p)=0$.
Now the expression~\eqref{A-exp} for the CHS wave operator takes the form:
\begin{equation}
\label{factor_example}
\begin{gathered}
A_{s,t}\phi(x,p)=\prod_{i=1}^{\frac{n-4}{2}+s -t +1}\lbrace \bar\nabla^2 +\frac{2J}{n}(-s + (n+s-t-i-1)(s-t-i) ) -\\- \frac{n+2s-4}{(t+i-1)(n+2s-t-i-2)}(p\cdot\bar\nabla)(\frac{\partial}{\partial p}\cdot\bar\nabla)+\\+\frac{1}{(t+i-1)(n+2s-t-i-2)}p^2(\frac{\partial}{\partial p}\cdot\bar\nabla)^2 \rbrace \phi(x,p)\,,
\end{gathered}
\end{equation}
where the operator with $i=0$ acts first, then the operator with $i=1$ and etc. In the special case $w=s-2$ we get the result obtained by Nutma and Taronna \cite{Nutma:2014pua}:
\begin{equation}
\label{Nutma_Taronna}
\begin{gathered}
A_{s,1}\phi(x, p)=\prod_{i=1}^{i = \frac{n-4}{2} +s}\lbrace \bar\nabla^2 +\frac{2J}{n}(-s + (n+s-i-2)(s-i-1) ) -\\- \frac{n+2s-4}{i(n+2s-i-3)}(p\cdot\bar\nabla)(\frac{\partial}{\partial p}\cdot\bar\nabla)+\frac{1}{i(n+2s-i-3)}p^2(\frac{\partial}{\partial p}\cdot\bar\nabla)^2 \rbrace \phi(x, p)\,.
\end{gathered}
\end{equation}
Note that although the formulas coincide in our derivation we assumed that $\phi$ is traceless.
\subsection{Gauge invariance}
Now we are ready to analyse gauge invariance of the CHS equations using its factorized representation in terms of $B_k$.
\begin{prop}
For any $\epsilon \in S[s-t,s-1]$
\begin{equation}
B^\ell_t (P\cdot \mathcal{D})^t\epsilon=0\,.
\end{equation}
\end{prop}
As $S[s-t,s-1]$ is one-to-one with traceless tensor densities on $M$, $(P\cdot \mathcal{D})^t$ determines a gauge symmetry of the CHS equations.
\begin{proof}
Observe that
\begin{equation}
\label{gauge_transf}
\begin{gathered}
B_t(P\cdot \mathcal{D})^t\epsilon(x, P, Y) = (I\cdot\mathcal{D} - \frac{1}{t}(P\cdot\mathcal{D}) (I\cdot\dl{P})(P\cdot\mathcal{D})^t\epsilon(x, P, Y) =\\
=-\frac{1}{t}(P\cdot\mathcal{D})^{t+1}(I\cdot\dl{P})\epsilon(x, P, Y)
\end{gathered}\,.
\end{equation}
Applying $B_{t+1}$ we get $(P\cdot\mathcal{D})^{t+2}(I\cdot\frac{\partial}{\partial P})^2\epsilon(x, P, Y)$ and so on. Because $\epsilon$ is of rank $s-t$, this procedure gives zero after $s-t+1$ iterations. CHS operator $B^{\ell}_t,\,\, \ell= \frac{n-2}{2} +s -t$ contains at least $s-t+1$ factors (for $n\geq 4$) and hence $(P\cdot\mathcal{D})^t\epsilon(x, P, Y)$ is in the kernel of $B^\ell_t$.
To make sure that the gauge transformation $\delta \Phi=(P\cdot\mathcal{D})^t\epsilon$ indeed coincides with the standard gauge transformation for CHS fields, one can check that $L^{-1}(P\cdot\mathcal{D})^t\Phi$ is traceless by construction and the leading term is proportional to $(p\cdot\bar\nabla)^t$ as it should be for CHS fields of this type.
\end{proof}
The above technique can be also used to study gauge invariance of $B_k$. Suppose that we subject the gauge parameter $\epsilon\in S[s-k,s-1]$ to the extra condition $I\cdot\dl{P}\epsilon = 0$ which encodes that $L^{-1}\epsilon$
satisfies $\dl p \cdot \bar\nabla(L^{-1}\epsilon)=0$. Then $B_t(P\cdot\mathcal{D})^t\epsilon = (P\cdot\mathcal{D})^{t+1}I\cdot\frac{\partial}{\partial P}\epsilon = 0$. This gives an additional argument that for $\Phi\in S[s,s-k-1]$ equation $B_k\Phi=0$ is a partially gauge-fixed version of the equations of motion of the partially-massless field of spin $s$ and depth $k$. Examples can be found in Appendix \ref{sec:gauge_pm}.
\subsection{CHS equations in terms of tractors}
Although we have described CHS fields and found the factorized form of the CHS equations by employing the parent formalism, it turns out that the resulting formulas can be written in terms of usual tractor fields. To see this let us find the conditions satisfied by
$\Phi_0(x,P)=(L_{s,w}\phi(x,p))\big|_{Y=0}$. These can be easily obtained by setting $Y^A=0$ in \eqref{FT2}, giving
\begin{equation}
\label{embed_tractors}
\begin{gathered}
V\cdot\dl P \Phi_0 = 0
\qquad P\cdot\dl{P}\Phi_0=s\Phi_0\,, \qquad \dl{P}\cdot \dl{P}\Phi_0=0\,,
\\
\dl{P}\cdot{D} \,\,\Phi_0 = 0\,,
\end{gathered}
\end{equation}
where $D_A$ is the usual Thomas-D derivative whose definition and explicit expression are given in respectively~\eqref{ThomasD} and \eqref{Thomas-D-comp}.
It turns out, that for relevant values of $w$ one can avoid constructing $\Phi=L_{s,w}\phi$ and obtain $\Phi_0$ directly by solving \eqref{embed_tractors} with boundary condition $\Phi_0\big|_{P^\pm=0}=\phi(x,p)$. More precisely,
with our choice of $V$ the first equation implies $\dl{P^{+}} \Phi_0(x, P)=0$ and hence the last one uniquely fixes the $P^-$ dependence. Indeed, for $w \neq -\frac{n}{2}$ the last equation is equivalent to $\left( \dl{P^-}(w-1+ n+s - P^-\dl {P^-}) + \dl{p} \cdot \bar\nabla) \right) \Phi(x,P) = 0$ and hence always admits a unique solution if $w$ corresponds to a CHS field.
It follows \eqref{embed_tractors} determines a particular embedding of totally symmetric traceless tensor densities into traceless symmetric tractors. Identifying off-shell CHS fields with the weight $w=s-t-1$ tractor fields satisfying~\eqref{embed_tractors} the CHS equations of motion take the following form:
\begin{equation}
\mathbb{B}_t^{\ell}\Phi(x, P) = 0,\qquad l = \frac{n-2}{2} +s -t\,,
\end{equation}
where $\mathbb{B}_t = I\cdot D - \frac{1}{t}(P\cdot D)(I\cdot {\dl P})$ is just a tractor version of $B_t$, i.e. where $\mathcal{D}$ is replaced with the conventional Thomas-D operator, and $\mathbb{B}_t^{\ell}\coloneqq \mathbb{B}_{t+\ell -1}\circ\mathbb{B}_{t+\ell -2}...\circ\mathbb{B}_{t}$. The operator $\mathbb{B}_t$ is well-defined on \eqref{embed_tractors} for $w=s-t-1$, i.e $\mathbb{B}_t\Phi_0$ satisfies \eqref{embed_tractors} for $w=s-t-2$. The gauge transformations are given by
\begin{equation}
\delta \Phi_0 = (P\cdot D)^t\epsilon
\end{equation}
where $\epsilon$ is a weight-$s-1$ and rank-$s-t$ tractor field satisfying~\eqref{embed_tractors}
with $s$ replaced by $s-t$.
Let us consider as a simple example Maxwell field in 4 dimensions, i.e. $n=4,s=t=1$.
Solving~\eqref{embed_tractors} with the initial condition $\Phi_0\big|_{P^\pm=0}=\phi^a p_a$ and analogous equations for the gauge parameter $\epsilon$ (these are satisfied trivially) gives:
\begin{equation}
\Phi_0^A =
\begin{pmatrix}
0\\
\phi^a\\
-\frac{1}{2}\bar\nabla_a\phi^a
\end{pmatrix}, \qquad D^A\epsilon = \begin{pmatrix}
0\\
2\bar\nabla^a\epsilon\\
-\bar\nabla^2\epsilon
\end{pmatrix}\,.
\end{equation}
Restricting for simplicity to the flat case and computing $\mathbb{B}_1\Phi_0$ gives
\begin{equation}
\label{Bmaxwell}
(\mathbb{B}_1\Phi_0)^A =
\begin{pmatrix}
0\\
\partial^2 \phi^a - \partial^a\partial_b\phi^b\\
0
\end{pmatrix}
\end{equation}
so that in accord with our general statements we indeed get just Maxwell equations.
If instead of $\mathbb{B}_1\Phi_0$ we consider $\nabla^2 \Phi_0$ (which is precisely $(\Box \Phi(x,P,Y))|_{Y=0}$) we arrive at~\cite{Eastwood:1985eh}
\begin{equation}
\nabla^2 \Phi =
\begin{pmatrix}
0\\
\partial^2 \phi^a - \partial^a\partial_b\phi^b\\
-\frac{1}{2}\partial^2\partial_a\phi^a
\end{pmatrix}
\,.
\end{equation}
The second slot still contains Maxwell equations themselves, while the last one is the conformal gauge \cite{Eastwood:1985eh} also known as Eastwood-Singer gauge. Of course, this is the same gauge as encoded in the system~\eqref{FT},\eqref{Box},\eqref{equiv-l} for $s=1,t=1$ on top of the Maxwell equations. For general CHS fields one gets higher-spin analogs of this gauge.
\section*{Acknowledgements}
We are grateful to K.~Alkalaev, A.~Chekmenev, R.~Metsaev for useful discussions.
M.G. also wishes to thank N.~Boulanger, X.~Bekaert, and especially A.~Waldron.
This work was supported by the Russian Science Foundation grant 18-72-10123.
|
\section{Introduction}
A rigorous definition of the lagrangean field theory, termed \textbf{the non-linear $\sigma$-model}, describing simple geometric dynamics of topologically charged material points, loops and higher ($p-$)dimensional extended objects ($p$-branes) in an ambient smooth metric manifold $\,(\mathscr{M},{\rm g})\,$ (termed \textbf{the target space}), modelled by smooth\footnote{In fact, in the most general scenario, $\,{\mathbb{X}}\,$ is only patchwise smooth over $\,\Omega_{p+1}\,$ and maps the latter into a disjoint sum of manifolds. This is the setting of a $\sigma$-model with defects, {\it cp} \Rcite{Runkel:2008gr}.} embeddings
\begin{eqnarray}\nn
{\mathbb{X}}\ :\ \Omega_{p+1}\longrightarrow\mathscr{M}
\end{eqnarray}
of the $(p+1)$-dimensional closed \textbf{worldvolume} $\,\Omega_{p+1},\ \partial\Omega_{p+1}=\emptyset\,$ of the extended object in the target space and affected by external fields coupling to their mass (the metric $\,{\rm g}\,$ on $\,\mathscr{M}$) and charge (a closed $(p+2)$-form gauge field $\,\underset{\tx{\ciut{(p+2)}}}{\chi},\ {\mathsf d}\underset{\tx{\ciut{(p+2)}}}{\chi}=0$), has long been known to call for a geometrisation of the de Rham $(p+2)$-cocycle $\,\underset{\tx{\ciut{(p+2)}}}{\chi}\,$ whose effect on the dynamics is captured by the so-called topological Wess--Zumino (WZ) term
\begin{eqnarray}\nn
S^{(p+1)}_{\rm WZ}[{\mathbb{X}}]=\int_{\Omega_{p+1}}\,{\mathbb{X}}^*{\mathsf d}^{-1}\underset{\tx{\ciut{(p+2)}}}{\chi}\,.
\end{eqnarray}
Here, $\,{\mathsf d}^{-1}\underset{\tx{\ciut{(p+2)}}}{\chi}\,$ is locally (that is for $\,{\mathbb{X}}\,$ that factors through a contractible open subset $\,\mathcal{O}\subset\mathscr{M}$) representable by a primitive of $\,\underset{\tx{\ciut{(p+2)}}}{\chi}$.\ The geometrisation takes the form of a principal ${\mathbb{C}}^\x$-bundle, an abelian bundle gerbe, or a $p$-gerbe with connection $\,\underset{\tx{\ciut{(p)}}}{\mathcal{G}}$,\ respectively. These are conveniently described by classes in the real Deligne--Beilinson cohomology of the target space in degree $p+2$ whose representatives are local differential-form data of a trivialisation of $\,\underset{\tx{\ciut{(p+2)}}}{\chi}\,$ over some ({\it e.g.}, good) open cover of $\,\mathscr{M}$.\ Accordingly, (the topological term of) \textbf{the Dirac--Feynman amplitude}
\begin{eqnarray}\nn
\cA_{\rm DF(WZ)}[{\mathbb{X}}]:={\rm e}^{{\mathsf i}\,S^{(p+1)}_{\rm WZ}[{\mathbb{X}}]}
\end{eqnarray}
acquires the precise interpretation of \textbf{the $(p+1)$-surface holonomy} of the $p$-gerbe $\,\underset{\tx{\ciut{(p)}}}{\mathcal{G}}\,$ along $\,{\mathbb{X}}(\Omega_{p+1})$,
\begin{eqnarray}\nn
\cA_{\rm DF(WZ)}\equiv{\rm Hol}_{\underset{\tx{\ciut{(p)}}}{\mathcal{G}}}(\cdot)\ :\ C^\infty(\Omega_{p+1},\mathscr{M})\longrightarrow{\rm U}(1)\,,
\end{eqnarray}
assigning to $\,\underset{\tx{\ciut{(p)}}}{\mathcal{G}}\,$ the image of the class
\begin{eqnarray}\nn
[{\mathbb{X}}^*\underset{\tx{\ciut{(p)}}}{\mathcal{G}}]\in\check{H}^{p+1}\bigl(\Omega_{p+1},\unl{\rm U}(1)\bigr)
\end{eqnarray}
of its pullback along $\,{\mathbb{X}}\,$ in the \v{C} ech cohomology group $\,\check{H}^{p+1}\bigl(\Omega_{p+1},\unl{\rm U}(1)\bigr)\,$ with values in the sheaf $\,\unl{\rm U}(1)\,$ of locally constant maps $\,\Omega_{p+1}\longrightarrow{\rm U}(1)\,$ under the isomorphism
\begin{eqnarray}\nn
\check{H}^{p+1}\bigl(\Omega_{p+1},\unl{\rm U}(1)\bigr)\cong{\rm U}(1)\,.
\end{eqnarray}
The relevance of the geometric object thus associated with $\,\underset{\tx{\ciut{(p+2)}}}{\chi}\,$ in the classical field theory hinges upon the fact that a $p$-gerbe canonically determines the prequantum bundle of the $\sigma$-model through the so-called cohomological transgression. Upon polarisation, square-integrable sections of that bundle become wave functionals of the field theory under consideration. This and other ramifications and merits of the rigorous higher-geometric and -cohomological formulation of the classical field theory have been established in a long sequence of works \cite{Gawedzki:1987ak,Gawedzki:1999bq,Gawedzki:2002se,Gawedzki:2003pm,Gawedzki:2004tu,Schreiber:2005mi,Recknagel:2006hp,Gawedzki:2007uz,Fuchs:2007fw,Runkel:2008gr,Gawedzki:2008um,Gawedzki:2010rn,Suszek:2011hg,Suszek:2012ddg,Gawedzki:2012fu,Suszek:2013,Gawedzki:2010G} and recalled, together with the relevant technicalities, in \Rcite{Suszek:2017xlw}, to be referred to as Part I henceforth (the reference being inherited by section, proposition and theorem labels).
Incorporation of supersymmetry into the $\sigma$-model picture, with a sound theoretical motivation (such as, {\it e.g.}, cancellation of the tachyonic mode of the bosonic string), leads to the emergence of novel cohomological and geometric phenomena. In the formulation in which supersymmetries form a Lie supergroup $\,{\rm G}\,$ acting transitively by automorphisms on the supermanifold $\,\mathscr{M}\,$ into which the previously introduced worldvolume $\,\Omega_{p+1}\,$ is mapped by the lagrangean field $\,{\mathbb{X}}\,$ of the ensuing \textbf{super-$\sigma$-model}, the cohomological novelty is due, in particular, to the discrepancy between the standard de Rham cohomology $\,H^\bullet(\mathscr{M})\,$ of the \textbf{supertarget} $\,\mathscr{M}$,\ identical with the de Rham cohomology of its body $\,|\mathscr{M}|\,$ in virtue of the Kostant Theorem of \Rcite{Kostant:1975},\ and its supersymmetry-invariant refinement, $\,H^\bullet(\mathscr{M})^{\rm G}$.\ A glaring example of the discrepancy is provided by the super-Minkowski space $\,{\rm sMink}^{d,1\,\vert\,D_{d,1}}\,$ (defined precisely as a supermanifold in Sec.\,I.4.1) with the trivial de Rham cohomology (of the $d+1$ dimensional body $\,{\rm Mink}^{d,1}\equiv{\mathbb{R}}^{d,1}$) and a non-trivial supersymmetry-invariant de Rham cohomology, containing, in particular, the Green--Schwarz (GS) super-$(p+2)$-cocycles $\,\underset{\tx{\ciut{(p+2)}}}{\chi}^{\rm GS}\,$ that define the standard (super-$p$-brane) super-$\sigma$-models with the supertarget $\,{\rm sMink}^{d,1\,\vert\,D_{d,1}}\,$ for distinguished values of $\,d\,$ and $\,p$.
The topological content of the supersymmetric refinement $\,H^\bullet(\mathscr{M})^{\rm G}\,$ of the de Rham cohomology discovered for $\,\mathscr{M}={\rm sMink}^{d,1\,\vert\,D_{d,1}}\,$ by Rabin and Crane in Refs.\,\cite{Rabin:1984rm,Rabin:1985tv} justifies a search for a (super)geometrisation of the physically distinguished GS super-$(p+2)$-cocycles, representing classes in $\,H^{p+2}(\mathscr{M})^{\rm G}$,\ fully analogous to the construction of $p$-gerbes in the standard (non-super-)geometry. Indeed, the nontrivial classes in $\,H^\bullet({\rm sMink}^{d,1\,\vert\,D_{d,1}})^{{\mathbb{R}}^{d,1\,\vert\,D_{d,1}}}\,$ that trivialise in $\,H^\bullet({\rm sMink}^{d,1\,\vert\,D_{d,1}})\,$ are to be regarded as duals of certain non-trivial cycles in the orbifold $\,{\rm sMink}^{d,1\,\vert\,D_{d,1}}/\Gamma_{\rm KR}\,$ of the super-Minkowski space by the action of the Kosteleck\'y--Rabin discrete supersymmetry group of \Rcite{Kostelecky:1983qu} engendered from trivial cycles in $\,{\rm sMink}^{d,1\,\vert\,D_{d,1}}\,$ through the orbifolding, and so we should actually think of the super-$\sigma$-model as a field theory on a supermanifold of the same type as $\,\mathscr{M}\,$ ({\it i.e.}, locally modelled on the same vector bundle over $\,|\mathscr{M}|\,$ in the sense of the Gaw\c{e}dzki--Batchelor Theorem of Refs.\,\cite{Gawedzki:1977pb,Batchelor:1979a}) but with the homological duals of the supersymmetric de Rham super-cocycles without supersymmetric primitives on $\,\mathscr{M}$.\ In particular, the new supertarget is anticipated to be compact in (some of) the Gra\ss mann-odd directions. When seen from this perspective, the geometrisation of the GS super-$p$-cocycles regains its purely topological nature. Independent motivation for the geometrisation was presented in \Rcite{Fiorenza:2013nha} where its formal aspects and its relation to the classification of consistent (supersymmetric) $p$-brane models were discussed at great length in the super-Minkowskian setting, in the context of the ubiquitous holographic principle (and so also in the context of the ${\rm AdS}$/CFT correspondence).
A constructive approach to the problem of geometrisation was initiated in Part I. The proposed scheme exploited the classical relation between the Cartan--Eilenberg cohomology $\,{\rm CaE}^\bullet({\rm G})\equiv H^\bullet({\rm G})^{\rm G}\,$ of the (supersymmetry) Lie supergroup $\,{\rm G}\,$ and the Chevalley--Eilenberg cohomology $\,{\rm CE}^\bullet(\gt{g},{\mathbb{R}})\,$ of its Lie superalgebra $\,\gt{g}\,$ (with values in the trivial module $\,{\mathbb{R}}$), in conjunction with the correspondence between the second cohomology group $\,{\rm CE}^2(\gt{g},{\mathbb{R}})\,$ of the latter and equivalence classes of supercentral extensions of the Lie superalgebra $\,\gt{g}$.\ Its underlying idea was to use the extended supersymmetry groups obtained through exponentiation of the supercentral extensions determined by the GS super-$(p+2)$-cocycles (for $\,p\in\{0,1,2\}$) over the Lie \emph{supergroup} $\,{\mathbb{R}}^{d,1\,\vert\,D_{d,1}}\equiv{\rm sMink}^{d,1\,\vert\,D_{d,1}}$,\ in a manner originally discussed by de Azc\'arraga {\it et al.} in Refs.\,\cite{Chryssomalakos:2000xd}, in an explicit construction of (or sometimes even directly as) the surjective submersions entering the definition of the supergeometric objects, dubbed (\textbf{Green--Schwarz}) \textbf{super-$p$-gerbes} over $\,{\rm sMink}^{d,1\,\vert\,D_{d,1}}\,$ in Part I. The construction proceeds in full structural analogy with the by now standard geometrisation scheme of cohomological descent for de Rham cocycles, due to Murray \cite{Murray:1994db}.
The super-$p$-gerbes were subsequently shown to possess the expected supersymmetry-(${\rm Ad}_\cdot$-)equi\-vari\-ant structure, in perfect analogy with their bosonic counterpart for $\,p=1$,\ {\it cp} Refs.\,\cite{Gawedzki:2010rn,Gawedzki:2012fu,Suszek:2011,Suszek:2012ddg,Suszek:2013}, whose appearance in this picture follows from the identification of the GS super-$3$-cocycle on $\,{\rm sMink}^{d,1\,\vert\,D_{d,1}}\,$ as a super-variant of the canonical Cartan 3-form on a Lie group, and that of the associated super-$\sigma$-model in the Polyakov formulation as the super-variant of the well-known Wess--Zumino--Witten $\sigma$-model of Refs.\,\cite{Witten:1983ar,Gawedzki:1990jc,Gawedzki:1999bq,Gawedzki:2001rm}. Finally, the geometrisation scheme was adapted to the setting of the equivalent Hughes--Polchinski formulation of \Rcite{Hughes:1986dn} of the same GS super-$\sigma$-model, whereby another supergeometric object, dubbed \textbf{the extended Green--Schwarz super-$p$-gerbe} in Part I, was associated with the super-$\sigma$-model. The object unifies the metric and topological (gerbe-theoretic) data of the previously considered Nambu--Goto formulation. The passage to the Hughes--Polchinski formulation opened the possibility for a straightforward geometrisation of the $\kappa$-symmetry of the GS super-$\sigma$-model, {\it i.e.}, the linearised gauge supersymmetry discovered in Refs.\,\cite{deAzcarraga:1982njd,Siegel:1983hh,Siegel:1983ke}, known to effectively implement suppersymmetric balance between the bosonic and fermionic degrees of freedom in the field theories under consideration. The geometrisation assumed the form of an incomplete $\kappa$-equivariant structure on the extended super-$p$-gerbe, defined in analogy with its (complete) bosonic counterpart of Ref.\,\cite{Gawedzki:2010rn,Gawedzki:2012fu,Suszek:2012ddg}, derived explicitly for $\,p\in\{0,1\}\,$ and termed \textbf{the weak $\kappa$-equivariant structure}.
The geometrisation scheme developed in Part I exploited largely the exceptional tractability of the super-Minkowskian superbackground, as well as its Cartan-geometric description as a homogeneous space
\begin{eqnarray}\nn
{\rm sMink}^{d,1\,\vert\,D_{d,1}}\cong{\rm sISO}(d,1\,\vert\,D_{d,1})/{\rm SO}(d,1)
\end{eqnarray}
of the super-Poincar\'e supergroup
\begin{eqnarray}\nn
{\rm sISO}(d,1\,\vert\,D_{d,1})(d,1)\equiv{\mathbb{R}}^{d,1\,\vert\,D_{d,1}}\rtimes{\rm SO}(d,1)\,,
\end{eqnarray}
and -- indeed -- as a Lie supergroup of supertranslations,
\begin{eqnarray}\nn
{\rm sMink}^{d,1\,\vert\,D_{d,1}}\cong{\mathbb{R}}^{d,1\,\vert\,D_{d,1}}\,.
\end{eqnarray}
The latter allowed to rephrase the differential calculus on the flat supertarget entirely in terms of the components of the supersymmetry-(left-)invariant Maurer--Cartan super-1-form and of the dual supersymmetry-(left-)invariant vector fields on\footnote{The Hughes--Polchinski formulation naturally calls for the full supersymmetry group $\,{\rm s}\mathscr{P}(d,1)$.} $\,{\mathbb{R}}^{d,1\,\vert\,D_{d,1}}\,$ and work with the Chevalley--Eilenberg model of the Lie-superalgebra cohomology for the Lie superalgebra $\,\gt{smink}^{d,1\,\vert\,D_{d,1}}\,$ of the Lie supergroup $\,{\rm sMink}^{d,1\,\vert\,D_{d,1}}\,$ -- hence the algebraisation of the supersymmetric de Rham cohomology of $\,{\rm sMink}^{d,1\,\vert\,D_{d,1}}$,\ explaining the definition of the geometrisation scheme in terms of extensions of the supertarget Lie superalgebra.
\medskip
The choice of the supertarget made in Part I masks a variety of topological and algebraic problems that arise over a generic homogeneous space $\,{\rm G}/{\rm H}\,$ of a supersymmetry group $\,{\rm G}\,$ (with a Lie subgroup $\,{\rm H}\subset{\rm G}$), such as, {\it e.g.}, a non-trivial topology and metric curvature of the body, absence of a Lie-supergroup structure on $\,{\rm G}/{\rm H}$,\ induction of a highly non-linear realisation of supersymmetry and inheritance of the associated supersymmetric differential calculus on $\,{\rm G}/{\rm H}\,$ from the Lie supergroup $\,{\rm G}\,$ along a family of \emph{locally} smooth sections of the principal ${\rm H}$-bundle $\,{\rm G}\longrightarrow{\rm G}/{\rm H}$,\ in the spirit of the theory of nonlinear realisations of (super)symmetries, developed in Refs.\,\cite{Schwinger:1967tc,Weinberg:1968de,Coleman:1969sm,Callan:1969sn,Salam:1969rq,Salam:1970qk,Isham:1971dv,Volkov:1972jx,Volkov:1973ix,Ivanov:1978mx,Lindstrom:1979kq,Uematsu:1981rj,Ivanov:1982bpa,Samuel:1982uh,Ferrara:1983fi,Bagger:1983mv} and recently revived in the string-theoretic context in Refs.\,\cite{McArthur:1999dy,West:2000hr,Gomis:2006xw,McArthur:2010zm}. In the present paper, we make the first step in this general direction by extending our geometrisation scheme to the supermanifold with the topologically non-trivial and metrically curved body
\begin{eqnarray}\nn
{\rm AdS}_5\x{\mathbb{S}}^5
\end{eqnarray}
and with the structure of a homogeneous space
\begin{eqnarray}\nn
{\rm G}/{\rm H}\equiv{\rm SU}(2,2\,\vert\,4)/\bigl({\rm SO}(4,1)\x{\rm SO}(5)\bigr)=:{\rm s}\bigl({\rm AdS}_5\x{\mathbb{S}}^5\bigr)
\end{eqnarray}
of the supersymmetry group $\,{\rm G}\equiv{\rm SU}(2,2\,\vert\,4)\,$ with the body $\,{\rm SO}(5,1)\x{\rm SO}(6)$.\ The homogeneous space is devoid of any (obvious) Lie-supergroup structure and embedded in $\,{\rm SU}(2,2\,\vert\,4)\,$ patchwise smoothly by a collection of local sections of the principal ${\rm SO}(4,1)\x{\rm SO}(5)$-bundle
\begin{eqnarray}\label{eq:AdSprinc}
{\rm SU}(2,2\,\vert\,4)\longrightarrow{\rm SU}(2,2\,\vert\,4)/\bigl({\rm SO}(4,1)\x{\rm SO}(5)\bigr)\,.
\end{eqnarray}
The supertarget is further endowed with a GS super-3-cocycle induced by the ${\rm SO}(4,1)\x{\rm SO}(5)$-basic \textbf{Metsaev--Tseytlin super-3-cocycle}
\begin{eqnarray}\nn
\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}=-\bigl(\widehat C\,\widehat\Gamma_{\widehat a}\otimes\sigma_3\bigr)_{\widehat\a\widehat\beta}\,\theta_{\rm L}^{\widehat\a}\wedge\theta_{\rm L}^{\widehat a}\wedge\theta_{\rm L}^{\widehat\beta}
\end{eqnarray}
on $\,{\rm SU}(2,2\,\vert\,4)$,\ first introduced in \Rcite{Metsaev:1998it} and written in terms of components $\,\theta_{\rm L}^{\widehat\a}\,$ and $\,\theta_{\rm L}^{\widehat a}\,$ of the $\gt{su}(2,2\,\vert\,4)$-valued Maurer--Cartan super-1-form along the direct-sum complement $\,\gt{t}\,$ of the Lie subalgebra $\,\gt{h}\equiv\gt{so}(4,1)\oplus\gt{so}(5)\,$ of the isotropy Lie subgroup $\,{\rm H}\equiv{\rm SO}(4,1)\x{\rm SO}(5)\subset{\rm SU}(2,2\,\vert\,4)\,$ defining a reductive decomposition
\begin{eqnarray}\nn
\gt{su}(2,2\,\vert\,4)\equiv\gt{g}=\gt{t}\oplus\gt{h}\,,\qquad\qquad[\gt{h},\gt{t}]\subset\gt{t}\,,
\end{eqnarray}
and of certain ${\rm H}$-invariant tensors $\,\bigl(\widehat C\,\widehat\Gamma_{\widehat a}\otimes\sigma_3\bigr)_{\widehat\a\widehat\beta}\,$ defined in Sec.\,\ref{sec:ssextaAdSS}. Altogether, these form, according to the general rules discovered and elucidated in the original literature on the subject of nonlinear realisations of (super)symmetries, cited above, and reviewed in Section \ref{sec:Cartsgeom}, a super-3-form on $\,{\rm G}\,$ that descends (through pullback) to the quotient supermanifold $\,{\rm s}\bigl({\rm AdS}_5\x{\mathbb{S}}^5\bigr)\,$ along the aforementioned local sections of \eqref{eq:AdSprinc}, with local pullbacks glueing smoothly over intersections of their domains. This time, the choice of the supertarget is motivated by the critical relevance of the associated super-$\sigma$-model, postulated by Metsaev and Tseytlin in \Rcite{Metsaev:1998it}, to the formulation and study of the celebrated ${\rm AdS}$/CFT correspondence of Refs.\,\cite{Maldacena:1997re,Maldacena:1998im}, of much significance in the research on the quantum dynamics of strongly coupled systems with a non-abelian gauge symmetry, such as, {\it e.g.}, the quark-gluon plasma. The \textbf{superbackground}
\begin{eqnarray}\nn
\bigl({\rm s}\bigl({\rm AdS}_5\x{\mathbb{S}}^5\bigr),\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}\bigr)
\end{eqnarray}
is also directly related to the formerly scrutinised flat superbackground
\begin{eqnarray}\nn
\bigl({\rm sMink}^{9,1\,\vert\,32},\underset{\tx{\ciut{(3)}}}{\chi}^{\rm GS}\bigr)
\end{eqnarray}
through the flattening limit
\begin{eqnarray}\label{eq:Rtoinf}
R\to\infty
\end{eqnarray}
of the \emph{common} radius $\,R\,$ of the generating 1-cycle of $\,{\rm AdS}_5\cong{\mathbb{S}}^1\x{\mathbb{R}}^{\x 4}\,$ and that of the 5-sphere in the body of the supertarget, with a dual algebraic realisation in the form of an \.In\"on\"u--Wigner contraction
\begin{eqnarray}\label{eq:IWcontrsAdS}
\gt{su}(2,2\,\vert\,4)\xrightarrow[{\rm rescaling}]{\ R-{\rm dependent}\ }\gt{su}(2,2\,\vert\,4)_R\xrightarrow{\ R\to\infty\ }\gt{smink}^{d,1\,\vert\,D_{d,1}}\,.
\end{eqnarray}
It is worth emphasising that the asymptotic relation between the two superbackgrounds was one of the basic guiding principles on which Metsaev and Tseytlin founded their construction of the two-dimensional super-$\sigma$-model for $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)$,\ and so it is apposite to expect that it should lift to the sought-after geometrisation of the GS super-3-cocycle descended from the Metsaev--Tseytlin (MT) super-3-cocycle -- this is the premise upon which much of the work reported herein has been based.
The latter super-3-cocycle admits a manifestly supersymmetric primitive $\,\underset{\tx{\ciut{(2)}}}{\beta}$,\ found by Roiban and Siegel in \Rcite{Roiban:2000yy}, that also descends to the supertarget $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,$ along the local sections of \eqref{eq:AdSprinc}. Thus, according to the geometrisation scheme formulated in Part I, the super-3-cocycle gives rise to a trivial GS super-1-gerbe over $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)$,\ which we construct explicitly in Section \ref{sec:trsAdSSext}. However, $\,\underset{\tx{\ciut{(2)}}}{\beta}\,$ does \emph{not} reproduce the (non-supersymmetric) primitive of the GS super-3-cocycle $\,\underset{\tx{\ciut{(3)}}}{\chi}^{\rm GS}\,$ on $\,{\rm sMink}^{9,1\,\vert\,32}\,$ in the limit \eqref{eq:Rtoinf}. Consequently, the trivial super-1-gerbe does \emph{not} contract to its super-Minkowskian counterpart. In Sections \ref{sec:KamSakext} and \ref{sec:KostRabdef}, we systematically examine deformations of the supersymmetry algebra $\,\gt{su}(2,2\,\vert\,4)\,$ within the category of Lie superalgebras which could trivialise -- through a mechanism originally devised by de Azc\'arraga {\it et al.} in the super-Minkowskian setting in \Rcite{Chryssomalakos:2000xd} and applied successfully in the same setting in Part I -- the super-3-cocycle $\,\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}\,$ in a manner compatible with the \.In\"on\"u--Wigner contraction \eqref{eq:IWcontrsAdS}. As the study of \emph{all} possible such deformations is well beyond the scope of the present work, we take guidance from an explicit asymptotic analysis of the charge deformation of the Poisson algebra of the Noether charges of supersymmetry in the MT super-$\sigma$-model, based -- in the sense made precise in Section \ref{sec:wrapanom} -- on the class of the supersymmetric primitive $\,\underset{\tx{\ciut{(2)}}}{\beta}$.\ The analysis has been carried out in Section \ref{sec:ssextaAdSS} and prepared by an abstract discussion of the r\^ole of the topological WZ term
in the action functional of the (super-)$\sigma$-model in the said deformation, presented in Section \ref{sec:wrapanom}. Its results shed light on the significance of the so-called pseudo-invariance of the WZ term for the existence and structure of the deformation. When concretised in the setting of the GS super-$\sigma$-model for $\,{\rm sMink}^{d,1\,\vert\,D_{d,1}}\,$ in Sections \ref{subsect:spartextMink} and \ref{subsect:sstringextsMink}, they provide direct evidence of the central r\^ole of the Kosteleck\'y--Rabin charges in the geometrisation of the Cartan--Eilenberg cohomology on the supersymmetry group through Lie-superalgebra extensions determined by the GS super-$(p+2)$-cocycles. In particular, we reobtain the superstring extension of the super-Minkowski superalgebra $\,\gt{smink}^{d,1\,\vert\,D_{d,1}}\,$ of Part I as a deformation of the Poisson algebra of the Noether charges of supersymmetry of the two-dimensional GS super-$\sigma$-model for $\,{\rm sMink}^{d,1\,\vert\,D_{d,1}}\,$ by the Gra\ss mann-odd Kosteleck\'y--Rabin charge. This, in conjunction with the knowledge of the topology of the body of the supertarget of main interest, $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)$,\ allows us to organise our search for extensions of $\,\gt{su}(2,2\,\vert\,4)\,$ into two natural directions:
\begin{enumerate}
\item a Gra\ss mann-even central extension deforming (exclusively) the anticommutator of the supercharges in an arbitrary manner, contemplated with view to recovering the desired non-supersymmetric correction to the Roiban--Siegel primitive of the Metsaev--Tseytlin super-3-cocycle as the leading term in an asymptotic expansion of a supersymmetric super-2-form on the resultant extended supersymmetry group, and
\item a generic extension determined by a Gra\ss mann-odd deformation of the commutator $\,[Q_{\a\a'I},P_{\widehat a}]\,$ engineered so as to allow for a trivialisation, on the resultant extended supersymmetry group, of a super-2-cocycle asymptoting to the Kosteleck\'y--Rabin super-2-cocycle of Section \ref{subsect:sstringextsMink} in the limit of an infinite radius of $\,{\rm AdS}_5\x{\mathbb{S}}^5$.
\end{enumerate}
The search returns negative results, and so -- at this stage -- the non-contractible trivial super-1-gerbe with curvature $\,\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}\,$ associated with the supersymmetric primitive $\,\underset{\tx{\ciut{(2)}}}{\beta}\,$ remains as the sole consistent geometrisation of the MT super-3-cocycle. The paper concludes with a discussion of an alternative approach to geometrisation founded on the assumption that it is \emph{not} the asymptotic relation between the fixed Green--Schwarz super-3-cocycles: $\,\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}\,$ and $\,\underset{\tx{\ciut{(2)}}}{\chi}^{\rm GS}\,$ (the latter on $\,{\rm Mink}^{9,1}$) but rather the asymptotic relation between the supertarget geometries: $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,$ and $\,{\rm Mink}^{9,1}$,\ modelled by the contraction mechanism for the (extended) super-${\rm AdS}_5\x{\mathbb{S}}^5\,$ Lie superalgebra, that ought to be regarded as fundamental. The approach takes as the point of departure a pair of extended Lie superalgebras: a Gra\ss mann-odd deformation of the super-${\rm AdS}_5\x{\mathbb{S}}^5\,$ superalgebra and the superstring deformation of the super-Minkowskian superalgebra, assumed to be related by an \.In\"on\"u--Wigner contraction to begin with, and \emph{defines} the Green--Schwarz super-3-cocycle on $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,$ as the exterior derivative of the manifestly (extended-)supersymmetric super-2-form on the Lie supergroup $\,\widetilde{{\rm SU}(2,2\,\vert\,4)}\,$ (integrating the former extended Lie superalgebra) constructed as a direct structural counterpart of the known supersymmetric primitive of the GS super-3-cocycle on the extended super-Minkowski superspace used in Part I. This logical possibility, and its potential consequences for the very definition of the curved supertarget, are analysed in the toy model of the super-${\rm AdS}$ Lie superalgebra with a built in Gra\ss mann-odd deformation. The analysis indicates an interesting and promising direction of the unfinished quest for the mechanism of trivialisation of the physically relevant (class in the) Chevalley--Eilenberg cohomology of $\,\gt{su}(2,2\,\vert\,4)\,$ compatible with the \.In\"on\"u--Wigner contraction and the associated geometric transition
\begin{eqnarray}\nn
{\rm s}\bigl({\rm AdS}_5\x{\mathbb{S}}^5\bigr)\xrightarrow[{\rm rescaling}]{\ R-{\rm dependent}\ }{\rm s}\bigl({\rm AdS}_5(R)\x{\mathbb{S}}^5(R)\bigr)\xrightarrow{\ R\to\infty\ }{\rm sMink}^{9,1\,\vert\,32}\,,
\end{eqnarray}
to which we hope to return in a future work.
\medskip
The paper is organised as follows:
\begin{itemize}
\item in Section \ref{sec:Cartsgeom}, we review systematically the logic of the construction of a lagrangean field theory on a homogeneous space $\,{\rm G}/{\rm H}\,$ of a (super)symmetry Lie (super)group $\,{\rm G}\,$ (corresponding to a reductive decomposition $\,\gt{g}=\gt{t}\oplus\gt{h}\,$ of the supersymmetry algebra $\,\gt{g}$) using suitable elements of the Cartan differential calculus on $\,{\rm G}$,\ and the ensuing induction of a non-linear realisation of supersymmetry;
\item in Section \ref{sec:wrapanom}, we discuss at length the canonical description of a field-theoretic realisation, in terms of the Poisson algebra of the relevant Noether charges, of supersymmetry in a super-$\sigma$-model with a pseudo-invariant Wess--Zumino term, whereby we discover the wrapping anomaly that quantifies the departure of that realisation from the original supersymmetry algebra $\,\gt{g}\,$ -- this we take as the germ of a physically motivated (normal) extension of $\,\gt{g}\,$ studied subsequently;
\item in Section \ref{sec:sMinkext}, we identify the field-theoretic source of the super-central extensions of the super-Minkowski Lie superalgebra encountered in Part I (and giving rise to the super-$p$-gerbes, for $\,p\in\{0,1\}$,\ associated with the Green--Schwarz super-$\sigma$-models with that supertarget) and provide concrete evidence of the r\^ole played in these deformations by the winding charges measuring the monodromy of the Gra\ss mann-odd coordinates along the non-contractible cycles in the Kosteleck\'y--Rabin quotient of the super-Minkowski superspace that topologises the non-trivial Cartan--Eilenberg cohomology of the underlying Lie supergroup;
\item in Section \ref{sec:ssextaAdSS}, we introduce the Metsaev--Tseytlin superbackground over the supertarget $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,$ and analyse in great detail the (super)geometric nature of the corresponding wrapping anomaly, including its asymptotics in the flat limit $\,R\to\infty\,$ -- this determines natural paths of deformation of the relevant supersymmetry algebra $\,\gt{su}(2,2\,\vert\,4)\,$ that we pursue in later sections;
\item in Section \ref{sec:trsAdSSext}, we describe the trivial Metsaev--Tseytlin super-1-gerbe associated with the manifestly supersymmetric Roiban--Siegel primitive of the Metsaev--Tseytlin super-3-cocycle, noting the incompatibility of its structure with the \.In\"on\"u--Wigner contraction;
\item in Section \ref{sec:KamSakext}, we study a class of super-central extensions of the supersymmetry algebra $\,\gt{su}(2,2\,\vert\,4)\,$ obtained through a Gra\ss mann-even deformation of the anticommutator of the supercharges, only to find out that the admissible ones do not allow for a supersymmetry-equivariant trivialisation of the Metsaev--Tseytlin super-3-cocycle compatible with the \.In\"on\"u--Wigner contraction;
\item in Section \ref{sec:KostRabdef}, we consider two classes of natural associative deformations of the supersymmetry algebra $\,\gt{su}(2,2\,\vert\,4)\,$ engendered by a Gra\ss mann-odd deformation, of the Kosteleck\'y--Rabin type, of the commutator of the supercharge and the momentum -- the deformations are proven algebraically inconsistent;
\item in Section \ref{sec:MTaway}, drawing inspiration from the previously encountered failures of geometrisation schemes compatible with the \.In\"on\"u--Wigner contraction, we conceive an alternative geometrisation scenario that takes the asymptotic relation between the two relevant supergeometries: $\,{\rm sMink}^{9,1\,\vert\,32}\,$ and $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)$,\ and between the respective supersymmetry algebras as the organising principle; the general idea is illustrated and tested on a toy example of a super-${\rm AdS}_d$ superspace and the associated supersymmetry algebra $\,\gt{sso}(d-1,2)$.
\item in Section \ref{sec:C&O}, we recapitulate the work reported in the present paper and indicate possible directions of future research that it motivates;
\item in the Appendices, we present the relevant conventions on and facts regarding the Clifford algebras employed in the paper, and gather various technical calculations, including proofs of the propositions and theorems stated in the main text.
\end{itemize}
\section{The Cartan geometry of homogeneous superspaces and super-$\sigma$-models thereon}\label{sec:Cartsgeom}
Let $\,{\rm G}\,$ be a Lie supergroup, with the Lie superalgebra $\,\gt{g}\,$ as defined in Part I, to be referred to as {\bf the supersymmetry group} (resp.\ {\bf the supersymmetry algebra}) in what follows and let $\,\mathscr{M}\,$ be a supermanifold endowed with a transitive (left) action of $\,{\rm G}$
\begin{eqnarray}\nn
\lambda_\cdot\ :\ {\rm G}\x\mathscr{M}\longrightarrow\mathscr{M}\ :\ (g,m)\longmapsto g\vartriangleright m\equiv\lambda_g(m)\,,
\end{eqnarray}
so that there exists a ${\rm G}$-equivariant diffeomorphism
\begin{eqnarray}\nn
\mu\ :\ \mathscr{M}\xrightarrow{\ \cong\ }{\rm G}/{\rm H}\,,
\end{eqnarray}
where $\,{\rm H}\equiv{\rm G}_m\,$ is the isotropy group $\,{\rm G}_m\,$ of an arbitrary point $\,m\in\mathscr{M}$.\ The manifold $\,\mathscr{M}\,$ shall be modelled on the homogeneous space $\,{\rm G}/{\rm H}\,$ henceforth, the latter being realised locally as a section of the principal bundle\footnote{The first arrow denotes the free and transitive action of the structure group on the fibre.}
\begin{eqnarray}\nn
\alxydim{@C=1cm@R=1cm}{{\rm H} \ar[r] & {\rm G} \ar[d]^{\pi_{{\rm G}/{\rm H}}} \\ & {\rm G}/{\rm H}}
\end{eqnarray}
with the structure group $\,{\rm H}$,\ a closed Lie subgroup of $\,{\rm G}$.\ Thus, we shall work with a family of submanifolds embedded in $\,{\rm G}\,$ by the respective (local) sections\footnote{That is, equivalently, by a collection of local trivialisations.}
\begin{eqnarray}\nn
\sigma_i\ :\ \mathcal{O}_i\longrightarrow{\rm G}\ :\ g{\rm H}\longmapsto g\cdot h_i(g)\,,\quad i\in I\,,
\end{eqnarray}
of the submersive projection on the base $\,\pi_{{\rm G}/{\rm H}}$,\ associated with a trivialising cover $\,\mathcal{O}=\{\mathcal{O}_i\}_{i\in I}\,$ of the latter,
\begin{eqnarray}\nn
{\rm G}/{\rm H}=\bigcup_{i\in I}\,\mathcal{O}_i\,.
\end{eqnarray}
The redundancy of such a realisation over any non-empty intersection, $\,\mathcal{O}_{ij}\equiv\mathcal{O}_i\cap\mathcal{O}_j\neq\emptyset$,\ is accounted for by a collection of locally smooth (transition) maps
\begin{eqnarray}\nn
h_{ij}\ :\ \mathcal{O}_{ij}\longrightarrow{\rm H}\subset{\rm G}
\end{eqnarray}
fixed by the condition
\begin{eqnarray}\label{eq:Htransmap}
\forall_{x\in\mathcal{O}_{ij}}\ :\ \sigma_j(x)=\sigma_i(x)\cdot h_{ij}(x)\,.
\end{eqnarray}
The original action $\,\lambda_\cdot$,\ with the simple model on $\,{\rm G}/{\rm H}$
\begin{eqnarray}\label{eq:cosetlact}
[\lambda]_\cdot\ :\ {\rm G}\x{\rm G}/{\rm H}\longrightarrow{\rm G}/{\rm H}\ :\ (g',g{\rm H})\longmapsto(g'\cdot g){\rm H}\,,
\end{eqnarray}
is transcribed, through the $\,\sigma_i^{\rm K}$,\ into a geometric realisation of $\,{\rm G}\,$ on the image of $\,{\rm G}/{\rm K}\,$ within $\,{\rm G}$,\ with the same obvious redundancy. Indeed, consider a point $\,x\in\mathcal{O}_i\,$ and an element $\,g\in{\rm G}$.\ Upon choosing an \emph{arbitrary} index $\,j\in I\,$ with the property
\begin{eqnarray}\label{eq:cosetactpt}
\widetilde x(x;g'):=\pi_{{\rm G}/{\rm H}}\bigl(g'\cdot\sigma_i(x)\bigr)\in\mathcal{O}_j\,,
\end{eqnarray}
we find a unique $\,\unl h_{ij}(x,g')\in{\rm H}\,$ defined (on some open neighbourhood of $\,(x,g')$) by the condition
\begin{eqnarray}\label{eq:cosetact}
g'\cdot\sigma_i(x)=\sigma_j\bigl(\widetilde x(x;g')\bigr)\cdot\unl h_{ij}(x;g')^{-1}\,.
\end{eqnarray}
Note that for $\,\widetilde x(x;g')\in\mathcal{O}_{jk}\,$ we have
\begin{eqnarray}\nn
\unl h_{ik}(x;g')=\unl h_{ij}(x;g')\cdot h_{jk}\bigl(\widetilde x(x;g')\bigr)\,,
\end{eqnarray}
so that the two realisations of the action are related by a compensating transformation from the structure group $\,{\rm H}$.
The realisation of the homogeneous space $\,{\rm G}/{\rm H}\,$ within $\,{\rm G}\,$ described above enables us to reconstruct the differential calculus on the former space (and so also on $\,\mathscr{M}$) from that on the Lie supergroup $\,{\rm G}$.\ To this end, we decompose the (super)vector space $\,\gt{g}\,$ as
\begin{eqnarray}\nn
\gt{g}=\gt{t}\oplus\gt{h}
\end{eqnarray}
into the Lie algebra $\,\gt{h}\supset[\gt{h},\gt{h}]\,$ of $\,{\rm H}\,$ and its direct vector-space complement $\,\gt{t}$,\ assuming, furthermore, the decomposition to be \emph{reductive}, in the convention of Sec.\,I.3 (with $\,\gt{r}\equiv\gt{h}$), so that $\,\gt{t}\,$ acquires the status of a super-graded $\gt{h}$-module,
\begin{eqnarray}\nn
[\gt{h},\gt{t}]\subset\gt{t}\,.
\end{eqnarray}
We make a choice of the basis of $\,\gt{g}\,$ compatible with the splitting, and -- accordingly -- mark the generators of $\,\gt{t}\,$ by an underline: $\,\{t_{\unl A}\}_{\unl A\in\ovl{1,{\rm dim}\,\gt{t}}}$,\ and denote those of $\,\gt{h}\,$ as $\,\{J_\kappa\}_{\kappa\in\ovl{1,{\rm dim}\,\gt{h}}}$.\ The generators are taken to be homogeneous with respect to the super-grading
\begin{eqnarray}\nn
\gt{g}=\gt{g}^{(0)}\oplus\gt{g}^{(1)}
\end{eqnarray}
in which $\,\gt{g}^{(0)}\,$ is the Gra\ss mann-even Lie subalgebra of $\,\gt{g}$,\ containing $\,\gt{h}$,\ and $\,\gt{g}^{(1)}\,$ is the Gra\ss mann-odd ${\rm ad}$-module thereof,
\begin{eqnarray}\nn
[\gt{g}^{(0)},\gt{g}^{(1)}]\subset\gt{g}^{(1)}\,.
\end{eqnarray}
The latter decomposition divides the set $\,\{t_A\}_{A\in\ovl{1,{\rm dim}\,\gt{g}}}\,$ of the homogeneous generators of $\,\gt{g}$,\ satisfying the defining supercommutation relations
\begin{eqnarray}\nn
[t_A,t_B\}=f_{AB}^{\ \ C}\,t_C\,,
\end{eqnarray}
into subsets with fixed Gra\ss mann parity: even ($\{B_a\}_{a\in\ovl{1,{\rm dim}\,\gt{g}^{(0)}}}$) and odd ($\{F_{\widehat\a}\}_{\widehat\a\in\ovl{1,{\rm dim}\,\gt{g}^{(1)}}}$), subject to relations
\begin{eqnarray}\nn
[B_a,B_b]=f_{ab}^{\ \ c}\,B_c\,,\qquad\qquad\{F_{\widehat\a},F_{\widehat\beta}\}=f_{\widehat\a\widehat\beta}^{\ \ c}\,B_c\,,\qquad\qquad[B_a,F_{\widehat\a}]=f_{a\widehat\a}^{\ \ \widehat\beta}\,F_{\widehat\beta}\,.
\end{eqnarray}
Accordingly, the generators of the module $\,\gt{t}\,$ further split into subsets: the Gra\ss mann-even ones $\,\{P_{\widehat{a}}\}_{\widehat{a}\in\ovl{1,{\rm dim}\,\gt{t}^{(0)}}}\,$ and the Gra\ss mann-odd ones $\,\{Q_{\widehat\a}\}_{\widehat\a\in\ovl{1,{\rm dim}\,\gt{t}^{(1)}}}$,\ where $\,\gt{t}^{(1)}\equiv\gt{g}^{(1)}\,$ in our considerations.
We may now span the tangent sheaf of $\,{\rm G}/{\rm H}\,$ over $\,\mathcal{O}_i\ni x\,$ on (restrictions of) the fundamental vector fields of $\,[\lambda]_\cdot$.\ These correspond to certain point-dependent (over $\,{\rm G}/{\rm H}$) linear combinations $\,\mathcal{K}_X\,$ of the right- and left-invariant vector fields on $\,\sigma_i(\mathcal{O}_i)\subset{\rm G}$,\ that is of
\begin{eqnarray}\nn
\mathcal{R}_X\bigl(\sigma_i(x)\bigr)=\tfrac{{\mathsf d}\ }{{\mathsf d} t}\mathord{\restriction}_{t=0}\,\bigl({\rm e}^{tX}\cdot\sigma_i(x)\bigr)\,,\quad X\in\gt{g}
\end{eqnarray}
and
\begin{eqnarray}\nn
\mathcal{L}_{Y_i(X;x)}\bigl(\sigma_i(x)\bigr)=\tfrac{{\mathsf d}\ }{{\mathsf d} t}\mathord{\restriction}_{t=0}\,\bigl(\sigma_i(x)\cdot{\rm e}^{tY_i(X;x)}\bigr)\,,\quad Y_i(X;x)\in\gt{h}\,,
\end{eqnarray}
respectively. The relevant combinations are readily read off from \Reqref{eq:cosetact}. Indeed, the left-regular translation of a point $\,\sigma_i(x),\ x\in\mathcal{O}_i\,$ by $\,{\rm e}^{tX},\ X\in\gt{g}\,$ fixes the right-invariant component of the (modelling) fundamental vector field in the form $\,\mathcal{R}_X(\sigma_i(x))$,\ whereas the compensating r\^ole of the right-regular translation $\,\unl h_{ii}(x,{\rm e}^{tX})\equiv{\rm e}^{tY_i(X;x)}\,$ along $\,{\rm H}\,$ identifies the corresponding left-invariant component, through imposition of the constraints
\begin{eqnarray}\label{eq:Yid}
{\rm e}^{tX}\cdot\sigma_i(x)\cdot{\rm e}^{tY_i(X;x)}=\sigma_i\circ\mathcal{G}_t^{\Xi_X}(x)\,,
\end{eqnarray}
written for $\,t\in]-\varepsilon,\varepsilon[,\ \varepsilon\gtrapprox 0\,$ and the fundamental vector field $\,\Xi_X\,$ associated with $\,X\in\gt{g}\,$ in the standard manner as
\begin{eqnarray}\nn
\Xi_X(x)=\tfrac{{\mathsf d}\ }{{\mathsf d} t}\mathord{\restriction}_{t=0}\,[\lambda]_{{\rm e}^{tX}}(x)\,,\qquad x\in{\rm G}/{\rm H}\,,
\end{eqnarray}
with the (local) flow
\begin{eqnarray}\nn
\mathcal{G}_\cdot^{\Xi_X}\ :\ ]-\varepsilon,\varepsilon[\longrightarrow{\rm Diff}_{\rm loc}({\rm G}/{\rm H})\,.
\end{eqnarray}
We shall write out $\,\Xi_X\,$ in the local coordinate basis of the tangent bundle engendered by local coordinates $\,Z_i^{\unl A}\equiv(X_i^{\widehat a},\theta_i^{\widehat\a})\,$ corresponding to a charting of $\,\mathcal{O}_i\,$ (and so also of $\,\sigma_i(\mathcal{O}_i)\subset{\rm G}$) by flows of left-invariant vector fields on $\,{\rm G}\,$ along $\,\gt{t}$ (in the standard manner, familiar from constructive proofs of the Frobenius Theorem) as
\begin{eqnarray}\nn
\Xi_X\mathord{\restriction}_{\mathcal{O}_i}=\delta_X Z_i^{\unl A}\,\tfrac{\partial\ }{\partial Z^{\unl A}_i}\,.
\end{eqnarray}
Differentiating both sides of \Reqref{eq:Yid} at $\,t=0$,\ we then obtain the result
\begin{eqnarray}\nn
\Xi_X\righthalfcup E(Z_i)={\mathsf T}_e{\rm Ad}_{\sigma_i(Z_i)^{-1}}(X)+Y_i(X;x)\,,
\end{eqnarray}
expressed in terms of the pullbacks
\begin{eqnarray}\label{eq:pullbackMC}\qquad\qquad\qquad
\sigma_i^*\theta_{\rm L}(Z_i)={\mathsf d} Z_i^{\unl A}\,E_{\unl A}^{\ A}(Z_i)\otimes t_A\equiv E^A(Z_i)\otimes t_A\equiv{\mathsf d} Z_i^{\unl A}\,E_{\unl A}(Z_i)\equiv E(Z_i)\in{\mathsf T}^*_{Z_i}({\rm G}/{\rm H})\otimes\gt{g}\,,
\end{eqnarray}
of the left-invariant $\gt{g}$-valued Maurer--Cartan super-1-form
\begin{eqnarray}\nn
\theta_{\rm L}=\theta_{\rm L}^A\otimes t_A
\end{eqnarray}
on $\,{\rm G}$,\ the latter being fixed by the standard condition
\begin{eqnarray}\nn
\mathcal{L}_X\righthalfcup\theta_{\rm L}=X\,.
\end{eqnarray}
It is these pullbacks of certain distinguished -- through the analysis that follows -- linear combinations of components of the Maurer--Cartan super-1-form that descend to the dual sheaf of (super)differential forms on $\,{\rm G}/{\rm H}$.
Our result translates into a pair of equations:
\begin{eqnarray}\label{eq:leftvarfiel}
\delta_X Z_i^{\unl A}\,E_{\unl A}^{\ \unl B}(Z_i)&=&{\mathsf T}_e{\rm Ad}_{\sigma_i(Z_i)^{-1}}(X)^{\unl B}\,,\\\cr
Y_i^\kappa(X;x)&=&\delta_X Z_i^{\unl A}\,E_{\unl A}^{\ \kappa}(Z_i)-{\mathsf T}_e{\rm Ad}_{\sigma_i(Z_i)^{-1}}(X)^\kappa \nn
\end{eqnarray}
of which the former determines (components of) the field $\,\Xi_{X;i}$,\ whereas the latter subsequently uses them to determine the compensating translation $\,Y_i(X;x)$.\ Note, in particular, that the dependence of the latter upon $\,X\,$ is linear, as expected, {\it i.e.},
\begin{eqnarray}\label{eq:YlinX}
Y_i(X;x)=X^A\,y_{i\,A}(x)
\end{eqnarray}
for some (smooth) maps
\begin{eqnarray}\nn
y_{i\,A}\ :\ \mathcal{O}_i\longrightarrow\gt{h}\,,\qquad A\in\ovl{1,{\rm dim}\,\gt{g}}\,.
\end{eqnarray}
The ensuing local vector fields
\begin{eqnarray}\nn
\mathcal{K}_{i\,X}\bigl(\sigma_i(x)\bigr)=\mathcal{R}_X\bigl(\sigma_i(x)\bigr)+\mathcal{L}_{Y_i(X;x)}\bigl(\sigma_i(x)\bigr)
\end{eqnarray}
are, by definition, the pushforwards, along $\,{\mathsf T}\sigma_i$,\ of the fundamental vector fields for the \emph{left} action $\,[\lambda]_\cdot\,$ of $\,{\rm G}\,$ on $\,{\rm G}/{\rm H}\,$ restricted to $\,\mathcal{O}_i$,
\begin{eqnarray}\label{eq:KXaspfXiX}
\mathcal{K}_{i\,X}\bigl(\sigma_i(x)\bigr)\equiv{\mathsf T}_x\sigma_i\bigl(\Xi_X(x)\bigr)\,,
\end{eqnarray}
and so they satisfy the relations
\begin{eqnarray}\label{eq:KKKh}
[\mathcal{K}_{i\,X_1},\mathcal{K}_{i\,X_2}]=-\mathcal{K}_{i\,[X_1,X_2]}\,.
\end{eqnarray}
Our hitherto considerations, in conjunction with the assumptions made, pave the way to the standard construction of supersymmetric ({\it i.e.}, globally ${\rm G}$-invariant) lagrangean field theories with the (typical) fibre of the covariant configuration bundle given by (or, to put it differently, with fields in the lagrangean density taking values in) $\,\mathscr{M}\cong{\rm G}/{\rm H}$.\ Indeed, denote, with view to subsequent analyses and for $\,i\,$ and $\,j\,$ as above (in a mild abuse of the notation),
\begin{eqnarray}\label{eq:cosetactptcoord}
Z_j\circ\pi_{{\rm G}/{\rm H}}\circ\lambda_{g'}\bigl(\sigma_i(Z_i)\bigr)=:\widetilde Z_j(Z_i,g')
\end{eqnarray}
and
\begin{eqnarray}\label{eq:cosetactcoord}
\bigl(g'\cdot\sigma_i(Z_i)\bigr)^{-1}\cdot\sigma_j\bigl(\widetilde Z_j(Z_i,g')\bigr)=:\unl h_{ij}(Z_i,g')\,.
\end{eqnarray}
We then compute
\begin{eqnarray}
E^A\bigl(\widetilde Z_j(Z_i;g')\bigr)\otimes t_A&\equiv&\sigma_j^*\theta_{\rm L}\bigl(\widetilde Z_j(Z_i,g')\bigr)\cr\cr
&=&E^{\unl B}(Z_i)\,\Gamma_{\unl B}^{\ \unl A}\bigl(Z_i,g';j\bigr)\otimes t_{\unl A}+\bigl(E^\lambda(Z_i)\,\Gamma_\lambda^{\ \kappa}\bigl(Z_i,g';j\bigr)+\bigl[\unl h_{ij}^*\theta_{\rm L}\bigl(Z_i,g'\bigr)\bigr]^\kappa\bigr)\otimes J_\kappa \label{eq:thetgtens}
\end{eqnarray}
in terms of the matrices
\begin{eqnarray}\nn
\Gamma_A^{\ B}\bigl(Z_i,g';j\bigr)\,t_B:={\mathsf T}_e{\rm Ad}_{\unl h_{ij}(Z_i,g')^{-1}}(t_A)
\end{eqnarray}
and the $\gt{h}$-valued 1-forms
\begin{eqnarray}\nn
\bigl[\unl h_{ij}^*\theta_{\rm L}\bigl(Z_i,g'\bigr)\bigr]^\kappa\otimes J_\kappa\equiv\unl h_{ij}^*\theta_{\rm L}\bigl(Z_i,g'\bigr)\,.
\end{eqnarray}
Thus, components of the Maurer--Cartan super-1-form along $\,\gt{h}\,$ transform under the induced action as a connection 1-form for $\,{\rm G}\longrightarrow{\rm G}/{\rm H}$,\ whereas those along $\,\gt{t}\,$ undergo tensorial transformations under the compensating right ${\rm H}$-translations. Consequently, linear combinations
\begin{eqnarray}\nn
\widehat T\equiv T_{\unl A_1\unl A_2\ldots\unl A_p}\,\theta_{\rm L}^{\unl A_1}\wedge\theta_{\rm L}^{\unl A_2}\wedge\cdots\wedge\theta_{\rm L}^{\unl A_p}
\end{eqnarray}
of (wedge products of) the latter with ${\rm H}$-invariant tensors $\,T_{\unl A_1\unl A_2\ldots\unl A_p}\equiv T_{[\unl A_1\unl A_2\ldots\unl A_p]}\in{\mathbb{C}}\,$ as coefficients can be used as building blocks of the sought-after supersymmetric lagrangean densities. Indeed, they are not only (right-)${\rm H}$-invariant (by construction), but also ${\rm H}$-horizontal as
\begin{eqnarray}\nn
\forall_{\unl A\in\ovl{1,{\rm dim}\,\gt{t}}}\ \forall_{Y\in\gt{h}}\ :\ \mathcal{L}_Y\righthalfcup\theta_{\rm L}^{\unl A}=0\,,
\end{eqnarray}
whence -- altogether -- ${\rm H}$-basic, and so they descend to the orbit supermanifold $\,{\rm G}/{\rm H}$.\ It ought to be emphasised that their pullbacks to $\,{\rm G}/{\rm H}\,$ do \emph{not} depend on the choice of the local section $\,\sigma_i\,$ over $\,\mathcal{O}_i\,$ (independently of the choice of the local coordinates $\,Z_i$) and hence, most importantly, glue smoothly over non-empty intersections $\,\mathcal{O}_{ij}\,$ to \emph{global} superdifferential forms on $\,{\rm G}/{\rm H}\,$ in consequence of the tensorial properties of the $\,\theta_{\rm L}^{\unl A}\,$ with respect to right translations by elements of $\,{\rm H}$.\void{The inherent redundancy of the realisation of the homogeneous space in $\,{\rm G}\,$ in terms of sections of the principal ${\rm H}$-bundle $\,{\rm G}\longrightarrow{\rm G}/{\rm H}\,$ does not affect the ensuing invariant differential calculus on $\,{\rm G}/{\rm H}$.\ Nevertheless, we may remove it for a \emph{given} family of local sections $\,\mathscr{R}=\{\sigma_i\}_{i\in I}\,$ covering the entire base $\,{\rm G}/{\rm H}\,$ ({\it i.e.}, for a complete family of local trivialisations of the principal ${\rm H}$-bundle) by fixing a map
\begin{eqnarray}\label{eq:trivindexmap}
\jmath_\mathscr{R}\ :\ {\rm G}/{\rm H}\longrightarrow I
\end{eqnarray}
with the property
\begin{eqnarray}\nn
x\in\mathcal{O}_{\jmath_\mathscr{R}(x)}\,.
\end{eqnarray}
We shall employ this map in our later considerations.}\ Furthermore, and importantly, they satisfy, for arbitrary $\,X\in\gt{g}$,\ the identities
\begin{eqnarray}\nn
\mathcal{K}_{i\,X}\righthalfcup\widehat T\equiv\mathcal{R}_X\righthalfcup\widehat T\,.
\end{eqnarray}
We conclude this part of our discussion by taking a closer look at the pullbacks \eqref{eq:pullbackMC} along the distinguished sections
\begin{eqnarray}\label{eq:embsec}
\sigma_i(Z_i)=\unl g_i\cdot g_i(X_i)\cdot{\rm e}^{\Theta_i(Z_i)}\,,\qquad\qquad\Theta_i(Z_i)=\Theta_i^{\widehat\a}(Z_i)\,Q_{\widehat\a}
\end{eqnarray}
with
\begin{eqnarray}\nn
g_i(X_i)={\rm e}^{X_i^{\widehat a}\,P_{\widehat a}}
\end{eqnarray}
and
\begin{eqnarray}\nn
\Theta_i^{\widehat\a}(Z_i)=\theta_i^{\widehat\beta}\,f_{i\,\widehat\beta}^{\ \ \widehat\a}(X_i)
\end{eqnarray}
in general depending upon the Gra\ss mann-even coordinate (through some functions $\,f_{i\,\widehat\beta}^{\ \ \widehat\a}$). Here, $\,\unl g_i\in{\rm G}\,$ defines a reference point $\,\unl g_i\,{\rm H}\,$ in the neighbourhood $\,\mathcal{O}_i\,$ on which the local coordinate system $\,Z_i\,$ is centred. In this setting, to be encountered in the sequel, explicit functional formul\ae ~for the restricted Vielbeine $\,E_{\unl A}^{\ A}(Z)\,$ are derived with the help of the retraction
\begin{eqnarray}\nn
\sigma_{i\,\cdot}\ :\ {\rm G}/{\rm H}\x[0,1]\longrightarrow{\rm G}\ :\ (Z_i,t)\longmapsto \unl g_i\cdot g_i(X_i)\cdot{\rm e}^{t\,\Theta_i(Z_i)}\equiv\sigma_{i\,t}(Z_i)
\end{eqnarray}
of $\,\sigma_i(\mathcal{O}_i)\,$ to the image of its body within $\,{\rm G}$.\ The result is stated in
\begin{Prop}\label{prop:Vielder}
Components of the pullback of the $\gt{g}$-valued Maurer--Cartan super-1-form along the local section \eqref{eq:embsec}, defined in \Reqref{eq:pullbackMC}, take the form
\begin{eqnarray}\nn
\left(\begin{smallmatrix} E^{\widehat\a} \\ E^{\widehat a} \end{smallmatrix}\right)(Z_i)=\vec\mathcal{E}_0(Z_i)+\tfrac{2\,{\rm sh}^2\tfrac{M_i(Z_i)}{2}}{M_i(Z_i)^2}\,\left(\begin{smallmatrix} 0 \\ [{\mathsf D}\Theta_i(Z_i),\Theta_i(Z_i)]^{\widehat a} \end{smallmatrix}\right)+\tfrac{{\rm sh}M_i(Z_i)}{M_i(Z_i)}\,\left(\begin{smallmatrix} {\mathsf D}\Theta_i^{\widehat\a}(Z_i) \\ 0 \end{smallmatrix}\right)\,,
\end{eqnarray}
where
\begin{eqnarray}\nn
\vec\mathcal{E}_0(Z_i):=\left(\begin{smallmatrix} 0 \\ e^{\widehat a}(X_i) \end{smallmatrix}\right)
\end{eqnarray}
is the value of the super-1-form at $\,\theta=0$,
\begin{eqnarray}\nn
{\mathsf D}\Theta_i(Z_i)\equiv{\mathsf D}\Theta_i^A(Z_i)\otimes t_A:=\bigl({\mathsf d}\Theta_i^A(Z_i)+f_{\widehat a\widehat\beta}^{\ \ A}\,e^{\widehat a}(X_i)\,\Theta_i^{\widehat\beta}(Z_i)\bigr)\otimes t_A
\end{eqnarray}
and
\begin{eqnarray}\nn
M_i(Z_i):=\left(\begin{smallmatrix} 0 & -\Theta_i^{\widehat\gamma}(Z_i)\,f_{\widehat\gamma b}^{\ \ \widehat\a} \\ \Theta_i^{\widehat\gamma}(Z_i)\,f_{\widehat\gamma\widehat\beta}^{\ \ a} & 0 \end{smallmatrix}\right)
\end{eqnarray}
\end{Prop}
\noindent\begin{proof}
A proof is given in App.\,\ref{app:Vielder}
\end{proof}
By now, we have all the tools necessary for the definition and subsequent study of the two-dimensional supersymmetric lagrangean field theory of interest, that is the two-dimensional Green--Schwarz super-$\sigma$-model of smooth embeddings\footnote{Here, we regard the target supermanifold $\,{\rm G}/{\rm H}\equiv\mathscr{M}\,$ as (the total space of) a Grassmann bundle of a vector bundle over a given base (body) $\,|{\rm G}/{\rm H}|$,\ in the spirit of the fundamental Gaw\c{e}dzki--Batchelor Theorem of Refs.\,\cite{Gawedzki:1977pb,Batchelor:1979a}.}
\begin{eqnarray}\nn
\xi\ :\ \Omega_2\longrightarrow{\rm G}/{\rm H}
\end{eqnarray}
of the compact worldsheet $\,\Omega_2\,$ in the homogeneous space $\,{\rm G}/{\rm H}\,$ of the Lie supergroup $\,{\rm G}$.\ The model uses components of the Maurer--Cartan super-1-form along $\,\gt{t}\,$ (contracted with suitable ${\rm H}$-invariant tensors) which we pull back to the embedded worldsheet $\,\xi(\Omega_2)\,$ along the previously considered family of (restrictions of) local sections of the principal ${\rm H}$-bundle $\,{\rm G}\longrightarrow{\rm G}/{\rm H}\,$ of the (local-)coordinate form
\begin{eqnarray}\nn
\sigma_i\ :\ \mathcal{O}_i\longrightarrow{\rm G}\ :\ Z_i\equiv(\theta_i,X_i)\longmapsto\unl g_i\cdot g_i(X_i)\cdot{\rm e}^{\Theta_i(Z_i)}\,,\qquad i\in I\,,
\end{eqnarray}
supported over elements of the trivialising open cover $\,\mathcal{O}=\{\mathcal{O}_i\}_{i\in I}\,$ of the base $\,{\rm G}/{\rm H}\,$ of the principal bundle $\,{\rm G}\longrightarrow{\rm G}/{\rm H}$.\ Here, the right-hand side is to be understood as the unital(-time) flow of the group element $\,\unl g_i\,$ first along the integral lines of the left-invariant vector field engendered by the Lie-superalgebra element $\,X_i^{\widehat a}\,P_{\widehat a}\,$ and subsequently along the one associated with $\,\Theta_i^{\widehat\a}(Z_i)\,Q_{\widehat\a}$.\ Next, we take an arbitrary tesselation $\,\triangle(\Omega_2)\,$ of $\,\Omega_2\,$ subordinate, for a given map $\,\xi$,\ to the open cover $\,\mathcal{O}$,\ as reflected by the existence of a map $\,i_\cdot\ :\ \triangle(\Omega_2)\longrightarrow I\,$ with the property
\begin{eqnarray}\nn
\forall_{\zeta\in\triangle(\Omega_2)}\ :\ \xi(\zeta)\subset\mathcal{O}_{i_\zeta}\,.
\end{eqnarray}
Let $\,\gt{P}\subset\triangle(\Omega_2)\,$ be the set of plaquettes of the tesselation,
\begin{eqnarray}\nn
\Omega_2=\bigcup_{\tau\in\gt{P}}\,\tau\,.
\end{eqnarray}
Given such data, we may write, in the Polyakov formulation,
\begin{eqnarray}\label{eq:supersimod}
&S_{{\rm GS}}[\xi]=\sum_{\tau\in\gt{P}}\,S^{(\tau)}_{{\rm GS}}[\xi_\tau]\,,\qquad\qquad\xi_\tau:=\xi\mathord{\restriction}_\tau&\\ \cr
&S^{(\tau)}_{{\rm GS}}[\xi_\tau]=-\tfrac{1}{2}\,\int_\tau\,\Vol(\Omega_2)\,\sqrt{\vert{\rm det}\,{\rm g}\vert}\,{\rm g}^{-1\,ij}\,{\rm g}_{\widehat a\widehat b}\,\bigl(\partial_i\righthalfcup(\sigma_{i_\tau}\circ\xi_\tau)^*\theta_{\rm L}^{\widehat a}\bigr)\,\bigl(\partial_j\righthalfcup(\sigma_{i_\tau}\circ\xi_\tau)^*\theta_{\rm L}^{\widehat b}\bigr)+S^{(\tau)}_{{\rm WZ},{\rm MT}}[\xi_\tau]\,,&\nn
\end{eqnarray}
in which the Wess--Zumino term
\begin{eqnarray}\nn
S^{(\tau)}_{{\rm WZ},{\rm MT}}[\xi_\tau]=\int_\tau\,\xi_\tau^*{\mathsf d}^{-1}\bigl(\sigma_{i_\tau}^*\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}\bigr)
\end{eqnarray}
has as its integrand a global primitive of the relevant Green--Schwarz super-3-cocycle on $\,{\rm G}/{\rm H}\,$ with local restrictions
\begin{eqnarray}\nn
\underset{\tx{\ciut{(3)}}}{{\rm H}}\mathord{\restriction}_{\mathcal{O}_i}\equiv\sigma_i^*\underset{\tx{\ciut{(3)}}}{\chi}\,,
\end{eqnarray}
where $\,\underset{\tx{\ciut{(3)}}}{\chi}\,$ is a super-3-cocycle on $\,{\rm G}\,$ defined as a linear combination
\begin{eqnarray}\nn
\underset{\tx{\ciut{(3)}}}{\chi}=\gamma_{\unl A\unl B\unl C}\,\theta_{\rm L}^{\unl A}\wedge\theta_{\rm L}^{\unl B}\wedge \theta_{\rm L}^{\unl C}
\end{eqnarray}
of the distinguished components of the Maurer--Cartan super-1-form along $\,\gt{t}\,$ with coefficients given by certain ${\rm H}$-invariant tensors $\,\gamma_{\unl A\unl B\unl C}$.\ The latter super-3-cocycle is assumed to be a de Rham coboundary on the $\,\sigma_i(\mathcal{O}_i)\,$ in what follows\footnote{In this manner, we isolate the difficulty in the construction of the Green--Schwarz super-$\sigma$-model resulting from the assumption of (nonlinearly realised) supersymmetry from that associated with the (de Rham-)cohomological non-triviality of a generic super-3-cocycle.} and we further presuppose the corresponding primitives $\,\underset{\tx{\ciut{(2)}}}{{\rm B}}{}_i:={\mathsf d}^{-1}(\sigma_i^*\underset{\tx{\ciut{(3)}}}{\chi})\,$ to form under the induced action of the supersymmetry group a pseudo-invariant family in the sense of the relation
\begin{eqnarray}\label{eq:pseudolocB}
[\lambda]_g^*\underset{\tx{\ciut{(2)}}}{{\rm B}}{}_j(x)=\underset{\tx{\ciut{(2)}}}{{\rm B}}{}_i(x)+{\mathsf d}\underset{\tx{\ciut{(1)}}}{\Delta}{}^g_{ij}(x)
\end{eqnarray}
valid for all $\,(g,x)\in{\rm G}\x\mathcal{O}_i$,\ for $\,j\in I\,$ such that $\,[\lambda]_g(x)\in\mathcal{O}_j$,\ and for some $\,\underset{\tx{\ciut{(1)}}}{\Delta}{}^g_{ij}\in\Omega^1(\mathcal{O}_i)$,\ such that the action functional on the \emph{closed} worldsheet is effectively invariant under that action.
\section{Super-$\sigma$-model extensions of supertarget algebras by wrapping charges}\label{sec:wrapanom}
In the setting of the preceding section, with $\,\mathscr{M}\cong{\rm G}/{\rm H}\,$ realised within $\,{\rm G}\,$ and in the notation
\begin{eqnarray}\nn
\sigma({\rm G}/{\rm H}):=\bigsqcup_{i\in I}\,\sigma_i(\mathcal{O}_i)\,,
\end{eqnarray}
assume given a ${\rm H}$-basic representative
\begin{eqnarray}\nn
\underset{\tx{\ciut{(p+2)}}}{\chi}\mathord{\restriction}_{\sigma({\rm G}/{\rm H})}={\mathsf d}\underset{\tx{\ciut{(p+1)}}}{\beta}
\end{eqnarray}
of a class $\,[\underset{\tx{\ciut{(p+2)}}}{\chi}]\in{\rm CaE}^{p+2}({\rm G})\,$ of the Cartan--Eilenberg cohomology of the Lie supergroup $\,{\rm G}$ restricted to the image of $\,{\rm G}/{\rm H}\,$ within $\,{\rm G}$,\ and -- further -- that the ${\rm H}$-horizontal de Rham primitive $\,\underset{\tx{\ciut{(p+1)}}}{\beta}\,$ of $\,\underset{\tx{\ciut{(p+2)}}}{\chi}\mathord{\restriction}_{\sigma({\rm G}/{\rm H})}\,$ (whose very existence is part of our assumptions) is \emph{pseudo}-invariant with respect to the transformations \eqref{eq:cosetact} of $\,{\rm G}\,$ in the sense rendered precise by the identity
\begin{eqnarray}\nn
\pLie{\mathcal{K}_X}\underset{\tx{\ciut{(p+1)}}}{\beta}={\mathsf d}\underset{\tx{\ciut{(p)}}}{\Gamma_X}\,,
\end{eqnarray}
to be satisfied for the vector field $\,\mathcal{K}_X\,$ on $\,\sigma({\rm G}/{\rm H})\,$ with restrictions $\,\mathcal{K}_X\mathord{\restriction}_{\sigma_i(\mathcal{O}_i)}:=\mathcal{K}_{i\,X}\,$ associated with an arbitrary element $\,X\in\gt{g}$,\ and for some $\,\underset{\tx{\ciut{(p)}}}{\Gamma_X}\in\Omega^p(\sigma({\rm G}/{\rm H}))$.\ The specific representative of the whole class of super-$p$-forms defined by the latter condition that we use hereunder is a \emph{choice}. In order to be able to proceed with our analysis, we consider henceforth, for the sake of transparency, (super-)linear combinations $\,X\,$ of the generators of $\,\gt{g}\,$ in which the coefficient in front of a given generator has the same Gra\ss man parity as the generator itself. Clearly, this does not, in any manner, affect the validity of our conclusions drawn from the ensuing analysis as the coefficients may always be removed at any stage of the analysis, with the sole effect that commutators become supercommutators, {\it e.g.}, $\,[\varepsilon_1^{\widehat\a}\,Q_{\widehat\a},\varepsilon_2^{\widehat\beta}\,Q_{\widehat\beta}]=-\varepsilon_1^{\widehat\a}\,\varepsilon_2^{\widehat\beta}\,\{Q_{\widehat\a},Q_{\widehat\beta}\}$.\ In this notation, we find, as a consequence of the pseudo-invariance, that the super-$p$-forms
\begin{eqnarray}\nn
\underset{\tx{\ciut{(p)}}}{\a_{X_1,X_2}}:=\pLie{\mathcal{K}_{X_1}}\underset{\tx{\ciut{(p)}}}{\Gamma_{X_2}}-\pLie{\mathcal{K}_{X_2}}\underset{\tx{\ciut{(p)}}}{\Gamma_{X_1}}+\underset{\tx{\ciut{(p)}}}{\Gamma_{[X_1,X_2]}}\,,
\end{eqnarray}
with the $\,X_1,X_2\,$ given by the super-linear combinations described above, are closed, and so whenever $\,H^p_{\rm dR}({\rm G}/{\rm H})=0$,\ {\it i.e.}\footnote{We are invoking the Kostant Theorem of \Rcite{Kostant:1975}.}, whenever $\,H^p_{\rm dR}(\vert{\rm G}/{\rm H}\vert)=0$,\ there exist smooth super-$(p-1)$-forms $\,{\mathsf d}^{-1}\underset{\tx{\ciut{(p)}}}{\a_{X_1,X_2}}\,$ on the image of $\,{\rm G}/{\rm H}\,$ within $\,{\rm G}\,$ such that
\begin{eqnarray}\label{eq:gammaXdel}
\underset{\tx{\ciut{(p)}}}{\a_{X_1,X_2}}={\mathsf d}\bigl({\mathsf d}^{-1}\underset{\tx{\ciut{(p)}}}{\a_{X_1,X_2}}\bigr)\,.
\end{eqnarray}
Consider, next, the previously defined Green--Schwarz super-$\sigma$-model with $\,{\rm G}/{\rm H}\,$ as the supertarget and with $\,\underset{\tx{\ciut{(p+2)}}}{\chi}\,$ as the Green--Schwarz super-$(p+2)$-cocycle. The corresponding (pre)symplectic form, derived in the first-order formalism of Refs.\,\cite{Gawedzki:1972ms,Kijowski:1973gi,Kijowski:1974mp,Kijowski:1976ze,Szczyrba:1976,Kijowski:1979dj} and evaluated on Cauchy data $\,(\sigma_{i_\cdot}\circ\unl\xi_\cdot,\pi_\cdot)\,$ (a degenerate canonical pair, with $\,\pi_{\cdot\,\widehat\a}\equiv 0$) of a classical configuration along a Cauchy hypersurface $\,\mathscr{C}_p\subset\Omega_{p+1},\ \partial\mathscr{C}_p=\emptyset$, with restrictions $\,\sigma_{i_\tau}\circ\unl\xi_\tau\equiv\sigma_{i_\tau}\circ\xi\mathord{\restriction}_{\mathscr{C}_p\cap\tau}\,$ and $\,\pi_\tau\equiv\pi\mathord{\restriction}_{\mathscr{C}_p\cap\tau}$,\ reads
\begin{eqnarray}\label{eq:Omsip}\qquad\qquad
\Omega^{({\rm NG})}_{{\rm GS},p}[\unl\xi,\pi]=\sum_{\tau\in\gt{P}}\,\bigg[\int_{\mathscr{C}_p\cap\tau}\,\Vol(\mathscr{C}_p)\,\delta\bigl(\pi_{\tau\,\unl A}(\cdot)\,\theta^{\unl A}_{\rm L}\bigl(\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)\bigr)+\int_{\mathscr{C}_p\cap\tau}\,{\rm ev}^*\underset{\tx{\ciut{(p+2)}}}{\chi}\bigl(\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)\bigg]\,,
\end{eqnarray}
and so we find the canonical lifts of the fundamental vector fields
\begin{eqnarray}\nn
\widetilde\mathcal{K}_X[\unl\xi,\pi]=\sum_{\tau\in\gt{P}}\,\int_{\mathscr{C}_p\cap\tau}\,\Vol(\mathscr{C}_p)\,\bigl[\mathcal{K}_X\bigl(\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)+\Delta_{\tau\,\unl A}\bigl(X;\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)\,\tfrac{\delta\ }{\delta\pi_{\tau\,\unl A}}(\cdot)\bigr]\,,
\end{eqnarray}
with the correction $\,\Delta_{\tau\,\unl A}\,$ determined by the condition
\begin{eqnarray}\label{eq:canlifunK}
\pLie{\widetilde\mathcal{K}_X}\vartheta\stackrel{!}{=} 0\,,\qquad\qquad\vartheta[\unl\xi,\pi]:=\sum_{\tau\in\gt{P}}\,\int_{\mathscr{C}_p\cap\tau}\,\Vol(\mathscr{C}_p)\,\pi_{\tau\,\unl A}(\cdot)\,\theta^{\unl A}_{\rm L}\bigl(\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)\,.
\end{eqnarray}
We have
\begin{Prop}\label{prop:canlifunK}
The canonical lift defined by \Reqref{eq:canlifunK} admits a solution
\begin{eqnarray}\nn
\Delta_{\tau\,\unl A}\bigl(X;\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)=\pi_{\tau\,\unl B}(\cdot)\,\bigl({\rm ad}_{Y_{i_\tau}(X;\unl\xi_\tau(\cdot))}\bigr)^{\unl B}_{\ \unl A}\,.
\end{eqnarray}
\end{Prop}
\noindent\begin{proof}
A proof is given in App.\,\ref{app:canlifunK}.
\end{proof}
The Noether hamiltonians (charges) corresponding to the above-defined lifts read
\begin{eqnarray}
&&h_X[\unl\xi,\pi]\cr\cr
&=&\sum_{\tau\in\gt{P}}\,\bigg[\int_{\mathscr{C}_p\cap\tau}\,\Vol(\mathscr{C}_p)\,\pi_{\tau\,\unl A}(\cdot)\,\bigl(\mathcal{K}_X\righthalfcup\theta^{\unl A}_{\rm L}\bigr)\bigl(\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)+\int_{\mathscr{C}_p\cap\tau}\,{\rm ev}^*\bigl(\mathcal{K}_X\righthalfcup\underset{\tx{\ciut{(p+1)}}}{\beta}-\underset{\tx{\ciut{(p)}}}{\Gamma_X}\bigr)\bigl(\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)\bigg]\cr\cr
&\equiv&\sum_{\tau\in\gt{P}}\,\bigg[\int_{\mathscr{C}_p\cap\tau}\,\Vol(\mathscr{C}_p)\,\pi_{\tau\,\unl A}(\cdot)\,\bigl(\mathcal{R}_X\righthalfcup\theta^{\unl A}_{\rm L}\bigr)\bigl(\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)+\int_{\mathscr{C}_p\cap\tau}\,{\rm ev}^*\bigl(\mathcal{R}_X\righthalfcup\underset{\tx{\ciut{(p+1)}}}{\beta}-\underset{\tx{\ciut{(p)}}}{\Gamma_X}\bigr)\bigl(\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)\bigg]\,.\cr && \label{eq:Noesusy}
\end{eqnarray}
As we want these hamiltonians to be well-defined, independently of the arbitrary choices of the representatives $\,\underset{\tx{\ciut{(p+1)}}}{\beta}\,$ and $\,\underset{\tx{\ciut{(p)}}}{\Gamma_X}\,$ of the respective classes of superdifferential forms, we should replace the latter with
\begin{eqnarray}\nn
\underset{\tx{\ciut{(p)}}}{\Gamma_X}\longmapsto\underset{\tx{\ciut{(p)}}}{\Gamma_X}+\mathcal{K}_X\righthalfcup\underset{\tx{\ciut{(p+1)}}}{\Delta\beta}
\end{eqnarray}
whenever the former is replaced by
\begin{eqnarray}\nn
\underset{\tx{\ciut{(p+1)}}}{\beta}\longmapsto\underset{\tx{\ciut{(p+1)}}}{\beta}+\underset{\tx{\ciut{(p+1)}}}{\Delta\beta}\,,
\end{eqnarray}
with, of necessity,
\begin{eqnarray}\nn
\underset{\tx{\ciut{(p+1)}}}{\Delta\beta}\in Z^{p+1}_{\rm dR}({\rm G})\,.
\end{eqnarray}
In general, the hamiltonians are \emph{not} in involution under the Poisson bracket induced by $\,\Omega^{({\rm NG})}_{{\rm GS},p}$.\ Instead, we establish
\begin{Prop}\label{prop:PoiNoe}
The Poisson bracket of the Noether charges \eqref{eq:Noesusy} associated with the (pre)symplectic form \eqref{eq:Omsip} is given by the formula
\begin{eqnarray}\nn
&&\{h_{X_1},h_{X_2}\}_{\Omega^{({\rm NG})}_{{\rm GS},p}}[\unl\xi,\pi]\cr\cr
&=&h_{-[X_1,X_2]}[\unl\xi,\pi]+\sum_{\tau\in\gt{P}}\,\bigg[\int_{\mathscr{C}_p\cap\tau}\,{\rm ev}^*{\mathsf d}\bigl(\mathcal{K}_{X_1}\righthalfcup\underset{\tx{\ciut{(p)}}}{\Gamma_{X_2}}-\mathcal{K}_{X_2}\righthalfcup\underset{\tx{\ciut{(p)}}}{\Gamma_{X_1}}+\mathcal{K}_{X_2}\righthalfcup\mathcal{K}_{X_1}\righthalfcup\underset{\tx{\ciut{(p+1)}}}{\beta}\bigr)\bigl(\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)\cr\cr
&&-\int_{\mathscr{C}_p\cap\tau}\,{\rm ev}^*\underset{\tx{\ciut{(p)}}}{\a_{X_1,X_2}}\bigl(\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)\bigg]\,.
\end{eqnarray}
\end{Prop}
\noindent\begin{proof}
A proof is given in App.\,\ref{app:PoiNoe}.
\end{proof}
From the above, we read off possible \emph{classical} field-theoretic corrections to the supertarget Lie superalgebra $\,\gt{g}$:\ First of all, the corrections vanish \emph{on the level of the Poisson algebra of Noether charges}\footnote{They survive on the level of the underlying Poisson algebra of Noether currents.} for all field configurations with a vanishing monodromy around the embedded Cauchy $p$-cycle $\,\bigcup_{\tau\in\gt{P}}\,\sigma_{i_\tau}\circ\unl\xi_\tau(\tau\cap\mathscr{C}_p)$.\ In particular, they are -- apparently -- absent whenever $\,H_p({\rm G}/{\rm H})=0$.\ Thus, the corrections are sourced by field configurations wrapping non-contractible $p$-cycles in $\,{\rm G}/{\rm H}\,$ -- the charges are none other than the wrapping ({\it e.g.}, winding) numbers of the classical configuration $\,\unl\xi$.\ Secondly, we may engineer such corrections without affecting the topology of the body $\,\vert{\rm G}/{\rm H}\vert\,$ of the supertarget $\,{\rm G}/{\rm H}\,$ by considering the Green--Schwarz super-$\sigma$-model -- in conformity with the interpretation of the Chevalley--Eilenberg cohomology proposed by Rabin and Crane in \Rcite{Rabin:1984rm} -- as a model of super-$p$-brane dynamics on the quotient of $\,{\rm G}/{\rm H}\,$ by the discrete Kosteleck\'y--Rabin supersymmetry group of \Rcite{Kostelecky:1983qu}, and -- consequently -- by taking into account field configurations with a non-vanishing integral monodromy in the Gra\ss mann-odd directions, {\it i.e.}, the twisted sector of the super-$\sigma$-model on the quotient. In what follows, this will serve to demonstrate, quite naturally, the wrapping nature of the charges that define the physically relevant deformations ({\it e.g.}, super-central extensions) of the supertarget algebras on $\,{\rm sMink}^{d,1\,\vert\,D_{d,1}}\,$ and $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,$ encoded by the Green--Schwarz super-$(p+2)$-cocycles, and through these -- also the associated Cartan--Eilenberg super-$p$-gerbes. To this end, we shall study at length the specific examples in which the said super-central extensions are known ({\it cp} Part I) to integrate to surjective submersions of the super-$p$-gerbes, {\it i.e.}, those with $\,p\in\{0,1\}\,$ on $\,{\rm sMink}^{d,1\,\vert\,D_{d,1}}$.\ We shall also consider, from this point of view, geometrisations of the Metsaev--Tseytlin super-3-cocycle defining the superstring model on $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)$.\ In so doing, we shall employ the computationally less cumbersome\footnote{Note that we merely have to compute the supersymmetry-variation super-$p$-forms $\,\Gamma_X\,$ for $\,X\in[\gt{g},\gt{g}\}$.} formula for the wrapping anomaly:
\begin{eqnarray}\nn
\mathscr{W}_{X_1,X_2}[\unl\xi]&:=&\sum_{\tau\in\gt{P}}\,\int_{\mathscr{C}_p\cap\tau}\,{\rm ev}^*\bigl({\mathsf d}\bigl(\mathcal{K}_{X_1}\righthalfcup\underset{\tx{\ciut{(p)}}}{\Gamma_{X_2}}-\mathcal{K}_{X_2}\righthalfcup\underset{\tx{\ciut{(p)}}}{\Gamma_{X_1}}+\mathcal{K}_{X_2}\righthalfcup\mathcal{K}_{X_1}\righthalfcup\underset{\tx{\ciut{(p+1)}}}{\beta}\bigr)-\underset{\tx{\ciut{(p)}}}{\a_{X_1,X_2}}\bigr)\bigl(\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)\cr\cr
&=&\sum_{\tau\in\gt{P}}\,\int_{\mathscr{C}_p\cap\tau}\,{\rm ev}^*\bigl(\bigl(\pLie{\mathcal{K}_{X_1}}\underset{\tx{\ciut{(p)}}}{\Gamma_{X_2}}-\pLie{\mathcal{K}_{X_2}}\underset{\tx{\ciut{(p)}}}{\Gamma_{X_1}}-\underset{\tx{\ciut{(p)}}}{\a_{X_1,X_2}}+\underset{\tx{\ciut{(p)}}}{\Gamma_{[X_1,X_2]}}\bigr)\cr\cr
&&\hspace{-.5cm}-\mathcal{K}_{X_1}\righthalfcup\pLie{\mathcal{K}_{X_2}}\underset{\tx{\ciut{(p+1)}}}{\beta}+\mathcal{K}_{X_2}\righthalfcup\pLie{\mathcal{K}_{X_1}}\underset{\tx{\ciut{(p+1)}}}{\beta}+{\mathsf d}\bigl(\mathcal{K}_{X_2}\righthalfcup\mathcal{K}_{X_1}\righthalfcup\underset{\tx{\ciut{(p+1)}}}{\beta}\bigr)-\underset{\tx{\ciut{(p)}}}{\Gamma_{[X_1,X_2]}}\bigr)\bigl(\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)\cr\cr
&=&\sum_{\tau\in\gt{P}}\,\int_{\mathscr{C}_p\cap\tau}\,{\rm ev}^*\bigl(-\mathcal{K}_{X_1}\righthalfcup\pLie{\mathcal{K}_{X_2}}\underset{\tx{\ciut{(p+1)}}}{\beta}+\mathcal{K}_{X_2}\righthalfcup\pLie{\mathcal{K}_{X_1}}\underset{\tx{\ciut{(p+1)}}}{\beta}+\pLie{\mathcal{K}_{X_2}}\bigl(\mathcal{K}_{X_1}\righthalfcup\underset{\tx{\ciut{(p+1)}}}{\beta}\bigr)\cr\cr
&&-\mathcal{K}_{X_2}\righthalfcup\pLie{\mathcal{K}_{X_1}}\underset{\tx{\ciut{(p+1)}}}{\beta}+\mathcal{K}_{X_2}\righthalfcup\mathcal{K}_{X_1}\righthalfcup\underset{\tx{\ciut{(p+2)}}}{\chi}-\underset{\tx{\ciut{(p)}}}{\Gamma_{[X_1,X_2]}}\bigr)\bigl(\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)\cr\cr
&=&\sum_{\tau\in\gt{P}}\,\int_{\mathscr{C}_p\cap\tau}\,{\rm ev}^*\bigl(\mathcal{K}_{X_2}\righthalfcup\mathcal{K}_{X_1}\righthalfcup\underset{\tx{\ciut{(p+2)}}}{\chi}+\mathcal{K}_{[X_1,X_2]}\righthalfcup\underset{\tx{\ciut{(p+1)}}}{\beta}-\underset{\tx{\ciut{(p)}}}{\Gamma_{[X_1,X_2]}}\bigr)\bigl(\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)\,.
\end{eqnarray}
Taking into account the ${\rm H}$-horizontality of the forms involved, we may further reduce the above to
\begin{eqnarray}\nn
\mathscr{W}_{X_1,X_2}[\unl\xi]=\sum_{\tau\in\gt{P}}\,\int_{\mathscr{C}_p\cap\tau}\,{\rm ev}^*\bigl(\mathcal{R}_{X_2}\righthalfcup\mathcal{R}_{X_1}\righthalfcup\underset{\tx{\ciut{(p+2)}}}{\chi}+\mathcal{R}_{[X_1,X_2]}\righthalfcup\underset{\tx{\ciut{(p+1)}}}{\beta}-\underset{\tx{\ciut{(p)}}}{\Gamma_{[X_1,X_2]}}\bigr)(\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)\,.
\end{eqnarray}
It deserves to be emphasised that the anomaly thus defined does \emph{not} depend on the choice of the primitive $\,\underset{\tx{\ciut{(p+1)}}}{\beta}\,$ of the Green--Schwarz super-$(p+2)$-cocycle $\,\underset{\tx{\ciut{(p+2)}}}{\chi}\,$ and of the corresponding variance super-$p$-form $\,\underset{\tx{\ciut{(p)}}}{\Gamma_X}\,$ for precisely the same reason as that for the well-definedness of the hamiltonians.
Let us extract from the above formula the contribution to the wrapping anomaly of a \emph{supersymmetric} (component of a) primitive $\,\underset{\tx{\ciut{(p+1)}}}{\beta}\,$ obtained through restriction from a super-$(p+1)$-form on $\,{\rm G}$,\ which we write in the form
\begin{eqnarray}\nn
\underset{\tx{\ciut{(p+1)}}}{\beta}=\theta_{\rm L}^{A_1}\wedge\theta_{\rm L}^{A_2}\wedge\cdots\wedge\theta_{\rm L}^{A_{p+1}}\,b_{A_1 A_2\ldots A_{p+1}}\mathord{\restriction}_{\sigma({\rm G}/{\rm H})}\,,
\end{eqnarray}
with functional coefficients $\,b_{A_1 A_2\ldots A_{p+1}}\,$ (of Gra\ss mann parity $\,\sum_{k=1}^{p+1}\,|A_k|$). As
\begin{eqnarray}\nn
\forall_{X\in\gt{g}} \ :\ \pLie{\mathcal{K}_X}\underset{\tx{\ciut{(p+1)}}}{\beta}=0\,,
\end{eqnarray}
we may take
\begin{eqnarray}\nn
\underset{\tx{\ciut{(p)}}}{\Gamma_X}=0=\underset{\tx{\ciut{(p)}}}{\a_{X_1,X_2}}\,.
\end{eqnarray}
We then obtain
\begin{eqnarray}\label{eq:wrappaninvB}
\mathscr{W}^{({\rm inv})}_{X_1,X_2}[\unl\xi]=\sum_{\tau\in\gt{P}}\,\int_{\mathscr{C}_p\cap\tau}\,{\rm ev}^*{\mathsf d}\bigl(\mathcal{K}_{X_2}\righthalfcup\mathcal{K}_{X_1}\righthalfcup\underset{\tx{\ciut{(p+1)}}}{\beta}\bigr)\bigl(\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)
\end{eqnarray}
Upon invoking \Reqref{eq:leftvarfiel}, we may render the last expression even more explicit, to wit,
\begin{eqnarray}\nn
&&\mathscr{W}_{X_1,X_2}^{({\rm inv})}[\unl\xi]\cr\cr
&=&p(p+1)\,\sum_{\tau\in\gt{P}}\,\int_{\mathscr{C}_p\cap\tau}\,{\rm ev}^*{\mathsf d}\bigl(\bigl(\mathcal{K}_{X_1}\righthalfcup\theta_{\rm L}^A\bigr)\,\bigl(\mathcal{K}_{X_2}\righthalfcup\theta_{\rm L}^B\bigr)\,\theta_{\rm L}^{A_1}\wedge\theta_{\rm L}^{A_2}\wedge\cdots\wedge\theta_{\rm L}^{A_{p-1}}\,b_{A B A_1 A_2\ldots A_{p-1}}\bigr)\bigl(\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)\cr\cr
&=&p(p+1)\,\sum_{\tau\in\gt{P}}\,\int_{\mathscr{C}_p\cap\tau}\,{\rm ev}^*{\mathsf d}\bigl(\bigl(\delta_{X_1}Z^{\unl A}\,E_{\unl A}^A\bigr)\,\bigl(\delta_{X_2}Z^{\unl B}\,E_{\unl B}^B\bigr)\,\theta_{\rm L}^{A_1}\wedge\theta_{\rm L}^{A_2}\wedge\cdots\wedge\theta_{\rm L}^{A_{p-1}}\,b_{A B A_1 A_2\ldots A_{p-1}}\bigr)\bigl(\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)\,.
\end{eqnarray}
Thus, whenever $\,\underset{\tx{\ciut{(p+1)}}}{\beta}\,$ is ${\rm H}$-horizontal,
\begin{eqnarray}\label{eq:Binvhor}
\underset{\tx{\ciut{(p+1)}}}{\beta}=\theta_{\rm L}^{\unl A_1}\wedge\theta_{\rm L}^{\unl A_2}\wedge\cdots\wedge\theta_{\rm L}^{\unl A_{p+1}}\,b_{\unl A_1\unl A_2\ldots\unl A_{p+1}}\mathord{\restriction}_{\sigma({\rm G}/{\rm H})}\,,
\end{eqnarray}
with, this time, constant ${\rm H}$-invariant tensors $\,b_{\unl A_1\unl A_2\ldots\unl A_{p+1}}\,$ as coefficients,
we have
\begin{eqnarray}\nn
\mathscr{W}_{X_1,X_2}^{({\rm inv})}[\unl\xi]&=&p(p+1)\,\sum_{\tau\in\gt{P}}\,\int_{\mathscr{C}_p\cap\tau}\,{\rm ev}^*{\mathsf d}\bigl({\mathsf T}_e{\rm Ad}_{\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)^{-1}}(X_1)^{\unl A}\,{\mathsf T}_e{\rm Ad}_{\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)^{-1}}(X_2)^{\unl B}\cr\cr
&&\hspace{3.75cm}\theta_{\rm L}^{\unl A_1}\wedge\theta_{\rm L}^{\unl A_2}\wedge\cdots\wedge\theta_{\rm L}^{\unl A_{p-1}}\bigl(\sigma_{i_\tau}\circ\unl\xi_\tau(\cdot)\bigr)\,b_{\unl A\unl B\unl A_1\unl A_2\ldots\unl A_{p-1}}\bigr)\,.
\end{eqnarray}
A word of comment is due at this point. Given that the $\,\mathcal{K}_{i\,X_\a},\ \a\in\{1,2\}\,$ are the local pushforwards \eqref{eq:KXaspfXiX} of the corresponding fundamental vector fields $\,\Xi_{X_\a}\,$ on $\,{\rm G}/{\rm H}\,$ and that the primitive $\,\underset{\tx{\ciut{(p+1)}}}{\beta}\,$ of \Reqref{eq:Binvhor} is the pullback, along $\,\pi_{{\rm G}/{\rm H}}$,\ of a globally smooth super-$(p+1)$-form $\,\underset{\tx{\ciut{(p+1)}}}{{\rm B}}\,$ on $\,{\rm G}/{\rm H}$,\ we readily see that the local super-$(p-1)$-form
\begin{eqnarray}\nn
\sigma_i^*\bigl(\mathcal{K}_{X_2}\righthalfcup\mathcal{K}_{X_1}\righthalfcup\underset{\tx{\ciut{(p+1)}}}{\beta}\bigr)
\end{eqnarray}
is, in fact, a (locally continuous) restriction of the super-$(p-1)$-form
\begin{eqnarray}\nn
{\rm B}_{X_1,X_2}\equiv\Xi_{X_2}\righthalfcup\Xi_{X_1}\righthalfcup\underset{\tx{\ciut{(p+1)}}}{{\rm B}}
\end{eqnarray}
to $\,\mathcal{O}_i$.\ This observation justifies our former identification of the anomaly $\,\mathscr{W}_{X_1,X_2}^{({\rm inv})}\,$ as the wrapping charge trapped by a non-contractible Cauchy $p$-cycle $\,\xi(\mathscr{C}_p)\,$ to which the latter super-$(p-1)$-form couples, or -- more formally -- the monodromy of $\,{\rm B}_{X_1,X_2}\,$ along $\,\xi(\mathscr{C}_p)$.
In particular for $\,p=1\,$ we readily establish that the presence of the wrapping anomaly indicates the physical relevance of a deformation of the original Lie superalgebra $\,\gt{g}\,$ of the general structure
\begin{eqnarray}\label{eq:wrapcentrext}
[t_A,t_B\}^\sim=f_{AB}^{\ \ C}\,t_C+Z_{AB}\,,\qquad\qquad Z_{BA}=-(-1)^{|A|\cdot|B|}\,Z_{AB}\qquad A,B\in\ovl{1,{\rm dim}\,\gt{g}}\,,
\end{eqnarray}
where the (wrapping) charges $\,Z_{AB}\,$ are induced by the monodromies
\begin{eqnarray}\label{eq:monofAB}
\mu\bigl(f_{AB};\xi(\mathscr{C}_1)\bigr)
\end{eqnarray}
around $\,\xi(\mathscr{C}_1),\ \mathscr{C}_1\cong{\mathbb{S}}^1$,\ of functions $\,f_{AB}\,$ with local representations
\begin{eqnarray}\label{eq:fAB}
f_{AB}\mathord{\restriction}_{\mathcal{O}_i}\equiv(-1)^{|\unl A|\cdot |B|+1}\,2\bigl({\mathsf T}_e{\rm Ad}_{\sigma_i(\cdot)^{-1}}\bigr)^{\unl A}_{\ A}\,\bigl({\mathsf T}_e{\rm Ad}_{\sigma_i(\cdot)^{-1}}\bigr)^{\unl B}_{\ B}\,b_{\unl A\unl B}\bigl(\sigma_i(\cdot)\bigr)\,.
\end{eqnarray}
The deformation may now take on the form of a super-central or some more general \emph{associative}\footnote{We do not consider non-associative deformations in which the super-Jacobi identity might fail. The lack of associativity renders the Cartan calculus, central to our considerations, ill-defined. While the alternative of working with an algebra loop with inverses seems an interesting theoretical possibility, we do not consider it in what follows, and, instead, \emph{impose} the super-Jacobi identity in the deformed algebra with the given, physically motivated `germ'.} extension of $\,\gt{g}\,$ with the `germ' \eqref{eq:wrapcentrext}. The idea behind it, advocated in \Rcite{Suszek:2017xlw}, is a trivialisation, in the spirit of Prop.\,I.C.4, of the class $\,[\varpi_p]\in{\rm CE}^2({\rm G})\,$ in the Chevalley--Eilenberg cohomology associated with it as per
\begin{eqnarray}\nn
\varpi_p(\mathcal{L}_{X_1},\mathcal{L}_{X_2})=\mathscr{W}_{X_1,X_2}\,.
\end{eqnarray}
This is attained on the Lie supergroup $\,\widetilde{\rm G}\,$ that integrates the new Lie superalgebra with the vector-space structure
\begin{eqnarray}\nn
\widetilde\gt{g}:=\gt{g}\oplus\corr{Z_{AB}\,\vert\,A,B\in\ovl{1,{\rm dim}\,\gt{g}}}_{\mathbb{C}}\equiv\gt{g}\oplus\gt{z}
\end{eqnarray}
and a Lie superbracket
\begin{eqnarray}\nn
[\cdot,\cdot\}^\sim\ :\ \widetilde\gt{g}\x\widetilde\gt{g}\longrightarrow\widetilde\gt{g}
\end{eqnarray}
determined (whenever possible, and then typically non-uniquely) by the imposition of the super-Jacobi constraints on the superalgebra with the `germ' specified. Given the straightforward geometric and physical interpretation of the Lie subalgebra $\,{\rm H}\,$ which we want to preserve under the extension, we are led to constrain the admissible deformations so that the commutation relations of its elements with the rest of $\,\gt{g}\subset\widetilde\gt{g}\,$ remain unchanged,
\begin{eqnarray}\nn
\forall_{(\kappa,A)\in\ovl{1,{\rm dim}\,\gt{h}}\x\ovl{1,{\rm dim}\,\gt{g}}}\ :\ [J_\kappa,t_A]^\sim=f_{\kappa A}^{\ \ \ B}\,t_B\,.
\end{eqnarray}
Moreover, we take the charges $\,Z_{AB}\,$ to transform linearly under the isotropy group $\,{\rm H}$,\ as do the remaining generators. In particular, $\,\gt{z}\,$ is assumed to be an ${\rm ad}$-module of the Lie algebra $\,\gt{h}$,
\begin{eqnarray}\nn
[\gt{h},\gt{z}]^\sim\subset\gt{z}\,,
\end{eqnarray}
so that the new decomposition
\begin{eqnarray}\nn
\widetilde\gt{g}=\widetilde\gt{t}\oplus\gt{h}\,,\qquad\qquad\widetilde\gt{t}\equiv\gt{t}\oplus\gt{z}
\end{eqnarray}
is reductive just as the original one. Accordingly, we may repeat the previous constructions over the new Lie supergroup which we, once again, regard as the total space of the principal ${\rm H}$-bundle
\begin{eqnarray}\nn
\alxydim{@C=1cm@R=1cm}{{\rm H} \ar[r] & \widetilde{\rm G} \ar[d]^{\pi_{\widetilde{\rm G}/{\rm H}}} \\ & \widetilde{\rm G}/{\rm H}}\,.
\end{eqnarray}
Its base, given by the homogeneous space $\,\widetilde{\rm G}/{\rm H}$,\ is locally charted by flows of the left-invariant vector fields on $\,\widetilde{\rm G}\,$ along $\,\widetilde\gt{t}$,\ whence the coordinates $\,(Z_i^{\unl A},\zeta_i^{AB})$,\ where the $\,Z_i^{\unl A}\,$ are as before and where the new ones, $\,\zeta_i^{AB}\equiv-(-1)^{|A|\cdot|B|}\,\zeta_i^{BA}$,\ of the same Gra\ss mann parity as the corresponding (independent) charges, $\,|\zeta_i^{AB}|\equiv |Z_{AB}|=|A|+|B|\,$ correspond to the left-invariant vector fields from $\,\gt{z}$.\ It is realised within $\,\widetilde{\rm G}\,$ by means of a collection of the distinguished local sections (the last sum is over the independent charges)
\begin{eqnarray}\nn
\widetilde\sigma_i\ :\ \widetilde\mathcal{O}_i\longrightarrow\widetilde{\rm G}\ :\ \widetilde Z_i\equiv(Z_i,\zeta)\longmapsto\sigma_i(Z_i)\cdot {\rm e}^{\zeta_i^{AB}\,Z_{AB}}\,.
\end{eqnarray}
Here, by a mild abuse of the notation, we take the $\,\sigma_i\,$ to denote the same products of flows of right-invariant vector fields (along $\,\gt{t}^{(0)}\,$ and $\,\gt{t}^{(1)}$) as before, but with the understanding that they are now subject to the deformed structure equations ({\it i.e.}, $\,\gt{t}^{(0)},\gt{t}^{(1)}\subset\widetilde\gt{g}$) and initiate at some $\,\widetilde{\unl g}_i\in\widetilde{\rm G}$.\ The choice of the embedding sections is the first step towards a reconstruction, in the same spirit as over $\,{\rm G}/{\rm H}$,\ of the differential calculus over the new homogeneous space $\,\widetilde{\rm G}/{\rm H}$.\ Inspection of the Schur--Poincar\'e formula for the fundamental Maurer--Cartan super-1-form,
\begin{eqnarray}\nn
\widetilde\sigma_i^*\theta_{\rm L}(\widetilde Z_i)&=&\sum_{k,l=0}^\infty\,\tfrac{(-1)^{k+l}}{k!(l+1)!}\,\bigl({\rm id}_{{\mathsf T}^*\mathcal{O}_i}\otimes\widetilde{\rm ad}^k_{\zeta_i^{AB}\,Z_{AB}}\bigr)\bigl({\mathsf d} X_i^{\widehat a}\otimes\sum_{m=0}^\infty\,\tfrac{(-1)^m}{m!}\,\widetilde{\rm ad}^m_{\Theta_i^{\widehat\beta}(Z_i)\,Q_{\widehat\beta}}\circ\widetilde{\rm ad}^l_{X_i^{\widehat b}\,P_{\widehat b}}(P_{\widehat a})\cr\cr
&&+{\mathsf d}\Theta_i^{\widehat\a}(Z_i)\otimes\widetilde{\rm ad}^l_{\Theta_i^{\widehat\beta}(Z_i)\,Q_{\widehat\beta}}(Q_{\widehat\a})\bigr)+\sum_{m=0}^\infty\,\tfrac{(-1)^m}{(m+1)!}\,{\mathsf d}\zeta_i^{AB}\otimes\widetilde{\rm ad}_{\zeta_i^{CD}\,Z_{CD}}^m(Z_{AB})\,,
\end{eqnarray}
in conjunction with the Baker--Campbell--Dynkin--Hausdorff formula, helps to determine the admissible structure of the deformation $\,\widetilde\gt{g}$,\ with the `germ' constrained by the foregoing analysis in the form
\begin{eqnarray}\nn
&\{Q_{\widehat\a},Q_{\widehat\beta}\}^\sim=f_{\widehat\a\widehat\beta}^{\ \ \ c}\,B_c+Z_{\widehat\a\widehat\beta}\,,\qquad\qquad[Q_{\widehat\a},P_{\widehat a}]^\sim= f_{\widehat\a\widehat a}^{\ \ \ \widehat\beta}\,Q_{\widehat\beta}+Z_{\widehat\a\widehat a}\,,\qquad\qquad[P_{\widehat a},P_{\widehat b}]^\sim=f_{\widehat a\widehat b}^{\ \ \ c}\,B_c+Z_{\widehat a\widehat b}\,,&\cr\cr
&[J_\kappa,Z_{\widehat\a\widehat\beta}]^\sim=\Lambda_{\kappa\,\widehat\a\widehat\beta}^{\ \ \ \ \widehat\gamma\widehat\delta}\,Z_{\widehat\gamma\widehat\delta}\,,\qquad\qquad[J_\kappa,Z_{\widehat\a\widehat a}]^\sim=\Lambda_{\kappa\,\widehat\a\widehat a}^{\ \ \ \ \widehat\beta\widehat b}\,Z_{\widehat\beta\widehat b}&
\end{eqnarray}
for some $\,\Lambda_{\kappa\,\widehat\a\widehat\beta}^{\ \ \ \widehat\gamma\widehat\delta},\Lambda_{\kappa\,\widehat\a\widehat a}^{\ \ \ \widehat\beta\widehat b}\in{\mathbb{C}}$,\ where $\,Z_{\widehat\a\widehat\beta}\equiv Z_{\widehat\beta\widehat\a},Z_{\widehat a\widehat b}=-Z_{\widehat b\widehat a}\in\gt{z}^{(0)}\,$ and $\,Z_{\widehat\a\widehat a}\in\gt{z}^{(1)}$,\ and where all other supercommutators are as in $\,\gt{g}$.\ Indeed, a deformation $\,\widetilde\gt{g}\,$ with $\,[\widetilde\gt{g},\gt{z}\}^\sim\cap\gt{g}\neq 0\,$ leads to an alteration of the coordinate expressions for the components of the Maurer--Cartan super-1-form along $\,\gt{t}\,$ entering the definition of the original physical model, and also for the induced supersymmetry on $\,{\rm G}/{\rm H}$,\ which is physically untenable, given that it is the canonical analysis of the original model that sources the deformation in the first place. Therefore, we are led to assume that the deformation $\,\widetilde\gt{g}\,$ is, in fact, \textbf{normal} in the sense expressed by the identities
\begin{eqnarray}\nn
[\widetilde\gt{g},\gt{z}\}^\sim\subset\gt{z}\,.
\end{eqnarray}
Let us denote the independent generators of the supervector space
\begin{eqnarray}\nn
\gt{z}=\gt{z}^{(0)}\oplus\gt{z}^{(1)}
\end{eqnarray}
as
\begin{eqnarray}\nn
\gt{z}^{(0)}=\bigoplus_{\widetilde a=1}^{{\rm dim}\,\gt{z}^{(0)}}\,\corr{Z_{\widetilde a}}_{\mathbb{C}}\,,\qquad\qquad\gt{z}^{(1)}=\bigoplus_{\widetilde\a=1}^{{\rm dim}\,\gt{z}^{(1)}}\,\corr{Z_{\widetilde\a}}_{\mathbb{C}}\,.
\end{eqnarray}
The Lie super-brackets of the normal extension $\,\widetilde\gt{g}\,$ of $\,\gt{g}\,$ may now be cast in the form
\begin{eqnarray}\nn
&[P_{\widehat a},P_{\widehat b}]^\sim=f_{\widehat a\widehat b}^{\ \ \widehat c}\,P_{\widehat c}+f_{\widehat a\widehat b}^{\ \ \kappa}\,J_\kappa+Z_{\widehat a\widehat b}\,,\qquad\qquad[J_\kappa,J_\lambda]^\sim=f_{\kappa\lambda}^{\ \ \mu}\,J_\mu\,,&\cr\cr
&[J_\kappa,P_{\widehat a}]^\sim=f_{\kappa\widehat a}^{\ \ \widehat b}\,P_{\widehat b}\,,\qquad\qquad[J_\kappa,Q_{\widehat\a}]^\sim=f_{\kappa\widehat\a}^{\ \ \widehat\beta}\,Q_{\widehat\beta}\,,&\cr\cr
&\{Q_{\widehat\a},Q_{\widehat\beta}\}^\sim=f_{\widehat\a\widehat\beta}^{\ \ c}\,B_c+\Delta_{\widehat\a\widehat\beta}^{\ \ \ \widetilde a}\,Z_{\widetilde a}\,,\qquad\qquad[Q_{\widehat\a},P_{\widehat a}]^\sim= f_{\widehat\a\widehat a}^{\ \ \widehat\beta}\,Q_{\widehat\beta}+\Delta_{\widehat\a\widehat a}^{\ \ \ \widetilde\beta}\,Z_{\widetilde\beta}\,,&\cr\cr
&[Q_{\widehat\a},Z_{\widetilde a}]^\sim=\Delta_{\widehat\a\widetilde a}^{\ \ \ \widetilde\beta}\,Z_{\widetilde\beta}\,,\qquad\qquad[P_{\widehat a},Z_{\widetilde b}]^\sim=\Delta_{\widehat a\widetilde b}^{\ \ \ \widetilde c}\,Z_{\widetilde c}\,,&\cr\cr
&\{Q_{\widehat\a},Z_{\widetilde\beta}\}^\sim=\Delta_{\widehat\a\,\widetilde\beta}^{\ \ \ \widetilde c}\,Z_{\widetilde c}\,,\qquad\qquad[P_{\widehat a},Z_{\widetilde\a}]^\sim=\Delta_{\widehat a\widetilde\a}^{\ \ \ \widetilde\beta}\,Z_{\widetilde\beta}\,,&\cr\cr
&[J_\kappa,Z_{\widetilde a}]^\sim=\Delta_{\kappa\widetilde a}^{\ \ \ \widetilde b}\,Z_{\widetilde b}\,,\qquad\qquad[J_\kappa,Z_{\widetilde\a}]^\sim=\Delta_{\kappa\widetilde\a}^{\ \ \ \widetilde\beta}\,Z_{\widetilde\beta}&
\end{eqnarray}
for some $\,\Delta_{\widehat\a\widehat\beta}^{\ \ \ \widetilde a},\Delta_{\widehat\a\widehat a}^{\ \ \ \widetilde\beta},\Delta_{\widehat\a\widetilde a}^{\ \ \ \widetilde\beta},\Delta_{\widehat a\widetilde b}^{\ \ \ \widetilde c},\Delta_{\widehat\a\widetilde\beta}^{\ \ \ \widetilde c},\Delta_{\widehat a\widetilde\a}^{\ \ \ \widetilde\beta},\Delta_{\kappa\widetilde a}^{\ \ \ \widetilde b},\Delta_{\kappa\widetilde\a}^{\ \ \ \widetilde\beta}\in{\mathbb{C}}$.\ Upon calculating the relevant commutators in the previous formula for the Maurer--Cartan super-1-form, we find the general structure
\begin{eqnarray}
\widetilde\sigma_i^*\theta_{\rm L}(Z_i,\zeta)-\sigma_i^*\theta_{\rm L}^A(Z_i)\otimes t_A&=&\bigl(e_{\widehat a}^{\ \widetilde a}(Z_i,\zeta_i)\,{\mathsf d} X_i^{\widehat a}+e_{\widehat\beta}^{\ \widetilde a}(Z_i,\zeta_i)\,{\mathsf d}\theta_i^{\widehat\beta}+e_{\widetilde b}^{\ \widetilde a}(Z_i,\zeta_i)\,{\mathsf d}\zeta_i^{\widetilde b}\bigr)\otimes Z_{\widetilde a}\cr\cr
&&+\bigl(e_{\widehat a}^{\ \widetilde\a}(Z_i,\zeta_i)\,{\mathsf d} X_i^{\widehat a}+e_{\widehat\beta}^{\ \widetilde\a}(Z_i,\zeta_i)\,{\mathsf d}\theta_i^{\widehat\beta}+e_{\widetilde b}^{\ \widetilde\a}(Z_i,\zeta_i)\,{\mathsf d}\zeta_i^{\widetilde b}\bigr)\otimes Z_{\widetilde\a}\,,\label{eq:MCsformext}
\end{eqnarray}
where the second (subtracted) term on the left-hand side is identical (in the functional sense) with its counterpart on $\,\mathcal{O}_i\,$ derived for $\,Z_{\widetilde a}=0=Z_{\widetilde\a}$,\ and where the nontrivial Vielbeine $\,e_x^{\widetilde a}\,$ and $\,e_x^{\widetilde\a},\ x\in\{\widehat a,\widehat\beta,\widetilde b\}\,$ in the component of the super-1-form along $\,\gt{z}\,$ reflect the deformation.
\brem
Our hitherto analysis gives us a particular choice of a surjective submersion over the original target superspace $\,{\rm G}/{\rm H}\,$ of the super-$\sigma$-model, to wit,
\begin{eqnarray}
\pi_{{\mathsf Y}({\rm G}/{\rm H})}\ &:&\ {\mathsf Y}({\rm G}/{\rm H}):=\bigsqcup_{i\in I}\,\widetilde\mathcal{O}_i\longrightarrow{\rm G}/{\rm H}\cr\cr
&:&\ \bigl(\pi_{\widetilde{\rm G}/{\rm H}}\circ\widetilde\sigma_j(\widetilde Z_j),j\bigr)\equiv(Z_j,\zeta_j,j)\longmapsto Z_j\equiv\pi_{{\rm G}/{\rm H}}\circ\sigma_j(Z_j)\,.\label{eq:surjsubmext}
\end{eqnarray}
It is this surjective submersion that we might take as the point of departure of a construction of the geometric object (a super-gerbe) over $\,{\rm G}/{\rm H}\,$ presenting, in a manner that generalises the previously considered mechanism of geometrisation of Green--Schwarz super-$(p+2)$-cocycles on Lie supergroups to the setting of their homogeneous spaces, the supersymmetric Green--Schwarz super-$(p+2)$-cocycle $\,\underset{\tx{\ciut{(p+2)}}}{{\rm H}}\,$ on $\,{\rm G}/{\rm H}\,$ pulling back to a given super-$(p+2)$-cocycle $\,\underset{\tx{\ciut{(p+2)}}}{\chi}\,$ on $\,{\rm G}\,$ as per
\begin{eqnarray}\nn
\underset{\tx{\ciut{(p+2)}}}{\chi}=\pi_{{\rm G}/{\rm H}}^*\underset{\tx{\ciut{(p+2)}}}{{\rm H}}\,.
\end{eqnarray}
Instrumental in this construction is the non-linear realisation of the extended supersymmetry group $\,\widetilde{\rm G}\,$ on the homogeneous space $\,\widetilde{\rm G}/{\rm H}\,$ with a definition, using the local sections $\,\widetilde\sigma_i$,\ fully analogous to that of the non-linear realisation of the original supersymmetry group $\,{\rm G}\,$ on $\,{\rm G}/{\rm H}\cong\mathscr{M}$.\ In consequence of the assumptions made as to the nature of the extension $\,\widetilde\gt{g}\longrightarrow\gt{g}$,\ integrating to a Lie-supergroup extension $\,\widetilde\pi\ :\ \widetilde{\rm G}\longrightarrow{\rm G}$,\ the realisation employs the same locally smooth mappings $\,\unl h_{ij}\,$ as before, {\it cp} \Reqref{eq:cosetact}. With these tools in hand, we may subsequently examine invariance of the various components of the $\widetilde\gt{g}$-valued Maurer--Cartan super-1-forms on $\,\widetilde{\rm G}\,$ restricted to the image of $\,\widetilde{\rm G}/{\rm H}\,$ within $\,\widetilde{\rm G}\,$ under the family $\,\{\widetilde\sigma_i\}_{i\in I}\,$ of sections, whereby we conclude that it is the linear combinations of wedge products of components along $\,\widetilde\gt{t}\,$ with ${\rm H}$-invariant tensors as coefficients that ought to be considered, upon pullback to $\,{\mathsf Y}({\rm G}/{\rm H})\,$ (along the $\,\widetilde\sigma_i$), as admissible trivialisations of the pullback of the Green--Schwarz super-$(p+2)$-cocycles along $\,\pi_{{\mathsf Y}({\rm G}/{\rm H})}$.\ The first step in the geometrisation would therefore consist in finding such a global primitive $\,\underset{\tx{\ciut{(p+1)}}}{\widetilde\beta}\,$ of the pullback of $\,\underset{\tx{\ciut{(p+2)}}}{\chi}\,$ to the extended supersymmetry group $\,\widetilde{\rm G}$,
\begin{eqnarray}\nn
\widetilde\pi^*\underset{\tx{\ciut{(p+2)}}}{\chi}={\mathsf d}\underset{\tx{\ciut{(p+1)}}}{\widetilde\beta}\,,
\end{eqnarray}
and checking that it be invariant under locally unique lifts to
\begin{eqnarray}\nn
\widetilde\sigma(\widetilde{\rm G}/{\rm H})=\bigsqcup_{i\in I}\,\widetilde\sigma_i(\widetilde\mathcal{O}_i)
\end{eqnarray}
of the flows of the fundamental vector fields $\,\widetilde\Xi_{\widetilde X}\,$ for the transitive action of the extended supersymmetry group $\,\widetilde{\rm G}\,$ on $\,\widetilde{\rm G}/{\rm H}$,\ or, equivalently, that the conditions
\begin{eqnarray}\nn
\pLie{\widetilde\mathcal{K}_{\widetilde X}}\underset{\tx{\ciut{(p+1)}}}{\widetilde\beta}=0\,,\qquad\widetilde X\in\widetilde\gt{g}
\end{eqnarray}
be satisfied for
\begin{eqnarray}\nn
\widetilde\mathcal{K}_{\widetilde X}\mathord{\restriction}_{\widetilde\sigma_i(\widetilde\mathcal{O}_i)}={\mathsf T}\widetilde\sigma_i(\widetilde\Xi_{\widetilde X})\,.
\end{eqnarray}
A meaningful continuation of the thus initiated generalisation of the definition of a Cartan--Eilenberg super-1-gerbe over a Lie supergroup to the setting of homogeneous spaces (without any obvious Lie-supergroup structure) would require further insights into some working examples of supersymmetry-equivariant trivialisation of Green--Schwarz super-3-cocycles on such homogeneous spaces that we currently lack. Hence, we abandon our considerations at this stage and postpone them to a future investigation.
\end{Rem}\medskip
The above supersymmetric trivialisation mechanism will be implicit in the examples scrutinised in Sections \ref{sec:sMinkext} and the one suggested in Section \ref{sec:MTaway}.
\brem A particular setting in which an explicit contribution to the charge and a resultant deformation of (super)\-sym\-me\-try sourced by the pseudo-invariance of the global primitive $\,\underset{\tx{\ciut{(p+1)}}}{\beta}\,$ of the Green--Schwarz super-$(p+2)$-cocycle $\,\underset{\tx{\ciut{(p+2)}}}{\chi}\mathord{\restriction}_{\sigma({\rm G}/{\rm H})}\,$ manifests itself is the realisation of supersymmetry on states from the Hilbert space $\,\mathcal{H}\,$ of the theory, induced, along the lines of Section I.2.2, in the standard procedure of geometric (pre)quantisation available in the field-theoretic setting in hand. In order to isolate and elucidate the phenomenon of interest, we make several simplifying assumptions. Thus, we consider a global primitive $\,\underset{\tx{\ciut{(p+1)}}}{{\rm B}}\,$ of the supersymmetric super-$(p+2)$-cocycle $\,\underset{\tx{\ciut{(p+2)}}}{{\rm H}}\,$ on $\,{\rm G}/{\rm H}$,\ the latter being descended from $\,\underset{\tx{\ciut{(p+2)}}}{\chi}\mathord{\restriction}_{\sigma({\rm G}/{\rm H})}$.\ We take its behaviour under the left action $\,[\lambda]_\cdot\,$ of \Reqref{eq:cosetlact} to be captured by the identities
\begin{eqnarray}\nn
[\lambda]_g^*\underset{\tx{\ciut{(p+1)}}}{{\rm B}}-\underset{\tx{\ciut{(p+1)}}}{{\rm B}}={\mathsf d}\underset{\tx{\ciut{(p)}}}{\jmath_g}\,,\qquad g\in{\rm G}\,,
\end{eqnarray}
with globally defined supersymmetry currents $\,\underset{\tx{\ciut{(p)}}}{\jmath_g}\in\Omega^p({\rm G}/{\rm H})$.\ In the above, we recognise a simplified (global) variant of relations \eqref{eq:pseudolocB}. The said action $\,R\ :\ {\rm G}\longrightarrow{\rm U}(\mathcal{H})$,\ is now readily derived, for any $\,g\in{\rm G}$,\ in the form
\begin{eqnarray}\nn
\bigl(R(g)\Psi\bigr)[\phi]:=c_g[\phi]\cdot\Psi\bigl[[\lambda]_{g^{-1}}\circ\phi\bigr]\,,\qquad\qquad c_g[\phi]:={\rm e}^{{\mathsf i}\,\int_{\mathscr{C}_{p,{\rm in}}}\,([\lambda]_{g^{-1}}\circ\phi)^*\underset{\tx{\ciut{(p)}}}{\jmath_g}}\,.
\end{eqnarray}
Accordingly, we find, for arbitrary $\,g_1,g_2\in{\rm G}$,\ the identity
\begin{eqnarray}\nn
R(g_1)\circ R(g_2)=(\delta_{\rm G} c)_{g_1,g_2}[\cdot]\cdot R(g_1\cdot g_2)\,,
\end{eqnarray}
with the homomorphicity 2-cocycle given by
\begin{eqnarray}\nn
(\delta_{\rm G} c)_{g_1,g_2}[\phi]={\rm e}^{{\mathsf i}\,\int_{\mathscr{C}_{p,{\rm in}}}\,([\lambda]_{(g_1\cdot g_2)^{-1}}\circ\phi)^*(\delta_{\rm G}\underset{\tx{\ciut{(p)}}}{\jmath})_{g_1,g_2}}
\end{eqnarray}
in terms of the current 2-cocycle
\begin{eqnarray}\nn
(\delta_{\rm G}\underset{\tx{\ciut{(p)}}}{\jmath})_{g_1,g_2}:=[\lambda]_{g_2}^*\underset{\tx{\ciut{(p)}}}{\jmath_{g_1}}-\underset{\tx{\ciut{(p)}}}{\jmath_{g_1\cdot g_2}}+\underset{\tx{\ciut{(p)}}}{\jmath_{g_2}}\,.
\end{eqnarray}
The latter defines a class in $\,H^p({\rm G}/{\rm H})\,$ as
\begin{eqnarray}\nn
{\mathsf d}(\delta_{\rm G}\jmath)_{g_1,g_2}=[\lambda]_{g_2}^*\bigl([\lambda]_{g_1}^*\underset{\tx{\ciut{(p+1)}}}{{\rm B}}-\underset{\tx{\ciut{(p+1)}}}{{\rm B}}\bigr)-\bigl([\lambda]_{g_1\cdot g_2}^*\underset{\tx{\ciut{(p+1)}}}{{\rm B}}-\underset{\tx{\ciut{(p+1)}}}{{\rm B}}\bigr)+\bigl([\lambda]_{g_2}^*\underset{\tx{\ciut{(p+1)}}}{{\rm B}}-\underset{\tx{\ciut{(p+1)}}}{{\rm B}}\bigr)=0\,,
\end{eqnarray}
and so the exponent of the homomorphicity 2-cocycle describes the standard pairing between the (de Rham) cohomology class $\,[([\lambda]_{(g_1\cdot g_2)^{-1}}\circ\phi)^*(\delta_{\rm G}\jmath)_{g_1,g_2}]\,$ of the pullback of that 2-cocycle and the homology class $\,[\mathscr{C}_{p,{\rm in}}]\,$ of the Cauchy $p$-cycle,
\begin{eqnarray}\nn
(\delta_{\rm G} c)_{g_1,g_2}[\phi]\equiv{\rm e}^{{\mathsf i}\,\corr{[([\lambda]_{(g_1\cdot g_2)^{-1}}\circ\phi)^*(\delta_{\rm G}\jmath)_{g_1,g_2}],[\mathscr{C}_{p,{\rm in}}]}}\,.
\end{eqnarray}
This demonstrates that we are dealing with a wrapping-charge extension of $\,{\rm G}$.\ The extension trivialises whenever the primitive $\,\underset{\tx{\ciut{(p+1)}}}{\beta}\,$ is invariant on the nose as we may then choose $\,\jmath_g=0$,\ whereby we obtain $\,(\delta_{\rm G} c)_{g_1,g_2}[\phi]=1$.
Note that the extension is \emph{not} independent of the the choice of the representative of the class of primitives of the Green--Schwarz super-$(p+2)$-cocycle. It depends, though, on the primitive $\,\underset{\tx{\ciut{(p+1)}}}{{\rm B}}\,$ solely up to a de Rham-exact correction. Indeed, upon replacement
\begin{eqnarray}\nn
\underset{\tx{\ciut{(p+1)}}}{{\rm B}}\longmapsto\underset{\tx{\ciut{(p+1)}}}{{\rm B}}+{\mathsf d}\underset{\tx{\ciut{(p)}}}{\eta}\,,
\end{eqnarray}
we may redefine the current as
\begin{eqnarray}\nn
\underset{\tx{\ciut{(p)}}}{\jmath_g}\longmapsto\underset{\tx{\ciut{(p)}}}{\jmath_g}+\bigl(\delta_{\rm G}\underset{\tx{\ciut{(p)}}}{\eta}\bigr)_g\,,
\end{eqnarray}
and so
\begin{eqnarray}\nn
(\delta_{\rm G} c)_{g_1,g_2}\longmapsto(\delta_{\rm G} c)_{g_1,g_2}\,.
\end{eqnarray}
In particular, and this is to be emphasised in the context of subsequent case studies, whenever there exists a ${\rm G}$-invariant primitive, a correction by an exact super-$(p+1)$-form -- whether ${\rm G}$-invariant or \emph{not} -- does not affect the homomorphicity 2-cocycle.
In summary, the dependence of the quantum realisation of the (super)symmetry group upon the choice of the primitive $\,\underset{\tx{\ciut{(p+1)}}}{{\rm B}}\,$ is, as usual, stronger than the one present in the canonical setting of the previous section.
\end{Rem}\medskip
\section{The Kosteleck\'y--Rabin extensions of the super-Minkowskian algebra}\label{sec:sMinkext}
We begin our case-by-case analysis of the wrapping charges and the associated deformations of the supertarget Lie superalgebras by reconsidering the super-Minkowski space (essentially in the notation of Part I)
\begin{eqnarray}\nn
{\rm sMink}^{d,1\,\vert\,D_{d,1}}\equiv{\rm sISO}(d,1\,\vert\,D_{d,1})/{\rm SO}(d,1)\,,
\end{eqnarray}
a homogeneous space of the super-Poincar\'e supergroup
\begin{eqnarray}\nn
{\rm sISO}(d,1\,\vert\,D_{d,1})\equiv{\mathbb{R}}^{d,1\,\vert\,D_{d,1}}\rtimes{\rm SO}(d,1)
\end{eqnarray}
with the Lie superalgebra
\begin{eqnarray}\nn
\gt{siso}(d,1\,\vert\,D_{d,1})=\bigg(\bigoplus_{\widehat\a=1}^{D_{d,1}}\,\corr{Q_{\widehat\a}}_{\mathbb{C}}\oplus\bigoplus_{\widehat a=0}^d\,\corr{P_{\widehat a}}_{\mathbb{C}}\bigg)\oplus\bigoplus_{\widehat a,\widehat b=0}^d\,\corr{J_{\widehat a\widehat b}\equiv-J_{\widehat b\widehat a}}_{\mathbb{C}}
\end{eqnarray}
defined by the structure equations (here, $\,(\eta_{\widehat a\widehat b})\equiv\textrm{diag}(-1,\underbrace{+1,+1,\ldots,+1}_{d\ {\rm times}})$)
\begin{eqnarray}\nn
&[P_{\widehat a},P_{\widehat b}]=0\,,\qquad\qquad\{Q_{\widehat\a},Q_{\widehat\beta}\}=\ovl\Gamma^{\widehat a}_{\widehat\a\widehat\beta}\,P_{\widehat a}\,,\qquad\qquad[P_{\widehat a},Q_{\widehat\a}]=0\,,&\cr\cr
&[J_{\widehat a\widehat b},J_{\widehat c\widehat d}]=\eta_{\widehat a\widehat d}\,J_{\widehat b\widehat c}-\eta_{\widehat a\widehat c}\,J_{\widehat b\widehat d}+\eta_{\widehat b\widehat c}\,J_{\widehat a\widehat d}-\eta_{\widehat b\widehat d}\,J_{\widehat a\widehat c}\,,&\cr\cr
&[J_{\widehat a\widehat b},P_{\widehat c}]=\eta_{\widehat b\widehat c}\,P_{\widehat a}-\eta_{\widehat a\widehat c}\,P_{\widehat b}\,,\qquad\qquad[J_{\widehat a\widehat b},Q_{\widehat\a}]=\tfrac{1}{2}\,\bigl(\Gamma_{\widehat a\widehat b}\bigr)^{\widehat\beta}_{\ \widehat\a}\,Q_{\widehat\beta}\,.&
\end{eqnarray}
The homogeneous space is embedded in the supersymmetry group $\,{\rm sISO}(d,1\,\vert\,D_{d,1})\,$ by a single section using the standard (global) coordinates $\,\{\theta^{\widehat\a},X^{\widehat a}\}^{\widehat\a\in\ovl{1,D_{d,1}},\ \widehat a\in\ovl{0,d}}\,$ on $\,{\rm sMink}^{d,1\,\vert\,D_{d,1}}$,
\begin{eqnarray}\nn
\sigma\ :\ {\rm sMink}^{d,1\,\vert\,D_{d,1}}\longrightarrow{\rm sISO}(d,1\,\vert\,D_{d,1})\ :\ (\theta,x)\longmapsto{\rm e}^{\theta^{\widehat\a}\,Q_{\widehat\a}+X^{\widehat a}\,P_{\widehat a}}\,.
\end{eqnarray}
It is a Lie supergroup itself, namely the supertranslation group $\,{\mathbb{R}}^{d,1\,\vert\,D_{d,1}}$.\ We read off its binary operation from the above embedding using its Lie (sub-)superalgebra -- from the (Baker--Campbell--Dynkin--Hausdorff) identity
\begin{eqnarray}\nn
{\rm e}^{\theta_1^{\widehat\a}\,Q_{\widehat\a}+X_1^{\widehat a}\,P_{\widehat a}}\cdot{\rm e}^{\theta_2^{\widehat\beta}\,Q_{\widehat\beta}+X_2^{\widehat b}\,P_{\widehat b}}&=&{\rm e}^{(\theta_1^{\widehat\a}+\theta_2^{\widehat\a})\,Q_{\widehat\a}+(X_1^{\widehat a}+X_2^{\widehat a}+\frac{1}{2}\,[\theta_1^{\widehat\a}\,Q_{\widehat\a}+X_1^{\widehat a}\,P_{\widehat a},\theta_2^{\widehat\beta}\,Q_{\widehat\beta}+X_2^{\widehat b}\,P_{\widehat b}])^{\widehat c}\,P_{\widehat c}}\cr\cr
&=&{\rm e}^{(\theta_1^{\widehat\a}+\theta_2^{\widehat\a})\,Q_{\widehat\a}+(X_1^{\widehat a}+X_2^{\widehat a}-\frac{1}{2}\,\ovl\theta_1\,\Gamma^{\widehat a}\,\theta_2)\,P_{\widehat a}}
\end{eqnarray}
we extract the explicit formula
\begin{eqnarray}\nn
(\theta_1^{\widehat\a},X_1^{\widehat a})\cdot(\theta_2^{\widehat\beta},X_2^{\widehat b})=\bigl(\theta_1^{\widehat\a}+\theta_2^{\widehat\a},X_1^{\widehat a}+X_2^{\widehat a}-\tfrac{1}{2}\,\ovl\theta_1\,\Gamma^{\widehat a}\,\theta_2\bigr)\,.
\end{eqnarray}
Given these, we take a closer look at the Green--Schwarz super-$(p+2)$-cocycles over the super-Minkowski space, for $\,p\in\{0,1\}$,\ (left-)invariant under the action of the super-Poincar\'e supergroup, alongside the corresponding non-supersymmetric primitives, found in Part I. For the latter supergroup, we compute the supersymmetry-variation super-$p$-forms $\,\underset{\tx{\ciut{(p)}}}{\Gamma_X}$.\ In our computations, we use the coordinate form of the relevant right-invariant vector fields
\begin{eqnarray}\nn
\mathcal{R}_{(\varepsilon,0)}(\theta,X)=\varepsilon^{\widehat\a}\,\tfrac{\vec\partial\ }{\partial\theta^{\widehat\a}}-\tfrac{1}{2}\,\ovl\varepsilon\,\Gamma^{\widehat a}\,\theta\,\tfrac{\partial\ }{\partial X^{\widehat a}}\,,\qquad\qquad\mathcal{R}_{(0,y)}(\theta,X)=y^{\widehat a}\,\tfrac{\partial\ }{\partial X^{\widehat a}}\,.
\end{eqnarray}
\subsection{The superparticle extension of $\,{\rm sMink}^{9,1\,\vert\,D_{9,1}}$}\label{subsect:spartextMink}
The relevant ${\rm sISO}(9,1\,\vert\,D_{9,1})$-invariant Green--Schwarz super-2-cocycle reads
\begin{eqnarray}\nn
\underset{\tx{\ciut{(2)}}}{\chi}(\theta,X)=\theta_{\rm L}^{\widehat\a}\wedge\ovl\Gamma_{11\,\widehat\a\widehat\beta}\,\theta^{\widehat\beta}_{\rm L}(\theta,X)\equiv\ovl{{\mathsf d}\theta}\wedge\Gamma_{11}\,{\mathsf d}\theta
\end{eqnarray}
and admits a primitive
\begin{eqnarray}\nn
\underset{\tx{\ciut{(1)}}}{\beta}(\theta,X)=\ovl{\theta}\,\Gamma_{11}\,{\mathsf d}\theta\,,
\end{eqnarray}
for which we get
\begin{eqnarray}\nn
\pLie{\mathcal{R}_{(\varepsilon,0)}}\underset{\tx{\ciut{(1)}}}{\beta}(\theta,X)=\ovl\varepsilon\,\Gamma_{11}\,{\mathsf d}\theta\,,\qquad\qquad\pLie{\mathcal{R}_{(0,y)}}\underset{\tx{\ciut{(1)}}}{\beta}(\theta,X)=0\,,
\end{eqnarray}
and so also
\begin{eqnarray}\nn
\underset{\tx{\ciut{(0)}}}{\Gamma_{(\varepsilon,0)}}(\theta,X)=\ovl\varepsilon\,\Gamma_{11}\,\theta\,,\qquad\qquad\underset{\tx{\ciut{(0)}}}{\Gamma_{(0,y)}}(\theta,X)=0\,,
\end{eqnarray}
whence
\begin{eqnarray}\nn
h_{(\varepsilon,0)}(\theta,X,p)=-\ovl\varepsilon\,(p_{\widehat a}\,\Gamma^{\widehat a}+2\Gamma_{11})\,\theta\,,\qquad\qquad h_{(0,y)}(\theta,X,p)=y^{\widehat a}\,p_{\widehat a}\,,
\end{eqnarray}
and
\begin{eqnarray}\nn
\mathscr{W}_{(\varepsilon_1,0),(\varepsilon_2,0)}[\theta,X]=2\ovl\varepsilon_1\,\Gamma_{11}\,\varepsilon_2\,,\qquad\qquad\mathscr{W}_{(0,y_1),(0,y_2)}[\theta,X]=0=\mathscr{W}_{(\varepsilon,0),(0,y)}[\theta,X]\,.
\end{eqnarray}
The ensuing deformation of the super-Minkowski superalgebra (obtained, {\it e.g.}, through canonical quantisation of the super-centrally extended Poisson--Lie algebra of Noether charges, and the obvious sign flip) is the familiar Lie superalgebra
\begin{eqnarray}\nn
\{Q_{\widehat\a},Q_{\widehat\beta}\}^\sim=\ovl\Gamma^{\widehat a}_{\widehat\a\widehat\beta}\,P_{\widehat a}+2\,\ovl\Gamma_{11\,\widehat\a\widehat\beta}\,Z\,,\qquad\qquad[P_{\widehat a},P_{\widehat b}]^\sim=0=[Q_{\widehat\a},P_{\widehat a}]^\sim
\end{eqnarray}
encountered in Part I, with an additional central generator $\,Z$.
The super-2-cocycle associated with the above anomaly,
\begin{eqnarray}\nn
\varpi_0=\ovl{\sigma}\wedge\Gamma_{11}\,\sigma\,,
\end{eqnarray}
is precisely the GS super-2-cocycle whose trivialisation over the Lie supergroup that integrates the super-centrally extended $\,{\mathbb{R}}^{9,1\,\vert\,D_{9,1}}\,$ determines the Green--Schwarz super-0-gerbe of Def.\,I.5.2.
\subsection{The superstring extension of $\,{\rm sMink}^{d,1\,\vert\,D_{d,1}}$}\label{subsect:sstringextsMink}
The ${\rm sISO}(d,1\,\vert\,D_{d,1})$-invariant Green--Schwarz super-$3$-cocycle of interest can be written in the following form:
\begin{eqnarray}\label{eq:sMinkGS3}
\underset{\tx{\ciut{(3)}}}{\chi}(\theta,X)=\theta_{\rm L}^{\widehat\a}\wedge\ovl\Gamma_{{\widehat a}\,\widehat\a\widehat\beta}\,\theta_{\rm L}^{\widehat\beta}\wedge\theta_{\rm L}^{\widehat a}(\theta,X)\equiv\ovl{{\mathsf d}\theta}\wedge\Gamma_{\widehat a}\,{\mathsf d}\theta\wedge{\mathsf d} X^{\widehat a}\,,
\end{eqnarray}
which immediately gives a primitive
\begin{eqnarray}\label{eq:sMincurv}
\underset{\tx{\ciut{(2)}}}{\beta}(\theta,X)=\ovl{\theta}\,\Gamma_{\widehat a}\,{\mathsf d}\theta\wedge{\mathsf d} X^{\widehat a}\,,
\end{eqnarray}
satisfying
\begin{eqnarray}\nn
\pLie{\mathcal{R}_{(\varepsilon,0)}}\underset{\tx{\ciut{(2)}}}{\beta}(\theta,X)=\ovl\varepsilon\,\Gamma_{\widehat a}\,{\mathsf d}\theta\wedge{\mathsf d} X^{\widehat a}+\tfrac{1}{2}\,\ovl\varepsilon\,\Gamma_{\widehat a}\,{\mathsf d}\theta\wedge\ovl\theta\,\Gamma^{\widehat a}\,{\mathsf d}\theta\,,\qquad\qquad\pLie{\mathcal{R}_{(0,y)}}\underset{\tx{\ciut{(2)}}}{\beta}(\theta,X)=0\,,
\end{eqnarray}
so that we obtain
\begin{eqnarray}\nn
\underset{\tx{\ciut{(1)}}}{\Gamma_{(\varepsilon,0)}}(\theta,X)=\ovl\varepsilon\,\Gamma_{\widehat a}\,\theta\,\bigl({\mathsf d} X^{\widehat a}+\tfrac{1}{6}\,\ovl\theta\,\Gamma^{\widehat a}\,{\mathsf d}\theta\bigr)\,,\qquad\qquad\underset{\tx{\ciut{(1)}}}{\Gamma_{(0,y)}}(\theta,X)=0
\end{eqnarray}
and the hamiltonians
\begin{eqnarray}\nn
h_{(\varepsilon,0)}[\theta,X,p]&=&-\int_0^{2\pi}\,{\mathsf d}\varphi\,(\ovl\varepsilon\,\Gamma_{\widehat a}\,\theta)\,\bigl(\eta^{\widehat a\widehat b}\,p_{\widehat b}+2\partial_\varphi X^{\widehat a}-\tfrac{1}{3}\,\ovl\theta\,\Gamma^{\widehat a}\,\partial_\varphi\theta\bigr)(\varphi)\,,\cr\cr
h_{(0,y)}[\theta,X,p]&=&\int_0^{2\pi}\,{\mathsf d}\varphi\,y^{\widehat a}\,\bigl(p_{\widehat a}-\ovl\theta\,\Gamma_{\widehat a}\,\partial_\varphi\theta\bigr)(\varphi)\,.
\end{eqnarray}
Accordingly, the wrapping anomaly reads
\begin{eqnarray}\nn
&\mathscr{W}_{(\varepsilon_1,0),(\varepsilon_2,0)}[\theta,X]=\int_0^{2\pi}\,{\mathsf d}\varphi\,\partial_\varphi\bigl(2\ovl\varepsilon_1\,\Gamma_{\widehat a}\,\varepsilon_2\,X^{\widehat a}\bigr)(\varphi)=0\,,\qquad\qquad\mathscr{W}_{(0,y_1),(0,y_2)}[\theta,X]=0\,,&\cr\cr
&\mathscr{W}_{(\varepsilon,0),(0,y)}[\theta,X]=\int_0^{2\pi}\,{\mathsf d}\varphi\,\partial_\varphi\bigl(-2y^{\widehat a}\,\ovl\varepsilon\,\Gamma_{\widehat a}\,\theta\bigr)(\varphi)\,.&
\end{eqnarray}
Owing to the topological triviality of the body $\,{\rm Mink}^{d,1}\equiv\vert{\rm sMink}^{d,1\,\vert\,D_{d,1}}\vert\,$ of the supertarget, the only non-vanishing contribution to the wrapping anomaly comes from the last term, and only if we take into account the Kosteleck\'y--Rabin states. Let $\,\nu\,$ be the (constant) Gra\ss mann-odd vector defining the Kosteleck\'y--Rabin lattice in $\,{\mathbb{R}}^{0\,\vert\,D_{d,1}}$,\ so that we may consider winding states with
\begin{eqnarray}\nn
\theta(2\pi)=\theta(0)+\nu\,.
\end{eqnarray}
We may then rewrite the last component of the anomaly in the form
\begin{eqnarray}\nn
\mathscr{W}_{(\varepsilon,0),(0,y)}[\theta,X]=-2y^{\widehat a}\,\ovl\varepsilon\,\Gamma_{\widehat a}\,\nu\,.
\end{eqnarray}
This produces ({\it e.g.}, through canonical quantisation, after a sign flip) -- as in \Reqref{eq:wrapcentrext} -- the super-centrally extended $\,{\rm sMink}^{d,1\,\vert\,D_{d,1}}\,$ algebra
\begin{eqnarray}
&\{Q_{\widehat\a},Q_{\widehat\beta}\}=\ovl\Gamma^{\widehat a}_{\widehat\a\widehat\beta}\,P_{\widehat a}\,,\qquad\qquad[P_{\widehat a},P_{\widehat b}]=0\,,\qquad\qquad[Q_{\widehat\a},P_{\widehat a}]=2\ovl\Gamma_{{\widehat a}\,\widehat\a\widehat\beta}\,Z^{\widehat\beta}\,,&\nonumber \\ \label{eq:ssextsMink} \\
&[Z^{\widehat\a},\xi\}=0\,,\quad \xi\in{\mathbb{R}}^{d,1\,\vert\,D_{d,1}}\,.&\nonumber
\end{eqnarray}
The anomaly corresponds to the family of super-2-cocycles
\begin{eqnarray}\nn
\varpi_{1\,\widehat\a}=2\sigma^{\widehat\beta}\wedge\ovl\Gamma_{\widehat a\,\widehat\beta\widehat\a}\,e^{\widehat a}\,,
\end{eqnarray}
whose trivialisation over the Lie supergroup that integrates the super-centrally extended $\,{\mathbb{R}}^{d,1\,\vert\,D_{d,1}}\,$ determines (the surjective submersion of) the Green--Schwarz super-1-gerbe of Def.\,I.5.9.
\section{The superstring deformation of the super-${\rm AdS}_5\x{\mathbb{S}}^5\,$ algebra}\label{sec:ssextaAdSS}
The point of departure of our subsequent considerations is the Metsaev--Tseytlin super-$\sigma$-model of superloop mechanics on the homogeneous space
\begin{eqnarray}\nn
{\rm SU}(2,2\,\vert\,4)/({\rm SO}(4,1)\x{\rm SO}(5))\equiv{\rm s}\bigl({\rm AdS}_5\x{\mathbb{S}}^5\bigr)
\end{eqnarray}
of the Lie supergroup $\,{\rm SU}(2,2\,\vert\,4)$,\ with the body given by the homogeneous space
\begin{eqnarray}\nn
{\rm SO}(4,2)/{\rm SO}(4,1)\x{\rm SO}(6)/{\rm SO}(5)\equiv{\rm AdS}_5\x{\mathbb{S}}^5
\end{eqnarray}
of the Lie group
\begin{eqnarray}\nn
{\rm SU}(2,2)\x{\rm SU}(4)\cong{\rm SO}(4,2)\x{\rm SO}(6)\,.
\end{eqnarray}
In the notation of the original paper \cite{Metsaev:1998it} and upon incorporation of the findings of Roiban and Siegel reported in \Rcite{Roiban:2000yy}, the model takes the form \eqref{eq:supersimod} with the supertarget (group) metric
\begin{eqnarray}\nn
&({\rm G}_{\widehat a\widehat b})\equiv(\eta_{\widehat a\widehat b})=(\eta_{ab})\oplus(\delta_{a'b'})\,,&\cr\cr &(\eta_{ab})\equiv\textrm{diag}(-1,+1,+1,+1,+1)\,,\qquad\qquad(\delta_{a'b'})\equiv\textrm{diag}(+1,+1,+1,+1,+1)&
\end{eqnarray}
and with the relevant Green--Schwarz super-3-cocycle
\begin{eqnarray}\label{eq:MT3scocyc}
\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}=-\Sigma_{\rm L}\wedge(\widehat C\otimes\sigma_3)\,\cancel e\wedge\Sigma_{\rm L}\,,
\end{eqnarray}
the latter being written, in the shorthand notation of Eqs.\,\eqref{eq:Gext1} and \eqref{eq:CCc}, in terms of the components
\begin{eqnarray}\nn
\Sigma_{\rm L}^{\a\a' I}=\theta_{\rm L}^{\a\a' I}\,,\qquad\qquad\qquad\cancel e=(\widehat\Gamma_{\widehat a}\otimes{\boldsymbol{1}})\,e^{\widehat a}\,,\quad e^{\widehat a}=\theta_{\rm L}^{\widehat a}
\end{eqnarray}
of the left-invariant Maurer--Cartan super-1-form on $\,{\rm SU}(2,2\vert 4)\,$ along the vector-space complement of the isotropy Lie algebra $\,\gt{so}(4,1)\oplus\gt{so}(5)\,$ within $\,\gt{su}(2,2\,\vert\,4)$,\ contracted with ${\rm SO}(4,1)\x{\rm SO}(5)$-invariant tensors. Here, the superalgebra and the supergeometry of the model (and so also its field-theoretic content) have been cast in a form compatible with the decomposition of the body into its independent constituents: $\,{\rm AdS}_5\,$ and $\,{\mathbb{S}}^5$,\ {\it i.e.}, we work with the Majorana--Weyl spinors of the product $\,{\rm Spin}\,$ group $\,{\rm Spin}(4,1)\x{\rm Spin}(5)\x{\rm Spin}(2,1)\,$ (the last factor accounts for the two species of chiral spinors entering the construction), whence the presence of the multi-indices $\,\widehat\a\equiv\a\a' I\,$ on them, with $\,\a,\a'\in\{1,2,3,4\}\,$ and $\,I\in\{1,2\}\,$ (and that of the tensor products of elements of the Clifford algebras of the quadratic spaces $\,{\mathbb{R}}^{4,1},{\mathbb{R}}^{5,0}\,$ and $\,{\mathbb{R}}^{2,1}$), and with tensors of the product isotropy group $\,{\rm SO}(4,1)\x{\rm SO}(5)$,\ whence the two subsets of vector indices: $\,a\in\ovl{0,4}\,$ and $\,a'\in\ovl{5,9}$.\ Important properties of the distinguished representations of the said Clifford algebras entering the construction, including the fundamental Fierz identities satisfied by their generators, have been recapitulated in App.\,\ref{app:CliffAdSS}.
The Lie superalgebra $\,\mathfrak{su}(2,2\,\vert\,4)\,$ of the supersymmetry group has generators
\begin{eqnarray}\nn
\gt{su}(2,2\,\vert\,4)&\equiv&\bigl(\gt{t}^{(0)}\oplus\gt{t}^{(1)}\bigr)\oplus\bigl(\gt{so}(4,1)\oplus\gt{so}(5)\bigr)\cr\cr
&=&\bigg(\bigl(\bigoplus_{a=0}^4\,\corr{P_a}_{\mathbb{C}}\oplus\bigoplus_{a'=5}^9\,\corr{P_{a'}}_{\mathbb{C}}\bigr)\oplus\bigoplus_{(\a,\a', I)\in\ovl{1,4}\x\ovl{1,4}\x\{1,2\}}\,\corr{Q_{\a\a' I}}_{\mathbb{C}}\bigg)\cr\cr
&&\oplus\bigg(\bigoplus_{a,b=0}^4\,\corr{J_{ab}=-J_{ba}}_{\mathbb{C}}\oplus\bigoplus_{a',b'=5}^9\,\corr{J_{a'b'}=-J_{b'a'}}_{\mathbb{C}}\bigg)
\end{eqnarray}
subject to the structure relations
\begin{eqnarray}
\{Q_{\a\a' I},Q_{\beta\b' J}\}={\mathsf i}\,\bigl(-2(\widehat C\,\widehat\Gamma^{\widehat a}\otimes{\boldsymbol{1}})_{\a\a'I\beta\b'J}\,P_{\widehat a}+(\widehat C\,\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2)_{\a\a'I\beta\b'J}\,J_{\widehat a\widehat b}\bigr)\,,\cr\cr\cr
[P_{\widehat a},P_{\widehat b}]=\varepsilon_{\widehat a\widehat b}\,J_{\widehat a\widehat b}\,,\qquad\varepsilon_{\widehat a\widehat b}=\left\{ \begin{array}{cl} +1 & \tx{if}\ \widehat a,\widehat b\in\ovl{0,4} \\
-1 & \tx{if}\ \widehat a,\widehat b\in\ovl{5,9} \\
0 & \tx{otherwise}\end{array}\right.\,,\cr\cr\cr
[J_{\widehat a\widehat b},J_{\widehat c\widehat d}]=\eta_{\widehat a\widehat d}\,J_{\widehat b\widehat c}-\eta_{\widehat a\widehat c}\,J_{\widehat b\widehat d}+\eta_{\widehat b\widehat c}\,J_{\widehat a\widehat d}-\eta_{\widehat b\widehat d}\,J_{\widehat a\widehat c}\,,\qquad\qquad[P_{\widehat a},J_{\widehat b\widehat c}]=\eta_{\widehat a\widehat b}\,P_{\widehat c}-\eta_{\widehat a\widehat c}\,P_{\widehat b}\,,\label{eq:AdSS}\\ \cr\cr
[Q_{\a\a' I},P_{\widehat a}]=-\tfrac{1}{2}\,(\widehat\Gamma_{\widehat a}\otimes\sigma_2)^{\beta\b'J}_{\ \ \a\a'I}\,Q_{\beta\b' J}\,,\qquad\qquad[Q_{\a\a' I},J_{\widehat a\widehat b}]=-\tfrac{1}{2}\,\varepsilon_{\widehat a\widehat b}\,(\widehat\Gamma_{\widehat a\widehat b}\otimes{\boldsymbol{1}})^{\beta\b'J}_{\ \ \a\a'I}\,Q_{\beta\b'J}\,.\nn
\end{eqnarray}
Upon rescaling
\begin{eqnarray}\label{eq:IWrescale}
\bigl(Q_{\a\a'I},P_{\widehat a},J_{\widehat a\widehat b}\bigr)\longmapsto\bigl(R^{\frac{1}{2}}\,Q_{\a\a'I},R\,P_{\widehat a},J_{\widehat a\widehat b}\bigr)\,,\quad R\in{\mathbb{R}}
\end{eqnarray}
the Lie superalgebra thus defined is readily seen to contract (in the sense of \.In\"on\"u and Wigner), in the limit $\,R\to\infty$,\ to the super-Minkowski algebra. The above structure relations are employed in the construction of the local sections that embed, in terms of local coordinates $\,\{\theta_i^{\a\a'I},X_i^{\widehat a}\}\,$ (around some $\,\unl g_i\,{\rm H}$), elements $\,\mathcal{O}_i\ni(\theta_i,X_i)\,$ of a trivialising cover $\,\{\mathcal{O}_i\}_{i\in I}\,$ of (the base of) the principal ${\rm SO}(4,1)\x{\rm SO}(5)$-bundle $\,{\rm SU}(2,2\,\vert\,4)\longrightarrow{\rm SU}(2,2\,\vert\,4)/({\rm SO}(4,1)\x{\rm SO}(5))\,$ in the total space of that bundle as per
\begin{eqnarray}\qquad\qquad
\sigma_i\ :\ \mathcal{O}_i\longrightarrow{\rm SU}(2,2\,\vert\,4)\ :\ Z_i\equiv\bigl(\theta_i^{\a\a' I},x_i^a,x_i'{}^{a'}\bigr)\equiv\bigl(\theta_i^{\a\a'I},X_i^{\widehat a}\bigr)\longmapsto\unl g_i\cdot{\rm e}^{X_i^{\widehat a}\,P_{\widehat a}}\cdot{\rm e}^{\theta_i^{\a\a' I}\,Q_{\a\a' I}}\,. \label{eq:MTsect}
\end{eqnarray}
The construction of the Wess--Zumino term of the Metsaev--Tseytlin super-$\sigma$-model calls for a global primitive of the Green--Schwarz super-3-cocycle \eqref{eq:MT3scocyc}. Among such primitives, we find the manifestly left-invariant one
\begin{eqnarray}\label{eq:susypriMT}
\underset{\tx{\ciut{(2)}}}{\beta}\equiv{\mathsf d}^{-1}\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}={\mathsf i}\,\ovl\Sigma_{\rm L}\wedge({\boldsymbol{1}}\otimes\sigma_1)\,\Sigma_{\rm L}\,.
\end{eqnarray}
The latter may be corrected by the addition of arbitrary closed super-2-forms, in a manner motivated by various physical and geometric considerations. In the light of our previous findings, the presence of such de Rham cocycles does not affect the Poisson bracket of the Noether charges of the ensuing (super-centrally extended) field-theoretic realisation of the supersymmetry algebra. Consequently, we may launch a systematic study of the structure of field-theoretic deformations of the latter Lie superalgebra for a large class of super-2-form potentials using the particularly simple form \eqref{eq:susypriMT} of the primitive. We shall perform the study with view to identifying those deformations motivated by the simple (super-$\sigma$-model) mechanics of extended charged objects introduced in Sec.\,\ref{sec:Cartsgeom} whose associative completions (derived through imposition of the super-Jacobi identity) geometrise, upon integration to the deformed Lie supergroup, the Green--Schwarz super-3-cocycle in that they support supersymmetric super-2-forms which descend to the (corrected) primitives of $\,\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}$.\ In order to better understand the topology quantified by these deformations through isolation of contributions from the non-trivial topology $\,{\mathbb{R}}^4\x{\mathbb{S}}^1\x{\mathbb{S}}^5\,$ of the body of the super-target and those from the twisted sector of the Kosteleck\'y--Rabin quotient, the latter in direct relation to the previously identified deformations of the flat super-Minkowskian limit of the supergeometry under consideration, we shall rescale the (local) coordinates on the homogeneous space $\,{\rm SU}(2,2\,\vert\,4)/({\rm SO}(4,1)\x{\rm SO}(5))\,$ in a standard manner dual to \eqref{eq:IWrescale}, {\it i.e.}, uniformly as
\begin{eqnarray}\label{eq:AdSSResc}
\bigl(\theta_i^{\a\a' I},X_i^{\widehat a}\bigr)\longmapsto\bigl(R^{-\frac{1}{2}}\,\theta_i^{\a\a' I},R^{-1}\,X_i^{\widehat a}\bigr)\,,\quad i\in I
\end{eqnarray}
and write out the corresponding restrictions of the Maurer--Cartan super-1-form
\begin{eqnarray}
\sigma_i^*\theta_{\rm L}(Z_i)&=&\sum_{k=0}^\infty\,\tfrac{(-1)^k}{(k+1)!\,R^{\frac{k+1}{2}}}\,\bigg(\tfrac{1}{R^{\frac{k+1}{2}}}\,{\mathsf d} X_i^{\widehat a}\otimes{\mathsf T}_e{\rm Ad}_{{\rm e}^{-R^{-\frac{1}{2}}\,\theta_i^{\a\a' I}\,Q_{\a\a' I}}}\circ{\rm ad}^k_{X_i^{\widehat b}\,P_{\widehat b}}(P_{\widehat a})\cr\cr
&&\hspace{2.5cm}+{\mathsf d}\theta_i^{\a\a' I}\otimes{\rm ad}^k_{\theta_i^{\beta\b' J}\,Q_{\beta\b' J}}(Q_{\a\a' I})\bigg)\,. \label{eq:MCs1f}
\end{eqnarray}
Upon taking into account the structure equations \eqref{eq:AdSS}, alongside the expansion
\begin{eqnarray}\nn
{\rm e}^{-R^{-\frac{1}{2}}\,\theta_i^{\a\a' I}\,Q_{\a\a' I}}&=&1-\frac{1}{R^{\frac{1}{2}}}\,\theta_i^{\a\a' I}\,Q_{\a\a' I}-\frac{1}{2R}\,\theta_i^{\a\a' I}\,\theta_i^{\beta\b' J}\,Q_{\a\a' I}\,Q_{\beta\b' J}+O\bigl(R^{-\frac{3}{2}}\bigr)\,,
\end{eqnarray}
we may, next, study the asymptotics of the various components of the above pullback super-1-forms. We do that with view to understanding the asymptotic relation between the left-invariant primitive $\,\underset{\tx{\ciut{(2)}}}{\beta}\,$ and the non-supersymmetric curving \eqref{eq:sMincurv} of the Green--Schwarz super-1-gerbe on ${\rm sMink}^{1,9\,\vert\,D_{1,9}}$,\ and -- in so doing -- to geometrising the Metsaev--Tseytlin super-3-cocycle in relation to the well-understood super-Minkowskian construction of Part I. Thus, we eventually arrive at the result
\begin{eqnarray}\nn
\sigma_i^*\theta_{\rm L}(Z_i)=\tfrac{1}{R^{\frac{1}{2}}}\,{\mathsf d}\theta_i^{\a\a'I}\otimes Q_{\a\a'I}+\tfrac{1}{R}\,\bigl[\bigl({\mathsf d} X_i^{\widehat a}-{\mathsf i}\,\ovl\theta_i\,(\widehat\Gamma^{\widehat a}\otimes{\boldsymbol{1}})\,{\mathsf d}\theta_i\bigr)\otimes P_{\widehat a}+\tfrac{{\mathsf i}}{2}\,\ovl\theta_i\,(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2)\,{\mathsf d}\theta_i\otimes J_{\widehat a\widehat b}\bigr]\cr\cr
+\tfrac{1}{2R^{\frac{3}{2}}}\,\bigl[\bigl({\mathsf d} X_i^{\widehat a}-\tfrac{{\mathsf i}}{3}\,\ovl\theta_i\,(\widehat\Gamma^a\otimes{\boldsymbol{1}})\,{\mathsf d}\theta_i\bigr)\,(\widehat\Gamma_{\widehat a}\otimes\sigma_2)^{\a\a'I}_{\ \beta\b'J}+\tfrac{{\mathsf i}\,\varepsilon_{\widehat a\widehat b}}{3!}\,\bigl(\ovl\theta_i\,(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2)\,{\mathsf d}\theta_i\,(\widehat\Gamma_{\widehat a\widehat b}\otimes{\boldsymbol{1}})^{\a\a'I}_{\ \beta\b'J}\bigr]\,\theta_i^{\beta\b'J}\otimes Q_{\a\a' I}\cr\cr
-\tfrac{1}{2R^2}\,\bigl[\bigl({\mathsf i}\,\bigl({\mathsf d} X_i^{\widehat b}-\tfrac{{\mathsf i}}{6}\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat b}\otimes{\boldsymbol{1}}\bigr)\,{\mathsf d}\theta_i\bigr)\,\ovl\theta_i\,\bigl(\{\widehat\Gamma^{\widehat a},\widehat\Gamma_{\widehat b}\}\otimes\sigma_2\bigr)\,\theta_i-\tfrac{\varepsilon_{\widehat b\widehat c}}{12}\,\ovl\theta_i\,\bigl(\widehat\Gamma_{\widehat b\widehat c}\otimes\sigma_2\bigr)\,{\mathsf d}\theta_i\cdot\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a}\,\widehat\Gamma^{\widehat b\widehat c}\otimes{\boldsymbol{1}}\bigr)\,\theta_i\bigr)\otimes P_{\widehat a}\cr\cr
-\bigl(\tfrac{{\mathsf i}}{2}\,\bigl({\mathsf d} X_i^{\widehat c}-\tfrac{{\mathsf i}}{3!}\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat c}\otimes{\boldsymbol{1}}\bigr)\,{\mathsf d}\theta_i\bigr)\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a\widehat b}\,\widehat\Gamma_{\widehat c}\otimes{\boldsymbol{1}}\bigr)\,\theta_i-\tfrac{\varepsilon_{\widehat a\widehat b}}{4!}\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat c\widehat d}\otimes\sigma_2\bigr)\,{\mathsf d}\theta_i\cdot\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a\widehat b}\,\widehat\Gamma_{\widehat c\widehat d}\otimes\sigma_2\bigr)\,\theta_i\cr\cr
-\varepsilon_{\widehat a\widehat b}\,X_i^{\widehat a}\,{\mathsf d} X_i^{\widehat b}\bigr)\otimes J_{\widehat a\widehat b}\bigr]+O\bigl(R^{-\frac{5}{2}}\bigr)\,,
\end{eqnarray}
written in the shorthand notation
\begin{eqnarray}\nn
\widehat\Gamma^{\widehat a}=\eta^{\widehat a\widehat b}\,\widehat\Gamma_{\widehat b}\,.
\end{eqnarray}
From the above, we read off, in particular, the $R^{-1}$-expansion of the spinorial component of the Maurer--Cartan super-1-form
\begin{eqnarray}\nn
\sigma_i^*\Sigma^{\a\a'I}_{\rm L}(Z_i)&=&\tfrac{1}{R^{\frac{1}{2}}}\,{\mathsf d}\theta_i^{\a\a'I}+\tfrac{1}{2R^{\frac{3}{2}}}\,\bigl[\bigl({\mathsf d} X_i^{\widehat a}-\tfrac{{\mathsf i}}{3}\,\ovl\theta_i\,(\widehat\Gamma^a\otimes{\boldsymbol{1}})\,{\mathsf d}\theta_i\bigr)\,(\widehat\Gamma_{\widehat a}\otimes\sigma_2)^{\a\a'I}_{\ \beta\b'J}\cr\cr
&&+\tfrac{{\mathsf i}\,\varepsilon_{\widehat a\widehat b}}{3!}\,\bigl(\ovl\theta_i\,(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2)\,{\mathsf d}\theta_i\,(\widehat\Gamma_{\widehat a\widehat b}\otimes{\boldsymbol{1}})^{\a\a'I}_{\ \beta\b'J}\bigr]\,\theta_i^{\beta\b'J}+O\bigl(R^{-\frac{5}{2}}\bigr)\,,
\end{eqnarray}
and so also the asymptotics of the Green--Schwarz super-3-cocycle
\begin{eqnarray}\nn
\sigma_i^*\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}(Z_i)=\tfrac{1}{R^2}\,\bigl({\mathsf d} X_i^{\widehat a}-{\mathsf i}\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a}\otimes{\boldsymbol{1}}\bigr)\,{\mathsf d}\theta_i\bigr)\wedge{\mathsf d}\ovl\theta_i\wedge\bigl(\widehat\Gamma_{\widehat a}\otimes\sigma_3\bigr)\,{\mathsf d}\theta_i+O\bigl(R^{-3}\bigr)
\end{eqnarray}
and of its left-invariant primitive
\begin{eqnarray}\nn
\sigma_i^*\underset{\tx{\ciut{(2)}}}{\beta}(Z_i)&=&\tfrac{{\mathsf i}}{R}\,{\mathsf d}\ovl\theta_i\wedge({\boldsymbol{1}}\otimes\sigma_1)\,{\mathsf d}\theta_i+\tfrac{1}{R^2}\,\bigl[\ovl\theta_i\,\bigl(\widehat\Gamma_{\widehat a}\otimes\sigma_3\bigr)\,{\mathsf d}\theta_i\wedge\bigl({\mathsf d} X_i^{\widehat a}-\tfrac{{\mathsf i}}{3}\,\ovl\theta_i\,(\widehat\Gamma^a\otimes{\boldsymbol{1}})\,{\mathsf d}\theta_i\bigr)\cr\cr
&&-\tfrac{\varepsilon_{\widehat a\widehat b}}{3!}\,\ovl\theta_i\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes\sigma_1\bigr)\,{\mathsf d}\theta_i\wedge\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2\bigr)\,{\mathsf d}\theta_i\bigr]+O\bigl(R^{-3}\bigr)\,.
\end{eqnarray}
The latter two reveal the precise relation between the superdifferential forms on the curved (super)target and their flat-superspace counterparts when rewritten in the notation of App.\,\ref{app:CliffAdSS},
\begin{eqnarray}\nn
\sigma_i^*\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}(Z_i)&=&\tfrac{1}{R^2}\,\bigl({\mathsf d} X_i^{\widehat a}-{\mathsf i}\,\theta_i^{\a\a'I}\,\bigl(\mathcal{C}\,\unl\gamma^{\widehat a}\,\Delta^1_{\widehat a}\bigr)_{\a\a'I\beta\b'J}\,{\mathsf d}\theta_i^{\beta\b'J}\bigr)\wedge{\mathsf d}\theta_i^{\gamma\g'K}\wedge\bigl(\mathcal{C}\,\unl\gamma_{\widehat a}\,\Delta^2_{\widehat a}\bigr)_{\gamma\g'K\delta\d'L}\,{\mathsf d}\theta_i^{\delta\d'L}\cr\cr
&&+O\bigl(R^{-3}\bigr)\,,\cr\cr\cr
\sigma_i^*\underset{\tx{\ciut{(2)}}}{\beta}(Z_i)&=&\tfrac{{\mathsf i}}{R}\,{\mathsf d}\ovl\theta_i\wedge({\boldsymbol{1}}\otimes\sigma_1)\,{\mathsf d}\theta_i\cr\cr
&&+\tfrac{1}{R^2}\,\bigl[\theta_i^{\a\a'I}\,\bigl(\mathcal{C}\,\unl\gamma_{\widehat a}\,\Delta^2_{\widehat a}\bigr)_{\a\a'I\beta\b'J}\,{\mathsf d}\theta_i^{\beta\b'J}\wedge\bigl({\mathsf d} X_i^{\widehat a}-\tfrac{{\mathsf i}}{3}\,\theta_i^{\a\a'I}\,\bigl(\mathcal{C}\,\unl\gamma^{\widehat a}\,\Delta^1_{\widehat a}\bigr)_{\a\a'I\beta\b'J}\,{\mathsf d}\theta_i^{\beta\b'J}\bigr)\cr\cr
&&-\tfrac{\varepsilon_{\widehat a\widehat b}}{3!}\,\ovl\theta_i\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes\sigma_1\bigr)\,{\mathsf d}\theta_i\wedge\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2\bigr)\,{\mathsf d}\theta_i\bigr]+O\bigl(R^{-3}\bigr)\,,
\end{eqnarray}
with
\begin{eqnarray}\nn
\Delta^1_{\widehat a}=\left\{ \begin{array}{cl} {\boldsymbol{1}}\otimes{\boldsymbol{1}}\otimes\sigma_3 & \tx{if}\ \widehat a\in\ovl{0,4} \\
{\boldsymbol{1}}\otimes{\boldsymbol{1}}\otimes{\boldsymbol{1}} & \tx{if}\ \widehat a\in\ovl{5,9}\end{array}\right.\,,\qquad\qquad\Delta^2_{\widehat a}=\left\{ \begin{array}{cl} {\boldsymbol{1}}\otimes{\boldsymbol{1}}\otimes{\boldsymbol{1}} & \tx{if}\ \widehat a\in\ovl{0,4} \\
{\boldsymbol{1}}\otimes{\boldsymbol{1}}\otimes\sigma_3 & \tx{if}\ \widehat a\in\ovl{5,9}\end{array}\right.
\end{eqnarray}
Thus, while the leading asymptotics of the Green--Schwarz super-3-cocycle reproduces the super-Minkowskian object \eqref{eq:sMinkGS3} upon evaluation on an ${\rm SU}(2,2\,\vert\,4)$-spinor $\,\theta=\psi\otimes\psi'\otimes\binom{1}{0}\,$ of positive chirality, along the lines of \Rcite{Metsaev:1998it}, its primitive may do that only upon substraction of the leading term of its $R^{-1}$-expansion. This prompts us to consider, in what follows, the distinguished (local) primitives
\begin{eqnarray}\label{eq:AdSprimcorr}\qquad
\underset{\tx{\ciut{(2)}}}{{\rm B}_i}=\sigma_i^*\underset{\tx{\ciut{(2)}}}{\beta}-\underset{\tx{\ciut{(2)}}}{{\rm D}_i}\,,\qquad\qquad\underset{\tx{\ciut{(2)}}}{{\rm D}_i}(Z_i)={\mathsf i}\,{\mathsf d}\ovl\theta_i\wedge({\boldsymbol{1}}\otimes\sigma_1)\,{\mathsf d}\theta_i
\end{eqnarray}
of the $\,\sigma_i^*\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}\,$ (written out above in their coordinate form prior to the rescaling). These are manifestly non-invariant, and so it seems apposite to look for an extension of the original supersymmetry algebra whose integration would allow for a global supersymmetric trivialisation of the Metsaev--Tseytlin super-3-cocycle. Our quest starts with an analysis of the natural field-theoretic sources of a sought-after deformation of $\,\gt{su}(2,2\,\vert\,4)$.
We begin our computation of the wrapping anomaly by noting that the ${\rm SU}(2,2\,\vert\,4)$-invariant primitive $\,\underset{\tx{\ciut{(2)}}}{\beta}\,$ is ${\rm SO}(4,1)\x{\rm SO}(5)$-horizontal in the sense of \Reqref{eq:Binvhor}, and so
\begin{eqnarray}\nn
\sigma_{i_\tau}^*\bigl(\mathcal{K}_{X_2}\righthalfcup\mathcal{K}_{X_1}\righthalfcup\underset{\tx{\ciut{(2)}}}{\beta}\bigr)=2{\mathsf i}\,{\rm Ad}_{\sigma_{i_\tau}(\cdot)^{-1}}(X_1)^{\a\a' I}\,{\rm Ad}_{\sigma_{i_\tau}(\cdot)^{-1}}(X_2)^{\beta\b' J}\,(\sigma_1)_{IJ}\,C_{\a\beta}\,C_{\a'\beta'}'\,.
\end{eqnarray}
Above, and in what follows, we are using the notation for the standard Pauli matrices
\begin{eqnarray}\nn
(\sigma_A)_{IJ}\equiv\delta_{IK}\,(\sigma_A)^K_{\ J}\equiv(\sigma_A)_I^{\ K}\,\delta_{KJ}\,,\quad A\in\{1,2,3\}\,.
\end{eqnarray}
In keeping with Eqs.\,\eqref{eq:wrappaninvB}, \eqref{eq:monofAB} and \eqref{eq:fAB}, we obtain the wrapping anomaly in the form
\begin{eqnarray}\nn
&\mathscr{W}^{({\rm inv})}_{X_1,X_2}[\unl\xi]=\mu\bigl(f_{X_1,X_2};\xi(\mathscr{C}_1)\bigr)\,,&\cr\cr
&f_{X_1,X_2}\mathord{\restriction}_{\mathcal{O}_i}\equiv 2{\mathsf i}\,{\mathsf T}_e{\rm Ad}_{\sigma_i(\cdot)^{-1}}(X_1)^{\a\a' I}\,\bigl(\widehat C\otimes\sigma_1\bigr)_{\a\a'I\ \beta\b'J}\,{\mathsf T}_e{\rm Ad}_{\sigma_i(\cdot)^{-1}}(X_2)^{\beta\b' J}\,.&
\end{eqnarray}
Explicit expressions for the $\,{\rm Ad}_{\sigma_{i_\tau}(\cdot)^{-1}}(X)^{\a\a' I}\,$ corresponding to the various vectors
\begin{eqnarray}\nn
\Delta\in\{\varepsilon\equiv\varepsilon^{\a\a' I}\,Q_{\a\a' I},Y\equiv Y^{\widehat a}\,P_{\widehat a}\equiv\bigl(y^a\,P_a,y'{}_{a'}\,P_{a'}\bigr),\Lambda\equiv\Lambda^{\widehat a\widehat b}\,J_{\widehat a\widehat b}\equiv\bigl(\lambda^{ab}\,J_{ab},\lambda'{}^{a'b'}\,J_{a'b'}\bigr)\}
\end{eqnarray}
in the Lie superalgebra $\,\mathfrak{su}(2,2\,\vert\,4)\,$ may be readily obtained from those for $\,\unl g_i=e\,$ ({\it i.e.}, computed in the neighbourhood of the supergroup unit) gathered in \Rcite{Hatsuda:2002hz}. These split as
\begin{eqnarray}\nn
{\mathsf T}_e{\rm Ad}_{\sigma_i(Z_i)^{-1}}(\Delta)^{\a\a' I}\equiv{\mathsf T}_e{\rm Ad}_{\sigma_i(Z_i)^{-1}}(\Delta)_Q^{\a\a' I}+{\mathsf T}_e{\rm Ad}_{\sigma_i(Z_i)^{-1}}(\Delta)_P^{\a\a' I}+{\mathsf T}_e{\rm Ad}_{\sigma_i(Z_i)^{-1}}(\Delta)_J^{\a\a' I}\,,
\end{eqnarray}
with components that read
\begin{eqnarray}\nn
{\mathsf T}_e{\rm Ad}_{\sigma_i(Z_i)^{-1}}(\Delta)_Q^{\a\a' I}&=&(\cos\Psi)^{\a\a' I}_{\ \beta\b' J}(\theta_i)\,{\rm exp}\bigl(-\tfrac{1}{2}\,X_i^{\widehat a}\,\widehat\Gamma_{\widehat a}\otimes\sigma_2\bigr)^{\beta\b' J}_{\ \gamma\g' K}\,\unl\Gamma_{i\ A}^{\gamma\g' K}\,\Delta^A\,,\cr\cr
{\mathsf T}_e{\rm Ad}_{\sigma_i(Z_i)^{-1}}(\Delta)_P^{\a\a' I}&=&\tfrac{1}{4}\,\bigl(\tfrac{\sin\Psi}{\Psi}\bigr)^{\a\a' I}_{\ \beta\b' J}(\theta_i)\,\bigl(2\widetilde Y_{i\ A}^{\widehat a}(X_i)\,\widehat\Gamma_{\widehat a}\otimes\sigma_2+\varepsilon_{\widehat a\widehat b}\,\widetilde{\widetilde Y}_{i\ A}^{\widehat a\widehat b}(X_i)\,\widehat\Gamma_{\widehat a\widehat b}\otimes{\boldsymbol{1}}\bigr)^{\beta\b'J}_{\ \gamma\g'K}\,\theta_i^{\gamma\g' K}\,\Delta^A\,,\cr\cr
{\mathsf T}_e{\rm Ad}_{\sigma_i(Z_i)^{-1}}(\Delta)_J^{\a\a' I}&=&\tfrac{1}{4}\,\bigl(\tfrac{\sin\Psi}{\Psi}\bigr)^{\a\a' I}_{\ \beta\b' J}(\theta_i)\,\bigl(2\widetilde\Lambda_{i\ A}^{\widehat a}(X_i)\,\widehat\Gamma_{\widehat a}\otimes\sigma_2+\varepsilon_{\widehat a\widehat b}\,\widetilde{\widetilde\Lambda}_{i\ A}^{\widehat a\widehat b}(X_i)\,\widehat\Gamma_{\widehat a\widehat b}\otimes{\boldsymbol{1}}\bigr)^{\beta\b'J}_{\ \gamma\g'K}\,\theta_i^{\gamma\g' K}\,\Delta^A\,,
\end{eqnarray}
where\footnote{It is to be noted that the functions with the matrix argument $\,\Psi\,$ appearing in the expressions listed can all be expressed as power series in $\,\Psi^2$.\ (In fact, these series are finite owing to the anticommutativity of the Gra\ss mann-odd coordinates.)}
\begin{eqnarray}\nn
\bigl(\Psi^2\bigr)^{\a\a' I}_{\ \beta\b' J}(\theta_i)&=&{\mathsf i}\,\bigl(\bigl(\widehat\Gamma^{\widehat a}\otimes\sigma_2\bigr)^{\a\a' I}_{\ \gamma\g' K}\,(\widehat C\,\widehat\Gamma_{\widehat a}\otimes{\boldsymbol{1}})_{\delta\d' L\ \beta\b' J}\cr\cr
&&-\tfrac{1}{2}\,\varepsilon_{\widehat a\widehat b}\,\bigl(\widehat\Gamma^{\widehat a\widehat b}\otimes{\boldsymbol{1}}\bigr)^{\a\a' I}_{\ \gamma\g' K}\,\bigl(\widehat C\,\widehat\Gamma_{\widehat a\widehat b}\otimes\sigma_2\bigr)_{\delta\d' L\ \beta\b' J}\bigr)\,\theta_i^{\gamma\g' K}\,\theta_i^{\delta\d' L}
\end{eqnarray}
and
\begin{eqnarray}\nn
&\widetilde Y_i^{\widehat a}(X_i)\equiv\bigl(\widetilde y_i^a(x_i^d),\widetilde y_i'{}^{a'}(x_i'{}^{d'})\bigr)\,,\qquad\qquad\widetilde{\widetilde Y}_i^{\widehat a\widehat b}(X_i)\equiv\bigl(\widetilde{\widetilde y}_i^{ab}(x_i^c),\widetilde{\widetilde y}_i'{}^{a'b'}(x_i'{}^{c'})\bigr)\,,&\cr\cr
&\widetilde\Lambda_i^{\widehat a}(X_i)\equiv\bigl(\widetilde\lambda_i^a(x_i^d),\widetilde\lambda_i'{}^{a'}(x_i'{}^{d'})\bigr)\,,\qquad\qquad\widetilde{\widetilde\Lambda}_i^{\widehat a\widehat b}(X_i)\equiv\bigl(\widetilde{\widetilde\lambda}_i^{ab}(x_i^c),\widetilde{\widetilde\lambda}_i'{}^{a'b'}(x_i'{}^{c'})\bigr)\,,&
\end{eqnarray}
with
\begin{eqnarray}\nn
\widetilde y_i^a(x_i^d)&=&\unl\Gamma^b_{i\ A}\bigl(\delta^a_{\ b}+({\rm ch}\,x_i-1)\,\bigl(\delta_{\ b}^a-\tfrac{\eta_{bc}\,x_i^c\,x_i^a}{x_i^2}\bigr)\bigr)\,,\qquad\qquad\widetilde{\widetilde y}^{ab}(x_i^c)=-x_i^{[a}\,\unl\Gamma_{i\ A}^{b]}\,\tfrac{{\rm sh}\,x_i}{x_i}\,,\cr\cr\cr
\widetilde y_i'{}^{a'}(x_i'{}^{d'})&=&\unl\Gamma^{b'}_{i\ A}\bigl(\delta_{\ b'}^{a'}+(\cos x_i'-1)\,\bigl(\delta_{\ b'}^{a'}-\tfrac{\delta_{b'c'}\,x_i'{}^{c'}\,x_i'{}^{a'}}{x_i'{}^2}\bigr)\bigr)\,,\qquad\qquad\widetilde{\widetilde y}_i'{}^{a'b'}(x_i'{}^{c'})=x_i'{}^{[a'}\,\unl\Gamma_{i\ A}^{b']}\,\tfrac{\sin x_i'}{x_i'}\cr\cr\cr
\widetilde\lambda_i^a(x_i^e)&=&\unl\Gamma_{i\ A}^{ab}\,\eta_{bc}\,x_i^c\,\tfrac{{\rm sh}\,x_i}{x_i}\,,\qquad\qquad\widetilde{\widetilde\lambda}_i^{ab}(x_i^c)=\unl\Gamma_{i\ A}^{ab}+\tfrac{{\rm ch}\,x_i-1}{x_i^2}\,x_i^{[a}\,\unl\Gamma_{i\ A}^{b]c}\,\eta_{cd}\,x_i^d\,,\cr\cr\cr
\widetilde\lambda_i'{}^{a'}(x_i'{}^{d'})&=&\unl\Gamma_{i\ A}^{a'b'}\,\delta_{b'c'}\,x_i'{}^{c'}\,\tfrac{\sin x_i'}{x_i'}\,,\qquad\qquad\widetilde{\widetilde\lambda}_i'{}^{a'b'}(x_i'{}^{e'})=\unl\Gamma_{i\ A}^{a'b'}+\tfrac{\cos x_i'-1}{x_i'{}^2}\,x_i'{}^{[a'}\,\unl\Gamma_{i\ A}^{b']c'}\,\delta_{c'd'}\,x_i'{}^{d'}\,,\cr\cr\cr
x_i&=&\sqrt{\eta_{ab}\,x_i^a\,x_i^b}\equiv\sqrt{x_i^2}\,,\qquad\qquad x_i'=\sqrt{\delta_{a'b'}\,x_i'{}^{a'}\,x_i'{}^{b'}}\equiv\sqrt{x_i'{}^2}\,,
\end{eqnarray}
and where the matrix elements
\begin{eqnarray}\nn
\unl\Gamma^A_{i\,\ B}\equiv\bigl({\mathsf T}_e{\rm Ad}_{\unl g_i^{-1}}\bigr)^A_{\ B}
\end{eqnarray}
capture conjugation of the variation vector by the reference supergroup element $\,\unl g_i$.\ The ensuing Poisson algebra of the Noether charges,
\begin{eqnarray}\nn
\{h_{\Delta_1},h_{\Delta_2}\}_{\Omega^{\rm (NG)}_{{\rm GS},1}}[X,\theta,\pi]=-h_{[\Delta_1,\Delta_2]}[X,\theta,\pi]+\mu\bigl(f_{\Delta_1,\Delta_2};(X,\theta)(\mathscr{C}_1)\bigr)\,,
\end{eqnarray}
has a fairly complicated local coordinate (or, equivalently, current) presentation. In order to draw concrete qualitative conclusions as to the (super)geometric nature of the wrapping-charge deformation present in it and, in so doing, establish a structural relation with the formerly discussed superstring deformation of the super-Minkowski superalgebra, understood as the flat-superspace limit of the superalgebra under consideration, we shall rescale the coordinates as previously ({\it cp} \Reqref{eq:AdSSResc}) and study the asymptotics of the above expressions in the r\'egime of large $\,R$,\ which corresponds to a correlated flattening of both: the 1-cycle in $\,{\rm AdS}_5\,$ and the 5-sphere $\,{\mathbb{S}}^5$,\ and an attendant uniform rescaling of the fibre of the Gra\ss mann bundle over them. We further zoom in on a neighbourhood of the group unit, that is we examine the local presentation of the functions $\,f_{\Delta_1,\Delta_2}\,$ over the (blown-up) coordinate patch $\,\mathcal{O}_{i*}\,$ centred on $\,\unl g_{i*}\equiv e$.\ We thus obtain
\begin{eqnarray}\nn
{\rm Ad}_{\sigma_{i*}(Z_{i*})^{-1}}(\varepsilon)^{\a\a' I}&=&\varepsilon^{\a\a' I}-\tfrac{1}{2R}\,\bigl(\bigl(\Psi^2\bigr)^{\a\a' I}_{\ \beta\b' J}(\theta_{i*})+X_{i*}^{\widehat a}\,\bigl(\widehat\Gamma_{\widehat a}\otimes\sigma_2\bigr)^{\a\a'I}_{\ \beta\b'J}\bigr)\,\varepsilon^{\beta\b'J}+O\bigl(R^{-2}\bigr)\,,\cr\cr\cr
{\rm Ad}_{\sigma_{i*}(Z_{i*})^{-1}}(Y)^{\a\a' I}&=&\tfrac{1}{2R^{\frac{1}{2}}}\,Y^{\widehat a}\,\bigl(\widehat\Gamma_{\widehat a}\otimes\sigma_2\bigr)^{\a\a'I}_{\ \beta\b'J}\,\theta_{i*}^{\beta\b' J}+O\bigl(R^{-\frac{3}{2}}\bigr)\,,\cr\cr\cr
{\rm Ad}_{\sigma_{i*}(Z_{i*})^{-1}}(\Lambda)^{\a\a' I}&=&\tfrac{1}{4R^{\frac{1}{2}}}\,\varepsilon_{\widehat a\widehat b}\,\Lambda^{\widehat a\widehat b}\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes{\boldsymbol{1}})^{\a\a'I}_{\ \beta\b'J}\,\theta_{i*}^{\beta\b' J}+O\bigl(R^{-\frac{3}{2}}\bigr)\,,
\end{eqnarray}
whence also the leading local asymptotics\footnote{In its derivation, we used the symmetry properties: $\,C^{\rm T}=-C,\ C'{}^{\rm T}=-C'\,$ and $\,(C\,\Gamma_a)^{\rm T}=-C\,\Gamma_a,\ (C'\,\Gamma_{a'})^{\rm T}=-C'\,\Gamma_{a'}$,\ as well as the absence of winding states on $\,{\mathbb{S}}^5$.}
\begin{eqnarray}\nn
f_{\varepsilon_1,\varepsilon_2}\mathord{\restriction}_{\mathcal{O}_{i*}}-2{\mathsf i}\,\ovl\varepsilon_1\,({\boldsymbol{1}}\otimes\sigma_1)\,\varepsilon_2\cr\cr
=-\tfrac{2{\mathsf i}}{R}\,\varepsilon_1^{\a\a' I}\,\varepsilon_2^{\beta\b' J}\,\bigl[\bigl(\widehat C\otimes\sigma_1\bigr)_{\gamma\g'K\delta\d'L}\,\delta^{\gamma\g'K}_{\ (\a\a'I}\,\delta^{\epsilon\ep'L}_{\ \beta\b'J)}\,\bigl(\Psi^2\bigr)^{\delta\d'L}_{\ \epsilon\ep'M}(\theta_{i*})+{\mathsf i}\,\bigl(\widehat C\,\widehat\Gamma_{\widehat a}\otimes\sigma_3\bigr)_{\a\a'I\beta\b'J}\,X_{i*}^{\widehat a}\bigr]+O\bigl(R^{-2}\bigr)\cr\cr\cr
f_{Y_1,Y_2}\mathord{\restriction}_{\mathcal{O}_{i*}}=\tfrac{{\mathsf i}}{2R}\,Y_1^{\widehat a}\,Y_2^{\widehat b}\,\ovl\theta_{i*}\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes\sigma_1\bigr)\,\theta_{i*}+O\bigl(R^{-2}\bigr)\,,\cr\cr\cr
f_{\Lambda_1,\Lambda_2}\mathord{\restriction}_{\mathcal{O}_{i*}}=-\tfrac{{\mathsf i}}{8R}\,\varepsilon_{\widehat a\widehat b}\,\varepsilon_{\widehat c\widehat d}\,\Lambda_1^{\widehat a\widehat b}\,\Lambda_2^{\widehat c\widehat d}\,\ovl\theta_{i*}\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\,\widehat\Gamma_{\widehat c\widehat d}\otimes\sigma_1\bigr)\,\theta_{i*}+O\bigl(R^{-2}\bigr)\cr\cr
=-\tfrac{{\mathsf i}}{8R}\,\bigl(\lambda_1^{ab}\,\lambda_2^{cd}\,\ovl\theta_{i*}\,\bigl(\widehat\Gamma_{ab}\,\widehat\Gamma_{cd}\otimes\sigma_1\bigr)\,\theta_{i*}+\lambda_1'{}^{a'b'}\,\lambda_2'{}^{c'd'}\,\ovl\theta_{i*}\,\bigl(\widehat\Gamma_{a'b'}\,\widehat\Gamma_{c'd'}\otimes\sigma_1\bigr)\,\theta_{i*}\bigr)+O\bigl(R^{-2}\bigr)\,,\cr\cr\cr
f_{\varepsilon,Y}\mathord{\restriction}_{\mathcal{O}_{i*}}=-\tfrac{1}{R^{\frac{1}{2}}}\,Y^{\widehat a}\,\ovl\varepsilon\,\bigl(\widehat\Gamma_{\widehat a}\otimes\sigma_3\bigr)\,\theta_{i*}+O\bigl(R^{-\frac{3}{2}}\bigr)\,,\cr\cr\cr
f_{\varepsilon,\Lambda}\mathord{\restriction}_{\mathcal{O}_{i*}}=\tfrac{{\mathsf i}\,\varepsilon_{\widehat a\widehat b}}{2R^{\frac{1}{2}}}\,\Lambda^{\widehat a\widehat b}\,\ovl\varepsilon\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes\sigma_1\bigr)\,\theta_{i*}+O\bigl(R^{-\frac{3}{2}}\bigr)\,,\cr\cr\cr
f_{Y,\Lambda}\mathord{\restriction}_{\mathcal{O}_{i*}}=-\tfrac{\varepsilon_{\widehat b\widehat c}}{4R}\,Y^{\widehat a}\,\Lambda^{\widehat b\widehat c}\,\ovl\theta_{i*}\,\bigl(\widehat\Gamma_{\widehat a}\,\widehat\Gamma_{\widehat b\widehat c}\otimes\sigma_3\bigr)\,\theta_{i*}+O\bigl(R^{-2}\bigr)\,.
\end{eqnarray}
Our asymptotic analysis identifies, on the qualitative level adopted, two independent sources of the anomaly: the Gra\ss mann-even winding modes that wrap (any representative of) the generator of $\,H_1({\rm AdS}_5\x{\mathbb{S}}^5)\,$ that sits in $\,{\rm AdS}_5\cong{\mathbb{R}}^{\x 4}\x{\mathbb{S}}^1\,$ (represented by the local coordinate functions $\,X^a\,$ with nontrivial monodromy), and the Gra\ss man-odd Kosteleck\'y--Rabin states (represented by the coordinate functions $\,\theta^{\a\a'I}$). Clearly, the wrapping anomaly is sourced by the latter in the leading order (in $\,R$). On the other hand, the subleading contribution from the standard winding charges from the $\,{\rm AdS}_5\cong{\mathbb{R}}^{\x 4}\x{\mathbb{S}}^1\,$ sector represents an irremovable geometric effect, and so it cannot be dropped.
At this stage, there are at least two inequivalent paths that can be taken to arrive at a geometrisation of the Metsaev--Tseytlin super-3-cocycle conformable with the supersymmetry present and leading through a `superstringy' deformation of the original Lie superalgebra $\,\gt{su}(2,2\,\vert\,4)$,\ to wit, we may
\begin{itemize}
\item[(i)] take the left-invariant primitive $\,\underset{\tx{\ciut{(2)}}}{\beta}\,$ of the Metsaev--Tseytlin super-3-cocycle as the point of departure of the geometrisation and work with the undeformed Lie superalgebra $\,\gt{su}(2,2\,\vert\,4)\,$ upon dropping the wrapping charges altogether, or
\item[(ii)] replace $\,\underset{\tx{\ciut{(2)}}}{\beta}\,$ with the non-invariantly corrected primitive \eqref{eq:AdSprimcorr} and subsequently look for deformations of $\,\gt{su}(2,2\,\vert\,4)\,$ corresponding, along the lines of Part I, to extended superspacetimes surjectively submersed over $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,$ on which left-invariant super-2-forms exist with the desired asymptotics of the corrections $\,\underset{\tx{\ciut{(2)}}}{{\rm D}_i}$.
\end{itemize}
The first path is accessible owing to the invariant nature of the primitive $\,\underset{\tx{\ciut{(2)}}}{\beta}\,$ but, as noted earlier, the resultant super-1-gerbe does not asymptote to the one over $\,{\rm sMink}^{1,9\,\vert\,D_{1,9}}\,$ constructed in Part I. The second one calls for a systematic exploration of deformations (or, indeed, extensions) admissible in the highly constrained category of Lie superalgebras in which we can draw hints from the previous studies of the mother supersymmetry (super)algebra $\,\gt{osp}(1\,\vert\,32)\,$ of the superstring and M-theory, {\it cp} \Rcite{Bergshoeff:2000qu}. We shall examine the former path in full detail, with our analysis culminating in a full-fledged definition of the relevant super-1-gerbe with curving $\,\underset{\tx{\ciut{(2)}}}{\beta}$.\ As for the latter one, we shall explore a large class of physically motivated deformations, only to conclude that there are no deformations with the desired properties (within the class considered). These results lead us to contemplate the third path along which we
\begin{itemize}
\item[(iii)] seek an alternative ($\neq\underset{\tx{\ciut{(2)}}}{\beta}$) left-invariant trivialisation of $\,\underset{\tx{\ciut{(3)}}}{\chi}\,$ only on the total space of a surjective submersion over $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,$ that integrates a (physically motivated) associative deformation of $\,\gt{su}(2,2\,\vert\,4)$.
\end{itemize}
The last path is -- by far -- the wildest, with no more than the flat-superspace intuition as guidance at best. Partial negative results in this direction will be obtained below in the course of our study of consistent `superstring' deformations of the basis Lie superalgebra. These are, ultimately, conducive to a rethinking of the very definition\footnote{One might, {\it e.g.}, replace the Metsaev--Tseytlin super-3-cocycle with another one, having the same asymptotic behaviour in the flat limit $\,R\to\infty$.} of the Green--Schwarz super-3-cocycle of the super-$\sigma$-model of interest along the most general guidelines of the original paper \cite{Metsaev:1998it}, an idea that we pursue in Section \ref{sec:MTaway}. The investigation thus initiated will be taken up in an upcoming work.
\void{The third one encompasses the largest, by far, spectrum of possibilities, justified by the variety of wrapping charges discovered above and otherwise unconstrained by the coherence conditions that follow from the super-Jacobi identity -- while tempting (and, with some hindsight, not all that intimidating), it confronts us with the challenge -- and a prerequisite -- of defining a meaningful extension of the purely geometric supersymmetry to the extended superspace based on a superalgebra with a non-vanishing super-Jacobiator, and subsequently identifying left-invariant super-1-forms on that superspace, projecting to super-1-forms on $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,$ with the desired flat-superspace asymptotics. In the remainder of the present work, we shall navigate carefully through the wealth of structures implied by each of the three paths.
where we have introduced the suggestive shorthand notation for the charges: $\,{}^{\rm KR}\hspace{-2pt}\mathcal{Z}^{\gamma\g' L}_{\ \lambda\la' M}\,$ and $\,\,{}^{\rm KR}\hspace{-2pt}\mathcal{Z}_2^{\a\a' I\,\beta\b' J}\,$ for those sourced by the Kosteleck\'y--Rabin winding $\,(\Psi^2)^{\gamma\g' L}_{\ \a\a' I}(\theta(\varphi))\vert^{\varphi=2\pi}_{\varphi=0}\,$ and $\,\theta^{\a\a' I}(\varphi)\,\theta^{\beta\b' J}(\varphi)\vert^{\varphi=2\pi}_{\varphi=0}$,\ respectively, and $\,{}^{{\mathbb{S}}^1}\hspace{-3pt}\mathcal{Z}_{\a\a'I\,\beta\b'J}\,$ for those sourced by the geometric winding\footnote{It makes sense to present it as a spacetime scalar with Gra\ss mannian indices as the charge corresponds to the \emph{single} generator of $\,H_1({\rm AdS}_5)$.} $\,2{\mathsf i}\,(\sigma_3)_{IJ}\,(C\,\Gamma_a)_{\a\beta}\,C_{\a'\beta'}'\,\Delta x^a$.\ Note that the latter is (the leading term of) the winding charge found in \Rcite{Hatsuda:2002hz}. Having derived the superalgebra extended in the manner dictated by the field theory of interest, we may reconsider the relative relevance of the super-central corrections of various origin upon rescaling the original generators as
\begin{eqnarray}\nn
(Q_{\a\a' I},P_{\widehat a},J_{\widehat b\widehat c})\longmapsto(R^{\frac{1}{2}}\,Q_{\a\a' I},R\,P_{\widehat a},J_{\widehat b\widehat c})\,,
\end{eqnarray}
that is, along the lines of the standard \.In\"on\"u--Wigner contraction. Upon rescaling, we arrive at the algebra
\begin{eqnarray}\nn
\{Q_{\a\a' I},Q_{\beta\b' J}\}=-2{\mathsf i}\,\delta_{IJ}\,\bigl((C\,\Gamma_a)_{\a\beta}\,C_{\a'\beta'}'\,\eta^{ab}\,P_b+C_{\a\beta}\,(C'\,\Gamma_{a'})_{\a' \beta'}\,\delta^{a'b'}\,P_{b'}\bigr)\cr\cr
+\tfrac{1}{R}\,(\sigma_2)_{IJ}\,\bigl((C\,\Gamma_{ab})_{\a'\beta}\,C_{\a'\beta'}\,\eta^{ac}\,\eta^{bd}\,J_{cd}+C_{\a\beta}\,(C'\,\Gamma_{a'b'})_{\a'\beta'}\,\delta^{a'c'}\,\delta^{b'd'}\,J_{c'd'}\bigr)\cr\cr
-\tfrac{1}{R^2}\,\bigl[\bigl((\sigma_1)_{IK}\,C_{\a\gamma}\,C_{\a'\gamma'}'\,\delta^{\lambda\la' L}_{\ \beta\b' J}-(\sigma_1)_{JK}\,C_{\beta\gamma}\,C_{\beta'\gamma'}'\,\delta^{\lambda\la' L}_{\ \a\a' I}\bigr)\,{}^{\rm KR}\hspace{-2pt}\mathcal{Z}^{\gamma\g' K}_{\ \lambda\la' L}+{}^{{\mathbb{S}}^1}\hspace{-3pt}\mathcal{Z}_{\a\a'I\,\beta\b'J}\bigr]+O\bigl(R^{-3}\bigr)\,,\cr\cr\cr
[P_a,P_b]=\tfrac{1}{R^2}\,J_{ab}+O\bigl(R^{-3}\bigr)\,,\qquad\qquad[P_{a'},P_{b'}]=-\tfrac{1}{R^2}\,J_{a'b'}+O\bigl(R^{-3}\bigr)\,,\qquad\qquad[P_a,P_{a'}]=O\bigl(R^{-4}\bigr)\,,\cr\cr\cr
[J_{ab},J_{cd}]=\eta_{ad}\,J_{bc}-\eta_{ac}\,J_{bd}+\eta_{bc}\,J_{ad}-\eta_{bd}\,J_{ac}+\tfrac{1}{16R}\,(\sigma_1)_{IJ}\,(C\,\Gamma_{ab}\,\Gamma_{cd})_{\a\beta}\,C_{\a'\beta'}'\,{}^{\rm KR}\hspace{-2pt}\mathcal{Z}_2^{\a\a' I\,\beta\b' J}+O\bigl(R^{-2}\bigr)\,,\cr\cr\cr
[J_{a'b'},J_{c'd'}]=\delta_{a'd'}\,J_{b'c'}-\delta_{a'c'}\,J_{b'd'}+\delta_{b'c'}\,J_{a'd'}-\delta_{b'd'}\,J_{a'c'}+\tfrac{1}{16R}\,(\sigma_1)_{IJ}\,C_{\a'\beta'}\,(C'\,\Gamma_{a'b'}'\,\Gamma_{c'd'}')_{\a'\beta'}\,{}^{\rm KR}\hspace{-2pt}\mathcal{Z}_2^{\a\a' I\,\beta\b' J}\cr\cr
+O\bigl(R^{-2}\bigr)\,,\cr\cr\cr
[J_{ab},J_{a'b'}]=O\bigl(R^{-2}\bigr)\,,\cr\cr\cr
[Q_{\a\a' I},P_a]=\tfrac{1}{2R^{\frac{3}{2}}}\,(\sigma_2)_I^{\ J}\,(\Gamma_a)^\beta_{\ \a}\,\delta^{\beta'}_{\ \a'}\,Q_{\beta\b' J}-\tfrac{{\mathsf i}}{R^2}\,(\sigma_3)_{IJ}\,(C\,\Gamma_a)_{\a\beta}\,C_{\a'\beta'}'\,{}^{\rm KR}\hspace{-2pt}\mathcal{Z}_1^{\beta\b'J}+O\bigl(R^{-3}\bigr)\,,\cr\cr\cr
[Q_{\a\a' I},P_{a'}]=\tfrac{1}{2R^{\frac{3}{2}}}\,(\sigma_2)_I^{\ J}\,\delta^\beta_{\ \a}\,(\Gamma_{a'})^{\beta'}_{\ \a'}\,Q_{\beta\b' J}-\tfrac{{\mathsf i}}{R^2}\,(\sigma_3)_{IJ}\,C_{\a\beta}\,(C'\,\Gamma_{a'})_{\a' \beta'}\,{}^{\rm KR}\hspace{-2pt}\mathcal{Z}_1^{\beta\b'J}+O\bigl(R^{-3}\bigr)\,,\cr\cr\cr
[Q_{\a\a' I},J_{ab}]=-\tfrac{1}{2R^{\frac{1}{2}}}\,(\Gamma_{ab})^\beta_{\ \a}\,\delta^{\beta'}_{\ \a'}\,Q_{\beta\b' I}-\tfrac{{\mathsf i}}{2R}\,(\sigma_3)_{IJ}\,(C\,\Gamma_{ab})_{\a\beta}\,C_{\a'\beta'}'\,{}^{\rm KR}\hspace{-2pt}\mathcal{Z}_1^{\beta\b'J}+O\bigl(R^{-2}\bigr)\,,\cr\cr\cr
[Q_{\a\a' I},J_{a'b'}]=-\tfrac{1}{2R^{\frac{1}{2}}}\,\delta^\beta_{\ \a}\,(\Gamma_{a'b'})^{\beta'}_{\ \a'}\,Q_{\beta\b' I}-\tfrac{{\mathsf i}}{2R}\,(\sigma_3)_{IJ}\,C_{\a\beta}\,(C'\,\Gamma_{a'b'})_{\a'\beta'}\,{}^{\rm KR}\hspace{-2pt}\mathcal{Z}_1^{\beta\b'J}+O\bigl(R^{-2}\bigr)\,,\cr\cr\cr
[P_a,J_{bc}]=\eta_{ab}\,P_c-\eta_{ac}\,P_b+\tfrac{1}{4R^2}\,(\sigma_1)_{IJ}\,(C\,\Gamma_a\,\Gamma_{bc})_{\a\beta}\,C_{\a'\beta'}'\,{}^{\rm KR}\hspace{-2pt}\mathcal{Z}_2^{\a\a' I\,\beta\b' J}+O\bigl(R^{-3}\bigr)\,,\cr\cr\cr
[P_{a'},J_{b'c'}]=\delta_{a'b'}\,P_{c'}-\delta_{a'c'}\,P_{b'}+\tfrac{1}{4R^2}\,(\sigma_1)_{IJ}\,C_{\a\beta}\,(C'\,\Gamma_{a'}\,\Gamma_{b'c'})_{\a'\beta'}\,{}^{\rm KR}\hspace{-2pt}\mathcal{Z}_2^{\a\a' I\,\beta\b' J}+O\bigl(R^{-3}\bigr)\,,\cr\cr\cr
[P_a,J_{a'b'}]=O\bigl(R^{-3}\bigr)=[P_{a'},J_{ab}]\,.
\end{eqnarray}
It is to be noted that from the point of view of the flattening procedure for $\,\mathfrak{su}(2,2\,\vert\,4)\,$ the two types of winding charges correspond to corrections of the same order (in $\,R$) of the super-Minkowskian algebra.}
\section{The trivial Green--Schwarz super-1-gerbe over the super-${\rm AdS}_5\x{\mathbb{S}}^5\,$ space}\label{sec:trsAdSSext}
The first of the three paths outlined in the previous section begins with the earlier observation that the Metsaev--Tseytlin super-3-cocycle admits a global manifestly (left-)${\rm SU}(2,2\,\vert\,4)$-invariant primitive $\,\underset{\tx{\ciut{(2)}}}{\beta}\in\Omega^2({\rm SU}(2,2\,\vert\,4))\,$ pulling back along the distinguished local sections $\,\sigma_i\,$ of \Reqref{eq:MTsect} to elements of the trivialising cover $\,\{\mathcal{O}_i\}_{i\in I}\subset\mathscr{T}({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5))\,$ of the base of the principal ${\rm SO}(4,1)\x{\rm SO}(5)$-bundle $\,{\rm SU}(2,2\,\vert\,4)\longrightarrow{\rm SU}(2,2\,\vert\,4)/({\rm SO}(4,1)\x{\rm SO}(5))\equiv{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,$ and thus defining restrictions
\begin{eqnarray}\nn
\underset{\tx{\ciut{(2)}}}{{\rm B}}\mathord{\restriction}_{\mathcal{O}_i}:={\mathsf i}\,\sigma_i^*\bigl(\ovl\Sigma_{\rm L}\wedge({\boldsymbol{1}}\otimes\sigma_1)\,\Sigma_{\rm L}\bigr)\,,\qquad i\in I
\end{eqnarray}
of a globally smooth super-2-form $\,\underset{\tx{\ciut{(2)}}}{{\rm B}}\,$ on $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)$.\ Consequently, we may take as the (trivial) surjective submsersion of the super-1-gerbe under reconstruction the supertarget itself,
\begin{eqnarray}\nn
\pi_{{\mathsf Y}{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)}\equiv{\rm id}_{{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)}\ :\ {\mathsf Y}{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5):={\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\longrightarrow{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,,
\end{eqnarray}
with $\,\underset{\tx{\ciut{(2)}}}{{\rm B}}\,$ as the corresponding curving. Over its fibred square
\begin{eqnarray}\nn
{\mathsf Y}^{[2]}{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\equiv{\mathsf Y}{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\x_{{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)}{\mathsf Y}{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5\cong{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,,
\end{eqnarray}
equipped with the canonical projections $\,{\rm pr}_\a,\ \a\in\{1,2\}$,\ identifiable with $\,{\rm id}_{{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)}$,\ we obtain the trivial identity
\begin{eqnarray}\nn
({\rm pr}_2^*-{\rm pr}_1^*)\underset{\tx{\ciut{(2)}}}{{\rm B}}\equiv0
\end{eqnarray}
which enables us to take the trivial principal ${\mathbb{C}}^\x$-bundle
\begin{eqnarray}\nn
\pi_\mathscr{L}\equiv{\rm pr}_1\ :\ \mathscr{L}:={\mathsf Y}^{[2]}{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\x{\mathbb{C}}^\x\longrightarrow{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\ :\ (Z_i,Z_i,z)\longmapsto Z_i
\end{eqnarray}
with the trivial principal connection
\begin{eqnarray}\nn
\nabla_\mathscr{L}:={\mathsf d}\,,
\end{eqnarray}
or, equivalently, with a principal connection super-1-form
\begin{eqnarray}\nn
\cA\bigl(Z_i,Z_i,z\bigr):={\mathsf i}\,\tfrac{{\mathsf d} z}{z}
\end{eqnarray}
as the data of the super-1-gerbe. The latter is manifestly invariant under the component-wise (non-linear) action of the \emph{product} Lie supergroup
\begin{eqnarray}\nn
\widetilde{{\rm SU}(2,2\,\vert\,4)}:={\rm SU}(2,2\,\vert\,4)\x{\mathbb{C}}^\x
\end{eqnarray}
which, in the notation of \Reqref{eq:cosetactptcoord}, takes the form
\begin{eqnarray}\nn
\widetilde{[\lambda]}_\cdot\ :\ \widetilde{{\rm SU}(2,2\,\vert\,4)}\x\mathscr{L}\longrightarrow\mathscr{L}\ :\ \bigl((g,\zeta),(Z_i,Z_i,z)\bigr)\longmapsto\bigl(\widetilde Z_j(Z_i,g),\widetilde Z_j(Z_i,g),\zeta\cdot z\bigr)\,.
\end{eqnarray}
Hence, the triple $\,(\mathscr{L},\pi_\mathscr{L},\cA)\,$ could be called a (trivial) super-0-gerbe in what seems a natural generalisation of Def.\,I.5.4. to the setting of a homogeneous space of a Lie supergroup.
In the last step, we consider the cartesian cube of the above trivial surjective submersion fibred over $\,({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)$,
\begin{eqnarray}\nn
{\mathsf Y}^{[3]}{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\equiv{\mathsf Y}{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\x_{{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)}{\mathsf Y}{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\x_{{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)}{\mathsf Y}{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\cong{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,,
\end{eqnarray}
(with its canonical projections $\,{\rm pr}_{\a,\beta}\equiv({\rm pr}_\a,{\rm pr}_\beta),\ (\a,\beta)\in\{(1,2),(2,3),(1,3)\}\,$ to the previously considered fibred square), and, over it, take the connection-preserving isomorphism
\begin{eqnarray}\nn
\mu_\mathscr{L}\ :\ {\rm pr}_{1,2}^*\mathscr{L}\otimes{\rm pr}_{2,3}^*\mathscr{L}\xrightarrow{\ \cong\ }{\rm pr}_{1,3}^*\mathscr{L}\ :\ \left[\bigl(Z_i,Z_i,z_{1,2}\bigr),\bigl(Z_i,Z_i,z_{2,3}\bigr)\right]\longmapsto\bigl(Z_i,Z_i,z_{1,2}\cdot z_{2,3}\bigr)\,,
\end{eqnarray}
where we have identified
\begin{eqnarray}\nn
\bigl(Z_i,Z_i,z_{\a,\beta}\bigr)\equiv\bigl(Z_i,(Z_i,Z_i,z_{\a,\beta})\bigr)\in{\rm pr}_{\a,\beta}^*\mathscr{L}\,.
\end{eqnarray}
A fibre-bundle map thus defined trivially satisfies the standard groupoid identity over $\,{\mathsf Y}^{[4]}{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\cong{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,$ and conforms structurally with the description of a super-0-gerbe isomorphism given in Def.\,I.5.4.
Our analysis results in
\begin{Def}\label{def:s1gerbeAdS}
The \textbf{trivial Metsaev--Tseytlin super-1-gerbe} over $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,$ of curvature $\,\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}\,$ is the sextuple
\begin{eqnarray}\nn
\mathcal{I}^{\rm MT}_{\underset{\tx{\ciut{(2)}}}{{\rm B}}}:=\bigl({\mathsf Y}{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5),\pi_{{\mathsf Y}{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)},\underset{\tx{\ciut{(2)}}}{{\rm B}},\mathscr{L},\nabla_\mathscr{L},\mu_\mathscr{L}\bigr)
\end{eqnarray}
constructed in the preceding paragraphs.
\begin{flushright}$\diamond$\end{flushright
Clearly, the latter does \emph{not} reproduce the non-supersymmetric curving \eqref{eq:sMincurv} of the Green--Schwarz super-1-gerbe on ${\rm sMink}^{1,9\,\vert\,D_{9,1}}\,$ (not even in restriction to positive-chirality spinors).
\section{The Kamimura--Sakaguchi supercentral charge extensions of $\,\gt{su}(2,2\,\vert\,4)$}\label{sec:KamSakext}
The most natural class of deformations encountered along the second path indicated at the end of Sec.\,\ref{sec:ssextaAdSS}, and well known independently from a variety of field-theoretic contexts, consists of super-central extensions. In order to identify, in this setting, the potential source of a (supersymmetric) correction to the previously considered primitive $\,\underset{\tx{\ciut{(2)}}}{\beta}$,\ we shall consider the most general such extension
\begin{eqnarray}\nn
\widetilde{\gt{su}(2,2\,\vert\,4)}=\gt{su}(2,2\,\vert\,4)\oplus\corr{Z_{\a\a'I\,\beta\b'J},Z_{\widehat a\,\widehat b},Z_{\widehat a\widehat b\,\widehat c\widehat d},Z_{\widehat a\,\widehat b\widehat c},Z_{\a\a'I\,\widehat a},Z_{\a\a'I\,\widehat a\widehat b}}_{\mathbb{C}}
\end{eqnarray}
of $\,\gt{su}(2,2\,\vert\,4)$,\ as in \Reqref{eq:wrapcentrext}, and study the asymptotics of the pullback of the Maurer--Cartan super-1-form from the Lie supergroup $\,\widetilde{{\rm SU}(2,2\,\vert\,4)}\,$ integrating the extension,
\begin{eqnarray}\nn
{\boldsymbol{1}}\longrightarrow{\rm exp}\bigl(\corr{Z_{\a\a'I\,\beta\b'J},Z_{\widehat a\,\widehat b},Z_{\widehat a\widehat b\,\widehat c\widehat d},Z_{\widehat a\,\widehat b\widehat c},Z_{\a\a'I\,\widehat a},Z_{\a\a'I\,\widehat a\widehat b}}_{\mathbb{C}}\bigr)\longrightarrow\widetilde{{\rm SU}(2,2\,\vert\,4)}\xrightarrow{\ \widetilde\pi\ }{\rm SU}(2,2\,\vert\,4)\longrightarrow{\boldsymbol{1}}\,,
\end{eqnarray}
along a distinguished (local) section
\begin{eqnarray}
\widetilde\sigma_i\ &:&\ \widetilde\mathcal{O}_i\longrightarrow\widetilde{{\rm SU}(2,2\,\vert\,4)}\cr\cr
&:&\ (Z_i,\zeta_i)\longmapsto{\rm e}^{X_i^{\widehat a}\,P_{\widehat a}}\cdot{\rm e}^{\theta_i^{\a\a' I}\,Q_{\a\a' I}}\cdot{\rm e}^{\zeta_i^{\a\a'I\,\beta\b'J}\,Z_{\a\a'I\,\beta\b'J}+\zeta_i^{\widehat a\,\widehat b}\,Z_{\widehat a\,\widehat b}+\zeta_i^{\widehat a\widehat b\,\widehat c\widehat d}\,Z_{\widehat a\widehat b\,\widehat c\widehat d}+\zeta_i^{\widehat a\,\widehat b\widehat c}\,Z_{\widehat a\,\widehat b\widehat c}+\zeta_i^{\a\a'I\,\widehat a}\,Z_{\a\a'I\,\widehat a}+\zeta_i^{\a\a'I\,\widehat a\widehat b}\,Z_{\a\a'I\,\widehat a\widehat b}}\,.\cr \label{eq:distlocsecext}
\end{eqnarray}
In so doing, we shall not be concerned with the potential linear dependence of the various super-central charges defining the extension, a feature of no bearing on our conclusions at our level of generality.
Calculating the leading asymptotics of expression \eqref{eq:MCs1f} once again, but this time for the super-centrally extended Lie superalgebra, we eventually arrive at
\begin{eqnarray}\nn
\widetilde\sigma_i^*\theta_{\rm L}(Z_i,\zeta_i)={\mathsf d}\zeta_i^{\a\a'I\,\beta\b'J}\otimes Z_{\a\a'I\,\beta\b'J}+{\mathsf d}\zeta_i^{\widehat a\,\widehat b}\otimes Z_{\widehat a\,\widehat b}+{\mathsf d}\zeta_i^{\widehat a\widehat b\,\widehat c\widehat d}\otimes Z_{\widehat a\widehat b\,\widehat c\widehat d}+{\mathsf d}\zeta_i^{\widehat a\,\widehat b\widehat c}\otimes Z_{\widehat a\,\widehat b\widehat c}+{\mathsf d}\zeta_i^{\a\a'I\,\widehat a}\otimes Z_{\a\a'I\,\widehat a}\cr\cr
+{\mathsf d}\zeta_i^{\a\a'I\,\widehat a\widehat b}\otimes Z_{\a\a'I\,\widehat a\widehat b}+\tfrac{1}{R^{\frac{1}{2}}}\,{\mathsf d}\theta_i^{\a\a'I}\otimes Q_{\a\a'I}\cr\cr
+\tfrac{1}{R}\,\bigl[\bigl({\mathsf d} X_i^{\widehat a}-{\mathsf i}\,\ovl\theta_i\,(\widehat\Gamma^{\widehat a}\otimes{\boldsymbol{1}})\,{\mathsf d}\theta_i\bigr)\otimes P_{\widehat a}+\tfrac{{\mathsf i}}{2}\,\ovl\theta_i\,(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2)\,{\mathsf d}\theta_i\otimes J_{\widehat a\widehat b}+\tfrac{1}{2}\,\theta_i^{\a\a'I}\,{\mathsf d}\theta_i^{\beta\b'J}\otimes Z_{\a\a'I\,\beta\b'J}\bigr]\cr\cr
+\tfrac{1}{R^{\frac{3}{2}}}\,\bigl\{\bigl[\tfrac{1}{2}\,\bigl({\mathsf d} X_i^{\widehat a}-\tfrac{{\mathsf i}}{3}\,\ovl\theta_i\,(\widehat\Gamma^a\otimes{\boldsymbol{1}})\,{\mathsf d}\theta_i\bigr)\,(\widehat\Gamma_{\widehat a}\otimes\sigma_2)^{\a\a'I}_{\ \beta\b'J}+\tfrac{{\mathsf i}\,\varepsilon_{\widehat a\widehat b}}{12}\,\bigl(\ovl\theta_i\,(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2)\,{\mathsf d}\theta_i\,(\widehat\Gamma_{\widehat a\widehat b}\otimes{\boldsymbol{1}})^{\a\a'I}_{\ \beta\b'J}\bigr]\,\theta_i^{\beta\b'J}\otimes Q_{\a\a' I}\cr\cr
-\theta_i^{\a\a'I}\,\bigl[\bigl({\mathsf d} X_i^{\widehat a}-\tfrac{{\mathsf i}}{3}\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a}\otimes{\boldsymbol{1}}\bigr)\,{\mathsf d}\theta_i\bigr)\otimes Z_{\a\a'I\,\widehat a}+\tfrac{{\mathsf i}}{3!}\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2\bigr)\,{\mathsf d}\theta_i\otimes Z_{\a\a'I\,\widehat a\widehat b}\bigr]\bigr\}\cr\cr
-\tfrac{1}{2R^2}\,\bigl\{\bigl({\mathsf i}\,\bigl({\mathsf d} X_i^{\widehat b}-\tfrac{{\mathsf i}}{6}\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat b}\otimes{\boldsymbol{1}}\bigr)\,{\mathsf d}\theta_i\bigr)\,\ovl\theta_i\,\bigl(\{\widehat\Gamma^{\widehat a},\widehat\Gamma_{\widehat b}\}\otimes\sigma_2\bigr)\,\theta_i-\tfrac{\varepsilon_{\widehat b\widehat c}}{12}\,\ovl\theta_i\,\bigl(\widehat\Gamma_{\widehat b\widehat c}\otimes\sigma_2\bigr)\,{\mathsf d}\theta_i\cdot\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a}\,\widehat\Gamma^{\widehat b\widehat c}\otimes{\boldsymbol{1}}\bigr)\,\theta_i\bigr)\otimes P_{\widehat a}\cr\cr
-\bigl(\tfrac{{\mathsf i}}{2}\,\bigl({\mathsf d} X_i^{\widehat c}-\tfrac{{\mathsf i}}{3!}\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat c}\otimes{\boldsymbol{1}}\bigr)\,{\mathsf d}\theta_i\bigr)\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a\widehat b}\,\widehat\Gamma_{\widehat c}\otimes{\boldsymbol{1}}\bigr)\,\theta_i-\tfrac{\varepsilon_{\widehat a\widehat b}}{4!}\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat c\widehat d}\otimes\sigma_2\bigr)\,{\mathsf d}\theta_i\cdot\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a\widehat b}\,\widehat\Gamma_{\widehat c\widehat d}\otimes\sigma_2\bigr)\,\theta_i-\varepsilon_{\widehat a\widehat b}\,X_i^{\widehat a}\,{\mathsf d} X_i^{\widehat b}\bigr)\otimes J_{\widehat a\widehat b}\cr\cr
-\tfrac{1}{2}\,\bigl[\bigl({\mathsf d} X_i^{\widehat a}-\tfrac{{\mathsf i}}{6}\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a}\otimes{\boldsymbol{1}}\bigr)\,{\mathsf d}\theta_i\bigr)\,\bigl(\widehat\Gamma_{\widehat a}\otimes\sigma_2\bigr)^{\beta\b'J}_{\ \gamma\g'K}+\tfrac{{\mathsf i}\varepsilon_{\widehat a\widehat b}}{12}\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2\bigr)\,{\mathsf d}\theta_i\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes{\boldsymbol{1}}\bigr)^{\beta\b'J}_{\ \gamma\g'K}\bigr]\,\theta_i^{\a\a'I}\,\theta_i^{\gamma\g'K}\otimes Z_{\a\a'I\,\beta\b'J}\cr\cr
+X^{\widehat a}_i\,{\mathsf d} X^{\widehat b}_i\otimes Z_{\widehat a\widehat b}\bigr\}+O\bigl(R^{-\frac{5}{2}}\bigr)\,.
\end{eqnarray}
Inspection of the above formula reveals that the unique left-invariant super-1-form whose exterior derivative possesses the desired asymptotics and could, consequently, be considered as the candidate for a supersymmertic extension of the corrections $\,\underset{\tx{\ciut{(2)}}}{{\rm D}_i}\,$ on $\,\widetilde{{\rm SU}(2,2\,\vert\,4)}\,$ (and so also on $\,\widetilde{{\rm SU}(2,2\,\vert\,4)}/({\rm SO}(4,1)\x{\rm SO}(5))$) is the linear combination of the components of the Maurer--Cartan super-1-form associated with the wrapping charges $\,Z_{\a\a'I\ \beta\b'J}$,
\begin{eqnarray}\nn
\widetilde\sigma_i^*\theta_{\rm L}^{\a\a'I\ \beta\b'J}(Z_i,\zeta_i)&=&{\mathsf d}\zeta_i^{\a\a'I\ \beta\b'J}+\tfrac{1}{4R}\,\bigl(\theta_i^{\a\a' I}\,{\mathsf d}\theta_i^{\beta\b' J}+\theta_i^{\beta\b' J}\,{\mathsf d}\theta_i^{\a\a' I}\bigr)\cr\cr
&&+\tfrac{1}{4R^2}\,\bigl[\bigl({\mathsf d} X_i^{\widehat a}-\tfrac{{\mathsf i}}{6}\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a}\otimes{\boldsymbol{1}}\bigr)\,{\mathsf d}\theta_i\bigr)\,\bigl(\widehat\Gamma_{\widehat a}\otimes\sigma_2\bigr)^{(\beta\b'J}_{\ \gamma\g'K}\,\theta_i^{\a\a'I)}\cr\cr
&&+\tfrac{{\mathsf i}\varepsilon_{\widehat a\widehat b}}{12}\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2\bigr)\,{\mathsf d}\theta_i\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes{\boldsymbol{1}}\bigr)^{(\beta\b'J}_{\ \gamma\g'K}\,\theta_i^{\a\a'I)}\bigr]\,\theta_i^{\gamma\g'K}+O\bigl(R^{-3}\bigr)\,,
\end{eqnarray}
with coefficients $\,2{\mathsf i}\,(\widehat C\otimes\sigma_1)_{\a\a'I\beta\b'J}$.\ Indeed, upon taking into account the Fierz identity that ensures the vanishing of the super-Jacobiator
\begin{eqnarray}\label{eq:sJacQQQ}
{\rm sJac}(Q_{\a\a'I},Q_{\beta\b'J},Q_{\gamma\g'K})=0
\end{eqnarray}
of $\,\gt{su}(2,2\,\vert\,4)$,\ or -- explicitly --
\begin{eqnarray}\nn
\varepsilon_{\widehat a\widehat b}\,\bigl(\widehat C\,\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2\bigr)_{(\a\a'I\beta\b'J}\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes{\boldsymbol{1}}\bigr)^{\delta\d'L}_{\ \gamma\g'K)}=2\bigl(\widehat C\,\widehat\Gamma^{\widehat a}\otimes{\boldsymbol{1}}\bigr)_{(\a\a'I\beta\b'J}\,\bigl(\widehat\Gamma_{\widehat a}\otimes\sigma_2\bigr)^{\delta\d'L}_{\ \gamma\g'K)}\,,
\end{eqnarray}
and the resultant identity (obtained through contraction with $\,(\widehat C\otimes\sigma_1)_{\epsilon\ep'M\delta\d'L}\,{\mathsf d}\theta_i^{\a\a'I}\wedge{\mathsf d}\theta_i^{\gamma\g'K}\,\theta_i^{\beta\b'J}\,\theta_i^{\epsilon\ep'M}$)
\begin{eqnarray}\nn
&&\varepsilon_{\widehat a\widehat b}\,{\mathsf d}\ovl\theta_i\wedge\bigl(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2\bigr)\,{\mathsf d}\theta_i\cdot\ovl\theta_i\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes\sigma_1\bigr)\,\theta_i\cr\cr
&=&-2\varepsilon_{\widehat a\widehat b}\,\ovl\theta_i\bigl(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2\bigr)\,{\mathsf d}\theta_i\wedge\,\ovl\theta_i\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes\sigma_1\bigr)\,{\mathsf d}\theta_i+4i\ovl\theta_i\bigl(\widehat\Gamma^{\widehat a}\otimes{\boldsymbol{1}}\bigr)\,{\mathsf d}\theta_i\wedge\,\ovl\theta_i\,\bigl(\widehat\Gamma_{\widehat a}\otimes\sigma_3\bigr)\,{\mathsf d}\theta_i\cr\cr
&&+2i{\mathsf d}\ovl\theta_i\wedge\bigl(\widehat\Gamma^{\widehat a}\otimes{\boldsymbol{1}}\bigr)\,{\mathsf d}\theta_i\cdot\ovl\theta_i\,\bigl(\widehat\Gamma_{\widehat a}\otimes\sigma_3\bigr)\,\theta_i\cr\cr
&=&-2\varepsilon_{\widehat a\widehat b}\,\ovl\theta_i\bigl(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2\bigr)\,{\mathsf d}\theta_i\wedge\,\ovl\theta_i\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes\sigma_1\bigr)\,{\mathsf d}\theta_i+4i\ovl\theta_i\bigl(\widehat\Gamma^{\widehat a}\otimes{\boldsymbol{1}}\bigr)\,{\mathsf d}\theta_i\wedge\,\ovl\theta_i\,\bigl(\widehat\Gamma_{\widehat a}\otimes\sigma_3\bigr)\,{\mathsf d}\theta_i\,,
\end{eqnarray}
we readily establish the equality
\begin{eqnarray}\nn
\widetilde\sigma_i^*\mathcal{E}(Z_i,\zeta_i)&:=&2{\mathsf i}\,\bigl(\widehat C\otimes\sigma_1)_{\a\a'I\beta\b'J}\,\widetilde\sigma_i^*\theta_{\rm L}^{\a\a'I\ \beta\b'J}(Z_i,\zeta_i)\cr\cr
&=&2{\mathsf i}\,\bigl(\widehat C\otimes\sigma_1)_{\a\a'I\beta\b'J}\,{\mathsf d}\zeta_i^{\a\a'I\beta\b'J}+\tfrac{{\mathsf i}}{R}\,\ovl\theta_i\,({\boldsymbol{1}}\otimes\sigma_1)\,{\mathsf d}\theta_i\cr\cr
&&-\tfrac{1}{2R^2}\,\bigl[\bigl({\mathsf d} X_i^{\widehat a}-\tfrac{{\mathsf i}}{6}\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a}\otimes{\boldsymbol{1}}\bigr)\,{\mathsf d}\theta_i\bigr)\,\ovl\theta_i\,\bigl(\widehat\Gamma_{\widehat a}\otimes\sigma_3\bigr)\,\theta+\tfrac{\varepsilon_{\widehat a\widehat b}}{12}\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2\bigr)\,{\mathsf d}\theta_i\cdot\ovl\theta_i\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes\sigma_1\bigr)\,\theta_i\bigr]\cr\cr
&&+O\bigl(R^{-3}\bigr)\cr\cr
&=&2{\mathsf i}\,\bigl(\widehat C\otimes\sigma_1)_{\a\a'I\beta\b'J}\,{\mathsf d}\zeta_i^{\a\a'I\beta\b'J}+\tfrac{{\mathsf i}}{R}\,\ovl\theta_i\,({\boldsymbol{1}}\otimes\sigma_1)\,{\mathsf d}\theta_i-\tfrac{\varepsilon_{\widehat a\widehat b}}{24R^2}\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2\bigr)\,{\mathsf d}\theta_i\cdot\ovl\theta_i\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes\sigma_1\bigr)\,\theta_i\cr\cr
&&+O\bigl(R^{-3}\bigr)
\end{eqnarray}
whence also
\begin{eqnarray}\nn
\underset{\tx{\ciut{(2)}}}{\widetilde{\mathsf D}_i}(Z_i,\zeta_i)&:=&{\mathsf d}\widetilde\sigma_i^*\mathcal{E}(Z_i,\zeta_i)=\tfrac{{\mathsf i}}{R}\,{\mathsf d}\ovl\theta_i\wedge({\boldsymbol{1}}\otimes\sigma_1)\,{\mathsf d}\theta_i-\tfrac{\varepsilon_{\widehat a\widehat b}}{24R^2}\,{\mathsf d}\ovl\theta_i\wedge\bigl(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2\bigr)\,{\mathsf d}\theta_i\cdot\ovl\theta_i\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes\sigma_1\bigr)\,\theta_i\cr\cr
&&+\tfrac{\varepsilon_{\widehat a\widehat b}}{12R^2}\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2\bigr)\,{\mathsf d}\theta_i\wedge\ovl\theta_i\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes\sigma_1\bigr)\,{\mathsf d}\theta_i+O\bigl(R^{-3}\bigr)\cr\cr
&=&\underset{\tx{\ciut{(2)}}}{{\mathsf D}}(Z_i)+\tfrac{{\mathsf i}}{6R^2}\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a}\otimes\sigma_3\bigr)\,{\mathsf d}\theta_i\wedge\ovl\theta_i\,\bigl(\widehat\Gamma_{\widehat a}\otimes{\boldsymbol{1}}\bigr)\,{\mathsf d}\theta_i\cr\cr
&&-\tfrac{\varepsilon_{\widehat a\widehat b}}{6R^2}\,\ovl\theta_i\,\bigl(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_1\bigr)\,{\mathsf d}\theta_i\wedge\ovl\theta_i\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes\sigma_2\bigr)\,{\mathsf d}\theta_i+O\bigl(R^{-3}\bigr)\,.
\end{eqnarray}
This yields the desired asymptotics
\begin{eqnarray}\nn
\widetilde\sigma_i^*\widetilde\pi^*\underset{\tx{\ciut{(2)}}}{\beta}(Z_i,\zeta_i)-\underset{\tx{\ciut{(2)}}}{\widetilde{\mathsf D}_i}(Z_i,\zeta_i)&=&\tfrac{1}{R^2}\,\ovl\theta_i\,\bigl(\widehat\Gamma_{\widehat a}\otimes\sigma_3\bigr)\,{\mathsf d}\theta_i\wedge\bigl({\mathsf d} X_i^{\widehat a}-\tfrac{{\mathsf i}}{2}\,\ovl\theta_i\,(\widehat\Gamma^a\otimes{\boldsymbol{1}})\,{\mathsf d}\theta_i\bigr)+O\bigl(R^{-3}\bigr)\,.
\end{eqnarray}
The terms of order $\,R^{-2}\,$ independent of $\,X^{\widehat a}\,$ sum up to zero for the asymptotic ten-dimensional spinors of positive chirality -- this is none other than the super-Minkowskian Fierz identity ensuring the vanishing of the super-Jacobiator \eqref{eq:sJacQQQ} for the super-Poincar\'e algebra. Thus, we have reproduced the (manifestly non-supersymmetric) primitive of the super-Minkowskian Green--Schwarz super-3-cocycle (and, in so doing, the result of \Rcite{Hatsuda:2002hz}) through our $\widehat{{\rm SU}(2,2\,\vert\,4)}$-invariant analysis. It remains to be checked whether such a correction can be obtained from a super-central extension of the original Lie superalgebra $\,\gt{su}(2,2\,\vert\,4)$.\ One may also rephrase this question of internal consistency so as to readily answer it in the negative for a large class of extensions. We begin by noting that in the above reasoning, we assumed that \emph{all} the $\,Z_{\a\a'I\,\beta\b'J}\,$ are allowed as Gra\ss mann-even charges extending the original Lie superalgebra $\,\gt{su}(2,2\,\vert\,4)\,$ in a consistent manner -- this is the assumption behind the left-invariance of the component super-1-forms which allows us to take the desired linear combination of the super-1-forms (with coefficients given by ${\rm SO}(4,1)\x{\rm SO}(5)$-invariant tensors) and obtain -- upon differentiation -- the sought-after left-invariant super-2-form. In fact, we need to assume less, to wit, that the specific linear combination $\,2{\mathsf i}\,(\widehat C\otimes\sigma_1)_{\a\a'I\beta\b'J}\,\theta_{\rm L}^{\a\a'I\beta\b'J}\,$ be left-invariant, or -- equivalently -- that a consistent Lie-superalgebraic deformation of the form
\begin{eqnarray}\nn
\{Q_{\a\a' I},Q_{\beta\b' J}\}^\sim={\mathsf i}\,\bigl(-2(\widehat C\,\widehat\Gamma^{\widehat a}\otimes{\boldsymbol{1}})_{\a\a'I\beta\b'J}\,P_{\widehat a}+(\widehat C\,\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2)_{\a\a'I\beta\b'J}\,J_{\widehat a\widehat b}\bigr)+2{\mathsf i}\,\bigl(\widehat C\otimes\sigma_1\bigr)_{\a\a'I\,\beta\b'J}\,Z
\end{eqnarray}
be allowed as the left-invariant super-1-form associated with charge $\,Z\,$ is $\,\mathcal{E}$.\ It is easy to see that such a supposition is untenable as it leads to a contradiction, at least as long as the anticommutator of the supercharges is the only place where $\,Z\,$ appears. Indeed, whenever this is the case, we find, in virtue of the Maurer--Cartan equation,
\begin{eqnarray}\nn
{\mathsf d}\mathcal{E}=\tfrac{1}{2}\,2{\mathsf i}\,\bigl(\widehat C\otimes\sigma_1\bigr)_{\a\a'I\,\beta\b'J}\,\theta^{\a\a'I}_{\rm L}\wedge\theta^{\beta\b'J}_{\rm L}\equiv\widetilde\pi^*\underset{\tx{\ciut{(2)}}}{\beta}\,,
\end{eqnarray}
which is manifestly at variance with the result derived previously (and ruins our plan of correcting the asymptotics of $\,\underset{\tx{\ciut{(2)}}}{\beta}\,$ through the substraction of $\,{\mathsf d}\mathcal{E}$). Below, we methodically check that (and see why) the desired deformation of $\,\gt{su}(2,2\,\vert\,4)\,$ is inconsistent with the assumption of associativity in a large class of geometrically motivated central-extensions.
A full-blown cartography of the \emph{entire} space of associative deformations of $\,\gt{su}(2,2\,\vert\,4)$,\ and even of the subspace of \emph{all} super-central extensions (without any further constraints) goes beyond the scope of the current report. We shall, instead, explore various corners of those spaces, guided by the physical intuition and algebraic hints from the analysis of the asymptotics of the Metsaev--Tseytlin super-3-cocycle and previous studies of the super-Minkowskian setting. We begin by considering an \emph{arbitrary} deformation of the anti-commutator of super-charges, keeping in mind its irremovable geometric germ expressible as the monodromy of the coordinate function $\,X^a\,$ in the vicinity of the unital coset $\,{\rm SO}(4,1)\x{\rm SO}(5)$.\ Thus, we write
\begin{eqnarray}\label{eq:KSextend}\qquad\qquad
\{Q_{\a\a' I},Q_{\beta\b' J}\}^\sim={\mathsf i}\,\bigl(-2(\widehat C\,\widehat\Gamma^{\widehat a}\otimes{\boldsymbol{1}})_{\a\a'I\beta\b'J}\,P_{\widehat a}+(\widehat C\,\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2)_{\a\a'I\beta\b'J}\,J_{\widehat a\widehat b}\bigr)+Z_{\a\a'I\,\beta\b'J}\,,
\end{eqnarray}
with the $\,Z_{\a\a'I\,\beta\b'J}=Z_{(\a\a'I\,\beta\b'J)}\,$ central, and keep all other super-commutators unchanged. A moment's thought convinces us that the only super-Jacobi identities to be imposed are the following ones:
\begin{eqnarray}\nn
{\rm sJac}(Q_{\a\a'I},Q_{\beta\b'J},P_{\widehat a})=0={\rm sJac}(Q_{\a\a'I},Q_{\beta\b'J},J_{\widehat a\widehat b})\,.
\end{eqnarray}
Upon invoking their undeformed counterparts, the former ones give us the following constraints:
\begin{eqnarray}\nn
Z_{\a\a'I\,\gamma\g'K}\,\bigl(\widehat\Gamma_{\widehat a}\otimes\sigma_2\bigr)^{\gamma\g'K}_{\ \beta\b'J}+Z_{\beta\b'J\,\gamma\g'K}\,\bigl(\widehat\Gamma_{\widehat a}\otimes\sigma_2\bigr)^{\gamma\g'K}_{\ \a\a'I}=0\,,\quad\widehat a\in\ovl{0,9}\,,
\end{eqnarray}
or, equivalently,
\begin{eqnarray}\nn
0&=&Z_{\a\a'I\,\beta\b'J}\,\bigl(\delta^{\a\a'I}_{\ \gamma\g'K}\,\delta^{\beta\b'J}_{\delta\d'L}+\bigl(-\widehat\Gamma_{\widehat a}\otimes{\mathsf i}\,\sigma_2\bigr)^{\a\a'I}_{\ \gamma\g'K}\,\bigl(\varepsilon_{\widehat a}\,\widehat\Gamma^{\widehat a}\otimes{\mathsf i}\,\sigma_2\bigr)^{\beta\b'J}_{\ \delta\d'L}\bigr)\cr\cr
&\equiv&Z_{\a\a'I\,\beta\b'J}\,\bigl({\boldsymbol{1}}\otimes{\boldsymbol{1}}+\bigl(\varepsilon_{\widehat a}\,\widehat\Gamma^{\widehat a}\otimes{\mathsf i}\,\sigma_2\bigr)^{-1}\otimes\bigl(\varepsilon_{\widehat a}\,\widehat\Gamma^{\widehat a}\otimes{\mathsf i}\,\sigma_2\bigr)\bigr)^{\a\a'I\beta\b'J}_{\ \gamma\g'K\delta\d'L}\,,
\end{eqnarray}
written for
\begin{eqnarray}\nn
\varepsilon_{\widehat a}=\left\{ \begin{array}{cl} +1 & \tx{if}\ \widehat a\in\ovl{0,4} \\
-1 & \tx{if}\ \widehat a\in\ovl{5,9}\end{array}\right.
\end{eqnarray}
(\emph{no} summation over the range of the spacetime index $\,\widehat a$!). The linear operators annihilating the central charges,
\begin{eqnarray}\nn
{\mathsf P}_{\widehat a}:=\tfrac{1}{2}\,\bigl({\boldsymbol{1}}\otimes{\boldsymbol{1}}+\pi_{\widehat a}^{-1}\otimes\pi_{\widehat a}\bigr)\,,\qquad\pi_{\widehat a}=\varepsilon_{\widehat a}\,\widehat\Gamma^{\widehat a}\otimes{\mathsf i}\,\sigma_2\,,\qquad\qquad\widehat a\in\ovl{0,9}\,,
\end{eqnarray}
are in fact projectors,
\begin{eqnarray}\nn
{\mathsf P}_{\widehat a}\cdot{\mathsf P}_{\widehat a}={\mathsf P}_{\widehat a}\,,
\end{eqnarray}
and we may summarise our first result in the concise form
\begin{eqnarray}\label{eq:ZinKerPs}
(Z_{\a\a'I\,\beta\b'J})\in\bigcap_{\widehat a=0}^9\,{\rm ker}\,{\mathsf P}_{\widehat a}\,.
\end{eqnarray}
Here, the kernels are to be understood as subspaces within the 528-dimensional space $\,{\mathbb{C}}(32)^{\rm sym}\,$ of (complex) symmetric matrices of size 32, a subspace in the 1024-dimensional ${\mathbb{C}}$-linear space
\begin{eqnarray}\nn
{\mathbb{C}}(32)\cong{\rm Cliff}\bigl({\mathbb{R}}^{9,1}\bigr)^{\mathbb{C}}\,.
\end{eqnarray}
The last isomorphism is not invoked accidetally -- indeed, it can actually be employed in a systematic anlysis of the problem in hand. To this end, we decompose the central charge $\,Z_{\a\a'I\,\beta\b'J}\,$ in the Clifford basis \eqref{eq:C32symbas} of $\,{\mathbb{C}}(32)^{\rm sym}\,$ as
\begin{eqnarray}\nn
Z_{\a\a'I\,\beta\b'J}=\mathcal{Z}_\lambda\,(\mathcal{C}\,\unl\Gamma^\lambda)_{\a\a'I\,\beta\b'J}\,,
\end{eqnarray}
and subsequently use the various properties of the Clifford algebras involved, {\it cp} App.\,\ref{app:CliffAdSS}, to identify $\,\bigcap_{\widehat a=0}^9\,{\rm ker}\,{\mathsf P}_{\widehat a}$.\ Prior to that, however, we take a closer look at the remaining set of constraints, following from the nullification of the other set of super-Jacobiators. We obtain
\begin{eqnarray}\nn
Z_{\a\a'I\,\gamma\g'K}\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes{\boldsymbol{1}}\bigr)^{\gamma\g'K}_{\ \beta\b'J}+Z_{\beta\b'J\,\gamma\g'K}\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes{\boldsymbol{1}}\bigr)^{\gamma\g'K}_{\ \a\a'I}=0\,,\qquad\widehat a\in\ovl{0,9}\,,
\end{eqnarray}
or, equivalently,
\begin{eqnarray}\nn
(Z_{\a\a'I\,\beta\b'J})\in\bigcap_{\widehat a,\widehat b=0}^9\,{\rm ker}\,{\mathsf P}_{\widehat a\widehat b}\,,
\end{eqnarray}
where
\begin{eqnarray}\nn
{\mathsf P}_{\widehat a\widehat b}:=\tfrac{1}{2}\,\bigl({\boldsymbol{1}}\otimes{\boldsymbol{1}}+\bigl(\varepsilon_{\widehat a\widehat b}\,\widehat\Gamma^{\widehat a\widehat b}\otimes{\boldsymbol{1}}\bigr)^{-1}\otimes\bigl(\varepsilon_{\widehat a\widehat b}\,\widehat\Gamma^{\widehat a\widehat b}\otimes{\boldsymbol{1}}\bigr)\bigr)\,,\qquad(\widehat a,\widehat b)\in\ovl{0,4}^{\x 2}\cup\ovl{5,9}^{\x 2}\,,\ \widehat a\neq\widehat b
\end{eqnarray}
form another set of projectors, written in terms of the inverses (as previously, no summation over repeated indices)
\begin{eqnarray}\nn
\bigl(\varepsilon_{\widehat a\widehat b}\,\widehat\Gamma^{\widehat a\widehat b}\otimes{\boldsymbol{1}}\bigr)^{-1}=-\eta_{\widehat a\widehat a}\,\eta_{\widehat b\widehat b}\,\varepsilon_{\widehat a\widehat b}\,\widehat\Gamma^{\widehat a\widehat b}\otimes{\boldsymbol{1}}\equiv-\varepsilon_{\widehat a\widehat b}\,\widehat\Gamma_{\widehat a\widehat b}\otimes{\boldsymbol{1}}\,.
\end{eqnarray}
We have, for $\,(\widehat a,\widehat b)\in\ovl{0,4}^{\x 2}\cup\ovl{5,9}^{\x 2}\,$ and $\,\widehat a\neq\widehat b$,\ the identities
\begin{eqnarray}\nn
{\mathsf P}_{\widehat a}\cdot{\mathsf P}_{\widehat b}&=&\tfrac{1}{4}\,\bigl({\boldsymbol{1}}\otimes{\boldsymbol{1}}-\bigl(\widehat\Gamma_{\widehat a}\otimes{\mathsf i}\,\sigma_2\bigr)\otimes\bigl(\varepsilon_{\widehat a}\,\widehat\Gamma^{\widehat a}\otimes{\mathsf i}\,\sigma_2\bigr)-\bigl(\widehat\Gamma_{\widehat b}\otimes{\mathsf i}\,\sigma_2\bigr)\otimes\bigl(\varepsilon_{\widehat b}\,\widehat\Gamma^{\widehat b}\otimes{\mathsf i}\,\sigma_2\bigr)+\bigl(\widehat\Gamma_{\widehat a}\,\widehat\Gamma_{\widehat b}\otimes{\boldsymbol{1}}\bigr)\otimes\bigl(\varepsilon_{\widehat a}\,\varepsilon_{\widehat b}\,\widehat\Gamma^{\widehat a}\,\widehat\Gamma^{\widehat b}\otimes{\boldsymbol{1}}\bigr)\bigr)\cr\cr
&=&\tfrac{1}{4}\,\bigl({\boldsymbol{1}}\otimes{\boldsymbol{1}}-\bigl(\widehat\Gamma_{\widehat a}\otimes{\mathsf i}\,\sigma_2\bigr)\otimes\bigl(\varepsilon_{\widehat a}\,\widehat\Gamma^{\widehat a}\otimes{\mathsf i}\,\sigma_2\bigr)-\bigl(\widehat\Gamma_{\widehat b}\otimes{\mathsf i}\,\sigma_2\bigr)\otimes\bigl(\varepsilon_{\widehat b}\,\widehat\Gamma^{\widehat b}\otimes{\mathsf i}\,\sigma_2\bigr)+\bigl(\widehat\Gamma_{\widehat a}\,\widehat\Gamma_{\widehat b}\otimes{\boldsymbol{1}}\bigr)\otimes\bigl(\widehat\Gamma^{\widehat a}\,\widehat\Gamma^{\widehat b}\otimes{\boldsymbol{1}}\bigr)\bigr)\cr\cr
&=&\tfrac{1}{4}\,\bigl({\boldsymbol{1}}\otimes{\boldsymbol{1}}-\bigl(\widehat\Gamma_{\widehat a}\otimes{\mathsf i}\,\sigma_2\bigr)\otimes\bigl(\varepsilon_{\widehat a}\,\widehat\Gamma^{\widehat a}\otimes{\mathsf i}\,\sigma_2\bigr)-\bigl(\widehat\Gamma_{\widehat b}\otimes{\mathsf i}\,\sigma_2\bigr)\otimes\bigl(\varepsilon_{\widehat b}\,\widehat\Gamma^{\widehat b}\otimes{\mathsf i}\,\sigma_2\bigr)+\bigl(\varepsilon_{\widehat a\widehat b}\,\widehat\Gamma_{\widehat a\widehat b}\otimes{\boldsymbol{1}}\bigr)\otimes\bigl(\varepsilon_{\widehat a\widehat b}\,\widehat\Gamma^{\widehat a\widehat b}\otimes{\boldsymbol{1}}\bigr)\bigr)\cr\cr
&=&\tfrac{1}{4}\,\bigl({\boldsymbol{1}}\otimes{\boldsymbol{1}}-\bigl(\widehat\Gamma_{\widehat a}\otimes{\mathsf i}\,\sigma_2\bigr)\otimes\bigl(\varepsilon_{\widehat a}\,\widehat\Gamma^{\widehat a}\otimes{\mathsf i}\,\sigma_2\bigr)\bigr)+\tfrac{1}{4}\,\bigl({\boldsymbol{1}}\otimes{\boldsymbol{1}}-\bigl(\widehat\Gamma_{\widehat b}\otimes{\mathsf i}\,\sigma_2\bigr)\otimes\bigl(\varepsilon_{\widehat b}\,\widehat\Gamma^{\widehat b}\otimes{\mathsf i}\,\sigma_2\bigr)\bigr)\cr\cr
&&-\tfrac{1}{4}\,\bigl({\boldsymbol{1}}\otimes{\boldsymbol{1}}-\bigl(\varepsilon_{\widehat a\widehat b}\,\widehat\Gamma_{\widehat a\widehat b}\otimes{\boldsymbol{1}}\bigr)\otimes\bigl(\varepsilon_{\widehat a\widehat b}\,\widehat\Gamma^{\widehat a\widehat b}\otimes{\boldsymbol{1}}\bigr)\bigr)\equiv\tfrac{1}{2}\,\bigl({\mathsf P}_{\widehat a}+{\mathsf P}_{\widehat b}-{\mathsf P}_{\widehat a\widehat b}\bigr)\,,
\end{eqnarray}
and so
\begin{eqnarray}\nn
2\bigl({\mathsf P}_{\widehat a}+{\mathsf P}_{\widehat b}\bigr)-\bigl({\mathsf P}_{\widehat a}+{\mathsf P}_{\widehat b}\bigr)^2={\mathsf P}_{\widehat a}+{\mathsf P}_{\widehat b}-{\mathsf P}_{\widehat a}\cdot{\mathsf P}_{\widehat b}-{\mathsf P}_{\widehat b}\cdot{\mathsf P}_{\widehat a}={\mathsf P}_{\widehat a\widehat b}\,.
\end{eqnarray}
From the above, we infer
\begin{eqnarray}\nn
\bigcap_{\widehat a=0}^9\,{\rm ker}\,{\mathsf P}_{\widehat a}\subset\bigcap_{\widehat a,\widehat b=0}^9\,{\rm ker}\,{\mathsf P}_{\widehat a\widehat b}\,,
\end{eqnarray}
which leaves us with the original constraint \eqref{eq:ZinKerPs} as the only one to be imposed. This we do in the aforementioned Clifford basis.
We begin by noting the symmetricity of
\begin{eqnarray}\nn
\mathcal{C}\,\pi_{\widehat a}\equiv-{\mathsf i}\,\widetilde\Gamma^{0\widehat a}\,,
\end{eqnarray}
and from that we derive the identity
\begin{eqnarray}\nn
\bigl(\mathcal{C}\,\unl\Gamma^\lambda\bigr)_{\a\a'I\beta\b'J}\,\bigl(\pi_{\widehat a}^{-1}\bigr)^{\a\a'I}_{\ \gamma\g'K}\,\bigl(\pi_{\widehat a}\bigr)^{\beta\b'J}_{\ \delta\d'L}=-\bigl(\mathcal{C}\,\pi_{\widehat a}\bigr)_{\epsilon\ep'M\delta\d'L}\,\bigl(\unl\Gamma^\lambda\,\pi_{\widehat a}^{-1}\bigr)^{\epsilon\ep'M}_{\ \gamma\g'K}=-\bigl(\mathcal{C}\,\pi_{\widehat a}\,\unl\Gamma^\lambda\,\pi_{\widehat a}^{-1}\bigr)_{\delta\d'L\gamma\g'K}\,,
\end{eqnarray}
which further yields
\begin{eqnarray}\nn
2\bigl(\mathcal{C}\,\unl\Gamma^\lambda\bigr)_{\a\a'I\beta\b'J}\,\bigl({\mathsf P}_{\widehat a}\bigr)^{\a\a'I\beta\b'J}_{\ \gamma\g'K\delta\d'L}=\mathcal{Z}_\lambda\,\bigl(\mathcal{C}\,[\unl\Gamma^\lambda,\pi_{\widehat a}]\,\pi_{\widehat a}^{-1}\bigr)_{\delta\d'L\gamma\g'K}\,.
\end{eqnarray}
Thus we may rephrase the original condition \eqref{eq:ZinKerPs} in the easily tractable form
\begin{eqnarray}\nn
\mathcal{Z}_\lambda\,[\unl\Gamma^\lambda,\pi_{\widehat a}]=0\,,\qquad\widehat a\in\ovl{0,9}\,.
\end{eqnarray}
As the $\,\pi_{\widehat a}\,$ are (up to a trivial rescaling) among the $\,\unl\Gamma^\mu$,\ to wit,
\begin{eqnarray}\nn
{\mathsf i}\,\pi_{\widehat a}\equiv\unl\Gamma^{0\widehat a}\,,
\end{eqnarray}
and the $\,\unl\Gamma^\lambda\,$ span, as linearly independent generators, a Lie algebra over $\,{\mathbb{C}}\,$ with certain structure constants $\,c^{\mu\nu}_{\ \ \kappa}\in{\mathbb{C}}$,\ {\it i.e.},
\begin{eqnarray}\nn
[\unl\Gamma^\mu,\unl\Gamma^\nu]=c^{\mu\nu}_{\ \ \kappa}\,\unl\Gamma^\kappa\,,
\end{eqnarray}
we may further rewrite the above condition as
\begin{eqnarray}\nn
\forall_{(\widehat a,\mu)\in\ovl{0,9}\x\ovl{0,527}}\ :\ \mathcal{Z}_\lambda\,c^{\lambda\,0\widehat a}_{\ \ \ \ \mu}=0\,.
\end{eqnarray}
It now suffices to calculate the relevant structure constants to prove the fundamental
\begin{Prop}\label{prop:KS}
In the notation introduced above,
\begin{eqnarray}\nn
\bigcap_{\widehat a=0}^9\,{\rm ker}\,{\mathsf P}_{\widehat a}=\corr{\widetilde\Gamma^0}_{\mathbb{C}}\,,
\end{eqnarray}
and so the only admissible central extension of the type discussed takes the form
\begin{eqnarray}\nn
\{Q_{\a\a' I},Q_{\beta\b' J}\}^\sim={\mathsf i}\,\bigl(-2(\widehat C\,\widehat\Gamma^{\widehat a}\otimes{\boldsymbol{1}})_{\a\a'I\beta\b'J}\,P_{\widehat a}+(\widehat C\,\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2)_{\a\a'I\beta\b'J}\,J_{\widehat a\widehat b}\bigr)+\bigl(\widehat C\otimes{\boldsymbol{1}}\bigr)_{\a\a'I\,\beta\b'J}\,\mathcal{Z}_0\,.
\end{eqnarray}
\end{Prop}
\noindent\begin{proof}
A proof, based on an explicit computation of all the structure constants $\,c^{\lambda\,0\widehat a}_{\ \ \ \ \mu}$,\ is given in App.\,\ref{app:KSproof}.
\end{proof}
\brem
The central extension derived above is one of the two Gra\ss mann-even deformations of the Lie superalgebra $\,\gt{su}(2,2\,\vert\,4)\,$ containing an undeformed Lie algebra $\,\gt{so}(4,2)\oplus\gt{so}(6)\,$ of isometries of the body $\,{\rm AdS}_5\x{\mathbb{S}}^5\,$ of the supertarget of interest as a subalgebra of its even subalgebra considered by Kamimura and Sakaguchi in \Rcite{Kamimura:2003rx}. Consequently, we propose to and do call it \textbf{the Kamimura--Sakaguchi central extension of} $\,\gt{su}(2,2\,\vert\,4)\,$ in what follows.
\end{Rem}\medskip
We have
\begin{Thm}
The Lie supergroup $\,\widetilde{{\rm SU}(2,2\,\vert\,4)}\,$ integrating the Kamimura--Sakaguchi central extension defined above does not support over the local sections $\,\widetilde\sigma_i\,$ of \Reqref{eq:distlocsecext} a left-invariant super-1-form whose exterior derivative would exhibit the asymptotics of $\,\underset{\tx{\ciut{(2)}}}{{\mathsf D}_i}$.\ Accordingly, there exists no central extension of $\,\gt{su}(2,2\,\vert\,4)\,$ with only the anticommutator of the super-charges deformed as in \Reqref{eq:KSextend} whose integration to a Lie supergroup would trivialise the Metsaev--Tseytlin super-3-cocycle in a manner compatible with both: supersymmetry and the super-Minkowskian asymptotics of Sec.\,\ref{subsect:sstringextsMink}.
\end{Thm}
\noindent\begin{proof}
The question of the existence of the relevant super-1-form may be rephrased as the question of the presence, among the central terms in the extension, now cast in the form
\begin{eqnarray}\nn
\zeta_i^{\a\a'I\beta\b'J}\,\mathcal{Z}_{\a\a'I\,\beta\b'J}\equiv\zeta_i^{\a\a'I\beta\b'J}\,\bigl(\mathcal{C}\,\unl\Gamma^\lambda\bigr)_{\a\a'I\beta\b'J}\,\mathcal{Z}_\lambda=:\zeta_i^\lambda\,\mathcal{Z}_\lambda
\end{eqnarray}
compatible with the choice of the basis in $\,{\mathbb{C}}(32)^{\rm sym}\,$ made in the proof of Prop.\,\ref{prop:KS}, of a non-zero term with index $\,\lambda=2\,$ corresponding to the generator $\,\unl\Gamma^2\equiv{\boldsymbol{1}}\otimes\gamma^2$.\ But the only non-zero term carries index $\,\lambda=0\,$ corresponding to the generator $\,\unl\Gamma^0\equiv{\boldsymbol{1}}\otimes\gamma^0$,\ whence the answer in the negative, and the thesis of the Theorem.
\end{proof}
\void{\section{The Cartan--Eilenberg super-1-gerbe over the super-${\rm AdS}_5\x{\mathbb{S}}^5\,$ space}\label{sec:sAdSSext}
The point of departure of our geometric construction is the manifestly (left-)$\widehat{{\rm SU}(2,2\,\vert\,4)}$-invariant primitive
\begin{eqnarray}\nn
\widetilde{\underset{\tx{\ciut{(2)}}}{{\rm b}}}=\pi^*\bigl(\ovl\Sigma_{\rm L}\wedge({\boldsymbol{1}}\otimes{\boldsymbol{1}}\otimes\sigma_1)\,\Sigma_{\rm L}\bigr)+{\mathsf d}\mathcal{E}\,,\qquad\qquad\mathcal{E}=-2C_{\a\beta}\,C_{\a'\beta'}'\,(\sigma_1)_{IJ}\,\theta_{\rm L}^{\a\a'I\ \beta\b'J}
\end{eqnarray}
of (the pullback of) the Metsaev--Tseytlin super-3-cocycle $\,\pi^*\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}$.\ We take it to be the curving of the Cartan--Eilenberg super-1-gerbe for $\,\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}\,$ upon choosing the restriction \eqref{eq:restrscentrext} of the super-central extension $\,\widehat{{\rm SU}(2,2\,\vert\,4)}\,$ as the surjective submersion of that super-1-gerbe, over the base $\,\sigma({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5))\subset{\rm SU}(2,2\,\vert\,4)$,\
\begin{eqnarray}\nn
{\mathsf Y}\sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr):=\pi^{-1}\bigl(\sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr)\bigr)\,.
\end{eqnarray}
Over the fibred square of this surjective submersion over its base, equipped with the canonical projections $\,{\rm pr}_\a\ :\ {\mathsf Y}^{[2]}\sigma({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5))\longrightarrow{\mathsf Y}\sigma({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,$ to its cartesian factors,
\begin{eqnarray}\nn
\alxydim{@C=-2cm@R=1cm}{& {\mathsf Y}^{[2]}\sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr)\equiv{\mathsf Y}\sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr)\x_{\sigma({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5))}{\mathsf Y}\sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr) \ar[rd]^{{\rm pr}_2} \ar[ld]_{{\rm pr}_1} & \\ {\mathsf Y}\sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr) \ar[rd]_{\pi_{\rm rstr}} & & {\mathsf Y}\sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr) \ar[ld]^{\pi_{\rm rstr}} \\ & \sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr) & }\,,
\end{eqnarray}
we obtain a Cartan--Chevalley super-2-cocycle
\begin{eqnarray}\nn
({\rm pr}_2^*-{\rm pr}_1^*)\widetilde{\underset{\tx{\ciut{(2)}}}{{\rm b}}}={\mathsf d}({\rm pr}_2^*-{\rm pr}_1^*)\mathcal{E}\,,
\end{eqnarray}
with a global \emph{left-invariant} primitive
\begin{eqnarray}\nn
{\rm A}:=({\rm pr}_2^*-{\rm pr}_1^*)\mathcal{E}\,.
\end{eqnarray}
Following the general scheme, we associate with the latter a trivial principal ${\mathbb{C}}^\x$-bundle
\begin{eqnarray}\nn
\pi_\mathscr{L}\equiv{\rm pr}_1\ :\ \mathscr{L}:={\mathsf Y}^{[2]}\sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr)\x{\mathbb{C}}^\x\longrightarrow{\mathsf Y}^{[2]}\sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr)\ :\ (y_1,y_2,z)\longmapsto(y_1,y_2)
\end{eqnarray}
with a principal connection
\begin{eqnarray}\nn
\nabla_\mathscr{L}:={\mathsf p}+\tfrac{1}{{\mathsf i}}\,{\rm A}\,,
\end{eqnarray}
or, equivalently, a principal connection super-1-form
\begin{eqnarray}\nn
\cA(\theta,X,\zeta_1,\zeta_2,z):={\mathsf i}\,\tfrac{{\mathsf d} z}{z}+{\rm A}(\theta,X,\zeta_1,\zeta_2)\,.
\end{eqnarray}
On the total space of the bundle, we have a component-wise (non-linear) action of the \emph{product} Lie supergroup
\begin{eqnarray}\nn
\widehat{\widehat{{\rm SU}(2,2\,\vert\,4)}}:=\widehat{{\rm SU}(2,2\,\vert\,4)}\x{\mathbb{C}}^\x\,,
\end{eqnarray}
induced from the binary group operation of the latter,
\begin{eqnarray}\nn
\widehat{\widehat{m}}\ :\ \widehat{\widehat{{\rm SU}(2,2\,\vert\,4)}}\x\widehat{\widehat{{\rm SU}(2,2\,\vert\,4)}}\longrightarrow\widehat{\widehat{{\rm SU}(2,2\,\vert\,4)}}\ :\ \bigl((g_1,z_1),(g_2,z_2)\bigr)\longmapsto\bigl(\widehat m(g_1,g_2),z_1\cdot z_2\bigr)\,.
\end{eqnarray}
The action reads
\begin{eqnarray}\nn
\widehat{\widehat{[\lambda]}}_\cdot\ :\ \widehat{\widehat{{\rm SU}(2,2\,\vert\,4)}}\x\mathscr{L}\longrightarrow\mathscr{L}\ :\ \bigl((g,\zeta),(y_1,y_2,z)\bigr)\longmapsto\bigl(\widehat{[\lambda]}_g(y_1),\widehat{[\lambda]}_g(y_2),\zeta\cdot z\bigr)\,,
\end{eqnarray}
and as such it preserves the super-1-form $\,\cA$.\ This means that we the triple $\,(\mathscr{L},\pi_\mathscr{L},\cA)\,$ is a super-0-gerbe.
In the last step, we consider the cartesian cube of the surjective submersion $\,{\mathsf Y}\sigma({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5))\,$ fibred over $\,\sigma({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5))$,\ with its canonical projections $\,{\rm pr}_{i,j}\equiv({\rm pr}_i,{\rm pr}_j),\ (i,j)\in\{(1,2),(2,3),(1,3)\}\,$ to $\,{\mathsf Y}^{[2]}\sigma({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5))\,$ that render the diagram
\begin{eqnarray}\nn
\alxydim{@C=0cm@R=1cm}{& & {\mathsf Y}^{[3]}\sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr) \ar[rd]^{{\rm pr}_{1,3}} \ar[ld]_{{\rm pr}_{1,2}} \ar[d]_{{\rm pr}_{2,3}} & & \\ & {\mathsf Y}^{[2]}\sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr) \ar[ld]_{{\rm pr}_1} \ar[d]_{{\rm pr}_2} & {\mathsf Y}^{[2]}\sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr) \ar[ld]_{{\rm pr}_1} \ar[rd]^{{\rm pr}_2} & {\mathsf Y}^{[2]}\sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr) \ar[d]^{{\rm pr}_2} \ar[rd]^{{\rm pr}_1} & \\ {\mathsf Y}\sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr) \ar@/_5.0pc/@{=}[rrrr] \ar[rrd]_{\pi_{\rm rstr}} & {\mathsf Y}\sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr) \ar[rd]^{\pi_{\rm rstr}} & & {\mathsf Y}\sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr) \ar[ld]_{\pi_{\rm rstr}} & {\mathsf Y}\sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr) \ar[lld]^{\pi_{\rm rstr}} \\ & & \sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr) & & }\cr\cr
\end{eqnarray}
commutative, and, over it, look for a connection-preserving isomorphism
\begin{eqnarray}\nn
\mu_\mathscr{L}\ :\ {\rm pr}_{1,2}^*\mathscr{L}\otimes{\rm pr}_{2,3}^*\mathscr{L}\xrightarrow{\ \cong\ }{\rm pr}_{1,3}^*\mathscr{L}\,.
\end{eqnarray}
Comparison of the pullbacks of the (global) connection 1-forms
\begin{eqnarray}\nn
({\rm pr}_{1,2}^*+{\rm pr}_{2,3}^*-{\rm pr}_{1,3}^*){\rm A}=0
\end{eqnarray}
lead us to set
\begin{eqnarray}\nn
\mu_\mathscr{L}\left(\bigl(\theta,X,\xi_1,\xi_2,z_{1,2}\bigr)\otimes\bigl(\theta,X,\xi_2,\xi_3,z_{2,3}\bigr)\right):=\bigl(\theta,X,\xi_1,\xi_3,z_{1,2}\cdot z_{2,3}\bigr)\,,
\end{eqnarray}
where we have identified
\begin{eqnarray}\nn
\bigl(\theta,X,\xi_i,\xi_j,z_{i,j}\bigr)\equiv\bigl((\theta,X,\xi_1,\xi_2,\xi_3),(\theta,X,\xi_i,\xi_j,z_{i,j})\bigr)\in{\rm pr}_{i,j}^*\mathscr{L}\,.
\end{eqnarray}
A fibre-bundle map thus defined trivially satisfies the standard groupoid identity over $\,{\mathsf Y}^{[4]}\sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr)\,$ (and conforms with description of a super-0-gerbe isomorphism given in Def.\,I.5.4).
We conclude our analysis with
\begin{Def}\label{def:s1gerbeAdS}
The \textbf{Metsaev--Tseytlin super-1-gerbe} over $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,$ of curvature $\,\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}\,$ \textbf{with an asymptotically super-Minkowskian curving} is the Cartan--Eilenberg super-1-gerbe, in the sense of Def.\,I.5.11,
\begin{eqnarray}\nn
\mathcal{sG}^{(1)}_{\rm MT}:=\bigl({\mathsf Y}\sigma\bigl({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\bigr),\pi_{\rm rstr},\widetilde{\underset{\tx{\ciut{(2)}}}{{\rm b}}},\mathscr{L},\nabla_\mathscr{L},\mu_\mathscr{L}\bigr)
\end{eqnarray}
constructed in the preceding paragraphs.
\begin{flushright}$\diamond$\end{flushright
\section{Obstruction against an \.In\"on\"u--Wigner contraction on the gerbe}
The most obvious alternative to the path taken in the previous section consists in using the manifestly left-invariant primitive $\,\underset{\tx{\ciut{(2)}}}{{\rm b}}\,$ of the Metsaev--Tseytlin super-3-cocycle as the curving of the super-1-gerbe, defined over the total space of the trivial surjective surjection $\,{\rm id}_{\sigma(\sigma({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)))}\ :\ \sigma({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5))\longrightarrow\sigma({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5))\,$ of the supertarget $\,\sigma({\rm s}({\rm AdS}_5\x{\mathbb{S}}^5))$.\ From the purely geometric point of view, this alternative is perfectly valid and meaningful, but from that of the associated field (or superstring) theory, it is not as we want to be able to reproduce the Green--Schwarz super-$\sigma$-model with the ten-dimensional super-Minkowskian target and the known primitive of the relevant Green--Schwarz super-3-cocycle in the flat limit $\,R\to\infty$.\ Clearly, the super-1-gerbe with the curving $\,\underset{\tx{\ciut{(2)}}}{{\rm b}}\,$ offers us no possibility to attain this goal.
Intuition developed in the super-Minkowskian setting in Part I immediately suggests another alternative to the invariant scenario laid out in Sec.\,\ref{sec:ssextaAdSS}, to wit, a reconstruction of the super-1-gerbe using the counterpart
\begin{eqnarray}\nn
\widetilde{\widetilde{\underset{\tx{\ciut{(2)}}}{{\rm b}}}}:={\mathsf i}\,\pi^*\Sigma_{\rm L}^{\a\a'I}\wedge E_{\a\a'I}
\end{eqnarray}
of the super-Minkowskian invariant
\begin{eqnarray}\nn
\pi_{01}^{(2)\,*}\sigma^\a\wedge e_\a^{(2)}
\end{eqnarray}
from Sec.\,I.5.1.2, built out of the spinorial components of the Maurer--Cartan super-1-form on $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)$:\ the formerly introduced $\,\Sigma_{\rm L}^{\a\a'I}\,$ and the linear combination
\begin{eqnarray}\nn
E_{\a\a'I}:=(\sigma_3)_{IJ}\,\bigl((C\,\Gamma_a)_{\a\beta}\,C_{\a'\beta'}'\,\theta_{\rm L}^{\a\a'I\ a}+C_{\a\beta}\,(C'\,\Gamma_{a'})\,\theta_{\rm L}^{\a\a'I\ a'}\bigr)
\end{eqnarray}
of the spinor-vector components $\,\theta_{\rm L}^{\a\a'I\ a}\,$ and $\,\theta_{\rm L}^{\a\a'I\ a'}\,$ of the same super-1-form on the super-central extension $\,\widehat{{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)}$,\ associated with the super-central charges $\,Z_{\a\a'I\ a}\,$ and $\,Z_{\a\a'I\ a'}$,\ respectively. The study of the asymptotics of the latter super-1-forms should take into account the desired flat-superspace limit \eqref{eq:ssextsMink} of $\,\widehat{\gt{su}(2,2\,\vert\,4)}$.\ Accordingly, we should rescale ($Z_{\rm other}\,$ denotes jointly all super-central charges other than $\,Z_{\a\a'I\ \widehat a}$)
\begin{eqnarray}\nn
(Q_{\a\a'I},P_{\widehat a},J_{\widehat a\widehat b},Z_{\a\a'I\ \widehat a},Z_{\rm other})\mapsto\bigl(R^{\frac{1}{2}}\,Q_{\a\a'I},R\,P_{\widehat a},J_{\widehat a\widehat b},R^{\frac{3}{2}}\,Z_{\a\a'I\ \widehat a},Z_{\rm other}\bigr)\,,
\end{eqnarray}
ensuring that the Kosteleck\'y--Rabin wrapping charge survives in the limit, or, dually,
\begin{eqnarray}\nn
\bigl(\theta^{\a\a' I},X^{\widehat a},\zeta^{\a\a'I\ \widehat a},\zeta^{\rm rest}\bigr)\longmapsto\bigl(R^{-\frac{1}{2}}\,\theta^{\a\a' I},R^{-1}\,X^{\widehat a},R^{-\frac{3}{2}}\,\zeta^{\a\a'I\ \widehat a},\zeta^{\rm other}\bigr)\,.
\end{eqnarray}
Upon such rescaling, we obtain the expansion
\begin{eqnarray}\nn
\theta_{\rm L}^{\a\a'I\ a}(\theta,X,\zeta)=\tfrac{1}{R^{\frac{3}{2}}}\,\bigl[{\mathsf d}\zeta^{\a\a'I\ a}+\theta^{\a\a'I}\,\bigl({\mathsf d} x^a-\tfrac{{\mathsf i}}{3}\,{\mathsf d}\ovl\theta\,(\Gamma^a\otimes{\boldsymbol{1}}\otimes{\boldsymbol{1}})\,\theta\bigr)\bigr]+O\bigl(R^{-\frac{5}{2}}\bigr)\,,
\end{eqnarray}
which implies the sought-after result
\begin{eqnarray}\nn
\widetilde{\widetilde{\underset{\tx{\ciut{(2)}}}{{\rm b}}}}(\theta,X,\zeta)&=&\tfrac{{\mathsf i}}{R^2}\,\bigl({\mathsf d}\theta^{\a\a'I}\wedge{\mathsf d}\xi_{\a\a'I}+{\mathsf d} x^a\wedge\ovl\theta\,(\Gamma_a\otimes{\boldsymbol{1}}\otimes\sigma_3)\,{\mathsf d}\theta+{\mathsf d} x'{}^{a'}\wedge\ovl\theta\,({\boldsymbol{1}}\otimes\Gamma_{a'}\otimes\sigma_3)\,{\mathsf d}\theta\cr\cr
&&+\tfrac{{\mathsf i}}{3}\,\ovl\theta\,(\Gamma_a\otimes{\boldsymbol{1}}\otimes{\boldsymbol{1}})\,{\mathsf d}\theta\wedge\ovl\theta\,(\Gamma^a\otimes{\boldsymbol{1}}\otimes\sigma_3)\,{\mathsf d}\theta+\tfrac{{\mathsf i}}{3}\,\ovl\theta\,({\boldsymbol{1}}\otimes\Gamma_{a'}\otimes{\boldsymbol{1}})\,{\mathsf d}\theta\wedge\ovl\theta\,({\boldsymbol{1}}\otimes\Gamma^{a'}\otimes\sigma_3)\,{\mathsf d}\theta\bigr)\cr\cr
&&+O\bigl(R^{-3}\bigr)\,,
\end{eqnarray}
with the all-important exact correction expressed in terms of the coordinate combinations
\begin{eqnarray}\nn
\xi_{\a\a'I}:=(\sigma_3)_{IJ}\,C_{\a\beta}\,C_{\a'\beta'}'\,\bigl((\Gamma_a)^\beta_{\ \gamma}\,\zeta^{\gamma\beta'J\ a}+(\Gamma_{a'})^{\beta'}_{\ \gamma'}\,\zeta^{\beta\gamma'J\ a'}\bigr)\,.
\end{eqnarray}
Thus, the super-2-form $\,\widetilde{\widetilde{\underset{\tx{\ciut{(2)}}}{{\rm b}}}}\,$ on the extension $\,\widehat{{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)}\,$ goes over to the familiar super-Minkowskian primitive of the relevant Green--Schwarz super-3-form in the flat limit recalled in Sec.\,\ref{subsect:sstringextsMink}.
The above observation seems to suggest that we should replace the previously considered super-2-form $\,\widetilde{\underset{\tx{\ciut{(2)}}}{{\rm b}}}$,\ with the desired \emph{non-invariant} flat limit (that is, the desired flat limit \emph{on the base} $\,{\rm sMink}^{1,9\,\vert\,D_{1,9}}\,$ of the extension), by $\,\widetilde{\widetilde{\underset{\tx{\ciut{(2)}}}{{\rm b}}}}\,$ as the supersymmetry-invariant curving of the Cartan--Eilenberg super-1-gerbe for $\,\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}$.\ The simple failure of this construction is readily revealed by a direct examination of the exterior derivative of the new super-2-form. This yields
\begin{eqnarray}\nn
{\mathsf d}\widetilde{\widetilde{\underset{\tx{\ciut{(2)}}}{{\rm b}}}}=\pi^*\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}+{\mathsf i}\,\pi^*{\mathsf d}\Sigma_{\rm L}^{\a\a'I}\wedge E_{\a\a'I}=\pi^*\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}-\tfrac{{\mathsf i}}{2}\,f_{AB}^{\ \a\a'I}\,\pi^*\theta_{\rm L}^A\wedge\theta_{\rm L}^B\wedge E_{\a\a'I}\,,
\end{eqnarray}
with $\,f_{AB}^{\ \a\a'I}\,$ the structure constants of $\,\gt{su}(2,2\,\vert\,4)$,\ and so we see that the derivative does \emph{not} descend to the base $\,{\rm s}({\rm AdA}_5\x{\mathbb{S}}^5)$.\ This demonstrates the existence, at least in the Cartan-geometric setting considered in this paper, of an obstruction against a geometrisation of the Metsaev--Tseytlin super-3-cocycle $\,\underset{\tx{\ciut{(3)}}}{\chi}^{\rm MT}\,$ compatible with the standard \.In\"on\"u--Wigner contraction $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\xrightarrow{R\to\infty}{\rm sMink}^{1,9\,\vert\,D_{1,9}}$.
\section{Another super-1-gerbe over $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,$ {\it via} a non-associative extension}\label{sec:sAdSSnJac}}
\section{A no-go for a class of Kosteleck\'y--Rabin deformations}\label{sec:KostRabdef}
In the next step, we pass to consider a class of Gra\ss mann-odd deformations of the Lie superalgebra $\,\gt{su}(2,2\,\vert\,4)\,$ motivated by our former experience with the super-Minkowskian (and super-Poincar\'e) algebra, which seems particularly well-founded and natural in the present context of the asymptotic analysis of associative deformations of the curved-superspacetime (super)algebra. Such considerations land us unavoidably on the third path indicated at the end of Sec.\,\ref{sec:ssextaAdSS}.
Taking into account the structure of the superstring extension of $\,{\rm sMink}^{d,1\,\vert\,D_{d,1}}\,$ recalled in Sec.\,\ref{subsect:sstringextsMink}, we postulate the structure relation
\begin{eqnarray}\nn
[Q_{\a\a' I},P_{\widehat a}]^\sim=-\tfrac{1}{2}\,(\widehat\Gamma_{\widehat a}\otimes\sigma_2)^{\beta\b'J}_{\ \ \a\a'I}\,Q_{\beta\b' J}-\tfrac{{\mathsf i}\,\a}{2}\,\bigl(\widehat C\,\widehat\Gamma_{\widehat a}\otimes\sigma_3\bigr)_{\a\a'I\beta\b'J}\,\mathcal{Z}^{\beta\b'J}\,,\qquad\a\in{\mathbb{C}}^\x
\end{eqnarray}
as the basis of the deformation. Given our previous identification of the source of the Gra\ss mann-odd charges (in the flat limit, and in the more general situation), the latter may rightly be dubbed a \textbf{Kosteleck\'y--Rabin deformation of} $\,\gt{su}(2,2\,\vert\,4)$.\ Indeed, in the 10-dimensional notation (recalled in App.\,\ref{app:CliffAdSS}), we have
\begin{eqnarray}\nn
[Q_{\a\a' I},P_{\widehat a}]^\sim=-\tfrac{{\mathsf i}}{2}\,(\unl\gamma_{\widehat a}\,\Delta^1_{\widehat a})^{\beta\b'J}_{\ \ \a\a'I}\,Q_{\beta\b' J}-\tfrac{{\mathsf i}\,\a}{2}\,\bigl(\mathcal{C}\,\unl\gamma_{\widehat a}\,\Delta^2_{\widehat a}\bigr)_{\a\a'I\beta\b'J}\,\mathcal{Z}^{\beta\b'J}\,,
\end{eqnarray}
so that if we rescale the positive-chirality spinors $\,\mathcal{Z}^{\beta\b'J}\,$ as
\begin{eqnarray}\nn
\mathcal{Z}^{\a\a'I}\longmapsto R^{\frac{3}{2}}\,\mathcal{Z}^{\a\a'I}
\end{eqnarray}
when performing the \.In\"on\"u--Wigner contraction \eqref{eq:IWrescale} the deformed relation asymptotes to the superstring deformation of the super-Minkowski superalgebra
\begin{eqnarray}\nn
[Q_{\a\a' I},P_{\widehat a}]^\sim=-\tfrac{{\mathsf i}\,\a}{2}\,\bigl(\mathcal{C}\,\unl\gamma_{\widehat a}\bigr)_{\a\a'I\beta\b'J}\,\mathcal{Z}^{\beta\b'J}
\end{eqnarray}
of \Reqref{eq:ssextsMink}. Incidentally, the above argument justifies our parametrisation of the deformation (by $\,\a$) \emph{independent} of the sector ($\widehat a\in\ovl{0,4}\,$ {\it vs} $\,\widehat a\in\ovl{5,9}$) in the decomposition of the body of the supertarget.
We still may, and -- speaking with hindsight -- actually need to accomodate the Gra\ss mann-even charge of the usual topological origin. As the latter corresponds to the presence of a single generator of $\,H_1({\rm AdS}_5\x{\mathbb{S}}^5)$,\ we are confronted with a choice: either we treat the charge as a ${\rm SO}(4,1)\x{\rm SO}(5)$-scalar, or as a ${\rm SO}(4,1)\x{\rm SO}(5)$-vector. We shall analyse both possibilities, calling the former, defined by the additional structure relation
\begin{eqnarray}\nn
\{Q_{\a\a' I},Q_{\beta\b' J}\}^\sim={\mathsf i}\,\bigl(-2(\widehat C\,\widehat\Gamma^{\widehat a}\otimes{\boldsymbol{1}})_{\a\a'I\beta\b'J}\,P_{\widehat a}+(\widehat C\,\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2)_{\a\a'I\beta\b'J}\,J_{\widehat a\widehat b}\bigr)+\beta^{\unl\mu}_{\rm s}\,\bigl(\widehat C\otimes\sigma_{\unl\mu}\bigr)_{\a\a'I\,\beta\b'J}\,\mathcal{Z}
\end{eqnarray}
with $\,\beta^{\unl\mu}_{\rm s}\in{\mathbb{C}}\,$ for some \emph{fixed} $\,\unl\mu\in\{0,1,3\}$,\ \textbf{the directional spinor-scalar Kosteleck\'y--Rabin deformation of} $\,\gt{su}(2,2\,\vert\,4)$,\ and the latter, with the additional structure relation
\begin{eqnarray}\nn
\{Q_{\a\a' I},Q_{\beta\b' J}\}^\sim={\mathsf i}\,\bigl(-2(\widehat C\,\widehat\Gamma^{\widehat a}\otimes{\boldsymbol{1}})_{\a\a'I\beta\b'J}\,P_{\widehat a}+(\widehat C\,\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_2)_{\a\a'I\beta\b'J}\,J_{\widehat a\widehat b}\bigr)+\beta^{\unl\mu}_{{\rm v}\,\widehat a}\,\bigl(\widehat C\,\widehat\Gamma_{\widehat a}\otimes\sigma_{\unl\mu}\bigr)_{\a\a'I\,\beta\b'J}\,\mathcal{Z}^{\widehat a}
\end{eqnarray}
with $\,\beta^{\unl\mu}_{{\rm v}\,\widehat a}\in{\mathbb{C}}\,$ for some \emph{fixed} $\,\unl\mu\in\{0,1,3\}$,\ \textbf{the directional spinor-vector Kosteleck\'y--Rabin deformation of} $\,\gt{su}(2,2\,\vert\,4)$.\ Here, it is presupposed that the formerly discussed rescaling of the original generators of the supersymmetry algebra and of the Gra\ss mann-odd charges is accompanied by
\begin{eqnarray}\nn
\mathcal{Z}^{(\widehat a)}\longmapsto\mathcal{Z}^{(\widehat a)}\,,
\end{eqnarray}
which ensures that the anticommutator of the supercharges asymptotes to the super-Minkowskian one, sourced by the surviving
\begin{eqnarray}\nn
-2{\mathsf i}\,\bigl(\widehat C\,\widehat\Gamma^{\widehat a}\otimes{\boldsymbol{1}}\bigr)\equiv-2{\mathsf i}\,\bigl(\mathcal{C}\,\unl\gamma^{\widehat a}\,\Delta^1_{\widehat a}\bigr)\,,
\end{eqnarray}
and we allow for two independent parameters $\,\beta^{\unl\mu}_{{\rm v}\,\widehat a}\,$ for $\,\widehat a\in\ovl{0,4}\,$ ($\beta^{\unl\mu}_{{\rm v}\,0}=\beta^{\unl\mu}_{{\rm v}\,1}=\beta^{\unl\mu}_{{\rm v}\,2}=\beta^{\unl\mu}_{{\rm v}\,3}=\beta^{\unl\mu}_{{\rm v}\,4}$) and $\,\widehat a\in\ovl{5,9}\,$ ($\beta^{\unl\mu}_{{\rm v}\,5}=\beta^{\unl\mu}_{{\rm v}\,6}=\beta^{\unl\mu}_{{\rm v}\,7}=\beta^{\unl\mu}_{{\rm v}\,8}=\beta^{\unl\mu}_{{\rm v}\,9}$), bearing in mind that the symmetry group has the product structure $\,{\rm SO}(4,1)\x{\rm SO}(5)\,$ (it is \emph{not} $\,{\rm SO}(9,1)$).
We shall deal with the spinor-scalar deformation first. Thus, assuming $\,\{Q_{\a\a'I},P_{\widehat a},J_{\widehat a\widehat b},\mathcal{Z}^{\a\a'I},\mathcal{Z}\}\,$ to be the generating set of the deformation $\,\widetilde{\gt{su}(2,2\,\vert\,4)}$,\ and -- as advocated earlier -- the body subalgebra $\,\gt{so}(4,2)\oplus\gt{so}(6)\,$ to be undeformed, we may write out the remaining \emph{deformed}\footnote{The undeformed structure relations have been left out.} structure relations:
\begin{eqnarray}\nn
&[\mathcal{Z}^{\a\a'I},J_{\widehat a\widehat b}]^\sim=\tfrac{\varepsilon_{\widehat a\widehat b}}{2}\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes{\boldsymbol{1}}\bigr)^{\a\a'I}_{\ \beta\b'J}\,\mathcal{Z}^{\beta\b'J}\,,\qquad\qquad[\mathcal{Z},J_{\widehat a\widehat b}]^\sim=0\,,&\cr\cr\cr
&[\mathcal{Z}^{\a\a'I},P_{\widehat a}]^\sim=\gamma^\mu_{\widehat a}\,\bigl(\widehat\Gamma_{\widehat a}\,\widehat C^{-1}\otimes\sigma_\mu\bigr)^{\a\a'I\beta\b'J}\,Q_{\beta\b'J}+\delta^\mu_{\widehat a}\,\bigl(\widehat\Gamma_{\widehat a}\otimes\sigma_\mu\bigr)\,^{\a\a'I}_{\ \beta\b'J}\,\mathcal{Z}^{\beta\b'J}\,,&\cr\cr
&\{\mathcal{Z}^{\a\a'I},Q_{\beta\b'J}\}^\sim=\varepsilon^{\widehat a,\mu}\,\bigl(\widehat\Gamma^{\widehat a}\otimes\sigma_\mu\bigr)_{\ \beta\b'J}^{\a\a'I}\,P_{\widehat a}+\zeta^{\widehat a\widehat b,\mu}\,\bigl(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_\mu\bigr)_{\ \beta\b'J}^{\a\a'I}\,J_{\widehat a\widehat b}+\eta^\mu\,\bigl({\boldsymbol{1}}\otimes\sigma_\mu\bigr)_{\ \beta\b'J}^{\a\a'I}\,\mathcal{Z}\,,&\cr\cr
&\{\mathcal{Z}^{\a\a'I},\mathcal{Z}^{\beta\b'J}\}^\sim=\theta^{\widehat a,K}\,\bigl(\widehat\Gamma^{\widehat a}\,\widehat C^{-1}\otimes\sigma_K\bigr)^{\a\a'I\beta\b'J}\,P_{\widehat a}+\iota^{\widehat a\widehat b}\,\bigl(\widehat\Gamma^{\widehat a\widehat b}\,\widehat C^{-1}\otimes\sigma_2\bigr)^{\a\a'I\beta\b'J}\,J_{\widehat a\widehat b}&\cr\cr
&\hspace{-2.25cm}+\kappa^K\,\bigl(\widehat C^{-1}\otimes\sigma_K\bigr)^{\a\a'I\beta\b'J}\,\mathcal{Z}\,,&\cr\cr\cr
&[\mathcal{Z},P_{\widehat a}]^\sim=\lambda_{\widehat a}\,P_{\widehat a}\,,\qquad\qquad[\mathcal{Z},Q_{\a\a'I}]^\sim=\mu^\nu\,\bigl({\boldsymbol{1}}\otimes\sigma_\nu\bigr)^{\beta\b'J}_{\ \a\a'I}\,Q_{\beta\b'J}+\nu^\mu\,\bigl(\widehat C\otimes\sigma_\mu\bigr)_{\a\a'I\beta\b'J}\,\mathcal{Z}^{\beta\b'J}\,,&\cr\cr
&[\mathcal{Z},\mathcal{Z}]^\sim=\rho\,\mathcal{Z}\,,&\cr\cr\cr
&[\mathcal{Z}^{\a\a'I},\mathcal{Z}]^\sim=\varsigma^\mu\,\bigl(\widehat C^{-1}\otimes\sigma_\mu\bigr)^{\a\a'I\beta\b'J}\,Q_{\beta\b'J}+\tau^\mu\,\bigl({\boldsymbol{1}}\otimes\sigma_\mu\bigr)^{\a\a'I}_{\ \beta\b'J}\,\mathcal{Z}^{\beta\b'J}\,,&
\end{eqnarray}
in which $\,\gamma^\mu_{\widehat a},\delta^\mu_{\widehat a},\varepsilon^{\widehat a,\mu},\zeta^{\widehat a\widehat b,\mu},\eta^\mu,\theta^{\widehat a,K},\iota^{\widehat a\widehat b},\kappa^K,\lambda_{\widehat a},\mu^\nu,\nu^\mu,\rho,\varsigma^\mu,\tau^\mu\in{\mathbb{C}}\,$ are parameters, with indices $\,\mu,\nu\in\{0,1,2,3\},K\in\{0,1,3\}\,$ and $\,\widehat b>\widehat a\in\ovl{0,9}\,$ summed over the respective ranges. The first two relations express the scalar and spinorial nature of the charges $\,\mathcal{Z}\,$ and $\,\mathcal{Z}^{\a\a'I}$,\ respectively. The remaining ones quantify the arbitrariness of the deformation within the bounds set by the former relations, and -- once again -- we are taking into account the product structure of the symmetry group by allowing $\,(\gamma^\mu_0,\delta^\mu_0)=(\gamma^\mu_1,\delta^\mu_1)=(\gamma^\mu_2,\delta^\mu_2)=(\gamma^\mu_3,\delta^\mu_3)=(\gamma^\mu_4,\delta^\mu_4)\,$ and $\,(\gamma^\mu_5,\delta^\mu_5)=(\gamma^\mu_6,\delta^\mu_6)=(\gamma^\mu_7,\delta^\mu_7)=(\gamma^\mu_8,\delta^\mu_8)=(\gamma^\mu_9,\delta^\mu_9)$,\ and similarly for the $\,\varepsilon^{\widehat a,\mu},\zeta^{\widehat a\widehat b,\mu},\theta^{\widehat a,K},\iota^{\widehat a\widehat b}$.\ We readily establish
\begin{Prop}\label{prop:ssKRproof}
There exist no associative directional spinor-scalar Kosteleck\'y--Rabin deformations of $\,\gt{su}(2,2\,\vert\,4)$.
\end{Prop}
\noindent\begin{proof}
A proof, based on a systematic imposition of the super-Jacobi identities, is given in App.\,\ref{app:ssKRproof}.
\end{proof}
We pass to examine the spinor-vector deformation. Here, the deformation is spanned on\footnote{We may consistently set the $\,\mathcal{Z}^{a'}\,$ to zero if we want to emphasise the geometric origin of the charge (there are no non-contractible 1-cycles in $\,{\mathbb{S}}^5$). This requires that some of the parameters of the deformation be nullified. As this scenario is subsumed by the more general one considered here, we do not discuss it separately.} $\,\{Q_{\a\a'I},P_{\widehat a},$ $J_{\widehat a\widehat b},\mathcal{Z}^{\a\a'I},\mathcal{Z}^{\widehat a}\}$,\ and so we end up with the \emph{deformed} structure relations:
\begin{eqnarray}\nn
&[\mathcal{Z}^{\a\a'I},J_{\widehat a\widehat b}]^\sim=\tfrac{\varepsilon_{\widehat a\widehat b}}{2}\,\bigl(\widehat\Gamma_{\widehat a\widehat b}\otimes{\boldsymbol{1}}\bigr)^{\a\a'I}_{\ \beta\b'J}\,\mathcal{Z}^{\beta\b'J}\,,\qquad\qquad[\mathcal{Z}^{\widehat a},J_{\widehat b\widehat c}]^\sim=\delta^{\widehat a}_{\ \widehat b}\,\mathcal{Z}_{\widehat c}-\delta^{\widehat a}_{\ \widehat c}\,\mathcal{Z}_{\widehat b}\,,&\cr\cr\cr
&[\mathcal{Z}^{\a\a'I},P_{\widehat a}]^\sim=\widetilde\gamma^\mu_{\widehat a}\,\bigl(\widehat\Gamma_{\widehat a}\,\widehat C^{-1}\otimes\sigma_\mu\bigr)^{\a\a'I\beta\b'J}\,Q_{\beta\b'J}+\widetilde\delta^\mu_{\widehat a}\,\bigl(\widehat\Gamma_{\widehat a}\otimes\sigma_\mu\bigr)\,^{\a\a'I}_{\ \beta\b'J}\,\mathcal{Z}^{\beta\b'J}\,,&\cr\cr
&\{\mathcal{Z}^{\a\a'I},Q_{\beta\b'J}\}^\sim=\widetilde\varepsilon^{\widehat a,\mu}\,\bigl(\widehat\Gamma^{\widehat a}\otimes\sigma_\mu\bigr)_{\ \beta\b'J}^{\a\a'I}\,P_{\widehat a}+\widetilde\zeta^{\widehat a\widehat b,\mu}\,\bigl(\widehat\Gamma^{\widehat a\widehat b}\otimes\sigma_\mu\bigr)_{\ \beta\b'J}^{\a\a'I}\,J_{\widehat a\widehat b}+\widetilde\eta_{\widehat a,\mu}\,\bigl(\widehat\Gamma_{\widehat a}\otimes\sigma_\mu\bigr)_{\ \beta\b'J}^{\a\a'I}\,\mathcal{Z}^{\widehat a}\,,&\cr\cr
&\{\mathcal{Z}^{\a\a'I},\mathcal{Z}^{\beta\b'J}\}^\sim=\theta^{\widehat a,K}\,\bigl(\widehat\Gamma^{\widehat a}\,\widehat C^{-1}\otimes\sigma_K\bigr)^{\a\a'I\beta\b'J}\,P_{\widehat a}+\widetilde\iota^{\widehat a\widehat b}\,\bigl(\widehat\Gamma^{\widehat a\widehat b}\,\widehat C^{-1}\otimes\sigma_2\bigr)^{\a\a'I\beta\b'J}\,J_{\widehat a\widehat b}&\cr\cr
&\hspace{-2cm}+\widetilde\kappa_{\widehat a}^K\,\bigl(\widehat\Gamma_{\widehat a}\,\widehat C^{-1}\otimes\sigma_K\bigr)^{\a\a'I\beta\b'J}\,\mathcal{Z}^{\widehat a}\,,&\cr\cr\cr
&[\mathcal{Z}^{\widehat a},P_{\widehat b}]^\sim=\eta^{\widehat a\widehat c}\,\widetilde\lambda_{\widehat c\widehat b}\,J_{\widehat c\widehat b}\,,\qquad\qquad[\mathcal{Z}^{\widehat a},Q_{\a\a'I}]^\sim=\widetilde\mu^{\nu,\widehat a}\,\bigl(\widehat\Gamma^{\widehat a}\otimes\sigma_\nu\bigr)^{\beta\b'J}_{\ \a\a'I}\,Q_{\beta\b'J}+\widetilde\nu^{\mu,\widehat a}\,\bigl(\widehat C\,\widehat\Gamma^{\widehat a}\otimes\sigma_\mu\bigr)_{\a\a'I\beta\b'J}\,\mathcal{Z}^{\beta\b'J}\,,&\cr\cr
&[\mathcal{Z}^{\widehat a},\mathcal{Z}^{\widehat b}]^\sim=\eta^{\widehat a\widehat c}\,\eta^{\widehat b\widehat d}\,\widetilde\rho^{\widehat c\widehat d}\,J_{\widehat c\widehat d}\,,&\cr\cr\cr
&[\mathcal{Z}^{\a\a'I},\mathcal{Z}^{\widehat a}]^\sim=\widetilde\varsigma^{\mu,\widehat a}\,\bigl(\widehat\Gamma^{\widehat a}\,\widehat C^{-1}\otimes\sigma_\mu\bigr)^{\a\a'I\beta\b'J}\,Q_{\beta\b'J}+\widetilde\tau^{\mu,\widehat a}\,\bigl(\widehat\Gamma^{\widehat a}\otimes\sigma_\mu\bigr)^{\a\a'I}_{\ \beta\b'J}\,\mathcal{Z}^{\beta\b'J}\,,&
\end{eqnarray}
in which $\,\widetilde\gamma^\mu_{\widehat a},\widetilde\delta^\mu_{\widehat a},\widetilde\varepsilon^{\widehat a,\mu},\widetilde\zeta^{\widehat a\widehat b,\mu},\widetilde\eta_{\widehat a,\mu},\widetilde\theta^{\widehat a}_K,\widetilde\iota^{\widehat a\widehat b},\widetilde\kappa_{a,K},\widetilde\lambda^{\widehat c\widehat d},\widetilde\mu^{\nu,\widehat a},\widetilde\nu^{\mu,\widehat a},\widetilde\rho^{\widehat c\widehat d},\widetilde\varsigma^{\mu,\widetilde a},\widetilde\tau^{\mu,\widetilde a}\in{\mathbb{C}}\,$ are parameters, with the repeated indices $\,\mu,\nu\in\{0,1,2,3\},K\in\{0,1,3\}\,$ and $\,\widehat b>\widehat a\in\ovl{0,9}\,$ summed over the respective ranges, and where -- as previously -- we are taking into account the product structure of the symmetry group. Once more, the first two relations express the vectorial and spinorial nature of the charges $\,\mathcal{Z}^{\widehat a}\,$ and $\,\mathcal{Z}^{\a\a'I}$,\ respectively. We find
\begin{Prop}\label{prop:svKRproof}
There exist no associative directional spinor-vector Kosteleck\'y--Rabin deformations of $\,\gt{su}(2,2\,\vert\,4)$.
\end{Prop}
\noindent\begin{proof}
A proof, once more based on a systematic imposition of the super-Jacobi identities, is given in App.\,\ref{app:svKRproof}.
\end{proof}
We conclude that the most natural (straightforward) deformations of the supersymmetry algebra $\,\gt{su}(2,2\,\vert\,4)\,$ of the supertarget under consideration asymptoting to the supercentral deformation \eqref{eq:ssextsMink} of the super-Minkowski superalgebra do not lead to a supersymmetric trivialisation of the Metsaev--Tseytlin super-3-cocycle as they leave the category of Lie superalgebras. This seems to point towards the absence of a trivialisation mechanism with the desired asymptotics, however\ldots
\section{Towards an \.In\"on\"u--Wigner-contractible super-1-gerbe on $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)$}\label{sec:MTaway}
Our hitherto analysis demonstrates the supersymmetry-invariant cohomology of the supertarget $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,$ (in degree 2) to be rather rigid and not amenable to trivialisation through associative deformation of the underlying Lie superalgebra compatible with the \.In\"on\"u--Wigner contraction to the super-Minkowski superalgebra. Indeed, the most natural deformations of $\,\gt{su}(2,2\,\vert\,4)\,$ either fail to trivialise the Metsaev--Tseytlin super-3-cocycle in such a way as to reproduce the super-Minkowskian trivialisation of the asymptotic Green--Schwarz super-3-cocycle in the flat limit or fail to define an extended associative structure with a Lie superbracket altogether. Thinking of the asymptotic transition between the curved geometry $\,{\rm AdS}_5\x{\mathbb{S}}^5\,$ and the flat geometry $\,{\rm Mink}^{9,1}\,$ and the tractable Lie-algebraic mechanism behind it as fundamental, we are led to invert the logic of our analysis and contemplate the possibility of deriving a Green--Schwarz super-3-cocycle, potentially different from the Metsaev--Tseytlin super-3-cocycle but with the same asymptotics, as the exterior derivative of a supersymmetric super-2-form on an extension of $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,$ (or even of another supermanifold with the same body and the structure of a homogeneous space of a Lie supergroup), both (the extension and the super-2-form) with the desired super-Minkowskian asymptotics. In order to give this rather non-specific and hence somewhat vague idea some flesh, we discuss a variant of a super-${\rm AdS}$ algebra conceived by Hatsuda, Kamimura and Sakaguchi in \Rcite{Hatsuda:2000mn}, with a built-in Gra\ss mann-odd deformation that makes it manifestly contractible, in the sense of \.In\"on\"u and Wigner, to the superstring-extended super-Minkowskian algebra of \Reqref{eq:ssextsMink}.
The point of departure is the so-called \textbf{${\rm AdS}$-algebra in $d+1$ dimensions}, {\it i.e.} the Lie algebra $\,\gt{so}(d,2)\,$ (of conformal transformations) which, in a suitable basis\footnote{The basis consists of simple linear combinations of the generators of translations and special conformal transformations, and those of the remaining conformal transformations: boosts, rotations and dilations.}, reads
\begin{eqnarray}\nn
&[P_a,P_b]=J_{ab}\,,\qquad\qquad[J_{ab},J_{cd}]=\eta_{ad}\,J_{bc}-\eta_{ac}\,J_{bd}+\eta_{bc}\,J_{ad}-\eta_{bd}\,J_{ac}\,,&\cr\cr
&[P_a,J_{bc}]=\eta_{ab}\,P_c-\eta_{ac}\,P_b\,,&\cr\cr
&(\eta_{ab})=\textrm{diag}(-1,\underbrace{1,1,\ldots,1}_d)\,,\qquad\qquad a,b,c,d\in\ovl{0,d}\,.&
\end{eqnarray}
We consider its Gra\ss mann-odd extension $\,\widetilde{\gt{sso}(d,2)}\,$ defined in terms of Majorana-spinor generators $\,Q_\a,\mathcal{Z}_\a,\ \a\in\ovl{1,2^{[\frac{d+1}{2}]}}\,$ through the relations
\begin{eqnarray}
&[Q_\a,J_{ab}]=-\tfrac{1}{2}\,\bigl(\Gamma_{ab}\bigr)^\beta_{\ \a}\,Q_\beta\,,\qquad\qquad[\mathcal{Z}_\a,J_{ab}]=-\tfrac{1}{2}\,\bigl(\Gamma_{ab}\bigr)^\beta_{\ \a}\,\mathcal{Z}_\beta\,,&\cr\cr
&[Q_\a,P_a]=-\tfrac{{\mathsf i}}{2}\,\bigl(\Gamma_a\bigr)^\beta_{\ \a}\,\mathcal{Z}_\beta\,,\qquad\qquad[\mathcal{Z}_\a,P_a]=\tfrac{{\mathsf i}}{2}\,\bigl(\Gamma_a\bigr)^\beta_{\ \a}\,Q_\beta\,,& \label{eq:sAdSext}\\ \cr
&\{Q_\a,Q_\beta\}=-2{\mathsf i}\,\bigl(C\,\Gamma^a\bigr)_{\a\beta}\,P_a\,,\qquad\qquad\{\mathcal{Z}_\a,\mathcal{Z}_\beta\}=2{\mathsf i}\,\bigl(C\,\Gamma^a\bigr)_{\a\beta}\,P_a\,,\qquad\qquad\{Q_\a,\mathcal{Z}_\beta\}=\bigl(C\,\Gamma^{ab}\bigr)_{\a\beta}\,J_{ab}\,.&\nn
\end{eqnarray}
Here, the $\,\Gamma_a\equiv\eta_{ab}\,\Gamma^b\,$ are generators of the Clifford algebra $\,{\rm Cliff}({\mathbb{R}}^{d,1})\,$ and
\begin{eqnarray}\nn
\Gamma_{ab}=\tfrac{1}{2}\,[\Gamma_a,\Gamma_b]\,,
\end{eqnarray}
all in a Majorana representation in which together with the charge-cojugation matrix $\,C\,$ they satisfy the identities
\begin{eqnarray}\nn
\bigl(C\,\Gamma_a\bigr)^{\rm T}=C\,\Gamma_a\,,\qquad\qquad\bigl(C\,\Gamma_{ab}\bigr)^{\rm T}=C\,\Gamma_{ab}\,.
\end{eqnarray}
Such a representation and the ensuing Lie-superalgebraic extension $\,\widetilde{\gt{sso}(d,2)}\,$ of the AdS-algebra are known to exist in dimension $\,d+1=3\,$ (with the charge-conjugation matrix $\,C^{\rm T}=-C\,$ determining the symmetry properties of the generators as $\,\Gamma_a^{\rm T}=-C\,\Gamma_a\,C^{-1}$). Basing on this observation, we simply \emph{assume} its existence for some $d$ and study the supergeometric consequences thereof. In so doing, we draw on our super-Minkowskian intuition. The latter is well justified as upon rescaling
\begin{eqnarray}\nn
\bigl(P_a,J_{ab},Q_\a,\mathcal{Z}_\a\bigr)\longmapsto\bigl(R\,P_a,J_{ab},R^{\frac{1}{2}}\,Q_\a,R^{\frac{3}{2}}\,\mathcal{Z}_\a\bigr)
\end{eqnarray}
the above algebra contracts, in the limit $\,R\to\infty$,\ to the superstring extension \eqref{eq:ssextsMink} of the Lie superalgebra $\,{\rm sMink}^{d,1\,\vert\,D_{d,1}}$,\ with
\begin{eqnarray}\label{eq:ZasZ}
Z^\a=\a\,\bigl(C^{-1}\bigr)^{\a\beta}\,\mathcal{Z}_\beta\,,
\end{eqnarray}
where $\,\a\in{\mathbb{C}}^\x\,$ is a constant that accounts for a suitable (constant) rescaling of the supercharges and momenta.
The Lie superalgebra \eqref{eq:sAdSext}, whenever it exists, integrates to a Lie supergroup which we denote as $\,\widetilde{{\rm sSO}(d,2)}\,$ in what follows. For the sake of the present discussion, we define \textbf{the extended super-${\rm AdS}_{d+1}\,$ space} to be the homogeneous space
\begin{eqnarray}\nn
\widetilde{{\rm sAdS}}_{d+1}:=\widetilde{{\rm sSO}(d,2)}/{\rm SO}(d,1)\,.
\end{eqnarray}
The latter is patchwise embedded in the total space $\,\widetilde{{\rm sSO}(d,2)}\,$ of the principal ${\rm SO}(d,1)$-bundle $\,\widetilde{{\rm sSO}(d,2)}\longrightarrow\widetilde{{\rm sSO}(d,2)}/{\rm SO}(d,1)\,$ by a collection of locally smooth sections
\begin{eqnarray}\nn
\widehat\sigma_i\ :\ \mathcal{O}_i\longrightarrow\widetilde{{\rm sSO}(d,2)}\ :\ \widehat Z_i\equiv\bigl(x^a_i,\theta^\a_i,\zeta^\beta_i\bigr)\longmapsto\widehat{\unl g}_i\cdot{\rm e}^{\zeta^\a_i\,\mathcal{Z}_\a}\cdot{\rm e}^{x^a_i\,P_a}\cdot{\rm e}^{\theta^\beta_i\,Q_\beta}\,,\qquad i\in I\,,
\end{eqnarray}
defined for an open cover $\,\{\mathcal{O}_i\}_{i\in I}\,$ of the base $\,\widetilde{{\rm sAdS}}_{d+1}$,\ with local coordinates $\,(x^a_i,\theta^\a_i,\zeta^\beta_i)\,$ associated with the direct-sum complement
\begin{eqnarray}\nn
\gt{t}=\bigoplus_{a=0}^d\,\corr{P_a}_{\mathbb{R}}\oplus\bigoplus_{\a=1}^{2^{[\frac{d+1}{2}]}}\,\corr{Q_\a}_{\mathbb{R}}\oplus\bigoplus_{\beta=1}^{2^{[\frac{d+1}{2}]}}\,\corr{\mathcal{Z}_\beta}_{\mathbb{R}}
\end{eqnarray}
of the Lorentz algebra $\,\gt{h}\equiv\gt{so}(d,1)\,$ within $\,\gt{g}\equiv\widetilde{\gt{sso}(d,2)}$,\ and for some reference elements $\,\widehat{\unl g}_i\in\widetilde{{\rm sSO}(d,2)}$.
We begin by writing out the pullback of the Maurer--Cartan super-1-form on $\,\widetilde{{\rm sSO}(d,2)}\,$ along one of the $\,\widehat\sigma_i$,
\begin{eqnarray}\nn
\widehat\sigma_i^*\theta_{\rm L}(\widehat Z_i)&=&\tfrac{1}{R^{\frac{1}{2}}}\,{\mathsf d}\theta^\a_i\otimes Q_\a+\tfrac{1}{R}\,\bigl({\mathsf d} x^a_i-{\mathsf i}\,\ovl\theta_i\,\Gamma^a\,{\mathsf d}\theta_i\bigr)\otimes P_a\cr\cr
&&+\tfrac{1}{R^{\frac{3}{2}}}\,\bigl({\mathsf d}\zeta^\a_i+\tfrac{{\mathsf i}}{2}\,{\mathsf d} x^a_i\,\theta^\beta_i\,\bigl(\Gamma_a\bigr)^\a_{\ \beta}+\tfrac{1}{3!}\,\ovl\theta_i\,\Gamma^a\,{\mathsf d}\theta_i\,\theta^\beta_i\,\bigl(\Gamma_a\bigr)^\a_{\ \beta}\bigr)\otimes\mathcal{Z}_\a\cr\cr
&&+\tfrac{1}{4R^2}\,\bigl(x^b_i\,{\mathsf d} x^a_i-x^a_i\,{\mathsf d} x^b_i+{\mathsf i}\,\ovl\theta_i\,\Gamma^{ab}\,\Gamma_c\,\theta_i\,\bigl({\mathsf d} x^c_i+\tfrac{{\mathsf i}}{3!}\,\ovl\theta_i\,\Gamma^c\,{\mathsf d}\theta_i\bigr)+4\ovl\theta_i\,\Gamma^{ab}\,{\mathsf d}\zeta_i\bigr)\otimes J_{ab}+O\bigl(R^{-\frac{5}{2}}\bigr)\,.
\end{eqnarray}
From the result, we infer the asymptotics of the various component super-1-forms (in order to distinguish the one associated with the supercharge $\,Q_\a\,$ from that associated with the Gra\ss mann-odd charge $\,\mathcal{Z}_\a$,\ we have put a tilde over the latter):
\begin{eqnarray}\nn
\widehat\sigma_i^*\theta_{\rm L}^\a(\widehat Z_i)&=&\tfrac{1}{R^{\frac{1}{2}}}\,{\mathsf d}\theta^\a+O\bigl(R^{-\frac{5}{2}}\bigr)\,,\cr\cr
\widehat\sigma_i^*\widetilde\theta_{\rm L}^\a(\widehat Z_i)&=&\tfrac{1}{R^{\frac{3}{2}}}\,\bigl({\mathsf d}\zeta^\a+\tfrac{{\mathsf i}}{2}\,{\mathsf d} x^a_i\,\theta^\beta_i\,\bigl(\Gamma_a\bigr)^\a_{\ \beta}+\tfrac{1}{3!}\,\ovl\theta_i\,\Gamma^a\,{\mathsf d}\theta_i\,\theta^\beta_i\,\bigl(\Gamma_a\bigr)^\a_{\ \beta}\bigr)+O\bigl(R^{-\frac{5}{2}}\bigr)\,,\cr\cr
\widehat\sigma_i^*\theta_{\rm L}^a(\widehat Z_i)&=&\tfrac{1}{R}\,\bigl({\mathsf d} x^a_i-{\mathsf i}\,\ovl\theta_i\,\Gamma^a\,{\mathsf d}\theta_i\bigr)+O\bigl(R^{-3}\bigr)\,,\cr\cr
\widehat\sigma_i^*\theta_{\rm L}^{ab}(\widehat Z_i)&=&\tfrac{1}{4R^2}\,\bigl(x^b_i\,{\mathsf d} x^a_i-x^a_i\,{\mathsf d} x^b_i+{\mathsf i}\,\ovl\theta_i\,\Gamma^{ab}\,\Gamma_c\,\theta_i\,\bigl({\mathsf d} x^c_i+\tfrac{{\mathsf i}}{3!}\,\ovl\theta_i\,\Gamma^c\,{\mathsf d}\theta_i\bigr)+4\ovl\theta_i\,\Gamma^{ab}\,{\mathsf d}\zeta_i\bigr)\otimes J_{ab}+O\bigl(R^{-3}\bigr)\,,
\end{eqnarray}
and so we conclude that the globally smooth left-invariant super-2-form
\begin{eqnarray}\nn
\underset{\tx{\ciut{(2)}}}{\widehat\beta}:=2{\mathsf i}\,\ovl\theta_{\rm L}\wedge\widetilde\theta_{\rm L}
\end{eqnarray}
asymptotes, upon pullback\footnote{By the previous reasoning, the pullbacks glue to a globally smooth super-2-form over $\,\widetilde{{\rm sAdS}}_{d+1}\,$ which we denote as $\,\underset{\tx{\ciut{(2)}}}{\widehat{\rm B}}$.} along $\,\widehat\sigma_i$,
\begin{eqnarray}\nn
\underset{\tx{\ciut{(2)}}}{\widehat{\rm B}}(\widehat Z_i)&\equiv&\widehat\sigma_i^*\underset{\tx{\ciut{(2)}}}{\widehat\beta}(\widehat Z_i)=\tfrac{2{\mathsf i}}{R^2}\,{\mathsf d}\theta_i^\a\wedge C_{\a\beta}\,\bigl({\mathsf d}\zeta^\beta_i+\tfrac{{\mathsf i}}{2}\,{\mathsf d} x^a_i\,\theta^\gamma_i\,\bigl(\Gamma_a\bigr)^\beta_{\ \gamma}+\tfrac{1}{3!}\,\ovl\theta_i\,\Gamma^a\,{\mathsf d}\theta_i\,\theta^\gamma_i\,\bigl(\Gamma_a\bigr)^\beta_{\ \gamma}\bigr)+O\bigl(R^{-3}\bigr)\cr\cr
&=&\tfrac{1}{R^2}\,\bigl(\ovl\theta_i\,\Gamma_a\,{\mathsf d}\theta_i\wedge{\mathsf d} x^a+2{\mathsf i}\,{\mathsf d}\ovl\theta_i\wedge{\mathsf d}\zeta_i\bigr)+O\bigl(R^{-3}\bigr)\,,
\end{eqnarray}
an overall rescaling by $\,R^2\,$ as well as a suitable redefinition (constant rescaling) of the $\,\theta^\a_i\,$ and $\,x^a_i\,$ and the identification
\begin{eqnarray}\nn
\xi_\a=\widetilde\a\,C_{\a\beta}\,\zeta^\beta_i
\end{eqnarray}
(the source of the constant $\,\widetilde\a\in{\mathbb{C}}^\x\,$ is similar to that of the constant $\,\a\,$ in \Reqref{eq:ZasZ}), to the supersymmetric primitive (I.5.19) of the Green--Schwarz super-3-cocycle $\,\underset{\tx{\ciut{(3)}}}{\chi}\,$ defined on the superstring-extended super-Minkowski space $\,\mathscr{M}_1^{(2)}\,$ of Prop.\,I.5.7. Thus, consistently with the postulate formulated at the beginning of the present section, we may take the left-invariant super-2-form as the point of departure of a (contractible, in the sense of \.In\"on\"u and Wigner) geometrisation of a Green--Schwarz super-3-cocycle
\begin{eqnarray}\nn
\underset{\tx{\ciut{(3)}}}{\widehat\chi}^{\rm GS}:={\mathsf d}\underset{\tx{\ciut{(2)}}}{\widehat\beta}
\end{eqnarray}
for the super-${\rm AdS}_d$ supertarget. The latter super-3-cocycle is readily derived with the help of the (super-)Maurer--Cartan equations,
\begin{eqnarray}\nn
\underset{\tx{\ciut{(3)}}}{\widehat\chi}^{\rm GS}&=&2{\mathsf i}\,\bigl({\mathsf d}\ovl\theta_{\rm L}\wedge\widetilde\theta_{\rm L}-\ovl\theta_{\rm L}\wedge{\mathsf d}\widetilde\theta_{\rm L}\bigr)\cr\cr
&=&2{\mathsf i}\,\bigl[\bigl(\tfrac{1}{2}\,\bigl(\Gamma_{ab}\bigr)^\a_{\ \gamma}\,\theta_{\rm L}^\gamma\wedge\theta_{\rm L}^{ab}-\tfrac{{\mathsf i}}{2}\,\bigl(\Gamma_a\bigr)^\a_{\ \gamma}\,\widetilde\theta_{\rm L}^\gamma\wedge\theta_{\rm L}^a\bigr)\wedge C_{\a\beta}\,\widetilde\theta_{\rm L}^\beta\cr\cr
&&-\theta_{\rm L}^\a\,C_{\a\beta}\wedge\bigl(\tfrac{1}{2}\,\bigl(\Gamma_{ab}\bigr)^\beta_{\ \gamma}\,\widetilde\theta_{\rm L}^\gamma\wedge\theta_{\rm L}^{ab}+\tfrac{{\mathsf i}}{2}\,\bigl(\Gamma_a\bigr)^\beta_{\ \gamma}\,\theta_{\rm L}^\gamma\wedge\theta_{\rm L}^a\bigr)\bigr]\cr\cr
&=&\ovl\theta_{\rm L}\wedge\Gamma_a\,\theta_{\rm L}\wedge\theta_{\rm L}^a+\ovl{\widetilde\theta}_{\rm L}\wedge\Gamma_a\,\widetilde\theta_{\rm L}\wedge\theta_{\rm L}^a\,.
\end{eqnarray}
While the two terms in the final expression are structurally identical, they behave differently in the flat-superspace limit $\,R\to\infty\,$ (upon pullback), to wit, the former scales as $\,R^{-2}$,\ whereas the latter scales as $\,R^{-4}$.\ Hence, after an overall rescaling by $\,R^2$,\ we wind up with the familiar Green--Schwarz super-3-cocycle \eqref{eq:sMinkGS3}. This means that the original guiding principles for the construction of the super-$\sigma$-model with the super-${\rm AdS}_{d+1}$ supertarget laid out in \Rcite{Metsaev:1998it} are obeyed, or -- in other words -- the super-$\sigma$-model with the super-3-form $\,\underset{\tx{\ciut{(3)}}}{\widehat\chi}^{\rm GS}\,$ as the Green--Schwarz super-3-cocycle is a valid\footnote{In fact, one should still verify the presence of a linearised gauged right supersymmetry, or Siegel's $\kappa$-symmetry. This issue was addressed in \Rcite{Hatsuda:2000mn}.} model of superloop mechanics over the homogeneous space $\,{\rm AdS}_d\,$ in the sense of Metsaev and Tseytlin.
On the other hand, at \emph{finite} $\,R$,\ the super-3-cocycle obtained above does \emph{not} descend to a sub(super)\-manifold within the extended super-${\rm AdS}_d$ space $\,\widetilde{{\rm sAdS}}_d\,$ which would correspond to a constant (zero) value of the extra coordinate $\,\zeta_i$.\ This implies that we should consider the super-$\sigma$-model with the extended supertarget $\,\widetilde{{\rm sAdS}}_d\,$ all along, and associate with it a trivial super-1-gerbe with the global curving $\,\underset{\tx{\ciut{(2)}}}{\widehat{\rm B}}$.\ It is only in the flat-superspace limit that the supports of the (asymptotic) Green--Schwarz super-3-cocycle and that of its supersymmetric primitive split and give rise to a non-trivial Cartan--Eilenberg super-1-gerbe of Part I. We hope to return to a systematic study of this supergeometric mechanism in a future work.
\section{Conclusions \& Outlook}\label{sec:C&O}
In the present paper, we have applied the general scheme, laid out and tested in the setting of the super-Minkowski space $\,{\rm sMink}^{d,1\,\vert\,D_{d,1}}\,$ in \Rcite{Suszek:2017xlw} (Part I), of a supersymmetry-equivariant geometrisation of (the physically distinguished) Green--Schwarz super-$(p+2)$-cocycles representing classes in the Cartan--Eilenberg cohomology of supersymmetry groups $\,{\rm G}$,\ through Lie-supergroup extensions defined by these classes, to the supergeometric data of the two-dimensional Metsaev--Tseytlin super-$\sigma$-model with the super-${\rm AdS}_5\x{\mathbb{S}}^5\,$ space as the supertarget. This places the scheme in a wider context of Cartan supergeometry of homogeneous spaces $\,\mathscr{M}\cong{\rm G}/{\rm H}\,$ of $\,{\rm G}\,$ corresponding to reductive decompositions of the supersymmetry algebra $\,\gt{g}=\gt{t}\oplus\gt{h}\,$ into a geometric (Gra\ss mann-even) Lie subalgebra $\,\gt{h}\,$ of the isotropy subgroup $\,{\rm H}\,$ of a point $\,x\in\mathscr{M}\,$ and its direct-sum complement $\,\gt{t}$,\ with a \emph{nontrivial} topology and a non-vanishing metric curvature in the body and \emph{no} Lie-supergroup structure assumed on $\,{\rm G}/{\rm H}$,\ and with the latter embedded patchwise smoothly in $\,{\rm G}\,$ by \emph{local} sections of the principal ${\rm H}$-bundle $\,{\rm G}\longrightarrow{\rm G}/{\rm H}\,$ and thus endowed with a non-linear realisation of supersymmetry. The relevance of the supertarget chosen for our case study hinges upon its r\^ole, as a critical superstring background, in the formulation and analysis of the celebrated ${\rm AdS}$/CFT correspondence.
The subtlety of the superbackground under consideration stems from the apparent incompatibility of the geometrisation with the \.In\"on\"u--Wigner contraction that underlies the transition to the flat geometry $\,{\rm sMink}^{9,1\,\vert\,D_{9,1}}$,\ dual to the flat limit of an infinite common radius of the generating 1-cycle of $\,{\rm AdS}_5\,$ and of the 5-sphere in the body $\,{\rm AdS}_5\x{\mathbb{S}}^5\,$ of the supertarget. Indeed, while the super-$\sigma$-model and the relevant (Metsaev--Tseytlin) super-3-cocycle do asymptote to their super-Minkowskian counterparts known from Part I, the manifestly supersymmetric primitive of the super-3-cocycle discovered in \Rcite{Roiban:2000yy} does not, and so neither does the associated trivial super-1-gerbe constructed in Section
\ref{sec:trsAdSSext}. The quest for a geometrisation (and hence also for an extension of the original Lie superalgebra $\,\gt{su}(2,2\,\vert\,4)$) consistent with the contraction has led us to analyse at some length, in Section \ref{sec:wrapanom}, the r\^ole of the (pseudo-supersymmetric) topological Wess--Zumino term in the action functional of a generic super-$\sigma$-model in the Polyakov formulation in the (classical) field-theoretic deformation of the supersymmetry (super)algebra furnished by the Poisson algebra of the corresponding Noether charges. The general mechanisms established in that section have been worked out explicitly, in Section \ref{sec:sMinkext}, for the one- and two-dimensional super-$\sigma$-models with the flat supertarget $\,{\rm sMink}^{d,1\,\vert\,D_{d,1}}\,$ and the Green--Schwarz super-$p$-cycles (for $\,p\in\{2,3\}$) studied in Part I, whereby the physical significance of the Gra\ss mann-odd Kosteleck\'y--Rabin charges, associated with a topological realisation of the Cartan--Eilenberg cohomology recalled and exploited in Part I, has been brought to light. This has prompted an asymptotic analysis, carried out in Section \ref{sec:ssextaAdSS}, of potential wrapping-charge deformations of the supersymmetry algebra $\,\gt{su}(2,2\,\vert\,4)\,$ of the super-$\sigma$-model with supertarget $\,{\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)\,$ induced by the de Rham cohomology class of the Roiban--Siegel supersymmetric primitive of the Metsaev--Tseytlin super-3-cocycle. Two types of (leading) contributions have been found: the Gra\ss mann-odd Kosteleck\'y--Rabin charges and purely geometric Gra\ss mann-even winding charges associated with the generating 1-cycle of $\,{\mathbb{S}}^1\x{\mathbb{R}}^{\x 4}\cong{\rm AdS}_5$.\ These have been used as motivation for a systematic exploration, in Sections \ref{sec:KamSakext} and \ref{sec:KostRabdef}, of the two most natural types of deformation of the supersymmetry algebra:
\begin{enumerate}
\item a Gra\ss mann-even central extension deforming (exclusively) the anticommutator of the supercharges in an arbitrary manner, contemplated with view to recovering the desired non-supersymmetric correction to the Roiban--Siegel primitive of the Metsaev--Tseytlin super-3-cocycle as the leading term in an asymptotic expansion of a supersymmetric super-2-form on the resultant extended supersymmetry group, and
\item a generic extension determined by a Gra\ss mann-odd deformation of the commutator $\,[Q_{\a\a'I},P_{\widehat a}]\,$ engineered so as to allow for a trivialisation, on the resultant extended supersymmetry group, of a super-2-cocycle asymptoting to the Kosteleck\'y--Rabin super-2-cocycle of Section \ref{subsect:sstringextsMink} in the limit of an infinite radius of $\,{\rm AdS}_5\x{\mathbb{S}}^5$.
\end{enumerate}
Both types of deformation have been ruled out as algebraically inconsistent. This has left us with the trivial super-1-gerbe of Section \ref{sec:trsAdSSext}, manifestly incompatible with the \.In\"on\"u--Wigner contraction, as the only working geometrisation of the Metsaev--Tseytlin super-3-cocycle to date. Finally, an alternative approach to the problem of geometrisation of the Green--Schwarz super-3-cocycle has been put forward in Section \ref{sec:MTaway}, promoting the asymptotic relation between the two extended supersymmetry algebras (for the super-Minkowski space and for the super-${\rm AdS}_5\x{\mathbb{S}}^5\,$ space, respectively), effected by an \.In\"on\"u--Wigner contraction, to the rank of the fundamental principle.
Our hitherto findings immediately suggest directions of further research. First and foremost, one should continue the search, initiated herein and motivated rather concretely in Section \ref{sec:MTaway}, for a super\-sym\-metry-equivariant geometrisation compatible with the \.In\"on\"u--Wigner contraction, of \emph{a} Green--Schwarz super-3-cocycle over \emph{a} super-${\rm AdS}_5\x{\mathbb{S}}^5\,$ space (understood in the spirit of the remarks made in Section \ref{sec:MTaway}) with the desired super-Minkowskian asymptotics. This would be expected, from a more general perspective, to give us insights into the higher-geometric realisation of the latter mechanism, acting in the tangent of the Lie supergroup and its homogeneous space, and, potentially, to bring further evidence of the significance, indicated strongly by our considerations, of the Kosteleck\'y--Rabin charges in the geometrisation scheme developed, the latter charges being an interesting subject of study in their own right. It is tempting to look for clues in this direction in the algebraically more tractable setting of the super-$\sigma$-models with supertarget of the form $\,{\rm s}({\rm AdS}_p\x{\mathbb{S}}^p)\,$ for $\,p\in\{2,3\}$,\ discussed in Refs.\,\cite{Zhou:1999sm,Rahmfeld:1998zn,Park:1998un}.
An absolutely fundamental -- from the physical point of view -- feature of any super-1-gerbe (to be constructed) over the super-${\rm AdS}_5\x{\mathbb{S}}^5\,$ space (or any other super-${\rm AdS}_p\x{\mathbb{S}}^p\,$ space, for that matter) is its weak $\kappa$-equivariance, as defined in the super-Minkowskian setting in \Rcite{Suszek:2017xlw}. This constatation determines yet another natural line of future research, understood as a continuation of the study initiated in the super-Minkowskian setting. As the symmetry is commonly regarded to be a basic building block in the construction of all the super-$\sigma$-models listed above, the existing literature forms a solid basis of a research thus oriented.
The relevance for our understanding of the physics of strongly coupled systems with gauge symmetry (such as, {\it e.g.}, the quark-gluon plasma) of the particular superstring background picked up for scrutiny in the present paper alone justifies pursuing the study taken up herein with view towards geometrising and thus elucidating the ${\rm AdS}$/CFT correspondence. This suggests that one ought to tackle the issue of classification and construction of supersymmetric bi-modules for any super-1-gerbe (to be constructed) over ${\rm s}({\rm AdS}_5\x{\mathbb{S}}^5)$.\ Given the nature of the Metsaev--Tseytlin super-$\sigma$-model, which is that of a descendant of a (supersymmetric) Wess--Zumino--Witten model for a supersymmetry group (to a homogeneous space thereof), it is natural to expect that the much-developed cohomological and group-theoretic techniques employed in the Lie-group setting in Refs.\,\cite{Fuchs:2007fw,Runkel:2009sp,Suszek:2018maxym} might prove to be of help.
Finally, there is a host of questions independent of the specific superstring background under consideration that were raised in the paper \cite{Suszek:2017xlw} and still await an in-depth study and elucidation. These include the question as to a structural relation between the intrinsically (Lie-super)algebraic geometrisation scheme furthered herein and the theory of Lie-$n$-superalgebras and $L_\infty$-superalgebras of Baez {\it et al.} considered in Refs.\,\cite{Baez:2004hda6,Baez:2010ye,Huerta:2011ic} and anchored firmly in the present field-theoretic context in \Rcite{Fiorenza:2013nha}, as well as the issue of correspondence between the Cartan--Eilenberg-cohomological framework developed for the two-dimensional super-$\sigma$-models and alternative approaches to supersymmetry in the context of superstring and related models, such as, in particular, the geometrisation, originally conceived by Killingback \cite{Killingback:1986rd} and Witten \cite{Witten:1988dls}, elaborated by Freed \cite{Yau:1987}, recently revived by Freed and Moore \cite{Freed:2004yc}, and ultimately concretised in the higher-geometric language by Bunke \cite{Bunke:2009} ({\it cp.}\ also \Rcite{Waldorf:2009uf} for an explicit construction), of the Pfaffian bundle of the target-space Dirac operator, associated with fermionic contributions to the superstring path integral, in terms of a differential ${\rm String}$-structure on the target space.
We hope to return to all the above ideas in a future work.
\newpage
|
\section{Introduction}
The neural network has been a successful learning machine during the past decade due to its highly expressive modeling capability, which is a consequence of multiple layers of non-linear transformations of input features.
Such transformations, however, make intermediate features ``latent,'' in the sense that they do not have explicit meaning and are not interpretable.
Therefore, neural networks are usually treated as black-box machinery.
Disentangling the latent space of neural networks has become an increasingly important research topic.
In the image domain, for example, \citeay{chen2016infogan} use adversarial and information maximization objectives to produce interpretable latent representations that can be tweaked to adjust writing style for handwritten digits, as well as lighting and orientation for face models.
\citeay{mathieu2016disentangling} utilize a convolutional autoencoder to achieve the same objective.
However, this problem is not well explored in natural language processing.
In this paper, we address the problem of disentangling the latent space of neural networks for text generation.
Our model is built on an autoencoder that encodes a sentence to the latent space (vector representation) by learning to reconstruct the sentence itself.
We would like the latent space to be disentangled with respect to different features, namely, \textit{style} and \textit{content} in our task.
To accomplish this, we propose a simple approach that combines multi-task and adversarial objectives.
We artificially divide the latent representation into two parts: the style space and content space. In this work, we consider the sentiment of a sentence as the style.
We design auxiliary losses, enforcing the separation of style and content latent spaces.
In particular, the multi-task loss operates on a latent space to ensure that the space does contain the information we wish to encode.
The adversarial loss, on the contrary, minimizes the predictability of information that should not be contained in that space.
In previous work, researchers typically work with the style, or specifically, sentiment space~\cite{hu2017toward,shen2017style,fu2018style}, but simply ignore the content space, as it is hard to formalize what ``content'' actually refers to.
In our paper, we propose to approximate the content information by bag-of-words (BoW) features, where we focus on style-neutral, non-stopwords.
Along with traditional style-oriented auxiliary losses, our BoW multi-task loss and BoW adversarial loss make the style and content spaces much more disentangled from each other.
The learned disentangled latent space can be directly used for text style-transfer~\cite{hu2017toward,shen2017style}, which aims to transform a given sentence to a new sentence with the same content but a different style.
Since it is difficult to obtain training sentence pairs with the same content and differing styles (i.e. parallel corpora), we follow the setting where we train our model on a non-parallel but style-labeled corpora. We call this \textit{non-parallel text style transfer}.
To accomplish this, we train an autoencoder with disentangled latent spaces.
For style-transfer inference, we simply use the autoencoder to encode the content vector of a sentence, but ignore its encoded style vector.
We then infer from the training data, an empirical embedding of the style that we would like to transfer.
The encoded content vector and the empirically-inferred style vector are concatenated and fed to the decoder.
This grafting technique enables us to obtain a new sentence similar in content to the input sentence, but with a different style.
We conducted experiments on two customer review datasets.
Qualitative and quantitative results show that both the style and content spaces are indeed disentangled well.
In the style-transfer evaluation, we achieve substantially better style-transfer strength, content preservation, and language fluency scores, compared with previous results.
Ablation tests also show that the auxiliary losses can be combined well, each playing its own role in disentangling the latent space.
\section{Related Work}
Disentangling neural networks' latent space has been explored in the image processing domain in the recent years, and researchers have successfully disentangled rotation features, color features, etc.~of images~\cite{chen2016infogan,luan2017deep}.
Some image characteristics (e.g., artistic style) can be captured well by certain statistics \cite{gatys2016image}.
In other work, researchers adopt data augmentation techniques to learn a disentangled latent space~\cite{kulkarni2015deep,champandard2016semantic}.
In natural language processing, the definition of ``style'' itself is vague, and as a convenient starting point, NLP researchers often treat sentiment as a salient style of text.
\citeay{hu2017toward} manage to control the sentiment by using discriminators to reconstruct sentiment and content from generated sentences.
However, there is no evidence that the latent space would be disentangled by this reconstruction.
\citeay{shen2017style} use a pair of adversarial discriminators to align the recurrent hidden decoder states of original and style-transferred sentences, for a given style.
\citeay{fu2018style} propose two approaches: training style-specific embeddings, and training separate style-specific decoders for style-transfer.
They apply an adversarial loss on the encoded space to discourage encoding style in the latent space of an autoencoding model. All the above approaches only deal with the style information and simply ignore the content part.
\citeay{zhao2018adversarially} extend the multi-decoder approach and use a Wasserstein-distance penalty to align content representations of sentences with different styles. However, the Wasserstein penalty is applied to empirical samples from the data distribution, and is more indirect than our BoW-based auxiliary losses.
Recently, \citeay{rao2018dear} treat the formality of writing as a style, and create a parallel corpus for style transfer with sequence-to-sequence models.
This is beyond the scope of our paper, as we focus on non-parallel text style transfer.
Our paper differs from previous work in that both our style space and content space are encoded from the input, and we design several auxiliary losses to ensure that each space encodes and only encodes the desired information.
Such disentanglement of latent space has its own research interest in the deep learning community.
The disentangled representation can be directly applied to non-parallel text style-transfer tasks, as in the aforementioned studies.
\section{Approach}
In this section, we describe our approach in detail, shown in Figure~\ref{fig:overview}.
Our model is built upon an autoencoder with a sequence-to-sequence neural network~\cite{sutskever2014sequence}, and we design multi-task and adversarial losses for both style and content spaces.
Finally, we present our approach to transfer style in the context of natural language generation.
\begin{figure}[!t]
\centering
\includegraphics[width=.9\linewidth]{model-overview}
\caption{Overview of our approach.}
\label{fig:overview}
\end{figure}
\subsection{Autoencoder} \label{ssec:seq2seq-autoencoder}
An autoencoder encodes an input to a latent vector space, from which it reconstructs the input itself.
The latent vector space is usually of much smaller dimensionality than input data, and the autoencoder learns salient and compact representations of data during the reconstruction process.
Let $\rmx=(x_1, x_2, \cdots x_n)$ be an input sequence with $n$ tokens.
The encoder recurrent neural network (RNN) with gated recurrent units (GRU) \cite{cho2014learning} encodes $\rm x$ and obtains a hidden vector representation $\bm h$, which is linearly transformed from the encoder RNN's final hidden state.
Then a decoder RNN generates a sentence, which ideally should be $\rmx$ itself.
Suppose at a time step $t$, the decoder RNN predicts the word $x_t$ with probability $p(x_t|\bm h, x_1\cdots x_{t-1})$. Then the autoencoder is trained with a sequence-aggregated cross-entropy loss, given by
\begin{equation}\label{eqn:ae}
\loss{AE}(\nnweight{E},\nnweight{D})= -\sum_{t=1}^n \log p(x_t|\bm h, x_1\cdots x_{t-1})
\end{equation}
where $\nnweight{E}$ and $\nnweight{D}$ are the parameters of the encoder and decoder, respectively.\footnote{For brevity, we only present the loss for a single data point (i.e., a sentence) throughout the paper. The total loss sums over all data points, and is implemented using mini-batches.} Both the encoder and decoder are deterministic functions in the original autoencoder model~\cite{rumelhart1985learning}, and thus we call it a \textit{deterministic autoencoder} (DAE).
\subsubsection{Variational Autoencoder.}
In addition to DAE, we also implement a variational autoencoder (VAE) \cite{kingma2013auto}, which imposes a probabilistic distribution on the latent vector. The Kullback-Leibler (KL) divergence \cite{kullback1951information} penalty is added to the loss function to regularize the latent space. The decoder reconstructs data based on the sampled latent vector from its posterior distribution.
Formally, the autoencoding loss in the VAE is
\begin{align}\label{eqn:vae}
\loss{AE}(\nnweight{E}, \nnweight{D}) = & - \mathbb{E}_{q_{E}(\bm h|\rmx)} [\log p(\rmx|\bm h)] \nonumber \\
& + \hyp{kl}\operatorname{KL}(q_{E}(\bm h|\rmx)\|p(\bm h))
\end{align}
where $\hyp{kl}$ is the hyperparameter balancing the reconstruction loss and the KL term. $p(\bm h)$ is the prior, set to the standard normal distribution $\mathcal{N}(\bm 0,\mathrm I)$. $q_E(\bm h|\mathrm x)$ is the posterior taking the form $\mathcal{N}(\bm \mu,\operatorname{diag} \bm\sigma)$, where $\bm\mu$ and $\bm\sigma$ are predicted by the encoder network.
The motivation for using VAE as opposed to DAE is that the reconstruction is based on the samples of the posterior, which populates encoded vectors to the neighborhood and thus smooths the latent space.
\citeay{bowman2016generating} show that VAE enables more fluent sentence generation from a latent space than DAE.
The autoencoding losses in Equations~(\ref{eqn:ae},\ref{eqn:vae}) serve as our primary training objective.
Besides, the autoencoder is also used for text generation in the style-transfer application.
We also design several auxiliary losses to disentangle the latent space. In particular, we hope that $\bm h$ can be separated into two spaces $\bm s$ and $\bm c$, representing style and content, respectively, i.e., $\bm h = [\bm s ; \bm c]$, where $[\cdot;\cdot]$ denotes concatenation.
This is accomplished by the auxiliary losses described in the rest of this section.
\subsection{Style-Oriented Losses}
We first design auxiliary losses that ensure the style information is contained in the style space $\bm s$.
This involves a multi-task loss that ensures $\bm s$ is discriminative for the style, as well as an adversarial loss that ensures $\bm c$ is not discriminative for the style.
\subsubsection{Multi-Task Loss for Style.} \label{ssec:multitask-style-objective}
Although the corpus we use is non-parallel, we assume that each sentence is labeled with its style. In particular, we treat the sentiment as the style of interest, following previous work~\cite{hu2017toward,shen2017style,fu2018style,zhao2018adversarially}, and each sentence is labeled with a binary sentiment tag (positive or negative).
We build a classifier on the style space that predicts the style label. Formally, a two-way softmax layer (equivalent to logistic regression) is applied to the style vector $\bm s$, given by
\begin{equation} \label{eqn:class-pred}
\bm y_s = \operatorname{softmax}(\weight{mul(s)} \bm s + \bias{mul(s)})
\end{equation}
where $\nnweight{mul(s)}=[\weight{mul(s)}; \bias{mul(s)}]$ are parameters for multi-task learning of style, and $\bm y_s$ is the output of softmax layer.
The classifier is trained with a simple cross-entropy loss against the ground truth distribution $t_s(\cdot)$, given by
\begin{equation} \label{eqn:style-multi-task-loss}
\loss{mul(s)}(\nnweight{E};\nnweight{mul(s)}) = - \sum\nolimits_{l\in\text{labels}} t_s(l)\log y_s(l)
\end{equation}
where $\nnweight{E}$ are the encoder's parameters.
We train the style classifier at the same time as the autoencoding loss.
Thus, this could be viewed as \textit{multi-task} learning, incentivizing the entire model to not only decode the sentence, but also predict its sentiment from the style vector $\bm s$.
We denote it by ``mul(s).''
The idea of multi-task losses is not new and has been used in previous work for sequence-to-sequence learning \cite{luong2015multi}, sentence representation learning \cite{jernite2017discourse} and sentiment analysis \cite{balikas2017multitask}, among others.
\subsubsection{Adversarial Loss for Style.}
\label{ssec:adversarial-style-objective}
The above multi-task loss only ensures that the style space contains style information.
However, the content space might also contain style information, which is undesirable for disentanglement.
We thus apply an adversarial loss to discourage the content space containing style information.
The idea is to first introduce a classifier, called an \textit{adversary}, that deliberately discriminates the true style label using the content vector $\bm c$.
Then the encoder is trained to learn a content vector space, from which its adversary cannot predict style information.
Concretely, the adversarial discriminator and its training objective have a similar form as Equations~\ref{eqn:class-pred} and~\ref{eqn:style-multi-task-loss}, but with different input and parameters, given by
\begin{align}
\label{eqn:adv-disc-loss}
\bm y_s & = \operatorname{softmax}(\weight{dis(s)} \bm c + \bias{dis(s)}) \\
\loss{dis(s)}(\nnweight{dis(s)}) & = - \sum\nolimits_{l\in\text{labels}} t_c(l)\log y_s(l)
\end{align}
where $\nnweight{dis(s)}=[\weight{dis(s)}; \bias{dis(s)}]$ are the parameters of the adversary.
It should be emphasized that, for the adversary, the gradients are not propagated back to the autoencoder, i.e., the variables in $\bm c$ are treated as shallow features. Therefore, we view $\loss{dis(s)}$ as a function of $\nnweight{dis(s)}$ only, whereas $\loss{mul(s)}$ is a function of both $\nnweight{E}$ and $\nnweight{mul(s)}$.
Having trained an adversary, we would like the autoencoder to be tuned in such an \textit{ad hoc} fashion, that $\bm c$ is not discriminative for style.
In existing literature, there could be different approaches, for example, maximizing the adversary's loss~\cite{shen2017style,zhao2018adversarially} or penalizing the entropy of the adversary's prediction~\cite{fu2018style}.
In our work, we adopt the latter, as it can be easily extended to multi-category classification, used for the content-oriented losses of our approach. Formally, the adversarial objective for the style is to maximize
\begin{equation} \label{eqn:advs}
\loss{adv(s)}(\nnweight{E})=\mathcal{H}(\bm y_s|\bm c; \nnweight{dis(s)})
\end{equation}
where $\mathcal{H}(\bm p)=-\sum_{i\in\text{labels}}p_i\log p_i$ and $\bm y_s$ is the predicted distribution over the style labels. Here, $\loss{adv(s)}$ is maximized with respect to the encoder, and attains maximum value when $\bm y_s$ is a uniform distribution. It is viewed as a function of $\nnweight{E}$, and we fix $\nnweight{dis(s)}$.
While adversarial loss has been explored in previous style-transfer papers~\cite{shen2017style,fu2018style}, it has not been combined with the multi-task loss. As we shall show in our experiments, combining these two losses is promisingly effective, achieving better style transfer performance than a variety of previous state-of-the-art methods.
\subsection{Content-Oriented Losses}
The above style-oriented losses only regularize style information, but they do not impose any constraint on how the content information should be encoded. This also happens in most previous work~\cite{hu2017toward,shen2017style,fu2018style}.
Although the style space is usually much smaller than the content space, it is unrealistic to expect that the content would not flow into the style space because of its limited capacity. Therefore, we need to design content-oriented auxiliary losses to regularize the content information.
Inspired by the above combination of multi-task and adversarial losses, we apply the same idea to the content space. However, it is usually hard to define what ``content'' actually refers to.
To this end, we propose to approximate the content information by bag-of-words (BoW) features.
The BoW features of an input sentence is a vector, each element indicating the probability of a word's occurrence in the sentence.
For a sentence $\rmx$ with $N$ words, the word $w_*$'s BoW probability is given by
$t_c(w_*)=\frac{\sum_{i=1}^{N}{\mathbb{I}\{w_i = w_*\}}}{N}$,
where $t_c(\cdot)$ denotes the target distribution of content, and $\mathbb{I\{\cdot\}}$ is an indicator function.
Here, we only consider content words, excluding stopwords and style-specific words, since we focus on ``content'' information. In particular, we exclude sentiment words from a curated lexicon \cite{hu2004mining} for sentiment style transfer.
The effect of using different vocabularies for BoW is analyzed in Supplemental Material A.
\subsubsection{Multi-Task Loss for Content.} \label{ssec:multitask-content-objective}
Similar to the style-oriented loss, the multi-task loss for content, denoted as ``mul(c)'', ensures that the content space $\bm c$ contains content information, i.e., BoW features.
We introduce a softmax classifier over the BoW vocabulary
\begin{equation} \label{eqn:bow-pred}
\bm y_c = \operatorname{softmax}({\weight{mul(c)}} \bm c + \bias{mul(c)})
\end{equation}
where $\nnweight{mul(c)}=[\weight{mul(c)}; \bias{mul(c)}]$ are the classifier's parameters, and $\bm y_c$ is the predicted BoW distribution.
The training objective is a cross-entropy loss against the ground truth distribution $t_c(\cdot)$, given by
\begin{equation}\label{eqn:content-multi-task-loss}
\loss{mul(c)}(\nnweight{E};\nnweight{mul(c)}) = - \sum\nolimits_{w\in\text{vocab}} t_c(w)\log y_c(w)
\end{equation}
where the optimization is performed with both encoder parameters $\nnweight{E}$ and the multi-task classifier $\nnweight{mul(c)}$.
Notice that although the target distribution is not one-hot as for BoW prediction, the cross-entropy loss (Equation~\ref{eqn:content-multi-task-loss}) has the same form.
It is also interesting that, at first glance, the multi-task loss for content appears to be redundant, given the autoencoding loss, when in fact, it is not. The multi-task loss only considers content words, which do not include stopwords and sentiment words, and is only applied to the content space $\bm c$. This ensures that the content information is captured in the content space. The autoencoding loss only requires that the model reconstructs the sentence based on the content and style space, and does not ensure their separation.
\subsubsection{Adversarial Loss for Content.} \label{ssec:adversarial-content-objective}
To ensure that the style space does not contain content information, we design our final auxiliary loss, the adversarial loss for content, denoted as ``adv(c).''
We build an adversary, a softmax classifier on the style space to predict BoW features, approximating content information, given by
\begin{align}
\label{eqn:adv-bow-disc-loss}
\bm y_c & = \operatorname{softmax}({\weight{dis(c)}}^\top \bm s + \bias{dis(c)}) \\
\loss{dis(c)}(\nnweight{dis(c)}) & = - \sum\nolimits_{w\in\text{vocab}} t_c(w)\log y_c(w)
\end{align}
where $\nnweight{dis(c)}=[\weight{dis(c)}; \bias{dis(c)}]$ are the classifier's parameters for BoW prediction.
The adversarial loss for the model is to maximize the entropy of the discriminator
\begin{equation}
\loss{adv(c)}(\nnweight{E}) = \mathcal{H}(\bm y_c | \bm s; \nnweight{dis(c)})
\end{equation}
Again, $\loss{dis(c)}$ is trained with respect to the discriminator's parameters $\nnweight{dis(c)}$, whereas $\loss{adv(c)}$ is trained with respect to $\nnweight{E}$, similar to the adversarial loss for style.
\subsection{Training Process}
The overall loss $\loss{ovr}$ for the autoencoder comprises several terms: the reconstruction objective, the multi-task objectives for style and content, and the adversarial objectives for style and content:
\begin{align}
\loss{ovr} = & \loss{AE}(\nnweight{E}, \nnweight{D}) \\
& + \hyp{mul(s)} \loss{mul(s)} (\nnweight{E},\nnweight{mul(s)}) - \hyp{adv(s)} \loss{adv(s)}(\nnweight{E}) \nonumber \\
& + \hyp{mul(c)} \loss{mul(c)} (\nnweight{E},\nnweight{mul(c)}) - \hyp{adv(c)} \loss{adv(c)}(\nnweight{E}) \nonumber
\end{align}
where $\lambda$'s are the hyperparameters that balance the autoencoding loss and these auxiliary losses.
To put it all together, the model training involves an alternation of optimizing discriminator losses $\loss{dis(s)}$ and $\loss{dis(c)}$, and the model's own loss $\loss{ovr}$, shown in Algorithm~\ref{alg:training-process}.
\begin{algorithm}[!t]
\ForEach{mini-batch}{
minimize $\loss{dis(s)}(\nnweight{dis(s)})$ w.r.t. $\nnweight{dis(s)}$\;
minimize $\loss{dis(c)}(\nnweight{dis(c)})$ w.r.t. $\nnweight{dis(c)}$\;
minimize $\loss{ovr}$ w.r.t. $\nnweight{E}, \nnweight{D}, \nnweight{mul(s)}, \nnweight{mul(c)}$\;
}
\caption{Training process.}\vspace{-.2cm}
\label{alg:training-process}
\end{algorithm}
\subsection{Generating Style-Transferred Sentences} \label{ssec:sentence-generation}
A direct application of our disentangled latent space is style-transfer for natural language generation.
For example, we can generate a sentence with generally the same meaning (content) but a different style (e.g., sentiment).
Let $\rmx^*$ be an input sentence with $\bm s^*$ and $\bm c^*$ being the encoded, disentangled style and content vectors, respectively.
If we would like to transfer its content to a different style, we compute an empirical estimate of the target style's vector $\hat{\bm s}$ using
\begin{equation*}
\hat{\bm s}=\frac{\sum_{i\in\text{target style}}\bm s_i}{\text{\# target style samples}}
\end{equation*}
The inferred target style $\hat{\bm s}$ is concatenated with the encoded content $\bm c^*$ for decoding style-transferred sentences, as shown in Figure~\ref{fig:overview}b.
\section{Experiments}
\subsection{Datasets}
We conducted experiments on two datasets, Yelp and Amazon reviews.
Both of these datasets comprise sentences accompanied by binary sentiment labels (positive, negative). They are used to train latent space disentanglement as well as to evaluate sentiment transfer.
\subsubsection{Yelp Service Reviews.}
We used a Yelp review dataset, following previous work \cite{shen2017style,zhao2018adversarially}.\footnote{The Yelp dataset is available at \url{https://github.com/shentianxiao/language-style-transfer}}
It contains 444101, 63483 and 126670 labeled reviews for train, validation, and test, respectively.
The maximum review length is 15 words, and the vocabulary size is approximately 9200.
\subsubsection{Amazon Product Reviews.}
We further evaluate our model with an Amazon review dataset, following another previous paper~\cite{fu2018style}.\footnote{The Amazon dataset is available at \url{https://github.com/fuzhenxin/text_style_transfer}}
It contains 559142, 2000 and 2000 labeled reviews for train, validation, and test, respectively.
The maximum review length is 20 words, and the vocabulary size is approximately 58000.
\subsection{Experiment Settings}
We used the Adam optimizer \cite{kingma2014adam} for the autoencoder and the RMSProp optimizer \cite{tieleman2012lecture} for the discriminators, following adversarial training stability tricks \cite{arjovsky2017wasserstein}.
Each optimizer has an initial learning rate of $10^{-3}$.
Our model is trained for 20 epochs, by which time it has mostly converged.
The word embedding layer was initialized by word2vec \cite{mikolov2013distributed} trained on respective training sets.
Both the autoencoder and the discriminators are trained once per mini-batch with $\hyp{mul(s)} = 10$, $\hyp{mul(c)} = 3$, $\hyp{adv(s)} = 1$, and $\hyp{adv(c)} = 0.03$.
These hyperparameters were tuned by performing a log-scale grid search within two orders of magnitude around the default value $1$, and choosing those that yielded the best validation results.
The recurrent unit size is $256$, the style vector size is $8$, and the content vector size is $128$.
We append the latent vector $\bm h$ to the hidden state at every time step of the decoder.
For the VAE model, we enforce the KL-divergence penalty on both the style and content posterior distributions, using $\hyp{kl(s)}$ and $\hyp{kl(c)}$, respectively.
We set $\hyp{kl(s)} = 0.03$ and $\hyp{kl(c)} = 0.03$ and use the $\operatorname{sigmoid}$ KL-weight annealing schedule following \citeay{bahuleyan2018probabilistic}.
They were tuned in the same manner as the other hyperparameters of the model.
\subsection{Experiment I: Disentangling Latent Space}
First, we analyze how the style (sentiment) and content of the latent space are disentangled.
We train classifiers on the different latent spaces, and report their inference-time classification accuracies in Table~\ref{tab:classification-yelp}.
We see that the 128-dimensional content vector $\bm c$ is not particularly discriminative for style.
It achieves accuracies slightly better than majority guess.
However, the 8-dimensional style vector $\bm s$, despite its low dimensionality, achieves substantially higher style classification accuracy.
When combining content and style vectors, we observe no further improvement.
These results verify the effectiveness of our disentangling approach, as the style space contains style information, whereas the content space does not.
We show t-SNE plots of both the deterministic autoencoder (DAE) and the variational autoencoder (VAE) models in Figure~\ref{fig:tsne-plots}.
As seen, sentences with different styles are noticeably separated in a clean manner in the style space (LHS), but are indistinguishable in the content space (RHS).
It is also evident that the latent space learned by the variational autoencoder is considerably smoother and continuous compared with the one learned by the deterministic autoencoder.
We show t-SNE plots for ablation tests with different combinations of auxiliary losses in Supplemental Material B.
\begin{table}[!t]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{|l||r|r||r|r|}
\hline
\multirow{2}{*}{\textbf{Latent Space}} & \multicolumn{2}{c||}{\textbf{Yelp}} & \multicolumn{2}{c|}{\textbf{Amazon}} \\
\cline{2-5}
& \textbf{DAE} & \textbf{VAE} & \textbf{DAE} & \tabh{VAE} \\
\hline \hline
None (majority guess) & \multicolumn{2}{c||}{0.602} & \multicolumn{2}{c|}{0.512} \\
\hline
Content space ($\bm c$) & 0.658 & 0.697 & 0.675 & 0.693 \\ \hline
Style space ($\bm s$) & 0.974 & 0.974 & 0.821 & 0.810 \\ \hline
Complete space ($[\bm s;\bm c]$) & 0.974 & 0.974 & 0.819 & 0.810 \\
\hline
\end{tabular}}
\caption{Classification accuracy on latent spaces.}
\label{tab:classification-yelp}
\end{table}
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{latent-spaces}
\caption{t-SNE plots of the disentangled style and content spaces (with all auxiliary losses on the Yelp dataset).}
\label{fig:tsne-plots}
\end{figure}
\begin{table*}[!t]
\centering
\resizebox{\textwidth}{!}{
\begin{tabular}{|l||c|c|c|c||c|c|c|c| }
\hline
\tabc{3}{Model} & \multicolumn{4}{c||}{\textbf{Yelp Dataset}} & \multicolumn{4}{c|}{\textbf{Amazon Dataset}} \\
\cline{2-9}
& \textbf{Transfer} & \textbf{Cosine} & \textbf{Word} & \textbf{Language} & \textbf{Transfer} & \textbf{Cosine} & \textbf{Word} & \textbf{Language} \\
& \textbf{Accuracy} & \textbf{Similarity} & \textbf{Overlap} & \textbf{Fluency} & \textbf{Accuracy} & \textbf{Similarity} & \textbf{Overlap} & \textbf{Fluency} \\
\hline
\hline
Style-Embedding \cite{fu2018style} & 0.182 & \textbf{0.959} & \textbf{0.666} & -16.17 & \ 0.400$^\dag$ & \ \textbf{0.930}$^\dag$ & \textbf{0.359} & -28.13 \\
\hline \hline
Cross-Alignment \cite{shen2017style} & \ 0.784$^\dag$ & 0.892 & 0.209 & -23.39 & 0.606 & 0.893 & 0.024 & -26.31 \\
\hline
Multi-Decoder \cite{zhao2018adversarially} & \ 0.818$^\dag$ & 0.883 & 0.272 & -20.95 & 0.552 & \textit{0.926} & 0.169 & -34.70 \\
\hline
Ours (DAE) & 0.883 & \textit{0.915} & \textit{0.549} & -10.17 & 0.720 & 0.921 & \textit{0.354} & -24.74 \\
\hline
Ours (VAE) & \textbf{0.934} & {0.904} & {0.473} & \textbf{-9.84} & \textbf{0.822} & 0.900 & 0.196 & \textbf{-21.70} \\
\hline
\end{tabular}}\vspace{-.2cm}
\caption{Performance of non-parallel text style transfer. The style-embedding approach achieves poor transfer accuracy, and should not be considered as an effective style-transfer model. Despite this, our model outperforms other previous methods in terms of all aspects (transfer strength, content preservation, and language fluency). Numbers with the $^\dag$ symbol are quoted from respective papers. Others are based on our replication using the published code in previous work. Our replicated experiments achieve 0.809 and 0.835 transfer accuracy on the Yelp dataset, close to the results in \protect\citeay{shen2017style} and \citeay{zhao2018adversarially}, respectively, showing that our replication is fair.}\vspace{-.2cm}
\label{tab:yelp-comparison-previous}
\end{table*}
\subsection{Experiment II: Non-Parallel Text Style Transfer}
We also conducted sentiment transfer experiments with our disentangled latent space.
\subsubsection{Metrics.} We evaluate competing models based on (1) style transfer strength, (2) content preservation and (3) quality of generated language. The evaluation of generated sentences is a difficult task in contemporary literature, so we adopt a few automatic metrics and use human judgment as well.
$\bullet$ \textit{Style-Transfer Accuracy.} We follow most previous work~\cite{hu2017toward,shen2017style,fu2018style} and train a separate convolutional neural network (CNN) to predict the sentiment of a sentence~\cite{kim2014convolutional}, which is then used to approximate the style transfer accuracy. In other words, we report the CNN classifier's accuracy on the style-transferred sentences, considering the target style to be the ground truth.
While the style classifier itself may not be perfect, it achieves a reasonable sentiment accuracy on the validation sets ($97\%$ for Yelp; $82\%$ for Amazon). Thus, it provides a quantitative way of evaluating the strength of style-transfer.
$\bullet$ \textit{Cosine Similarity.}
We followed \citeay{fu2018style} and computed a sentence embedding by concatenating the $\operatorname{min}$, $\operatorname{max}$, and $\operatorname{mean}$ of its word embeddings (sentiment words removed).
Then, we computed the cosine similarity between the source and generated sentence embeddings, which is intended to be an indicator of content preservation.
$\bullet$ \textit{Word Overlap.} We find that the cosine similarity measure, although correlated to human judgment, is not a sensitive measure, and we propose a simple yet effective measure that counts the unigram word overlap rate of the original sentence $\mathrm x$ and the style-transferred sentence $\mathrm y$, computed by $\frac{count(w_{\mathrm x} \cap w_{\mathrm y})}{count(w_{\mathrm x} \cup w_{\mathrm y})}$.
$\bullet$ \textit{Language Fluency.}
We use a trigram Kneser-Ney (KL) smoothed language model \cite{kneser1995improved} as a quantitative and automated metric to evaluate the fluency of a sentence.
It estimates the empirical distribution of trigrams in a corpus, and computes the log-likelihood of a test sentence.
We train the language model on the respective dataset, and report the Kneser-Ney language model's log-likelihood. A larger (closer to zero) number indicates a more fluent sentence.
$\bullet$ \textit{Manual Evaluation.}
Despite the above automatic metrics, we also conduct human evaluations to further confirm the performance of our model.
This was done on the Yelp dataset only, due to the amount of manual effort involved.
We asked 6 human evaluators to rate each sentence on a 1--5 Likert scale~\cite{stent2005evaluating} in terms of transfer strength, content similarity, and language quality.
This evaluation was conducted in a strictly blind fashion: samples obtained from all evaluated models are randomly shuffled, so that the evaluator would be unaware of which model generated a particular sentence.
The inter-rater agreement---as measured by Krippendorff's alpha \cite{krippendorf2004content} for our Likert scale ratings---is 0.74, 0.68, and 0.72 for transfer strength, content preservation, and language quality, respectively.
According to \citeay{krippendorf2004content}, this is an acceptable inter-rater agreement.
\begin{table}[!t]
\centering\vspace{-.2cm}
\resizebox{\linewidth}{!}{
\begin{tabular}{| l || c | c | c | }
\hline
\tabc{2}{Model} & \tabh{Transfer} & \tabh{Content} & \tabh{Language} \\
& \tabh{Strength} & \tabh{Preservation} & \tabh{Quality} \\
\hline
\hline \citeay{fu2018style} & 1.67 & \textbf{3.84} & 3.66 \\\hline\hline
\citeay{shen2017style} & 3.63 & 3.07 & 3.08 \\
\hline
\citeay{zhao2018adversarially} & 3.55 & 3.09 & 3.77 \\
\hline
Ours (DAE) & 3.67 & 3.64 & 4.19 \\
\hline
Ours (VAE) & \textbf{4.32} & \textit{3.73} & \textbf{4.48} \\
\hline
\end{tabular}}
\caption{Manual evaluation on the Yelp dataset.}
\label{tab:manual-evaluation}
\end{table}
\subsubsection{Results and Analysis.}
We compare our approach with previous state-of-the-art work in Table \ref{tab:yelp-comparison-previous}.
For baseline methods, we quoted results from existing papers whenever possible, and replicated the experiments to report other metrics with publicly available code \cite{shen2017style,fu2018style,zhao2018adversarially}.\footnote{\citeay{fu2018style} propose another model using multiple decoders; the method is further developed in \citeay{zhao2018adversarially}, and we adopt the latter for comparison.} As discussed in Table~\ref{tab:yelp-comparison-previous}, our replication involves reasonable efforts and is fair for comparison.
\begin{table*}[ht]
\centering\vspace{-.2cm}
\begin{tabular}{| l || c | c | c | c |}
\hline
\tabc{2}{Objectives} & \tabh{Transfer} & \tabh{Cosine} & \tabh{Word} & \tabh{Language} \\
& \tabh{Accuracy} & \tabh{Similarity} & \tabh{Overlap} & \tabh{Fluency} \\
\hline
\hline
$\loss{AE}$ & 0.106 & \textbf{0.939} & 0.472 & -12.58 \\
\hline
$\loss{AE}$, $\loss{mul(s)}$ & 0.767 & 0.911 & 0.331 & -12.17 \\
\hline
$\loss{VAE}$, $\loss{adv(s)}$ & 0.782 & 0.886 & 0.230 & -12.03 \\
\hline
$\loss{VAE}$, $\loss{mul(s)}$, $\loss{adv(s)}$ & 0.912 & 0.866 & 0.171 & -9.59 \\
\hline
$\loss{VAE}$, $\loss{mul(s)}$, $\loss{adv(s)}$, $\loss{mul(c)}$, $\loss{adv(c)}$ & \textbf{0.934} & 0.904 & \textbf{0.473} & \textbf{-9.84} \\
\hline
\end{tabular}\vspace{-.2cm}
\caption{Ablation tests on the Yelp dataset. In all variants, we follow the same protocol of style transfer by substituting an empirical estimate of the target style vector.}
\label{tab:ablation-results}
\end{table*}
\begin{table*}[!t]
\centering
\resizebox{.95\textwidth}{!}{\footnotesize
\begin{tabular}{| p{0.3\linewidth} || p{0.3\linewidth} | p{0.3\linewidth} |}
\hline
\tabc{1}{Original (Positive)} & \tabh{DAE Transferred (Negative)} & \tabh{VAE Transferred (Negative)} \\
\hline
\hline
the food is excellent and the service is exceptional & the food was a bit bad but the staff was exceptional & the food was bland and i am not thrilled with this \\
\hline
the waitresses are friendly and helpful & the guys are rude and helpful & the waitresses are rude and are lazy \\
\hline
the restaurant itself is romantic and quiet & the restaurant itself is awkward and quite crowded & the restaurant itself was dirty \\
\hline
great deal & horrible deal & no deal \\
\hline
both times i have eaten the lunch buffet and it was outstanding & their burgers were decent but the eggs were not the consistency & both times i have eaten here the food was mediocre at best \\
\hline
\hline
\tabc{1}{Original (Negative)} & \tabh{DAE Transferred (Positive)} & \tabh{VAE Transferred (Positive)} \\
\hline
\hline
the desserts were very bland & the desserts were very good & the desserts were very good \\
\hline
it was a bed of lettuce and spinach with some italian meats and cheeses & it was a beautiful setting and just had a large variety of german flavors & it was a huge assortment of flavors and italian food \\
\hline
the people behind the counter were not friendly whatsoever & the best selection behind the register and service presentation & the people behind the counter is friendly caring \\
\hline
the interior is old and generally falling apart & the decor is old and now perfectly & the interior is old and noble \\
\hline
they are clueless & they are stoked & they are genuinely professionals \\
\hline
\end{tabular}}\vspace{-.2cm}
\caption{Examples of style transferred sentence generation.}\vspace{-.2cm}
\label{tab:transfer-samples}
\end{table*}
We observe that the style embedding model \cite{fu2018style} performs poorly on the style-transfer objective,\footnote{It should be noted that the transfer accuracy is lower bounded by 0\% as opposed to 50\%, because we always transfer a sentence to the opposite sentiment. The lower-bound, zero transfer accuracy, is achieved by a trivial model that copies the input.} resulting in inflated cosine similarity and word overlap scores. We also examined the number of times each model generates exact copies of the source sentences during style transfer. We notice that the style-embedding model simply reconstructs the exact source sentence $24\%$ of the time, whereas all other models do this less than $6\%$ of the time. Therefore, we do not think that the style embedding approach is an effective model for text style transfer.
The other two competing methods~\cite{shen2017style,zhao2018adversarially} achieve reasonable transfer accuracy and cosine similarity. However, our model outperforms them by 10\% transfer accuracy as well as content preserving scores (measured by cosine similarity and the word overlap rate). This shows our model is able to generate high-quality style transferred sentences, which in turn indicates that the latent space is well disentangled into style and content subspaces. Regarding language fluency, we see that VAE is better than DAE in both experiments. This is expected as VAE regularizes the latent space by imposing a probabilistic distribution. We also see that our method achieves considerably more fluent sentences than competing methods, showing that our multi-task and adversarial losses are more ``natural'' than other methods, for example, aligning RNN hidden states \cite{shen2017style}.
Table~\ref{tab:manual-evaluation} presents the results of human evaluation. Again, we see that the style embedding model \cite{fu2018style} is ineffective as it has a very low transfer strength, and that our method outperforms other baselines in all aspects. The results are consistent with the automatic metrics in both experiments (Table~\ref{tab:yelp-comparison-previous}). This implies that the automatic metrics we used are reasonable; it also shows consistent evidence of the effectiveness of our approach.
We conducted ablation tests on the Yelp dataset, and show results in Table~\ref{tab:ablation-results}. With $\loss{VAE}$ only, we cannot achieve reasonable style transfer accuracy by substituting an empirically estimated style vector of the target style. This is because the style and content spaces would not be disentangled spontaneously with the autoencoding loss alone.
With either $\loss{mul(s)}$ or $\loss{adv(s)}$, the model achieves reasonable transfer accuracy and cosine similarity. Combining them together improves the transfer accuracy to 90\%, outperforming previous methods by a margin of 10\% (Table~\ref{tab:yelp-comparison-previous}). This shows that the multi-task loss and the adversarial loss work in different ways. Our insight of combining the two auxiliary losses is a simple yet effective way of disentangling latent space.
However, $\loss{mul(s)}$ and $\loss{adv(s)}$ only regularize the style information, leading to gradual drop of content preserving scores. Then, we have another insight of introducing content-oriented auxiliary losses, $\loss{mul(c)}$ and $\loss{adv(c)}$, based on BoW features, which regularize the content information in the same way as the style information. By incorporating all these auxiliary losses, we achieve high transfer accuracy, high content preservation, as well as high language fluency.
Table~\ref{tab:transfer-samples} provides several examples of our style-transfer model. Results show that we can successfully transfer the sentiment while preserving the content of a sentence.
We see that, with the empirically estimated style vector, we can reliably control the sentiment of generated sentences.
\section{Conclusion}
In this paper, we propose a simple yet effective approach for disentangling the latent space of neural networks.
We combine multi-task and adversarial objectives to separate content and style information from each other, and propose to approximate content information with bag-of-words features of style-neutral, non-stopword vocabulary.
Both qualitative and quantitative experiments show that the latent space is indeed separated into style and content parts.
This disentangled space can be directly applied to text style-transfer tasks.
It achieves substantially better style-transfer strength, content-preservation scores, as well as language fluency, compared with previous state-of-the-art work.
|
\section{Introduction}
Reservoir simuations are important tools for petroleum engineering, which have been applied to model
underground flows and interations between reservoirs and wells to predict the well
performance such as oil rates, water rates and bottom hole pressure. When a reservoir model is large enough,
it may take severals days or even longer for the simulator to complete a single model. Therefore, the numerical
methods and parallel techniques are essential to accelerate the simulations.
As a type of unconventional reservoirs, the reservoir with natural fractures and hydraulic fractures which are
commonly seen in tight and shale oil and gas reservoir, the dual-porosity/dual-permeability model and the
multiple iteration contiua (MINC) model\cite{pruess,wu-pruess} homogenize the fractures and use
superpositioned cells to represent the fractures and the matrix. The multiple continuum approaches have been
successfully employed in the unconventional reservoir problems. Wu et al.\cite{wu-qin,wu-di} regarded the
fractured vuggy rocks as a triple- or multiple-continuum medium with highly permeability and well-connected
fractures, low-permeabilty rock matrix and various-sized vugs. Wu et al. also developed a hybrid
multiple-continuum-medium modeling appoach to describe different types of fractures including hydrautic
fractures, natural fracture network and micro fractures \cite{wu-li}. Jiang and Moinfar and their
collaborators designed explicit fracture
models coupled with MINC model to simulate the unconventional reservoirs \cite{moinfar,jiang}. In reference
\cite{wu-li-ding}, the authors presented a unified
framework model to incorporate the known mechanisms and processes including gas adsorption and desorption,
geomechanics effect, Klinkenberg or gas-slippage effect and non-Darcy flow. They also used a hybrid
fracture-modeling appoach to simulate the unconventional gas reservoirs. Jiang and Younis developed two
alternative hybrid approaches to capture the effects of the multiscale fracture systems by combining the
advantages of the multi-continuum and discrete-fracture/matrix representations \cite{jiang-younis}.
The numerical methods for reservoir simulations have been explored by both mathematicians and engineers for
decades. Killough et al. \cite{killough} studied local refinement
techniques to improve the accuracy and reduce computational cost comparing to the global grid refinement.
Those local refinement techniques are useful for the complex models such as in-situ combustion. Dogru and his
group \cite{dogru} developed parallel simulators using structured and unstructured grids to handle faults,
pinchouts, complex wells, polymer flooding in non-fractured and fractured reservoirs. Zhang et al. developed
a general-purpose parallel platform for large-scale scientific applications which was designed using the
adaptive element methods and the adaptive finite volumn methods \cite{zhang,zhang-cui-liu}. It has
been applied to black oil simulation using discontinuous Galerkin methods\cite{wang-zhang-chen}. Chen et al.
studied finite element emtjods and finite different methods for black oil, compositional and thermal
models \cite{chen-huan-ma}. They studied Newton methods, linear solvers and preconditioners as well.
Chen and his collaborators also developed a parallel platform to support the new-generation simulator
development, such as black oil, compositional, thermal, polymer flooding models
\cite{para-frame,WLC-JCP,full-field}. Wheeler studied
discretization methods, linear solvers, preconditioner techniques and developed a parallel black oil
simulator \cite{wheeler}. For the linear solver and preconditioners, many preconditioning methods have been
proposed and applied to reservoir simulations such as constrained pressure residual (CPR)
methods \cite{cpr,cao}, multi-stage methods \cite{dogru-wheeler}, multiple level preconditioners \cite{turkey},
fast auxiliary space preconditioners (FASP) \cite{auxiliary} and a family of parallel CPR-like
methods\cite{cpr}.
In this paper, we present the mathematical model of the dual-porosity reservoirs.
After that, we give the numerical methods used in our simulation inlcuding the nonlinear solvers,
the preconditioner and the linear solver and the parallel technique.
Finally, we show the numerical results obtained from our simulator.
\section{Mathematical Model}
Darcy's law describes the flow of a fluid through porous media, establishing the relationship
between the volumtric flow rate and the pressure gradient:
\begin{equation}
Q = -\frac{KA\Delta p}{\mu L}
\end{equation}
where $K$ is the permeability of a given reservoir, $A$ is the area in the flow direction,
$\Delta p$ is the pressure difference, $\mu$ is the viscosity of the fluid and $L$ is the
length of the reservoir. In three dimensional space, its differential form is
\begin{equation}
q =\frac{Q}{A}= -\frac{K\nabla p}{\mu }
\end{equation}
Combining Darcy's law and the mass conservation law for oil and water components, the
two-phase model is as follows:
\begin{equation}
\left\{
\begin{array}{ll}
\frac{\partial}{\partial t}(\phi s_o\rho_o) &= \nabla \cdot (\frac{{\bm K}K_{ro}\rho_o}{\mu_o}\nabla \Phi_o) + q_o\\
\frac{\partial}{\partial t}(\phi s_w \rho_w ) &= \nabla \cdot (\frac{{\bm K}K_{rw}\rho_w}{\mu_w} \nabla \Phi_w) + q_w\\
\end{array}
\right.
\end{equation}
where $\phi$ and ${\bm K}$ are the porosity and the permeabilities (in $x$-, $y$- and $z$-directions) respectively,
$\Phi_\alpha$ is the potential, $s_\alpha$, $\mu_\alpha$, $K_{r\alpha}$ and $q_\alpha$
($\alpha = o, w$) are the saturation, viscosity, relative permeability and production/injection rate,
respectively, which satisfy the following constraints:
\begin{equation}
\left\{
\begin{array}{l}
\Phi_\alpha = p_\alpha - \rho_\alpha g z,\\
s_o+s_w = 1,\\
p_w = p_o - p_{cow}(s_w),\\
\end{array}
\right.
\end{equation}
where $p_\alpha$ and $\rho_\alpha$ are the pressure and the mass density of phase $\alpha$($\alpha=o,~w$). $p_{cow}$
is the capillary pressure between oil phase and water phase, $g$ is the gravitational constant and $z$
is the reservoir depth.
\begin{figure}[!ht]\centering
\includegraphics[scale = 0.2]{matrix-fracture.eps}
\caption{The flow commmunication for dual porosity model \cite{cmg}}
\label{matrix-fracture}
\end{figure}
For the dual-porosity model, fluids can flow between matrix and fracture, as well as fracture and fracture,
but there is no flow between matrices. There flow directions are described by Fig \ref{matrix-fracture}.
The following equations are used to describe this model:
\begin{equation}\label{prob}
\left\{
\begin{array}{l}
\frac{\partial}{\partial t}({\phi_f s_{o,f} \rho_{o,f}}) = \nabla \cdot (\frac{{\bm K}_fK_{ro_f}
\rho_{o,f}}{\mu_{o,f}}\nabla \Phi_{o,f})- q_{o,mf}+ q_o,\\
\frac{\partial}{\partial t}({\phi_m s_{o,m} \rho_{o,f}}) = q_{o,mf},\\
\frac{\partial}{\partial t}({\phi_f s_{w,f} \rho_{w,f}}) = \nabla \cdot (\frac{{\bm K}_fK_{rw_f}
\rho_{w,f}}{\mu_{w,f}b_{w,f}}\nabla \Phi_{w,f})-q_{w,mf}+ q_w,\\
\frac{\partial}{\partial t}({\phi_m s_{w,m} \rho_{w,m}}) = q_{w,mf}
\end{array}
\right.
\end{equation}
where $(\cdot)_f$ denotes the variables for the fracture and $(\cdot)_m$ for the matrix.
Note that here $q_{\alpha,mf} = \omega (\frac{K_{r\alpha} \rho_{\alpha,t}}{\mu_\alpha})_t(p_f - p_m)$ is a source
term which represents the net addition of the fluid to the fracture from the matrix with $\omega$ the shape
factor. In our simulator, the shape factor we use is based on the work of Warren and Root\cite{warren-root}
and Gilman and Kazemi\cite{gilman-kazemi}. Note that $(\cdot)_t$ denoted the value at $t
= m$ or $t = f$ depending on which is the upsteam. The well rate at a perforation, $i$, is modelled by the sink-source method,
\begin{equation}
q_{\alpha,i} = W_i \frac{{\bm K}_fK_{r\alpha_f} \rho_{\alpha,f}}{\mu_{\alpha,f}}(p_b - p), \quad \alpha = o,~w
\end{equation}
where $W_i$ is the well index at the perforation, $p_b$ is the bottom-hole pressure of the well at the perforation, $p$ is the
block pressure from oil phase at the perforation.
For a producer or an injection well, its well rate is calculated as,
\begin{equation}
q_\alpha = \sum_{i}q_{\alpha,i}, \quad \alpha = o,~w.
\end{equation}
The well index is calculated as follows:
\begin{equation}
W_i = 2\pi* k_h*w_{frac}/(\ln(r_e/r_w)+s).
\end{equation}
Here $k_h$ is a given value and $w_{frac}$ is the facture between 0.0 and 1.0 specifying the fraction of a circle
that the well models, $r_w$ is the wellbore radius, $s$ is a real number specifying the well skin factor,
$r_e$ is the well effective radius calculated from
\begin{equation}
r_e = w_{g}*\sqrt{A_{r}/(\pi*w_{frac})}
\end{equation}
with $w_{g}$ the geometric factor for the well element, $A_{r}$ the area perpendicular to the
wellbore (e.g., $D_x*D_y$ for a vertical well) and $w_{frac}$ as same above.
The other option for the well effective radius is the following Peaceman's form \cite{peaceman}
\begin{equation}\displaystyle
\left\{
\begin{array}{l}
r_{e,x} = 0.28\frac{\displaystyle[D_x^2(\frac{k_y}{k_x})^{1/2} + D_y^2(\frac{k_x}{k_y})^{1/2}]^{\frac{1}{2}}}{\displaystyle(\frac{k_y}{k_x})^{1/4}+ (\frac{k_x}{k_y})^{1/4}},\\
r_{e,y} = 0.28\frac{\displaystyle[D_x^2(\frac{k_z}{k_x})^{1/2} + D_z^2(\frac{k_x}{k_z})^{1/2}]^{\frac{1}{2}}}{\displaystyle(\frac{k_z}{k_x})^{1/4}+ (\frac{k_x}{k_z})^{1/4}},\\
r_{e,z} = 0.28\frac{\displaystyle[D_y^2(\frac{k_z}{k_y})^{1/2} + D_z^2(\frac{k_y}{k_z})^{1/2}]^{\frac{1}{2}}}{(\displaystyle\frac{k_z}{k_y})^{1/4}+ (\frac{k_y}{k_z})^{1/4}},\\
\end{array}
\right.
\end{equation}
where $r_{e,x}$, $r_{e,y}$ and $r_{e,z}$ are the well effective radius in $x$, $y$ and $z$ direction respectively,
$D_x$, $D_y$ and $D_z$ are the block sizes in $x$, $y$ and $z$ direction respectively, $k_x$, $k_y$ and $k_z$
are the permeabilities in $x$, $y$ and $z$ direction respectively.
\section{Numerical Methods and Parallel Computing}
In this paper, the fully implicit method (FIM) is applie, and the oil phase pressure $p$, water saturation
$s_w$ and the well bottom hole pressure $p_b$ are unknowns. For the time differential and the space
differential term, we use the backward Euler scheme and cell-centered finite difference method as
discretization methods.
The system is highly nonlinear,
\begin{equation}\label{nonlinear}
F(x) = 0
\end{equation}
where $F$ is a nonlinear map from $\mathcal {R}^N$ to $\mathcal {R}^N$ with $N = 2
\times n + \tau$. Here $n$ is the number of grid blocks and $\tau$ is the well size.
In this nonlinear equation, the properties related to the saturation are strongly nonlinear
while the properties related to the pressure are weakly nonlinear.
In our simulation, we use Newton method (or inexact Newton method) to solve it.
\subsection{Nonlinear Solver}
The inexact Newton method can be regarded as an extension of the standard Newton method. Let $A$ be the Jacobian
matrix of $F(x)$ and $y$ is the correction between the last step approximate solution and the current step one.
Its algorithm \cite{chen-gewecke} is as follows:
\begin{algorithm}[!htb]
\caption{The Inexact Newton Method}
\label{inewton-alg}
\begin{algorithmic}[1]
\STATE Given an initial guess $x^0$ and a termination tolerance $\epsilon$, let $l = 0$
and assemble the right-hand side $b$.
\WHILE{$\left\|b \right\| \ge \epsilon$}
\STATE Compute the Jacobian matrix $A$.
\STATE Determine $\eta_l$.
\STATE Find a solution $y$ such that
\begin{equation}
\label{inexact-newton-alg}
\left\| A y - b \right\| \leq \eta_l \left\| b \right\|,
\end{equation}
\STATE Let $l = l+1$ and $x^l = x^{l-1} + y$.
\STATE Compute the new right-hand side $b$
\ENDWHILE
\STATE $x^* = x^l$ is the approximate solution of the nonlinear system, $F(x) = 0$.
\end{algorithmic}
\end{algorithm}
We can see that the algorithm is the same as Newton method except the choice
of the tolerance $\theta$. The standard Newton method usually applies a small constant such as $1.0e-6$, while
the {\bf Algorithm 1} uses adaptive $\theta_l$ tolerance to avoid over solution. Common methods are defined
as follows \cite{chen-gewecke}:
\begin{equation}
\theta_l =\left\{
\begin{array}{l}
\displaystyle\frac{\|b(x^l) - r^{l-1}\|}{\|b(x^{l-1})\|},\\
\displaystyle\frac{\|b(x^l)\| - \|r^{l-1}\|}{\|b(x^{l-1})\|},\\
\displaystyle\gamma (\frac{\|b(x^l)\|}{\|b(x^{l-1})\|})^\beta, \quad \gamma\in [0,~1], \quad \beta\in (1,2]
\end{array}
\right.
\end{equation}
where $r^l$ is the residual of the $l$-th iteration.
\subsection{Preconditioner and Linear Solver}
For each Newton iteration, we need to solve a linear system $Ax = b$ which is very time
comsuming. To improve the efficiency, we choose a proper preconditioner, CPR-FPF method \cite{cpr}, for the
model. Each grid block has four unknowns, pressures ($p_f$ and $p_m$) and saturations ($s_{w,f}$ and
$s_{w,m}$) for matrix and fracture, and four
equations. In our linear system, each unknown and each row are arranged cell by cell, in this case, the
Jacobian matrix is a block matrix. It is also well-known that block matrix has better convergence that
point-wise matrix.
Before solving the linear system $Ax=b$, a decoupling operation $D$ is applied, and an equivalent
system is obtained:
\begin{equation}\label{decoupled-eq}
D^{-1}Ax = D^{-1} b.
\end{equation}
A proper decoupling method can improve linear solver dramatically. Many
decoupling strategies have been proposed, such as Quasi-IMPES\cite{quasi-impes} method and ABF\cite{abf}
method. In this paper, a modified Gauss-Jordan elimination method is applied.
\subsection{Parallel Computing}
The simulator is based on our parallel platform PRSI \cite{para-frame}, which is developed using C language
and MPI (Message Passing Interface). The platform has implemented many modules, such as
grid generation, load balancing, well management, parallel input and output, distributed matrix and vector, linear solver
and preconditioner, communication management and visualization.
Based on the platform, physical modules, such as rock properties and rock-fluid properties, are implemented.
More details can be read from reference \cite{para-frame}.
\section{Numerical Results}
\begin{example}
\label{ex1}
The grid dimension is $10\times 10\times 1$ with sizes 102.04 ft. in $x$ and $y$ directions
and 100.0 ft. in $z$ direction from top to bottom. The depth of the top layer center is 2000.0 ft.
The permeabilities for the matrix in $x$, $y$ and $z$ directions are 100, 100, 10 mD respectively.
The permeability for the fracture in $x$, $y$ and $z$ directions are 395.85 mD.
The reference porosity for the matrix is 0.1392. The reference porosity for the fracture is 0.039585.
The rock compressibilities for the matrix and fracture are both 3e-06 (1/psi). The reference pressure
is 15.0 psi for both the matrix and the fracture.
Component properties: densities of oil and water are 58.0 lbm/ft3 and 62.4 lbm/ft3 respectively.
The reference pressure is 15 psi at which the oil formation volume factor is 1.036 RB/STB and the oil viscosity is 40.0 cp.
The oil compressibility is const 1.313e-5 l/psi.
The initial conditions are as follows: initial pressure for the matrix is 2000 psi, initial pressure for the fracture is 1980 psi,
and initial water saturations are 0.08 and 0.01 in matrix and fracture respectively.
There are one injection well and two production wells. All of them are vertical.
Injection well has maximum water injection rate 500.0 bbl/day, maximum bottom
hole pressure 5.0e+4 psi, well index 200.0 with perforation at cell [5 1 1].
Both of the production well has maximum oil production rate 300.0 STB/day, minimum bottom hole pressure 15 psi
with well radius 0.25ft. The perforation of Produer 1 is at cell [1 10 1] while the perforation of Produer 2 is at cell [10 10 1].
The simulation period is 800 days.
\end{example}
The results of oil production rate, bottom-hole-pressure and water rate are shown in
Fig. \ref{fig-dp-oil-rate}, \ref{fig-dp-bhp}, \ref{fig-dp-water-rate}, from which we can see that the results
from our simulator and from CMG IMEX match very well. This proves our methods and implementation are
correct.
\begin{figure}[!ht]
\centering
\includegraphics[width = 0.5\textwidth]{dp-oil-rate.eps}
\caption{Example \ref{ex1}: Oil production rate (unit: STB/day).}
\label{fig-dp-oil-rate}
\end{figure}
\begin{figure}[!ht]
\centering
\includegraphics[width = 0.5\textwidth]{dp-bhp.eps}
\caption{Example \ref{ex1}: Well bottom-hole pressure (pressure unit: psi).}
\label{fig-dp-bhp}
\end{figure}
\begin{figure}[!ht]\centering
\includegraphics[width = 0.5\textwidth]{dp-water-rate.eps}
\caption{Example \ref{ex1}: Water rate (unit: STB/day).}
\label{fig-dp-water-rate}
\end{figure}
\begin{example}
\label{ex2}
The grid dimension is $10\times 10 \times 3$ with mesh size 102 ft. in $x$ and $y$ directions
and 100.0 ft. in $z$ direction from top to bottom. The depth of the top layer
center is 2000.0 ft.
The permeabilities for the matrix in $x$, $y$ and $z$ directions are 100, 100, 100 mD respectively.
The permeability for the fracture in $x$, $y$ and $z$ directions are 395.85 mD.
The reference porosity for the matrix is 0.1392. The reference porosity for the fracture is 0.039585.
The rock compressibilities for the matrix and fracture are both 3e-06 (1/psi). The reference pressure
is 15.0 for both the matrix and the fracture.
Component properties: densities of oil and water are 58.0 lbm/ft3 and 62.4 lbm/ft3 respectively.
PVT: the reference pressure is 15 psi at which the oil formation volume factor is 1.036 RB/STB and the oil viscosity is 40.0 cp.
The oil compressibility is const 1.313e-5 l/psi.
The initial conditions are as follows: initial pressure for the matrix is 800 psi, initial pressure for the fracture is 500 psi,
and initial water saturations are 0.08 and 0.01 in matrix and fracture respectively.
There are one injection well and two production wells. All of them are vertical.
Injection well has maximum water injection rate 200.0 bbl/day, maximum bottom
hole pressure 5.0e+4 psi, well index 200.0 with perforation
at cell [5 5 1].
Both of the production well has maximum oil production rate 500.0 STB/day, minimum
bottom hole pressure 15 psi with well radius 0.25ft. The perforation of Produer 1 is at
cell [1 1 1] while the perforation of Produer 2 is at cell [10 10 1].
The simulation period is 1600 days.
\end{example}
The results of oil production rate, bottom-hole-pressure and water rate are shown in
Fig. \ref{fig-mxfrr3-revised-oil-rate}, \ref{fig-mxfrr3-revised-bhp}, \ref{fig-mxfrr3-revised-water-rate}.
Again, these figures show that our simulator match CMG simulator.
\begin{figure}[!ht]
\centering
\includegraphics[width = 0.5\textwidth]{mxfrr003-revised-oil-rate.eps}
\caption{Example \ref{ex2}: Oil production rate (unit: STB/day).}
\label{fig-mxfrr3-revised-oil-rate}
\end{figure}
\begin{figure}[!ht]
\centering
\includegraphics[width = 0.5\textwidth]{mxfrr003-revised-bhp.eps}
\caption{Example \ref{ex2}: Well bottom-hole pressure (pressure unit: psi).}
\label{fig-mxfrr3-revised-bhp}
\end{figure}
\begin{figure}[!ht]
\centering
\includegraphics[width = 0.5\textwidth]{mxfrr003-revised-wrate.eps}
\caption{Example \ref{ex2}: Water rate (unit: STB/day).}
\label{fig-mxfrr3-revised-water-rate}
\end{figure}
\begin{example}
\label{ex3}
This example tests the scalability of our two-phase dual porosity simulator by computing Example 1 with
grid dimension $500\times 500\times 50$.
\end{example}
As we introducerd before, we employ the GMRES linear solver and CPR preconditioner. Table \ref{tab-scale}
shows the numerical summaries, which show that our numerical methods are stable when increasing CPU cores.
Figure \ref{fig-scale} presents the scalability of our simulator, which demonstrates that the simulator
and parallel implementation are scalable.
\begin{table}[!ht]
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
\#MPIs & 8 & 64 \\
\#Time steps & 68 & 68 \\
\#Newton iterations& 273 & 267 \\
\#total linear iterations & 641& 623 \\
\#total running time(s) &20009.71 & 2531.91 \\
\hline
\end{tabular}\label{tab-scale}
\end{table}
\begin{figure}[!ht]\centering
\includegraphics[width = 0.5\textwidth]{speedup.eps}
\caption{Example \ref{ex3}: speed-up.}
\label{fig-scale}.
\end{figure}
\section{Conclusion}
This paper presents our work on development of two-phase oil-water simulator for natural fractured
reservoirs using dual porosity method. Effective numerical methods and parallel computing techniques
are introduced.
From the numerical experiments, we can see that our results match commercial simulator, CMG IMEX,
and the simulatro has good parallel scalability and is capable of handling large scale reservoir simulation
problems.
\section{Acknowledgements}
This work is partially supported by Department of Chemical Petroleum Engineering,
University of Calgary, NSERC, AIEES, Foundation CMG, AITF iCore, IBM Thomas J. Watson
Research Center, Frank and Sarah Meyer FCMG Collaboration Center,
WestGrid (www.westgrid.ca),
SciNet (www.scinetpc.ca) and Compute Canada Calcul Canada (www.computecanada.ca).
|
\section{Introduction} \label{sec:intro}
Cosmological, hydrodynamical simulations have seen dramatic advances in recent years, producing well-resolved stellar disks with sizes, scale lengths, and stellar masses in reasonable agreement with observations of late-type galaxies \citep[e.g.][]{Schaye2015, Grand2017, Garrison2017, Genel2018}. These simulations have wildly different feedback models, suggesting that galaxy properties may not discriminate between them. The circumgalactic medium (CGM; i.e.\ the gaseous haloes around galaxies) will enable us to study how galaxies' gas reservoirs are replenished and how galactic winds alter the galaxies and their environments. Studying this regime can also help test theoretical models by comparing them to available observations of the CGM \citep[e.g.][]{Putman2012, Tumlinson2017}.
It is often assumed that the hydrodynamical processes outside the galaxy's interstellar medium (ISM) are well-resolved. However, the resolution in state-of-the-art simulations is generally adaptive in a (quasi-)Lagrangian sense, such that the mass resolution is kept fixed. The spatial resolution therefore drops quickly with galactocentric radius in the CGM, in lock-step with the much lower densities there. Worryingly, many idealized studies (which give up the cosmological context in favour of much higher spatial resolution) found that the properties of `halo gas' change considerably with improved resolution \citep[e.g.][]{Scannapieco2015, Schneider2017, Mandelker2018, McCourt2018, Sparre2018}. This suggests that the CGM in cosmological simulations is under-resolved, which could affect our ability to predict or reproduce CGM observables and also have major consequences for the amount of gas accretion on to galaxies and for the mixing of metals.
Questions surrounding the physical and observable properties of the CGM, the role of feedback, the re-accretion of previously expelled gas, and the distribution of heavy elements would greatly benefit from simulations that offer higher spatial resolution in the CGM than permitted by standard simulation techniques. We therefore present zoom-in simulations centred on a Milky Way-mass galaxy taken from the `Auriga' project \citep{Grand2017}, resimulated with fixed spatial resolution within the gaseous haloes of all galaxies within the zoom-in region. A similar approach was taken by \citet{Miniati2014} to study turbulence in a merging galaxy cluster at fixed spatial resolution, achieving 10~kpc uniform spatial resolution in the intracluster medium.
In this letter, we present properties of the halo gas around a simulated Milky Way analogue with direct implications for observations of atomic hydrogen (also called ``neutral hydrogen'' or ``H\,\textsc{i}'') around the Milky Way. The new simulation refinement method is described in Section~\ref{sec:sim}. In Section~\ref{sec:results} we present our results on the density and H\,\textsc{i} structure in the CGM and we conclude in Section~\ref{sec:concl}.
\section{Method} \label{sec:sim}
This work is an extension of the Auriga project \citep{Grand2017}, which consists of a large number of zoom-in magnetohydrodynamical, cosmological simulations of isolated Milky Way-mass galaxies and their environments. These simulations produce realistic disc-dominated galaxies with stellar masses, galaxy sizes, rotation curves, star formation rates, and metallicities in agreement with observations. The simulations were carried out with the quasi-Lagrangian moving-mesh code \textsc{arepo} \citep{Springel2010}. This original suite employs a fixed mass resolution, which means that mesh cells within the zoom-in region are refined (i.e.\ split in two) or derefined (i.e.\ merged with their neighbours) if their mass differs by more than a factor of two from the target mass resolution.
We resimulate, using \textsc{arepo}, one of the Auriga galaxies (halo 6) at standard mass resolution, i.e.\ the target cell mass for baryons is $5.4\times10^4$~M$_{\astrosun}$ and dark matter particle masses are $2.9\times10^5$~M$_{\astrosun}$. In two other simulations, we impose an additional refinement criterion based on the physical cell volume. In this case, a cell is also refined when its volume is more than twice the target volume resolution and only derefined if its mass \emph{and} volume are less than half the target resolution. In the ISM of the galaxy, the density is high, which means that the cells are smaller than the imposed spatial resolution. There is thus no change in the resolution of the galaxy itself, but the resolution is greatly improved in the lower density CGM.
In order to select the region that is spatially refined, we run \textsc{subfind} \citep{Springel2001, Dolag2009} on-the-fly. Dark matter haloes are identified using a Friends-of-Friends (FoF) algorithm with a linking length $b=0.2$. Each spherical region of radius $1.2R_\mathrm{vir}$ centred on every FoF halo with $M_\mathrm{halo}>10^{8.7}$~M$_\odot$ is tagged for additional spatial refinement. In this work, we define the virial radius, $R_\mathrm{vir}$, as the radius within which the mean overdensity is 200 times the mean density of the Universe at its redshift.
We create a \textsc{subfind} catalogue at each of the 128 output redshifts between $z=47$ and $z=0$, with time intervals of $4-204$~Myr, and `dye' cells within $1.2R_\mathrm{vir}$ of sufficiently massive haloes with a passive scalar of value unity. As the simulation continues between two consecutive outputs, the scalar dye is advected with the fluid and used as a marker for the CGM of the galaxies. At each computational time-step, a cell is spatially refined (as described in \citealt{Springel2010}) if its refinement scalar is larger than 0.9, ensuring that the virial regions remain well-resolved until the dye is reinitialized at the next output redshift.
The three simulations presented here have
\begin{enumerate}
\item standard mass refinement, i.e.\ all cells contain approximately $5.4\times10^4$~M$_{\astrosun}$ of gas;
\item standard mass refinement plus fixed spatial refinement where $\rho<\rho_\mathrm{threshold}/512$, resulting in gas cell volumes of approximately 8~proper kpc$^3$ or spatial resolution $\approx2$~kpc;
\item standard mass refinement plus fixed spatial refinement where $\rho<\rho_\mathrm{threshold}/64$, resulting in gas cell volumes of approximately 1~proper kpc$^3$ or spatial resolution $\approx1$~kpc,
\end{enumerate}
where $\rho_\mathrm{threshold}$ is the gas density above which star formation occurs (i.e.\ $n_\mathrm{H}^\star=0.11$~cm$^{-3}$).
Our simulations use the same physical model as the original Auriga suite (see \citealt{Grand2017}). Besides the additional spatial refinement, the only other significant difference with the original Auriga project is that we inject mass and metals ejected by a star particle only into its host cell, rather than into its 64 neighbours. However, the global galaxy properties are not affected by this choice.
In this letter, we are mainly interested in properties of the CGM and we focus on the most common element: hydrogen. The neutral hydrogen fraction is calculated on-the-fly, using the ionization and recombination rates from \citet{Katz1996} and the ionizing ultra-violet (UV) background from \citet{Faucher2009}. Self-shielding from the UV background is implemented using the fits to the self-shielding fraction in radiative transfer simulations by \citet{Rahmati2013}. Gas above $\rho_\mathrm{threshold}$ is put on an artificial equation of state, because the simulations cannot resolve the multi-phase nature of the ISM. The density of the star-forming gas should therefore be treated with caution.
\section{Results}\label{sec:results}
\begin{table}
\begin{center}
\caption{\label{tab:prop} \small Properties of the galaxy and halo in our simulations at $z=0$: simulation refinement strategy, total stellar mass within 30~kpc from the centre ($M_\mathrm{star}$), total ISM mass within 30~kpc from the centre ($M_\mathrm{ISM}$), total CGM mass ($M_\mathrm{CGM}$), total H\,\textsc{i} mass in the CGM ($M_\mathrm{CGM}^{\rm H\,\textsc{i}}$), number of gas cells in the CGM ($N_\mathrm{CGM}^\mathrm{cell}$).}
\vspace{-4mm}
\begin{tabular}[t]{rrrrrr}
\hline \\[-3mm]
simulation & $M_\mathrm{star}$ & $M_\mathrm{ISM}$ & $M_\mathrm{CGM}$ & $M_\mathrm{CGM}^{\rm H\,\textsc{i}}$ & $N_\mathrm{CGM}^\mathrm{cell}$ \\
refinement & (M$_{\astrosun}$) & (M$_{\astrosun}$) & (M$_{\astrosun}$) & (M$_{\astrosun}$) & \\
\hline \\[-4mm]
mass only & $10^{10.73}$ & $10^{9.48}$ & $10^{10.97}$ & $10^{10.26}$ & 1.6M \\
+ 2~kpc & $10^{10.72}$ & $10^{9.41}$ & $10^{10.91}$ & $10^{10.24}$ & 16.3M \\
+ 1~kpc & $10^{10.67}$ & $10^{9.52}$ & $10^{10.96}$ & $10^{10.29}$ & 132.4M \\[-1mm]
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure*}
\center
\includegraphics[scale=0.65]{figures/imagesxz_los600_pix0p5los0p1_127.pdf}
\caption {\label{fig:img} $200\times200$~kpc$^2$ images of the gas in and around a Milky Way-mass galaxy at $z=0$ in a simulation with only mass refinement (left-hand panels), a simulation with additional spatial refinement of $\approx2$~kpc (middle panels), and a simulation with additional spatial refinement of $\approx1$~kpc (right-hand panels). Top panels: an infinitesimally thin slice of the hydrogen number density. Middle panels: total hydrogen column density in a 600~kpc column. Bottom panels: neutral hydrogen column density in a 600~kpc column. Higher resolution in the halo results in more small-scale structure, including more dense gas clumps and thin filaments. The covering fraction of high-column density H\,\textsc{i} is highest in the 1~kpc simulation.}
\vspace{-3mm}
\end{figure*}
Gas accretes on to galaxies from the intergalactic medium after it has passed through the CGM. A better resolved CGM may exhibit different properties, which could affect cooling and thus the growth of the central galaxy. We find, however, that the global properties of this galaxy and its halo remain unchanged. The virial radius is approximately 337~kpc and varies by less than 0.5~per cent. Table~\ref{tab:prop} lists some of the global properties of the galaxy and its CGM, which is defined as all gas cells within $R_\mathrm{vir}$ with $\rho<\rho_\mathrm{thresh}$. The ISM of a galaxy is defined as all star-forming gas (i.e.\ $\rho>\rho_\mathrm{thresh}$) within 30~kpc of its centre. The CGM values include gas associated with satellites, while excluding their ISM, but this choice does not affect our conclusions.
The stellar mass and ISM mass vary by only 0.06~dex and 0.11~dex, respectively, which is consistent with being due to chaotic behaviour \citep{Genel2019}. The bulge-to-disc ratio varies by just 1 per cent. The total mass in the CGM and the H\,\textsc{i} mass in the CGM also remain approximately the same when the resolution in the CGM is enhanced. However, these are all dominated by the dense gas in the central regions, where the spatial refinement is not in effect. H\,\textsc{i} makes up $\approx20$ per cent of the total gas mass in the CGM (and $\approx28$ per cent of the total hydrogen in the CGM).
Compared to the simulation with only mass refinement, the number of resolution elements in the CGM in the 2~kpc and 1~kpc spatially refined simulations is enhanced substantially, by a factor of 10 and 82, respectively. The computational effort, however, only increases by a factor of 2~and~8, respectively. This new method is therefore cheap compared to increasing the resolution by using a higher mass resolution.
Simulations of this type are well suited for high-resolution studies of the gas in galaxy haloes and potentially in other regimes. When adding spatial refinement, most of the additional resolution elements are added to the low-density outer halo. The resolution increase within 150~kpc ($\approx0.45R_\mathrm{vir}$) is only a factor of 3 (19) for the 2~kpc (1~kpc) spatially refined simulations.
Fig.~\ref{fig:img} shows $200\times200$~kpc$^2$ images centred on the Milky Way-mass galaxy in our three simulations. The left-hand panels show a simulation with standard mass refinement ($5\times10^4$~M$_{\astrosun}$), which means that the resolution decreases with decreasing density. The middle and left-hand panels show simulations with the same mass refinement, but additional spatial refinement, resulting in a maximum cell volume of about (2~kpc)$^3$ and (1~kpc)$^3$, respectively.
The top panels show the hydrogen number density in an infinitesimally thin slice through the centre of the halo. The CGM shows significantly more small-scale structure at high resolution, including a larger number of dense clumps and smaller turbulent eddies. The middle panels show the total hydrogen column density integrated over 600~kpc along the line-of-sight. Although the contrast in the CGM is not as large as for the density slice, because the fluctuations average out in projection, the amount of structure on small scales is still clearly enhanced in the spatially refined simulations. The neutral hydrogen column density is shown in the bottom panels. The small-scale structure of dense clumps and thin filaments is much more visible even though the dynamic range spans a much larger range. Also striking is that the average $N_\mathrm{H\,\textsc{i}}$ is higher in the 1~kpc spatial refinement simulation and that the covering fraction of high-column density systems is larger by a substantial amount. We quantify this result below. Some of the high-column density systems result from gas stripped from a satellite flyby, but most of the covering fraction is not associated with such events.
\begin{figure}
\center
\includegraphics[scale=0.45]{figures/radial_los600_median_z0p3to0.pdf}
\caption {\label{fig:rad} Median 2D radial profile for the total hydrogen column density (top panel) and neutral hydrogen column density (bottom panel) at $z=0.3-0$ in our three simulations, described in Section~\ref{sec:sim}. Shaded blue and red regions show the $1\sigma$ scatter for the mass refinement only simulation and for the 1~kpc spatially refined simulation. The bottom panel includes observations of H\,\textsc{i} absorption systems from the COS-Halos survey \citep{Tumlinson2013, Prochaska2017} at $z\approx0.2$, where black circles indicate detections, blue upward triangles indicate lower limits, and red downward triangles indicate upper limits. Regions that are shaded grey could be dominated by ISM gas or project multiple H\,\textsc{i} systems and should therefore be treated with caution. The total hydrogen column density (and thus also the density profile of the CGM) does not depend on the resolution of our simulation. However, the neutral hydrogen column density is much higher in the 1~kpc spatially refined simulation. The $N_\mathrm{H\,\textsc{i}}$ enhancement is largest between 40~and 150~kpc, by up to 1.6~dex.}
\vspace{-3mm}
\end{figure}
In Fig.~\ref{fig:rad}, we quantify the median projected two-dimensional surface density profile for all hydrogen (top panel) and for only neutral hydrogen (bottom panel) in a 600~kpc column. Note that the column density range in the bottom panel is much larger than in the top panel. The solid, blue curves are radial profiles for our standard mass refinement simulation, whereas the dot-dashed, green curves and dashed, red curves are based on simulations with additional spatial refinement of 2~kpc and 1~kpc, respectively. Shaded regions show the 16th and 84th percentiles of the distribution, or the $1\sigma$ scatter, for the lowest and highest resolution simulation.
To limit the impact of satellites and stochasticity due to galactic winds, we show the median of 3 orthogonal projections in the 22 simulation outputs between $z=0.3$ and $z=0$, but this choice does not affect our conclusions.
The central 30~kpc is dominated by the unresolved ISM and is shaded grey, because we cannot trust the column densities within this region. High-column density systems ($N_\mathrm{H\,\textsc{i}}\gtrsim10^{15}$~cm$^{-2}$) are rare, so a 600~kpc sightline will be dominated by a single neutral hydrogen system in the vast majority of cases. However, because H\,\textsc{i} systems with smaller column densities are so numerous, it becomes likely that a 600~kpc sightline intersects multiple clouds. For a fair comparison to observations in the low-column density regime, it is therefore necessary to create synthetic spectra and isolate individual H\,\textsc{i} systems, which we leave for future work. In Fig.~\ref{fig:rad}, $N_\mathrm{H\,\textsc{i}}<10^{15}$~cm$^{-2}$ is shaded grey to indicate that these column densities may be overestimated compared to observations.
The total projected density profile is relatively shallow outside the galaxy, decreasing by one order of magnitude from 30 to 200~kpc. $N_\mathrm{H\,\textsc{i}}$, on the other hand, decreases by 5 orders of magnitude over the same radial range. This means that neutral hydrogen is much more centrally concentrated than the mass. Furthermore, there are only small differences in the median $N_\mathrm{H}$ between the 3 simulations and the scatter is small in all three simulations. This shows that the average CGM density is not affected by resolution effects. However, we calculated the average clumping factor, i.e.\ $\langle\rho^2\rangle / \langle\rho\rangle^2$, at $30-200$~kpc and find a 10~per cent (30~per cent) increase with additional 2~kpc (1~kpc) spatial refinement.
The H\,\textsc{i} column density does not change substantially with additional 2~kpc refinement, probably because the resolution in the central 150~kpc is only slightly higher than the standard simulation with only mass refinement.
However, the median neutral hydrogen column density is much higher in the 1~kpc spatially refined simulation, by up to 1.6~dex. The H\,\textsc{i} column density could potentially increase even further, because the results are not yet converged. The scatter between H\,\textsc{i} sightlines is large in all simulations, especially compared to the scatter in the total hydrogen column, and does not appear to depend significantly on resolution. Interestingly, the 1~kpc spatially refined simulation reproduces the observed high-column density systems better than the standard simulation.
\begin{figure}
\center
\includegraphics[scale=0.45]{figures/HI_fcov_los600_median_z0p3to0.pdf}
\caption {\label{fig:fcov} The covering fraction of neutral hydrogen column density systems above the value indicated on the $x$-axis, within a projected radius of 150~kpc from the central galaxy, between $z=0.3$ and $z=0$. The line styles of the curves are identical to those used in Fig.~\ref{fig:rad}. The vertical dotted, grey line indicates the minimum column density for LLSs ($N_\mathrm{H\,\textsc{i}}\geq10^{17.2}$~cm$^{-2}$) and the black error bars show the covering fractions in the COS-Halos survey \citep{Tumlinson2013, Prochaska2017}. The covering fraction of all systems with $N_\mathrm{H\,\textsc{i}}\lesssim10^{19}$~cm$^{-2}$ is much higher with additional 1~kpc spatial refinement and the covering fraction of LLSs within 150~kpc increases from 18~to 30~per cent.}
\vspace{-3mm}
\end{figure}
Another way to quantify the increase in high-column density systems with resolution is using covering fractions of gas with $N_\mathrm{H\,\textsc{i}}$ above a certain value. Fig.~\ref{fig:fcov} shows the covering fraction of H\,\textsc{i} column densities above the value indicated on the $x$-axis, within a projected radius of 150~kpc of the central galaxy. We use the same simulation outputs, projection axes, and line styles as in Fig.~\ref{fig:rad}.
At the highest column densities, $N_\mathrm{H\,\textsc{i}}\gtrsim10^{20}$~cm$^{-2}$, the simulations exhibit very similar covering fractions. This is due to the fact that these column densities are dominated by dense gas for which the standard mass refinement already reduces the cell size to below 1~kpc, resulting in no difference between the simulations in this regime. However, at lower column densities, $N_\mathrm{H\,\textsc{i}}\lesssim10^{19}$~cm$^{-2}$, the additional spatial refinement becomes important and the covering fraction of these systems is much higher with 1~kpc spatial refinement. An increase in the H\,\textsc{i} covering fraction was also seen in simulations with increasingly better mass refinement, both in zoom-in simulations \citep{Faucher2016} as well as in lower resolution full cosmological volume simulations with a statistical sample of galaxies \citep{Rahmati2015}.
This increase in the covering fraction does not mean the total amount of H\,\textsc{i} is increased substantially, as also demonstrated in Table~\ref{tab:prop}, because most of the H\,\textsc{i} mass is located in large H\,\textsc{i} discs around the galaxies and only roughly 1~per cent of the H\,\textsc{i} is located between 50~kpc and $R_\mathrm{vir}$ (excluding extended discs around satellites).
Lyman-Limit Systems (LLSs) are those with $N_\mathrm{H\,\textsc{i}}>10^{17.2}$~cm$^{-2}$ (indicated by a vertical dotted, grey line). The COS-Halos observations, shown in Fig.~\ref{fig:rad}, exhibit a covering fraction of LLSs of $18-36$ per cent (including uncertainties due to $1\sigma$ errors and upper and lower limits), shown by a black error bar. The covering fraction of LLSs within 150~kpc rises from 18 per cent in the standard mass refined simulation to 30 per cent with additional 1~kpc spatial refinement. Overall, our highest spatial resolution simulation agrees best with the observations, especially at $N_\mathrm{H\,\textsc{i}}>10^{15-16}$~cm$^{-2}$. However, it remains to be seen what the covering fractions would be in a simulation with even higher resolution.
\section{Discussion and conclusions} \label{sec:concl}
We have presented a new refinement technique to simulate the CGM at uniform spatial resolution. This enables us to achieve much higher resolution at relatively low cost. We ran three cosmological simulations of the same Milky Way-mass halo: one with standard mass refinement, one with mass refinement and additional 2~kpc spatial refinement within $1.2R_\mathrm{vir}$, and one with additional 1~kpc spatial refinement. We find that the global properties of the galaxy are not affected by the better sampling of the halo gas. Similarly, the median density profile and its $1\sigma$ scatter are also robust to changes in the resolution of the halo gas. However, we find a large impact on the H\,\textsc{i} column density in the CGM. The median $N_\mathrm{H\,\textsc{i}}$ 2D radial profile is substantially higher (by up to 1.6~dex) in the simulation with additional 1~kpc spatial refinement as compared to the simulation with only mass refinement. As a result, the covering fraction of LLSs within 150 kpc from the galaxy centre increases from 18 to 30 per cent when including 1~kpc spatial refinement. The large reservoirs of cool CGM detected in observations \citep[e.g.][]{Werk2014} arise naturally in cosmological simulations.
However, there are some uncertainties remaining that affect the normalization of our results. The correction due to self-shielding as given in \citet{Rahmati2013} was derived from simulations with much lower resolution and may not be applicable to the small clouds we can resolve here. Therefore, we repeated our calculations in the optically thin regime. This reduces the column densities in $N_\mathrm{H\,\textsc{i}}\gtrsim10^{16}$~cm$^{-2}$ systems and strongly affects the normalization. However, qualitatively our results are unchanged and the simulations with 1~kpc spatial resolution have substantially higher $N_\mathrm{H\,\textsc{i}}$. The H\,\textsc{i} fraction would also be lower across the density range if our simulations used ionization and recombination rates given by \citet{Hui1997}, as done by \citet{Rahmati2013}, instead of following \citet{Katz1996}. This again affects the absolute normalization of our H\,\textsc{i} column density results, but none of the reported relative differences between our simulations.
We conclude that the growth of the central galaxy and the bulk properties of the halo gas are not strongly affected by improved spatial resolution in the CGM. Simulations with only mass refinement therefore seem adequate to capture these aspects of galaxy evolution. However, observables that probe the extremes of the halo gas distribution, such as the low-temperature CGM, are strongly resolution dependent. Cosmological, hydrodynamical simulations with additional spatial refinement in the CGM can help bridge the gap between simulations with a realistic cosmological environment and high-resolution idealized simulations. Our results suggest that future observations with telescopes such as ASKAP, MeerKAT, and SKA are likely to detect more H\,\textsc{i} 21~cm emission than predicted by low-resolution cosmological simulations. In combination with observations, a well-resolved CGM will allow us to test, refine, and potentially rule out feedback models. In future work, we will study the impact of our improved uniform spatial resolution on metal-line emission and absorption in the CGM.
\section*{Acknowledgements}
This work is part of the HITS-Yale Program in Astrophysics (HYPA), sponsored by the Klaus Tschira Foundation. We would like to thank the referee for helpful comments as well as the Simons Foundation and the organizers and participants of the Simons Symposium `Galactic Superwinds: Beyond Phenomenology', for stimulating discussions and inspiration for this work.
FvdB is supported by the National Aeronautics and Space Administration under Grant No. 17-ATP17-0028 issued through the Astrophysics Theory Program and by the US National Science Foundation through grant AST 1516962.
\bibliographystyle{mnras}
|
\section{Introduction}
For a long time, attempts to multi-compartment modeling in brain white matter (WM) with simple single diffusion encodings \cite{fieremans2011white,zhang2012noddi,novikov2018rotationally,reisert2017disentangling}
led to ambiguous results \cite{jelescu2016degeneracy,novikov2018rotationally}.
For example, it was argued in
\cite{fieremans2011white} that intra axonal diffusion is substantially smaller than
extra axonal diffusion along the axons, while others argued for the opposite
\cite{zhang2012noddi,dhital2017absence}.
Multiple diffusion encodings offer
substantially more information than ordinary single diffusion encoding schemes \cite{jespersen2013orientationally,westin2014measurement}.
However, most efforts in understanding the additional information gained by
such methods were focused on dispersed single-compartment systems thus revealing apparent measures like eccentricity, microscopic and fractional anisotropy \cite{jespersen2013orientationally,westin2014measurement,szczepankiewicz2015quantification}.
Recent studies have investigated the benefits of using multiple diffusions encodings to resolve white matter compartmental parameters \cite{lampinen2017neurite, dhital2018}.
For example, spherical diffusion encodings \cite{dhital2017absence} shows very low kurtosis in white and gray matter, which gives rise to the assumption that traces of the tissue compartments are similar.
In \cite{fieremans2018} an additional spherical encodings were used to stabilize fits and release constraints.
Or, in \cite{coelo2017,reisert2018} a combination of linear and planar encodings was used with the same intention.
Thus, the question arises, what kind of protocol is sufficient to solve the problem
uniquely?
This short note contributes to the answer of this question.
We will show that a combination of linear and planar encodings is indeed enough to provide
a unique solution of the full 3-compartment model of brain white matter using $\mathcal{O}(b^2)$ measurements.
The key ingredient of the approach is that a combination of linear and planar measurements provide
a direct estimate of the mesoscopic orientation dispersion, without relying on any other concurrent estimates.
We further discuss an inherent model property, that, under special conditions,
this solution still shows an ambiguity.
Finally, we demonstrate by a few counterexamples the inadequacy of linear and spherical encoding to resolve the problem ( taking only $\mathcal{O}(b^2)$ coefficients).
\section{The White Matter Model}
\label{sec:model}
We follow the standard tissue model as proposed in \cite{novikov2018rotationally,reisert2017disentangling}.
In contrast to \cite{zhang2012noddi}, in this model both intra and extra-axonal compartments undergo the same convolution with the mesoscopic orientation
distribution.
For a general encoding matrix $\mv B$ the signal for this model looks as follows
\begin{eqnarray}
S(\mv B) &=& \int_{\mv n \in S_2} d^2 \mv n\ M(\mv n, \mv B) f(\mv n) \\
&=& \int_{\mv n \in S_2} d^2 \mv n \left( v_i e^{-\tr(\mv B \mv D_i^{\mv n})} + v_e e^{-\tr(\mv B \mv D_e^{\mv n})} + v_\text{f} e^{-\tr(\mv B)D_\text{f}} \right) f(\mv n)
\label{model}
\end{eqnarray}
where $f(\mv n)$ is an arbitrary, normalized orientation distribution function and
$M(\mv n, \mv B)$ the axially symmetric, multi-exponential microstructural model
with symmetry axis $\mv n$. The diffusion tensor of intra- and extra-axonal fractions are
parametrized as
\[
\mv D_i^{\mv n} = \mv n \mv n^T D_i ,\ \mv D_e^{\mv n} = \mv n \mv n^T \Delta_e + \mv I_3 D_e
\]
We now focus on linear encoding $\mv B_\text{lin} = b \mv q \mv q^T$ and planar encoding $\mv B_\text{pla} = b (\mv I_3 - \mv q \mv q^T)/2$, where $\mv q$ is the diffusion
gradient direction of modulus one and the b-value $b$ is defined as the trace of the b-matrix.
Rewriting the microstructural model in terms of the cosine $t = \mv q^T\mv n$ between encoding direction
and axon orientation gives,
\begin{eqnarray}
M_\text{lin}(t,b) &=& v_\text{i} e^{-b D_i t^2} + v_\text{e} e^{-b \Delta_e t^2 - bD_e} + v_\text{f} e^{-bD_\text{f}}
\label{Mlin}\\
M_\text{pla}(t,b) &=& v_\text{i} e^{-b D_i (1-t^2)/2} + v_\text{e} e^{-b \Delta_e (1-t^2)/2 - bD_e} + v_\text{f} e^{-bD_\text{f}}
\label{Mpla}
\end{eqnarray}
In this formulation, the convolution with the mesostructural orientation distribution $f(\mv n)$ takes the form
\begin{eqnarray}
S_\alpha(\mv q,b) &=& \int_{\mv n\in S_2} d^2\mv n \ f(\mv n)\ M_\alpha(\mv q^T\mv n,b)
\label{S=fnMq}
\end{eqnarray}
where $\alpha = \text{linear}$ or $\alpha = \text{planar}$ depending on the gradient waveform.
Note that $S_\alpha(\mv q,b)$ is normalized in the sense $S_\alpha(\mv q,0)=1$.
The key to decouple micro and mesostructural contribution is
to work in the domain of spherical harmonics.
The spherical convolution turns out to be a product of the two spherical harmonic representations, $f_{l,m}$ and $M_\alpha^{l}(b)$, of
$f(\mv n)$ and $M_\alpha(t,b)$, respectively.
\begin{equation}
S_\alpha^{l,m}(b) = f_{l,m}\ M_\alpha^{l}(b) \,.
\label{S=fM}
\end{equation}
We used here in semi-Schmidt normalization\footnote{
In this normalization $\sum_{m=-l}^l |Y_l^m(\mv n)|^2 = 1$ and
$\int d^2 \mv n\, Y_l^m(\mv n) Y_{l'}^{m'}(\mv n)^* = \frac{4\pi}{2l+1}\delta_{l,l'}\delta_{m,m'}$
and $Y_l^0(\mv n) = P_l(\cos\theta)$, where
$Y_l^m$ are the spherical harmonics and $P_l$ the Legendre polynomials and $\theta$ the polar angle of $\mv n$. The axial symmetry implies that the spherical harmonics expansion of $M_\alpha(\mv q^T\mv n,b)$ contains only components with $l=0$, $M_\alpha(\mv q^T\mv n,b)=\sum_l \frac{2l+1}{4\pi}M^l_\alpha(b)P_l(\mv q^T\mv n)$. } as in \cite{reisert2017disentangling}.
The signal is characterized by a set of quantities that are rotationally invariant for any signal-generating tissue.
\begin{equation}
S_\alpha^{l}(b)
= \sqrt{\sum_{m=-l}^l |S_\alpha^{l,m}(b)|^2}
= f_l\ |M_\alpha^{l}(b)|
\label{eq:powers}
\end{equation}
Here $f_l = \sqrt{\sum_m | f_{l,m}|^2}>0$ is the rotation invariant mesoscopic dispersion.
For both linear and planar encodings, we define the moments
\begin{align}
W_\alpha^{l,k} &:= \frac{1}{4\pi}\left. \frac{d^k}{db^k} \right|_{b=0} S_\alpha^{l}(b) \label{W=} \\
&= f_l \,\sgn (M_\alpha^{l}(0)) \left. \int_{-1}^1 \frac{dt}{2}\ P_l(t) \frac{d^k}{db^k} \right|_{b=0} M_\alpha(t,b) \,,
\end{align}
where $\sgn(x)=x/|x|$ and we do not write the delta-functional term for $l\geq 2$, since $M_\alpha^{l}(b)$ have definite signs.
This follows from their physical meaning of the signal from the idealized unidirectional fiber bundle, since diffusion is faster along such a bundle, $M^2_\text{lin}<0$ and $M^2_\text{pla}>0$, for all meaningful constellations of microstructural parameters.
Introduction of these definite signs is sufficient to resolve the ambiguity borne by taking the square of \eq{eq:powers}, which is necessary to build rotation invariant quantities.
Note that the moments defined in \eq{W=} generalize the moments used by
\cite{novikov2018rotationally}; for linear encoding
$W_\text{lin}^{l,k} \propto M^{(2k),l} $ following definitions in
\cite{novikov2018rotationally} equations (15-18).
\subsection{Finding the solution}
The white matter model described above includes one known (the free water diffusivity, $D_\text{f}$) and five unknown scalar parameters: intra-axonal difusivity ($D_i$), extra-axonal radial diffusivity ($D_e$), difference between extra-axonal parallel and radial diffusivity ($\Delta_e$), and the volume fractions ($v_\text{i}$, $v_\text{e}$, $v_\text{f}$) with the constraint $v_i+v_e+v_f=1$. The orientation distribution function $f(\mv n)$ contains an infinite set of coefficients. In this section we show that resolving the signal for both linear and planar encoding up to the order $b^2$ and $l=2$ enables unambiguous determination of the scalar parameters and the first non-trivial coefficient, $f_2$, of $f(\mv n)$.
The corresponding moments are expressed via the model parameters as follows:
\begin{eqnarray}
W_\text{lin}^{0,1}&=& - \frac{1}{3} \Delta_e v_\text{e} - D_e v_\text{e} - \frac{1}{3} D_i v_\text{i} - D_\text{f} v_\text{f} \label{weq1} \\
W_\text{lin}^{2,1}&=& \frac{2}{15} f_2 [\Delta_e v_\text{e} + D_i v_\text{i}] \label{weq2} \\
W_\text{lin}^{0,2}&=& \frac{1}{5} \Delta_e^2 v_\text{e} + D_e^2 v_\text{e} + \frac{1}{5} D_i^2 v_\text{i} + D_\text{f}^2 v_\text{f} + \frac{2}{3} \Delta_e D_e v_\text{e} \label{weq3} \\
W_\text{lin}^{2,2}&=& -f_2 \left[ \frac{4}{35} \Delta_e^2 v_\text{e} + \frac{4}{35} D_i^2 v_\text{i} + \frac{4}{15} \Delta_e D_e v_\text{e} \right] \label{weq4} \\
W_\text{pla}^{0,2}&=& \frac{2}{15} \Delta_e^2 v_\text{e} + D_e^2 v_\text{e} + \frac{2}{15} D_i^2 v_\text{i} + D_\text{f}^2 v_\text{f} + \frac{2}{3} \Delta_e D_e v_\text{e} \label{weq5} \\
W_\text{pla}^{2,2}&=& -f_2 \left[ \frac{4}{105} \Delta_e^2 v_\text{e} + \frac{4}{105} D_i^2 v_\text{i} + \frac{2}{15} \Delta_e D_e v_\text{e} \right] \label{weq6}
\end{eqnarray}
The calculations straightforwardly follow from, \eq{W=}.
Note the absence of the moments $W_\text{pla}^{l,1}$ - in this order (linear in $b$) measurements with any shape of $\mv B$ is equivalent to a set of single-direction measurements and thus do not add any extra information.
For example, the signal obtained using the planar encoding in the $x,y$ plane is equivalent to the mean of signals encoded linearly in the $x$ and $y$ directions.
In particular,
$W_\text{lin}^{0,1} = W_\text{pla}^{0,1} \label{Wlin=Wpla}$ and
$W_\text{lin}^{2,1} = 2 W_\text{pla}^{2,1} \label{Wlin=2Wpla}$,
which can be observed from the fact that $M_\text{pla}(t,b)$, \eq{Mpla}, can be obtained from $M_\text{lin}(t,b)$, \eq{Mlin}, by substituting $t^2$ with $(1-t^2)/2=[P_0(t)-P_2(t)]/3$.
Complimentary information can be found in the second (or higher) order of $b$.
The dispersion parameter, $f_2$, can be easily found from \eq{weq3}-\eq{weq6}
\begin{eqnarray}
f_2 = -\frac{7}{4} \, \frac{W_\text{lin}^{2,2} - 2W_\text{pla}^{2,2}}{ W_\text{lin}^{0,2} - W_\text{pla}^{0,2} }
\label{f2=}
\end{eqnarray}
Note that the denominator is just the average eccentricity
of the compartments \cite{jespersen2013orientationally} (or microstructural fractional anisotropy), namely
$W_\text{lin}^{0,2} - W_\text{pla}^{0,2} =
\Delta_e^2 v_\text{e} + D_i^2 v_\text{i}$.
Finding other parameters is not so straightforward.
Assuming $f_2$ is known, we define a set of auxiliary variables $x_i$ as follows
\begin{alignat}{3}
x_1&= \frac{15}{2 f_2} W_\text{lin}^{2,1} &=\Delta_e v_\text{e} + D_i v_\text{i} \label{x1=} \\
x_2&= W_\text{lin}^{0,2} - W_\text{pla}^{0,2} &=\Delta_e^2 v_\text{e} + D_i^2 v_\text{i} \label{x2=} \\
x_3&= - W_\text{lin}^{0,1} - \frac{5}{2 f_2} W_\text{lin}^{2,1} &=D_e v_\text{e} + D_\text{f} v_\text{f} \label{x3=} \\
x_4&= W_\text{lin}^{0,2} + \frac{1}{4 f_2}W_\text{lin}^{2,2} + \frac{9}{2 f_2} W_\text{pla}^{2,2} &=D_e^2 v_\text{e} + D_\text{f}^2 v_\text{f} \label{x4=} \\
x_5&= \frac{15}{2 f_2} W_\text{lin}^{2,2} - \frac{45}{2 f_2} W_\text{pla}^{2,2} &=\Delta_e D_e v_\text{e} \label{x5=}
\end{alignat}
This system including the constraint on the compartment water fractions
\begin{equation}
v_\text{i}+v_\text{e}+v_\text{f}=1 \label{volfrac}
\end{equation}
defines all scalar parameters.
In the following derivation all parameters are restricted to be strictly positive.
Let's express all unknowns in terms of $v_\text{f}$. From simple algebra applied to \eqs{x3=}{x4=} and then \eq{x5=} we find
\begin{eqnarray}
D_e = \frac{x_4-D_\text{f}^2v_\text{f}}{x_3-D_\text{f}v_\text{f}}\,,\ v_\text{e} = \frac{(x_3-D_\text{f}v_\text{f})^2}{x_4-D_\text{f}^2v_\text{f}}\,,\ \Delta_e = \frac{x_5}{x_3-D_\text{f}v_\text{f}} \label{eq:sol1}
\end{eqnarray}
The intra-axonal parameters are expressed from \eqs{x1=}{x2=},
\begin{eqnarray}
D_i = \frac{x_2-\Delta_e^2v_\text{e}}{x_1-\Delta_ev_\text{e}} = \frac{x_2 x_4 -x_5^2 - D_\text{f}^2 v_\text{f} x_2}{x_1 x_4 - x_3 x_5 + D_\text{f} v_\text{f} x_5 - D_\text{f}^2 v_\text{f} x_1} \label{eq:sol2}
\end{eqnarray}
and
\begin{eqnarray}
v_\text{i} = \frac{(x_1-\Delta_ev_\text{e})^2}{x_2-\Delta_e^2v_\text{e}} = \frac{(x_1 x_4 - x_3 x_5 + D_\text{f} v_\text{f} x_5 - D_\text{f}^2 v_\text{f} x_1)^2}{(x_4 - D_\text{f}^2 v_\text{f}) (x_5^2 - x_2 x_4 + D_\text{f}^2 v_\text{f} x_2)} \label{eq:sol3}
\end{eqnarray}
Now, all expressions depend exclusively on the unknown $v_\text{f}$. The last equation $v_\text{i}+v_\text{e}+v_\text{f}=1$ solves for $v_\text{f}$ as follows
\begin{eqnarray}
1-v_\text{f} &=& v_\text{i} + v_\text{e} \\
&=& \frac{D_\text{f}^2 v_\text{f} x_1^2 - x_1^2 x_4 - x_2 x_3^2 - D_\text{f}^2 v_\text{f}^2 x_2 + 2 x_1 x_3 x_5 + 2 D_\text{f} v_\text{f} x_2 x_3 - 2 D_\text{f} v_\text{f} x_1 x_5}{x_5^2 - x_2 x_4 + D_\text{f}^2 v_\text{f} x_2}
\end{eqnarray}
Multiplying both sides by the denominator $(x_5^2 - x_2 x_4 + D_\text{f}^2 v_\text{f} x_2)$ (which is allowed since $x_5^2 - x_2 x_4 + D_\text{f}^2 v_\text{f} x_2 = -D_i^2 D_e^2 v_\text{e} v_\text{i} \neq 0$)
leads to the following equation in $v_\text{f}$
\begin{equation}
(x_2 D_\text{f}^2 - D_\text{f}^2 x_1^2 + 2 D_\text{f} x_1 x_5 - 2 x_2 x_3 D_\text{f} - x_5^2 + x_2 x_4) v_\text{f} + (x_4 x_1^2 - 2 x_1 x_3 x_5 + x_2 x_3^2 + x_5^2 - x_2 x_4) = 0 \label{eq:last}
\end{equation}
This equation is linear, since the quadratic terms in $v_\text{f}$ cancel, which results in the unique final solution
\begin{eqnarray}
v_\text{f} = \frac{x_2 x_4 -x_2 x_3^2 - x_1^2 x_4 - x_5^2 + 2 x_1 x_3 x_5}{
x_2 x_4 + D_\text{f}^2 x_2 - x_5^2 - D_\text{f}^2 x_1^2 - 2 D_\text{f} x_2 x_3 + 2 D_\text{f} x_1 x_5} \label{eq:sol_vf}
\end{eqnarray}
By inserting this $v_\text{f}$ into \eqss{eq:sol1}{eq:sol3} we obtain the full solution for
all parameters, which is our main result.
We now analyze the case of zero denominator in \eq{eq:sol_vf}, which
will express an ambiguity inherent to the model itself.
Substituting the defining expression for the $x_i$'s, \eqss{x1=}{x5=}, gives for the denominator the form
$v_\text{e} v_\text{i} (D_i D_e - D_i D_\text{f} + \Delta_e D_\text{f})^2$
and the same form multiplied with $v_\text{f}$ for the numerator.
This means that for the special case
\begin{eqnarray}
D_i D_e - D_i D_\text{f} + \Delta_e D_\text{f} = 0 \label{eqspecial}
\end{eqnarray}
there is no information about $v_\text{f}$, since \eq{eq:last} turns into an identity. In other words, the constraint $v_\text{i}+v_\text{e}+v_\text{f}=1$ is automatically fulfilled for any $v_\text{f}$.
Note that the system of \eqss{x1=}{x5=} is linear in the volume fractions.
In particular, given the diffusivities, \eqS{x1=}{x3=}{volfrac} can be used to build a linear system for the volume fractions.
The determinant of the so constructed system is just
the left-hand side of \eq{eqspecial} -- its zero value implies a linear dependency, thus resulting in an infinite number of solutions.
To understand the physics behind the degeneracy condition, \eq{eqspecial}, consider first two special cases. If $D_e=0$, \eq{eqspecial} gives $\Delta_e=D_i$, which means that the extra-axonal compartment is indistinguishable from the intra-axonal one. Another case is the isotropic extra-axonal compartment, $\Delta_e=0$, in which case it is indistinguishable from free water, $D_e=D_\text{f}$. We found a family of solutions that interpolates between
these two special cases, which is shown in Figure \ref{fig:family}, as a functions of $v_\text{f}$.
This solution only exists for
a specific choice of diffusivities obeying \eq{eqspecial}. All the shown solutions
have exactly the same moments up to second order.
Outside the displayed interval, the solution is unphysical with several negative parameters. Interestingly, the intra-axonal diffusivity is not subjected to the ambiguity.
In that case, one can find $D_i = \frac{D_\text{f} x_2}{D_\text{f} x_1 -x_5 }$.
Note the similarity of the above degeneracy to the bi-exponential model when the diffusivities in two compartments are equal. \Eq{eqspecial} expresses this inherent drawback of multi-exponential models exemplified by the standard white matter model.
\begin{figure}[t]
\centering
\includegraphics[width=12cm]{family.pdf}
\caption{An example for a family of solutions where $D_i D_e - D_i D_\text{f} + \Delta_e D_\text{f} = 0$ for varying $v_\text{f}$.
All these solutions have the same linear and planar moments up to order two.
}
\label{fig:family}
\end{figure}
\subsection{Determination of mesoscopic dispersion $f_2$}
\Eq{f2=} operates with the moments of the order $b^2$. Here we show that the dispersion can also be expressed directly in terms of the signal.
Recall that the moments $W_\alpha^{l,k}$ define the Taylor expansion of $S_\alpha^{l}(b)$ in powers of $b$ according to \eq{W=}. Therefore the function
\begin{eqnarray}
F(b) := -\frac{7}{4}\, \frac{S_\text{lin}^{2}(b) - 2 S_\text{pla}^{2}(b)}{S_\text{lin}^{0}(b) - S^0_\text{pla}(b)}
\end{eqnarray}
reproduces \eq{f2=} with account for the identities $W_\text{lin}^{0,1} = W_\text{pla}^{0,1} $ and $W_\text{lin}^{2,1} = 2 W_\text{pla}^{2,1}$. Practically, one has to consider the function
\[
F(b) = f_2 + \mathcal{O}(b) \,
\]
and fit it linearly to find its value for $b=0$.
\subsection{Linear and spherical encodings are not sufficient}
For spherical encoding we have
\begin{eqnarray}
S_\text{sph}(b) &=& v_\text{i} e^{-b D_i/3} + v_\text{e} e^{-b (\Delta_e/3+D_e)} + v_\text{f} e^{-D_\text{f} b} \\
W_\text{sph}^k &=& \left. \frac{d^k}{db^k} \right|_{b=0} S_\text{sph}(b)
\end{eqnarray}
We assume that only moments up to $\mathcal{O}(b^2)$ are observable, i.e.
$W_\text{lin}^{0,1},W_\text{lin}^{0,2},W_\text{lin}^{2,1},W_\text{lin}^{2,2},W^1_\text{sph},W^2_\text{sph}$
are known. We know that $W^1_\text{sph}$ is linearly dependent on $W_\text{lin}^{0,1}$
and $W_\text{lin}^{0,2}$, so linear and spherical encodings give five equations up to order 2.
In fact, with these equations, one can find analytically a solution for the two-compartment
model without the fast water fraction. However, this solution has two roots and is, hence, ambigious.
We do not show here the solutions, but give a few numeric examples, where both roots lead to physical meaningful results:
\begin{center}
\begin{tabular}{l|l|l|l|l|l}
solution & $D_i$ & $\Delta_e$ & $D_e$ & $f_2$ & $v_\text{i}$ \\ \hline
1.a & 2.00 & 0.60 & 0.50 & 0.80 & 0.60 \\
1.b & 2.11 & 1.29 & 0.24 & 0.74 & 0.31 \\ \hline
2.a & 2.00 & 0.60 & 0.50 & 0.80 & 0.40 \\
2.b & 2.17 & 1.11 & 0.31 & 0.72 & 0.17 \\ \hline
3.a & 2.40 & 1.00 & 0.50 & 0.80 & 0.40 \\
3.b & 2.58 & 1.51 & 0.31 & 0.75 & 0.14 \\ \hline
4.a & 2.00 & 0.60 & 0.50 & 0.50 & 0.50 \\
4.b & 2.14 & 1.20 & 0.27 & 0.46 & 0.24 \\ \hline
\end{tabular}
\end{center}
where mainly $v_\text{i},\Delta_e$ and $D_e$ are confused. The parameters $D_i,f_2$ and
$\Delta_e+D_e$
are rather stable. This goes in line with the observation that a spherical
encoding can resolve the ambiguity of the parallel
diffusivities \cite{fieremans2011white,fieremans2018, dhital2017absence}, but still has to struggle with $\Delta_e$, $D_e$
and $v_\text{i}$.
In Figure \ref{fig:example} we show signal courses for the counterexamples.
\begin{figure}[t]
\includegraphics[width=14cm]{counterexample.pdf}
\caption{Signal courses for the four examples, where linear and spherical moments are
identical up to order two. First notable differences appear
above $b=2$. Note that for $S^2_\text{lin}$ differences are enlarged
by a factor of ten.
}
\label{fig:example}
\end{figure}
\section{Conclusion}
We have constructively shown that linear and planar diffusion encodings can fully resolve the three-compartment model of white matter using
data up to the order $\mathcal{O}(b^2)$ and $l=2$.
The common experience with the diffusional kurtosis imaging \cite{jensen2005diffusional} indicates the practical availability of $\mathcal{O}(b^2)$ terms. While in principle, these terms include information for $l \leq 4$, the order $l=4$ is spoiled by noise as it was shown for a typical two-shell measurement on an advanced scanner with the maximal gradient strength $80\units{mT/m}$ \cite[Fig.\,2]{reisert2017disentangling}.
Our analysis highlighted a special situation of ambiguous solution due to an inherent inability of multiexponential models to resolve compartments with similar parameters.
The only way to distinguish such compartments is measuring in a domain where their differences get
apparent, for example in the large b-value regime, where stable estimates of higher order information becomes possible. Without such information, a stable parameter estimate is only possible relying on prior knowledge.
We have also shown that a combination of spherical and linear encoding is not
enough to find a unique solution in order $\mathcal{O}(b^2)$. In fact,
$\mathcal{O}(b^2)$ information delivered by a spherical encoding is
fully contained in the combination of linear and planar information,
namely $W_\text{sph}^{0,2} = (4W_\text{pla}^{0,2} - W_\text{lin}^{0,2})/3$, which
renders a spherical encoding in the presence of linear and planar encodings
in the low b-value regime superfluous.
In fact, it is a fortunate coincidence that $\mathcal{O}(b^2)$ information spanned by
linear and planar diffusion encoding (it is actually the
'full' encoding in $\mathcal{O}(b^2)$) is 6 dimensional (\eqss{weq1}{weq6}) and the parameter space
of the three compartment white matter model has also 6 free parameters, \eq{model}.
The derived mapping is only valid for noiseless signals, i.e., when the
signal is in the image of the modeling equation. For practical applications the obtainable signal-to-noise ratios are too low.
A recent preprint \cite{coelho2018double} shows by numerical
simulations that in a slightly simplified setting (two-compartments and Watson distribution) also in the noisy case the degeneracy is resolved.
The importance of
the analytical solution lies in its justification for parameter estimators that rely on unimodal posterior distributions.
Additionally, the solution can give certain hints for the construction of such parameter estimators. In fact, the expression
of the parameters are all low-order rational functions of the moments (which are all linear projections of the signal).
This suggests to make a similar approach for the estimator (e.g. as found in
\cite{reisert2017disentangling}), i.e.\ using functions, which are rational in linear combinations of the signal.
\bibliographystyle{apalike}
|
\section*{Abstract (Not appropriate in this style!)}%
\else \small
\begin{center}{\bf Abstract\vspace{-.5em}\vspace{\z@}}\end{center}%
\quotation
\fi
}%
}{%
}%
\@ifundefined{endabstract}{\def\endabstract
{\if@twocolumn\else\endquotation\fi}}{}%
\@ifundefined{maketitle}{\def\maketitle#1{}}{}%
\@ifundefined{affiliation}{\def\affiliation#1{}}{}%
\@ifundefined{proof}{\def\proof{\noindent{\bfseries Proof. }}}{}%
\@ifundefined{endproof}{\def\endproof{\mbox{\ \rule{.1in}{.1in}}}}{}%
\@ifundefined{newfield}{\def\newfield#1#2{}}{}%
\@ifundefined{chapter}{\def\chapter#1{\par(Chapter head:)#1\par }%
\newcount\c@chapter}{}%
\@ifundefined{part}{\def\part#1{\par(Part head:)#1\par }}{}%
\@ifundefined{section}{\def\section#1{\par(Section head:)#1\par }}{}%
\@ifundefined{subsection}{\def\subsection#1%
{\par(Subsection head:)#1\par }}{}%
\@ifundefined{subsubsection}{\def\subsubsection#1%
{\par(Subsubsection head:)#1\par }}{}%
\@ifundefined{paragraph}{\def\paragraph#1%
{\par(Subsubsubsection head:)#1\par }}{}%
\@ifundefined{subparagraph}{\def\subparagraph#1%
{\par(Subsubsubsubsection head:)#1\par }}{}%
\@ifundefined{therefore}{\def\therefore{}}{}%
\@ifundefined{backepsilon}{\def\backepsilon{}}{}%
\@ifundefined{yen}{\def\yen{\hbox{\rm\rlap=Y}}}{}%
\@ifundefined{registered}{%
\def\registered{\relax\ifmmode{}\r@gistered
\else$\m@th\r@gistered$\fi}%
\def\r@gistered{^{\ooalign
{\hfil\raise.07ex\hbox{$\scriptstyle\rm\RIfM@\expandafter\text@\else\expandafter\mbox\fi{R}$}\hfil\crcr
\mathhexbox20D}}}}{}%
\@ifundefined{Eth}{\def\Eth{}}{}%
\@ifundefined{eth}{\def\eth{}}{}%
\@ifundefined{Thorn}{\def\Thorn{}}{}%
\@ifundefined{thorn}{\def\thorn{}}{}%
\def\TEXTsymbol#1{\mbox{$#1$}}%
\@ifundefined{degree}{\def\degree{{}^{\circ}}}{}%
\newdimen\theight
\def\Column{%
\vadjust{\setbox\z@=\hbox{\scriptsize\quad\quad tcol}%
\theight=\ht\z@\advance\theight by \dp\z@\advance\theight by \lineskip
\kern -\theight \vbox to \theight{%
\rightline{\rlap{\box\z@}}%
\vss
}%
}%
}%
\def\qed{%
\ifhmode\unskip\nobreak\fi\ifmmode\ifinner\else\hskip5\p@\fi\fi
\hbox{\hskip5\p@\vrule width4\p@ height6\p@ depth1.5\p@\hskip\p@}%
}%
\def\cents{\hbox{\rm\rlap/c}}%
\def\miss{\hbox{\vrule height2\p@ width 2\p@ depth\z@}}%
\def\vvert{\Vert
\def\tcol#1{{\baselineskip=6\p@ \vcenter{#1}} \Column} %
\def\dB{\hbox{{}}
\def\mB#1{\hbox{$#1$}
\def\nB#1{\hbox{#1}
\def\note{$^{\dag}}%
\defLaTeX2e{LaTeX2e}
\def\chkcompat{%
\if@compatibility
\else
\usepackage{latexsym}
\fi
}
\ifx\fmtnameLaTeX2e
\DeclareOldFontCommand{\rm}{\normalfont\rmfamily}{\mathrm}
\DeclareOldFontCommand{\sf}{\normalfont\sffamily}{\mathsf}
\DeclareOldFontCommand{\tt}{\normalfont\ttfamily}{\mathtt}
\DeclareOldFontCommand{\bf}{\normalfont\bfseries}{\mathbf}
\DeclareOldFontCommand{\it}{\normalfont\itshape}{\mathit}
\DeclareOldFontCommand{\sl}{\normalfont\slshape}{\@nomath\sl}
\DeclareOldFontCommand{\sc}{\normalfont\scshape}{\@nomath\sc}
\chkcompat
\fi
\def\alpha{\Greekmath 010B }%
\def\beta{\Greekmath 010C }%
\def\gamma{\Greekmath 010D }%
\def\delta{\Greekmath 010E }%
\def\epsilon{\Greekmath 010F }%
\def\zeta{\Greekmath 0110 }%
\def\eta{\Greekmath 0111 }%
\def\theta{\Greekmath 0112 }%
\def\iota{\Greekmath 0113 }%
\def\kappa{\Greekmath 0114 }%
\def\lambda{\Greekmath 0115 }%
\def\mu{\Greekmath 0116 }%
\def\nu{\Greekmath 0117 }%
\def\xi{\Greekmath 0118 }%
\def\pi{\Greekmath 0119 }%
\def\rho{\Greekmath 011A }%
\def\sigma{\Greekmath 011B }%
\def\tau{\Greekmath 011C }%
\def\upsilon{\Greekmath 011D }%
\def\phi{\Greekmath 011E }%
\def\chi{\Greekmath 011F }%
\def\psi{\Greekmath 0120 }%
\def\omega{\Greekmath 0121 }%
\def\varepsilon{\Greekmath 0122 }%
\def\vartheta{\Greekmath 0123 }%
\def\varpi{\Greekmath 0124 }%
\def\varrho{\Greekmath 0125 }%
\def\varsigma{\Greekmath 0126 }%
\def\varphi{\Greekmath 0127 }%
\def\Greekmath 0272 {\Greekmath 0272 }
\def\FindBoldGroup{%
{\setbox0=\hbox{$\mathbf{x\global\edef\theboldgroup{\the\mathgroup}}$}}%
}
\def\Greekmath#1#2#3#4{%
\if@compatibility
\ifnum\mathgroup=\symbold
\mathchoice{\mbox{\boldmath$\displaystyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\textstyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\scriptstyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\scriptscriptstyle\mathchar"#1#2#3#4$}}%
\else
\mathchar"#1#2#3#
\fi
\else
\FindBoldGroup
\ifnum\mathgroup=\theboldgroup
\mathchoice{\mbox{\boldmath$\displaystyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\textstyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\scriptstyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\scriptscriptstyle\mathchar"#1#2#3#4$}}%
\else
\mathchar"#1#2#3#
\fi
\fi}
\newif\ifGreekBold \GreekBoldfalse
\let\SAVEPBF=\pbf
\def\pbf{\GreekBoldtrue\SAVEPBF}%
\@ifundefined{theorem}{\newtheorem{theorem}{Theorem}}{}
\@ifundefined{lemma}{\newtheorem{lemma}[theorem]{Lemma}}{}
\@ifundefined{corollary}{\newtheorem{corollary}[theorem]{Corollary}}{}
\@ifundefined{conjecture}{\newtheorem{conjecture}[theorem]{Conjecture}}{}
\@ifundefined{proposition}{\newtheorem{proposition}[theorem]{Proposition}}{}
\@ifundefined{axiom}{\newtheorem{axiom}{Axiom}}{}
\@ifundefined{remark}{\newtheorem{remark}{Remark}}{}
\@ifundefined{example}{\newtheorem{example}{Example}}{}
\@ifundefined{exercise}{\newtheorem{exercise}{Exercise}}{}
\@ifundefined{definition}{\newtheorem{definition}{Definition}}{}
\@ifundefined{mathletters}{%
\newcounter{equationnumber}
\def\mathletters{%
\addtocounter{equation}{1}
\edef\@currentlabel{\arabic{equation}}%
\setcounter{equationnumber}{\c@equation}
\setcounter{equation}{0}%
\edef\arabic{equation}{\@currentlabel\noexpand\alph{equation}}%
}
\def\endmathletters{%
\setcounter{equation}{\value{equationnumber}}%
}
}{}
\@ifundefined{BibTeX}{%
\def\BibTeX{{\rm B\kern-.05em{\sc i\kern-.025em b}\kern-.08em
T\kern-.1667em\lower.7ex\hbox{E}\kern-.125emX}}}{}%
\@ifundefined{AmS}%
{\def\AmS{{\protect\usefont{OMS}{cmsy}{m}{n}%
A\kern-.1667em\lower.5ex\hbox{M}\kern-.125emS}}}{}%
\@ifundefined{AmSTeX}{\def\AmSTeX{\protect\AmS-\protect\TeX\@}}{}%
\ifx\ds@amstex\relax
\message{amstex already loaded}\makeatother\endinpu
\else
\@ifpackageloaded{amstex}%
{\message{amstex already loaded}\makeatother\endinput}
{}
\@ifpackageloaded{amsgen}%
{\message{amsgen already loaded}\makeatother\endinput}
{}
\fi
\let\DOTSI\relax
\def\RIfM@{\relax\ifmmode}%
\def\FN@{\futurelet\next}%
\newcount\intno@
\def\iint{\DOTSI\intno@\tw@\FN@\ints@}%
\def\iiint{\DOTSI\intno@\thr@@\FN@\ints@}%
\def\iiiint{\DOTSI\intno@4 \FN@\ints@}%
\def\idotsint{\DOTSI\intno@\z@\FN@\ints@}%
\def\ints@{\findlimits@\ints@@}%
\newif\iflimtoken@
\newif\iflimits@
\def\findlimits@{\limtoken@true\ifx\next\limits\limits@true
\else\ifx\next\nolimits\limits@false\else
\limtoken@false\ifx\ilimits@\nolimits\limits@false\else
\ifinner\limits@false\else\limits@true\fi\fi\fi\fi}%
\def\multint@{\int\ifnum\intno@=\z@\intdots@
\else\intkern@\fi
\ifnum\intno@>\tw@\int\intkern@\fi
\ifnum\intno@>\thr@@\int\intkern@\fi
\int
\def\multintlimits@{\intop\ifnum\intno@=\z@\intdots@\else\intkern@\fi
\ifnum\intno@>\tw@\intop\intkern@\fi
\ifnum\intno@>\thr@@\intop\intkern@\fi\intop}%
\def\intic@{%
\mathchoice{\hskip.5em}{\hskip.4em}{\hskip.4em}{\hskip.4em}}%
\def\negintic@{\mathchoice
{\hskip-.5em}{\hskip-.4em}{\hskip-.4em}{\hskip-.4em}}%
\def\ints@@{\iflimtoken@
\def\ints@@@{\iflimits@\negintic@
\mathop{\intic@\multintlimits@}\limits
\else\multint@\nolimits\fi
\eat@
\else
\def\ints@@@{\iflimits@\negintic@
\mathop{\intic@\multintlimits@}\limits\else
\multint@\nolimits\fi}\fi\ints@@@}%
\def\intkern@{\mathchoice{\!\!\!}{\!\!}{\!\!}{\!\!}}%
\def\plaincdots@{\mathinner{\cdotp\cdotp\cdotp}}%
\def\intdots@{\mathchoice{\plaincdots@}%
{{\cdotp}\mkern1.5mu{\cdotp}\mkern1.5mu{\cdotp}}%
{{\cdotp}\mkern1mu{\cdotp}\mkern1mu{\cdotp}}%
{{\cdotp}\mkern1mu{\cdotp}\mkern1mu{\cdotp}}}%
\def\RIfM@{\relax\protect\ifmmode}
\def\RIfM@\expandafter\text@\else\expandafter\mbox\fi{\RIfM@\expandafter\RIfM@\expandafter\text@\else\expandafter\mbox\fi@\else\expandafter\mbox\fi}
\let\nfss@text\RIfM@\expandafter\text@\else\expandafter\mbox\fi
\def\RIfM@\expandafter\text@\else\expandafter\mbox\fi@#1{\mathchoice
{\textdef@\displaystyle\f@size{#1}}%
{\textdef@\textstyle\tf@size{\firstchoice@false #1}}%
{\textdef@\textstyle\sf@size{\firstchoice@false #1}}%
{\textdef@\textstyle \ssf@size{\firstchoice@false #1}}%
\glb@settings}
\def\textdef@#1#2#3{\hbox{{%
\everymath{#1}%
\let\f@size#2\selectfont
#3}}}
\newif\iffirstchoice@
\firstchoice@true
\def\Let@{\relax\iffalse{\fi\let\\=\cr\iffalse}\fi}%
\def\vspace@{\def\vspace##1{\crcr\noalign{\vskip##1\relax}}}%
\def\multilimits@{\bgroup\vspace@\Let@
\baselineskip\fontdimen10 \scriptfont\tw@
\advance\baselineskip\fontdimen12 \scriptfont\tw@
\lineskip\thr@@\fontdimen8 \scriptfont\thr@@
\lineskiplimit\lineskip
\vbox\bgroup\ialign\bgroup\hfil$\m@th\scriptstyle{##}$\hfil\crcr}%
\def\Sb{_\multilimits@}%
\def\endSb{\crcr\egroup\egroup\egroup}%
\def\Sp{^\multilimits@}%
\let\endSp\endSb
\newdimen\ex@
\[email protected]
\def\rightarrowfill@#1{$#1\m@th\mathord-\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\mathord-\mkern-2mu$}\hfill
\mkern-6mu\mathord\rightarrow$}%
\def\leftarrowfill@#1{$#1\m@th\mathord\leftarrow\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\mathord-\mkern-2mu$}\hfill\mkern-6mu\mathord-$}%
\def\leftrightarrowfill@#1{$#1\m@th\mathord\leftarrow
\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\mathord-\mkern-2mu$}\hfill
\mkern-6mu\mathord\rightarrow$}%
\def\overrightarrow{\mathpalette\overrightarrow@}%
\def\overrightarrow@#1#2{\vbox{\ialign{##\crcr\rightarrowfill@#1\crcr
\noalign{\kern-\ex@\nointerlineskip}$\m@th\hfil#1#2\hfil$\crcr}}}%
\let\overarrow\overrightarrow
\def\overleftarrow{\mathpalette\overleftarrow@}%
\def\overleftarrow@#1#2{\vbox{\ialign{##\crcr\leftarrowfill@#1\crcr
\noalign{\kern-\ex@\nointerlineskip}$\m@th\hfil#1#2\hfil$\crcr}}}%
\def\overleftrightarrow{\mathpalette\overleftrightarrow@}%
\def\overleftrightarrow@#1#2{\vbox{\ialign{##\crcr
\leftrightarrowfill@#1\crcr
\noalign{\kern-\ex@\nointerlineskip}$\m@th\hfil#1#2\hfil$\crcr}}}%
\def\underrightarrow{\mathpalette\underrightarrow@}%
\def\underrightarrow@#1#2{\vtop{\ialign{##\crcr$\m@th\hfil#1#2\hfil
$\crcr\noalign{\nointerlineskip}\rightarrowfill@#1\crcr}}}%
\let\underarrow\underrightarrow
\def\underleftarrow{\mathpalette\underleftarrow@}%
\def\underleftarrow@#1#2{\vtop{\ialign{##\crcr$\m@th\hfil#1#2\hfil
$\crcr\noalign{\nointerlineskip}\leftarrowfill@#1\crcr}}}%
\def\underleftrightarrow{\mathpalette\underleftrightarrow@}%
\def\underleftrightarrow@#1#2{\vtop{\ialign{##\crcr$\m@th
\hfil#1#2\hfil$\crcr
\noalign{\nointerlineskip}\leftrightarrowfill@#1\crcr}}}%
\def\qopnamewl@#1{\mathop{\operator@font#1}\nlimits@}
\let\nlimits@\displaylimits
\def\setboxz@h{\setbox\z@\hbox}
\def\varlim@#1#2{\mathop{\vtop{\ialign{##\crcr
\hfil$#1\m@th\operator@font lim$\hfil\crcr
\noalign{\nointerlineskip}#2#1\crcr
\noalign{\nointerlineskip\kern-\ex@}\crcr}}}}
\def\rightarrowfill@#1{\m@th\setboxz@h{$#1-$}\ht\z@\z@
$#1\copy\z@\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\box\z@\mkern-2mu$}\hfill
\mkern-6mu\mathord\rightarrow$}
\def\leftarrowfill@#1{\m@th\setboxz@h{$#1-$}\ht\z@\z@
$#1\mathord\leftarrow\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\copy\z@\mkern-2mu$}\hfill
\mkern-6mu\box\z@$}
\def\qopnamewl@{proj\,lim}{\qopnamewl@{proj\,lim}}
\def\qopnamewl@{inj\,lim}{\qopnamewl@{inj\,lim}}
\def\mathpalette\varlim@\rightarrowfill@{\mathpalette\varlim@\rightarrowfill@}
\def\mathpalette\varlim@\leftarrowfill@{\mathpalette\varlim@\leftarrowfill@}
\def\mathpalette\varliminf@{}{\mathpalette\mathpalette\varliminf@{}@{}}
\def\mathpalette\varliminf@{}@#1{\mathop{\underline{\vrule\@depth.2\ex@\@width\z@
\hbox{$#1\m@th\operator@font lim$}}}}
\def\mathpalette\varlimsup@{}{\mathpalette\mathpalette\varlimsup@{}@{}}
\def\mathpalette\varlimsup@{}@#1{\mathop{\overline
{\hbox{$#1\m@th\operator@font lim$}}}}
\def\tfrac#1#2{{\textstyle {#1 \over #2}}}%
\def\dfrac#1#2{{\displaystyle {#1 \over #2}}}%
\def\binom#1#2{{#1 \choose #2}}%
\def\tbinom#1#2{{\textstyle {#1 \choose #2}}}%
\def\dbinom#1#2{{\displaystyle {#1 \choose #2}}}%
\def\QATOP#1#2{{#1 \atop #2}}%
\def\QTATOP#1#2{{\textstyle {#1 \atop #2}}}%
\def\QDATOP#1#2{{\displaystyle {#1 \atop #2}}}%
\def\QABOVE#1#2#3{{#2 \above#1 #3}}%
\def\QTABOVE#1#2#3{{\textstyle {#2 \above#1 #3}}}%
\def\QDABOVE#1#2#3{{\displaystyle {#2 \above#1 #3}}}%
\def\QOVERD#1#2#3#4{{#3 \overwithdelims#1#2 #4}}%
\def\QTOVERD#1#2#3#4{{\textstyle {#3 \overwithdelims#1#2 #4}}}%
\def\QDOVERD#1#2#3#4{{\displaystyle {#3 \overwithdelims#1#2 #4}}}%
\def\QATOPD#1#2#3#4{{#3 \atopwithdelims#1#2 #4}}%
\def\QTATOPD#1#2#3#4{{\textstyle {#3 \atopwithdelims#1#2 #4}}}%
\def\QDATOPD#1#2#3#4{{\displaystyle {#3 \atopwithdelims#1#2 #4}}}%
\def\QABOVED#1#2#3#4#5{{#4 \abovewithdelims#1#2#3 #5}}%
\def\QTABOVED#1#2#3#4#5{{\textstyle
{#4 \abovewithdelims#1#2#3 #5}}}%
\def\QDABOVED#1#2#3#4#5{{\displaystyle
{#4 \abovewithdelims#1#2#3 #5}}}%
\def\tint{\mathop{\textstyle \int}}%
\def\tiint{\mathop{\textstyle \iint }}%
\def\tiiint{\mathop{\textstyle \iiint }}%
\def\tiiiint{\mathop{\textstyle \iiiint }}%
\def\tidotsint{\mathop{\textstyle \idotsint }}%
\def\toint{\mathop{\textstyle \oint}}%
\def\tsum{\mathop{\textstyle \sum }}%
\def\tprod{\mathop{\textstyle \prod }}%
\def\tbigcap{\mathop{\textstyle \bigcap }}%
\def\tbigwedge{\mathop{\textstyle \bigwedge }}%
\def\tbigoplus{\mathop{\textstyle \bigoplus }}%
\def\tbigodot{\mathop{\textstyle \bigodot }}%
\def\tbigsqcup{\mathop{\textstyle \bigsqcup }}%
\def\tcoprod{\mathop{\textstyle \coprod }}%
\def\tbigcup{\mathop{\textstyle \bigcup }}%
\def\tbigvee{\mathop{\textstyle \bigvee }}%
\def\tbigotimes{\mathop{\textstyle \bigotimes }}%
\def\tbiguplus{\mathop{\textstyle \biguplus }}%
\def\dint{\mathop{\displaystyle \int}}%
\def\diint{\mathop{\displaystyle \iint }}%
\def\diiint{\mathop{\displaystyle \iiint }}%
\def\diiiint{\mathop{\displaystyle \iiiint }}%
\def\didotsint{\mathop{\displaystyle \idotsint }}%
\def\doint{\mathop{\displaystyle \oint}}%
\def\dsum{\mathop{\displaystyle \sum }}%
\def\dprod{\mathop{\displaystyle \prod }}%
\def\dbigcap{\mathop{\displaystyle \bigcap }}%
\def\dbigwedge{\mathop{\displaystyle \bigwedge }}%
\def\dbigoplus{\mathop{\displaystyle \bigoplus }}%
\def\dbigodot{\mathop{\displaystyle \bigodot }}%
\def\dbigsqcup{\mathop{\displaystyle \bigsqcup }}%
\def\dcoprod{\mathop{\displaystyle \coprod }}%
\def\dbigcup{\mathop{\displaystyle \bigcup }}%
\def\dbigvee{\mathop{\displaystyle \bigvee }}%
\def\dbigotimes{\mathop{\displaystyle \bigotimes }}%
\def\dbiguplus{\mathop{\displaystyle \biguplus }}%
\def\stackunder#1#2{\mathrel{\mathop{#2}\limits_{#1}}}%
\begingroup \catcode `|=0 \catcode `[= 1
\catcode`]=2 \catcode `\{=12 \catcode `\}=12
\catcode`\\=12
|gdef|@alignverbatim#1\end{align}[#1|end[align]]
|gdef|@salignverbatim#1\end{align*}[#1|end[align*]]
|gdef|@alignatverbatim#1\end{alignat}[#1|end[alignat]]
|gdef|@salignatverbatim#1\end{alignat*}[#1|end[alignat*]]
|gdef|@xalignatverbatim#1\end{xalignat}[#1|end[xalignat]]
|gdef|@sxalignatverbatim#1\end{xalignat*}[#1|end[xalignat*]]
|gdef|@gatherverbatim#1\end{gather}[#1|end[gather]]
|gdef|@sgatherverbatim#1\end{gather*}[#1|end[gather*]]
|gdef|@gatherverbatim#1\end{gather}[#1|end[gather]]
|gdef|@sgatherverbatim#1\end{gather*}[#1|end[gather*]]
|gdef|@multilineverbatim#1\end{multiline}[#1|end[multiline]]
|gdef|@smultilineverbatim#1\end{multiline*}[#1|end[multiline*]]
|gdef|@arraxverbatim#1\end{arrax}[#1|end[arrax]]
|gdef|@sarraxverbatim#1\end{arrax*}[#1|end[arrax*]]
|gdef|@tabulaxverbatim#1\end{tabulax}[#1|end[tabulax]]
|gdef|@stabulaxverbatim#1\end{tabulax*}[#1|end[tabulax*]]
|endgroup
\def\align{\@verbatim \frenchspacing\@vobeyspaces \@alignverbatim
You are using the "align" environment in a style in which it is not defined.}
\let\endalign=\endtrivlist
\@namedef{align*}{\@verbatim\@salignverbatim
You are using the "align*" environment in a style in which it is not defined.}
\expandafter\let\csname endalign*\endcsname =\endtrivlist
\def\alignat{\@verbatim \frenchspacing\@vobeyspaces \@alignatverbatim
You are using the "alignat" environment in a style in which it is not defined.}
\let\endalignat=\endtrivlist
\@namedef{alignat*}{\@verbatim\@salignatverbatim
You are using the "alignat*" environment in a style in which it is not defined.}
\expandafter\let\csname endalignat*\endcsname =\endtrivlist
\def\xalignat{\@verbatim \frenchspacing\@vobeyspaces \@xalignatverbatim
You are using the "xalignat" environment in a style in which it is not defined.}
\let\endxalignat=\endtrivlist
\@namedef{xalignat*}{\@verbatim\@sxalignatverbatim
You are using the "xalignat*" environment in a style in which it is not defined.}
\expandafter\let\csname endxalignat*\endcsname =\endtrivlist
\def\gather{\@verbatim \frenchspacing\@vobeyspaces \@gatherverbatim
You are using the "gather" environment in a style in which it is not defined.}
\let\endgather=\endtrivlist
\@namedef{gather*}{\@verbatim\@sgatherverbatim
You are using the "gather*" environment in a style in which it is not defined.}
\expandafter\let\csname endgather*\endcsname =\endtrivlist
\def\multiline{\@verbatim \frenchspacing\@vobeyspaces \@multilineverbatim
You are using the "multiline" environment in a style in which it is not defined.}
\let\endmultiline=\endtrivlist
\@namedef{multiline*}{\@verbatim\@smultilineverbatim
You are using the "multiline*" environment in a style in which it is not defined.}
\expandafter\let\csname endmultiline*\endcsname =\endtrivlist
\def\arrax{\@verbatim \frenchspacing\@vobeyspaces \@arraxverbatim
You are using a type of "array" construct that is only allowed in AmS-LaTeX.}
\let\endarrax=\endtrivlist
\def\tabulax{\@verbatim \frenchspacing\@vobeyspaces \@tabulaxverbatim
You are using a type of "tabular" construct that is only allowed in AmS-LaTeX.}
\let\endtabulax=\endtrivlist
\@namedef{arrax*}{\@verbatim\@sarraxverbatim
You are using a type of "array*" construct that is only allowed in AmS-LaTeX.}
\expandafter\let\csname endarrax*\endcsname =\endtrivlist
\@namedef{tabulax*}{\@verbatim\@stabulaxverbatim
You are using a type of "tabular*" construct that is only allowed in AmS-LaTeX.}
\expandafter\let\csname endtabulax*\endcsname =\endtrivlist
\def\@@eqncr{\let\@tempa\relax
\ifcase\@eqcnt \def\@tempa{& & &}\or \def\@tempa{& &}%
\else \def\@tempa{&}\fi
\@tempa
\if@eqnsw
\iftag@
\@taggnum
\else
\@eqnnum\stepcounter{equation}%
\fi
\fi
\global\@ifnextchar*{\@tagstar}{\@tag}@false
\global\@eqnswtrue
\global\@eqcnt\z@\cr}
\def\endequation{%
\ifmmode\ifinner
\iftag@
\addtocounter{equation}{-1}
$\hfil
\displaywidth\linewidth\@taggnum\egroup \endtrivlist
\global\@ifnextchar*{\@tagstar}{\@tag}@false
\global\@ignoretrue
\else
$\hfil
\displaywidth\linewidth\@eqnnum\egroup \endtrivlist
\global\@ifnextchar*{\@tagstar}{\@tag}@false
\global\@ignoretrue
\fi
\else
\iftag@
\addtocounter{equation}{-1}
\eqno \hbox{\@taggnum}
\global\@ifnextchar*{\@tagstar}{\@tag}@false%
$$\global\@ignoretrue
\else
\eqno \hbox{\@eqnnum
$$\global\@ignoretrue
\fi
\fi\fi
}
\newif\iftag@ \@ifnextchar*{\@tagstar}{\@tag}@false
\def\@ifnextchar*{\@tagstar}{\@tag}{\@ifnextchar*{\@tagstar}{\@tag}}
\def\@tag#1{%
\global\@ifnextchar*{\@tagstar}{\@tag}@true
\global\def\@taggnum{(#1)}}
\def\@tagstar*#1{%
\global\@ifnextchar*{\@tagstar}{\@tag}@true
\global\def\@taggnum{#1
}
\makeatother
\endinput
\section{Introduction}
\IEEEPARstart{A}{rguably}, one of the emerging disruptive technologies of
the 21st century is what the Harvard Business Review \cite{Anderson17} has
recently coined the \textquotedblleft Internet of Flying
Things\textquotedblright , referring to the latest generation of consumer
grade drones or UAVs, capable of carrying imaging, thermal or even
chemical/radiation/biological sensors. Drones are touted to be
transformational for tasks from wildlife monitoring, agricultural
inspection, building inspection, to threat detection, as they have the
potential to dramatically reduce both the time and cost associated with a
traditional manual tasking based on human operators. Realizing this
potential requires equipping UAVs with the ability to carry out missions
autonomously.
In this work, we consider the problem of online path planning for UAV based
localization or tracking of a time-varying number of radio-tagged objects. This
is an important basic problem if UAVs are to be able to autonomously
gather spatial-temporal information about the objects of interest such as
animals in wildlife monitoring \cite{kays2011tracking,thomas2012wildlife,cliff2015online,hoa2017icra},
or safety beacons in search-and-rescue missions \cite{gerasenko2001beacon,Murphy2008}. Signals received by the UAV's on-board radio receiver
are used for the detection and tracking of multiple objects in the region of
interest. However, the radio receiver has a limited range, hence, the UAV---with
limited energy supply---needs to move within range of the mobile objects that
are scattered throughout the region. This is extremely challenging because
neither the exact number nor locations of the objects of interest are
available to the UAV.
Detecting and tracking an unknown and time-varying number of moving objects
in low signal-to-noise ratio (SNR) environments is a challenging problem in
itself. Objects of interest such as wildlife and people tend to switch
between various modes of movements in an unpredictable manner. Constraints
on the transmitters such as cost and battery life mean that signals emitted
from radio-tagged objects have very low power, and become unreliable due to
receiver noise, even when they are within receiving range.
The traditional approach of detection before tracking incurs information loss, and is not
feasible in such low SNR environments. Reducing information loss introduces
far too many false alarms, while reducing the false alarms increases
misdetections and information loss~\cite{lehmann2012recursive}.
Planning the path for a UAV to effectively detect and track multiple objects
in such environments poses additional challenges. Path planning techniques
for tracking a single object are not applicable. Since there are multiple
moving objects appearing and disappearing in the region, following only
certain objects to localize them accurately means that the UAV is likely to
miss many other objects. The important question is: \textit{which objects should the
UAV follow, and for how long before switching to follow other objects or to
search for new objects?} In addition to detection and tracking, the UAV needs
to maintain a safe distance from the objects without exact knowledge of
their locations. For example, in wildlife monitoring, UAV noise will scare
animals away if they move within a close range. We also need to keep in
mind that the UAV itself has limited power supply as well as computing and
communication resources.
Well-known bio-inspired planning algorithms such as genetic algorithm (GA)
and particle swarm optimization (PSO) \cite{Roberge2013} are computationally
expensive and not suitable for online applications. Markov decision process and partially observable Markov decision process (POMDP) are receiving
increasing attention as online planning algorithms over the last few decades
with techniques such as grid-based MDP \cite{Baek2013}, or POMDP with
nominal belief state optimization \cite{ragi2013uav}. Furthermore, at a
conceptual level, the POMDP framework enables direct generalization to
multiple objects via the use of random finite set (RFS) models \cite%
{mahler2007statistical}. Random finite set can be regarded as a special case of point process when the points are not repeated. For more information related to point process theory, please see [30],\cite{moller2003statistical},\cite{daley2007introduction}. This so-called RFS-POMPD is a POMDP with the
information state being the filtering density of the RFS of objects.
RFS-POMDP provides a natural framework that addresses all the challenges of our online UAV path planning problem. Indeed, RFS-POMDP for
multi-object tracking with various information theoretic reward functions
and task-based reward functions have been proposed in \cite%
{ristic2010sensor, ristic2011anote, hoang2014sensor, hoang2015cauchy,
beard2015void} and \cite{Reza0,Reza1, Reza2}, respectively. This framework
accommodates path planning for tracking an unknown and time-varying number
of objects in a conceptually intuitive manner. In addition, RFS constructs
such as the void probabilities facilitate the incorporation of a safe distance
between the UAV and objects (whose exact locations are unknown) into the
POMDP \cite{beard2015void}. However, these algorithms require detection to
be performed before tracking and hence not applicable to our problem due to
the low SNR.
In our earlier work~\cite{hoa2017icra}, we presented a path planning solution for tracking one object at a time, in a high SNR environment with a fixed number of objects. This solution, also based on a detection before tracking formulation, is not applicable to the far more challenging problem of simultaneously tracking an unknown and time-varying number of objects in low SNR.
In this work, we propose an online path planning algorithm for joint
detection and tracking of multiple objects directly from the received radio
signal in low SNR environments.
This is accomplished by formulating it as a POMDP with an RFS-based track-before-detect (TBD) multi-object filter.
TBD methods operate on raw, un-thresholded data \cite{ebenezer2016generalized} and are well-suited for tracking in low SNR environments such as infrared, optical \cite{barniv1985dynamic,tonissen1998maximum,rutten2005recursive,vo2010joint}, and radar \cite{buzzi2005track,buzzi2008track,lehmann2012recursive,dunne2013multiple,papi2015generalized}. However, TBD methods are computationally intensive, and TBD for range-only (received signal strength) tracking has not been developed. One of the main innovations of our solution is to convert the raw signals received by the UAV receiver into time-frequency input measurements for the multi-object TBD filter (using the short time Fourier transform). Such signal representation enables us to derive a separable measurement likelihood function that yields a numerically efficient multi-object TBD filter.
\begin{comment}
TBD methods operate on raw, un-thresholded data \cite{ebenezer2016generalized} and are well-suited for tracking in low SNR environments such as infrared, optical \cite{barniv1985dynamic,tonissen1998maximum,rutten2005recursive,vo2010joint}, and radar \cite{buzzi2005track,buzzi2008track,lehmann2012recursive,dunne2013multiple,papi2015generalized}. However, TBD methods are computationally intensive, and TBD for range-only (received signal strength) tracking has not been developed. To this end, we propose to convert the raw signals received by the UAV into time-frequency input measurements for the multi-object TBD filter (using the short time Fourier transform). Such signal representation enables us to derive a separable measurement likelihood function that yields a numerically efficient multi-object TBD filter.
\end{comment}
In order to accommodate the time-varying modes of movements of the objects, we use a jump Markov system (JMS) to model their dynamics. Further, to maintain a safe distance from the objects, we impose an object avoidance constraint based on the void probability functional in~\cite{beard2015void} for the planning formulation.
The paper is organized as follows. Section \ref{sec_background} provides the necessary background: problem statement, RFS and POMDP. Section \ref{sec_problem_formulation} establishes the track-before-detect measurement model, and its implementation to track multiple radio-tagged objects using POMDP under constraints. Section \ref{sec_simulation_experiments} details numerical results and comparisons with detection-based methods. Section \ref{sec_conclusion} reports concluding remarks.
\section{Background} \label{sec_background}
In this work, we consider the problem of online trajectory planning for a UAV to optimally detect and track an unknown and time-varying number of radio-tagged objects.
Our solution to the problem can be formulated in an RFS-POMDP framework with the multi-object filtering density as the information state. Therefore in the following sections we provide an overview of: i) RFS theory; ii) multi-object filtering using RFS; and iii) the POMDP framework. We start with the problem statement.
\subsection{Problem Statement} \label{sec_problem_statement}
The sensor system under consideration consists of a UAV with antenna elements, and a signal processing module. Following the sensor hardware description in~\cite{hoa2017icra}, we present some of its basic components:
\begin{itemize}
\item UAVs used are commercial, civilian, low cost, and small form factor platform with physical constraints on maximum linear and rotation speeds and onboard battery life.
\item The main payload on a UAV is a directional antenna (\textit{e.g.}, Yagi antenna) to capture radio signals.
\item The signal processing module is a hardware component embodying a software defined radio capable of receiving and processing multiple radio-tag signals simultaneously.
\end{itemize}
The objects of interest are equipped with radio transmitters with on-off-keying signaling with low transmit power settings.
This strategy is commonly used in numerous applications such as very high frequency (VHF) collared tags for wildlife tracking\cite{kays2011tracking,thomas2012wildlife,cliff2015online,hoa2017icra}, or safety beacons for search and rescue missions \cite{gerasenko2001beacon, Murphy2008}. The transmitter design and signaling methods are designed to conserve battery power, reduce the cost of the transmitters, increase the transmitters' lifespan as well as reduce installation and maintenance costs. Such a transmitter usually emits a pulse train of period $T_0$. Within this period, the pulse consists of a truncated sine wave with frequency $f$ over the interval $[\tau,\tau + P_w]$, as illustrated later in Fig.~\ref{fig_received_pulse_stft_illustration}. Low power on-off-keying signals are difficult to detect in noisy environments.
\begin{comment}
\begin{figure}[t]
\centering
\includegraphics[width=6cm]{figures/Illustration_on_off_keying_signal.eps}
\caption{An illustration of an on-off-keying signal with period $T_0$. $P_w$ is the pulse width, $\tau$ is the time-offset and $f$ is the frequency of the sine wave.}
\label{fig_on_off_keying_signal}
\end{figure}
\end{comment}
The objects of interest, \textit{e.g.}, people, wildlife, do not follow very predictable trajectories (such as cars, or planes), and most objects, wildlife, for instance, are afraid of the presence of the UAV in their territories. Consequently, the UAV also needs to maintain a safe distance from objects, although getting close to the objects of interest improves tracking accuracy. Consequently, the received signals from the objects of interest are even harder to detect.
\subsection{Random Finite Set Models}
For notational consistency, we use lowercase letters (%
\textit{e.g. }$x$) for single-object states; capital letters (\textit{e.g. }$X$) represent the multi-object states; bold letters (\textit{e.g. }$\mathbf{x,X}$) are used for labeled states; blackboard letters (\textit{e.g.}, $~\mathbb{X}$) denote state spaces. Let $1_A(\cdot)$ denote the inclusion function of a given set $A$, and $\mathcal{F}(A)$ denote the class of finite subset of $A$. If $X = \{ x\}$, for convenience, write $1_{A}(x)$ instead of $1_{A}(\{x\})$. For simplicity, albeit with a slight abuse of notation, we use the symbol $\Phi(\cdot |\cdot )$ to denote the single-object and multi-object transition kernels, and the symbol $g(\cdot |\cdot )$ to denote the single-object and multi-object measurement likelihood functions.
An RFS $X$ on $\mathbb{X}$ is a random variable taking values in the finite subsets of $\mathbb{X}$. Using Mahler's finite set statistic (FISST), an RFS is fully described by its FISST density. The FISST density is not a probability density \cite{mahler2007statistical}, but it is equivalent to a probability density as shown in \cite{vo2005sequential}. We introduce three common RFS, Bernoulli RFS, multi-Bernoulli RFS and labeled multi-Bernoulli RFS used in our work.
\subsubsection{Bernoulli RFS}
A Bernoulli RFS $X$ on $\mathbb{X}$ has at most one element with probability $r$ for being a singleton distributed over the state space $\mathbb{X}$ according to PDF $p(x)$, and probability $1-r$ for being empty. Its FISST density is defined as follows \cite[p.~351]{mahler2007statistical}:
\begin{align}
\pi(X) =
\begin{cases}
1-r & X= \emptyset, \\ \notag
r\cdot p(x) & X = \{x\}
\end{cases}%
\end{align}
while its cardinality distribution $\rho(n)$ is also a Bernoulli distribution parameterized by $r$.
\subsubsection{Multi-Bernoulli RFS}
is a union of fixed $N$ numbers of independent Bernoulli RFSs: $X =
\bigcup\limits_{i=1}^{N} X^{(i)}$, where $X^{(i)}$ is a random variable on $\mathcal{F}(\mathbb{X}) $ characterized by the existence probability $r^{(i)}$
and probability density $p^{(i)}$ defined on $\mathbb{X}$. Its FISST density is given by~\cite[pp.~368]{mahler2007statistical}:
\begin{align}
\pi(\{ x^{(1)}, \dots,x^{(n)} \}) &= \pi(\emptyset)
\sum\limits_{1 \leq i_1 \neq \dots \neq i_n \leq N} \prod\limits_{j=1}^{n} \nonumber
\dfrac{r^{(i_{j})} \cdot p^{(i_{j})}(x^{(j)})}{1-r^{(i_{j})}},
\end{align}
where $\pi(\emptyset) = \prod\limits_{i=1}^{N}(1-r^{(i)})$, and its
cardinality distribution is also a multi-Bernoulli distribution ~\cite[pp.~369]{mahler2007statistical}:
\begin{align}
\rho(n) = \pi(\emptyset) \sum\limits_{1 \leq i_1 < \dots < i_n \leq N} \notag
\prod\limits_{j=1}^{n} \dfrac{r^{(i_{j})} }{1-r^{(i_{j})}}.
\end{align}
\subsubsection{Labeled Multi-Bernoulli RFS}~\label{sec_LMB_RFS} A labeled RFS with state space $\mathbb{X}$ and label space $\mathbb{L}$ is an RFS on $\mathbb{X} \times \mathbb{L}$ where all realizations of labels are distinct. Similar to the multi-Bernoulli RFS, a labeled multi-Bernoulli (LMB) RFS is completely defined by a parameter set $\{(r^{(\lambda)},p^{(\lambda)}):\lambda \in \Psi\}$ with index set $\Psi$. Its FISST density is given by: \cite{reuter2014lmb}
\begin{align}
\mathbf{\pi}(\mathbf{X}) & =\delta_{|\mathbf{X}|}(\mathbf{|\mathcal{L}(\mathbf{X})|})w(\mathcal{L}(\mathbf{X}))p^{\mathbf{X}}, \notag
\end{align}
where
$\delta$ is the Kronecker delta, $\mathcal{L}(\mathbf{X})$
denotes the set of labels extracted from $\mathbf{X} ~\in \mathcal{F}(\mathbb{X} \times \mathbb{L}) $, $p(\mathbf{x})=p(x,\lambda) = p^{(\lambda)}(x)$, $p^{\mathbf{X}}=\prod_{(x,\lambda)\in\mathbf{X}}p^{(\lambda)}(x)$,
$w(L)\triangleq\prod_{i\in\mathbb{L}}(1-r^{(i)})\prod_{\lambda\in L}\dfrac{1_{\mathbb{L}}(\lambda)r^{(\lambda)}}{(1-r^{(\lambda)})}$.
$\\$
\subsection{Multi-object Filtering Using RFS Theory}
In the FISST approach, the multi-object state at time $k$ is modeled as a (labeled) RFS $\mathbf{X}_k$. The representation of a multi-object state by a finite set
provides consistency with the notion of estimation error distance~\cite{vo2010joint}. Let $z_{1:k}$ denote the history of measurement data from time 1
to $k$. Then using the FISST concept of density and integration, the filtering densities can be propagated using prediction and update steps of the Bayes multi-object filter \cite%
{mahler2007statistical}:
\begin{align}
& \mathbf{\pi} _{k|k-1}(\mathbf{X}_{k}|z_{1:k-1}) \notag
\label{eq_fisst_bayes_filter_predict} \\
& =\int \mathbf{\Phi}_{k|k-1}(\mathbf{X}_{k}|\mathbf{X}_{k-1})\mathbf{\pi}_{k-1}(%
\mathbf{X}_{k-1}|z_{1:k-1})\delta \mathbf{X}_{k-1}, \\
& \mathbf{\pi} _{k}(\mathbf{X}_{k}|z_{1:k})=\dfrac{g(z_{k}|%
\mathbf{X}_{k})\mathbf{\pi}_{k|k-1}(\mathbf{X}_{k}|z_{1:k-1})}{\int %
g(z_{k}|\mathbf{X})\mathbf{\pi}_{k|k-1}(\mathbf{X}|z%
_{1:k-1})\delta \mathbf{X}}, \label{eq_fisst_bayes_filter_update}
\end{align}%
where $\mathbf{\pi}_{k|k-1}(\cdot|z_{1:k-1})$ denotes a multi-object predicted density; $\mathbf{\pi}_{k}(\cdot|z_{1:k})$ denotes a multi-object filtering density; $\mathbf{\Phi}_{k|k-1}(\cdot |\cdot )$ denotes a transition kernel from time $k-1$ to $k$; $g(z_{k}|\cdot )$ denotes a measurement likelihood function at time $k$. Note that the multi-object transition kernel $\mathbf{\Phi}_{k|k-1}(\cdot |\cdot )$ incorporates all of dynamic aspects of objects including death, birth and transition to new states.
The integral is a \textit{set integral} defined for any function $\mathbf{p}:\mathcal{F}(\mathbb{X} \times \mathbb{L}) \rightarrow \mathbb{R}$, given by:
\begin{equation}
\resizebox{0.5\textwidth}{!}{ $\begin{aligned} \label{eq_fisst_definition}
&\int \mathbf{p(X)} \delta \mathbf{X} = \notag\\
&\sum_{n=0}^{\infty}\dfrac{1}{n!} \sum_{(l_1,...,l_n) \in \mathbb{L}^n} \int\limits_{\mathbb{X}^{n}} p(\{(x^{(1)},l^{(1)}), \dots,(x^{(n)},l^{(n)})\}) d(x^{(1)},\dots, x^{(n)})
\end{aligned}$ }
\end{equation}
Generally, the FISST Bayes multi-object recursion is intractable. However, considerable interest in the field has lead to a number of filtering solutions such as the probability hypothesis density (PHD) filter~\cite{mahler2003phd}, the cardinalized PHD (CPHD) filter~\cite{mahler2007cphd}, the multi-object multi-Bernoulli (MeMBer) filter~\cite{mahler2007statistical,vo2009cardinality}, the
generalized labeled multi-Bernoulli (GLMB) filter~\cite{vo2013glmb,vo2014glmb}, and the labeled multi-Bernoulli (LMB) filter~\cite{reuter2014lmb}.
\subsection{Partially Observable Markov Decision Process} \label{sec_bg_POMDP}
POMDP (partially observable Markov decision process) is a theoretical framework for stochastic control problems and is described by the 6-tuple
$\Big[ \mathcal{F}(\mathbb{T}) \times \mathbb{U}, \mathbb{A}, \mathcal{T},\mathcal{R}, \mathcal{F}(\mathbb{Z}), g(\cdot|\cdot) \Big]$ where \cite{monahan1982state,lovejoy1991survey,bertsekas1996dynamic,Hsu2008}
\begin{itemize}
\item $\mathbb{T} = \mathbb{X} \times \mathbb{L}$ is the labeled state space;
\item $\mathcal{F}(\mathbb{T}) \times \mathbb{U}$ is the space where each of its elements is an ordered pair $(\mathbf{X},u)$, with $\mathbf{X}$ is an object state (possibly a multi-object state) and $u$ is an observer state;
\item $\mathbb{A}$: a set of control actions;
\item $\mathcal{T}$: a state-transition function on $\big[\mathcal{F}(\mathbb{T}) \times \mathbb{U}\big] \times \mathbb{A} \times \big[ \mathcal{F}(\mathbb{T}) \times \mathbb{U}\big]$ where $\mathcal{T}((\mathbf{X},u),a,(\mathbf{X'},u'))$ is the probability density of next state $(\mathbf{X'},u')$ given current state $(\mathbf{X},u)$ and action taken $a$ by the observer;
\item $\mathcal{R}$: a real-valued reward function defined on $\mathbb{A}$;
\item $\mathcal{F}(\mathbb{Z})$: a set of observations;
\item $g(\cdot|\cdot)$: an observation likelihood function on $\mathcal{F}(\mathbb{Z}) \times \big[\mathcal{F}(\mathbb{T}) \times \mathbb{U}\big] \times \mathbb{A}$ where $g(z|(\mathbf{X},u),a)$ is the likelihood of an observation
$z$ given the state $(\mathbf{X},u)$, after the observer takes the action $a$.
\end{itemize}
The main goal in a POMDP is to find an optimal action $a^{*}_k$ that generates an optimal trajectory (a sequence of observer's positions) by maximizing the total expected reward over $H$ look-ahead steps. Specifically, the total expected reward is $\mathbb{E}[\sum_{j=1}^{H} \gamma^{j-1}\mathcal{R}_{k+j}(a_k)]$ with a discount factor \(\gamma \in (0,1] \) to modulate the effects of future rewards on current actions, and $\mathbb{E}[\cdot]$ is the expectation operator.
In this work, we propose using an information-based reward function. For the purpose of joint detection and tracking, where reducing overall uncertainty is the main objective, such a reward function is more appropriate because more information implies less uncertainty~\cite{beard2015void}. There are other reward functions, such as cardinality variances~\cite{Reza0}, which are good at estimating the number of objects or the OSPA-based method in~\cite{Reza1} that depends on user-defined threshold values. In contrast, the information-based methods capture overall cardinality and position information, and can be efficiently computed in a closed-form. A detailed comparison between task-based and information-based reward functions can be found in~\cite{kreucher2005comparison}.
Suppose $\mathbf{\pi}_{k+H|k}(\cdot|z_{1:k})$ is the predicted density to time $k+H$ given measurements up to time $k$, which can be calculated recursively by only using the Bayes prediction step in \eqref{eq_fisst_bayes_filter_predict} from time $k$ to $k+H$. Now, suppose $a_{k}$ is the control action applied to the UAV at time $k$; then, the UAV follows a trajectory consisting of a sequence of discrete positions $u_{1:H}(a_k) = [u_{k+1}(a_k),\dots, u_{k+H}(a_k)]^T$ with corresponding hypothesized measurements $z_{1:H}(a_k) = [z_{k+1}(a_k),\dots, z_{k+H}(a_k)]^T$.
Then the filtering density $\mathbf{\pi}_{k+H}(\cdot|z_{1:k}, z_{1:H}(a_{k}))$ can be computed recursively using the Bayes filter in \eqref{eq_fisst_bayes_filter_predict}, \eqref{eq_fisst_bayes_filter_update} from time $k$ to $k+H$. The reward function can be specified in terms of information divergence between the filtering density and the predicted density. The rationale is that a more informative filtering density yields better estimation results. Thus, it is appropriate to choose an optimal policy that generates a more informative filtering density. Since the filtering density is equally or more informative than the predicted density, maximizing the information divergence between the filtering density and the predicted density often results in a more informative filtering density, and consequently, a better tracking performance. In particular, the information-based reward function is given by~\cite{beard2015sensor}:
\begin{equation
\resizebox{0.45\textwidth}{!}{ $\begin{aligned}
\mathcal{R}_{k+H}(a_{k}) = D(\mathbf{\pi}_{k+H}(\cdot|z_{1:k},z_{1:H}(a_{k}),\mathbf{\pi}_{k+H|k}(\cdot|z_{1:k})) , \notag
\end{aligned}$ }
\end{equation}
where $D(\mathbf{\pi}_2,\mathbf{\pi}_1)$ is the information divergence between two FISST densities, $\mathbf{\pi}_2$ and $\mathbf{\pi}_1$. Some information divergence candidates are R\'{e}nyi divergence (including Kullback-Leibler divergence) or Cauchy-Schwarz divergence described below:
\subsubsection{R\'{e}nyi divergence} between any two FISST
densities, $\mathbf{\pi}_2$ and $\mathbf{\pi}_1$, is defined as \cite{ristic2010sensor}:
\begin{align}
D_{\RIfM@\expandafter\text@\else\expandafter\mbox\fi{R\'{e}nyi}}(\mathbf{\pi}_2,\mathbf{\pi}_1) = \dfrac{1}{\alpha-1} \log \int \mathbf{\pi}_2^{\alpha}(\mathbf{X}) \mathbf{\pi}_1^{1-\alpha}(\mathbf{X}) \delta \mathbf{X} \notag
\end{align}
where $\alpha\geq0$ is a parameter which determines the emphasis of the tails of two distributions in the metric. When $\alpha \rightarrow 1$, we obtain the well known Kullback-Leibler (KL) divergence.
\subsubsection{Cauchy-Schwarz divergence} between any two FISST
densities, $\mathbf{\pi}_2$ and $\mathbf{\pi}_1$, is defined as \cite{hoang2015cauchy}:
\begin{equation}
\resizebox{0.5\textwidth}{!}{ $\begin{aligned}
D_{CS}(\mathbf{\pi}_2,\mathbf{\pi}_1) = -\log \begin{pmatrix}
\dfrac{\int K^{|\mathbf{X}|} \mathbf{\pi}_2(\mathbf{X}) \mathbf{\pi}_1(\mathbf{X}) \delta \mathbf{X} }{\sqrt{\int K^{|\mathbf{X}|} \mathbf{\pi}_2^2(\mathbf{X}) \delta \mathbf{X} \int K^{|\mathbf{X}|} \pi_1^2(\mathbf{X}) \delta \mathbf{X}}} \notag
\end{pmatrix}
\end{aligned}$ }
\end{equation}
where $K$ denotes the unit of hyper-volume on $\mathbb{T}$.
\section{Problem Formulation} \label{sec_problem_formulation}
The problem we formulate involves tracking multiple radio-tagged objects of interest. The state of a single object of interest comprises of all of its kinematic state (denoted as $\zeta=[x,s]^T \in \mathbb{R}^4 \times \mathbb{S} $), including its position and velocity $x \in \mathbb{R}^4$, and its unknown dynamic model $s \in \mathbb{S}$ (\textit{e.g.}, wandering, constant velocity). Furthermore, each object of interest transmits an on-off-keying signal, as illustrated later in Fig. \ref{fig_received_pulse_stft_illustration}, with unknown offset time $\tau \in \mathbb{R}^+_{0}$ (a non-negative real number), and an unknown unique frequency index $\lambda \in \mathbb{L} \subset \mathbb{N}$ (a natural number). Thus, the state of a single object of interest is $\mathbf{x} = [\zeta,\tau,\lambda]^T \in \mathbb{T} = \mathbb{X} \times \mathbb{L}$, where $\mathbb{X} \subseteq \mathbb{R}^4 \times \mathbb{S} \times \mathbb{R}^+_{0}$.
We begin with a model of the received radio signal in Section~\ref{sec_meas_model} and derive its separable measurement likelihood function in Section \ref{sec_meas_llm}. We apply our proposed measurement likelihood function to track multiple radio-tagged maneuvering objects in Section \ref{sec_MTT}. We formulate our UAV trajectory planning problem as a POMDP in Section \ref{sec_pomdp_tbd}.
\subsection{Measurement Model} \label{sec_meas_model}
Given a multi-object state $\mathbf{X} \in \mathcal{F}(\mathbb{T})$, each object $\mathbf{x}=[\zeta,\tau,\lambda]^T \in \mathbf{X}$, uniquely identified by frequency index $\lambda$, transmits an on-off-keying signal within a frequency band (\textit{e.g.}, $150-151$ MHz VHF band in Australia for wildlife transmitters \cite{kenward2000manual}) to a directional antenna mounted on an observer.
The receiver model of the observer is illustrated in Fig.~\ref{signal_received_diagram}. Here, a Software Defined Radio (SDR) collects received signals from the antenna and down-converts the received signal $v$ via the Hilbert transform and a mixer to a baseband signal $y$, which is subsequently digitized via an embedded analog-to-digital converter (ADC)~\cite{Ossmann2015}. The digitized signal is then transformed to the time-frequency domain via a short time Fourier transform (STFT) algorithm (Fig. \ref{signal_received_diagram}c). In practice, the following holds for the receiver:
\begin{itemize}
\item The required safety distance between the observer and each object of interest is sufficiently large, so that the transmitted signal can be treated as a far-field signal and the effect of multipath is negligible \cite{hoa2017icra}.
\item The receiver noise $\eta$, which may come from the outside environment or thermal noise generated from electronic devices within the receiver, is narrowband wide-sense-stationary (WSS) Gaussian because the bandwidth $B_w$ is small compared to the center frequency $f_c$, $B_w \ll f_c$ \cite[pp.116]{orfanidis2002electromagnetic}.
\end{itemize}
In the following, we construct a model of the received signals captured by the receiver, and we begin with the antenna model.
$\\$
\textbf{Antenna Model }(Fig. \ref{signal_received_diagram}a): For a single object with state $\mathbf{x}=[\zeta,\tau,\lambda]^T$, the signal $s^{(\mathbf{x})}$ measured at a reference distance $d_0>0$ in the far field region can be modeled as:
\begin{equation}
\begin{aligned} s^{(\mathbf{x})}(t) = \dfrac{A^{(\lambda)}}{d_0^{\kappa}}\cos[2\pi(f_c + f^{(\lambda)})t +
\phi^{(\lambda)}] \textrm{rect}^{T_0}_{P_w}(t - \tau) \notag, \end{aligned}
\end{equation}
where $A^{(\lambda)}, f^{(\lambda)},\phi^{(\lambda)}$ are the signal amplitude, baseband frequency and phase, respectively, corresponding to frequency index $\lambda$ of object $\mathbf{x}$; $\kappa $ is a dimensionless path loss exponent that depends on environment and typically ranges from 2 to 4; $f_c$ is the center frequency of band of interest; where
\begin{equation} \label{eq_rect_def}
\mathrm{rect}^{T_0}_{P_w}(t - \tau) = \sum\limits_{n=-\infty}^{\infty} \mathrm{boxcar}_{\tau}^{\tau+P_w}(t+nT_0)
\end{equation}
is a periodic rectangular pulse train with period $T_0$, pulse width $P_w$; $\mathrm{boxcar}_{a}^{b}(\cdot)$ is a function which is zero over the entire real line except the interval $[a,b]$, where it is equal to unity.
\begin{figure}[!tb]
\centering
\includegraphics[width=0.5\textwidth]{figures/Receiver_diagram.pdf}
\caption{The receiver model. $|X|$ objects transmit on-off-keying analog signals in time domain. These signals are captured by the antenna and subsequently digitized through a software defined radio device, and converted to time-frequency domain measurements using an STFT algorithm.}
\label{signal_received_diagram}
\end{figure}
At the output of directional antenna, the noiseless received signal from a given set $\mathbf{X}$ of objects of interest is modeled as:
\begin{align} \notag
v^{(u)}(t) = \sum_{\mathbf{x} \in \mathbf{X}}v^{(\mathbf{x},u)}(t).
\end{align}
Here, $v^{(\mathbf{x},u)}$ is the individual signal contribution of object with state $\mathbf{x}$ measured by the observer with state $u$, given by \cite{hoa2017icra}:
\begin{align} \label{eq_received_signal_at_antenna_continuous}
&v^{(\mathbf{x},u)}(t) = \\
&\gamma(\zeta,u)\cos[2\pi(f_c + f^{(\lambda)})t +
\psi(\zeta,u)]\mathrm{rect}^{T_0}_{{P_w}}(t -
\tau)\notag,
\end{align}
where
\begin{itemize}
\item $u = [p^u; \theta^u]$ is the observer state which
comprises of its position $p^u$ and heading angle $\theta^u$;
\item $\gamma(\zeta,u) = A^{(\lambda)}G_{r}G_{a}(\zeta,u)(d_{0}/d(p^{\zeta},p^u)
)^{\kappa}$ is the received signal magnitude when distance between the position of object $\mathbf{x}$ ($p^{\zeta}$) and the position of observer $u$ ($p^u$) is $d(p^{\zeta},p^u)$;
\item $G_{r}$ is the receiver gain to amplify the received signal;
\item $G_{a}(\zeta,u)$ is the directional antenna gain that depends on a UAV's heading angle $\theta^u$ and its relative position with respect to the position of object $\mathbf{x}$;
\item $\psi(\zeta,u) = \phi^{(\lambda)} - (f_c +
f^{(\lambda)}) d(p^{\zeta},p^u)/c$ is the received signal
phase, where $c$ is the signal velocity.
\end{itemize}
\begin{remark} \label{remark1}\textnormal{
Notably, the measured signal $v^{(\mathbf{x},u)}$ always depends on the observer state $u$. Hereafter, for notational simplicity, $u$ is suppressed. \textit{e.g.}, $v^{(\mathbf{x})} \triangleq v^{(\mathbf{x},u)}$; $\gamma(\zeta) \triangleq \gamma(\zeta,u) $. }
\end{remark}
\noindent\textbf{Software Defined Radio (SDR)} (Fig. \ref{signal_received_diagram}b):
The received signal $v$ is down-converted from the VHF band to the baseband via the Hilbert transform and the mixer. This down-conversion step implemented on the SDR's hardware components is a linear operation and is presented here for completeness. The baseband signal, $\tilde{v}$, is given by:
\begin{align}
\tilde{v}(t) = \sum\limits_{\mathbf{x}\in \mathbf{X}} \tilde{v}^{(\mathbf{x})}(t),
\end{align}
where
\begin{align} \label{eq_tilde_Gamma_i}
\tilde{v}^{(\mathbf{x})}(t) &\triangleq [v^{(\mathbf{x})}(t) + j [v^{(\mathbf{x})}]^*(t)] e^{-j2\pi f_{c}t} \\
&= \gamma(\zeta)
e^{j\psi(\zeta)} e^{j2\pi f^{(\lambda)} t } \mathrm{rect}^{T_0}_{{P_w}}(t -
\tau); \notag
\end{align}
$j$ is the imaginary unit; $[v^{(\mathbf{x})}]^*$ is the complex conjugate of $v^{(\mathbf{x})}$. Since the received signal is corrupted by receiver noise $\eta \sim \mathcal{N}(\cdot;0,\Sigma_{\eta})$, the total baseband signal $y$ can be written as:
\begin{align} \label{eq_meas_time}
y(t)
&= \sum\limits_{\mathbf{x}\in \mathbf{X}} \tilde{v}^{(\mathbf{x})}(t)+ \eta(t).
\end{align}
This continuous baseband signal $y(\cdot)$ in \eqref{eq_meas_time} is sampled at rate $f_s$ by the ADC component, which generates a discrete-time signal $y[\cdot]$, given by $y[n] \triangleq y(n/f_s)$.
$\\$
\noindent\textbf{Short-Time Fourier Transform} (Fig. \ref{signal_received_diagram}c): The short time Fourier transform (STFT) converts the received signal to a time-frequency measurement.
Since the on-off keying pulse offset time $\tau$ is unknown, we apply STFT to divide the measurement interval into shorter segments of equal length to capture the sinusoidal component of the received signal to estimate $\tau$ from the measurement. Fig.~\ref{fig_received_pulse_stft_illustration}
illustrates how the STFT is implemented over one measurement interval $[t_{k-1}, t_k)$ of a discrete on-off keying signal (the dash line in Fig. \ref{fig_received_pulse_stft_illustration}) with period $T_0$ and pulse width $P_w$.
To capture the characteristic of the entire signal, we choose the $k^{\RIfM@\expandafter\text@\else\expandafter\mbox\fi{th}}$ measurement interval to be $[t_k - T_0, t_k)$ to fully contain one cycle of the periodic pulse train. The discrete-time signal on $[t_k - T_0, t_k)$, at the STFT window frame $m \in \{0,\ldots,M-1\}$, is given by:
\begin{align} \label{eq_discrete_signal}
y_k^{(m)}[n] \triangleq y(t_k - T_0 + mR/f_s + n/f_s),
\end{align}
where $n = \{0,1,\dots, N_w-1\}$.
\begin{figure}[tb]
\centering
\includegraphics[width=0.5\textwidth]{figures/Pulse_with_STFT.pdf}
\caption{Illustration for an on-off-keying discrete-time signal $\tilde{v}^{(\mathbf{x})}[\cdot]$ and a STFT windowing method at the $k^\RIfM@\expandafter\text@\else\expandafter\mbox\fi{th}$ measurement interval $[t_{k-1}, t_{k})$. $R$ is the hop size, $%
N_{w}$ is the window width, $P_{w}$ is the pulse width, $\tau$ is the pulse time offset, $T_{0}$ is the period of the pulse. The STFT window frame is indexed at $mR/f_s$ where $m \in \{0,\dots,M-1\}$, and $M$ is the number of window frames in one measurement interval. $m^{(\tau)} = \lceil \tau f_s/R \rceil $ is the time frame index of the signal transmitted from object $\mathbf{x}$. }
\label{fig_received_pulse_stft_illustration}
\end{figure}
\begin{figure*}[tb]
\centering
\includegraphics[width=1.0\textwidth]{figures/Illustration_two_signals_time_freq.pdf}
\caption{a) Illustration of two on-off-keying signals superpositioned in the time domain: $v(t) = [\cos(40t) + 2\cos(100t)]\mathrm{rect}^{1}_{0.35}(t-0.4)$ at the sampling rate $f_s = 1$~kHz; b) The signals are well-separated in frequency domain when using a 4-term Blackman Harris window with $N_m =8$, where $N_w = 150$ samples and the main-lobe width-in-Hz $N_m f_s/N_w = 53.33$ Hz < $\triangle f = 60$ Hz; c) However, it is not separable when $N_w = 42$ samples where the main-lobe width-in-Hz $N_m f_s/N_w = 190.47$ Hz > $\triangle f = 60$ Hz.}
\label{fig_freq_resolution}
\end{figure*}
We set the hop size $R$ and the STFT window width $N_w$ to meet the following condition,
\begin{align} \label{eq_hop_size_Nw_upper_bound}
1/f^{(\lambda)} \leq N_w < R = P_w f_s /2,
\end{align}
to ensure that the rectangular pulse of signal $\tilde{v}^{(\mathbf{x})}$ in \eqref{eq_tilde_Gamma_i} over the interval $[t_{k}-T_0+\tau,t_{k}-T_0+\tau+P_w)$ contains two non-overlapping STFT window indices, $\{m^{(\tau)},m^{(\tau)}+1\}$ as illustrated in Fig. \ref{fig_received_pulse_stft_illustration}, such that these two STFT windows are only composed of the sinusoidal part of the signal. Thus, the number of window frames in one measurement interval is
\begin{align} \label{eq_window_frames}
M = \lceil 2T_0/P_w\rceil,
\end{align}
where $\lceil \cdot \rceil$ is the ceiling operator.
The corresponding $L$-point STFT of $y^{(m)}_k[\cdot]$ using the windowing function $w[\cdot]$ is:
\begin{align} \label{eq_meas_DTFT}
Y^{(m)}_k[l] &= \sum\limits_{n=0}^{N_w-1}
y^{(m)}_{k}[n]w[n]e^{-j(n+mR)2\pi l/L},
\end{align}
for $l = \{0,1,\dots,L-1\}$ (definitions of different window functions for extracting short-time signal segments and their properties can be found in~\cite{smith2011spectral}).
At the $k^{\RIfM@\expandafter\text@\else\expandafter\mbox\fi{th}}$ measurement interval, let $\mathbf{X}_k$ denote the multi-object state and $\mathbf{x}_k=[\zeta_k,\tau_k,\lambda_k]^T$ be an element of $\mathbf{X}_k$. By substituting \eqref{eq_rect_def}, \eqref{eq_tilde_Gamma_i}, \eqref{eq_meas_time}, \eqref{eq_discrete_signal} into \eqref{eq_meas_DTFT}, and combining with conditions in \eqref{eq_hop_size_Nw_upper_bound}, $Y^{(m)}_k[l]$ can be written in term of signal and noise components as:
\begin{align} \label{eq_stft_out_signal_noise}
Y^{(m)}_k[l] &= \sum_{\mathbf{x}_k \in \mathbf{X}_k} G^{(m,l)}(\mathbf{x}_k) + H^{(m)}_k[l]
\end{align}
where
\begin{equation}
\resizebox{0.5\textwidth}{!}{ $\begin{aligned}
G^{(m,l)}(\mathbf{x}_k)&=
\begin{cases}
\gamma(\zeta_k)e^{j\psi(\zeta_k)}W[l-l^{(\lambda_k)}] & \RIfM@\expandafter\text@\else\expandafter\mbox\fi{if } m \in \{m^{(\tau_k)},m^{(\tau_k)}+1\} \label{eq_intensity_func} \\
0 & \RIfM@\expandafter\text@\else\expandafter\mbox\fi{otherwise,}
\end{cases}\\ \end{aligned} $ }
\end{equation}
\begin{align}
W[m] &= \sum\limits_{n=0}^{N_w-1} w[n] e^{-jn2\pi l/L}, \label{eq_fft_win_def} \\
l^{(\lambda_k)} &= \lfloor L f^{(\lambda_k)}/f_s \rfloor, \label{eq_freq_idx_def1} \\
m^{(\tau_k)} &= \lceil \tau_k f_s/R \rceil, \\
H^{(m)}_k[l] &=\sum\limits_{n=0}^{N_w-1} \eta^{(m)}_{k}[n]w[n] e^{-j(n+mR)2 \pi l/L}, \label{eq_timefreq_noise_frames} \\
\eta^{(m)}_{k}[n] &\triangleq \eta(t_k - T_0 + mR/f_s + n/f_s). \label{eq_stft_of_noise}
\end{align}
Now the measurement data $z_k$ at the $k^{\RIfM@\expandafter\text@\else\expandafter\mbox\fi{th}}$ measurement interval is an $M\times L$ matrix, with each element $z^{(m,l)}_k = |Y^{(m)}_k[l]|$ is the magnitude of $Y^{(m)}_k[l]$ defined in \eqref{eq_stft_out_signal_noise}.
Notably, to increase the estimation accuracy of the number of transmitted signals, we need to reduce the interference among signal signatures in the frequency domain. Let $N_m$ denote the main-lobe width-in-bins, where each windowing function $w[\cdot]$ affects $N_m$ differently, as shown in Table \ref{table_L_scale_window_type} \cite{smith2011spectral}. Denote $\triangle f$ as the minimum frequency separation among all transmitted signals, given by $\triangle f = \min\limits_{i,j \in \{1,\dots,|\mathbf{X}|\}} |f^{(\lambda^i)} - f^{(\lambda^j)}|$ where $i \neq j$. A criterion to ensure resolvability of signal frequencies requires the main-lobe width-in-Hz of signal signatures be well-separated~\cite{smith2011spectral}, as illustrated in Fig. \ref{fig_freq_resolution}b; hence $N_m f_s/N_w \leq \triangle f$, which implies
\begin{align} \label{eq_Nw_lower_bound}
N_w \geq \lceil N_m \dfrac{f_s}{\triangle f} \rceil.
\end{align}
\begin{table}[!tbp]
\centering
\renewcommand{\arraystretch}{1.3}
\caption{Main-lobe width-in-bins $N_m$ for various windowing functions }
\label{table_L_scale_window_type}
\begin{adjustbox}{max width=0.47\textwidth}
\begin{tabular}{L{1.5cm}|C{1.5cm}C{1.2cm}C{1.2cm}C{2cm}}
\hline
\textbf{Windowing Function} & Rectangular & Hamming & Blackman & $B$-term Blackman-Harris \\ \hline
$N_m$ & $2$ & $4$ & $6$ & $2B$ \\ \hline
\end{tabular}
\end{adjustbox}
\end{table}
Next, we derive the measurement likelihood given measurement $z_k$ and the condition in~\eqref{eq_Nw_lower_bound}.
\subsection{Measurement Likelihood Function} \label{sec_meas_llm}
Let $C(\mathbf{x}_k) $ denote the influence region of an object with state $\mathbf{x}_k$, given by:
\begin{align}
C(\mathbf{x}_k) \triangleq \{(m,l): |G^{(m,l)}(\mathbf{x}_k)| >0 \},
\end{align}
where $G^{(m,l)}(\mathbf{x}_k)$ is defined in \eqref{eq_intensity_func}. We have the following proposition:
\begin{proposition} \label{prop_1}
\textit{Given a multi-object $\mathbf{X}_k$, and its corresponding measurement $z_k$ at the $k^{\RIfM@\expandafter\text@\else\expandafter\mbox\fi{th}}$ measurement interval. If the influence region of each object does not overlap, \textit{i.e.},
\begin{align}
C(\mathbf{x}_k)\cap C(\mathbf{x}'_k)=\emptyset ~ \forall~ \mathbf{x}_k,\mathbf{x}'_k \in \mathbf{X}_k,
\end{align}
then the measurement likelihood function is given by: }
\begin{align} \label{eq_separable_llh}
g(z_{k}|\mathbf{X}_{k})\propto
\prod\limits_{\mathbf{x}_k \in \mathbf{X}_k}g_{z_{k}}(\mathbf{x}_{k}),
\end{align}
where
\begin{itemize}
\item $g_{z_{k}}(\mathbf{x}_{k})=\prod\limits_{(m,l)\in C(\mathbf{x}_{k})}\dfrac{\varphi (%
z_{k}^{(m,l)};|G^{(m,l)}(\mathbf{x}_{k})|,\Sigma _{z})}{\phi (%
z_{k}^{(m,l)};\Sigma _{z})};$
\item $\varphi(\cdot;|G^{(m,l)}(\mathbf{x}_{k})|,\Sigma_z)$ is the Ricean distribution with mean $|G^{(m,l)}(\mathbf{x}_{k})|$ and covariance $\Sigma_z$;
\item $\phi(\cdot;\Sigma_z)$ is the Rayleigh distributions with covariance $\Sigma_z$;
\item $\Sigma_z = E_w \Sigma_{\eta}/2$ is the receiver noise covariance in frequency domain;
\item $E_w = \sum\limits_{n=0}^{N_w-1}|w[n]|^2$ is the window energy;
\end{itemize}
\end{proposition}
\indent \textit{Proof:} See the Appendix.
For our particular problem, given a multi-object $\mathbf{X}$, a single object $\mathbf{x} = [\zeta,~\tau,~\lambda]^T \in \mathbf{X}$ is uniquely identified by the unique frequency index $\lambda$.
Furthermore, the condition in \eqref{eq_Nw_lower_bound} ensures negligible interference in the frequency domain between the signals emitted from objects with different $\lambda$. As shown in \cite{Harris1978}, using the $4$-term Blackman Harris window, the side-lobe level is less than $-92$ dB compared to the main-lobe level.
Consequently, for all practical purposes, we can consider that the influence region of each object does not overlap, \textit{i.e.}, $ C(\mathbf{x}) \cap C(\mathbf{x}') = \emptyset ~ \forall~\mathbf{x},\mathbf{x}' \in \mathbf{X}$. Thus, Proposition \ref{prop_1} applies to our measurement model.
\subsection{Multi-object Tracking} \label{sec_MTT}
Tracking an unknown number of objects of interest under noisy measurements is a difficult problem. It is even more challenging when the number of objects of interest may change over time. Due to low power characteristics of signals from radio-tagged objects, the detection-based approaches often fail to detect objects in low signal-to-noise ratio (SNR) environments, especially when objects appear or disappear frequently, which lead to higher tracking errors. Thus, detection based approaches may not be applicable for tracking radio-tagged objects in low SNR environments due to the information loss during the thresholding process to detect objects' signals. On the other hand, the TBD method using raw received signals as measurements, preserves all of the signals' information and has been successfully proven to be an effective filter under low SNR environments in \cite{barniv1985dynamic,tonissen1998maximum,rutten2005recursive,vo2010joint,buzzi2005track,buzzi2008track,lehmann2012recursive,dunne2013multiple, Papi2013,papi2015generalized}.
We propose using the TBD-LMB filter~\cite{papi2015generalized} to track a multiple, unknown and time-varying number of objects.
For our particular problem, the single object state $\mathbf{x} = [\zeta,\tau,\lambda]^T = [\bar{\zeta},\lambda]^T\in \mathbf{X}$ is uniquely identified by $\lambda \in \mathbb{L}$, where $\mathbb{L}$ (assumed to be known)\footnote{In practice, the assumption that $\mathbb{L}$ is known holds; for example, conservation biologists possess a collection of radio-tagged wildlife captured, tagged and released back into the wild. However, $\lambda \in \mathbb{L}$ itself cannot be directly inferred from the measurements, especially under the low signal-to-noise ratio scenarios where existing object signals may or may not be received by the sensor and the sensor also receives interfering measurements (from other users) and thermal noise generated measurement artifacts not originating from any object.} is a discrete label space containing all frequency indices $\lambda$, and $\bar{\zeta}=[\zeta,\tau]^T \in \mathbb{X}$ is the object state without label.
Hence, the multi-object $\mathbf{X}~\in \mathcal{F}(\mathbb{T})$ is in fact a labeled RFS. Our initial prior is an LMB density with label space $\mathbb{L}$ and an LMB birth model with label space $\mathbb{B}$ to accommodate an increase in the label space that can occur during UAV path planning for tracking objects\footnote{Notably, in an application where no new objects are introduced into the system over time, the label space $\mathbb{L}$ remains unchanged and the set of LMB birth parameters as expressed in \eqref{eq_MB_predict_params} vanishes. In a practical application, the birth model can accommodate, for example, newly released wildlife during the operation of a tracking task by a UAV.}. Since we use the LMB birth model, TBD-GLMB filter in \cite{papi2015generalized} reduces to a TBD-LMB filter.
TBD-LMB filter provides a simple and elegant solution for a multi-object tracking approach in a low SNR environment with various tracking uncertainties. However, existing applications of TBD-LMB filters do not make use of a jump Markov system (JMS).
Following \cite{reuter2015multiple}, we apply a JMS to the proposed TBD-LMB filter by augmenting the discrete mode into the state vector: $\zeta = [x,s]^T $, where $x$ is the object position and velocity, $ s \in \mathbb{S} = \{1,2,...,S_0 \}$ is the object dynamic mode, $S_0 \in \mathbb{N}^+$ is a positive natural number. Moreover, the mode variable is modeled as first-order Markov chain with transitional probability $t_{k|k-1}(s_k|s_{k-1})$. Hence, the
state dynamics and measurement likelihood for a single augmented state vector are given by:
\begin{align} \label{eq_jms_tbd_mem_update}
\mathbf{\Phi}_{k|k-1}(\mathbf{x}_{k}|\mathbf{x}_{k-1}) &=
\Phi_{k|k-1}(\bar{\zeta}_{k}|\bar{\zeta}_{k-1}) \delta_{\lambda_{k-1}}(\lambda_k), \notag \\
g_{z_{k}}(\mathbf{x}_{k}) &= g_{z_{k}}(x_{k},\tau_k,\lambda_k) = g_{z_{k}}^{(\lambda_k)}(x_{k},\tau_k)\notag,
\end{align}
where
\begin{align}
\Phi_{k|k-1}(\bar{\zeta}_{k}|\bar{\zeta}_{k-1}) &= \mathcal{N}(x_k;F_{k-1}^{(s_{k-1})}x_{k-1},Q^{(s_{k-1})}) \notag \\
& \times \mathcal{N}(\tau_{k};\tau_{k-1},Q^{(\tau)}) t_{k|k-1}(s_k|s_{k-1}) \nonumber;
\end{align}
$\mathcal{N}(\cdot;\mu,Q)$ denotes a Gaussian density with mean $\mu$ and covariance $Q$; $F_{k-1}^{(s_{k-1})}$ is the single-object dynamic kernel on the discrete mode $s_{k-1}$. The offset time $\tau$ is estimated using a zero mean Gaussian random walk method
with its covariance $Q^{(\tau)} = \sigma^2_{\tau} {T_0^2}$, where $\sigma^2_{\tau}$ is the standard deviation of the time offset noise. The frequency index $\lambda_k \in \mathbb{L}$ is unique and static, thus the transition kernel for $\lambda_k$ is given by:
\begin{align}
\delta_{\lambda_{k-1}}(\lambda_k) =
\begin{cases}
1 & \lambda_k = \lambda_{k-1}, \\
0 & \RIfM@\expandafter\text@\else\expandafter\mbox\fi{otherwise}.%
\end{cases}
\end{align}
\noindent\textbf{LMB Prediction:} At time $k-1$, suppose the filtering density is an LMB RFS described by a parameter set $\mathbf{\pi}_{k-1} = \{r^{(\lambda)}_{k-1},p^{(\lambda)}_{k-1}\}_{\lambda \in \mathbb{L}_{k-1}} $ with state space $\mathbb{X}$ and label space $\mathbb{L}_{k-1}$ (for notational brevity and convenience, $\mathbf{\pi}_{k-1} = \{r^{(\lambda)}_{k-1},p^{(\lambda)}_{k-1}\}_{\lambda \in \mathbb{L}_{k-1}} $ is also used to denote the density of an LMB RFS), and the birth model is also an LMB RFS with a parameter set $\mathbf{\pi}_{B,k} = \{r_{B,k}^{(\lambda)},p_{B,k}^{(\lambda)}\}_{\lambda \in \mathbb{B}_{k}}$ with state space $\mathbb{X}$ and label space $\mathbb{B}_k$, then the predicted multi-object density is also an LMB RFS $\mathbf{\pi}_{k|k-1} = \{r_{k|k-1}^{(\lambda)},p_{k|k-1}^{(\lambda)}\}_{\lambda \in \mathbb{L}_{k|k-1}}$ with state space $\mathbb{X}$ and label space $\mathbb{L}_{k|k-1} = \mathbb{L}_{k-1} \cup \mathbb{B}_k ~(\RIfM@\expandafter\text@\else\expandafter\mbox\fi{with } \mathbb{L}_{k-1} \cap \mathbb{B}_k = \emptyset )$, given by the parameter set~\cite{reuter2014lmb}:
\begin{align} \label{eq_MB_predict_params}
\mathbf{\pi}_{k|k-1} &=
\{r_{E,k|k-1}^{(\lambda)},p_{E,k|k-1}^{(\lambda)}\}_{\lambda \in \mathbb{L}_{k-1}} \cup \{r_{B,k}^{(\lambda)},p_{B,k}^{(\lambda)}\}_{\lambda \in \mathbb{B}_{k}}
\end{align}
where
\begin{align} \label{eq_tbd_member_predict_r}
r_{E,k|k-1}^{(\lambda)} =& r_{k-1}^{(\lambda)}\cdot\langle p_{k-1}^{(\lambda)},p^{(\lambda)}_{S,k} \rangle,
\\
p_{E,k|k-1}^{(\lambda)}(\bar{\zeta}) =& \dfrac{\langle \Phi_{k|k-1}(\bar{\zeta}|\cdot),
p_{k-1}^{(\lambda)} p^{(\lambda)}_{S,k} \rangle} {\langle p_{k-1}^{(\lambda)},p^{(\lambda)}_{S,k} \rangle}, \label{eq_tbd_member_predict_p}
\end{align}
and $\langle \cdot \rangle$ is the inner product calculated on the previous state $%
\bar{\zeta}_{k-1}$, given by:
\begin{align} \label{eq_inner_prod_def}
\langle \alpha, \beta\rangle
= \sum\limits_{s} \int \alpha(x,\tau|s) \beta(x,\tau|s) d(x,\tau).
\end{align}
\noindent\textbf{LMB Update:} Given the predicted LMB parameters $\mathbf{\pi}_{k|k-1} =
\{r_{k|k-1}^{(\lambda)},p_{k|k-1}^{(\lambda)}\}_{\lambda \in \mathbb{L}_{k|k-1}}$ defined in \eqref{eq_MB_predict_params}, and the measurement likelihood function is separable as in \eqref{eq_separable_llh}, then the filtering LMB parameters follows~\cite{vo2010joint}:
\begin{align} \label{eq_MB_update_params}
\mathbf{\pi}_k = \{r_{k}^{(\lambda)},p_{k}^{(\lambda)}\}_{\lambda \in \mathbb{L}_{k|k-1}}
\end{align}
where
\begin{align} \label{eq_MB_update_params_r}
r_{k}^{(\lambda)} &= \dfrac{r_{k|k-1}^{(\lambda)}\langle p_{k|k-1}^{(\lambda)}, g_{z_k}^{(\lambda)}\rangle}{1 - r_{k|k-1}^{(\lambda)}+ r_{k|k-1}^{(\lambda)} \langle
p_{k|k-1}^{(\lambda)}, g^{(\lambda)}_{z_k}\rangle}, \\
p_{k}^{(\lambda)} &= \dfrac{ p_{k|k-1}^{(\lambda)} g_{z_k}^{(\lambda)}}{%
\langle p_{k|k-1}^{(\lambda)}, g_{z_k}^{(\lambda)}\rangle}, \label{eq_MB_update_params_p}
\end{align}
with $\langle \cdot \rangle$ is the inner product on the current state $\bar{\zeta}_k$.
\subsection{Path Planning Under Constraints} \label{sec_pomdp_tbd}
We formulate the online UAV path planning problem for joint detection and tracking as a partially observable Markov decision process (POMDP) which has been proven
as an efficient and optimal technique for trajectory planning problems~\cite{kaelbling1998planning,castanon2008stochastic}. In the POMDP framework, the purpose of path planning is to find the
optimal policy (\textit{e.g. } a sequence of actions) to maximize the total expected
reward \cite{Reza1}. Hence, we first focus on evaluating the reward functions. Second, we incorporate a void constraint to maintain a safe distance between the UAV and objects of interest
\subsubsection{Reward Functions for Path Planning}\label{sec_rewardfunctionsforplaning}
Let $\mathcal{A}_k \in \mathbb{A}$ denote a set of possible control vectors $a_k$ at time $k$. A common approach is to
calculate an optimal action that maximizes the total expected reward over a look ahead horizon $H$~\cite{ristic2010sensor,hoang2014sensor,beard2015void}---see Section~\ref{sec_bg_POMDP}:
\begin{align} \label{eq_reward_argmax}
a^{*}_{k}&=\argmax\limits_{a_{k} \in \mathcal{A}_{k}}
\mathbb{E}\Big[\sum_{j=1}^{H} \gamma^{j-1}\mathcal{R}_{k+j}(a_{k})\Big]
\end{align}
Since an analytical solution for the expectation of \eqref{eq_reward_argmax} is not available in general, two popular alternatives are to use
Monte-Carlo integration \cite{beard2015void,ristic2010sensor} or the predicted ideal measurement set (PIMS) as in \cite{ristic2011anote,hoang2014sensor,Amirali2016Cauchy}. Using PIMS, the computationally lower cost approach, we only generate one ideal future measurement at each measurement interval \cite{hoang2014sensor,Amirali2016Cauchy}. Hence, instead of \eqref{eq_reward_argmax}, the optimal action is defined by:
\begin{align} \label{eq_reward_pims}
a^{*}_{k}&=\argmax\limits_{a_{k} \in \mathcal{A}_{k}} \sum_{j=1}^{H} \gamma^{j-1} \hat{\mathcal{R}}_{k+j}(a_{k}),
\end{align}
where
\begin{equation}
\resizebox{0.42\textwidth}{!}{ $\begin{aligned} \label{eq_reward_sampling_def}
\hat{\mathcal{R}}_{k+j}(a_{k}) = D(\mathbf{\pi}_{k+j}(\cdot|z_{1:k},\hat{z}_{1:j}(a_{k}), \mathbf{\pi}_{k+j|k}(\cdot|z_{1:k})).
\end{aligned} $ }
\end{equation}
In~\eqref{eq_reward_sampling_def}, the predicted density $\mathbf{\pi}_{k+j|k}(\cdot|z_{1:k})$ is calculated by propagating the filtering density $\mathbf{\pi}_k(\cdot|z_{1:k})$ in \eqref{eq_MB_update_params} using the prediction step\footnote{The prediction step generally includes birth, death and object motion. For improving computational time and tractability, we limit this to object motion only as in~\cite{beard2015void}.} in \eqref{eq_tbd_member_predict_r}, \eqref{eq_tbd_member_predict_p} repeatedly, from time $k$ to $k+j$. In contrast, the filtering density $\mathbf{\pi}_{k+j}(\cdot|z_{1:k},\hat{z}_{1:j}(a_{k}))$ is computed recursively by propagating $\mathbf{\pi}_k(\cdot|z_{1:k})$ in \eqref{eq_MB_update_params} from $k$ to $k+j$ using both prediction in \eqref{eq_tbd_member_predict_r}, \eqref{eq_tbd_member_predict_p} and update steps in \eqref{eq_MB_update_params_r}, \eqref{eq_MB_update_params_p} with the ideal measurement $\hat{z}_{1:H}(a_{k})$. The ideal measurement $\hat{z}_{1:j}(a_{k})$ is computed by the following steps \cite{hoang2014sensor}:
\begin{itemize}
\item[\textit{i})] Sampling from the filtering density $\mathbf{\pi}_k(\cdot|z_{1:k})$ in \eqref{eq_MB_update_params};
\item[\textit{ii})] Propagating it to $k+j$ using the prediction step in \eqref{eq_tbd_member_predict_r}, \eqref{eq_tbd_member_predict_p};
\item[\textit{iii})] Calculating the number of objects $\hat{n}_{k+j|k}$ and the estimated multi-object state $\hat{\mathbf{X}}_{k+j|k} = \{\hat{\mathbf{x}}^{(i)}_{k+j|k}\}_{i=1}^{\hat{n}_{k+j|k}}$;
\item[\textit{iv})] Simulating the ideal measurement at $k +j$ based on the measurement model in \eqref{eq_stft_out_signal_noise} with the estimated state $\hat{\mathbf{X}}_{k+j|k}$.
\end{itemize}
\begin{comment}
$\hat{\mathcal{R}}_{k+j}(a_{k+j})$ in \eqref{eq_reward_sampling_def} usually does not have an analytical form, hence, we approximate it using Monte-Carlo sampling
\textbf{Monte-Carlo Approximation: } Since the measurement likelihood function is separable, the predicted density $\mathbf{\pi}_{k|k-1}$ \eqref{eq_MB_predict_params} and the filtering density $\mathbf{\pi}_{k}$ in \eqref{eq_MB_update_params} have the same number of LMB components. Consequently, the number of LMB components for the predicted density $\mathbf{\pi}_{k+j|k}(\cdot|z_{1:k})$ and the filtering density $\mathbf{\pi}_{k+j}(\cdot|z_{1:k},\hat{z}_{k+1:k+j}(a_{k+1:k+j})$ are the same. For notational simplicity, let $\mathbf{\pi}_1 \triangleq \mathbf{\pi}_{k+j|k}(\cdot|z_{1:k}),~\mathbf{\pi}_2 \triangleq \mathbf{\pi}_{k+j}(\cdot|z_{1:k},\hat{z}_{k+1:k+j}(a_{k+1:k+j})$ are two LMB distributions on $\mathbb{X}$ with the same label space $\mathbb{L}$, given by:
\begin{align}
\mathbf{\pi}_1 = \{r_1^{(\lambda)},p_1^{(\lambda)}\}_{\lambda \in \mathbb{L}}; ~ \mathbf{\pi}_2 = \{r_2^{(\lambda)},p_2^{(\lambda)}\}_{\lambda \in \mathbb{L}};
\end{align}
and rewriting $\mathbf{\pi}_{1}$ and $\mathbf{\pi}_{2}$ in terms of LMB densities:
\begin{align}
\mathbf{\pi}_{1}(\mathbf{X}) & =\delta_{|\mathbf{X}|}(\mathbf{|\mathcal{L}(\mathbf{X})|})w_{1}(\mathcal{L}(\mathbf{X}))p_{1}^{\mathbf{X}}\label{eq:pi_1}\\
\mathbf{\pi}_{2}(\mathbf{X}) & =\delta_{|\mathbf{X}|}(|\mathcal{L}(\mathbf{X})|)w_{2}(\mathcal{L}(\mathbf{X}))p_{2}^{\mathbf{X}}.\label{eq:pi_2}
\end{align}
In this work, we use Monte Carlo sampling \cite{Amirali2016Cauchy} to calculate the divergence between two LMB densities. Each $\lambda$ component of $\mathbf{\pi}_{j}$ ($j=1,2$), the continuous density $p_{j}^{(\lambda)}(\cdot)$, is approximated by a probability mass function $\hat{p_{j}}^{(\lambda)}(\cdot)$ using the same set of samples
$\{\bar{\zeta}^{(\lambda,i)}\}_{i=1}^{N_s}$ with different weights $\{\omega_{j}^{(\lambda,i)}\}_{i=1}^{N_s}$:
\begin{equation}
p_{j}^{(\lambda)}(\bar{\zeta})\approx\hat{p}_{j}^{(\lambda)}(\bar{\zeta})=\sum_{i=1}^{N_s}\omega_{j}^{(\lambda,i)}\delta_{\bar{\zeta}^{(\lambda,i)}}(\bar{\zeta}).
\end{equation}
\end{comment}
The number of LMB components for the predicted density $\mathbf{\pi}_{k+j|k}(\cdot|z_{1:k})$ and the filtering density $\mathbf{\pi}_{k+j}(\cdot|z_{1:k},\hat{z}_{1:j}(a_{k}))$ are the same because the measurement likelihood function is separable. For notational simplicity, $\mathbf{\pi}_1 \triangleq \mathbf{\pi}_{k+j|k}(\cdot|z_{1:k})$ and $\mathbf{\pi}_2 \triangleq \mathbf{\pi}_{k+j}(\cdot|z_{1:k},\hat{z}_{1:j}(a_{k}))$ are two LMB densities on $\mathbb{X}$ with the same label space $\mathbb{L}$ (see Section~\ref{sec_LMB_RFS} for a definition of an LMB density), given by:
\begin{align}
\mathbf{\pi}_1 = \{r_1^{(\lambda)},p_1^{(\lambda)}\}_{\lambda \in \mathbb{L}}; ~ \mathbf{\pi}_2 = \{r_2^{(\lambda)},p_2^{(\lambda)}\}_{\lambda \in \mathbb{L}};
\end{align}
and rewriting $\mathbf{\pi}_{1}$ and $\mathbf{\pi}_{2}$ in terms of LMB densities:
\begin{align}
\mathbf{\pi}_{1}(\mathbf{X}) & =\delta_{|\mathbf{X}|}(\mathbf{|\mathcal{L}(\mathbf{X})|})w_{1}(\mathcal{L}(\mathbf{X}))p_{1}^{\mathbf{X}}\label{eq:pi_1}\\
\mathbf{\pi}_{2}(\mathbf{X}) & =\delta_{|\mathbf{X}|}(|\mathcal{L}(\mathbf{X})|)w_{2}(\mathcal{L}(\mathbf{X}))p_{2}^{\mathbf{X}}.\label{eq:pi_2}
\end{align}
Hence, evaluating $\hat{\mathcal{R}}_{k+j}(a_{k+j})$ requires calculating the divergence between the two LMB densities $\mathbf{\pi}_{2}$ and $\mathbf{\pi}_{1}$. We consider two candidates to measure divergence: \textit{i)} R\'{e}nyi divergence; and \textit{ii)} Cauchy-Schwarz divergence described in Section~\ref{sec_bg_POMDP}. However, given the non-linearity of our measurement likelihood, both divergence measures have no closed form solution. Therefore, we approximate the divergence between two LMB densities using Monte-Carlo sampling. In contrast to~\cite{Amirali2016Cauchy} where Monte Carlo sampling was used to approximate the first moment, we approximate the full distribution.
\vspace{2mm}
\noindent1)~\textbf{R\'{e}nyi Divergence Approximation} From the definition in Section~\ref{sec_bg_POMDP}, we have:
\begin{comment}
Each $\lambda$ component of $\mathbf{\pi}_{j}$ ($j=1,2$), the continuous density $p_{j}^{(\lambda)}(\cdot)$, is approximated by a probability mass function $\hat{p_{j}}^{(\lambda)}(\cdot)$ using the same set of samples
$\{\bar{\zeta}^{(\lambda,i)}\}_{i=1}^{N_s}$ with different weights $\{\omega_{j}^{(\lambda,i)}\}_{i=1}^{N_s}$:
\begin{equation}
p_{j}^{(\lambda)}(\bar{\zeta})\approx\hat{p}_{j}^{(\lambda)}(\bar{\zeta})=\sum_{i=1}^{N_s}\omega_{j}^{(\lambda,i)}\delta_{\bar{\zeta}^{(\lambda,i)}}(\bar{\zeta}).
\end{equation}
\end{comment}
\begin{align}
&D_{\RIfM@\expandafter\text@\else\expandafter\mbox\fi{R\'{e}nyi}}(\mathbf{\pi}_{2},\mathbf{\pi}_{1}) =\dfrac{1}{\alpha-1}\log\int\mathbf{\pi}_{2}^{\alpha}(\mathbf{X})\mathbf{\pi}_{1}^{1-\alpha}(\mathbf{X})\delta\mathbf{X} \notag\\
&=\dfrac{1}{\alpha-1}\log\int\Big[\big(\delta_{|\mathbf{X}|}(\mathbf{|\mathcal{L}(\mathbf{X})|})w_{2}(\mathcal{L}(\mathbf{X}))[p_{2}(\cdot)]^{\mathbf{X}}\big)^{\alpha} \\
&~~~~~~~~~~~~~~~~~\times
\big(\delta_{|\mathbf{X}|}(\mathbf{|\mathcal{L}(\mathbf{X})|})w_{1}(\mathcal{L}(\mathbf{X}))[p_{1}(\cdot)]^{\mathbf{X}}\big)^{1-\alpha}\Big]\delta\mathbf{X}. \notag
\end{align}
Since $[p^{\mathbf{X}}]^{\alpha} = \big[\prod_{\mathbf{x}\in \mathbf{X}}p(\mathbf{x})\big]^{\alpha} = \prod_{\mathbf{x}\in \mathbf{X}}[p(\mathbf{x})]^{\alpha} = [p^{\alpha}]^{\mathbf{X}}$, using Lemma 3 in \cite{vo2013glmb}, this becomes:
\begin{align} \label{eq:Renyi_gen}
&D_{\RIfM@\expandafter\text@\else\expandafter\mbox\fi{R\'{e}nyi}}(\mathbf{\pi}_{2},\mathbf{\pi}_{1}) =\dfrac{1}{\alpha-1}\log\Bigg[\sum_{L\subseteq\mathbb{L}}w_{2}^{\alpha}(L)w_{1}^{1-\alpha}(L) \\
&~~~~~~~~~~~~\times \prod_{\lambda\in L}\Big[\int\big[p_{2}^{(\lambda)}(\bar{\zeta})\big]^{\alpha}\big[p_{1}^{(\lambda)}(\bar{\zeta})\big]^{1-\alpha}d\bar{\zeta}\Big]\Bigg]. \notag
\end{align}
Each $\lambda$ component of $\mathbf{\pi}_{j}$ ($j=1,2$), the continuous density $p_{j}^{(\lambda)}(\cdot)$, is approximated by a probability mass function $\hat{p_{j}}^{(\lambda)}(\cdot)$ using the same set of samples
$\{\bar{\zeta}^{(\lambda,i)}\}_{i=1}^{N_s}$ with different weights $\{\omega_{j}^{(\lambda,i)}\}_{i=1}^{N_s}$:
\begin{equation}
p_{j}^{(\lambda)}(\bar{\zeta})\approx\hat{p}_{j}^{(\lambda)}(\bar{\zeta})=\sum_{i=1}^{N_s}\omega_{j}^{(\lambda,i)}\delta_{\bar{\zeta}^{(\lambda,i)}}(\bar{\zeta}).
\end{equation}
Using Monte Carlo sampling, the product between the two continuous
densities in \eqref{eq:Renyi_gen} can be approximated by the product of two probability mass functions on the finite samples $\{\bar{\zeta}^{(\lambda,i)}\}_{i=1}^{N_s}$, given by:
\begin{equation}
\resizebox{0.5\textwidth}{!}{ $\begin{aligned} \label{eq:Renyi_prod}
&\int\big[p_{2}^{(\lambda)}(\bar{\zeta})\big]^{\alpha}\big[p_{1}^{(\lambda)}(\bar{\zeta})\big]^{1-\alpha}d\bar{\zeta} \approx\sum_{i=1}^{N_s}\big[\hat{p}{}_{2}^{(\lambda)}(\bar{\zeta}^{(\lambda,i)})\big]^{\alpha}\big[\hat{p}_{1}^{(\lambda)}(\bar{\zeta}^{(\lambda,i)})\big]^{1-\alpha}
\\
& \approx\sum_{i=1}^{N_s}\big[\sum_{j=1}^{N_s}\omega_{2}^{(\lambda,j)}\delta_{\bar{\zeta}^{(\lambda,j)}}(\bar{\zeta}^{(\lambda,i)})\big]^{\alpha}\big[\sum_{k=1}^{N_s}\omega_{1}^{(\lambda,k)}\delta_{\bar{\zeta}^{(\lambda,k)}}(\bar{\zeta}^{(\lambda,i)})\big]^{1-\alpha}\\
& \approx\sum_{i=1}^{N_s}\big[\omega_{2}^{(\lambda,i)}\big]^{\alpha}\big[\omega_{1}^{(\lambda,i)}\big]^{1-\alpha},
\end{aligned} $ }
\end{equation}
Substituting \eqref{eq:Renyi_prod} into \eqref{eq:Renyi_gen}, the R\'{e}nyi divergence becomes:
\begin{align}
&D_{\textnormal{R\'{e}nyi}}(\mathbf{\pi}_{2},\mathbf{\pi}_{1})\label{eq:Renyi_divergence} \approx\dfrac{1}{\alpha-1} \\
&\times \log\Bigg[\sum_{L\subseteq\mathbb{L}}w_{2}^{\alpha}(L)w_{1}^{1-\alpha}(L)\prod_{\lambda\in L}\Big[\sum_{i=1}^{N_s}\big(\omega_{2}^{(\lambda,i)}\big)^{\alpha}\big(\omega_{1}^{(\lambda,i)}\big)^{1-\alpha}\Big]\Bigg].\notag
\end{align}
\noindent2)~\textbf{Cauchy-Schwarz Divergence Approximation}
From the definition in Section~\ref{sec_bg_POMDP} and following~ \cite{beard2015void}, we have
\begin{align}
D_{CS}(\mathbf{\pi}_{2},\mathbf{\pi}_{1}) & =-\log\Big(\dfrac{\langle\mathbf{\pi}_{2},\mathbf{\pi}_{1}\rangle_{K}}{\sqrt{\langle\mathbf{\pi}_{2},\mathbf{\pi}_{2}\rangle_{K}\langle\mathbf{\pi}_{1},\mathbf{\pi}_{1}\rangle_{K}}}\Big),\label{eq:cs_divergence}
\end{align}
where
\begin{align}
\langle\mathbf{\pi}_{i},\mathbf{\pi}_{j}\rangle_{K} & =\sum_{L\subseteq\mathbb{L}}w_{i}(L)w_{j}(L)\prod_{\lambda\in L}K\langle p_{i}^{(\lambda)}(\cdot),p_{j}^{(\lambda)}(\cdot)\rangle,\label{eq:Cauchy_gen}
\end{align}
for $i,j \in \{1,2\}$. Using the approach in \eqref{eq:Renyi_prod}, we have
\begin{align} \notag
\langle\mathbf{\pi}_{i},\mathbf{\pi}_{j}\rangle_{K} & \approx\sum_{L\subseteq\mathbb{L}}w_{i}(L)w_{j}(L)\prod_{\lambda\in L}K\Big(\sum_{k=1}^{N_s}\omega_{i}^{(\lambda,k)}\omega_{j}^{(\lambda,k)}\Big).\label{eq:cs_divergence_element}
\end{align}
The UAV needs to maintain a safe distance from objects, although getting close to the objects of interest improves tracking accuracy. Therefore, in the following section, we derive a void constraint for the path planning formulation.
\subsubsection{Void Probability Functional}
Let $V(u_{k+j}(a_k),r_{\min})$
denote the void region of objects
based on a UAV's position at time $k+j$ if an action $a_k$ is taken. This leads to a cylinder shape where the ground distance between a UAV and any objects should be smaller than $r_{\min}
, given by:
\begin{equation}
\begin{aligned}
&V(u_{k+j}(a_k),r_{\min}) = \Big\{ \mathbf{x} \in \mathbb{X}: \label{eq_exclusion_zone_def1} \\
&\sqrt{ (p^{(\mathbf{x})}_x - p_x^{(u_{k+j}(a_k))})^2 + (p^{(\mathbf{x})}_y - p_y^{(u_{k+j}(a_k))})^2} < r_{\min} \Big\},
\end{aligned}
\end{equation}
where $p^{(\mathbf{x})}_x,p^{(\mathbf{x})}_y$ and $p^{(u_{k+j}(a_k))}_x,p^{(u_{k+j}(a_k))}_y$ denote positions of $\mathbf{x}$ and $u_{k+j}(a_k)$ in $x-y$ coordinates, respectively.
Using the closed form expression for the void probability functional\footnote{Here, we use the notion of void probabilities as defined in \cite{kendall1995stochastic}.} of the GLMB in \cite{beard2015void}, we impose the constraint in (44) on the trajectory planning problem as formulated below.
Given a region $S \subseteq \mathbb{X}$ and an LMB density $\mathbf{\pi}$ on $\mathbb{X}$ parameterized as
$\mathbf{\pi} = \delta_{|\mathbf{X}|}(\mathbf{|\mathcal{L}(\mathbf{X})|})w(\mathcal{L}(\mathbf{X}))p^{\mathbf{X}} = \{r^{(\lambda)},p^{(\lambda)}\}_{\lambda \in \mathbb{L}}$ where each $\lambda$ component is approximated by a set of weighted samples $\{\omega^{(\lambda,i)},\bar{\zeta}^{(\lambda,i)}\}_{i=1}^{N_s}: p^{(\lambda)}(\bar{\zeta}) \approx \sum_{i=1}^{N_s} \omega^{(\lambda,i)} \delta_{\bar{\zeta}^{(\lambda,j)}}(\bar{\zeta})$, the void functional of $S$ given the multi-object density $\mathbf{\pi}$, $B_{\pi}(S)$, can be approximated as:
\begin{align} \notag
B_{\mathbf{\pi}}(S) \approx \sum_{L \subseteq \mathbb{L}} w(\mathcal{L})
\prod_{\lambda \in \mathbb{L}} \big(1-\sum_{i=1}^{N_s}w^{(\lambda,i)} \delta_{\bar{\zeta}^{(\lambda,i)}}(\bar{\zeta}) 1_{S}(\bar{\zeta}) \big)
\end{align}
using the expression of the void probability functional in \cite{beard2015void}.
\begin{comment}
Here, $1_S(\cdot)$ denotes the indicator function of $S$, given by
\begin{align}
1_S(\bar{\zeta}) = \begin{cases}
1 & \RIfM@\expandafter\text@\else\expandafter\mbox\fi{if } \bar{\zeta} \in S \\
0 & \RIfM@\expandafter\text@\else\expandafter\mbox\fi{otherwise. }
\end{cases}
\end{align}
\end{comment}
Now the maximization problem in \eqref{eq_reward_pims} becomes:
\begin{align}
a^{*}_{k}&=\argmax\limits_{a_{k} \in \mathcal{A}_{k}}
\sum_{j=1}^{H} \gamma^{j-1}\hat{\mathcal{R}}_{k+j}(a_{k})
\end{align}
subject to the constraint
\begin{equation}\notag
\min\limits_{j\in \{1,\dots ,H\}}[B_{\pi_{k+j}(\cdot|z_{1:k})}(V(u_{k+j}(a_{k}),r_{\min}))]>P_{\vmin}
\end{equation}
\noindent where $P_{\vmin}$ denotes a void probability threshold.
\begin{comment}
\subsubsection{Void Probability Functional}
Using the closed form formulation of the void probability functional in \cite{beard2015void}, we construct two constraints illustrated in Fig.~\ref{fig_UAV_exclusion_zone}.
Given a region $S \subseteq \mathbb{X}$ and an LMB density $\mathbf{\pi}$ on $\mathbb{X}$ parameterized as $\mathbf{\pi} = \{r^{(\lambda)},p^{(\lambda)}\}_{\lambda \in \mathbb{L}}$ where each $\lambda$ component is approximated by a set of weighted samples $\{w^{(\lambda,i)},\bar{\zeta}^{(\lambda,i)}\}_{i=1}^{N_s}: p^{(\lambda)}(\bar{\zeta}) \approx \sum_{i=1}^{N_s} w^{(\lambda,i)} \delta_{\bar{\zeta}^{(\lambda,j)}}(\bar{\zeta})$, the belief functional of $S$ given the multi-object density $\mathbf{\pi}$, $B_{\pi}(S)$, can be approximated as:
\begin{align}
B_{\mathbf{\pi}}(S) \approx \prod_{\lambda \in \mathbb{L}} \big(r^{(\lambda)} \sum_{i=1}^{N_s}w^{(\lambda,i)} \delta_{\bar{\zeta}^{(\lambda,i)}}(\bar{\zeta}) 1_{S}(\bar{\zeta}) \big)
\end{align}
using the expression of the void probability functional in \cite{beard2015void}. Here, $1_S(\cdot)$ denotes the indicator function of $S$, given by
\begin{align}
1_S(\bar{\zeta}) = \begin{cases}
1 & \RIfM@\expandafter\text@\else\expandafter\mbox\fi{if } \bar{\zeta} \in S \\
0 & \RIfM@\expandafter\text@\else\expandafter\mbox\fi{otherwise. }
\end{cases}
\end{align}
Now, as illustrated in Fig. \ref{fig_UAV_exclusion_zone}, let $V(a_k,r_{\min},r_{\max})$ denote the preferable region of objects based on UAV's positions at time $k$ if an action $a_k$ is taken. This leads to an annulus where a ground distance between UAV and any objects should be larger than $r_{\min}$ but smaller than $r_{\max}$, given by:
\begin{align}
V(a_k,&r_{\min},r_{\max}) = \Big\{ \mathbf{x} \in \mathbb{X}: \\ &r_{\min} < \sqrt{ (p^{(\mathbf{x})}_x - p_x^{(a_k)})^2 + (p^{(\mathbf{x})}_y - p_y^{(a_k)})^2} < r_{\max} \Big\}, \notag
\end{align}
where $p^{(\mathbf{x})}_x,p^{(\mathbf{x})}_y$ and $p^{(a_k)}_x,p^{(a_k)}_y$ denote positions of $\mathbf{x}$ and $a_k$ in $x-y$ coordinates, respectively;
$P_{\vmin}$ denotes a void probability threshold; the maximization problem in \eqref{eq_reward_pims} becomes:
\begin{align}
a^{*}_{k+1:k+H}&=\argmax\limits_{a_{k+1:k+H} \in \mathcal{A}_{k+1:k+H}}
\sum_{j=1}^{H} \gamma^{j}\hat{\mathcal{R}}_{k+j}(a_{k+j})
\end{align}
subject to the constraint
\begin{equation}\notag
\min\limits_{j\in \{1,\dots ,H\}}[B_{\pi_{k+j}(\cdot|z_{1:k})}(V(a_{k+j},r_{\min},r_{\max}))]>P_{\vmin}.
\end{equation}
\begin{figure}[!tb]
\centering
\includegraphics[width=4cm]{figures/UAV_exclusion_zone.pdf}
\caption{Preferable region for objects based on a UAV's position.
\label{fig_UAV_exclusion_zone}
\end{figure}
\end{comment}
\section{Simulation Experiments} \label{sec_simulation_experiments}
In this section, we evaluate the proposed online path planning strategy for joint detection and tracking of multiple radio-tagged objects using a UAV.
\subsection{Experimental Settings} \label{sec_dynamic_model}
A two-dimensional area of $[0,1500]~\RIfM@\expandafter\text@\else\expandafter\mbox\fi{m}\times \lbrack 0,1500]~\RIfM@\expandafter\text@\else\expandafter\mbox\fi{m}$ is
investigated to demonstrate the proposed approach. The UAV's height is maintained at $30$~m while the objects' heights are fixed at $1$~m to limit the scope to a two-dimensional (2D)
problem\footnote{It can be easily extended to 3D; however, to save computational power, we limit our problem to the 2D domain. }. The total flight time is $400~$s for all experiments.
\begin{figure*}[!tb]
\centering
\includegraphics[width=\textwidth]{figures/Ex1_Raw_Signal_with_Spectrogram.pdf}
\caption{An illustration of received signals from four transmitting objects at distances of $[120,~515,~400,~920]$~m for object 1, object 2, object 3 and object 4 respectively, to the UAV in the presence of complex receiver noise covariance $\Sigma_{\eta} = 0.02^2~V^2$. a) The received signal without noise in time domain; b) the received signal in the presence of the complex white noise in time domain. c) Spectrogram of the received signal in discrete time and frequency domain ($111\times 256$ frames) where the bright spots represent an object's signal in a time-frequency frame.}
\label{fig_Ex1_Raw_Signal_and_Spectrogram}
\end{figure*}
We also follow the same practical constraints mentioned in \cite{hoa2017icra} for our simulations. The UAV cannot turn its heading instantly, hence its maximum turning rate is limited to $\triangle \theta^u_k = |\theta_k^u -\theta^u_{k-1}| \leq \theta^u_{max}$ (rad/s). In addition, since the planning step normally consumes more time than the tracking step, we apply a cruder planning interval $N_p$ compared to measurement interval $T_0$, such that $N_p = n T_0$ where $n \geq 2,~ n \in \mathbb{N}$ (\textit{i.e.}, $T_0 = 1$ s, $N_p = 5$ s, the planning algorithm calculates the best trajectory for the UAV in next five seconds at each five-measurement-intervals instead of every measurement-interval.)
An object's dynamic mode $s$ follows the jump Markov system where its motion model is either: \textit{i)} a \textit{Wandering} (WD) mode where an object moves short distances without any clear purpose or direction; or \textit{ii)} a constant velocity (CV) mode.
\vspace{2mm}
\noindent\textbf{The Wandering (WD) Model}:
\begin{align} \label{eq_RW_model}
x_k &= F^{WD}_{k-1}x_{k-1} + q^{WD}_{k-1}
\end{align}
where $F^{WD}_{k-1} = \diag([1~0~1~0]^T)$, $q^{WD}_{k-1} \sim \mathcal{N}(0,Q^{WD})$ is a zero mean Gaussian process noise with covariance $Q^{WD} =\diag([0.25~\RIfM@\expandafter\text@\else\expandafter\mbox\fi{m}^2,2.25~\RIfM@\expandafter\text@\else\expandafter\mbox\fi{(m/s)}^2,~0.25~\RIfM@\expandafter\text@\else\expandafter\mbox\fi{m}^2,2.25~\RIfM@\expandafter\text@\else\expandafter\mbox\fi{(m/s)}^2]^T) $.
In the wandering model, the velocity components are instantly forgotten and then sampled from covariance $Q^{WD}$ at each time step $k$. However, the sampled velocity components do not influence an object's position. Further, the velocity components in $Q^{WD}$ are significantly larger than the position components therein. This is necessary to achieve the fast moving behavior of objects in the constant velocity dynamic mode when an object switches from the wandering mode to the constant velocity mode.
\vspace{2mm}
\noindent\textbf{The Constant Velocity (CV) Model}:
\begin{align} \label{eq_CV_model}
x_k &= F^{CV}_{k-1} x_{k-1} + q^{CV}_{k-1}, \nonumber
\\
F^{CV}_{k-1} &=
\begin{pmatrix}
1 & T_0 \\
0 & 1%
\end{pmatrix}
\otimes I_2,
\end{align}
where $\otimes$ denotes the Kronecker tensor product operator between
two matrices, and $q^{CV}_{k-1} \sim \mathcal{N}(0,Q^{CV})$ is a $%
4\times 1$ zero mean Gaussian process noise, with covariance $Q^{CV} =
\sigma_{CV}^2%
\begin{pmatrix}
T_0^3/3 & T_0^2/2 \\
T_0^2/2 & T_0%
\end{pmatrix}%
\otimes I_2$, where $\sigma_{CV}$ is the standard deviation of
the process noise parameter.
There are four objects with different birth and death times, listed in pairs as $(t_{\rm birth},t_{\rm death})$: $(1,250)$, $(50,300)$, $(100,350)$, $(150,400)$~s. The four objects initially follow the wandering model (WD) with initial state vectors $[800,0.13,300,-1.44]^T$, $[200,0.18,700,-2.17]^T$, $[1200,-1.94,1000,0.42]^T$, $[900,1.91,1300,-2.04]^T$ (with appropriate standard units) at birth. One second period after birth, object 1 and object 3 switch their dynamic mode to the constant velocity mode while object 2 and object 4 continue to follow the wandering model for $65$~s. We detail the mode changes (later) in Fig.~\ref{fig_Ex1_Estimated_Mode}.
For each newly born object, we assume an initial birth state described by a Gaussian distribution with means at $[800,0,300,0]^T$, $[200,0,700,0]^T$, $[1200,0,1000,0]^T$, $[900,0,1300,0]^T$ (with appropriate standard units) and covariance $Q^B=\diag([100~\RIfM@\expandafter\text@\else\expandafter\mbox\fi{m}^2,4~\RIfM@\expandafter\text@\else\expandafter\mbox\fi{(m/s)}^2,
100~\RIfM@\expandafter\text@\else\expandafter\mbox\fi{m}^2,4~\RIfM@\expandafter\text@\else\expandafter\mbox\fi{(m/s)}^2]^T)$. In practice, such a setting is reasonable and captures the prior knowledge about an object's location. For example, in applications such as wildlife tracking, conservation biologists know the location of newly released wildlife or the locations of entry and exit points of animals that can suddenly appear in a scene from underground animal dwellings.
\begin{comment}
There \hlnew{is} a maximum of four objects with different birth and death times listed in pairs $(t_{\rm birth},t_{{\rm death}})$ as follows: $(1,250),~
(50,300),~(100,350),~(150,400)$ s. \hlnew{The four objects initially follow the wandering model with their initial state vectors are $[800,0.13,300,-1.44]^T,[200,0.18,700,-2.17]^T,$ $[1200,-1.94,1000,0.42]^T,$ $[900,1.91,1300,-2.04]^T$ when they are born.
After being born for $1$~s, object 1 and object 3 change its motion model to the constant velocity model while object 2 and object 4 continues to follow the wandering model for $65$~s (see later in Fig.~\ref{fig_Ex1_Estimated_Mode}).
For each object birth position, we assume it follows initial Gaussian distributions with position means at \hlnew{$[800,300]^T,~[200,700]^T,~[1200,1000]^T,~[900,1300]^T$}~m with its birth position covariance $Q^B_{p_x} = Q^B_{p_y} = 100$~m$^2$ given we have some prior knowledge about its birth positions. For the object birth velocity, since there is no prior information because all of four objects are initialized with the wandering model, its birth velocity is sampled from $Q^{WD}$ and is unknown to the filter, the birth velocity follows a zero-mean Gaussian distribution with its birth velocity covariance $Q^B_{\dot{p}_x} = Q^B_{\dot{p}_y} = 4$~(m/s)$^2$.}
\end{comment}
All the common parameters used in the following experiments are listed in Tables~\ref{table_birth_death_mode}, \ref{table_signal_parameters}, and \ref{table_measurement_parameter}.
In addition, Fig. \ref{fig_Ex1_Raw_Signal_and_Spectrogram}a
illustrates a raw received signal without noise from four transmitted objects along with a noisy received signal in Fig. \ref{fig_Ex1_Raw_Signal_and_Spectrogram}b. Furthermore, a single measurement set of the noisy received signal after going through the STFT process consists of $111 \times 256$ time-frequency frames is illustrated in Fig. \ref{fig_Ex1_Raw_Signal_and_Spectrogram}c.
\begin{comment}
\begin{figure*}[!tb]
\centering
\includegraphics[width=0.8\textwidth]{figures/Ex1_scenario_setup.pdf}
\caption{\hlnew{Experimental setup with true trajectories of four objects. The start/stop locations for each object are denoted by $\circ/\Box$. The start location for the UAV is denoted by $\bigstar$. Note that objects are born and died at different time, which is not shown on the figure. }}
\label{fig_Ex1_scenario_setup}
\end{figure*}
\end{comment}
\begin{table}[!tb]
\centering
\caption{Birth, death, and dynamic mode parameters}
\label{table_birth_death_mode}
\begin{tabular}{L{3.6cm}|L{3.6cm}} \hline
\textbf{Parameter} & \textbf{Value} \\ \hline
Birth probability ($r_B$) & $10^{-6}$ \\ \hline
Survival probability ($p_S$) & 0.99 \\ \hline
Initial mode probability & $[0.5~0.5]^T$ \\ \hline
Mode transitional probability & $[0.99~0.01;0.01~0.99]$ \\ \hline
Constant velocity noise ($\sigma_{CV}$) & $0.05$ m/$\RIfM@\expandafter\text@\else\expandafter\mbox\fi{s}^2$ \\ \hline
\end{tabular}
\end{table}
\begin{table}[!tb]
\centering
\caption{Signal parameters}
\label{table_signal_parameters}
\begin{adjustbox}{max width=0.45\textwidth}
\begin{tabular}{L{3cm}|C{1.5cm}|L{3cm}}
\hline
\textbf{Parameter} & \textbf{Symbol} & \textbf{Value} \\ \hline
Center frequency & $f_c$ & $150$ MHz \\ \hline
Baseband frequencies & $f^{(\lambda)}$ & $131$ kHz, $201$ kHz, $401$kHz , $841$ kHz \\ \hline
Sampling frequency & $f_s$ & $2$ MHz \\ \hline
Pulse period & $T_0$ & $1$ s \\ \hline
Pulse offset time & $\tau^{(\lambda)} $ & $0.1$ s, $0.2$ s, $0.3$ s, $0.4$ s \\ \hline
Pulse width & $P_w$ & $18$ ms \\ \hline
Reference distance & $d_0$ & $1$ m \\ \hline
Pulse amplitude & $A$ & $0.0059$ V \\ \hline
Path loss constant & $\kappa$ & $3.1068$ \\ \hline
\end{tabular}
\end{adjustbox}
\end{table}
\begin{table}[!tb]
\centering
\caption{Measurement parameters}
\label{table_measurement_parameter}
\begin{tabular}{L{3.5cm}|C{1.5cm}|L{2cm}}
\hline
\textbf{Parameter} & \textbf{Symbol} & \textbf{Value} \\ \hline
Receiver gain & $Gr$ & $72$ dB \\ \hline
Receiver noise covariance & $\Sigma_{\eta}$ & $0.025^2$ $\RIfM@\expandafter\text@\else\expandafter\mbox\fi{V}^2$ \\ \hline
Number of window frames & $M$ & $111$ \\ \hline
Number of frequency samples & $L$ & $256$ \\ \hline
Window width & $N_w$ & $256$ \\ \hline
Number of particles & $N_s$ & $50,000$ \\ \hline
UAV's max heading angle & $\theta^u_{\max}$ & $\pi/3$ rad/s \\ \hline
UAV's velocity & $v_u$ & $20$ m/s \\ \hline
UAV's initial position & $u_1$ & $[0;0;30;\pi/4]$ \\ \hline
Planning interval & $N_p$ & $5$~s \\ \hline
Look-a-head horizon & $H$ & 3 \\ \hline
Minimum distance & $r_{\min}$ & $50$ m \\ \hline
Void threshold & $P_{\vmin}$ & $0.9$ \\ \hline
OSPA (order, cut-off) & $(p,c)$ & $(1,100$ m$)$ \\ \hline
\end{tabular}
\end{table}
\subsection{Experiments and Results}
We conduct two different experiments: \textit{i)}
to validate and evaluate our proposed planning method for joint detection and tracking; \textit{ii)} compare performance against planning for tracking with conventional detection-then-track methods.
\begin{figure*}[!tb]
\centering
\includegraphics[width=0.88\textwidth]{figures/Ex1_Estimiated_track_xy_coordinate_And_OSPA.pdf}
\caption{Tracking four objects in various locations with different birth and death times and motion dynamics. Estimated positions and truth in: a) x-coordinate; b) y-coordinate; c) cardinality---its truth versus mean $\mu$ and its variance ($\mu \pm 3\sigma)$; d) OSPA---the cutoff and order parameters are given in Table~\ref{table_measurement_parameter}.}
\label{fig_Ex1_Estimiated_track_xy_coordinate}
\end{figure*}
\begin{figure}[!ht]
\centering
\includegraphics[width=0.5\textwidth]{figures/Ex1_Estimated_Mode.pdf}
\caption{The estimated mode probability for four objects with mode WD:Wandering and mode CV:Constant Velocity: a) object $1$; b) object $2$; c) object $3$; and d) object $4$. }
\label{fig_Ex1_Estimated_Mode}
\end{figure}
\vspace{2mm}
\noindent\textbf{Experiment 1--Validating Planning for Joint Detection and Tracking: } The first experiment is conducted with four objects in various locations and moving in different directions where birth and death times and motion dynamics are described in Section~\ref{sec_dynamic_model}. We employ R\'{e}nyi divergence based reward function with receiver noise covariance $\Sigma_{\eta} = 0.025^2~\RIfM@\expandafter\text@\else\expandafter\mbox\fi{V}^2$ and the UAV undergoes trajectory changes every 15~s, \textit{i.e.}, the planning interval $N_p = 5$~s with a look ahead horizon $H = 3$ (see Table~\ref{table_measurement_parameter}).
Fig. \ref{fig_Ex1_Estimiated_track_xy_coordinate}a-b depict true object trajectories, birth and death times together with the estimated tracking accuracy for a typical experiment run. The results show that the proposed planning for joint detection and tracking accurately estimates position and cardinality of the objects.
Fig.~\ref{fig_Ex1_Estimiated_track_xy_coordinate}c depicts the ground truth changes in the number of objects over time with the estimated cardinality. We used optimal sub-pattern assignment (OSPA)~\cite{schuhmacher2008consistent} to quantify the error between the filter estimates and the ground truth to evaluate the multi-object miss distance. The spikes in Fig.~\ref{fig_Ex1_Estimiated_track_xy_coordinate}c indicate a high uncertainty in the estimated cardinality distribution. The high uncertainty is due to low signal-to-noise ratio (SNR) of received measurements. During path planning, noisy signals lead to poor control decision that result in the UAV navigating to positions further from objects of interest where the signal incident on the UAV sensor antenna is often at an angle where the antenna gain is poor. Further, planning decisions are also subject to void constraints. Consequently, the existence probability of objects of interest can suddenly increase or decrease after a poor control action.
The OSPA distance performance over the tracking period for these objects is depicted in Fig.~\ref{fig_Ex1_Estimiated_track_xy_coordinate}d. We see changes in OSPA distance during birth and death events and its subsequent reduction as the planning algorithm undergoes course changes to improve tracking accuracy. These results confirm that our trajectory planning algorithm consistently tracks the number of object change over time whilst making course changes to improve estimation accuracy of all the objects.
Fig.~\ref{fig_Ex1_Estimated_Mode} depicts the multiple motion modes of objects and how it changes over time. The results show that although the received signals are noisy, the filter can still accurately estimate the correct mode of objects most of the time.
\begin{figure*}[!tb]
\centering
\centering
\includegraphics[width=\textwidth]{figures/Ex1_Estimated_Positions.pdf}
\caption{A typical UAV trajectory (green path) under the proposed path planning for joint detection and tracking algorithm for multiple radio-tagged objects. Here, `$\circ$' locations of object births; `$\Box$' locations of object deaths; `$\Diamond$' current locations of the UAV. Faint tracks show objects subject to a death process.}
\label{fig_Ex1_Estimated_Positions}
\end{figure*}
Fig.~\ref{fig_Ex1_Estimated_Positions} depicts the evolution of true and estimated object trajectories under the control of the path planning scheme subject to the void constraint. From these snapshots in time, we can see that a typical trajectory to track objects under the birth and death process agrees with our intuition. Initially, the UAV navigates towards object~$1$. At time $t=50$~s (Fig. \ref{fig_Ex1_Estimated_Positions}a), object~$2$ is born; subsequently, the UAV maintains a trajectory between the two objects with course changes to track both objects. Object~$3$ is born at $t=100$~s, the UAV undertakes course changes to estimate the position of all three moving objects with a maneuver to follow object~$1$ and $2$ whilst moving closer to object $3$ (Fig. \ref{fig_Ex1_Estimated_Positions}b~and~c). We can observe a similar planning strategy evolving when object~$4$ is born at time $150$~s. The UAV navigates to a position to be closer to all four objects and maintain a position at the center of the four objects to estimate the position of all four objects (Fig.~\ref{fig_Ex1_Estimated_Positions}d~and ~e). At time $250$~s, object $1$ vanishes, thus the UAV moves up toward to a position at the center of object $2$, object $3$ and object $4$ to track the remaining objects (Fig.~\ref{fig_Ex1_Estimated_Positions}f). Beyond $300$~s, both object $1$ and object $2$ are no longer in existence; therefore we can observe the UAV heading to a position between objects $3$---whilst maintaining the void constraint illustrated by the dashed circle at the UAV position---and object $4$ (Fig.~\ref{fig_Ex1_Estimated_Positions}g). After time $350$~s, only object $4$ is in existence; thus, the UAV undertakes trajectory changed to move towards object $4$ (Fig. \ref{fig_Ex1_Estimated_Positions}h). The results show that the proposed planning strategy is able to detect and track all objects whilst dynamically acting upon different birth and death events to maneuver the UAV to move to positions that minimize the overall tracking error.
\vspace{2mm}
\noindent\textbf{Experiment 2--Comparing Performance: }
In this experiment, we compare our proposed online path planning for joint detection and tracking formulation with the TBD-LMB filter with planning for detection-then-track (DTT) methods using a DTT-LMB filter \cite{reuter2014lmb}. We compare three trajectory planning approaches for tracking: \textit{i)} a straight path---direct the UAV back and forth along a diagonal line between $(0,0)$~m and $(1500,1500)$~m;
\textit{ii)} planning with R\'{e}nyi divergence as the reward function; and \textit{iii)} planning
with Cauchy divergence as the reward function
The measurements for DTT are extracted based on a peak detection algorithm to find the prominent peak such that the minimum peak separation is $N_m=8$ frequency bins---i.e. the number of main-lobe width-in-bin for a 4-term Blackman Harris window as listed in Table~\ref{table_L_scale_window_type}. Since we examine the filter performance under various receiver noise levels, it is more appropriate to use a peak detection method compared to a fixed threshold value. Further, the peak detection method is robust against different noise levels, considering false alarm and misdetection rates \cite{scholkmann2012efficient}. The planning for DTT methods uses the same PIMS approach for TBD methods listed in Section~\ref{sec_rewardfunctionsforplaning}.
We use OSPA cardinality and OSPA distance to compare performance across the three planning strategies for TBD and DTT approaches. We perform 100 Monte Carlo runs for each of the six cases and receiver noise levels $\Sigma_{\eta} =
\{0.010^2, 0.015^2,\dots,0.050^2\}~\RIfM@\expandafter\text@\else\expandafter\mbox\fi{V}^2$ for the the scenario shown in Fig.~\ref{fig_Ex1_Estimated_Positions}.
OSPA distance and cardinality results in Fig.~\ref{fig_Ex2_OSPA} show that the proposed path planning for TBD strategy provides significantly better estimation performance over planning for DTT based strategies as demonstrated by the lower OSPA distance metrics in the presence of increasing receiver noise. The TBD approaches are more effective than DTT approaches, especially because of the failure of DTT methods to detect changes in the number of objects in the presence of birth and death processes as evident in Fig.~\ref{fig_Ex2_OSPA}b.
\begin{figure}[!tb]
\centering
\includegraphics[width=0.47\textwidth]{figures/Ex4_OSPA_D_C_vs_Sigma_Noise.pdf}
\caption{Mean OSPA performance comparison across increasing receiver noise values.
Here, -Straight, -R\'{e}nyi and -Cauchy denote straight path, R\'{e}nyi divergence and Cauchy divergence based planning strategies, respectively: a) OSPA distance; b) OSPA cardinality.}
\label{fig_Ex2_OSPA}
\end{figure}
Intuition suggests that information based approach should execute control actions to continually position the UAV to locations with the best ability to track multiple objects undergoing motion changes. Information based planning strategies outperforming the straight path approaches in both the TBD and DTT methods agrees with this intuition. Although, R\'{e}nyi or Cauchy divergence as reward functions improve the overall tracking performance compared to the straight path method, we also observe that R\'{e}nyi divergence is more discriminative than Cauchy divergence in our task and yields better OSPA distance values and hence the best performance.
\section{Conclusion} \label{sec_conclusion}
In this paper, we have proposed an online path planning algorithm for joint detection and tracking of multiple radio-tagged objects under low SNR conditions. The planning for multi-object tracking problem was formulated as a POMDP with two information-based reward functions and the JMS TBD-LMB filter. In particular, the planning formulation incorporates the practical constraint to maintain a safe distance between the UAV and objects of interest to minimize the disturbances from the UAV. We have derived a measurement likelihood for the TBD-LMB filter and proved that the likelihood is separable in practice for multiple radio-tagged objects; thus deriving an accurate multi-object TBD filter. The results demonstrated that our approach is highly effective in reducing the estimation error of multiple-objects in the presence of low signal-to-noise ratios compared to both detection-then-track approaches and tracking without planning.
|
\section{Experiments}
\label{sec:experiments}
To derive a baseline set of metrics between pairs of images, we selected $\alpha$ to maximize mutual information between registered images. This metric was then used to define the kernel in the KLDA classifier, the results of which we used as a baseline accuracy for our proposed method.
\begin{figure} [ht]
\begin{center}
\begin{tabular}{c}
\includegraphics[height=6.5cm]{pics/synthetic_data.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:synthetic_data_ex}
\textbf{Synthetic data generation}
}
\end{figure}
Our initial experiments were based on 100 images of rectangles, and 100 images of ellipses, each generated with a random locally affine deformation sufficiently noisy to obscure the original class of the image to the naked eye \textbf{(Figure \ref{fig:synthetic_data_ex})}.
Using 50 deformed ellipses and 50 deformed rectangles as a training dataset, we optimized $\alpha$ until hinge loss convergence. \textbf{Figure \ref{fig:roc_auc}) (left)}
shows that EM converges stably after several iterations based on ROC AUC. The final model, chosen based on the best training ROC AUC, performed nearly as well on the synthetic test dataset: ROC AUC = 0.84.
\begin{figure}[!htb]
\centering
\begin{minipage}[b]{0.49\textwidth}
\includegraphics[width=\textwidth]{pics/synthetic.pdf}
\end{minipage}
\hfill
\begin{minipage}[b]{0.49\textwidth}
\includegraphics[width=\textwidth]{pics/roc_acu_hippo_best.pdf}
\end{minipage}
\caption
{ \label{fig:roc_auc} ROC area under the curve vs. EM iterations on \textbf{(left) synthetic data} and \textbf{(right) hippocampal shape}
}.
\end{figure}
Our 3D hippocampal shape sample was derived from the SchizConnect brain MRI dataset \cite{schizconnect}. We used right hippocampal segmentations extracted with FreeSurfer \cite{fischl2012freesurfer} from 227 Schizophrenia (SCZ) patients and 496 controls (CTL). All shapes were affinely registered to the ENIGMA hippocampal shape atlas \cite{ShapeNatureComm}, and their binary mask was computed from the transformed mesh model.
We again used 100 training examples (50 CTL, 50 SCZ) in all our experiments below, using the remaining sample as a test dataset. To derive baseline results to compare with our algorithm's performance on hippocampal shapes, we constructed two additional discriminative models.
(1) A logistic regression model simply using the vectorized binary mask. No spatial information is used in this model.
(2) A KLDA model constructed using LDDMM metrics optimized for registration quality.
ROC AUC scores for the three models are shown in \ref{tb:result}.
\begin{table}[h]
\centering
\begin{tabular}{|p{1.6cm}|p{2.8cm}|p{2.3cm}|p{4.2cm}|}
\hline
{} & Logistic Regression & Maximum MI & Optimized LDDMM-kernel\\ \hline
ROC AUC &0.36$\pm$0.02 & 0.72 $\pm$ 0.06 & 0.75 $\pm$ 0.06 \\\hline
\end{tabular}
\caption{ROC AUC scores for three models}
\label{tb:result}
\end{table}
As expected, ignoring spatial information leads to significant drop in performance. It is also encouraging to see improvement in the classification accuracy when the LDDMM metric is optimized for this explicitly. The stability of the EM algorithm trained on hippocampal shapes is comparable to stability when synthetic data, as seen in \textbf{Figure \ref{fig:roc_auc}) (right)}. To visualize the difference in the kernel-based models, we project the mean difference between SCZ subjects and controls in the scalar momenta defining the registration velocity fields \cite{Mang2016DistributedMemoryLD}, as seen in \textbf{Figure \ref{fig:momenta}}.
\begin{figure} [ht]
\begin{center}
\begin{tabular}{c}
\includegraphics[height=10cm]{pics/Momentum.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:momenta}
\textbf{Mean momentum difference between Schizophrenics and healthy subjects}, using (A) a classification-optimized metric and (B) a metric optimized for pairwise mutual information. The effect in the latter is diffuse, while the classification-aware metric focuses on the hippocampal tail.
}
\end{figure}
\section{Introduction}
\label{sec:intro}
Image registration, and more generally geometric alignment underlies a large number of analyses in medical imaging, particularly neuroimaging. As one of the mainstays of medical image analysis, the problem has been addressed extensively over the last 2+ decades, with several flavors of robust algorithms \cite{Klein_compare_14_registrations}. A number of registration approaches develop an explicit metric space comprised of the geometric objects of interest~--- anatomical shapes, diffusion tensors, images, etc. Prominent among these is the Large Deformation Diffeomorphic Metric Mapping (LDDMM) framework \cite{Original_LDDMM_Beg}. Instead of treating images as objects of interest directly, LDDMM builds a space of image-matching diffeomorphisms using a Riemannian metric on velocity fields. This metric is induced by a differential operator which at once controls the nature of the metric space and regularizes the registration.
The structure of such a space~--- a manifold of smooth mappings with well-defined geodesics~--- enables generalizations of several standard statistical analysis methods. These methods adapted to the Riemannian setting has been repeatedly shown to improve their sensitivity and ability to discern population dynamics when compared to projecting the data onto Euclidean domains. Works in this area include computation of the geometric median and metric optimization for robust atlas estimation \cite{Fletcher_Geometric_Median,Zhang_Bayesian_LDDMM_atlas}, time series geodesic regression \cite{Hong_Geodesic_regression}, and principal geodesic analysis \cite{Zhang_Prob_PGA}.
With the exception of \cite{Zhang_Bayesian_LDDMM_atlas}, in the works above the metric is assumed to be fixed. Further, the metricity and Riemannian inner product with which the LDDMM space is endowed has not been used explicitly in predictive modeling up to now. In this work, we strive for two complementary aims: (1) to exploit the Riemannian metric on registration-defining velocities as a kernel in a classification task and (2) to optimize the metric to improve classification. We follow an Expectation-Maximization (EM) approach similar to \cite{Zhang_Bayesian_LDDMM_atlas}, alternating between minimizing image misalignment for kernel estimation, and optimizing model quality over the kernel parameters. In this work, we choose the kernel Fischer linear discriminant classifier for simplicity, though other predictive models are admissible in our framework as well. It is our hope that by explicit tuning the diffeomorphism metric to questions of biological interest, the carefully crafted manifold properties of LDDMM will gain greater practical utility.
Our experiments consist of synthetic 2-dimensional shape classification, as well as classifying hippocampal shapes extracted from brain MRI of the Schizconnect schizophrenia study. In both cases, the classification accuracy and ROC area under the curve (AUC) improved significantly compared to default baseline kernel parameters.
\section{Methods}
\subsection{Metric on diffeomorphisms}
\label{sec:lddmm}
The Large Deformation Diffeomorphic Metric Mapping (LDDMM) was first introduced in \cite{Original_LDDMM_Beg}. The goal of the registration is to compute a diffeomorphism $\phi\colon \Omega \rightarrow \Omega$, where the $\Omega$ is the image domain. The diffeomorphism $\phi$ is generated by the flow of a time-dependent velocity vector field $v$, defined as follows:
\begin{equation} \label{eq:diff_velocity}
\begin{gathered}
\frac{\partial \phi(t, x)}{\partial t} = v(t, \phi(t, x)),\\
\phi(0, x) = id,
\end{gathered}
\end{equation}
where $id$ is the identity transformation: $id(x) = x$, $\forall x \in \Omega$.
This equation gives a path $\phi_t\colon \Omega \rightarrow \Omega$, $t \in [0, 1]$, in the space of diffeomorphisms. Estimation of the optimal diffeomorphism via the basic variational problem in the space of smooth velocity fields $V$ on $\Omega$ takes the following form, constrained by \ref{eq:diff_velocity}:
\begin{equation} \label{eq:opt_LDDMM}
v^{*} = \argmin_{v} \left(\int\limits_0^1 \norm{v_t}^2_L \,d t + \frac{1}{\sigma^2} \norm{I_0 \circ \phi - I_1}^2\right).
\end{equation}
The required smoothness is enforced by defining the norm $\norm{\cdot}_L$ on the space $V$ of smooth velocity vector fields through a Riemannian metric $L$.
The Riemannian metric $L$ should be naturally defined by the geometric structure of the domain. The inner product $\norm{v}^2_L = \langle L v, v \rangle$ can also be thought of as a metric between images, i.e.\ the minimal diffeomorphism required to transform the appearance of $I_0$ to be as similar as possible to $I_1$. Since the diffeomorphism space is a Lie Group with respect to eq.\ref{eq:diff_velocity}, the Riemannian metric defined suggests a right-invariance property. The original LDDMM work \cite{Original_LDDMM_Beg} defines $L$ as a smooth differential self-adjoint operator $L =(\alpha \Delta + \beta E)$, where E is identity operator. Here, we choose to use an $L$ based on the biharmonic operator $L =(\alpha \Delta + \beta E)^2$, as e.g. in \cite{Mang2016DistributedMemoryLD}. The parameters $(\alpha, \beta)$ correspond to convexity and normalization terms, respectively. These parameters significantly affect the quality of the registration. It is not obvious how to select $\alpha$ and $\beta$, though they effectively define the geometric structure of the primal domain. Indeed, these are the parameters we optimize in our EM scheme below.
\subsection{Predictive Model}
\label{sec:klda
Consider a standard binary classification problem: given a sample $\bigl(x_i, y(x_i)\bigr)_{i = 1}^n$, where $y(x_i) \in \{1, -1\}$ is a class label, find a classification function $\hat{y}$ approximating the true one $y$.
One of the standard linear techniques in statistical data analysis is the Fisher's linear discriminant analysis (LDA). The kernel Fisher Discriminant Analysis (KLDA) introduced in \cite{KLDA_Original} is a generalization of the classical LDA. There are several approaches to derive more general class separability criteria. KLDA derives a linear classification in the embedding feature space (RKHS\cite{aronszajn1950theory}) induced by a kernel $k$, what corresponds to a non-linear decision function in the original, or ``input'' space. The main idea of LDA is to find a one-dimensional projection $w$ in the feature space that maximizes the between-class variance while minimizing the within-class variance. KLDA seeks an analogous projection in the embedding space, where means $(M_z)$ and covariance matrices $(\Sigma_z)$ for each class $z \in \{-1, 1\}$ are computed. The (K)LDA cost function takes the following quadratic rational form:
\begin{equation} \label{eq:opt_KLDA}
J(w) = \frac{w^T (M_1 - M_{-1}) (M_1 - M_{-1})^T w}{w^T (\Sigma_{-1} + \Sigma_1) w} = \frac{w^{T} M w}{w^T N w},
\end{equation}
where
\begin{gather*}
(M_z)_i = \frac{1}{n_z} \sum_{x_\ell : y(x_\ell) = z} k(x_i, x_\ell),\\
(\Sigma_z)_{i, j} = \frac{1}{n_z} \sum_{x_\ell : y(x_\ell) = z} k(x_i, x_\ell) k(x_j, x_\ell) - (M_z)_i (M_z)_j.
\end{gather*}
Here $n_z$ is a number of objects from class $z$ in the sample.
The solution of the problem $J(w) \rightarrow \min$ is known to be $\hat{w} = N^{-1} (M_1 - M_{-1})$. The decision function for a new observation $x$ is based on the projected distance to the training sample means, $w^T (M_z - x)$.
The M-step in an EM formulation requires a differentiable measure of model quality which in our case is the accuracy of classification. The more common approach is to formulate a probabilistic model which leads to a log-likelihood optimization. Such an approach is used e.g.\ in \cite{Zhang_Bayesian_LDDMM_atlas}. In our case, this can be done by modeling the classifier's output with a parametric distribution.
However, we found that such a formulation using the sigmoid distribution function leads to an unstable solution. Instead, we propose to use the hinge loss defined for KLDA as
\begin{equation} \label{eq:hinge_loss}
\begin{gathered}
h(x', z) = \max\{0,\, 1 - z \, y(x')\}\\
y(x') = \sum\limits_{i = 1}^n w_i \left(k(x_i, x') - \frac{(M_1)_i + (M_{-1})_i}{2}\right),
\end{gathered}
\end{equation}
where $z \in \{-1, 1\}$ is a true label for the new observation $x'$ and $k(x_i, x')$ is the inner product (i.e.\ the kernel) between $x'$ and the training observation $x_i$. While both hinge loss and log-likelihood formulations eventually lead to some locally optimal solutions on simple problems, such as our synthetic dataset, the former exhibits greater stability. For the hippocampal data, only hinge loss minimization leads to a stable solution.
\subsection{Learning the diffeomorphic metric}
\label{sec:reg+ml}
The main goal of this work is to use the registration-derived metric to classify images. Let us denote the Riemannian metric by $K_L(\alpha, \beta) = \langle Lv, v \rangle$. In practice, $\beta$ plays an insignificant role and can be fixed, as multiplication of the velocity by a constant does not change the optimization problem in LDDMM. We focus on optimizing $\alpha$, fixing $\beta = 1$ as a normalization term.
We optimize $\alpha$ in the EM framework as follows.
$\bullet$ E-step:
Register each pair of images in our training sample optimizing equation \ref{eq:opt_LDDMM} to derive $K_L(x_i, x_j)$. Define $K(x_i, x_j) = \exp\{-\gamma K_{L}(x_i, x_j)\}$
and apply KLDA using $K(x_i, x_j)$. The parameter gamma is estimated by grid search to make a computation easier, but it can be also estimated by gradient descent. Estimate the hinge loss \ref{eq:hinge_loss} given a fixed $\alpha$.
$\bullet$ M-step:
Minimize the hinge loss \ref{eq:hinge_loss} with respect to $\alpha$.
The primary computational challenge above is in the M-step. Though the decision function is non-convex with respect to $\alpha$, we seek a local minimum via gradient descent. We give the gradient direction with respect to $\theta = (\alpha,\beta)$ below, keeping in mind that $\beta$ is fixed.
\begin{equation} \label{eq:derivative_hinge_loss}
\begin{gathered}
\frac{d h(x',z)}{d \theta} = \begin{cases}
-z \frac{d y(x')}{d \theta},& \text{if } z y(x') < 1,\\
0,& \text{otherwise}.
\end{cases}
\end{gathered}
\end{equation}
Using the matrix notation $y(x') = w^T \bigl(k(x') - (M_1 + M_{-1}) / 2\bigr)$ with $k_i(x') = K(x_i, x')$ one can obtain
\begin{equation} \label{eq:derivative_hinge_loss}
\begin{gathered}
\frac{d y(x')}{d \theta} = \left(k(x') - \frac{M_1 + M_{-1}}{2}\right)^T \frac{d w}{d \theta} + w^T \left(\frac{d k(x')}{d \theta} - \frac{1}{2} \frac{d (M_1 + M_{-1})}{d \theta}\right),\\
d w = - N^{-1} (d N) N^{-1} (M_1 - M_{-1}) + N^{-1} (d M_1 - d M_{-1}),\\
\frac{d K(x_i, x')}{d \theta} = -\gamma K(x_i, x') \frac{d K_{L}(x_i, x')}{d \theta}\\
\frac{d K_{L}(x_i, x')}{d \theta} = \begin{pmatrix}
2 \bigl\langle (\alpha \Delta^2 + \beta \Delta) v_{x, x'}, v_{x, x'}\bigr\rangle,\\
2 \bigl\langle (\alpha \Delta + \beta E) v_{x, x'}, v_{x, x'}\bigr\rangle.
\end{pmatrix}
\end{gathered}
\end{equation}
The resulting algorithm requires $n \times (n - 1)$ registrations at each EM step to train, and $n$ registrations to a new image from each of $n$ images in training sample to apply.
\section{Acknowledgements}
This work was funded in part by the Russian Science Foundation grant 17-11-01390.
\label{sec:ack}
\section{Conclusion}
We have presented a method to optimize registration parameters for improved classification performance. Method exploits the geodesic distance on the space of diffeomorphisms as an image similarity measure to be learned in the fashion of traditional metric learning \cite{MetricLearningSurvey}. Our aim in this work was twofold: 1. to show that the metricity of a high dimensional space of geometric objects can be successfully used to improve predictive modeling, and (2) to suggest a means of making the sophisticated mathematical machinery of constructions such a LDDMM more useful in medical imaging practice. As a first attempt, we believe this work shows progress towards both goals. A stable LDDMM metric optimization is devised, and classification accuracy in our real-world application is indeed improved. The main drawback is the significant computational burden, as $N \times N$ training registrations are required. One approach to alleviate this problem is to lift the classification problem onto the tangent space at identity, thus requiring only $N$ training registrations to an atlas, similar to \cite{Zhang_Bayesian_LDDMM_atlas}. Other generalizations of the idea presented here are possible both in LDDMM and other metric frameworks. We hope our work will inspire these generalizations to be developed.
\label{sec:conclu}
|
\section{Introduction}\label{sec: intro}
Let $\mathcal{H}$ be a separable Hilbert space and let $A$ be a densely defined closed operator on $\mathcal{H}$. Let $\rho(A)$ denote the resolvent set of $A$. Assume that $\rho(A)\neq\emptyset$.
For $z\in \rho(A)$ the resolvent is denoted by $R_A(z)=(A-zI)^{-1}$. If for some $z\in\rho(A)$ the operator $S(z)=R_A(z)^{\ast}R_A(z)$ satisfies the spectral gap condition stated below we obtain a local growth estimate for the resolvent norm $\norm{R_A(z)}$. We apply the estimate to the question of (non-)existence of local extrema in the resolvent norm and the related question whether the level sets of $\norm{R_A(\cdot)}$ can have interior points. We give sufficient criteria for a local minimum in the resolvent norm and give a number of examples satisfying these criteria.
\begin{assumption}\label{assum}
Let $z\in\rho(A)$ be given. Assume that there exist $a(z)>0$ and $\lambda_{\rm max}(z)>a(z)$ such that $\sigma(S(z))\subseteq[0,a(z)]\cup\{\lambda_{\rm max}(z)\}$ and such that $\lambda_{\rm max}(z)$ is an eigenvalue of $S(z)$.
\end{assumption}
Note that this is an assumption on a single point $z \in \rho(A)$.
Assumption~\ref{assum} is satisfied for all $z\in\rho(A)$ in at least two generic cases. The first case is when $R_A(z_0)$ is compact for some $z_0\in \rho(A)$. Then $S(z)$ is compact and self-adjoint for all $z\in\rho(A)$.
The second case is an operator of the form $A=\alpha I+K$ where $\alpha\in\IC$ and $K$ is a compact operator. If $\dim\mathcal{H} = \infty$, $R_A(z)$ is never compact but $S(z)-|\alpha-z|^{-2} I$ is compact and self-adjoint for all $z\in\rho(A)$. In both cases, the norm of $S(z)$ is a discrete positive eigenvalue, but our results also hold true when the norm of $S(z)$ is an infinitely degenerate eigenvalue. In subsection \ref{ex-assum} we introduce a class of operators for which Assumption \ref{assum} is only satisfied on a subset of $\rho(A)$.
For $z,z'\in\IC$ we denote by $[z,z']$ the line segment from $z$ to $z'$.
\begin{theorem}\label{thm-horia}
Let $z\in\rho(A)$ be a point for which Assumption~{\rm\ref{assum}} holds.
Then there exist a constant $C>0$ and a point $z'\in \rho(A)$, $z'\neq z$, such that $[z,z']\subset\rho(A)$ and such that for every point $\zeta\in[z,z']$ we have
\begin{equation}\label{eq-thm1(i)}
\norm{R_A(\zeta)} \geq \norm{R_A(z)} + C \abs{\zeta-z}^2.
\end{equation}
\end{theorem}
It is well-known that the resolvent norm is subharmonic on $\rho(A)$, see e.g.~\cite[Theorem~4.2]{TE}, which by the maximum principle implies that it cannot have a local maximum unless it is constant in an open set.
The theorem implies that for $z\in \rho(A)$ satisfying Assumption~\ref{assum} the resolvent norm $\norm{R_A(\cdot)}$ cannot have a local maximum at $z$. Hence a level set of $\norm{R_A(\cdot)}$ cannot have $z$ as an interior point. We emphasize that the result is local since Assumption~\ref{assum} may be satisfied only in a subset of the resolvent set, see the example given in subsection~\ref{ex-assum}.
If $A$ is an unbounded closed operator with compact resolvent on a Hilbert space (or more generally a complex strictly convex Banach space) then it was proved recently in \cite[Theorem~2.2]{DS} that the resolvent level sets cannot have interior points.
An example of a closed unbounded operator on a Hilbert space with resolvent norm constant in a neighborhood of the origin was given in~\cite[Theorem ~3.2]{S2}.
Since the proof of Theorem~\ref{thm-horia} is based on $\norm{R_A(z)}^2=\norm{S(z)}=\lambda_{\rm max}(z)$, the Schur complement and perturbation theory, the non-existence of local maxima of the resolvent norm can be shown without using a maximum principle.
As a consequence of the proof we have the following two results.
\begin{corollary}\label{cor12}
Assume that there exists $z\in\rho(A)$ satisfying Assumption~{\rm\ref{assum}} with the following properties:
\begin{itemize}
\item[\rm(i)] $\ip{\psi}{R_A(z) \psi} = 0$ for every eigenvector $\psi$ corresponding to the eigenvalue $\lambda_{\rm max}(z)$ of $R_A(z)^* R_A(z)$.
\item[\rm(ii)] $\ip{\psi}{R_A(z)^2 \psi} = 0$ for at least one of these eigenvectors.
\end{itemize}
Then the resolvent norm $\norm{R_A(\cdot)}$ has a local minimum at $z$.
\end{corollary}
In subsections~\ref{sec: block diagonals}--\ref{infinite} we present examples of non-normal matrices and unbounded closed operators such that the resolvent norm has a local minimum.
\begin{corollary}\label{cor13}
Assume that there exists $z\in\rho(A)$ satisfying Assumption~{\rm\ref{assum}} with the following property:
\begin{itemize}
\item[\rm(i)] $\ip{\psi}{R_A(z) \psi} \neq 0$ for an eigenvector $\psi$ corresponding to the eigenvalue $\lambda_{\rm max}(z)$ of $R_A(z)^* R_A(z)$.
\end{itemize}
Then there exist a constant $C>0$ and a point $z'\in \rho(A),\ z'\neq z,$ such that $[z,z']\subset\rho(A)$ and such that for every point $\zeta\in[z,z']$ we have
\begin{equation}\label{cor1}
\norm{R_A(\zeta)} \geq \norm{R_A(z)} + C \abs{\zeta-z}.
\end{equation}
\end{corollary}
If $A$ is normal and has compact resolvent, then condition (i) in Corollary~\ref{cor13} holds for all $z\in\rho(A)$. We note that in this case there is a simple direct proof of \eqref{cor1}, see Remark~\ref{normal-remark}. There may exist local minima in the case where the conditions in Corollary~\ref{cor12} are not satisfied. An example is the normal $3\times 3$-matrix $A=\diag(1,\e^{2\I\pi/3},\e^{-2\I\pi/3})$. Since $A$ is normal we have $\norm{R_A(z)}=1/\dist(z,\sigma(A))$. Then it is geometrically obvious that $z=0$ is a local minimum.
In the case $\dim \mathcal{H}=2$ the resolvent norm cannot have any local extremum at all.
This result is stated as follows:
\begin{theorem} \label{thm-henrik}
Let $A\in \IC^{2\times 2}$.
Then the resolvent norm $\norm{R_A(\cdot)}$ is symmetric with respect to $\tr(A)/2$ which is the average of the eigenvalues of $A$, i.e.
\begin{equation*}
\norm{R_A(\tr(A)/2+z)} = \norm{R_A(\tr(A)/2-z)}.
\end{equation*}
Furthermore,
\begin{enumerate}[\rm(i)]
\item If $A$ has an eigenvalue with algebraic multiplicity $2$, then $\norm{R_A(\cdot)}$ is a strictly decreasing radial function with center at the eigenvalue $\tr(A)/2$.
\item If $A$ has distinct eigenvalues then $\norm{R_A(\cdot)}$ has a saddle point at $\tr(A)/2$ and the following results hold.
\begin{enumerate}[\rm(a)]
\item If $A$ is normal, then the critical points consist of a line through $\tr(A)/2$ perpendicular to the line through the two eigenvalues of $A$. $\norm{R_A(\cdot)}$ is not real-differentiable on this line.
\item If $A$ is not normal then $\norm{R_A(\cdot)}$ is real-differentiable on $\rho(A)$ and $\tr(A)/2$ is the only critical point.
\end{enumerate}
None of the critical points are local extrema.
\end{enumerate}
\end{theorem}
We can use the result in Theorem~\ref{thm-horia} to give a result on the pseudospectra. We recall the definition. For $\varepsilon>0$ the $\varepsilon$-pseudospectrum of $A$ is defined as
\begin{equation*}
\sigma_\varepsilon(A)=\{z\in \IC\,|\, \norm{R_A(z)} > \varepsilon^{-1}\}.
\end{equation*}
See~\cite{TE} for further information on pseudospectra.
We use the convention that $\norm{R_A(z)} = \infty$ if $(A - zI)$ is not invertible. We state the result in the finite dimensional case.
\begin{theorem}\label{path}
Let $A\in\IC^{N\times N}$. Then any point $z\in \sigma_\varepsilon(A)$ can be linked to one of the eigenvalues of $A$ through a finite polygonal path contained in $\sigma_\varepsilon(A)$.
\end{theorem}
Note that this theorem implies the well-known result \cite[Theorem 2.4]{TE} that there must be at least one eigenvalue of $A$ in each connected component of $\sigma_{\varepsilon}(A)$.
Thus a pseudospectrum consists of at most $J$ connected components where $J$ is the number of distinct eigenvalues of $A$.
The novelty here is that we prove this fact by constructing a path inside a given pseudospectrum which connects any $z\in\sigma_{\varepsilon}(A)$ to an eigenvalue. This constructive approach is a consequence of the local growth estimate in \eqref{eq-thm1(i)}.
This article is organized as follows. The proofs of Theorems~\ref{thm-horia} and~\ref{thm-henrik} are given in sections~\ref{sec: proof horia} and~\ref{sec: proof henrik}. We construct the polygonal path for Theorem~\ref{path} in section~\ref{sec: connected components}.
In section~\ref{sec: local minima} we provide the examples already mentioned above. In subsection~\ref{ex-assum} we give a class of operators for which Assumption~\ref{assum} only holds in a proper subset of the resolvent set. The examples in subsections~\ref{sec: block diagonals} and~\ref{sec: general example} are non-normal matrices satisfying the conditions in Corollary~\ref{cor12} such that the resolvent norm has a local minimum at the origin. In subsection~\ref{infinite} we finally give examples of unbounded closed operators with or without compact resolvent that satisfy Assumption~\ref{assum} such that the resolvent norm has a local minimum at the origin.
\section{Proof of Theorem~\ref{thm-horia}}%
\label{sec: proof horia}
Let $z\in\rho(A)$ such that Assumption~{\rm\ref{assum}} holds. This $z$ is fixed throughout the proof. As above we let $S(z)=R_A(z)^{\ast}R_A(z)$. Note that $S(z)$ is a bounded, self-adjoint, and strictly positive operator.
For $\zeta\in\IC$ we introduce the notation $\Delta\zeta=\zeta-z$. Fix $\delta_1>0$ such that $\abs{\Delta\zeta}\leq\delta_1$ implies $\zeta\in\rho(A)$. Then a simple computation shows that we have
\begin{equation}\label{eq1}
\norm{S(\zeta)-S(z)}\leq C\abs{\Delta\zeta}, \quad \abs{\Delta\zeta}\leq\delta_1,
\end{equation}
where $C$ depends on $z$ and $\delta_1$.
Take $\delta_2=\frac12(\lambda_{\rm max}(z)-a(z))$ with $a(z)$ from Assumption~\ref{assum}. We can find $\delta_3>0$ such that for all $\abs{\Delta\zeta}\leq\delta_3$ and all $\lambda$ with $\abs{\lambda-\lambda_{\rm max}(z)}=\delta_2$ we have $\lambda\in\rho(S(\zeta))$ and
\begin{align}
(S(\zeta)-\lambda I)^{-1}&=(S(z)-\lambda I)^{-1}\bigl[
I+(S(\zeta)-S(z))(S(z)-\lambda I)^{-1}\bigr]^{-1}\notag\\
&=(S(z)-\lambda I)^{-1}\notag\\
&\quad-(S(z)-\lambda I)^{-1}(S(\zeta)-S(z))(S(z)-\lambda I)^{-1}\notag\\
&\qquad\cdot\bigl[
I+(S(\zeta)-S(z))(S(z)-\lambda I)^{-1}\bigr]^{-1}.\label{neumann}
\end{align}
We now use some standard arguments from perturbation theory, see~\cite{Kato}. Note that the map $z\mapsto S(z)$ is norm continuous but not analytic.
We define the Riesz projections
\begin{equation}
P(\zeta)=\frac{-1}{2\pi \I}\int_{\abs{\lambda-\lambda_{\rm max}(z)}=\delta_2} (S(\zeta)-\lambda I)^{-1}\di\lambda.
\end{equation}
We write $P=P(z)$, which is the eigenprojection of the eigenvalue $\lambda_{\rm max}(z)$.
Using \eqref{neumann} we can find a $\delta_4$, $0<\delta_4\leq\delta_3$, such that for all $\abs{\Delta\zeta}\leq\delta_4$ we have $\norm{P(\zeta)-P}<1$. There exists a family of unitary operators $U(\zeta)\colon\ran P(\zeta)\to\ran P$ such that
$U(\zeta)P(\zeta)=P U(\zeta)$, see~\cite[I-\S6.8]{Kato}. Note that this result holds in infinite dimensions. Together with the upper semi-continuity of the parts of the spectrum of $S(\zeta)$, see~\cite[IV-\S3.4]{Kato}, we conclude that
$\norm{S(\zeta)}=\norm{S(\zeta)P(\zeta)}$.
Let us define
\begin{equation}\label{defD}
D=\{\lambda\in \IC\,|\, \abs{\lambda-\lambda_{\rm max}(z)}<\delta_2\},
\end{equation}
such that $\sigma(S(\zeta)P(\zeta))=\sigma(S(\zeta))\cap D$.
We have
\begin{equation}
\dist(\lambda_{\rm max}(z),\sigma(S(\zeta)P(\zeta)))\leq C\abs{\Delta\zeta},\quad
\abs{\Delta\zeta}\leq\delta_4.
\end{equation}
Let $P^{\perp}=I-P$. Then for all $\lambda$ satisfying $\abs{\lambda-\lambda_{\rm max}(z)}\leq C\abs{\Delta\zeta}$ and all $\abs{\Delta\zeta}\leq\delta_4$ we see that $P^{\perp}(S(\zeta)-\lambda I)P^{\perp}$ is invertible in $\ran P^{\perp}$. We now use the Schur complement based on $P$ and $P^{\perp}$, also known as the Feshbach formula, see e.g.\ \cite[Equations~(6.1)-(6.2)]{Nen}.
We have that $\lambda\in\rho(S(\zeta))$ if and only if the Schur complement
\begin{equation}
F(\zeta,\lambda)=PS(\zeta)P-\lambda P+
PS(\zeta)P^{\perp}\bigl(\lambda P^{\perp}-P^{\perp} S(\zeta)P^{\perp}\bigr)^{-1}P^{\perp} S(\zeta)P
\end{equation}
is invertible in $\ran P$. In the affirmative case we have
\begin{equation}\label{inverseF}
F(\zeta,\lambda)^{-1}=P(S(\zeta)-\lambda I)^{-1}P.
\end{equation}
We note that with
$\abs{\lambda-\lambda_{\rm max}(z)}\leq C \abs{\Delta\zeta}$
we have
\begin{equation}\label{perp-comp}
\norm{
\bigl(P^{\perp}(S(\zeta)-\lambda I)P^{\perp}\bigr)^{-1}-
\bigl(P^{\perp}(S(z)-\lambda_{\rm max}(z) I)P^{\perp}\bigr)^{-1}
}\leq C\abs{\Delta\zeta},
\end{equation}
where the norm is the operator norm on $\ran P^{\perp}$.
\smallskip
Let $M$ and $N$ be two non-empty compact sets in $\IC$.
The Hausdorff distance between $M$ and $N$ is defined as
\begin{equation*}
d_{\rm H}(M,N) = \max\bigl\{\sup_{\mu\in N} \dist(\mu,M),\; \sup_{\mu\in M} \dist(\mu,N)\bigr\}.
\end{equation*}
The geometric interpretation is that given any $\mu\in M$ we can find at least one
$\nu\in N$ such that $\abs{\mu-\nu} \leq d_{\rm H}(M,N)$, and vice versa.
\begin{lemma}\label{lemma1}
Let $\zeta \in \rho(A)$ and write $\Delta S(\zeta) = S(\zeta) - S(z)$.
Define the operator
\begin{align*}
W(\zeta)&= \lambda_{\rm max}(z) P + P\Delta S(\zeta)P\\
&\quad
+P\Delta S(\zeta)P^{\perp} \bigl(\lambda_{\rm max}(z)P^{\perp}-P^{\perp} S(z)P^{\perp} \bigr)^{-1} P^{\perp} \Delta S(\zeta)P \end{align*}
on $\ran P$.
Then the spectrum of $W(\zeta)$ is at a Hausdorff distance of order $\abs{\Delta\zeta}^3$ from $\sigma(S(\zeta))\cap D$.
\end{lemma}
\begin{proof}
Let $\zeta \in \rho(A)$.
In the first part of the proof we will show that if $\lambda\in D$ is located at a distance larger than some constant times $\abs{\Delta\zeta}^3$ from the spectrum of $S(\zeta)$ then it must belong to the resolvent set of $W(\zeta)$.
In other words, no points of $D$ belonging to the spectrum of $W(\zeta)$ can be at a distance larger than $C \abs{\Delta\zeta}^3$ from the spectrum of $S(\zeta)$.
Indeed, let us assume that $\dist(\lambda,\sigma(S(\zeta)))>0$.
Due to \eqref{inverseF} and the self-adjointness of $S(\zeta)$ we have
\begin{equation*}
\norm{F(\zeta,\lambda)^{-1}} \leq \frac{1}{\dist(\lambda,\sigma(S(\zeta)))}.
\end{equation*}
Note that since {$P^{\perp} S(\zeta)P= P^{\perp} \Delta S(\zeta)P$
and $P S(\zeta)P^\perp=P \Delta S(\zeta)P^\perp$,}
the norms of the off-diagonal components $P^\perp S(\zeta)P$ and $P S(\zeta)P^\perp$ are of order $\abs{\Delta \zeta}$ by \eqref{eq1}, which combined with \eqref{perp-comp} gives the estimate
\begin{equation*}
\norm{F(\zeta,\lambda)-(W(\zeta)-\lambda P)} \leq C \abs{\Delta \zeta}^3.
\end{equation*}
Hence
\begin{equation*}
W(\zeta)-\lambda P = \bigl( I -[F(\zeta,\lambda)-(W(\zeta)-\lambda P)]F(\zeta,\lambda)^{-1}\bigr)F(\zeta,\lambda)
\end{equation*}
is invertible in $\ran (P)$ if
\begin{equation*}
\frac{C \abs{\Delta \zeta}^3}{\dist(\lambda,\sigma(S(\zeta)))}<1.
\end{equation*}
Thus if {$\lambda\in D$} and $\dist(\lambda,\sigma(S(\zeta))) > C \abs{\Delta \zeta}^3$, then $\lambda$ is not in the spectrum of $W(\zeta)$ and {the first part of the proof is finished}.
Now we prove the second part, i.e.\ we show that any point $\lambda\in D$ which is located at a distance larger than $C \abs{\Delta \zeta}^3$ from the spectrum of $W(\zeta)$ must belong to the resolvent set of $S(\zeta)$.
Indeed, let us assume that $\dist(\lambda,\sigma(W(\zeta)))>0$.
Then
\begin{equation*}
F(\zeta,\lambda)=\bigl(I +[F(\zeta,\lambda)-(W(\zeta)-\lambda P)](W(\zeta)-\lambda P)^{-1}\bigr)(W(\zeta)- \lambda P)
\end{equation*}
is invertible if
\begin{equation*}
\frac{C \abs{\Delta \zeta}^3}{\dist(\lambda,\sigma(W(\zeta)))}<1.
\end{equation*}
Here we used that $W(\zeta)$ is self-adjoint.
We conclude that $F(\zeta,\lambda)$, and therefore $S(\zeta)-\lambda I$, is invertible for {such $\lambda$'s, hence no element of $D$ which belongs to the spectrum of $S(\zeta)$ can be located at a distance larger than $C \abs{\Delta \zeta}^3$ from the spectrum of $W(\zeta)$}.
\end{proof}
The following proposition is a direct consequence of Lemma~\ref{lemma1}. We abuse notation slightly and write $\lambda_{\rm max}(\zeta)=\norm{S(\zeta)}$. We emphasize that in the case where $\lambda_{\rm max}(z)$ has infinite multiplicity $\lambda_{\rm max}(\zeta)$ need not be an eigenvalue of $S(\zeta)$.
\begin{proposition}\label{lemma2}
We have that
\begin{equation}\label{hc54}
\bigl\lvert\lambda_{\rm max}(\zeta)-\norm{W(\zeta)}\bigr\rvert \leq C \abs{\Delta \zeta}^3.
\end{equation}
\end{proposition}
\begin{proof}
If $\abs{\Delta\zeta}$ is small enough, then both $\norm{W(\zeta)}$ and $\lambda_{\rm max}(\zeta)$ belong to $D$.
Assume without loss of generality that $\lambda_{\rm max}(\zeta) > \norm{W(\zeta)}$.
Since $\norm{W(\zeta)}$ is the element of $\sigma(W(\zeta))$ which is closest to $\lambda_{\rm max}(\zeta)$ we have
\begin{equation*}
0 < \lambda_{\rm max}(\zeta)-\norm{W(\zeta)} \leq d_{\rm H}\big (\sigma(W(\zeta)){\cap D},\sigma(S(\zeta)){\cap D}\big ) \leq C \abs{\Delta \zeta}^3.
\end{equation*}
\end{proof}
\begin{proposition}\label{propo9}
There exist $C>0$ and a point $z'\in\rho(A)\setminus\{z\}$ such that $[z,z']\subset\rho(A)$ and $\lambda_{\rm max}(\zeta)-\lambda_{\rm max}(z) \geq C\abs{\zeta-z}^{2}$ for every $\zeta \in [z,z']$.
\end{proposition}
\begin{remark}
Recall that $\norm{R_A(z)}^2 = \lambda_{\rm max}(z)$ and $\norm{R_A(\zeta)}^2 = \lambda_{\rm max}(\zeta)$.
If {$\abs{z'-z}$} is small enough we have $\norm{R_A(\zeta)} \leq 2 \norm{R_A(z)}$,
$\zeta\in[z,z']$.
This implies for $\zeta\in[z,z']$
\begin{equation*}
\norm{R_A(\zeta)} - \norm{R_A(z)} = \frac{\norm{R_A(\zeta)}^2
-\norm{R_A(z)}^2}{\norm{R_A(\zeta)} + \norm{R_A(z)}} \geq
\frac{\lambda_{\rm max}(\zeta)-\lambda_{\rm max}(z)}{3 \norm{R_A(z)}} \geq C \abs{z-\zeta}^2.
\end{equation*}
Thus this proposition implies Theorem~\ref{thm-horia}.
\end{remark}
\begin{proof}
Use the Taylor expansion to get
\begin{equation*}
R_A(\zeta) = R_A(z) + (\Delta\zeta)R_A(z)^2 + (\Delta\zeta)^2R_A(z)^3 + \mathcal{O}(\abs{\Delta \zeta}^3).
\end{equation*}
Let $\Delta S(\zeta)$ and $W(\zeta)$ be defined as in Lemma~\ref{lemma1}.
Then we get
\begin{align}\label{hc10}
\Delta S(\zeta) &= (\Delta \zeta) S(z) R_A(z) + (\overline{\Delta\zeta}) R_A(z)^* S(z)\nonumber \\
&\quad +(\Delta\zeta)^2 S(z) R_A(z)^2 + (\overline{\Delta\zeta})^2 (R_A(z)^*)^2 S(z)\nonumber \\
&\quad +\abs{\Delta \zeta}^2 R_A(z)^* S(z) R_A(z) + \mathcal{O}(\abs{\Delta \zeta}^3).
\end{align}
Next we introduce the operator
\begin{align}\label{hc12}
\widetilde{W}(\zeta) &= \lambda_{\rm max}(z) P + \lambda_{\rm max}(z) (\Delta\zeta) P R_A(z) P + \lambda_{\rm max}(z) (\overline{\Delta\zeta}) P R_A(z)^* P \nonumber \\
&\phantom{=.} + \lambda_{\rm max}(z) (\Delta\zeta)^2 PR_A(z)^2 P + \lambda_{\rm max}(z) (\overline{\Delta\zeta})^2 P (R_A(z)^*)^2 P\nonumber \\
&\phantom{=.} + \abs{\Delta\zeta}^2 P R_A(z)^* S(z) R_A(z) P \nonumber \\
&\phantom{=.} + P \Delta S(\zeta) P^\perp \bigl(\lambda_{\rm max}(z) P^\perp - P^\perp S(z) P^\perp \bigr)^{-1} P^\perp \Delta S(\zeta) P.
\end{align}
The estimate
\begin{equation*}
\norm{W(\zeta)-\widetilde{W}(\zeta)} \leq C \abs{\Delta\zeta}^3
\end{equation*}
yields that the Hausdorff distance between the spectra of $W(\zeta)$ and $\widetilde{W}(\zeta)$ is of order $\abs{\Delta\zeta}^3$.
Using Proposition~\ref{lemma2} we conclude that $\lambda_{\rm max}(\zeta)$ is at a distance of order $\abs{\Delta \zeta}^3$ from $\norm{\widetilde{W}(\zeta)}$.
The main idea is now to find $z' \in \rho(A)$ such that for every $\zeta \in [z,z']$,
\begin{equation*}
\norm{\widetilde{W}(\zeta)} \geq \lambda_{\rm max}(z) + C_1 \abs{\Delta \zeta} + C_2 \abs{\Delta \zeta}^2
\end{equation*}
where $C_1, C_2\geq0$ with $\max\{C_1,C_2\}>0$.
This will prove that for $\abs{z-z'}$ small enough, either $(\lambda_{\rm max}(\zeta)-\lambda_{\rm max}(z))/\abs{\Delta\zeta}$ or $(\lambda_{\rm max}(\zeta)-\lambda_{\rm max}(z))/\abs{\Delta\zeta}^{2}$ is bounded from below by a positive constant on $[z,z']$
First assume that there exists $\psi=P\psi$ with $\norm{\psi}=1$, such that
\begin{equation*}
\ip{\psi}{R_A(z) \psi} = \eta \e^{\I\varphi},\quad \eta>0,\quad \varphi\in [0,2\pi[.
\end{equation*}
Now choose $\zeta\in [z,z']$ with $z'=z+r\e^{-\I\varphi}$.
Then
\begin{equation}\label{order1}
\norm{\widetilde{W}(\zeta)} \geq \ip{\psi}{\widetilde{W}(\zeta)\psi} \geq \lambda_{\rm max}(z) + 2\lambda_{\rm max}(z) \abs{\Delta\zeta} \eta + \mathcal{O}(\abs{\Delta\zeta}^2).
\end{equation}
In this case we can choose $C_1 = 2 \lambda_{\rm max}(z) \eta$, $C_2=0$, and we are done.
Next assume that $\ip{\psi}{R_A(z)\psi} =0$ for all $\psi=P\psi$ of norm one.
Using
\begin{equation*}
\ip{\Delta S(\zeta)\psi}{P^\perp \bigl(\lambda_{\rm max}(z) P^\perp - P^\perp S(z)P^\perp \bigr)^{-1} P^\perp \Delta S(\zeta)\psi} \geq 0
\end{equation*}
in \eqref{hc12} we obtain
\begin{align}\label{hc13}
\norm{\widetilde{W}(\zeta)} &\geq \ip{\psi}{\widetilde{W}(\zeta)\psi}\nonumber \\
&\geq \lambda_{\rm max}(z)+2\lambda_{\rm max}(z) \re \bigl( (\Delta\zeta)^2 \ip{\psi}{R_A(z)^2 \psi} \bigr)\nonumber \\
&\quad + \abs{\Delta \zeta}^2 \ip{R_A(z) \psi}{S(z) R_A(z)\psi}.
\end{align}
Let $C_2=\ip{R_A(z) \psi}{S(z)R_A(z)\psi}$. Since $S(z)$ is strictly positive and $R_A(z)\psi\neq0$, we get $C_2>0$.
This result inserted into \eqref{hc13} gives
\begin{equation*}
\norm{\widetilde{W}(\zeta)} \geq \lambda_{\rm max}(z)+2\lambda_{\rm max}(z) \re \bigl((\Delta\zeta)^2\ip{\psi}{R_A(z)^2\psi} \bigr) + C_2\abs{\Delta \zeta}^2.
\end{equation*}
Write
\begin{equation*}
\ip{\psi}{R_A(z)^2 \psi} = \eta \e^{\I\varphi}, \quad \eta\geq 0,\quad \varphi\in [0,2\pi[.
\end{equation*}
Then choosing $\zeta\in[z,z']$ with $z'=z+r\e^{-\I\varphi/2}$ we get
$\re \bigl((\Delta\zeta)^2\ip{\psi}{R_A(z)^2\psi} \bigr)\geq0$ and then
\begin{equation}\label{estimate}
\norm{\widetilde{W}(\zeta)} \geq \lambda_{\rm max}(z) + C_2\abs{\Delta\zeta}^2.
\end{equation}
It remains to consider the case when $\mathcal{H}$ is finite dimensional and $\lambda_{\max}(z)$ is the only eigenvalue of $S(z)$. In this case $P=I$ and $P^{\perp}=0$. If one omits all terms involving $P^{\perp}$ in the proof above then it is valid also in this case.
This concludes the proof of Theorem~\ref{thm-horia}.
\end{proof}
Corollary~\ref{cor13} is an immediate consequence of the proof above. If the conditions in Corollary~\ref{cor12} are satisfied
and we take $\psi$ satisfying (ii), then we get the estimate \eqref{estimate} for any choice of direction, such that $z$ is a local minimum point of the resolvent norm.
\begin{remark}\label{normal-remark}
We note that in the case $A$ normal with compact resolvent there is a simple direct proof of the estimate~\eqref{cor1} in Corollary~\ref{cor13}. Let $z\in\rho(A)$. Since $\sigma(A)$ is discrete, there exists $\lambda\in\sigma(A)$ such that $\dist(z,\sigma(A))=\abs{\lambda-z}$ and therefore $\{z+t(\lambda-z) \,|\, t\in[0,1[\,\}\subset\rho(A)$. Take $z'\neq z$ in this set, sufficiently close to $z$. Since $A$ is normal, we have
\begin{equation}
\norm{R_A(z)}=\frac{1}{\dist(z,\sigma(A))}.
\end{equation}
Thus the result follows by a simple geometrical argument.
We can also prove the estimate~\eqref{cor1} for $A$ normal with compact resolvent by verifying the condition (i) in Corollary~\ref{cor13}. We have that $R_A(z)$ is normal for all $z\in\rho(A)$. Fix $z\in\rho(A)$. Assume that $\ip{\psi}{R_A(z)\psi}=0$ for all $\psi\in\ran P$. By polarization we get that $\ip{\psi'}{R_A(z)\psi}=0$ for all $\psi',\psi\in\ran P$. Since $R_A(z)$ and $R_A(z)^{\ast}$ commute we get that $R_A(z)\colon\ran P\to\ran P$ is an isomorphism. Take $\psi'=R_A(z)\psi$. Then it follows from $\ip{\psi'}{R_A(z)\psi}=0$ that $\psi=0$, a contradiction. Thus there exists $\psi\in\ran P$ with
$\ip{\psi}{R_A(z)\psi}\neq0$.
\end{remark}
\section{Proof of Theorem~\ref{thm-henrik}}%
\label{sec: proof henrik}
Let
\begin{equation*}
A = \begin{bmatrix}
a & b \\
c & d
\end{bmatrix} \in \IC^{2\times 2}.
\end{equation*}
Let $z\in\rho(A)$ and set $T(z) = (A-zI)^*(A-zI)$. Define
\begin{align*}
w(z) &= \tr(T(z)) = \sum_{i,j=1}^2\abs{(A-zI)_{ij}}^2 = \abs{a-z}^2+\abs{d-z}^2+\abs{b}^2 + \abs{c}^2, \\
h(z) &= \det(T(z)) = \abs{\det(A-zI)}^2 = \abs{z^2-\tr(A)z+\det(A)}^2.
\end{align*}
The resolvent norm $\norm{R_A(z)}$ equals the reciprocal to the smallest singular value $s(z)$ of $A-zI$, i.e.\ $s(z)^2$ is the smallest eigenvalue of the positive definite matrix $T(z)$.
This leads to the expression
\begin{equation*}
\norm{R_A(z)}^2 = \frac{1}{s(z)^2} = \frac{2}{w(z)-\sqrt{w(z)^2-4h(z)}}.
\end{equation*}
The average of the eigenvalues of $A$ is $\tr(A)/2$, thus
\begin{equation*}
A_1 = A - \frac{\tr(A)}{2}I
\end{equation*}
has eigenvalues $\pm \lambda$ for some $\lambda\in \IC$.
We consider the two cases (i) $\lambda = 0$ and (ii) $\lambda \neq 0$ separately.
\paragraph{Case (i)}
We start with the case $\lambda=0$. Then $\tr(A_1) = \det(A_1) = 0$ which implies that either
\begin{equation}
A_1 = \begin{bmatrix}
0 & 0 \\ c & 0
\end{bmatrix}
\qquad\text{or}\qquad
A_1 = \begin{bmatrix}
a & b \\ -\frac{a^2}{b} & -a
\end{bmatrix}
\label{eq:type1}
\end{equation}
with $a,b,c\in\IC$ and $b \neq 0$.
The corresponding expressions for $w$ and $h$ for the matrices in \eqref{eq:type1} are $w(z) = 2t+k$ and $h(z) = t^2$ where $t = \abs{z}^2$ and $k = \abs{c}^2$, respectively $k = 2\abs{a}^2+\abs{b}^2+\frac{\abs{a}^4}{\abs{b}^2}$.
Thus in both cases $\norm{R_{A_1}(\cdot)}$ is a radial function
\begin{equation*}
\norm{R_{A_1}(z)}^2 = \frac{2}{2t+k-(4kt+k^2)^{1/2}}.
\end{equation*}
For $k = 0$ we have $\norm{R_{A_1}(z)} = \abs{z}^{-1}$, so we may assume $k>0$. The derivative is
\begin{equation*}
\frac{d}{dt}\norm{R_{A_1}(z)}^2 = \frac{-4+4k(4kt+k^2)^{-1/2}}{(2t+k-(4kt+k^2)^{1/2})^2}.
\end{equation*}
As the eigenvalues of $A_1$ are 0 then $t>0$ and $k>0$, i.e.\ we have $\frac{d}{dt}\norm{R_{A_1}}^2 < 0$.
Since the resolvent for $A$ is given by
\begin{equation}
R_A(z) = R_{A_1}(z-\tfrac{1}{2}\tr(A)) \label{eq:resnormtype1}
\end{equation}
we conclude that $\norm{R_A(\cdot)}$ is a radial function with center at its eigenvalue $\tr(A)/2$ and is strictly decreasing away from $\tr(A)/2$.
\paragraph{Case (ii)} Assume $\lambda\neq 0$. As the eigenvalues of $A_1$ are distinct
\begin{equation*}
A_2 = \frac{1}{\lambda}A_1
\end{equation*}
has eigenvalues $\pm 1$. Thus $\tr(A_2) = 0$ and $\det(A_2) = -1$ whence either
\begin{equation}
A_2 = \pm\begin{bmatrix}
1 & 0 \\ c & -1
\end{bmatrix}
\qquad\text{or}\qquad
A_2 = \begin{bmatrix}
a & b \\ \frac{1-a^2}{b} & -a
\end{bmatrix}
\label{eq:type2}
\end{equation}
with $a,b,c\in\IC$ and $b \neq 0$.
The corresponding expressions for $w$ and $h$ are $w(z) = 2\abs{z}^2 + k$ and $h(z) = \abs{z}^4+1-z^2-\overline{z}^2$ where $k = 2+\abs{c}^2$ and $k = 2\abs{a}^2+\abs{b}^2+\frac{\abs{1-a^2}^2}{\abs{b}^2}$, respectively.
Note that since $w^2-4h\geq 0$ a priori, $k\geq 2$. In fact $k = 2$ if and only if $A_2$ is self-adjoint, {which is the case if and only if $A$ is normal.}
For $k = 2$ we have for $z = x_1 + \I x_2$ with $x_1,x_2\in\IR$,
\begin{equation*}
\norm{R_{A_2}(z)}^2 = \frac{1}{(1-\abs{x_1})^2+x_2^2}.
\end{equation*}
This function is not differentiable at $x_1=0$, but clearly it increases away from the imaginary axis for each fixed $x_2$ between $z = \pm 1+\I x_2$, and decreases away from the origin on the imaginary axis.
It is easily checked that there are no critical points for $\abs{x_1}>0$ as $\pm 1\in\sigma(A_2)$.
In particular, $z=0$ is a saddle point of $\norm{R_{A_2}(\cdot)}$ and there are no local extrema.
Now assume $k>2$.
Then
\begin{align}
\norm{R_{A_2}(z)}^2 &= \frac{2}{2\abs{z}^2+k-((k+2)(k-2)+4(k\abs{z}^2+z^2+\overline{z}^2))^{1/2}} \notag\\
&= \frac{2}{2x_1^2+2x_2^2+k-((k+2)(k-2)+4(k+2)x_1^2+4(k-2)x_2^2)^{1/2}}. \label{eq:resnormA2}
\end{align}
It is evident that $\norm{R_{A_2}(\cdot)}$ is symmetric with respect to the origin.
By a straightforward direct calculation
\begin{align*}
\frac{\partial}{\partial x_1}\norm{R_{A_2}(z)}^2 &= 2x_1\norm{R_{A_2}(z)}^4 ((k+2)(w(z)^2-4h(z))^{-1/2}-1), \\
\frac{\partial}{\partial x_2}\norm{R_{A_2}(z)}^2&= 2x_2\norm{R_{A_2}(z)}^4 ((k-2)(w(z)^2-4h(z))^{-1/2}-1).
\end{align*}
Note that as $k>2$ then $w^2-4h\geq (k+2)(k-2) > (k-2)^2$.
Thus we must have $x_2 = 0$ at a critical point.
If $x_2 = 0$, then $w^2-4h = (k+2)^2$ if and only if $x_1 = \pm 1$.
However, as $z = \pm 1 \in \sigma(A_2)$ the only critical point of $\norm{R_{A_2}(\cdot)}$ is at $z = 0$.
From \eqref{eq:resnormA2} and writing $z = \sqrt{t}\e^{\I\theta}$, we get
\begin{equation*}
\norm{R_{A_2}(z)}^2 = \frac{2}{2t+k-((k+2)(k-2)+4t(k+2\cos(2\theta)))^{1/2}}.
\end{equation*}
From this we immediately obtain
\begin{align*}
\frac{d}{dt}\norm{R_{A_2}(z)}^2 &= \norm{R_{A_2}(z)}^4 ((k+2\cos(2\theta))(w(z)^2-4h(z))^{-1/2}-1) \\
&= \frac{4\norm{R_{A_2}(z)}^4 (k+2\cos(2\theta))
(w(z)^2-4h(z))^{-1/2}}{k+2\cos(2\theta)+(w(z)^2-4h(z))^{1/2}}(g(k,\theta)-t)
\end{align*}
where
\begin{equation*}
g(k,\theta) = \frac{1}{4}(k+2\cos(2\theta)) - \frac{(k+2)(k-2)}{4(k+2\cos(2\theta))}.
\end{equation*}
For $k>2$ all the terms in $\frac{d}{dt}\norm{R_{A_2}(z)}^2$ are positive except $g(k,\theta)-t$.
As a consequence, $g(k,\theta)$ determines the sign of the derivative
\begin{equation}
\begin{cases}
\dfrac{d}{dt}\norm{R_{A_2}(z)}^2 > 0, & 0<t<g(k,\theta), \\[10pt]
\dfrac{d}{dt}\norm{R_{A_2}(z)}^2 < 0, & t > \max\{g(k,\theta),0\}.
\end{cases} \label{eq:deriv}
\end{equation}
Since $k > 2$ we have in particular $\frac{d}{dt}\norm{R_{A_2}(x_1)}^2>0$ for $0< t < 1$ and $\frac{d}{dt}\norm{R_{A_2}(\I x_2)}^2<0$ for $t>0$.
Thus $\norm{R_{A_2}(\cdot)}$ is increasing away from the origin on the real axis between $z=\pm 1$ and decreasing away from the origin on the imaginary axis.
The resolvent for $A$ is now given by
\begin{equation}
{R_A(z) = \lambda^{-1}R_{A_2}\bigl( \lambda^{-1}(z-\tfrac{1}{2}\tr(A))\bigr).} \label{eq:resnormtype2}
\end{equation}
This implies that for $k>2$, $\norm{R_A(\cdot)}$ has exactly one critical point at $\tr(A)/2$ which is a saddle point, and is furthermore symmetric with respect to the point $\tr(A)/2$.
{For $k=2$ (when $A$ is normal) we also have a line of critical points where $\norm{R_A(\cdot)}$ is non-differentiable, as stated in the theorem, none of which are local extrema.}
\section{Proof of Theorem~\ref{path}}%
\label{sec: connected components}
Assume that $A\in \IC^{N\times N}$ has the distinct eigenvalues $\lambda_1,\ldots, \lambda_J$, $1\leq J\leq N$. Let $\psi_j$ denote a normalized eigenvector corresponding to the eigenvalue $\lambda_j$, $j=1,\ldots,J$.
Then we have for any $z \in \rho(A)$ and $j=1,\ldots,J$,
$\norm{R_A(z)} \geq \norm{R_A(z) \psi_j} = 1/{\abs{\lambda_j-z}}$.
Thus
\begin{equation*}
\norm{R_A(z)} \geq \frac{1}{\dist(z,\sigma(A))}.
\end{equation*}
For $\epsilon>0$ this estimate implies
$B_{\varepsilon/2}(\lambda_j) \subset \sigma_\varepsilon(A)$ for each $j=1,\ldots,J$.
We have that $\sigma_{\varepsilon}(A)$ is bounded,
since $\norm{R_A(z)}\to0$ as $\abs{z}\to \infty$.
Thus $\overline{\sigma_\varepsilon(A)}$ is compact.
If $z \in \sigma(A)$ then $\norm{R_A(z)}$ is interpreted to be $\infty$. Actually we have $\overline{\sigma_\varepsilon(A)}=\{z\in\IC \,|\, \norm{R_A(z)} \geq 1/{\varepsilon}\}$; see e.g.\ \cite{CCH,DS,H,S} for this result and for results on $\{z\in\IC \,|\, \norm{R_A(z)} = 1/{\varepsilon}\}$.
Fix $\varepsilon>0$. Define $f\colon \overline{\sigma_\varepsilon(A)} \to [\varepsilon^{-1},\infty]$ by $f(z) = \norm{R_A(z)}$.
Consider the compact set
\begin{equation*}
K = \overline{\sigma_\varepsilon(A)} \setminus \bigl(\bigcup_{j=1}^J B_{\varepsilon/2}(\lambda_j) \bigr).
\end{equation*}
$f$ is bounded on $K$ and $\dist(z,\sigma(A)) \geq \varepsilon/2$ for any $z \in K$.
Next fix $z \in \sigma_{\varepsilon}(A)$.
The remainder of the proof constructs a sequence $(x_n)_{n\in\IN}\subset \sigma_\varepsilon(A)$ which starts at the point $x_1=z$ and at a finite index $m$ enters $B_{\varepsilon/2}(\lambda_j)$ for some $j=1,\dots,J$.
Moreover, each line segment $[x_n,x_{n+1}]$, $1 \leq n \leq m-1$, lies completely inside $\sigma_\varepsilon(A)$.
{If $z\in\sigma_\varepsilon(A)\setminus K$ we are done, as $z$ can be connected to an eigenvalue by a single line segment. Hence we may assume $z$ belongs to the interior of $K$.} Set $\delta = (f(z)-\varepsilon^{-1})/2>0$ and for any $x\in \overline{\sigma_{\varepsilon}(A)}\setminus\sigma(A)$ define $r_x$ to be the supremum over all $r>0$ such that there exists a $y$ with $\abs{x-y}=r$, such that {$f(x)<f(y)$} and $f(x)-\delta\leq f(\zeta)$ for all $\zeta\in[x,y]$.
Note that this implies $y \in \sigma_{\varepsilon}(A)$, but not necessarily $[x,y] \subset \overline{\sigma_{\varepsilon}(A)}$.
By Theorem~\ref{thm-horia} the supremum is taken over a non-empty set,
such that $r_x > 0$.
As $\overline{\sigma_{\varepsilon}(A)}$ is bounded the point $y$ cannot be located arbitrarily far away from $x$ so $r_x$ is finite.
For each $x\in \overline{\sigma_{\varepsilon}(A)}\setminus\sigma(A)$ there exists $y_x \in \sigma_{\varepsilon}(A)$ such that $r_x/2 < \abs{y_x - x} \leq r_x$, $f(x)<f(y_x)$, and $f(x) - \delta \leq f(\zeta)$ for all $\zeta \in [x,y_x]$.
Define a sequence $(x_n)_{n \in \IN} \subset \sigma_\varepsilon(A)$ by $x_1=z$ and $x_{n+1}=y_{x_n}$, $n\geq 1$, which implies
\begin{equation*}
\frac{1}{\varepsilon} < f(z)=f(x_1)<f(x_2)<\dots
\end{equation*}
Next we show that $x_n$ must escape from $K$ for some $n$.
Assume that this is not true such that $x_n\in K$ for all $n$.
Then the increasing sequence given by $f(x_n)$ is bounded. Thus there exists $M>0$ such that $f(x_n)<M$ for all $n$ and $\lim_{n\to\infty}f(x_n)=M$.
As $K$ is compact we may assume (by passing to a subsequence) that $x_n$ converges to $a\in K$.
From
Theorem~\ref{thm-horia} we know that there exists some $a'$ located at a positive distance from $a$ where $f(a)<f(a')$ and $f(a)<f(\zeta)$ for all $\zeta$ in the open line segment $]a,a'[$.
Since $f$ is uniformly continuous on $K$, we have for $n$ larger than some $m$, $f(x_n)<f(a')$ and $f(x_n)-\delta\leq f(\zeta_n)$ for every $\zeta_n\in[x_n,a']$.
Furthermore, we may assume that $\abs{x_{n+1}-x_n} \leq \abs{a-a'}/10$ and $\abs{a-x_{n+1}}\leq \abs{a-a'}/10$.
Then $a'$ fulfils the criteria for $y$ which were used to define $r_{x_n}$. If $n$ is large enough we have
\begin{equation*}
\abs{a'-x_n} \geq \abs{a'-a}-\abs{a-x_{n+1}} - \abs{x_{n+1}-x_n} > 2\abs{x_{n+1}-x_n} = 2 \abs{y_{x_n}-x_n} > r_{x_n},
\end{equation*}
contradicting the definition of $r_{x_n}$.
Thus the $x_n$ must lie outside $K$ if $n \geq m$ for some $m$, and they must lie in $\bigcup_{j=1}^J B_{\varepsilon/2}(\lambda_j)$.
It remains to show that the polygonal path connecting $z$, via the points $z=x_1,\dots,x_m$, to $\bigcup_{j=1}^J B_{\varepsilon/2}(\lambda_j)$ is contained in $\sigma_{\varepsilon}(A)$. From the definition of $\delta$ and the construction of $x_1,\dots,x_m$, every $\zeta$ on the polygonal path satisfies
\begin{equation*}
f(\zeta)\geq f(z)-\delta= \frac{f(z)+\varepsilon^{-1}}{2} > \frac{1}{\varepsilon}.
\end{equation*}
Thus the path lies in $\sigma_\varepsilon(A)$, and can with one additional line segment be connected to one of the eigenvalues of $A$.
\begin{remark}
The connected components of $\sigma_{\varepsilon}(A)$ are not necessarily simply connected. In the case of $A$ normal an example is constructed as follows.
Let $N\geq2$ and let $A\in\IC^{N\times N}$ be the diagonal matrix with $A_{jj}=j+\I(-1)^j\sqrt{3}/2$, $j=1,2,\ldots,N$. Elementary geometry shows that for $1<\varepsilon<2/\sqrt{3}$ the set $\sigma_{\varepsilon}(A)$ is $N-1$ connected.
In section~\ref{sec: local minima} we give an example with a non-normal matrix, see Figure~\ref{fig:pseudo} in subsection~\ref{sec: block diagonals}.
\end{remark}
\section{Examples}%
\label{sec: local minima}
In this section we give a number of examples.
In subsection~\ref{ex-assum} we give an example of an operator with an eigenvalue $\lambda_{\rm max}(z)$ that has infinite multiplicity. In this example Assumption~\ref{assum} holds for some but not all $z\in\rho(A)$.
In subsections~\ref{sec: block diagonals} and~\ref{sec: general example} we give examples of non-normal matrices such that the resolvent norm has a local minimum at the origin. In subsection~\ref{infinite} we give an example of a non-normal operator in an infinite dimensional Hilbert space which satisfies Assumption~\ref{assum} at the origin and such that the resolvent norm has a local minimum at the origin.
\subsection{On Assumption \ref{assum}}
\label{ex-assum}
We give examples showing that Assumption~\ref{assum} may only be satisfied in a proper subset of the resolvent set.
For $N\in\IN$ let $\mathcal{H}_N=L^2([0,1])\oplus \IC^N$ and for $N=\infty$ let
$\mathcal{H}_{\infty}=L^2([0,1])\oplus\ell^2(\IN)$. Let $A_1$ be multiplication by $x$ on $L^2([0,1])$, such that $\sigma(A_1)=\sigma_{\rm ac}(A_1)=[0,1]$, and let $A_2=2I$ on $\IC^N$ or $\ell^2(\IN)$. Let $A=A_1\oplus A_2$ on $\mathcal{H}_N$, $N\in\IN\cup\{\infty\}$. Then $A$ is a bounded self-adjoint operator with $\sigma(A)=[0,1]\cup\{2\}$. Using $\norm{S(z)}=\norm{R_A(z)}^2=1/\dist(z,\sigma(A))^2$ we see that Assumption~\ref{assum} is satisfied in
$\mathcal{G}=\{z\in\IC\,|\,\re z>\frac32, z\neq2\}$. It is not satisfied in $\rho(A)\setminus\mathcal{G}$. For $z\in\mathcal{G}$ we have $\lambda_{\rm max}(z)=\abs{2-z}^{-2}$ with multiplicity $N$.
\subsection{Local minimum for block diagonals with $\mathbf{2\times 2}$-blocks}%
\label{sec: block diagonals}
Let $A \in \IC^{2\times 2}$ have distinct eigenvalues $\frac{\tr(A)}{2}\pm r_A \e^{\I \phi}$ for $r_A>0$.
By \eqref{eq:resnormA2} and \eqref{eq:resnormtype2} we have
\begin{equation*}
\norm{R_{A}(\tr(A)/2)}^2 = \frac{2{r_A}^{-2}}{k_A-\sqrt{{k_A}^2-4}} = \frac{1}{r_A^2\gamma_A}
\end{equation*}
for $\gamma_A = (k_A-\sqrt{{k_A}^2-4})/2 \in~]0,1]$ and some $k_A\geq 2$ as given in the proof of Theorem~\ref{thm-henrik}(ii) (any real value $\geq 2$ may be obtained, depending on the structure of the matrix).
Note that $k_A$ is independent of the eigenvalues of $A$.
\begin{lemma} \label{lemma3}
Let $A\in \IC^{2\times 2}$ with distinct eigenvalues $\frac{\tr(A)}{2}\pm r_A \e^{\I\phi}$ for $r_A>0$ and define
\begin{equation*}
\theta_A = \frac{\pi}{2} - \frac{1}{2}\arccos(\gamma_A) \in~]\tfrac{\pi}{4},\tfrac{\pi}{2}].
\end{equation*}
The angles for which $\norm{R_A(\cdot)}$ is increasing away from $\tr(A)/2$ {\rm(}in a small neighborhood{\rm)} are precisely the two arcs $]\phi-\theta_A,\phi+\theta_A[\;\cup\;]\phi+\pi-\theta_A,\phi+\pi+\theta_A[$.
\end{lemma}
\begin{proof}
For simplicity consider $A_2 = {r_A}^{-1} \e^{-\I\phi}(A-\frac{\tr(A)}{2}I)$ such that $A_2$ is as in the proof of Theorem~\ref{thm-henrik}(ii).
For $k_A=2$ let $z=\sqrt{t} \e^{\I\theta}$ as in the proof of Theorem~\ref{thm-henrik}(ii).
We arrive again at \eqref{eq:deriv} under the condition that $t>0$ and $\cos(2\theta)\neq -1$ (avoiding the imaginary axis where we know $\norm{R_{A_2}}$ is decreasing), i.e.
\begin{equation*}
g(2,\theta) = \frac{1+\cos(2\theta)}{2} > 0,\quad \theta\in~]-\pi,\pi]\setminus \{\pm\tfrac{\pi}{2}\}.
\end{equation*}
So for $k_A=2$ the resolvent norm $\norm{R_{A_2}(\cdot)}$ is increasing away from $z=0$ (in a neighborhood) in all directions except along the imaginary axis where it decreases, corresponding to $\theta_A = \frac{\pi}{2}$.
Now assume $k_A>2$. Thus $\gamma_A\in~]0,1[$ and therefore
$\theta_A\in~]\tfrac{\pi}{4},\frac{\pi}{2}[$.
By \eqref{eq:deriv} it holds for $\theta_0\in~]-\pi,\pi]$ that $g(k_A,\theta_0)=0$, or equivalently $\cos(2\theta_0)=-\gamma_A$, if and only if
\begin{equation*}
\theta_0 = \begin{cases}
\phantom{-}\frac{\pi}{2}\pm \frac{1}{2}\arccos(\gamma_A) \\
-\frac{\pi}{2}\pm \frac{1}{2}\arccos(\gamma_A)
\end{cases} = \begin{cases}
\pm \theta_A, \\
\pm \pi \mp \theta_A.
\end{cases}
\end{equation*}
From the proof of Theorem~\ref{thm-henrik} we know that $\norm{R_{A_2}(\cdot)}$ is increasing away from $z=0$ on the real line (in a neighborhood) and decreasing away from $z=0$ on the imaginary line.
Hence, \eqref{eq:deriv} implies that $g(k_A,\theta)>0$ for $\theta\in~]-\theta_A,\theta_A[\;\cup\; ]\pi-\theta_A,\pi+\theta_A[$ and $g(k_A,\theta) \leq 0$ for $\theta\in [\theta_A,\pi-\theta_A]\cup[\pi+\theta_A,2\pi -\theta_A]$.
To recapitulate, for a small enough neighborhood of $0$, the directions for which $\norm{R_{A_2}(\cdot)}$ is increasing away from the origin are precisely the angles $]-\theta_A,\theta_A[\;\cup\; ]\pi-\theta_A,\pi+\theta_A[$.
Now the rotation by $\e^{\I\phi}$ occurring in \eqref{eq:resnormtype2} concludes the proof.
\end{proof}
From Lemma~\ref{lemma3} we can conclude that if $A\in \IC^{2\times 2}$ has eigenvalues $z_0 \pm r_A \e^{\I\phi}$ with $r_A>0$, then $\norm{R_A(\cdot)}$ is increasing near $z_0$ at least in a $\frac{\pi}{2}$-arc centered at $\phi$ and in a $\frac{\pi}{2}$-arc centered at $\phi+\pi$.
Thus, we directly get the following result by constructing a block diagonal matrix, where for each direction from a point $z_0$ there is at least one block for which the resolvent norm is increasing. We denote the torus by $\IT = \{z\in \IC\,|\, \abs{z}=1 \}$.
\begin{theorem}\label{thm: henrik 2}
Let $\{A_j\}_{j=1}^M\subset \IC^{2\times 2}$ such that $A_j$ has eigenvalues $z_0\pm r_j \e^{\I\phi_j}$ for $r_j>0$.
Let $B = \oplus_{j= 1}^M A_j$ such that $B$ has the eigenvalues of each $A_j$ and
\begin{equation*}
\norm{R_{B}(\cdot)} = \max_{j=1,\dots,M}\norm{R_{A_j}(\cdot)}.
\end{equation*}
Then $\norm{R_{B}(\cdot)}$ has a local minimum at $z_0$ if and only if there is a subset of indices $J\subseteq \{1,\dots,M\}$ such that
\begin{enumerate}[\rm(i)]
\item $\norm{R_{A_{j'}}(z_0)} < K = \norm{R_{A_{j}}(z_0)},
\quad j\in J,\quad j'\not\in J$,
\item $\IT = \{\e^{\I\phi} \,|\, \phi\in \Phi\cup(\pi+\Phi)\}$, where $\Phi = \bigcup_{j\in J} ]\phi_j-\theta_{A_j},\phi_j +\theta_{A_j}[$.
\end{enumerate}
\end{theorem}
\begin{proof}
We note that (i) implies, by continuity, that the value of $\norm{R_{B}(\cdot)}$ near $z_0$ is determined by $\{A_j\}_{j\in J}$.
(ii) is a necessary condition since otherwise there is a direction for which all $\norm{R_{A_j}(\cdot)}$ are decreasing near $z_0$ for $j\in J$.
On the other hand, if (i) and (ii) hold then for each direction near $z_0$ we have that $\norm{R_{B}(\cdot)}$ equals $\max_{j\in J}\norm{R_{A_j}(\cdot)}$ for which at least one of the resolvent norms $\norm{R_{A_j}(\cdot)}$ is increasing. Since $\{\norm{R_{A_j}(\cdot)}\}_{j\in J}$ coincide at $z_0$ then $\max_{j\in J} \norm{R_{A_j}(\cdot)}$ is increasing near $z_0$.
\end{proof}
As a sufficient condition in Theorem~\ref{thm: henrik 2} we may simply have $\norm{R_{A_j}(z_0)} = K$, $j=1,\dots,M$, for a constant $K>0$ and
\begin{equation*}
\IT = \left\{\e^{\I\phi} \,|\, \phi\in \bigcup{}_{j=1}^M \left( [\phi_j-\tfrac{\pi}{4},\phi_j +\tfrac{\pi}{4}]\cup[\phi_j+\tfrac{3\pi}{4},\phi_j+\tfrac{5\pi}{4}] \right)\right\}.
\end{equation*}
\begin{example}
Let $A_1,A_2\in\IC^{2\times 2}$ have eigenvalues, respectively $z_0\pm r_{A_1} \e^{\I\phi}$ and $z_0\pm r_{A_2} \I\e^{\I\phi}$ for $r_{A_1},r_{A_2}>0$.
I.e. the eigenvalues are placed on orthogonal lines intersecting at $z_0$.
For $B=A_1\oplus A_2$ then $\norm{R_{B}(\cdot)}$ has a local minimum at $z_0$ if and only if $\norm{R_{A_1}(z_0)} = \norm{R_{A_2}(z_0)}$.
This is the case if $\gamma_{A_1}/\gamma_{A_2} = \left({r_{A_2}}/{r_{A_1}}\right)^2$.
\end{example}
\begin{example} \label{ex:last}
Consider the following matrices
\begin{equation*}
A_1 = \begin{bmatrix}
\I & 0 \\ 0 & -\I
\end{bmatrix}, \quad
A_2 = \sqrt{\tfrac{2}{6-\sqrt{32}}}\,\e^{-\frac{\pi}{6}\I}\begin{bmatrix}
1 & 2 \\ 0 & -1
\end{bmatrix}, \quad A_3 = \sqrt{\tfrac{2}{30-\sqrt{896}}}\,\e^{\frac{\pi}{6}\I} \begin{bmatrix}
2\I & -1 \\ -5 & -2\I
\end{bmatrix}.
\end{equation*}
These are all scaled and rotated versions of matrices of the type \eqref{eq:type2} such that they satisfy the conditions of Theorem~\ref{thm: henrik 2} with $z_0 = 0$, $\phi_1 = \frac{\pi}{2}$, $\phi_2= -\frac{\pi}{6}$ and $\phi_3 = \frac{\pi}{6}$.
Moreover, we have $\norm{R_{A_1}(0)} = \norm{R_{A_2}(0)} = \norm{R_{A_3}(0)} = 1$.
Thus with $B = A_1\oplus A_2\oplus A_3$ we get that $\norm{R_{B}(\cdot)}$ has a local minimum at the origin. Note that $B$ is not normal.
\begin{figure}[htb]
\centering
\includegraphics[width=.6\textwidth]{fig1}
\caption{Pseudospectrum $\sigma_\varepsilon(B)$ with $\varepsilon = 0.97$ and eigenvalues of $B$ (black dots) from Example~\ref{ex:last}.}
\label{fig:pseudo}
\end{figure}
This is furthermore observed in Figure~\ref{fig:pseudo} where for appropriate values of $\varepsilon$, $\sigma_\varepsilon(B)$ is a connected set, but not simply connected.
In particular it excludes a region near the origin.
\end{example}
\subsection{Examples for any finite $\mathbf{\textit N > 2}$}%
\label{sec: general example}
Let $N>2$ and $\{a_j\}_{j=1}^N \subset \IC\setminus\{0\}$ such that $\abs{a_1} > \abs{a_j}$ for $j=2,\dots,N$. Define
\begin{equation}
A = \begin{bmatrix}
0 & \ldots & \ldots & 0 & \frac{1}{a_1} \\
\frac{1}{a_2} & \ddots & & & 0 \\
& \frac{1}{a_3} & \ddots & & \vdots \\
& & \ddots & \ddots & \vdots\\
& & & \frac{1}{a_N} & 0
\end{bmatrix}. \label{eq:genexample}
\end{equation}
Then we have:
\begin{equation*}
A^{-1} = \begin{bmatrix}
0 & a_2 & & & \\
\vdots & \ddots & a_3 & & \\
\vdots & & \ddots & \ddots & \\
0 & & & \ddots & a_N\\
a_1 & 0 & \ldots & \ldots & 0
\end{bmatrix}, \quad S(0) = (A^{-1})^{\ast}A^{-1} = \begin{bmatrix}
\abs{a_1}^2 & & & & \\
& \abs{a_2}^2 & & & \\
& & \abs{a_3}^2 & & \\
& & & \ddots &\\
& & & & \abs{a_N}^2
\end{bmatrix}.
\end{equation*}
Thus the normalized eigenvectors for $S(0)$ corresponding to the largest eigenvalue $\abs{a_1}^2$ are
\begin{equation*}
\psi = (\e^{\I\theta},0,\dots,0),\quad \theta\in\IR.
\end{equation*}
Furthermore, as $N>2$ we have
\begin{equation*}
A^{-2} = \begin{bmatrix}
0 & 0 & a_2a_3 & & & \\
\vdots & \ddots & \ddots & a_3a_4 & \\
\vdots & & \ddots & \ddots & \ddots& \\
0 & & & \ddots & \ddots & a_{N-1}a_N\\
a_1a_N & 0 & & & \ddots& 0 \\
0 & a_1a_2 & 0 & \ldots & \ldots & 0
\end{bmatrix}.
\end{equation*}
Thus we have
\begin{equation*}
\ip{\psi}{A^{-1}\psi} = \ip{\psi}{A^{-2}\psi} = 0,
\end{equation*}
which by Corollary~\ref{cor12} implies that $\norm{R_A(\cdot)}$ has a local minimum at the origin.
Note that this example fails for $N=2$, since in that case $A^{-2} = \diag(a_1a_2,a_1a_2)$, so $\ip{\psi}{A^{-2}\psi} \neq 0$.
See Figure~\ref{fig:pseudo2} for a specific choice of $\{a_j\}_{j=1}^6$ for $N=6$.
\begin{figure}[htb]
\centering
\includegraphics[width=.5\textwidth]{fig2}
\caption{Pseudospectrum $\sigma_\varepsilon(A)$ with $\varepsilon = 9.9966\cdot 10^{-7}$ and eigenvalues of $A$ (black dots) of the matrix in \eqref{eq:genexample} where $N=6$, $a_1 = 10^6$ and $a_j = 1$ for $j = 2,\dots,6$.}
\label{fig:pseudo2}
\end{figure}
\subsection{An infinite dimensional example}
\label{infinite}
We give an example of a non-normal operator on an infinite dimensional Hilbert space sa\-tis\-fy\-ing Assumption~\ref{assum} such that its resolvent norm has a local minimum at the origin.
Let $\mathcal{H}=\ell^2(\IZ)$. Let $a_j\in\IC\setminus\{0\}$, $j\in\IZ$, be a sequence which satisfies
\begin{equation}\label{a-assum}
\abs{a_0} > \sup_{j \neq 0} \abs{a_j}.
\end{equation}
Define an operator $A$ by
\begin{equation*}
(Ax)_j=a_{j+1}^{-1}x_{j+1},\quad x\in D(A)
\end{equation*}
where $D(A)=\{x\in\mathcal{H}\,|\, Ax\in \mathcal{H}\}$ is the maximal domain. It is easy to verify that $A$ is densely defined and closed. We have that $A$ is invertible with a bounded inverse
\begin{equation*}
(A^{-1}x)_j=a_jx_{j-1},\quad x\in\mathcal{H}.
\end{equation*}
A computation shows that $(A^{-1})^{\ast}A^{-1}$ is given by $((A^{-1})^{\ast}A^{-1}x)_j=\abs{a_{j+1}}^2x_j$, $j\in\IZ$. Thus a basis of eigenvectors is given by the canonical basis $\{e_j\}_{j\in\IZ}$. By construction the largest eigenvalue is simple and equals $\abs{a_0}^2$, with $\psi=e_{-1}$ as a normalized eigenvector. We have
\begin{align*}
\ip{\psi}{A^{-1}\psi}&=a_0\ip{e_{-1}}{e_0}=0,\\
\ip{\psi}{A^{-2}\psi}&=a_0a_1\ip{e_{-1}}{e_{1}}=0.
\end{align*}
Thus by Corollary~\ref{cor12} $\norm{R_A(\cdot)}$ has a local minimum at the origin.
We note that if $\lim_{\abs{j}\to\infty}a_j=0$, then $A$ has compact resolvent. If $\inf_{j\in\IZ}\abs{a_j}>0$ then $A$ is bounded.
\subsection*{Acknowledgements}
HC, AJ, and HKK were partially supported by the Danish Council for Independent Research $|$ Natural Sciences, Grant DFF--4181-00042. AJ thanks Kenji Yajima, Gakushuin University, Tokyo, Japan, for useful comments.
|
\section{Conclusion}
\label{sec:conclusion}
We propose \mechanism, a mechanism for extracting true random numbers with high
throughput from unmodified commodity DRAM devices on any system that allows
manipulation of DRAM timing parameters in the memory controller. \mechanism\
harvests fully non-deterministic random numbers from DRAM row activation
failures, which are \jk{bit errors} induced by intentionally \jk{accessing DRAM
with lower latency than required for correct row activation.} Our TRNG is based
on two key observations: 1) activation failures can be induced quickly and 2)
repeatedly accessing certain DRAM cells with reduced activation latency results
in reading true random data. We validate the quality of our TRNG with the
commonly-used NIST statistical test suite for randomness. Our evaluations show
that \mechanism\ significantly outperforms the previous highest-throughput
DRAM-based TRNG by up to 211x \jkthree{(128x on average)}. We conclude that
DRAM row activation failures can be effectively exploited to efficiently
generate true random numbers with high throughput on a wide range of devices
that use commodity DRAM chips.
\section{Introduction}
Random number generators (RNGs) are critical components in many different
applications, including cryptography, scientific simulation, industrial
testing, and recreational entertainment~\mpt{\cite{bagini1999design,
rock2005pseudorandom, ma2016quantum, stipvcevic2014true,
barangi2016straintronics, tao2017tvl, gutterman2006analysis, von2007dual,
kim2017nano, drutarovsky2007robust, kwok2006fpga, cherkaoui2013very,
zhang2017high, quintessence2015white}}. These applications require a mechanism
capable of rapidly generating random numbers across \mpt{a wide variety of
operating conditions (e.g., temperature/voltage fluctuations, manufacturing
variations, malicious external attacks)~\cite{yang2016all}}. In particular, for
modern cryptographic applications, a random (i.e., completely
unpredictable) number generator is critical to prevent information leakage to
a potential adversary~\cite{kocc2009cryptographic, gutterman2006analysis,
von2007dual, kim2017nano, drutarovsky2007robust, kwok2006fpga,
cherkaoui2013very, zhang2017high, quintessence2015white}.
Random number generators can be broadly classified into two
categories~\cite{tilborgencyclopedia, chevalier1974random, knuth1998art,
tsoi2003compact}: 1) \emph{pseudo-random number generators
(PRNGs)}~\cite{matsumoto1998mersenne, blum1986simple, mascagni2000algorithm,
steele2014fast, marsaglia2003xorshift}, which deterministically generate
numbers starting from a \emph{seed value} with the goal of approximating a true
random sequence, and 2) \emph{true random number generators
(TRNGs)}~\cite{pyo2009dram, keller2014dynamic, sutar2018d, hashemian2015robust,
tehranipoor2016robust, wang2012flash, ray2018true, holcomb2007initial,
holcomb2009power, van2012efficient, chan2011true, tzeng2008parallel,
teh2015gpus, majzoobi2011fpga, wieczorek2014fpga, chu1999design,
amaki2015oscillator, mathew20122, brederlow2006low, tokunaga2008true,
bucci2003high, bhargava2015robust, kinniment2002design, holleman20083,
gutterman2006analysis, dorrendorf2007cryptanalysis, lacharme2012linux,
pareschi2006fast, yang2016all}, which generate random numbers based on
sampling non-deterministic random variables inherent in various physical
phenomena (e.g., electrical noise, atmospheric noise, clock jitter, Brownian
motion).
PRNGs are popular due to their flexibility, low cost, and fast pseudo-random
number generation time~\cite{chan2011true}, but their output is \emph{fully
determined} by the starting seed value. This means that the output of a PRNG
may be predictable given complete information about its operation.
Therefore, a PRNG falls short for applications that require high-entropy
values~\cite{corrigan2013ensuring, von2007dual, cherkaoui2013very}. In
contrast, because a TRNG mechanism relies on sampling entropy inherent in
\emph{non-deterministic} physical phenomena, the output of a TRNG is
\emph{fully unpredictable} even when complete information about the underlying
mechanism is available~\cite{kocc2009cryptographic}.
Based on analysis done by prior work on TRNG
design~\cite{kocc2009cryptographic, jun1999intel, schindler2002evaluation},
we argue that an \emph{effective} TRNG must: \hhthree{1) produce truly
random (i.e., completely unpredictable) numbers, 2) provide a high
throughput of random numbers at low latency, and 3) be practically
implementable at low cost.} Many prior works study different methods of
generating true random numbers that can be implemented using CMOS
devices~\cite{pyo2009dram, keller2014dynamic, sutar2018d,
hashemian2015robust, tehranipoor2016robust, wang2012flash, ray2018true,
holcomb2007initial, holcomb2009power, van2012efficient, chan2011true,
tzeng2008parallel, teh2015gpus, majzoobi2011fpga, wieczorek2014fpga,
chu1999design, amaki2015oscillator, mathew20122, brederlow2006low,
tokunaga2008true, bucci2003high, bhargava2015robust, kinniment2002design,
holleman20083, gutterman2006analysis, dorrendorf2007cryptanalysis,
lacharme2012linux, pareschi2006fast, yang2016all}. We provide a thorough
discussion of these past works in Section~\ref{related}.
\mpt{Unfortunately, most of these proposals fail to satisfy all of the
properties of an effective TRNG because they} either require specialized
hardware to implement (e.g., free-running
oscillators~\cite{amaki2015oscillator, yang2016all}, metastable
circuitry~\cite{mathew20122, brederlow2006low, tokunaga2008true,
bhargava2015robust}) or are unable to sustain continuous high-throughput
operation on the order of Mb/s (e.g., memory startup
values~\cite{tehranipoor2016robust, holcomb2007initial, holcomb2009power,
van2012efficient, eckert2017drng}, memory data retention
failures~\cite{keller2014dynamic, sutar2018d}). These limitations preclude the
widespread adoption of such TRNGs, thereby limiting the overall impact of these
proposals.
Commodity DRAM chips offer a promising substrate to overcome these
limitations due to three major reasons. First, DRAM
operation is highly sensitive to changes in access timing, which means
that we can easily induce failures by manipulating manufacturer-recommended
DRAM access timing parameters. \jkfour{These failures have been shown to exhibit
non-determinism~\cite{yaney1987meta, nair2016xed, khan2014efficacy,
chang2016understanding, qureshi2015avatar, patel2017reaper, kim2018dram,
lee2015adaptive, lee-sigmetrics2017, kim2018solar} and therefore \jkfive{they}
may be exploitable for true random number generation.} Second, commodity DRAM
devices already provide an interface capable of transferring data continuously
with high throughput in order to support a high-performance TRNG. Third, DRAM
devices are already prevalently in use throughout modern computing systems,
ranging from simple microcontrollers to sophisticated supercomputers.
\textbf{Our goal} in this paper is to design a TRNG
that:
\begin{enumerate}
\setlength{\itemsep}{0pt
\item is implementable on commodity DRAM devices today
\item is fully non-deterministic (i.e., it is impossible to predict the next output even with complete information about the underlying mechanism)
\item provides continuous (i.e., constant rate), high-throughput random values \hhtwo{at low latency}
\item provides random values while minimally affecting concurrently-running applications
\end{enumerate}
Meeting these four goals would enable a TRNG design that is suitable
for applications requiring high-throughput true random number generation in
commodity devices today.
Prior approaches to DRAM-based TRNG design successfully use DRAM data retention
failures~\cite{keller2014dynamic, sutar2018d, hashemian2015robust}, DRAM
startup values~\cite{tehranipoor2016robust, eckert2017drng}, and non-determinism in DRAM
command scheduling~\cite{pyo2009dram} to generate true random numbers.
Unfortunately, these approaches do not fully satisfy our four goals because
they either \hhtwo{do not exploit a fundamentally non-deterministic entropy source
(e.g., DRAM command scheduling~\cite{pyo2009dram}) or are too slow for
continuous high-throughput operation (e.g., DRAM data retention
failures~\cite{keller2014dynamic, sutar2018d, hashemian2015robust}, DRAM
startup values~\cite{tehranipoor2016robust, eckert2017drng})}.
Section~\ref{comparison} provides a detailed comparative analysis of these
prior works.
In this paper, we propose a new way to leverage DRAM cells as an entropy source
for true random number generation by intentionally violating the access
timing parameters and using the resulting errors as the source of randomness.
Our technique specifically extracts randomness from \emph{activation
failures}, i.e., DRAM errors caused by intentionally decreasing the row
activation latency (timing parameter $t_{RCD}$) below manufacturer-recommended
specifications. Our proposal is based on \textbf{two key observations}:
\begin{enumerate}
\setlength{\itemsep}{0pt
\item \hh{Reading} certain DRAM cells with a reduced activation latency returns
\emph{true random} values.
\item An activation failure can be induced very quickly (i.e., \emph{even
faster} than a normal DRAM row activation).
\end{enumerate}
Based on these key observations, we propose \mechanism, a new methodology for
extracting true random numbers from commodity DRAM devices with high
throughput. \mpt{\mechanism~consists of two steps: 1) identifying specific
DRAM cells that are vulnerable to activation failures using a
\emph{low-latency} profiling step and 2) generating a continuous stream (i.e.,
constant rate) of random numbers by repeatedly inducing activation failures in
the previously-identified vulnerable cells.} \mechanism~runs entirely in
software and is capable of immediately running on any commodity system that
provides the ability to manipulate DRAM timing parameters within the memory
controller~\cite{bkdg_amd2013, opteron_amd}. For most other devices, a simple
software API must be exposed without any hardware changes to the commodity DRAM
device (e.g., similarly to SoftMC~\cite{hassan2017softmc,
softmc-safarigithub}), which makes \mechanism\ suitable for implementation on
most existing systems today.
In order to demonstrate \mechanism's effectiveness, we perform a rigorous
experimental characterization of activation failures using 282 state-of-the-art
LPDDR4~\cite{2014lpddr4} DRAM devices from three major DRAM manufacturers. We
also verify our observations using four additional DDR3~\cite{jedec2012} DRAM
devices from a single manufacturer. Using the standard NIST statistical test
suite for randomness~\cite{rukhin2001statistical}, we show that \mechanism~is
able to maintain high-quality true random number generation both over 15 days
of testing and across the entire reliable testing temperature range of our
infrastructure (55$^{\circ}$C-70$^{\circ}$C). Our results show that
\mechanism's maximum (average) throughput is \changes{$717.4 Mb/s$ ($435.7
Mb/s$) using four \hhtwo{LPDDR4} DRAM channels, which is over two orders of
magnitude} higher than that of the best prior DRAM-based TRNG.
We make the following \textbf{key contributions}:
\begin{enumerate}
\setlength{\itemsep}{0pt
\item We introduce \mechanism, a new methodology for extracting true random
numbers from a commodity DRAM device at \hhtwo{high} throughput \hhtwo{and low
latency}. The key idea of \mechanism~is to use DRAM cells as entropy sources to
generate true random numbers by accessing them with a latency
that is \hhtwo{lower than}
manufacturer-recommended specifications.
\item Using experimental data from 282 state-of-the-art LPDDR4 DRAM devices
from three major DRAM manufacturers, we present a rigorous characterization of
\hhtwo{randomness in errors induced by accessing DRAM with low latency}.
Our analysis demonstrates that \mechanism~is able
to maintain high-quality random number generation both over 15 days of testing
and across the entire reliable testing temperature range of our
infrastructure (55$^{\circ}$C-70$^{\circ}$C). We verify our observations from
this study with prior works' observations on DDR3 DRAM
devices~\cite{chang2016understanding, lee-sigmetrics2017, lee2015adaptive,
kim2018solar}. Furthermore, we experimentally demonstrate on four DDR3 DRAM
devices, from a single manufacturer, that \mechanism~is suitable for implementation
in a \hhtwo{wide} range of commodity DRAM devices.
\item We evaluate the quality of \mechanism's output bitstream using the
\mpo{standard} NIST statistical test suite for
randomness~\cite{rukhin2001statistical} and find that it successfully passes
every test. We also compare \mechanism's performance to four previously
proposed DRAM-based TRNG designs (Section~\ref{comparison}) and show that
\mechanism\ outperforms the best prior DRAM-based TRNG design by over two orders
of magnitude in terms of maximum and average throughput.
\end{enumerate}
\section{Background}
\label{sec:background}
\hh{We} provide the necessary background on DRAM and true
random number generation that is required to understand our idea of true
random number generation \hasan{using the inherent properties of DRAM}.
\subsection{Dynamic Random Access Memory (DRAM)}
\label{subsec:dram}
We briefly \hh{describe} DRAM organization \hh{and basics}. We refer the
reader to past works~\cite{lee2015adaptive, khan2016parbor,
seshadri2016simple, hassan2016chargecache, hassan2017softmc, zhang2014half,
lee2013tiered, kim2012case, seshadri2013rowclone, chang2016understanding,
lee2016reducing, chang2017thesis, chang2016low, lee-sigmetrics2017,
seshadri2017ambit, lee2015decoupled, kim2016ramulator, kim2014flipping,
bhati2015flexible, chang2017understanding, liu2013experimental,
khan2014efficacy, qureshi2015avatar, zhang2014half, khan2017detecting,
patel2017reaper, mukundan2013understanding, chang2014improving, raidr,
kim2018dram, kim2018solar, vampire2018ghose} for more detail.
\subsubsection{DRAM System Organization}
\label{subsubsec:dram_org}
~In a typical system configuration, a CPU chip includes a set of memory
controllers, where each memory controller interfaces \hhtwo{with} a DRAM channel to
perform read and write operations. As we show in Figure~\ref{fig:dram_org}
\hh{(left)}, a DRAM channel has its own I/O bus and operates independently
of other channels in the system. \hh{To achieve high memory capacity, a
channel can host multiple DRAM modules by sharing the I/O bus between the
modules. A DRAM module implements a single or multiple DRAM ranks.}
Command and data transfers are serialized between ranks in the same channel due
to the shared I/O bus. A DRAM rank consists of multiple DRAM chips that operate
in lock-step, i.e., all chips simultaneously perform the same operation\hhtwo{,
but they do so on different bits}. The number of DRAM chips per rank depends on
the data bus width of the DRAM chips and the channel width. For example, a
typical system has a 64-bit wide DRAM channel. Thus, four 16-bit or eight 8-bit
DRAM chips are needed to build a DRAM rank.
\begin{figure}[h] \centering
\includegraphics[width=0.85\linewidth]{dram_system3.pdf}
\caption{A typical DRAM-based system~\cite{kim2018solar}.}
\label{fig:dram_org}
\end{figure}
\subsubsection{DRAM Chip Organization}
\label{subsubsec:dram_bank}
~At a high-level, a DRAM chip consists of \hh{billions} of DRAM cells that are
hierarchically organized to maximize storage density and performance. We
describe each level of the hierarchy of a modern DRAM chip.
A modern DRAM chip is composed of multiple DRAM banks (shown in
Figure~\ref{fig:dram_org}\hh{, right}). The chip communicates with the memory controller
through the \emph{I/O circuitry}. The I/O circuitry is connected to the
\emph{internal command and data bus} that is shared among all banks in the
chip.
Figure~\ref{subfig:dram_bank} illustrates the \hasan{organization} of a DRAM bank. In a
bank, the \emph{global row decoder} partially decodes the address of the
accessed \emph{DRAM row} to select the corresponding \emph{DRAM subarray}. A
DRAM subarray is a 2D array of DRAM cells, where cells are horizontally
organized into multiple DRAM rows. A DRAM row is a set of DRAM
cells that share a wire called the \emph{wordline}, which the \emph{local row
decoder} of the subarray drives after fully decoding the row address. In a
subarray, a column of cells shares a wire, referred to as the \emph{bitline},
that connects the column of cells to a \emph{sense amplifier}. The sense
amplifier is the circuitry used to read and modify the data of a DRAM cell.
\hh{The row} of sense amplifiers in the subarray is referred to as the \emph{local
row-buffer}. To access a DRAM cell, the corresponding DRAM row first needs to
be copied into the local row-buffer, which connects to the internal I/O bus via
the \emph{global row-buffer}.
\begin{figure}[h]
\centering
\begin{subfigure}[b]{.600\linewidth}
\centering
\includegraphics[width=\linewidth]{dram_bank.pdf}
\caption{DRAM bank.}
\label{subfig:dram_bank}
\end{subfigure}
\quad
\begin{subfigure}[b]{.345\linewidth}
\centering
\includegraphics[width=\linewidth]{dram_cell.pdf}
\caption{DRAM cell.}
\label{subfig:dram_cell}
\end{subfigure}
\caption{DRAM bank and cell architecture~\cite{kim2018solar}.}
\label{figure:dram_bank_and_cell}
\end{figure}
Figure~\ref{subfig:dram_cell} illustrates a DRAM cell, which is composed of a
\emph{storage capacitor} and \emph{access transistor}. A DRAM cell stores a
single bit of information based on the charge level of the capacitor. The data
stored in the cell is interpreted as a ``1'' or ``0'' depending on whether the
charge stored in the cell is above or below a certain threshold. Unfortunately,
the capacitor and the access transistor are not ideal circuit components and
have \emph{charge leakage paths}. Thus, to ensure that the cell does not leak
charge to the point where the bit stored in the cell flips, the cell needs to
be periodically \emph{refreshed} to fully restore its original charge.
\subsubsection{DRAM Commands}
\label{subsubsec:dram_cmds}
~The memory controller issues a set of DRAM commands to access data in the DRAM
chip. To perform a read or write operation, the memory controller first needs
to \emph{open} a row, i.e., copy the data of the cells in the row to the
row-buffer. To open a row, the memory controller issues an \emph{activate
(ACT)} command to a bank by specifying the address of the row to open. The
memory controller can issue \emph{ACT} commands to different banks in
consecutive DRAM bus cycles to operate on \emph{multiple banks in parallel}.
After opening a row \hh{in a bank}, the memory controller issues either a \emph{READ} or a
\emph{WRITE} command to read or write a DRAM word (which is typically equal to
64 bytes) \hh{within the open row}.
\hh{An open row can serve multiple \emph{READ} and \emph{WRITE} requests
without incurring precharge and activation delays.}
A DRAM row typically contains 4-8 KiBs of data. To access data from
another DRAM row \hh{in} the same bank, the memory controller must first close the
currently open row by issuing a \emph{precharge (PRE)} command. The memory
controller also periodically issues \emph{refresh (REF)} commands to prevent
data loss due to charge leakage.
\subsubsection{DRAM Cell Operation}
\label{subsubsec:dram_operation}
~We describe DRAM operation by explaining the steps involved in reading data
from \hh{a DRAM cell}.\footnote{Although we focus only on reading data, steps
involved in a write operation are similar.} The memory controller initiates
each step by issuing a DRAM command. Each step takes a certain amount of time
to complete, and thus, a DRAM command is typically associated with one or more
timing constraints known as \emph{timing parameters}. It is the responsibility
of the memory controller to satisfy these timing parameters in order to ensure
\emph{correct} DRAM operation.
In Figure~\ref{fig:dram_op}, we show how the state of a DRAM cell changes
during the steps involved in a read operation. Each DRAM cell diagram
corresponds to the state of the cell at exactly the tick \hh{mark} on the time
axis. Each command (\jkfour{shown} in purple boxes below the \hhtwo{time} axis)
is issued by the memory controller at the \hh{corresponding} tick \hh{mark}.
Initially, the cell is in a \emph{precharged} state~\incircle{1}. When
precharged, the capacitor of the cell is disconnected from the bitline since
the wordline is not asserted and thus the access transistor is off. The bitline
voltage is stable at $\frac{V_{dd}}{2}$ and is ready to \hh{be perturbed}
towards the \hh{voltage level} of the cell \hh{capacitor} upon enabling the
access transistor.
\begin{figure}[h] \centering
\includegraphics[width=\linewidth]{dram_cmd_timeline2.pdf}
\caption{Command sequence for reading data from DRAM and the state of a DRAM cell during each \hhtwo{related} step.}
\label{fig:dram_op}
\end{figure}
To read data from a cell, the memory controller first needs to perform
\emph{row activation} by issuing an \emph{ACT} command. During row activation
\hh{(\incircle{2}),} the row decoder asserts the wordline that connects the
storage capacitor of the cell to the bitline by enabling the access transistor.
At this point, the capacitor charge perturbs the bitline \hh{via the}
\emph{charge sharing} \hh{process}. Charge sharing continues until the
capacitor and bitline voltages reach an equal value of $\frac{V_{dd}}{2} +
\delta$. After charge sharing \hh{(\incircle{3}),} the sense amplifier begins
driving the bitline towards \hh{either} $V_{dd}$ or $0V$ depending on the
direction of the perturbation in the charge sharing step. This step\hh{, which
amplifies} the voltage level \hh{on the bitline as well as the cell} is called
\emph{charge restoration}. Although charge restoration continues until the
original capacitor charge is fully replenished (\incircle{4}), the memory
controller can issue a \emph{READ} command to safely read data from the
activated row \hh{before the capacitor charge is fully replenished}. A
\emph{READ} command can \hasan{reliably} be issued when the bitline voltage
reaches \hh{the} voltage level $V_{read}$. To ensure that the read occurs
after the bitline reaches $V_{read}$, the memory controller inserts a time
interval $t_{RCD}$ between the \emph{ACT} and \emph{READ} commands. It is the
responsibility of the DRAM \hh{manufacturer} to ensure that their DRAM chip
operates safely as long as the \hh{memory controller obeys the} $t_{RCD}$
timing parameter, which is defined in the DRAM standard~\cite{2014lpddr4}. If
the memory controller issues a \emph{READ} command before $t_{RCD}$ elapses,
the bitline voltage may be below $V_{read}$, which can lead to \hh{the} reading
\hh{of} a wrong value.
To return a cell to its precharged state, the voltage in the cell must first be
fully restored. A cell is expected to be fully \hh{restored when the memory
controller satisfies} \hh{a} time \hh{interval dictated by} $t_{RAS}$
\hh{after} issuing the \emph{ACT} command. Failing to satisfy $t_{RAS}$ may
lead to \hh{insufficient amount of charge \hh{to be} restored} in the cells of
the accessed row. A subsequent activation of the row can then result in the
reading \hh{of} incorrect data from the cells.
Once the cell is successfully \emph{restored} \hh{(\incircle{4}),} the
memory controller can issue a \emph{PRE} command to close the
\hh{currently-open} row to prepare the bank for an access to another row.
The cell returns to the precharged state (\incircle{5}) after waiting for
\hh{the} timing parameter $t_{RP}$ following the \emph{PRE} command.
Violating $t_{RP}$ \hh{may prevent} the sense amplifiers from fully driving
the bitline back to $\frac{V_{dd}}{2}$, \hh{which may later result in the row
to be activated with too small amount of charge in its cells,} \hhtwo{potentially
preventing the sense amplifiers to read the data correctly}.
For correct DRAM operation, it is critical for the memory controller to ensure
that the DRAM timing parameters defined in the DRAM specification are
\emph{not} violated. Violation of the timing parameters may lead to
\hh{incorrect data to be read from the DRAM,} and \hh{thus cause} unexpected
program behavior~\cite{lee2015adaptive, khan2016parbor, hassan2017softmc,
chang2016understanding, chang2017understanding, chang2017thesis,
lee-sigmetrics2017}. In this work, we study \hhtwo{the} failure modes \hhtwo{due to
violating DRAM timing parameters} and explore their application \hh{to}
reliably generating true random numbers.
\subsection{True Random Number Generators}
\label{subsec:trng}
A \emph{true random number generator (TRNG)} requires physical processes (e.g.,
radioactive decay, thermal noise, Poisson noise) to construct a bitstream of
random data. Unlike pseudo-random number generators, the random numbers
generated by a TRNG do \emph{not} depend on the \hh{previously-generated} numbers
and \emph{only} depend on the random noise obtained from physical processes.
TRNGs are usually validated using statistical tests such as
NIST~\cite{rukhin2001statistical} or DIEHARD~\cite{marsaglia2008marsaglia}. A
TRNG typically consists of 1) an \emph{entropy source}, 2) a \emph{randomness
extraction technique}, and sometimes 3) a \emph{post-processor}, which improves
the randomness of the extracted data often at the expense of throughput. These
three components are typically used to reliably generate true random
numbers\hh{~\cite{stipvcevic2014true, sunar2007provably}}.
\textbf{Entropy Source.} The entropy source is a critical component of a random
number generator, as its amount of entropy affects the unpredictability and the
throughput of the generated random data. Various physical phenomena can be
used as entropy sources. In the domain of electrical circuits, thermal and
Poisson noise, jitter, and circuit metastability have been proposed as processes
that have high entropy~\cite{wang2012flash, ray2018true, holcomb2007initial,
holcomb2009power, van2012efficient, mathew20122, brederlow2006low,
tokunaga2008true, bhargava2015robust}. To ensure robustness, the entropy source
should not be visible or modifiable by an adversary. Failing to satisfy that
requirement would result in generating predictable data, and thus put the
system into a state susceptible to security attacks.
\textbf{Randomness Extraction Technique.} The randomness extraction
technique harvests random data from an entropy source. A good randomness
extraction technique should have two key properties. First, it should have
high throughput, i.e., extract as much as randomness possible in a short amount
of time~\cite{kocc2009cryptographic, stipvcevic2014true}\hh{, especially
important for applications that require high-throughput random number
generation (e.g., security applications~\cite{gutterman2006analysis, von2007dual, kim2017nano,
drutarovsky2007robust, kwok2006fpga, cherkaoui2013very, zhang2017high,
quintessence2015white, bagini1999design, rock2005pseudorandom, ma2016quantum,
stipvcevic2014true, barangi2016straintronics, tao2017tvl, botha2005gammaray,
mathew20122, yang201416}, scientific simulation~\cite{ma2016quantum,
botha2005gammaray})}. Second, it should not disturb the physical
process~\cite{kocc2009cryptographic, stipvcevic2014true}. Affecting the entropy
source during the randomness extraction process would make the harvested data
predictable, lowering the reliability of the TRNG.
\textbf{Post-processing.} Harvesting randomness from a physical phenomenon
\emph{may} produce bits that are biased or
correlated~\cite{kocc2009cryptographic, rahman2014ti}. In such a case, a
post-processing step, which is also known as \emph{de-biasing}, is applied to
eliminate the bias and correlation. The post-processing step also provides
protection against environmental changes and adversary
tampering~\cite{kocc2009cryptographic, rahman2014ti, stipvcevic2014true}.
\hhtwo{Well-known} post-processing techniques are the von Neumann
corrector~\cite{jun1999intel} and cryptographic hash functions such as
SHA-1~\cite{eastlake2001us} or MD5~\cite{rivest1992md5}. These post-processing
steps work well, but generally result in decreased throughput (e.g., up to
80\%~\cite{kwok2011comparison}).
\section{Motivation and Goal}
\label{sec:motivation}
True random numbers sampled from physical phenomena have a number of real-world
applications from system security~\hh{\cite{bagini1999design,
rock2005pseudorandom, stipvcevic2014true}} to recreational
entertainment~\hh{\cite{stipvcevic2014true}}. As user data
privacy becomes a \emph{highly-sought} commodity in Internet-of-Things (IoT)
and mobile devices, enabling primitives that provide security on such systems
becomes \jk{critically important}~\cite{lin2017survey, ponemon2017sec, zhang2017high}.
Cryptography is one typical method for securing systems against various attacks
by encrypting the system's data with keys generated with true random values.
Many cryptographic algorithms require random values to generate keys in many
standard protocols (e.g., TLS/SSL/RSA/VPN keys) to either 1) encrypt network
packets, file systems, and data, 2) select internet protocol sequence numbers
(TCP), or 3) generate data padding values~\cite{gutterman2006analysis,
von2007dual, kim2017nano, drutarovsky2007robust, kwok2006fpga,
cherkaoui2013very, zhang2017high, quintessence2015white}. TRNGs are also
commonly used in authentication protocols and in countermeasures against
hardware attacks~\cite{cherkaoui2013very}, in which psuedo-random number
generators (PRNGs) \hh{are shown to be insecure}~\cite{von2007dual,
cherkaoui2013very}. To keep up with the \emph{ever-increasing} rate of
secure data creation, especially with the growing number of commodity
data-harvesting devices (e.g., IoT and mobile devices), the ability to generate
true random numbers with \emph{high throughput and low latency} becomes ever
more relevant to maintain user data privacy. In addition,
\emph{high-throughput} TRNGs are \hh{already} \emph{essential} components of
various important applications such as scientific
simulation~\hh{\cite{ma2016quantum, botha2005gammaray}}, industrial testing,
statistical sampling, randomized algorithms, and recreational
entertainment~\cite{bagini1999design, rock2005pseudorandom, ma2016quantum,
stipvcevic2014true, barangi2016straintronics, tao2017tvl, botha2005gammaray,
zhang2017high, mathew20122, yang201416}.
A \emph{\hh{widely-available,} high-throughput, low-latency} TRNG will enable
all previously mentioned applications that rely on TRNGs, including improved
\jktwo{security and privacy} in most systems that are known to be vulnerable to
attacks~\cite{lin2017survey, ponemon2017sec, zhang2017high}, as well as enable
research that we may not anticipate at the moment. One such direction is using
a one-time pad \hhthree{(i.e., a private key used to encode and decode only a single message)} with quantum key
distribution, which requires at least \emph{4Gb/s} \hh{of true random number
generation throughput}~\cite{wang2016theory, clarke2011robust, lu2015fpga}.
Many \emph{high-throughput} TRNGs have been recently
proposed~\cite{kwok2006fpga, tsoi2007high, zhang201568, cherkaoui2013very,
barangi2016straintronics, zhang2017high, ning2015design, quintessence2015white,
wang2016theory, mathew20122, yang201416, bae20173, fischer2004high,
gyorfi2009high}, and the availability of these high-throughput TRNGs can enable
\hh{a wide range of new applications with improved security} \jktwo{and privacy}.
DRAM offers a promising substrate for developing an effective and
widely-available TRNG due to the prevalence of DRAM throughout all modern
computing systems ranging from microcontrollers to supercomputers. A
high-throughput DRAM-based TRNG would help enable widespread adoption of
applications that are today limited to only select architectures equipped with
dedicated high-performance TRNG engines. Examples of such applications include
high-performance scientific simulations and cryptographic applications for
securing devices and communication protocols, both of which would run
\hhthree{much} more
\jk{efficiently on \hhthree{mobile devices, embedded devices,} or microcontrollers with the availability
of higher-throughput TRNGs \hhthree{in} the system.}
In terms of the CPU architecture itself, a high-throughput DRAM-based TRNG
\jkfour{could} help the memory controller to improve scheduling
decisions~\cite{usui2016dash, mutlu2009parallelism, ausavarungnirun2012staged,
subramanian2016bliss, kim2010thread, mutlu2007stall, subramanian2013mise,
subramanian2015application} and \jkfour{enable the implementation} a
truly-randomized version of PARA~\cite{kim2014flipping} (i.e., a protection
mechanism against the RowHammer \jktwo{vulnerability~\cite{kim2014flipping,
mutlu2017rowhammer}}). \jktwo{Furthermore, a DRAM-based TRNG would likely have
additional hardware and software applications as system designs become more
capable and increasingly security-critical.}
In addition to traditional computing paradigms, DRAM-based TRNGs can
benefit processing-in-memory (PIM) architectures~\cite{ghose2018enabling,
seshadri2017simple}, which co-locate logic \hhthree{within or near} memory \jktwo{to overcome the
large bandwidth and energy bottleneck caused by} the memory bus and leverage
the \emph{significant} data parallelism available within the DRAM chip itself.
Many prior works provide primitives for PIM or exploit PIM-enabled systems for
workload acceleration~\cite{ahn2015scalable, ahn2015pim, lee2015simultaneous,
seshadrifast, seshadri2013rowclone, seshadri2015gather, seshadri2017ambit,
liu2017concurrent, seshadri2017simple, pattnaik2016scheduling,
babarinsa2015jafar, farmahini2015nda, gao2015practical, gao2016hrl,
hassan2015near, hsieh2016transparent, morad2015gp, sura2015data, zhang2014top,
hsieh2016accelerating, boroumand2017lazypim, chang2016low, kim2018grim,
ghose2018enabling, boroumand2018google, seshadri2016simple}. A low-latency,
high-throughput DRAM-based TRNG can enable PIM applications to source random
values \emph{directly within the memory itself}, thereby enhancing the overall
potential, \jktwo{security, and privacy,} of PIM-enabled architectures. For
example, \hhf{in applications that require true random numbers, a DRAM-based
TRNG can enable large contiguous code segments to execute in memory, which
would reduce communication with the CPU, and thus improve system efficiency. A
DRAM-based TRNG can also enable security tasks to run completely in memory.
This would remove the dependence of PIM-based security tasks on an I/O channel
and would increase \jkfive{overall} system security.}
We posit, based on analysis done in prior works~\cite{kocc2009cryptographic,
jun1999intel, schindler2002evaluation}, that an \emph{effective} TRNG must
satisfy \emph{six} key properties: \jktwo{it} must 1) have \hh{low} implementation
cost, 2) be fully non-deterministic such that it is impossible to predict the
next output given complete information about how the mechanism operates, 3)
provide a continuous stream of true random numbers with high throughput, 4)
provide true random numbers with low latency, 5) exhibit low system
interference, i.e., not significantly \hh{slow down} concurrently-running
applications, and 6) generate random values with low energy overhead.
To this end, our \textbf{goal} in this work, is to provide a
\hh{widely-available} TRNG for DRAM
devices that satisfies all six key properties of an effective TRNG.
\section{Testing Environment}
\label{sec:methodology}
In order to test our hypothesis that DRAM cells are an effective source of
entropy when accessed with reduced DRAM timing parameters, we developed an
infrastructure to characterize modern LPDDR4 DRAM chips.
We also use an infrastructure for DDR3 DRAM chips,
SoftMC~\cite{hassan2017softmc, softmc-safarigithub}, to demonstrate
empirically that our proposal is applicable beyond the LPDDR4
technology. Both testing environments give us precise control over DRAM
commands and DRAM timing parameters as verified with a logic analyzer probing
the command bus.
We perform all tests, unless otherwise specified, using a total of 282 2y-nm
LPDDR4 DRAM chips from three major manufacturers in a thermally-controlled
chamber held at 45$^{\circ}$C. For consistency across results, we
precisely stabilize the ambient temperature using heaters and fans
controlled via a microcontroller-based proportional-integral-derivative (PID)
loop to within an accuracy of 0.25$^{\circ}$C and a reliable range of
40$^{\circ}$C to 55$^{\circ}$C. We maintain DRAM temperature at 15$^{\circ}$C
above ambient temperature using a separate local heating source. We use
temperature sensors to smooth out temperature variations caused by self-induced
heating.
We also use a separate infrastructure, based on open-source
SoftMC~\cite{hassan2017softmc, softmc-safarigithub}, to validate our mechanism
on 4 DDR3 DRAM chips from a single manufacturer. SoftMC enables precise control
over timing parameters, and we house the DRAM chips inside \jktwo{another}
temperature chamber to maintain a stable ambient testing temperature (with the
same temperature range as the temperature chamber \jktwo{used for the LPDDR4
devices}).
To explore the various effects of temperature, short-term aging, and
circuit-level interference (in Section~\ref{sec:dlrng_characterization}) on
activation failures, we reduce the $t_{RCD}$ parameter from the default
$18ns$ to $10ns$ for all experiments, unless otherwise stated.
Algorithm~\ref{alg:testing_lat} explains the general testing methodology we
use to induce activation failures. First, we write a data pattern to the
\begin{algorithm}[tbh]\footnotesize
\SetAlgoNlRelativeSize{0.7}
\SetAlgoNoLine
\DontPrintSemicolon
\SetAlCapHSkip{0pt}
\caption{DRAM Activation Failure Testing}
\label{alg:testing_lat}
\textbf{DRAM\_ACT\_failure\_testing($data\_pattern$, $DRAM\_region$):} \par
~~~~write $data\_pattern$ (e.g., solid 1s) into all cells in $DRAM\_region$ \par
~~~~set low $t_{RCD}$ for ranks containing $DRAM\_region$ \par
~~~~\textbf{foreach} $col$ in $DRAM\_region$: \par
~~~~~~~~\textbf{foreach} $row$ in $DRAM\_region$: \par
~~~~~~~~~~~~$activate(row)$ \textcolor{gray}{~~~~// fully refresh cells } \par
~~~~~~~~~~~~$precharge(row)$ \textcolor{gray}{~// ensure next access activates the row} \par
~~~~~~~~~~~~$activate(row)$ \par
~~~~~~~~~~~~$read(col)$ \textcolor{gray}{~~~~~~~~~~~// induce activation failure on col} \par
~~~~~~~~~~~~$precharge(row)$ \par
~~~~~~~~~~~~record activation failures to storage \par
~~~~set default $t_{RCD}$ for DRAM ranks containing $DRAM\_region$ \par
\end{algorithm}
region of DRAM under test (Line~2). Next, we reduce the $t_{RCD}$
parameter to begin inducing activation failures (Line~3). We then access
the DRAM region in column order (Lines~4-5) in order to ensure that each DRAM
access is to a closed DRAM row and thus requires an activation. This enables
each access to induce activation failures in DRAM. Prior to each
reduced-latency read, we first refresh the target row such that each cell has
the same amount of charge each time it is accessed with a reduced-latency read.
We effectively refresh a row by issuing an activate (Line~6) followed by a
precharge (Line~7) to that row. We then induce the activation failures by
issuing consecutive activate (Line~8), read (Line~9), and precharge (Line~10)
commands. Afterwards, we record any activation failures that we observe
(Line~11). We find that this methodology enables us to quickly induce
activation failures across \emph{all} of DRAM, and \jktwo{minimizes} testing
time.
\section{Activation Failure Characterization}
\label{sec:dlrng_characterization}
To demonstrate the viability of using DRAM cells as an entropy source for
random data, we explore and characterize DRAM failures when employing a
reduced DRAM activation latency ($t_{RCD}$) across 282 LPDDR4 DRAM chips. We
also compare our findings against \mpx{those of} prior works that study an
older generation of DDR3 DRAM chips~\cite{chang2016understanding,
lee-sigmetrics2017, lee2015adaptive, kim2018solar} to \jk{cross-}validate our
infrastructure. \jk{To understand the effects of changing environmental
conditions on a DRAM cell that is used as a source of entropy, we rigorously
characterize DRAM cell behavior as we vary four environmental conditions.
First, we study the effects of DRAM array design-induced variation (i.e., the
spatial distribution of activation failures in DRAM). Second, we study data
pattern dependence (DPD) effects on DRAM cells. Third, we study the effects of
temperature variation on DRAM cells. Fourth, we study a DRAM cell's activation
failure probability over time.} We present several key observations that
support the viability of a mechanism that generates random numbers by
accessing DRAM cells with a reduced $t_{RCD}$. In
Section~\ref{section:DLRNG_mechanism}, we discuss a mechanism to effectively
\jk{sample DRAM cells to extract true random numbers while minimizing the effects
of environmental condition variation (presented in this section) on the DRAM
cells.}
\subsection{Spatial Distribution of Activation Failures}
\label{subsec:spatial_char}
To study \mpx{which} regions of DRAM \mpx{are better suited to} generating
random data, we first visually inspect the spatial distributions of activation
failures \mpx{both} across DRAM chips and within each chip \mpx{individually}.
Figure~\ref{fig:spatial_failures} plots the spatial distribution of activation
failures in a \emph{representative} $1024\times1024$ array of DRAM cells
\mpx{taken} from a single DRAM chip\mpx{. Every} observed activation
failure is marked in black. We make two observations. First, we observe that
each contiguous region of 512 DRAM rows\footnote{\jk{We note that subarrays
have either 512 or 1024 (not shown) rows depending on the manufacturer of the
DRAM device.}} \mpx{consists} of \mpx{repeating} rows with the same set (or
subset) of \jk{column bits} that are prone to activation failures. As shown in
the figure, rows 0 to 511 have the same 8 (or a subset of the 8) \jk{column
bits} failing in the row, and rows 512 to 1023 have the same 4 (or a subset of
the 4) \jk{column bits} failing in the row. We hypothesize that these
contiguous regions reveal the DRAM subarray architecture \mpx{as a result of
variation across} the local sense amplifiers in the subarray. \jkthree{We
indicate the two subarrays in Figure~\ref{fig:spatial_failures} as Subarray A
and Subarray B.} A ``weaker'' local sense amplifier results in cells that share
\mpx{its} respective \emph{local bitline} in the subarray \mpx{having} an
increased probability of failure. For this reason, we observe \mpx{that}
activation failures \mpx{are} localized to a few columns within a DRAM subarray
as shown in Figure~\ref{fig:spatial_failures}. Second, we observe that within a
subarray, the activation failure probability \mpx{increases} across rows (i.e.,
activation failures are \emph{more} likely to occur in \jk{higher-numbered}
rows in the subarray and \mpx{are} \emph{less} likely in \jk{lower-numbered}
rows in the subarray). This can be seen \jkthree{from} the fact that more cells
fail in \jk{higher-numbered} rows in the subarray (i.e., there are more black
marks higher in each subarray). We hypothesize that the \jkthree{failure
probability of a cell attached to a local bitline} correlates \mpx{with} the
distance between the row and the local sense amplifiers, and further rows have
less time to amplify their data due to the signal propagation delay in a
bitline. These observations are similar to \mpx{those made in} prior
studies~\cite{lee-sigmetrics2017, chang2016understanding, lee2015adaptive,
kim2018solar} on DDR3 devices.
\begin{figure}[h]
\centering \includegraphics[width=0.7\linewidth]{spatial_array3.pdf}
\caption{Activation failure bitmap in $1024\times1024$ cell array.}
\label{fig:spatial_failures}
\end{figure}
We next study the granularity at which \mpx{we can induce} activation failures
when accessing a row. We observe (not shown) that activation failures
\mpx{occur} \emph{only} within the \mpx{first} cache line that is \mpx{accessed}
immediately following an activation. No subsequent access to an
\jk{already} \emph{open} row results in activation failures. This is because
cells within the \emph{same} row have a longer time to restore their cell
charge (Figure~\ref{fig:dram_op}) when they are accessed after the row
\mpx{has already been opened}. We draw two key conclusions: 1) the region
\emph{and} bitline of DRAM being accessed affect the number of observable
activation failures, and 2) different \mpx{DRAM} subarrays \emph{and}
different local bitlines exhibit varying levels of entropy.
\subsection{Data Pattern Dependence}
\label{subsec:dpd}
To understand the data pattern dependence of activation failures and DRAM cell
entropy, we study \mpy{how effectively we can discover failures using
different data patterns across multiple rounds of testing.} Our \emph{goal} in
this \mpy{experiment} is to determine \mpy{which} data pattern results in the
highest entropy such that we can generate random values with high throughput.
Similar to prior works~\cite{patel2017reaper, liu2013experimental} that
extensively describe the data patterns, we analyze a total of 40 unique data
patterns: solid 1s, checkered, row stripe, column stripe, 16 walking 1s,
\emph{and} the inverses of all 20 aforementioned data patterns.
Figure~\ref{fig:dpd_coverage} plots the \mpy{ratio of activation failures
discovered by a particular data pattern after 100 iterations of
Algorithm~\ref{alg:testing_lat} relative to the \emph{total} number of
failures discovered by \emph{all} patterns for a representative chip from each
manufacturer}. \mpy{We call this metric \emph{coverage} because it
indicates the effectiveness of a single data pattern to identify all possible
DRAM cells that are prone to activation failure. We show results for each
pattern individually except for the WALK1 and WALK0 patterns, for which we
show the mean (bar) and minimum/maximum (error bars) coverage across all 16
iterations of each walking pattern.}
\begin{figure}[h]
\centering \includegraphics[width=\linewidth]{dpd_figure3.pdf}
\caption{Data pattern dependence of DRAM cells prone to activation failure over 100 iterations}
\label{fig:dpd_coverage}
\end{figure}
\mpy{We make three key observations from this experiment.} \mpy{First}, we
find that testing with different data patterns identifies different subsets of
the total set of possible activation failures. This indicates that \jk{1)
different data patterns cause different DRAM cells to fail and 2) specific
data patterns induce more activation failures than others. Thus,} certain data
patterns may extract more entropy from a DRAM cell array than other data
patterns. \mpy{Second}, we find that, of \emph{all} 40 tested data patterns,
each of the 16 \emph{walking 1s}, for a given device, provides a \jk{similarly
high} coverage, regardless of the manufacturer. This high coverage is
similarly provided by only one other data pattern per manufacturer: solid 0s
for manufacturers A and B, and walking 0s for manufacturer C. \mpy{Third, if
we repeat this experiment (i.e., Figure~\ref{fig:dpd_coverage}) while varying
the number of iterations of Algorithm~\ref{alg:testing_lat}, the \emph{total
failure count} across all data patterns \emph{increases} as we increase the
number of iterations of Algorithm~\ref{alg:testing_lat}.} This indicates that
\mpy{not all DRAM cells fail deterministically when accessed with a reduced
$t_{RCD}$, providing a potential source of entropy for random number
generation.}
We next analyze \mpy{each cell's probability of failing when accessed with a
reduced $t_{RCD}$ (i.e., its \emph{activation failure probability}) to
determine which data pattern most effectively identifies cells that provide
high entropy}. We note that DRAM cells with an activation failure probability
$F_{prob}$ of 50\% provide high entropy when accessed many times. With the
same data used to produce Figure~\ref{fig:dpd_coverage}, we study the
different data patterns with regard to the number of cells they cause to fail
50\% of the time. Interestingly, we find that the data pattern that
\jkfour{induces} the most failures overall does not necessarily find the most
number of cells that fail 50\% of the time. In fact, when searching for cells
with an $F_{prob}$ between 40\% and 60\%, \jk{we observe that} the data
\mpy{patterns that find} the \jk{highest number of cells} \mpy{are} solid 0s,
checkered 0s, and solid 0s for manufacturers A, B, and C, respectively. We
conclude that: 1) due to manufacturing and design variation across DRAM
devices from different manufacturers, different data patterns \mpy{result in}
different failure probabilities in our DRAM devices, and 2) to provide high
entropy \mpy{when} accessing DRAM cells with a reduced $t_{RCD}$, we should
use the respective data pattern that finds the most number of cells with an
$F_{prob}$ of 50\% for DRAM devices from a given manufacturer.
Unless otherwise stated, in the rest of this paper, we use \mpy{the} solid 0s,
checkered 0s, and solid 0s data patterns for manufacturers A, B, and C,
respectively, to analyze $F_{prob}$ at the granularity of a single cell
\mpy{and} to study the effects of temperature and time on our sources of
entropy.
\subsection{Temperature Effects}
\label{subsec:temp_effects}
In this section, we study whether temperature \mpy{fluctuations affect} a DRAM
cell's activation failure probability and thus the entropy that can be
extracted from the DRAM cell. To analyze \mpy{temperature effects}, we record
the $F_{prob}$ of cells \mpy{throughout} our DRAM devices \mpy{across 100
iterations of Algorithm~\ref{alg:testing_lat} at 5$^{\circ}$C increments}
between 55$^{\circ}$C and 70$^{\circ}$). Figure~\ref{fig:temperature_effects}
aggregates \mpy{results} across 30 DRAM modules from each DRAM manufacturer.
\jk{Each point in the figure represents how the $F_{prob}$ of a DRAM cell
changes as the temperature changes (i.e., ${\Delta}F_{prob}$). The x-axis
shows the $F_{prob}$ of a single cell at temperature $T$ (i.e., the baseline
temperature), and the y-axis shows the $F_{prob}$ of the same cell at
temperature $T+5$ (i.e., 5$^{\circ}$C above the baseline temperature).}
Because we test each cell at each temperature across 100 iterations, the
granularity \mpy{of} $F_{prob}$ on both the x- and y-axes is 1\%. For a given
$F_{prob}$ at temperature $T$ \jk{(x\% on the x-axis)}, we aggregate
\emph{all} respective $F_{prob}$ points at temperature $T+5$ \jk{(y\% on the
y-axis)} with box-and-whiskers plots\footnote{A box-and-whiskers plot
emphasizes the important metrics of a dataset's distribution. The box is
lower-bounded by the first quartile (i.e., the median of the first half of the
ordered set of data points) and upper-bounded by the third quartile (i.e., the
median of the second half of the ordered set of data points). The median falls
within the box. The \emph{inter-quartile range} (IQR) is the distance between
the first and third quartiles (i.e., box size). Whiskers extend an additional
$1.5 \times IQR$ on either sides of the box. We indicate outliers, or data
points outside of the range of the whiskers, with pluses.} to show how the
given $F_{prob}$ is affected by the increased DRAM temperature. The
\emph{box} is drawn in blue and contains the \emph{median} drawn in red. The
\emph{whiskers} are drawn in gray, and the \emph{outliers} are indicated with
orange pluses.
\begin{figure}[h]
\centering \includegraphics[width=\linewidth]{temperature_effects_fixed2.pdf}
\caption{Effect of temperature variation on failure \hhf{probability}}
\label{fig:temperature_effects}
\end{figure}
We observe that $F_{prob}$ at temperature $T+5$ tends to be higher than
$F_{prob}$ at temperature $T$\jkfour{,} \mpy{as shown by} the blue region of the figure
(i.e., the boxes of the box-and-whiskers plots) \mpy{lying} above the $x~=~y$
line. However, \mpy{fewer} than 25\% of all data points fall below the $x~=~y$
line, indicating that a portion of cells have a lower $F_{prob}$ as
temperature is increased.
We observe that DRAM devices from different manufacturers are affected by
temperature differently. DRAM cells of manufacturer A have the \emph{least}
variation of ${\Delta}F_{prob}$ when temperature is increased since the boxes
of the box-and-whiskers plots are strongly correlated with the $x~=~y$ line.
\jk{How a DRAM cell's activation failure probability changes in DRAM devices
from \emph{other} manufacturers is unfortunately \emph{less} predictable
under} temperature change (i.e., a DRAM cell from manufacturers B or C has
higher variation in $F_{prob}$ change), but \mpy{the data still shows a strong
positive correlation between temperature and $F_{prob}$}. We conclude that
temperature affects \jk{cell failure probability ($F_{prob}$)} to different
degrees depending on the manufacturer of the DRAM device, \mpy{but increasing
temperature generally increases the activation failure probability.}
\subsection{Entropy Variation over Time}
\label{subsec:short_term_variation}
To determine whether the failure probability of a DRAM cell \jk{changes over
time}, we complete 250 \emph{rounds} of recording the activation failure
probability of DRAM cells over the span of 15 days. Each round consists of
accessing every cell in DRAM 100 times with a reduced $t_{RCD}$ value and
recording the failure probability for each individual cell (out of 100
iterations). We \mpy{find} that a DRAM cell's activation failure probability
does \emph{not} change significantly over time. This means that, \mpy{once we
identify} a DRAM cell that exhibits high entropy, we can rely on the cell to
maintain its high entropy over time. We hypothesize that this is because
\jkfour{a DRAM cell fails with high entropy} when process manufacturing
variation in peripheral and DRAM cell circuit elements combine such that, when
we read the cell using a reduced $t_{RCD}$ value, we induce a metastable state
resulting from the cell voltage falling between the reliable sensing margins
(i.e., falling close to $\frac{V_{dd}}{2}$)~\cite{chang2016understanding}.
Since manufacturing variation is fully determined at manufacturing time, a
\jkfour{DRAM cell's activation failure probability} is stable over time given
the same experimental conditions. In Section~\ref{subsec:cell_selection}, we
discuss our \mpy{methodology} for selecting DRAM cells for extracting stable
entropy, such that we can preemptively avoid longer-term aging effects that we
do not study in this paper.
\delete{\subsection{Effects of Reduced $t_{RCD}$ on DRAM Writes}}
\delete{We rely on inducing activation failures on DRAM cells having the same data
pattern each time. This means that between every reduced $t_{RCD}$ access,
we must restore, i.e., rewrite, the data pattern in the recently accessed DRAM
word. Therefore, we would like to study the effects that a reduced $t_{RCD}$
has on DRAM write accesses to see whether we need to reset $t_{RCD}$ to its
default value before restoring the data pattern in the DRAM cells. To test
whether DRAM writes are affected by a reduced $t_{RCD}$ value, we run two
experiments with our set of DRAM devices. First, we sweep the value of
$t_{RCD}$ between $10ns$ and $18ns$ and write a known data pattern across DRAM. We
then subsequently check the DRAM array for unexpected data and we observe no
failures in this range of $t_{RCD}$. }
\delete{In our second experiment, we would like to verify that a shortened write
command does not result in side-effects such as lower voltage levels in the
written DRAM cells. Thus, for a range of $t_{RCD}$ values, we run the following
test. For a $t_{RCD}$ value, we write a known data pattern into the DRAM
array, and disable DRAM refresh, so the charge in every cell will drain over
time. We then compare the failure rates across the DRAM array for a set
interval with refresh disabled. We note highly similar failure rates regardless
of the value of $t_{RCD}$ and conclude that writing to DRAM with a
significantly reduced $t_{RCD}$ has no impact on the quality of the writes.
This enables DRAM to service application writes between reduced $t_{RCD}$
accesses for true random number generation to minimize system interference.}
\section{\mechanism: A DRAM-based TRNG}
\label{section:DLRNG_mechanism}
\jkfour{Based on} our rigorous analysis of \mpy{DRAM} activation failures
(presented in Section~\ref{sec:dlrng_characterization}), we propose \mechanism,
a flexible \mpy{mechanism} that provides high-throughput DRAM-based true random
number \mpy{generation} (TRNG) by sourcing entropy from a subset of DRAM cells
\jk{and is built fully within the memory controller.} \mechanism~is \mpy{based
on the \textbf{key observation}} that DRAM cells fail \mpy{probabilistically}
when accessed with reduced DRAM timing parameters\jk{, and this probabilistic
failure mechanism can be used as a source of true random numbers}. While there
are many other timing parameters that we could \mpy{reduce} to induce failures
in DRAM~\cite{chang2016understanding, lee-sigmetrics2017, lee2015adaptive,
kim2018solar, lee2016reducing, chang2017thesis}, we focus specifically on
reducing $t_{RCD}$ below manufacturer-recommended values to study the resulting
activation failures.\footnote{We believe that reducing other timing parameters
could be used to generate true random values, but we leave their exploration to
future work.}
Activation failures occur as a result of reading the value from a DRAM cell
\emph{too soon} after sense amplification. This results in reading the value
at the sense amplifiers before the bitline voltage is amplified to an
\jk{I/O-readable} voltage level. The probability of reading incorrect data
from the DRAM cell therefore depends largely \jk{on the bitline's voltage} at
the time of reading the sense amplifiers. Because there is significant process
variation across the DRAM cells and I/O
circuitry~\cite{chang2016understanding, lee-sigmetrics2017, lee2015adaptive,
kim2018solar}, we observe a wide \mpy{variety} of failure probabilities for
different DRAM cells (as discussed in
Section~\ref{sec:dlrng_characterization}) for a given $t_{RCD}$ value\mpy{,
ranging from 0\% probability to 100\% probability}.
\textbf{We discover that a subset of cells fail at \mpy{$\jkfour{\sim}$}50\%
probability, and a subset of these cells fail randomly with high entropy
(shown in Section~\ref{subsec:diff_chips}).} In this section, we first discuss
our method of identifying such cells, \jk{which} we refer to as \emph{RNG
cells} (in Section~\ref{subsec:cell_selection}). Second, we describe the
mechanism with which D-RaNGe \emph{samples} RNG cells to extract random data
(Section~\ref{subsec:sampling_rng_cells}). Finally, we discuss a potential
design for integrating D-RaNGe \jk{in} a full system
(Section~\ref{subsec:full_system_integration}).
\subsection{RNG Cell Identification}
\label{subsec:cell_selection}
Prior to generating random data, we must first identify cells that are capable
of producing truly random output (i.e., RNG cells). Our process of
identifying RNG cells involves reading every cell in the DRAM array 1000 times
with a \emph{reduced} $t_{RCD}$ and approximating each cell's Shannon
entropy~\cite{shannon1948mathematical} by counting the occurrences of 3-bit
symbols across its 1000-bit stream. We identify cells that \mpy{generate an
approximately equal} number of every \mpy{possible} 3-bit symbol ($\pm10\%$ of
the number of expected symbols) as RNG cells.
We find that \mpy{RNG cells provide unbiased output, meaning that a}
post-processing step (described in Section~\ref{subsec:trng}) is \emph{not}
necessary \mpy{to provide sufficiently high entropy for random number
generation}. We also find that RNG cells \emph{maintain high entropy across
system reboots}. In order to account for our observation that entropy from an
RNG cell changes depending on the DRAM temperature
\jk{(Section~\ref{subsec:temp_effects})}, we identify reliable RNG cells at
each temperature and store their locations in the memory controller. Depending
on the DRAM temperature \jk{at the time an application requests random
values,} D-RaNGe samples the appropriate RNG cells. To ensure that DRAM aging
does not negatively impact the reliability of RNG cells, we require
re-identifying the set of RNG cells at regular intervals. From our observation
that entropy does not change significantly over a tested 15 day period of
sampling RNG cells
\jk{(Section~\ref{subsec:short_term_variation})}, we expect the interval of
re-identifying RNG cells to be at least 15 days long. Our RNG cell
identification process is effective at identifying cells that are reliable
entropy sources for random number generation, and we quantify their randomness
using the NIST test suite for randomness~\cite{rukhin2001statistical} in
Section~\ref{subsec:nist}.
\subsection{Sampling RNG Cells for Random Data}
\label{subsec:sampling_rng_cells}
Given the availability of these RNG cells, we use our observations in
Section~\ref{sec:dlrng_characterization} to design a high-throughput TRNG that
quickly and repeatedly samples RNG cells with reduced DRAM timing parameters.
Algorithm~\ref{alg:lat_rng} demonstrates the key components of \mechanism~that
enable us to generate random numbers with high throughput.
\begin{algorithm}[tbh]\footnotesize
\SetAlgoNlRelativeSize{0.7}
\SetAlgoNoLine
\DontPrintSemicolon
\SetAlCapHSkip{0pt}
\caption{\mechanism:~A DRAM-based TRNG}
\label{alg:lat_rng}
\textbf{\mechanism($num\_bits$):} \textcolor{gray}{~~~// \emph{num\_bits}: number of random bits requested} \par
~~~~$DP$: a known data pattern that results in high entropy \par
~~~~select 2 DRAM words with RNG cells in distinct rows in each bank \par
~~~~write $DP$ to chosen DRAM words and their neighboring cells \par
~~~~get exclusive access to rows of chosen DRAM words and nearby cells \par
~~~~set low $t_{RCD}$ for DRAM ranks containing chosen DRAM words \par
~~~~\textbf{for} each bank: \par
~~~~~~~~read data in $DW_1$ \textcolor{gray}{~~// induce activation failure} \par
~~~~~~~~write the read value of $DW_1$'s RNG cells to $bitstream$ \par
~~~~~~~~write original data value back into $DW_1$ \par
~~~~~~~~memory barrier \textcolor{gray}{~~~~// ensure completion of write to $DW_1$} \par
~~~~~~~~read data in $DW_2$ \textcolor{gray}{~// induce activation failure} \par
~~~~~~~~write the read value of $DW_2$'s RNG cells to $bitstream$ \par
~~~~~~~~write original data value back into $DW_2$ \par
~~~~~~~~memory barrier \textcolor{gray}{~~~// ensure completion of write to $DW_2$} \par
~~~~~~~~\textbf{if} $bitstream_{size}$$~\geq~num\_bits$: \par
~~~~~~~~~~~~break \par
~~~~set default $t_{RCD}$ for DRAM ranks of the chosen DRAM words \par
~~~~release exclusive access to rows of chosen words and nearby cells \par
\end{algorithm}
\mechanism~ takes in $num\_bits$ as an argument, which is defined as the
number of random bits desired (Line~1). \mechanism~then prepares to generate
random numbers in Lines~2-6 by first selecting DRAM words (i.e., the
granularity at which \jk{a DRAM module} is accessed) containing known RNG cells
for generating random data (Line~3). To maximize \jkfour{the} throughput of
random number generation, \mechanism~\mpy{chooses} DRAM words with the highest
density of RNG cells in each bank (to exploit DRAM parallelism). Since each
DRAM access can induce activation failures \emph{only} in the accessed DRAM
word, the density of RNG cells per DRAM word determines the number of random
bits D-RaNGe can generate per access.
For each available DRAM bank, \mechanism~selects two DRAM words (in
distinct DRAM rows) containing RNG cells. The purpose of selecting two DRAM
words in \emph{different} rows is to \emph{repeatedly} cause \emph{bank conflicts}, or
\jk{issue} requests to \emph{closed} DRAM rows so that every read request will
\emph{immediately} follow an activation. This is done by alternating accesses
to the chosen DRAM words in different DRAM rows. After selecting DRAM words
for generating random values, \mechanism~ writes a known data pattern that
results in high entropy to each chosen DRAM word and its neighboring cells
(Line~4) and gains exclusive access to rows containing the two chosen DRAM
words as well as their neighboring cells (Line~5).\footnote{Ensuring
exclusive access to DRAM rows can be done by remapping rows to 1) redundant
DRAM rows or 2) buffers in the memory controller so that these rows are hidden
from the \jkfour{system software} and only accessible by the memory controller for
generating random numbers. } This ensures that the data pattern surrounding
the RNG cell and the original value of the RNG cell stay constant prior to each
access such that the failure probability of each RNG cell remains reliable (as
observed to be necessary in Section~\ref{subsec:dpd}). To begin generating
random data (i.e., sampling RNG cells), \mechanism~reduces the value of
$t_{RCD}$ (Line~6). From every available bank \jk{(Line~7)},
\mechanism~generates random values in parallel (Lines~8-15). Lines 8 and 12
indicate the commands to alternate accesses to two DRAM words in distinct rows
of a bank to both 1) induce activation failures and 2) precharge the
\hhtwo{recently-accessed} row. After inducing activation failures in a DRAM
word, \mechanism~extracts the value of the RNG cells within the DRAM word
(Lines~9 and 13) to use as random data and restores the DRAM word to its
original data value (Lines~10 and 14) to maintain the original data pattern.
Line~15 ensures that writing the original data value is complete before
attempting to \jk{sample} the DRAM words again. Lines~16 and 17 simply end the
loop if enough random bits of data have been harvested. Line~18 sets the
$t_{RCD}$ timing parameter back to its default value, so other applications can
access DRAM without corrupting data. Line~19 releases exclusive access to the
rows containing the chosen DRAM words and their neighboring rows.
We find that this methodology maximizes the opportunity for activation failures
in DRAM, thereby maximizing the rate of generating random data from RNG
cells.
\subsection{Full System Integration}
\label{subsec:full_system_integration}
In this work, we focus on developing a flexible substrate for sampling RNG
cells fully from within the memory controller. D-RaNGe generates random
numbers using a simple firmware routine running entirely within the memory
controller. The firmware executes the sampling algorithm
(Algorithm~\ref{alg:lat_rng}) whenever \mpx{an application requests random
samples and there is available DRAM} bandwidth (i.e., DRAM is not servicing other
\jk{requests} or maintenance commands). In order to minimize latency between
requests for samples and their \mpx{corresponding} responses, a small queue of
already-harvested random data may be maintained in the memory controller for
use by the system. Overall performance overhead can be minimized by tuning
\mpx{both 1) the queue size and 2) how the memory
controller prioritizes requests for random numbers relative to normal memory
requests}.
In order to integrate \mechanism~with the rest of the system, the system
designer needs to decide how to best expose an interface by which an
application can \jk{leverage \mechanism~to generate true random numbers} on
their system. There are many ways to achieve this, including, but not limited
to:
\begin{itemize}
\item \mpx{Providing a simple \texttt{REQUEST} and \texttt{RECEIVE} interface
for applications to request and receive the random numbers using memory-mapped
configuration status registers (CSRs)~\cite{wolrich2004mapping} or other existing I/O
datapaths (e.g., x86 \texttt{IN} and \texttt{OUT} opcodes, Local Advanced
Programmable Interrupt Controller (LAPIC configuration \cite{intel32intel}).}
\item \mpx{Adding a new ISA instruction (e.g., Intel
\texttt{RDRAND}~\cite{hamburg2012analysis}) that retrieves random numbers from
the memory controller and stores them into processor registers.}
\end{itemize}
The operating system may then expose \mpx{one or more of these interfaces to
user applications through standard kernel-user interfaces (e.g., system calls,
file I/O, operating system APIs). The system designer has complete freedom to
choose between these (and other) mechanisms that expose an interface for user
applications to interact with D-RaNGe. We expect that the best option will be
\hhtwo{system specific}, depending both on the desired D-RaNGe use cases and the ease
with which the design can be implemented.}
\section{\mechanism~Evaluation}
\label{sec:evaluation}
We evaluate \mpy{three key aspects of} \mechanism. \mpy{First, we} show that
the random data \jkthree{obtained from RNG cells identified by D-RaNGe}
\mpy{passes} all of the tests in the NIST test suite for randomness
(Section~\ref{subsec:nist}). Second, we analyze the
\jkthree{existence} of RNG cells across 59 LPDDR4 and 4 DDR3 DRAM chips (due to long
testing time) randomly sampled from the overall population of DRAM chips across
all three major DRAM manufacturers (Section~\ref{subsec:diff_chips}). Third, we
evaluate \mechanism~in terms of the six key properties of an ideal TRNG as
explained in Section~\ref{sec:motivation}
(Section~\ref{subsec:trng_key_char_eval}).
\subsection{NIST Tests}
\label{subsec:nist}
First, we identify RNG cells using our RNG cell \jkthree{identification}
process (Section~\ref{subsec:cell_selection}). Second, we sample \jkfour{\emph{each}}
identified RNG \hhf{cell} \jkthree{one million times} to generate large amounts of
random data (\mpy{i.e.,} 1~Mb \emph{bitstreams}). Third, we evaluate the
entropy of the bitstreams from the identified RNG cells with the NIST test
suite for randomness~\cite{rukhin2001statistical}. Table~\ref{Tab:NIST} shows
the average results of 236 1~Mb bitstreams\footnote{We test data obtained from
4 RNG cells from each of 59 DRAM chips, to maintain a reasonable NIST testing
time and \jkfour{thus} show that RNG cells across all \jkfour{tested} DRAM
chips reliably generate random values.} across the 15 tests of the full NIST
\begin{table}[h!]
\footnotesize
\begin{center}
\begin{tabular}{ |c||c|c|c }
\cline{1-3}
\textbf{NIST Test Name} & \textbf{P-value} & \textbf{Status} \\
\cline{1-3}
\hhline{|=|=|=|}
monobit & 0.675 & PASS \\
frequency\_within\_block & 0.096 & PASS \\
runs & 0.501 & PASS \\
longest\_run\_ones\_in\_a\_block & 0.256 & PASS \\
binary\_matrix\_rank & 0.914 & PASS \\
dft & 0.424 & PASS \\
non\_overlapping\_template\_matching & >0.999 & PASS \\
overlapping\_template\_matching & 0.624 & PASS \\
maurers\_universal & 0.999 & PASS \\
linear\_complexity & 0.663 & PASS \\
serial & 0.405 & PASS \\
approximate\_entropy & 0.735 & PASS \\
cumulative\_sums & 0.588 & PASS \\
random\_excursion & 0.200 & PASS \\
random\_excursion\_variant & 0.066 & PASS \\
\cline{1-3}
\end{tabular}
\caption{\hhf{\mechanism\ results} with NIST randomness test suite. \vspace{-10pt}}
\label{Tab:NIST}
\end{center}
\end{table}
test suite for randomness. P-values are calculated for each test,\footnote{A
p-value close to 1 indicates that we must accept the null hypothesis, while a
p-value close to 0 and below a small threshold, e.g., $\alpha=0.0001$
(\jkthree{recommended by the NIST Statistical Test Suite
documentation~\cite{rukhin2001statistical}}), indicates that we must reject the
null hypothesis.} where the null hypothesis for each test is that a perfect
random number generator would \emph{not} have produced random data with
\emph{better} characteristics for the given test than the tested
sequence~\cite{marton2015interpretation}. Since the resulting P-values for each
test in the suite are greater than our chosen level of significance,
$\alpha=0.0001$, we accept our null hypothesis for each test. We note that all
236 bitstreams pass all 15 tests with similar P-values. Given our
$\alpha=0.0001$, our proportion of passing sequences (1.0) falls within the
range of acceptable proportions of sequences that pass each test
(\jkthree{[0.998,1]} calculated by \mpy{the} NIST statistical test suite using
$(1-\alpha)\pm3\sqrt{\frac{\alpha(1-\alpha)}{k}}$, where $k$ is the number of
tested sequences). This \emph{strongly} indicates that \mechanism~can generate
unpredictable, \mpy{truly} random values. Using the proportion of 1s and 0s
generated from each RNG cell, we calculate \mpy{Shannon}
entropy~\cite{shannon1948mathematical} and find the \emph{minimum} entropy
across all RNG cells to be 0.9507.
\subsection{RNG Cell Distribution}
\label{subsec:diff_chips}
The throughput at which \mechanism~generates random numbers is a function of
the 1) density of RNG cells per DRAM word and 2) bandwidth \jkthree{with which we
can access} DRAM words when using our methodology for inducing activation
failures. Since each DRAM access can induce activation failures \emph{only} in
the accessed DRAM word, the density of RNG cells per DRAM word indicates the
number of random bits D-RaNGe can sample per access. We first study the
density of RNG cells per word across DRAM \jkthree{chips}.
Figure~\ref{fig:bits_per_word} plots the distribution of the number of words
containing x RNG cells (\jkthree{indicated by the value on the} x-axis) per
\emph{bank} across 472 banks from 59 DRAM devices from \jkthree{all manufacturers}.
The distribution is presented as a box-and-whiskers plot where the y-axis
\jkthree{has} a logarithmic scale with a 0 point. The three plots respectively
show the distributions for DRAM devices from the three manufacturers
(indicated at the bottom left corner of each plot).
\begin{figure}[h]
\centering \includegraphics[width=\linewidth]{bits_per_word2.pdf}
\caption{Density of RNG cells in DRAM words per bank.}
\label{fig:bits_per_word}
\end{figure}
We make three key observations. First, RNG cells are \emph{widely available in
every bank} across many chips. This means that we can use the available DRAM
access parallelism that multiple banks offer and sample RNG cells from each
DRAM bank in parallel to improve random number generation throughput. Second,
\emph{every} bank that we \mpy{analyze} has \emph{multiple DRAM words}
containing at least one RNG cell. The DRAM bank with the smallest
\mpy{occurrence} of RNG cells has 100 DRAM words containing only 1 RNG cell
(manufacturer B). Discounting this point, the distribution of the number of
DRAM words containing only 1 RNG cell is tight with a high number of RNG cells
(e.g., tens of thousands) in each bank, regardless of the manufacturer. Given
our random sample of DRAM chips, we expect that the existence of RNG cells in
DRAM banks will hold true for all DRAM chips. Third, we observe that a single
DRAM word can contain as many as 4 RNG cells. Because the throughput of
accesses to DRAM is fixed, the number of RNG cells in the accessed words
essentially acts as a multiplier for the throughput of random numbers
generated (e.g., accessing DRAM words containing 4 RNG cells results in
\emph{4x} the throughput of random numbers compared to accessing DRAM words
containing 1 RNG cell).
\subsection{TRNG Key Characteristics Evaluation}
\label{subsec:trng_key_char_eval}
We now evaluate \mechanism~in terms of the six key properties of an effective
TRNG as explained in Section~\ref{sec:motivation}.
\textbf{Low Implementation Cost.} To induce activation failures, we must be
able to reduce the DRAM timing parameters below manufacturer-specified values.
Because memory controllers issue memory accesses according to the timing
parameters specified in a set of internal registers, D-RaNGe requires
\jkfour{simple} software support to be able to programmatically modify the
memory controller's registers. Fortunately, there exist some
processors~\cite{lee2015adaptive, opteron_amd, bkdg_amd2013, ram_overclock}
that \emph{already} enable software to directly change memory controller
register values, i.e., the DRAM timing parameters. These processors can easily
generate random numbers with \mechanism.
All other processors that do \emph{not} currently support direct changes to
memory controller registers require \emph{minimal} software changes to expose
an interface for changing the memory controller
registers~\cite{ARM2016memrefman, samsung2014refman, hassan2017softmc,
softmc-safarigithub}. To \jkthree{enable} a more efficient implementation, the
memory controller could be programmed such that it issues DRAM accesses with
distinct timing parameters on a per-access granularity to reduce the overhead
in 1) changing the DRAM timing parameters and 2) allow concurrent DRAM
accesses by other applications. In the rare case where these registers are
unmodifiable by even the hardware, the hardware changes necessary to enable
register modification \mpy{are} minimal and \mpy{are} simple to
implement~\cite{lee2015adaptive, hassan2017softmc, softmc-safarigithub}.
We experimentally find that we can induce activation failures \mpy{with}
$t_{RCD}$ between $6ns$ and $13ns$ (\mpy{reduced} from the default of $18ns$). Given
this wide range of failure-inducing $t_{RCD}$ values, most memory controllers
should be able to adjust their timing parameter registers to a value within
this range.
\textbf{Fully Non-deterministic.} As we have shown in
Section~\ref{subsec:nist}, the bitstreams extracted from the
\jkthree{D-RaNGe-identified} RNG cells pass \emph{all} 15 NIST tests\jkthree{. We have}
full reason to believe that we are inducing a metastable state of the sense
amplifiers (as hypothesized by~\cite{chang2016understanding}) such that we are
effectively sampling random physical phenomena to extract unpredictable random
values.
\textbf{High Throughput of Random Data.} Due to the various use cases of random
number generation discussed in Section~\ref{sec:motivation}, different
applications have different throughput requirements for random number
generation, and applications may tolerate a reduction in performance so
that \mechanism~can quickly generate true random numbers. Fortunately,
\mechanism~provides flexibility to tradeoff between the \textit{system
interference} it causes, i.e., the slowdown experienced by concurrently running
applications, and the random number generation throughput it provides. To
demonstrate this flexibility, Figure~\ref{fig:trng_throughput} plots the TRNG
throughput of \mechanism~when using varying numbers of banks (\emph{x} banks on
the x-axis) across the three DRAM manufacturers (indicated at the top left
corner of each plot). For each number of banks used, we plot the distribution
of TRNG throughput that we observe \emph{real} DRAM devices to provide\jkthree{. The
available density of RNG cells in a DRAM device (provided in
Figure~\ref{fig:bits_per_word}) dictates the TRNG throughput that the DRAM
device can provide. We plot each distribution} as
a box-and-whiskers plot. For each number of banks used, we select x
banks with the greatest sum of RNG cells across each banks' two DRAM words with
the highest density of RNG cells (that are \emph{not} in the same DRAM row).
We select two DRAM words per bank because we must alternate accesses between
two DRAM rows (as shown in \jkfour{Lines 8 and 12} of Algorithm~\ref{alg:lat_rng}).
The sum of the RNG cells available across the two selected DRAM words for each
bank is considered each bank's \emph{TRNG data rate}, and we use this value to
obtain \mechanism's throughput. We use Ramulator~\cite{ramulatorgithub,
kim2016ramulator} to obtain the rate at which we can execute the core loop of
Algorithm~\ref{alg:lat_rng} with varying numbers of banks. We obtain the random
number generation throughput for x banks with the following equation: \vspace{-4pt}
\begin{equation} \label{eq:x_bank_throughput}
TRNG\_Throughput_{x\_Banks} = \sum_{n=1}^{x} \frac{TRNG\_data\_rate_{Bank\_n}}{Alg2\_Runtime_{x\_banks}}
\end{equation}
where $TRNG\_data\_rate_{Bank\_n}$ is the TRNG data rate for the selected bank,
and $Alg2\_Runtime_{x\_banks}$ is the runtime of the core loop of
Algorithm~\ref{alg:lat_rng} when using $x$ Banks. We note that because we
observe small variation in the density of RNG cells per word (between 0 and 4),
we see that \jkthree{TRNG throughput across different chips is} generally very similar.
For this reason, we see that the box and whiskers are condensed into a single
point for distributions of manufacturers B and C. We find that when
\emph{fully} using \emph{all} 8 banks in a single DRAM channel, every
device can produce \emph{at least} 40 Mb/s of random data regardless of
manufacturer. The highest throughput we observe from devices of manufacturers
A/B/C respectively are \changes{179.4/134.5/179.4 Mb/s}. On average, across all
manufacturers, we find that \mechanism~can provide a throughput of
\changes{108.9 Mb/s.}
\begin{figure}[h]
\centering \includegraphics[width=\linewidth]{trng_throughput3.pdf}
\caption{Distribution of TRNG throughput across chips.}
\label{fig:trng_throughput}
\end{figure}
We draw two key conclusions. First, due to the \emph{parallelism} of multiple
banks, the throughput of random number generation increases linearly as we
use more banks. Second, there is variation of TRNG throughput across
different DRAM devices, but the medians across manufacturers are very
similar.
We note that any throughput sample point on this figure can be multiplied by
the number of available channels in a memory hierarchy for a better TRNG
throughput estimate for a system with multiple DRAM channels. For an example
memory hierarchy comprised of 4 DRAM channels, \mechanism~results in a maximum
(average) throughput of 717.4 Mb/s (435.7 Mb/s).
\textbf{Low Latency.} \jkfour{Since D-RaNGe's sampling mechanism consists of a
single DRAM access, the latency of generating random values is directly related
to the DRAM access latency. Using the timing parameters specified in the JEDEC
LPDDR4 specification~\cite{2014lpddr4}, we calculate D-RaNGe's latency to
generate a 64-bit random value. To \hhf{calculate the} \emph{maximum latency} for D-RaNGe,
we assume that 1) each DRAM access provides only 1 bit of random data (i.e.,
each DRAM word contains \emph{only} 1 RNG cell) and 2) we can use only a single
bank within a single channel to generate random data. We find that D-RaNGe can
generate 64 bits of random data with a \emph{maximum} latency of $960 ns$. If
D-RaNGe takes full advantage of DRAM's \hhf{channel- and bank-level}
parallelism in a system with 4 DRAM channels \jkfive{and 8 banks per channel,} D-RaNGe
can generate 64 bits of random data by issuing 16 DRAM accesses per channel in
parallel. This results in a latency of $220 ns$. To \hhf{calculate the}
\emph{empirical minimum latency} for D-RaNGe, we fully parallelize D-RaNGe
\hhf{across banks in all 4 channels} while \hhf{also} assuming that each DRAM
access provides 4 bits of random data, since we find a maximum density of 4 RNG
cells per DRAM word in the LPDDR4 DRAM devices that we characterize
(Figure~\ref{fig:bits_per_word}). We find the empirical minimum latency to be
\emph{only} $100 ns$ in our tested devices.}
\textbf{Low System Interference.} The flexibility of using a different number
of banks across the available channels in a system's memory hierarchy allows
\mechanism~to \jkthree{cause} varying levels of system interference at the
expense of TRNG throughput. \jkfour{This enables application developers to
generate random values with D-RaNGe at varying tradeoff points} depending on
the running applications' \jkfour{memory access} requirements. We analyze
\mechanism's system interference with respect to DRAM storage overhead and DRAM
latency.
In terms of storage overhead, \mechanism~simply requires exclusive access
rights to \jkfour{six DRAM rows} per bank, consisting of the two rows
containing the RNG cells and each row's two \jkfour{physically-adjacent} DRAM
rows containing the chosen data pattern.\footnote{As in prior
work~\cite{kim2014flipping, bains2015method, mutlu2017rowhammer}, we argue that
manufacturers can disclose \jkfour{which rows are physically adjacent to each
other.}} This results in an insignificant 0.018\% DRAM storage overhead cost.
\jkthree{To evaluate} \mechanism's effect on \mpy{the} DRAM \jkthree{access
latency of regular memory requests}, we present one implementation of
\mechanism. For a single DRAM channel, which is the granularity at which DRAM
timing parameters are applied, \mechanism~can alternate between using a
reduced $t_{RCD}$ and the default $t_{RCD}$. When using a reduced $t_{RCD}$,
\mechanism~\mpy{generates} random numbers across every bank in the channel. On
the other hand, when using the default $t_{RCD}$, memory requests from running
applications \mpy{are} serviced to ensure application progress. The length of
these time intervals (with default/reduced $t_{RCD}$) can both be adjusted
according to the applications' random number \jkthree{generation}
requirements. Overall,
\mechanism~\jkthree{provides} significant flexibility in \jkthree{trading off} its system
overhead \jkthree{with its TRNG} throughput. \jkthree{However,} it is
up to the system designer to use and exploit the flexibility for their
requirements. To show the potential throughput of \mechanism~without impacting
\jkthree{concurrently-running} applications, we run simulations with
\jkfour{the SPEC CPU2006}~\cite{spec2006} workloads, and calculate the idle DRAM bandwidth
available that we can use to issue D-RaNGe commands. We find \mpy{that,}
across all workloads, we can obtain an average (maximum, minimum) random-value
throughput of 83.1 (98.3, 49.1) Mb/s \mpy{with \emph{no} significant impact on
overall system performance}.
\textbf{Low Energy Consumption.} To evaluate the energy consumption of D-RaNGe,
we use DRAMPower~\cite{drampowergithub} to analyze the output traces of
Ramulator~\cite{ramulatorgithub, kim2016ramulator} when DRAM is (1)
generating random numbers (Algorithm~\ref{alg:lat_rng}), and (2) idling and not
servicing memory requests. We subtract quantity (2) from (1) to
obtain the estimated energy consumption of D-RaNGe. We then divide the value by
the total number of random bits found during execution and find that, on
average, D-RaNGe finds random bits at the cost of 4.4 nJ/bit.
\section{Comparison with Prior DRAM TRNGs}
\label{comparison}
To our knowledge, this paper provides the highest-throughput TRNG \hh{\emph{for
commodity DRAM devices}} by exploiting activation failures as a sampling
mechanism for observing entropy in DRAM cells. There are a number of proposals
to construct TRNGs using commodity DRAM devices, which \jkthree{we summarize} in
Table~\ref{tab:prior_dram_works} based on their entropy sources. In this
section, we compare each of these works with \mechanism. \jkthree{We} show how
\mechanism\ fulfills the six key properties of an ideal TRNG
(Section~\ref{sec:motivation}) better than any prior DRAM-based TRNG proposal.
We group our comparisons by the entropy source of each prior DRAM-based TRNG
proposal.
\begin{table*}[h!]
\footnotesize
\begin{center}
\begin{tabular}{ |c||c|c|c|c|c|c|c| }
\hline
\textbf{Proposal}
& \textbf{Year}
& \begin{tabular}{@{}c@{}}\textbf{Entropy} \\ \textbf{Source}\end{tabular}
& \begin{tabular}{@{}c@{}}\textbf{True} \\ \textbf{Random}\end{tabular}
& \begin{tabular}{@{}c@{}}\textbf{Streaming} \\ \textbf{Capable}\end{tabular}
& \begin{tabular}{@{}c@{}}\textbf{\hhf{64-bit TRNG}} \\ \textbf{Latency}\end{tabular}
& \begin{tabular}{@{}c@{}}\textbf{Energy} \\ \textbf{Consumption}\end{tabular}
& \begin{tabular}{@{}c@{}}\textbf{Peak} \\ \textbf{Throughput}\end{tabular} \\
\hline \hline
Pyo+~\cite{pyo2009dram}
& 2009
& Command Schedule
& \xmark
& \cmark
& $18{\mu}s$
& N/A
& $3.40 Mb/s$ \\
\hline
Keller+~\cite{keller2014dynamic}
& 2014
& Data Retention
& \cmark
& \cmark
& $40 s$
& $6.8mJ/bit$
& $0.05 Mb/s$ \\
\hline
Tehranipoor+~\cite{tehranipoor2016robust}
& 2016
& Startup Values
& \cmark
& \xmark
& $>60ns$ \jkfour{(optimistic)}
& $>245.9 pJ/bit$ \jkfour{(optimistic)}
& N/A \\
\hline
Sutar+~\cite{sutar2018d}
& 2018
& Data Retention
& \cmark
& \cmark
& $40 s$
& $6.8 mJ/bit$
& $0.05 Mb/s$ \\
\hline
\textbf{\mechanism}
& 2018
& Activation Failures
& \cmark
& \cmark
& $100 ns < x < 960 ns$
& $4.4 nJ/bit$
& $717.4 Mb/s$ \\
\hline
\end{tabular}
\caption{Comparison to previous DRAM-based TRNG proposals. \vspace{-10pt}}
\label{tab:prior_dram_works}
\end{center}
\end{table*}
\subsection{DRAM Command Scheduling}
Prior work~\cite{pyo2009dram} proposes using non-determinism in DRAM command
scheduling for true random number generation. In particular, since pending
access commands contend with regular refresh operations, the latency of a DRAM
access is hard to predict and is useful for random number generation.
Unfortunately, this method fails to satisfy two important properties of an
ideal TRNG. First, it harvests random numbers from the instruction and DRAM
command scheduling decisions made by the processor and memory controller, which
does \emph{not} constitute a fully non-deterministic entropy source. Since the
quality of the harvested random numbers depends directly on the quality of the
processor and memory controller implementations, the entropy source is visible
to and potentially modifiable by an adversary \hh{(e.g., by simultaneously
running a memory-intensive workload on another processor
core~\cite{moscibroda2007memory})}. Therefore, this method does not meet our
design goals as it does not securely generate random numbers.
Second, although this technique has a higher throughput than those based on
DRAM data retention (Table~\ref{tab:prior_dram_works}), \mechanism\ still
outperforms this method in terms of throughput by 211x (maximum) and 128x
(average) because a single byte of random data requires a \emph{significant}
amount of time to generate. Even if we scale the throughput results provided
by~\cite{pyo2009dram} to a modern day system (e.g., $5 GHz$ processor, 4 DRAM
channels\footnote{The authors do not provide their DRAM configuration, so we
optimistically assume that they evaluate their proposal using one DRAM channel.
We also assume that by utilizing 4 DRAM channels, the authors can harvest four
times the entropy, which gives the benefit of the doubt
to~\cite{pyo2009dram}.}), the theoretical maximum throughput of \jkthree{Pyo et
al.'s} approach\footnote{We base our estimations on \cite{pyo2009dram}'s claim
that they can harvest one byte of random data every 45000 cycles. However,
using these numbers along with the authors' stated processor configuration
(i.e., $2.8 GHz$) leads to a discrepancy between our calculated maximum
throughput ($\approx 0.5 Mb/s$) and that reported in~\cite{pyo2009dram}
($\approx 5 Mb/s$). We believe our estimation methodology and calculations are
sound. In our work, we compare \mechanism's peak throughput against that of
\cite{pyo2009dram} using a more modern system configuration (i.e., $5 GHz$
processor, 4 DRAM channels) than used in the original work, which gives the
benefit of the doubt to~\cite{pyo2009dram}.} is \emph{only} $3.40 Mb/s$ as
compared with the maximum (average) throughput of $717.4 Mb/s$ ($435.7 Mb/s$)
for \mechanism. To calculate the latency of generating random values, we assume
the same system configuration with \hh{\cite{pyo2009dram}'s} claimed number of
cycles \hh{45000} to generate random bits. To provide 64 bits of random data,
\cite{pyo2009dram} takes 18$\mu$s, which is significantly higher than D-RaNGe's
\jkthree{minimum/maximum latency of $100ns/960ns$}. Energy
consumption for ~\cite{pyo2009dram} depends heavily on the entire system that
it is running on, so we do not compare against this \hh{metric}.
\subsection{DRAM Data Retention}
Prior works~\cite{keller2014dynamic, sutar2018d} propose using DRAM data
retention failures to generate random numbers. Unfortunately, this approach is
\emph{inherently too slow} for high-throughput operation due to the long wait
times required to induce \jkthree{data retention failures in DRAM}. While the
failure rate can be increased by increasing the operating temperature, a wait
time on the order of seconds is required to induce \jkthree{enough}
failures~\cite{liu2013experimental, khan2014efficacy, qureshi2015avatar,
patel2017reaper, kim2018dram} to achieve \jkfour{high-throughput} random number
generation, which is orders of \jkfour{magnitude} slower than D-RaNGe.
Sutar et al.~\cite{sutar2018d} report that they are able to generate {256-bit}
random numbers using a hashing algorithm (e.g., {SHA-256}) on a $4~MiB$ DRAM
block that contains \mpo{data retention errors resulting from having disabled
DRAM refresh} for 40~seconds. \mpf{Optimistically assuming a large DRAM
capacity of $32~GiB$ and ignoring the time required to read out and hash the
erroneous data, a waiting time of 40~seconds to induce data retention errors
allows for an estimated maximum random number throughput of $0.05~Mb/s$. This
throughput is already far \hh{smaller} than \mechanism's measured maximum
throughput of \changes{$717.4 Mb/s$}, and it would decrease linearly with
DRAM capacity. Even if we were able to induce a large number of
data retention errors by waiting only 1 second, the \hh{maximum} random number
generation throughput would be $2~Mb/s$, i.e., orders of \jkfour{magnitude}
\hh{smaller} than \jkthree{that of} \mechanism.}
Because \cite{sutar2018d} requires a wait time of 40~seconds before
producing any random values, its latency for random number generation is
extremely high \jkfour{(40s)}. \hh{In contrast,} D-RaNGe can produce random values
very quickly since it generates random values \jkthree{potentially with} each DRAM access
(10s of nanoseconds). D-RaNGe therefore has a latency many orders of \jkfour{magnitude}
lower than \hh{Sutar et al.'s mechanism~\cite{sutar2018d}.}
We estimate the energy consumption of retention-time \jkthree{based TRNG} mechanisms
with Ramulator\hh{~\cite{kim2016ramulator, ramulatorgithub}} and
DRAMPower\hh{~\cite{drampowergithub, chandrasekar2011improved}}. We model first
\mpy{writing data to} a 4MiB DRAM \jkfour{region} (to constrain the energy consumption
estimate to the region of interest), waiting for 40 seconds, and then reading
from that region. We then divide the energy consumption of these operations by
the number of bits found (256 bits). \hh{We find that the energy consumption
is} around $6.8 mJ$ per bit\hh{, which} is orders of \jkfour{magnitude} more costly
than \jkthree{that of} \mechanism, which provides random \jkthree{numbers} at
$4.4 nJ$ per bit.
\subsection{DRAM Startup Values}
Prior work~\cite{tehranipoor2016robust, eckert2017drng} proposes using DRAM
startup values as random numbers. Unfortunately, this method is unsuitable for
continuous high-throughput operation since it requires a DRAM power cycle in
order to obtain random data. We are unable to accurately model the latency of
this mechanism since it relies on the startup time of DRAM (i.e., bus frequency
\hh{calibration}, temperature \hh{calibration}, timing register
initialization~\cite{ddr4operationhynix}). This \jkthree{heavily depends} on
the implementation of the system and \jkfour{the} DRAM device in use. Ignoring
these components, \hh{we estimate the throughput of generating random numbers
using startup values by taking into account only the latency of a single} DRAM
read (\emph{after} all initialization is complete), which \jkthree{is $60ns$}.
We model energy consumption ignoring the initialization phase as well, by
modeling the energy to read a MiB of DRAM and \hh{dividing} that quantity by
\cite{tehranipoor2016robust}'s claimed number of random bits found in that
region (420Kbit). \hh{Based on this calculation, we estimate energy consumption
as} $245.9pJ$ per bit. While \hh{the energy consumption of
\cite{tehranipoor2016robust}} is smaller than the energy cost of \mechanism, we
note that \hh{our energy estimation for \cite{tehranipoor2016robust}} does
\emph{not} account for the energy consumption required for initializing DRAM to
be able to read out the random values. Additionally,
~\cite{tehranipoor2016robust} requires a full system reboot which is often
impractical for applications and for effectively providing a \emph{steady
stream of random values}. \cite{eckert2017drng} suffers from the same issues
since it uses the same mechanism as \cite{tehranipoor2016robust} to
generate random numbers and is strictly worse since \cite{eckert2017drng}
results in $31.8x$ less entropy.
\subsection{Combining DRAM-based TRNGs}
We note that \mechanism's method for sampling random values from DRAM is
entirely distinct from prior DRAM-based TRNGs that we have discussed in this
section. This makes it possible to combine \mechanism~with prior work to
produce random values at an even higher throughput.
\section{Other Related Works}
\label{related}
In this work, we focus on the design of a DRAM-based hardware mechanism to
implement a TRNG, which makes the focus of our work orthogonal to those that
design PRNGs. In contrast to prior DRAM-based TRNGs discussed in
Section~\ref{comparison}, we propose using \emph{activation failures} as an
entropy source. Prior works characterize activation failures in order to
exploit the resulting error patterns for overall DRAM latency
reduction~\hh{\cite{chang2016understanding, kim2018solar, lee-sigmetrics2017,
lee2015adaptive}} and to implement physical unclonable functions
(PUFs)~\cite{kim2018dram}. However, none of these works measure the randomness
inherent in activation failures or propose using them to generate random
numbers.
\hh{Many} TRNG designs have been proposed that exploit sources of entropy
\mpf{that are \emph{not}} based on DRAM. \mpf{Unfortunately, these proposals
either 1) require custom hardware modifications that preclude their application
to commodity devices, or 2) do not sustain continuous (i.e., constant-rate)
high-throughput operation}. We briefly discuss different entropy sources with
examples.
\textbf{Flash Memory Read Noise.} Prior proposals use random
telegraph noise in flash memory devices as an entropy source (up to 1
Mbit/s)~\cite{wang2012flash, ray2018true}. Unfortunately, flash memory is
orders of magnitude slower than DRAM, making flash unsuitable for
high-throughput \jkthree{and low-latency} operation.
\textbf{SRAM-based Designs.} SRAM-based TRNG designs exploit
randomness in startup values \cite{
holcomb2007initial, holcomb2009power,
van2012efficient}.
Unfortunately, these proposals are unsuitable for continuous, high-throughput
operation since they require a power cycle.
\textbf{GPU- and FPGA-Based Designs.} Several works harvest random numbers
from GPU-based (up to 447.83 Mbit/s)~\cite{chan2011true, tzeng2008parallel,
teh2015gpus} and FPGA-based (up to 12.5 Mbit/s)~\cite{majzoobi2011fpga,
wieczorek2014fpga, chu1999design, hata2012fpga} entropy sources. \hh{These}
\mpf{proposals do not require modifications to commodity GPUs or FPGAs\hh{.
Yet,} GPUs and FPGAs are not as prevalent as DRAM in commodity devices
today.}
\textbf{Custom Hardware.} Various works propose TRNGs based in part or fully on
non-determinism \jkthree{provided by custom hardware designs} (\jkfour{with
TRNG throughput} up to 2.4 Gbit/s)~\cite{
amaki2015oscillator,
yang2016all,
bucci2003high,
bhargava2015robust,
petrie2000noise,
mathew20122, brederlow2006low, tokunaga2008true,
kinniment2002design, holleman20083, holcomb2009power,
pareschi2006fast,
stefanov2000optical}.
Unfortunately, \jkthree{the need for custom hardware limits the widespread use of such
proposals in} commodity hardware devices (today).
\delete{\textbf{System-Level RNGs.} Operating systems often provide interfaces
for harvesting entropy from devices running on a system. Unfortunately, the
entropy source of a system-level RNG depends on 1) the particular devices
attached to the system (e.g., I/O peripherals, disk drives) and 2) the quality
of the drivers that directly harvest entropy from these devices. This means
that a system-level RNG is not guaranteed to be harvesting random numbers from
a fully non-deterministic entropy source; therefore, system-level RNGs do not
meet our design goals.}
\section*{Acknowledgments} \jksix{We thank Ivan Puddu for useful comments and
\section*{Acknowledgments} \hh{We} thank the anonymous reviewers \jk{of
HPCA 2019 and MICRO 2018} for feedback and the SAFARI group members for
feedback and the stimulating intellectual environment they provide.
\SetTracking
[ no ligatures = {f},
outer kerning = {*,*} ]
{ encoding = * }
{ -40 }
{
\let\OLDthebibliography\thebibliography
\renewcommand\thebibliography[1]{
\OLDthebibliography{#1}
\setlength{\parskip}{0pt}
\setlength{\itemsep}{0pt}
}
\bibliographystyle{IEEEtranS}
|
\section{Introduction}
The relativistic heavy-ion collider (RHIC) at the Brookhaven National
Laboratory, USA, with the center-of-mass energy, $\sqrt{s}$= 200 GeV per
nucleon in Au - Au collisions and the large hadron
collider (LHC) at the European Organization for Nuclear Research,
Geneva, with $\sqrt{s}$ = 2.76 TeV per nucleon in Pb-Pb collisions may have
produced intensely strong magnetic field
at very early stages of collisions, when the event is off-central
\cite{Shovkovy:LNP871_2013,Elia:LNP871_2013,
Fukushima:LNP871_2013,Mulller:PRD89_2014,
Miransky:PR576_2015}.
Depending on the centralities, the strength of the magnetic field may
reach between $m_{\pi}^2$ ($\simeq 10^{18}$ Gauss) at
RHIC \cite{Kharzeev:NPA803_2008} to 15 $m_{\pi}^2$ at
LHC \cite{Skokov:IJMPA24_2009}.
At extreme cases it may reach values of 50 $m_{\pi}^2$ at LHC.
A very strong magnetic field ($\sim 10^{23}$~Gauss) may have existed
in the early universe during the electroweak
phase transition due to the gradients in Higgs field
\cite{vachaspati:PLB265'1991} or at the core of magnetars~\cite{magnetars}.
Thus we are motivated in this work to study the different processes
in electron-muon and electron-electron scattering in the lowest order in the
presence of strong magnetic
field. However, the above processes in the lowest as well as higher
order are well studied theoretically in
vacuum~\cite{bonciani_ferroglia,ilyichev_zykunov,halzen_martin} with
the solutions of Dirac equation in vacuum, {\em i.e.} with the free Dirac
spinors for positive and negative energy and their
corresponding completeness relation.
However, the scenario in the
presence of strong magnetic field is different because the form of
the Dirac
spinors are going to change, where, apart from the momentum dependence,
the spinors also depend on the spatial coordinates due to the
gauge used to solve the Dirac equation in magnetic field. As a result,
the completeness relations are also going to change.
For the sake of simplicity we assume the magnetic field to be uniform
and stationary. Moreover the strength of the magnetic field is
strong enough so that
only the lowest Landau level is sufficient for the calculation
of the matrix element and the crosssection.
This paper is divided into following sections. We first
nomenclature the
four-momentum, Mandelstam variables, magnetic Mandelstam variables,
suitable for the description in a strong magnetic field and then
revisit the Dirac equation in a strong and homogeneous magnetic
field in
section 2.1 and 2.2, respectively. Using those notations,
we calculate the spin-summed matrix element squared for the
electron-muon scattering $s$-channel (annihilation process) and $t$-
channel, Bhabha scattering and M\o ller scattering in sections
2.3-2.5,
respectively. In section 3, we first revisit for the formula
for calculating the crosssection by constructing the Lorentz
invariant phase space, flux factor, energy-momentum conserving
Dirac-Delta
function etc. in the presence of strong magnetic field and then
using the
the matrix elements for the above processes from the above sections 2.3-2.5,
the corresponding crosssections have been evaluated in sections 3.1-3.4,
respectively. Finally we conclude our results and discussion
in section 4.
\section{Electron-Electron and Electron-Muon Scattering in Strong Magnetic
field}
Our aim in this section is to calculate the square of matrix
element for the electron-muon scattering in both $s$- and
$t$-channel,
Bhabha scattering, and M\o ller scattering
in a strong homogeneous magnetic field. For the sake of simplicity we
work in the extreme relativistic limit, where we neglect the masses
of electrons as well as muons. In the presence of magnetic field
the form of spinor and hence the form of electron propagator is
changed
but the form of photon propagator remains the same. So first we are
going to revisit the Dirac equation in the strong magnetic field to
obtain
the form of positive energy and negative energy spinors and their
completeness relations.
\subsection{Notations}
\begin{figure}[h]
\begin{center}
\includegraphics[height=12cm,width=8cm,angle=-90]{notation.ps}
\vspace{-2cm}
\caption{}\label{notation}
\end{center}
\end{figure}
The dynamics of an electron in a magnetic field is factorized into
transverse and longitudinal plane with respect to the direction of
magnetic
field, as a result its momentum ($\mathbf{p}$) is separated into
perpendicular and longitudinal components with respect to the direction
of magnetic field
(say, $z$-direction). Hence the dispersion relation is modified
quantum mechanically into
\begin{eqnarray}
E_n(p_z)=\sqrt{p_z^2+m_f^2+2n\left|eB\right|}\quad,\label{landau_quantization}
\end{eqnarray}
In a strong magnetic field ($|eB|>>m^2$), electrons
prefer to lie in the lowest Landau level, hence the electron
momentum becomes
purely longitudinal~\cite{Gusynin:1998nh}, i.e $\vec{p}_\perp \approx 0$.
The aforesaid observation motivates
to construct the following kinematic variables, {\em viz.}
momentum, Mandelstam variables etc. which will be advantageous
to express matrix element squared, crosssection etc. in a strong magnetic
field. Thus using the following convention of the metric tensor
\begin{eqnarray}
g^{\mu\nu}=(1,-1,-1,-1), g^{\mu_{\perp}\nu_{\perp}}=(0,-1,-1,0)~ {\rm and} ~
g^{\mu_{\parallel}\nu_{\parallel}}=(1,0,0,-1)~,
\end{eqnarray}
we will first denote the four-momentum for a generic Feynman diagram in
Figure \ref{notation},
\begin{eqnarray}
p_\perp^\mu&=&(0,p^1,p^2,0)=(0,p_x,p_y,0),\\
p_\parallel^\mu&=&(p^0,0,0,p^3)=(E_0,0,0,p_z), \\
\widetilde{p_\parallel}^\mu&=&(\widetilde{p}^0,0,0,\widetilde{p}^3)
=(p^3,0,0,p^0),
\end{eqnarray}
therefore the usual Mandelstam variables take the following form:
\begin{eqnarray}
s&=&(p_\parallel+k_\parallel)^2=(P_\parallel+K_\parallel)^2,\label{mandelstam-s}\\
t&=&(p_\parallel-P_\parallel)^2=(K_\parallel-k_\parallel)^2,\label{mandelstam-t}\\
u&=&(p_\parallel-K_\parallel)^2=(P_\parallel-k_\parallel)^2.\label{mandelstam-u}
\end{eqnarray}
We define some new variables
$s_p,s_k,s_P,s_K,t_p,t_P,t_K,t_k,u_p,u_k,u_P,u_K$, dubbed as the magnetic
Mandelstam variables, which are defined as
\begin{eqnarray}
s_p&=&(\widetilde p_\parallel+k_\parallel)^2,~s_k=(\widetilde k_\parallel+p_\parallel)^2,\label{magnetic-mandelstam-s}\\
s_P&=&(\widetilde P_\parallel+K_\parallel)^2,~s_K=(\widetilde K_\parallel+P_\parallel)^2,\label{magnetic-mandelstam-ss}\\
t_p&=&(\widetilde p_\parallel-P_\parallel)^2,~t_P=(\widetilde P_\parallel-p_\parallel)^2,\label{magnetic-mandelstam-t}\\
t_K&=&(\widetilde K_\parallel-k_\parallel)^2,~t_k=(\widetilde k_\parallel-K_\parallel)^2,\label{magnetic-mandelstam-tt}\\
u_p&=&(\widetilde p_\parallel-K_\parallel)^2,~u_K=(\widetilde K_\parallel-p_\parallel)^2,\label{magnetic-mandelstam-u}\\
u_P&=&(\widetilde P_\parallel-k_\parallel)^2,~u_k=(\widetilde k_\parallel-P_\parallel)^2.\label{magnetic-mandelstam-uu}
\end{eqnarray}
Although the form of the magnetic Mandelstam variables (or the
half-tilde
Mandelstam variables, where one momentum is tilde) are same as the
Mandelstam
variables but they are completely different from them. In the
extreme
relativistic limit, they satisfy the following relations among
themselves
\begin{eqnarray}
s_p=-s_k,&&s_P=-s_K,\\
t_p=-t_P,&&t_K=-t_k,\\
u_p=-u_K,&&u_P=-u_k.
\end{eqnarray}
Furthermore we use other notations for the full-tilde Mandelstam
variables,
which are defined as
\begin{eqnarray}
\widetilde s&=&(\widetilde p_\parallel+\widetilde k_\parallel)^2=(\widetilde P_\parallel+\widetilde K_\parallel)^2,\\
\widetilde t&=&(\widetilde p_\parallel-\widetilde P_\parallel)^2=(\widetilde K_\parallel-\widetilde k_\parallel)^2,\\
\widetilde u&=&(\widetilde p_\parallel-\widetilde K_\parallel)^2=(\widetilde P_\parallel-\widetilde k_\parallel)^2.
\end{eqnarray}
We can directly relate the full-tilde Mandelstam variables to the
Mandelstam variables as
\begin{eqnarray}
s=-\widetilde s,~t=-\widetilde t,~u=-\widetilde u~.
\end{eqnarray}
\subsection{Dirac Spinors in a strong magnetic field}
The methods of Ritus eigenfunction~\cite{ritus} along with
the Schwinger Proper-time formalism~\cite{schwinger} are commonly
used to
solved the Dirac equation of charged fermions in the presence of a
constant magnetic field. There are different ways which have been
adopted in the literature~\cite{p_meszaros,herold_ruder_wunner,
sokolov_ternov} to obtain the spinor in a magnetic field, however,
we have mainly adopted to solve the
Dirac equation in a constant external field
from Ref.~\cite{Bhattacharya:2007vz}.
For the sake of simplicity, we assume a static and
homogeneous magnetic field, which is along the $z$-direction,
$\vec{B} =B \hat{z}$. Such a magnetic field can be
obtained from a vector potential $A^\mu = (0,0,Bx,0)$.
The choice of vector potential is not unique as the same magnetic
field can also be obtained from a symmetric potential given by
$A^\mu = (0,\frac{-By}{2},\frac{Bx}{2},0)$. Thus the positive energy
Dirac spinors with the gauge $A^\mu=(0,-By,0,0)$ are given by
the shifted coordinate, $\xi$ (=$\sqrt{eB}\left(y-\frac{p_x}{eB}\right)$)
~\cite{Furry,Bhattacharya:2007vz}
\begin{eqnarray}
U_+ (y,n,\vec p_{\omit{y}})=N\left( \begin{array}{c} I_{n-1}(\xi) \\[2ex] 0 \\[2ex]
{\strut\textstyle p_z \over \strut\textstyle E_n+m} I_{n-1}(\xi) \\[2ex]
-\, {\strut\textstyle \sqrt{2neB} \over \strut\textstyle
E_n+m} I_n (\xi)
\end{array} \right);\hspace{2mm}
U_- (y,n,\vec p_{\omit{y}})=N\left( \begin{array}{c} 0 \\[2ex] I_n (\xi) \\[2ex]
-\, {\strut\textstyle \sqrt{2neB} \over \strut\textstyle E_n+m}
I_{n-1}(\xi) \\[2ex]
-\,{\strut\textstyle p_z \over \strut\textstyle E_n+m} I_n(\xi)
\end{array} \right)
\end{eqnarray}
Similarly the negative energy Dirac spinors with $\widetilde \xi$
(=$\sqrt{eB}\left(y+\frac{p_x}{eB}\right)$) are given
by~\cite{Furry,Bhattacharya:2007vz}
\begin{eqnarray}
V_- (y,n,\vec p_{\omit{y}}) = N\left( \begin{array}{c}
{\strut\textstyle p_z \over \strut\textstyle E_n+m}
I_{n-1}(\widetilde\xi) \\[2ex]
{\strut\textstyle \sqrt{2neB} \over \strut\textstyle E_n+m}
I_n (\widetilde\xi) \\[2ex]
I_{n-1}(\widetilde\xi) \\[2ex] 0
\end{array} \right);\hspace{2mm}
V_+ (y,n,\vec p_{\omit{y}}) =N \left( \begin{array}{c}
{\strut\textstyle \sqrt{2neB} \over \strut\textstyle E_n+m}
I_{n-1}(\widetilde\xi) \\[2ex]
-\,{\strut\textstyle p_z \over \strut\textstyle E_n+m}
I_n(\widetilde\xi) \\[2ex]
0 \\[2ex] I_n (\widetilde\xi)
\end{array} \right) \,
\end{eqnarray}
where the normalization constant ($N$) is $N=\sqrt{E_n+m}$ and the
symbol,
$p_{\omit y}$ denotes the absence of the $y$-component of momentum
in the
spinors. The energy eigenvalues are given by the above Landau
quantization
\eqref{landau_quantization},
where $n$ denotes the Landau levels and the energy eigenfunctions,
$I_n (\xi)$
are expressed in terms of Hermite polynomials, $H_n (\xi)$
\begin{eqnarray}
I_n(\xi)&=&\frac{\sqrt{eB}}{n!2^n\sqrt{\pi}}e^{\frac{-\xi^2}{2}}H_n(\xi),
\end{eqnarray}
with the properties: $I_{-1}(\xi)=0$ and $I_0^2(\xi)=1$.
As mentioned earlier, in a strong magnetic field, only the lowest
Landau level (n=0) is populated.
We can now calculate the spin sums for the particles ($P_U$) and
anti-particles ($P_V$) in the presence of external magnetic field
as~\cite{Furry,Bhattacharya:2007vz}
\begin{eqnarray}
P_U (y,y' ,n,\vec p\omit y) &=&\sum_{\rm s}U_s (y,n,\vec p_{\omit{y}})
\overline{U}_s (y',n,\vec p_{\omit{y}})\nonumber\\
&=& {1\over 2} \bigg[\left\{ m(1+\Sigma_z) +
\rlap/p_\parallel -
\widetilde{\rlap/p}_\parallel \gamma_5 \right\} I_{n-1}(\xi)
I_{n-1}(\xi')\nonumber\\
&+& \left\{ m(1-\Sigma_z) + \rlap/p_\parallel -
\gamma_5 \widetilde{\rlap/p}_\parallel \right\} I_n(\xi)
I_n (\xi')\nonumber\\
&-& \sqrt{2neB} (\gamma_1 - i\gamma_2) I_n(\xi) I_{n-1}(\xi') \nonumber\\
&-& \sqrt{2neB} (\gamma_1 + i\gamma_2) I_{n-1}(\xi) I_n(\xi')
\bigg],\label{PUs}
\end{eqnarray}
\begin{eqnarray}
P_V(\widetilde y,\widetilde{y'} ,n,\vec p\omit y) &=&\sum_{\rm s}V_s (\widetilde{y},n,
\vec p_{\omit{y}})\overline{V}_s (\widetilde{y'},n,\vec p_{\omit{y}})
\nonumber\\
&=& {1\over 2} \Bigg[ \left\{ -m(1+\Sigma_z) +
\rlap/p_\parallel -
\widetilde{\rlap/p}_\parallel \gamma_5 \right\} I_{n-1}(\widetilde\xi)
I_{n-1} (\widetilde\xi')\nonumber\\
&+& \left\{ -m(1-\Sigma_z) + \rlap/p_\parallel -
\gamma_5 \widetilde{\rlap/p}_\parallel \right\} I_n(\widetilde\xi)
I_n(\widetilde\xi')\nonumber\\
&+&\sqrt{2neB} (\gamma_1 - i\gamma_2) I_n(\widetilde\xi)
I_{n-1}(\widetilde\xi')\nonumber\\
&+& \sqrt{2neB} (\gamma_1 + i\gamma_2) I_{n-1}(\widetilde\xi)
I_n(\widetilde\xi') \Bigg], \,
\label{PVs}
\end{eqnarray}
where $\Sigma_z=i\gamma^1\gamma^2$,
$\slashed{p}_\parallel=p^0\gamma^0-p^3\gamma^3$,
$\widetilde{\slashed{p}}_\parallel=p^3\gamma^0-p^0\gamma^3$ and $m$
is the mass of electron.
\subsection{Matrix element for electron-muon scattering at the lowest-order}
There are two different processes for electron-muon
scattering, {\em viz.} $\ems$ and $\emt$ represented by $s$ and $t$-channel
diagrams, respectively.
We will first evaluate the matrix element ($\mathfrak {M}$) for the
$s$-channel diagram and calculate the $\overline{\left|\mathfrak{M}\right|^2}$
by summing over the spin states. Finally we calculate the
same for the $t$-channel process.\\
\noindent {\large \bf $s$-Channel Process:}\\
The electron-muon scattering in $s$-channel represents the $e^- e^+
\rightarrow \mu^- \mu^+$ process by the following Feynman diagram
at the lowest-order in Figure \ref{s_e_mu}.
\begin{figure}[h]
\begin{center}
\includegraphics[height=12cm,width=8cm,angle=-90]{s_e_mu.ps}
\vspace{-2cm}
\caption{$s$-channel}\label{s_e_mu}
\end{center}
\end{figure}
where $U$ and $\overline U$ denote the Dirac spinors for the
incoming and the outgoing fermions,
respectively whereas $V$ and $\overline V$ in Figure \ref{s_e_mu}
represent the spinors for the outgoing and the incoming
anti-fermions,
respectively. Therefore the invariant amplitude for the $e^- e^+
\rightarrow \mu^- \mu^+$ process at the lowest-order
is given by the matrix element,
\begin{eqnarray}
-i\mathfrak{M}_{\rm s} \ems &=&\big[\overline{V}(y_B,k_{\omit{y}})ie\gamma^\mu U(y_A,p_{\omit{y}})\big]
\left(\frac{-ig^{\mu\nu}}{q^2}\right)\big[\overline{U}(Y_C,P_{\omit{y}})ie\gamma^\nu
V(Y_D,K_{\omit{y}})\big].\label{s-matrix}
\end{eqnarray}
We will now calculate the square of the matrix element,
${\left|\mathfrak{M}_s\right|}^2$
and then sum over the spin-states. It is convenient and easy to
solve if we
separately write the spin sums for electron and muon
\begin{equation}
\overline{\left|\mathfrak{M}_{\rm s}\right|^2}{\ems}=\frac{e^4}{q^4}L_{e}^{\mu\nu}
L^{muon}_{\mu\nu},\label{matrix-element-square}
\end{equation}
where the $L_e^{\mu \nu}$ and $L^{\rm{muon}}_{\mu \nu}$ are given by
\begin{eqnarray}
L_{e}^{\mu\nu}&=&\sum_{e~spins}\big[\overline{V}(y_B,k_{\omit{y}})
\gamma^\mu U(y_A,p_{\omit{y}})\big]\big[\overline{V}(y_B,k_{\omit{y}})\gamma^\nu U(y_A,p_{\omit{y}})
\big]^*~,\\
L^{muon}_{\mu\nu}&=&\sum_{\mu~ spins}\big[\overline{U}(Y_C,P_{\omit{y}})\gamma_\mu
V(Y_D,K_{\omit{y}})\big]\big[\overline{U}(Y_C,P_{\omit{y}})\gamma_\nu V(Y_D,K_{\omit{y}})\big]^*~,
\end{eqnarray}
respectively.
Using the properties of gamma matrices, {\em like}
\begin{equation*}
{\gamma^\mu}^\dag=\gamma^0\gamma^\mu\gamma^0 \hspace{3mm}\mbox{and}\hspace{3mm}\gamma^0\gamma^0=I,
\end{equation*}
and the cyclic property of trace, we further simplify $L_{e}^{\mu\nu}$ as
\begin{eqnarray}
L_{e}^{\mu\nu}=Tr\big[P_V(y_B,k_{\omit{y}})\gamma^\mu P_U(y_A,p_{\omit{y}})\gamma^\nu\big]
~.
\end{eqnarray}
The above spin-sums, $P_V$ and $P_U$ for the positron and the electron
~\cite{Furry,Bhattacharya:2007vz} in strong
magnetic field ({\em i.e.} for the lowest Landau levels, $n=0$) can
be calculated as
\begin{eqnarray}
P_V(y_B,n=0,k_\parallel)&=&\frac{1}{2}I_0^2(\xi_B)\big[-m(1-\Sigma_z)+
\slashed{k}_\parallel-\gamma_5\widetilde{\slashed{k}}_\parallel\big],\\
P_U(y_A,n=0,p_\parallel)&=&\frac{1}{2}I_0^2(\xi_A)\big[m(1-\Sigma_z)+
\slashed{p}_\parallel-\gamma_5\widetilde{\slashed{p}}_\parallel\big],\\
\end{eqnarray}
respectively.
Thus, after calculating the traces \footnote{which
is calculated in Appendix A}, the tensor at the electron vertex
is simplified into
\begin{equation}\label{L-electron}
L_{e}^{\mu\nu}=\Big[p^\mu_\parallel k^\nu_\parallel+p^\nu_\parallel k^\mu_\parallel-(p_\parallel\cdot k_\parallel) g^{\mu\nu}
+\widetilde p^\mu_\parallel \widetilde k^\nu_\parallel+\widetilde p^\nu_\parallel \widetilde k^\mu_\parallel-
(\widetilde p_\parallel\cdot \widetilde k_\parallel) g^{\mu\nu}\Big]-2m^2(g^{\mu\nu}-g^{\mu_{\perp}\nu_{\perp}}).
\end{equation}
In a similar way the tensor at the muon vertex is also calculated as
\begin{equation}\label{L-muon}
L^{muon}_{\mu\nu}=\Big[K_{\parallel \mu} P_{\parallel \nu}+K_{\parallel \nu} P_{\parallel \mu}-(K_\parallel\cdot P_\parallel) g_{\mu\nu}+
\widetilde K_{\parallel \mu} \widetilde P_{\parallel \nu}+\widetilde K_{\parallel \nu} \widetilde P_{\parallel \mu}-
(\widetilde K_\parallel\cdot \widetilde P_\parallel) g_{\mu\nu}\Big]-
2M^2(g_{\mu\nu}-g_{\mu_{\perp}\nu_{\perp}}).
\end{equation}
However in the extreme relativistic limit, where the mass terms
could be
neglected, the above squared matrix element
(\ref{matrix-element-square})
becomes simplified and is given by the short-hand notation
\begin{equation}\label{matrix-element-squre-2}
\overline{\left| \mathfrak{M}_{\rm s} \right|^2}{\ems}=\frac{e^4}{q^4}(T^{\mu\nu}R_{\mu\nu}+T^{\mu\nu}\widetilde{R}_{\mu\nu}+\widetilde{T}^{\mu\nu}R_{\mu\nu}
+\widetilde{T}^{\mu\nu}\widetilde{R}_{\mu\nu})~,
\end{equation}
where $T^{\mu\nu},R_{\mu\nu},\widetilde{T}^{\mu\nu},\widetilde{R}_{\mu\nu}$ are defined by
\begin{eqnarray}
T^{\mu\nu}&=&p^\mu_\parallel k^\nu_\parallel+p^\nu_\parallel k^\mu_\parallel-(p_\parallel\cdot k_\parallel) g^{\mu\nu},\\
\widetilde{T}^{\mu\nu}&=&\widetilde p^\mu_\parallel \widetilde k^\nu_\parallel+\widetilde p^\nu_\parallel
\widetilde k^\mu_\parallel-(\widetilde p_\parallel\cdot \widetilde k_\parallel) g^{\mu\nu},\\
R_{\mu\nu}&=&K_{\parallel \mu} P_{\parallel \nu}+K_{\parallel \nu} P_{\parallel \mu}-(K_\parallel\cdot P_\parallel) g_{\mu\nu},\\
\widetilde{R}_{\mu\nu}&=&\widetilde K_{\parallel \mu} \widetilde P_{\parallel \nu}+\widetilde K_{\parallel \nu} \widetilde P_{\parallel \mu}-
(\widetilde K_\parallel\cdot \widetilde P_\parallel) g_{\mu\nu}.
\end{eqnarray}
Thus the products are calculated as
\begin{eqnarray}
T^{\mu\nu}R_{\mu\nu}&=&2\Big[(p_\parallel\cdot K_\parallel)(k_\parallel\cdot P_\parallel)+(p_\parallel\cdot P_\parallel)(k_\parallel\cdot K_\parallel)\Big],\\
\widetilde{T}^{\mu\nu}R_{\mu\nu}&=&2\Big[(\widetilde p_\parallel\cdot K_\parallel)(\widetilde k_\parallel\cdot P_\parallel)+(\widetilde
p_\parallel\cdot P_\parallel)(\widetilde k_\parallel\cdot K_\parallel)\Big],\\
T^{\mu\nu}\widetilde{R}_{\mu\nu}&=&2\Big[(p_\parallel\cdot \widetilde K_\parallel)(k_\parallel\cdot \widetilde P_\parallel)+(p_\parallel\cdot
\widetilde P_\parallel)(k_\parallel\cdot \widetilde K_\parallel)\Big],\\
\widetilde{T}^{\mu\nu}\widetilde{R}_{\mu\nu}&=&2\Big[(\widetilde p_\parallel\cdot \widetilde K_\parallel)(\widetilde k_\parallel\cdot
\widetilde P_\parallel)+(\widetilde p_\parallel\cdot \widetilde P_\parallel)(\widetilde k_\parallel\cdot
\widetilde K_\parallel)\Big].
\end{eqnarray}
Thus after substituting the products, the matrix element squared
\eqref{matrix-element-squre-2}
for the electron-muon scattering for $s$-channel becomes
\begin{eqnarray}
\overline{\left|\mathfrak{M}_{\rm s}\right|^2}{\ems}&=&\frac{2e^4}{q^4}\Big[(p_\parallel\cdot K_\parallel)(k_\parallel\cdot P_\parallel)+(p_\parallel\cdot P_\parallel)(k_\parallel\cdot K_\parallel)+(p_\parallel\cdot \widetilde K_\parallel)(k_\parallel\cdot \widetilde P_\parallel)\nonumber\\
&&+(p_\parallel\cdot \widetilde P_\parallel)(k_\parallel\cdot \widetilde K_\parallel)+(\widetilde p_\parallel\cdot K_\parallel)(\widetilde k_\parallel\cdot P_\parallel)+(\widetilde p_\parallel\cdot P_\parallel)(\widetilde k_\parallel\cdot K_\parallel)\nonumber\\
&&+(\widetilde p_\parallel\cdot \widetilde K_\parallel)(\widetilde k_\parallel\cdot \widetilde P_\parallel)+(\widetilde p_\parallel\cdot \widetilde P_\parallel)(\widetilde k_\parallel\cdot \widetilde K_\parallel)\Big]. \label{matrix-element-square-3}
\end{eqnarray}
Using the notations for the Mandelstam and magnetic Mandelstam
variables
mentioned in equations \eqref{mandelstam-s}-\eqref{mandelstam-u}
and
\eqref{magnetic-mandelstam-s}-\eqref{magnetic-mandelstam-uu},
respectively,
the above matrix element squared becomes
\begin{eqnarray}
{\overline{\left|\mathfrak{M}_{\rm s}\right|^2}}_{\rm B \neq 0} (e^- e^+ \rightarrow \mu^-
\mu^+)&=&
\frac{e^4}{2q^4}\Big[u^2+t^2+u_Ku_P+t_Pt_K+u_pu_k+t_pt_k+\widetilde{u}^2+\widetilde{t}^2\Big]\nonumber\\
&=&\frac{e^4}{s^2}\Big[u^2+t^2+u_pu_k+t_pt_k\Big].\label{matrix-element-square-4}
\end{eqnarray}
For the sake of completeness, the $e$-$\mu$ scattering in
$s$-channel,
{\em i.e.} $\ems$ process in vacuum can be calculated
as~\cite{halzen_martin}
\begin{eqnarray}
{\overline{\left|\mathfrak{M}_{\rm s}\right|^2}}_{\rm B =0} \ems=
\frac{2e^4}{s^2}\Big[u^2+t^2\Big].\label{s-vacuum}
\end{eqnarray}
\noindent {\large \bf $t$-Channel Process:}\\
The electron-muon scattering in the $t$-channel, {\em i.e.} $e^- \mu^-
\rightarrow e^- \mu^-$ process in the lowest-order is represented
by the following Feynman diagram in Figure \ref{t_e_mu}.
We will now evaluate the matrix element for it as
\begin{figure}[h]
\begin{center}
\includegraphics[height=12cm,width=8cm,angle=-90]{t_e_mu.ps}
\vspace{-1cm}
\caption{$t$-channel}\label{t_e_mu}
\end{center}
\end{figure}
\begin{eqnarray}
\mathfrak{M}_{\rm t} \emt &=& \frac{-e^2}{q^2}\big[\overline{U}(Y_C,P_{\omit{y}})
\gamma^\mu U(y_A,p_{\omit{y}})\big]\big[\overline{V}(y_B,k_{\omit{y}})
\gamma_\mu V(Y_D,K_{\omit{y}})\big].
\end{eqnarray}
As we know that the $e$-$\mu$ scattering in $s$-channel, {\em i.e.}
$e^- e^+ \rightarrow \mu^- \mu^+$ process is related to
the $e$-$\mu$ scattering in $t$-channel, {\em i.e.}
$e^- \mu^- \rightarrow e^-\mu^- $ process by the
crossing symmetry. This facilitates us to obtain the spin-summed
squared matrix element for the $t$-channel directly from the
$e$-$\mu$ scattering in $s$-channel in following way.
As mentioned earlier, in the strong magnetic field (along the
$z$-direction), the spinors do not have the spatial $y$ dependence
and the perpendicular component of momentum ($p_\perp$) becomes
vanishingly small. Therefore the matrix element in $s$-channel \eqref{s-matrix}
can be rewritten as
\begin{eqnarray}
\mathfrak{M}_{\rm s} \ems &=&\frac{-e^2}{(p_\parallel+k_\parallel)^2}\big[\overline{V}(k_\parallel)\gamma^\mu U(p_\parallel)\big]
\big[\overline{U}(P_\parallel)\gamma_\mu
V(K_\parallel)\big].
\end{eqnarray}
Now if we interchange the momentum $k_\parallel$ with $-P_\parallel$ due to
the crossing symmetry, the above matrix element becomes
\begin{eqnarray}
\mathfrak{M}_{\rm s} \ems &=&\frac{-e^2}{(p_\parallel-P_\parallel)^2}\big[\overline{V}(-P_\parallel)\gamma^\mu U(p_\parallel)\big]
\big[\overline{U}(-k_\parallel)\gamma_\mu
V(K_\parallel)\big].
\end{eqnarray}
Using the Feynman-St\"uckelberg interpretation, where by reversing the
direction of momentum (say, p) in the spinors, $\overline{V}(-p)$ becomes
$\overline{U}(p)$ and $\overline{U}(-p)$ becomes $\overline{V}(p)$, the above $s$-channel matrix element gives the desired
$t$-channel matrix element
\begin{eqnarray}
\mathfrak{M}_{\rm t} \emt &=& \frac{-e^2}{(p_\parallel-P_\parallel)^2}\big[\overline{U}(P_\parallel)
\gamma^\mu U(p_\parallel)\big]\big[\overline{V}(k_\parallel)
\gamma_\mu V(K_\parallel)\big].
\end{eqnarray}
Therefore the spin-summed matrix element squared for the $t$-channel diagram
(Figure 3) is easily derived as
\begin{eqnarray}
\overline{\left|\mathfrak{M}_{\rm t}\right|^2}_{\rm B \neq 0} {\emt}&=&\frac{2e^4}{q^4}\Big[(p_\parallel\cdot k_\parallel)(P_\parallel\cdot K_\parallel)+(p_\parallel\cdot K_\parallel)(P_\parallel\cdot k_\parallel)+(p_\parallel\cdot \widetilde k_\parallel)(P_\parallel\cdot \widetilde K_\parallel)\nonumber\\
&&+(p_\parallel\cdot \widetilde K_\parallel)(P_\parallel\cdot \widetilde k_\parallel)+(\widetilde p_\parallel\cdot k_\parallel)(\widetilde P_\parallel\cdot K_\parallel)+(\widetilde p_\parallel\cdot K_\parallel)(\widetilde P_\parallel\cdot k_\parallel)\nonumber\\
&&+(\widetilde p_\parallel\cdot \widetilde k_\parallel)(\widetilde P_\parallel\cdot \widetilde K_\parallel)
+(\widetilde p_\parallel\cdot \widetilde K_\parallel)(\widetilde P_\parallel
\cdot \widetilde k_\parallel)\Big],\label{e-u-t-ch}
\end{eqnarray}
which, in turn, will be expressed in terms of the Mandelstam
variables from the
$s$-channel expression \eqref{matrix-element-square-4} by
interchanging
$s$ with $t$, $u_k$ with $u_P$, $s_k$ with $t_P$, and $s_P$ with $t_k$ and is
given by
\begin{eqnarray}
{\overline{\left|\mathfrak{M}_{\rm t}\right|^2}}_{\rm B \neq 0} {(e^- \mu^-
\rightarrow e^- \mu^-)}&=&
\frac{e^4}{4q^4}\Big[s^2+u^2+s_ks_K+u_Ku_k+s_ps_P+u_pu_P+\widetilde{s}^2+\widetilde{u}^2\Big]\nonumber\\
&=&\frac{e^4}{t^2}\Big[s^2+u^2+s_ps_P+u_pu_P\Big].\label{matrix-element-squared-t}
\end{eqnarray}
However, the above squared-matrix element in vacuum
is~\cite{halzen_martin}
\begin{eqnarray}
{\overline{\left|\mathfrak{M}_{\rm t}\right|^2}}_{\rm B=0} \emt=
\frac{2e^4}{t^2}\Big[s^2+u^2\Big].\label{t-vacuum}
\end{eqnarray}
\subsection{Matrix Element for Bhabha Scattering: $e^- e^+ \rightarrow
e^- e^+$}
There are possible $s$- and $t$-channel diagrams, which contribute
to the Bhabha scattering, in Figure(s) \ref{s-ch} and \ref{fd4},
respectively in the lowest order.
\begin{figure}[h]
\begin{minipage}{0.5\textwidth}
\includegraphics[height=10cm,width=8cm,angle=-90]{s_ep.ps}
\vspace{-2cm}
\caption{$s$-channel}\label{s-ch}
\end{minipage}
\begin{minipage}{0.5\textwidth}
\includegraphics[height=9cm,width=7cm,angle=-90]{t_ep.ps}
\vspace{-1cm}
\caption{$t$-channel}\label{fd4}
\end{minipage}
\end{figure}
Therefore the matrix element for the Bhabha scattering in the
lowest order is
\begin{eqnarray}
{\mathfrak{M}}\bhabha = {\mathfrak{M}}_{\rm{s}}\bhabha+
{\mathfrak{M}}_{\rm{t}}\bhabha ,
\end{eqnarray}
where the $s$- and $t$-channel
contributions to the matrix element in a strong magnetic field are
given by
\begin{eqnarray}
{\mathfrak{M}}_{\rm{s}}\bhabha
&=& \frac{-e^2}{q_1^2}\big[\overline{V}(y_B,k_{\omit{y}})
\gamma^\mu U(y_A,p_{\omit{y}})\big]\big[\overline{U}(Y_C,P_{\omit{y}})
\gamma_\mu V(Y_D,K_{\omit{y}})\big], \label{bhabha-m-s}\\
\mathfrak{M}_{\rm{t}} (e^- e^+ \rightarrow e^- e^+)
&=& \frac{-e^2}{q_2^2}\big[\overline{U}(Y_C,P_{\omit{y}})
\gamma^\mu U(y_A,p_{\omit{y}})\big]\big[\overline{V}(y_B,k_{\omit{y}})
\gamma_\mu V(Y_D,K_{\omit{y}})\big],\label{bhabha-m-t}
\end{eqnarray}
respectively. Hence the total matrix element squared becomes
\begin{equation}
|\mathfrak{M}|^2=|\mathfrak{M}_{\rm{s}}|^2+|\mathfrak{M}_{\rm{t}}|^2 +
\mathfrak{M}_{\rm{s}}\mathfrak{M}_{\rm{t}}^*+\mathfrak{M}_{\rm{s}}^*
\mathfrak{M}_{\rm{t}}.
\end{equation}
Similar to the electron-muon scattering in the $s$-channel~
\eqref{matrix-element-square-3}, we can now directly write the
spin-summed
squared matrix element of the Bhabha scattering for the $s$-channel
in extreme relativistic limit
\begin{eqnarray}
\overline{\left|\mathfrak{M}_{\rm{s}}\right|^2}{\bhabha}&=&\frac{2e^4}{q_1^4}\Big[(p_\parallel\cdot K_\parallel)(k_\parallel\cdot P_\parallel)+(p_\parallel\cdot P_\parallel)(k_\parallel\cdot K_\parallel)+(p_\parallel\cdot \widetilde K_\parallel)(k_\parallel\cdot \widetilde P_\parallel)\nonumber\\
&&+(p_\parallel\cdot \widetilde P_\parallel)(k_\parallel\cdot \widetilde K_\parallel)+(\widetilde p_\parallel\cdot K_\parallel)(\widetilde k_\parallel\cdot P_\parallel)+(\widetilde p_\parallel\cdot P_\parallel)(\widetilde k_\parallel\cdot K_\parallel)\nonumber\\
&&+(\widetilde p_\parallel\cdot \widetilde K_\parallel)(\widetilde k_\parallel\cdot \widetilde P_\parallel)+(\widetilde p_\parallel\cdot \widetilde P_\parallel)(\widetilde k_\parallel\cdot \widetilde K_\parallel)\Big],
\end{eqnarray}
which in terms of Mandelstam and Magnetic Mandelstam variables can
be written as
\begin{equation}
\overline{\left|\mathfrak{M}_{\rm{s}}\right|^2}{\bhabha}=\frac{e^4}{s^2}\Big[u^2+t^2+u_pu_k+t_pt_k\Big].\label{bhabha_s}
\end{equation}
The contribution for the Bhabha scattering in $t$-channel can be obtained in
extreme relativistic limit, in analogy with
$e$-$\mu$ scattering for $t$-channel in \eqref{e-u-t-ch},
\begin{eqnarray}
\overline{\left|\mathfrak{M}_{\rm{t}}\right|^2}{\bhabha}&=&\frac{2e^4}{q_2^4}\Big[(p_\parallel\cdot k_\parallel)(P_\parallel\cdot K_\parallel)+(p_\parallel\cdot K_\parallel)(P_\parallel\cdot k_\parallel)+(p_\parallel\cdot \widetilde k_\parallel)(P_\parallel\cdot \widetilde K_\parallel)\nonumber\\
&&+(p_\parallel\cdot \widetilde K_\parallel)(P_\parallel\cdot \widetilde k_\parallel)\nonumber+(\widetilde p_\parallel\cdot k_\parallel)(\widetilde P_\parallel\cdot K_\parallel)+(\widetilde p_\parallel\cdot K_\parallel)(\widetilde P_\parallel\cdot k_\parallel)\nonumber\\
&&+(\widetilde p_\parallel\cdot \widetilde k_\parallel)(\widetilde P_\parallel\cdot \widetilde K_\parallel)
+(\widetilde p_\parallel\cdot \widetilde K_\parallel)(\widetilde P_\parallel\cdot \widetilde k_\parallel)\Big],
\end{eqnarray}
which can be expressed in terms of Mandelstam variables as
\begin{equation}
\overline{\left|\mathfrak{M}_{\rm{t}}\right|^2}{\bhabha}=
\frac{e^4}{t^2}\Big[s^2+u^2+s_ps_P+u_pu_P\Big].\label{bhabha_t}
\end{equation}
The interference term is given by
\begin{eqnarray}
\mathfrak{M}_{\rm{s}}\mathfrak{M}_{\rm{t}}^*&=&\frac{e^4}{q_1^2q_2^2}\big[\overline{V}(y_B,k_{\omit{y}})\gamma^\mu U(y_A,p_{\omit{y}})\big]\big[\overline{U}(Y_C,P_{\omit{y}})\gamma_\mu V(Y_D,K_{\omit{y}})\big]\nonumber\\
&&\times \big[\overline{V}(Y_D,K_{\omit{y}})\gamma_\nu V(y_B,k_{\omit{y}})
\big] \big[\overline{U}(y_A,p_{\omit{y}})\gamma^\nu U(Y_C,P_{\omit{y}})
\big].
\end{eqnarray}
However, in the strong magnetic field limit, all the spatial ($y$)
dependence
of the Dirac spinors are gone and also the $p_\perp$
is zero, so the above interference term is rewritten as
\begin{eqnarray}
\mathfrak{M}_{\rm{s}}\mathfrak{M}_{\rm{t}}^*&=&\frac{e^4}{q_1^2q_2^2}\big[\overline{V}(k_\parallel)\gamma^\mu U(p_\parallel)\big]\big[\overline{U}(P_\parallel)\gamma_\mu V(K_\parallel)\big]\big[\overline{V}(K_\parallel)\gamma_\nu V(k_\parallel)\big] \big[\overline{U}(p_\parallel)\gamma^\nu U(P_\parallel)\big],
\end{eqnarray}
which becomes, after summing over the spin states
\begin{eqnarray}
\overline{\mathfrak{M}_s\mathfrak{M}_t^*}
=\frac{e^4}{q_1^2q_2^2}Tr\Big[P_V(k_\parallel)\gamma^\mu P_U
(p_\parallel)\gamma^\nu P_U(P_\parallel)\gamma_\mu P_V(K_\parallel)
\gamma_\nu\Big].
\end{eqnarray}
Using the property of $\gamma$-matrices~
$\gamma^\mu\slashed{a}\slashed{b}\slashed{c}\gamma_\mu=-2\slashed{c}
\slashed{b}\slashed{a}$, the above interference term can be further
simplified in terms of Mandelstam and the magnetic Mandelstam
variables\footnote{Calculated in Appendix B} and finally
we find that it vanishes,
\begin{eqnarray}
\overline{\mathfrak{M}_{\rm s}\mathfrak{M}_{\rm t}^*} &=&0\label{M1M2}.
\end{eqnarray}
Finally using the Mandelstam and magnetic Mandelstam variables, the matrix
element squared for the Bhabha scattering is obtained by the $s$- \eqref{bhabha_s}
and $t$-channel \eqref{bhabha_t} contributions only in a strong magnetic field
\begin{eqnarray}
{\overline{\left|\mathfrak{M}\right|^2}}_{\rm B \neq 0} \bhabha&=&
\overline{\left|\mathfrak{M}_{\rm s}\right|^2} +
\overline{\left|\mathfrak{M}_{\rm t}\right|^2} \nonumber\\
&=&\frac{e^4}{s^2}\Big[u^2+t^2+u_pu_k+t_pt_k\Big]+\frac{e^4}{t^2}\Big[s^2+u^2+s_ps_P+u_pu_P\Big].\label{Bhabha_M}
\end{eqnarray}
The above crucial observation in a strong magnetic field can be
understood
as follows: In vacuum, Bhabha scattering in $s$ channel gives the
forward peak
whereas in $t$-channel it gives the backward peak. In the presence of
strong magnetic field, the dynamics of the electron is restricted
to one
dimension so the interference of two peaks in vacuum may not be
feasible in the presence of strong magnetic field.
However, in vacuum, the interference term does not vanish, thus the above
matrix element squared for Bhabha scattering in the
absence of strong magnetic field is given by~\cite{halzen_martin}
\begin{eqnarray}
{\overline{\left|\mathfrak{M}\right|^2}}_{\rm B =0} \bhabha=
\frac{2e^4}{s^2}\Big[u^2+t^2\Big]+\frac{4e^4u^2}{ts}+
\frac{2e^4}{t^2}\Big[s^2+u^2\Big].\label{bhabha-vacuum-m}
\end{eqnarray}
\subsection{Matrix Element for M\o ller Scattering }
The Feynman diagrams for the M\o ller scattering ($e^-e^- \rightarrow
e^-e^-$) are shown below in Figure (s) \ref{moller_fd_1} and \ref{moller_fd_2},
\begin{figure}[h]
\begin{minipage}{0.5\textwidth}
\includegraphics[height=9cm,width=7cm,angle=-90]{t_ee.ps}
\vspace{-1cm}
\caption{$t$-channel}\label{moller_fd_1}
\end{minipage}
\begin{minipage}{0.5\textwidth}
\includegraphics[height=9cm,width=7cm,angle=-90]{u_ee.ps}
\vspace{-1cm}
\caption{$u$-channel}\label{moller_fd_2}
\end{minipage}
\end{figure}
which are related to the Bhabha scattering ($e^-e^+
\rightarrow e^-e^+$) by the crossing symmetry, {\em namely}
by simply crossing the incoming positron to outgoing positron
\begin{eqnarray}
&&e^-(p)+e^+(k) \rightarrow e^-(P)+e^+(K),\label{bhabha-eq}\\
&&e^-(p)+e^-(-K) \rightarrow e^-(P)+e^-(-k).\label{moller-eq}
\end{eqnarray}
As we know already, in strong magnetic field, the interchange of momenta due to the crossing
symmetry is effectively translated in terms of their longitudinal
component only. Thus the momentum exchange between $k_\parallel$ and $K_\parallel$ in
\eqref{bhabha-eq}-\eqref{moller-eq} helps to obtain the matrix element for the
M\o ller scattering in u and t-channel by identifying the same for the Bhabha scattering in s- and
t-channels \eqref{bhabha-m-s}-\eqref{bhabha-m-t}, respectively.
The above interchange of momenta can be equivalently expressed in terms of the Mandelstam
and magnetic Mandelstam variables as
\begin{eqnarray}
s (=(p_\parallel+k_\parallel)^2) &\longrightarrow& u (=(p_\parallel-
K_\parallel)^2)\label{moller-cross-1}\\
u (=(p_\parallel-K_\parallel)^2) &\longrightarrow& s(=(p_\parallel+
k_\parallel)^2)\\
t_k (=(\widetilde{k}_\parallel-K_\parallel)^2) &\longrightarrow& t_K (=
(k_\parallel-\widetilde{K}_\parallel)^2)\\
s_k(=(p_\parallel+\widetilde k_\parallel)^2) &\longrightarrow& u_K(=
(p_\parallel-\widetilde K_\parallel)^2).\label{moller-cross-2}
\end{eqnarray}
Therefore the above crossing symmetry \eqref{moller-cross-1}-\eqref{moller-cross-2} helps
us to obtain the squared
matrix element for the M\o ller scattering from the Bhabha
scattering~\eqref{Bhabha_M}, {\em namely}
\begin{eqnarray}
{\overline{\left|\mathfrak{M}\right|^2}}_{\rm B \neq 0} {\moller} &=&
{\overline{\left|\mathfrak{M}_{\rm u}\right|^2}}_{\rm B \neq 0} {\moller}+
{\overline{\left|\mathfrak{M}_{\rm t} \right|^2}}_{\rm B \neq 0}{\moller},
\end{eqnarray}
where the $u$ and $t$-channel matrix element squared are given by
\begin{eqnarray}
{\overline{\left|\mathfrak{M}_{\rm u}\right|^2}}_{\rm B \neq 0} \moller &=&
\frac{e^4}{u^2}\Big[s^2+t^2+s_ps_K+t_pt_K\Big],\label{moller-u} \\
{\overline{\left|\mathfrak{M}_{\rm t} \right|^2}}_{\rm B \neq 0} {\moller} &=&
\frac{e^4}{t^2}\Big[u^2+s^2+u_pu_P+s_ps_P\Big],\label{Moller_M}
\end{eqnarray}
respectively. However, the above matrix element in vacuum can also
be calculated as~\cite{halzen_martin}
\begin{eqnarray}
{\overline{\left|\mathfrak{M}\right|^2}}_{\rm B =0} \moller=
\frac{2e^4}{u^2}\Big[s^2+t^2\Big]+\frac{4e^4s^2}{tu}+
\frac{2e^4}{t^2}\Big[s^2+u^2\Big].\label{moler-vacuum-m}
\end{eqnarray}
\section{Crosssection}
Let us illustrate the usual procedure to compute the
crosssection from the transition amplitude for the
above mentioned processes in the presence of strong magnetic field.
For that we choose a generic $e^-\mu^-
\rightarrow e^-\mu^-$ process in Figure \ref{lifs}, for which we will calculate the
transition amplitude.
\begin{figure}[h]
\begin{center}
\includegraphics[height=12cm,width=8cm,angle=-90]{lifs.ps}
\vspace{-1cm}
\caption{Feynman diagram}\label{lifs}
\end{center}
\end{figure}
The solution of the Dirac equation for the $e^-$ in an external
magnetic field in the $z$-direction is given
by~\cite{Furry,Bhattacharya:2007vz}
\begin{equation*}
\Psi(X,p)=U(y,n,p_{\omit{y}}) e^{-ip_{\omit{y}}\cdot X},
\end{equation*}
where $U$ is the $e^-$ spinor in the presence of an external magnetic field, $n$
labels
the Landau levels, $X^\mu=(t,x,y,z)$ and $p_{\omit{y}}^\mu=(E,p_x,0,p_z)$,
where the $y$-component of momentum is missing. The transition
matrix
element for the above process is thus given by
\begin{eqnarray}
T_{fi}&=&\int (\overline{U}_D ie\gamma^\mu U_B) \left(\frac{-i}{q^2}\right)(\overline{U}_C ie\gamma_\mu U_A ) e^{i(p^D_{\omit{y}}+p^C_{\omit{y}}-p^A_{\omit{y}}-p^B_{\omit{y}})\cdot X}d^4X\\
&\equiv &\int- i \hspace{1mm}\mathfrak{M}\hspace{1mm}
e^{i(p^D_{\omit{y}}+p^C_{\omit{y}}-p^A_{\omit{y}}-p^B_{\omit{y}})\cdot X}
d^4X.
\end{eqnarray}
Therefore the matrix element is given by
\begin{equation}
-i \mathfrak{M}= (\overline{U}_D ie\gamma^\mu U_B) \left(\frac{-i}{q^2}\right)
(\overline{U}_C ie\gamma_\mu U_A ),
\end{equation}
where the main problem arises that $
\mathfrak{M}$ now becomes a function of $y$ in the presence of
an external magnetic field, hence we can not take it outside the
integration. We circumvent this issue by taking the strong magnetic
field limit, where all the $y$ dependence
in $\mathfrak{M}$ is gone. The transition matrix element,
in the strong magnetic field becomes
\begin{eqnarray}
T_{fi} =- i \hspace{1mm}\mathfrak{M} \hspace{1mm}(2\pi)^4
\delta^4(p^D_{\omit{y}}+p^C_{\omit{y}}-p^A_{\omit{y}}-p^B_{\omit{y}}),
\end{eqnarray}
and the transition rate per unit volume is given by
\begin{eqnarray}
W_{fi}&=&\frac{|T_{fi}|^2}{TV}\nonumber\\
&=&(2\pi)^4|\mathfrak{M}|^2 \delta^4(p^D_{\omit{y}}+p^C_{\omit{y}}-
p^A_{\omit{y}}-p^B_{\omit{y}}).
\end{eqnarray}
Therefore using the definition of crosssection
\begin{equation*}
\mbox{crosssection}=\frac{W_{fi}}{\mbox{(initial flux)}}\mbox{(number of
final states)},
\end{equation*}
the crossection is given by
\begin{eqnarray}
d\sigma=\frac{(2\pi)^4|\mathfrak{M}|^2 \delta^4(p^D_{\omit{y}}+p^C_{\omit{y}}-p^A_{\omit{y}}-p^B_{\omit{y}})}{F}d^4p^Cd^4p^D.
\end{eqnarray}
All the particles A,B,C and D are real particles so they
must satisfy the on-shell mass condition, which in the
strong magnetic field becomes $p_\parallel^2=m^2$
\cite{Gusynin:1998nh}
with $p_\parallel^\mu=(E,0,0,p_z)$. This gives
the crosssection
\begin{eqnarray}
d\sigma=\frac{(2\pi)^4|\mathfrak{M}|^2 \delta^4(p^D_{\omit{y}}+
p^C_{\omit{y}}-p^A_{\omit{y}}-p^B_{\omit{y}})}{F}
\frac{d^3\vec{p^C}}{(2\pi)^32E_C}\frac{d^3\vec{p^D}}{(2\pi)^32E_D}. \label{liff}
\end{eqnarray}
To calculate the unpolarized crosssection, we need to average over the
quantum states of incoming particles and sum over the final states,
therefore we replace the above matrix element squared
\begin{equation}
|\mathfrak{M}|^2 \rightarrow \frac{1}{(2s_A+1)(2s_B+1)}\sum_{all ~states}|\mathfrak{M}|^2,
\end{equation}
where $s_A$ and $s_B$ are the spin of incoming particles.
We have already summed over the particle states in the matrix
element squared,
denoted as $\overline {\left|\mathfrak{M}\right|^2}$, so we just
need to divide
$\overline {\left|\mathfrak{M}\right|^2}$ by the degeneracy factor, 4, to get the
desired crosssection.
Now we have all the ingredients to compute the crosssection for the
aforesaid
processes with the corresponding matrix element squared.
\subsection{Electron-Muon scattering in $s$ channel:
Annihilation Process ($e^-e^+\rightarrow \mu^-\mu^+$)}
With the help of~\eqref{liff} the crosssection for the
electron-muon scattering in
$s$-channel ($e^-e^+\rightarrow \mu^-\mu^+$)
in Figure \ref{s_e_mu} with the matrix element~\eqref{matrix-element-square-4} is
\begin{eqnarray}
d\sigma^{\rm{s}}_{\rm B \neq 0 } \ems
&=&\frac{(2\pi)^4}{F} \overline{\left|\mathfrak{M}_{\rm s}\right|^2} \ems \delta^4(K_{\omit{y}}+P_{\omit{y}}-k_{\omit{y}}-p_{\omit{y}})\frac{d^3\vec{P}}{(2\pi)^32E_P}\frac{d^3\vec{K}}{(2\pi)^32E_K}\nonumber\\
&=&\frac{ \overline{\left|\mathfrak{M}_{\rm s}\right|^2}\ems}{(2\pi)^24F} \delta(E_P+E_K-E_p-E_k)\delta^3(\vec{K}_{\omit{y}}+\vec{P}_{\omit{y}}-\vec{k}_{\omit{y}}-\vec{p}_{\omit{y}})\nonumber\\
&&\hspace{10cm}\times\frac{d^3\vec{P}}{E_P}\frac{d^3\vec{K}}{E_K}.\nonumber
\end{eqnarray}
In the center-of-mass (cm) frame, $\vec{p}+\vec{k}=0$, this also implies, $\vec{p}_{\omit{y}}+\vec{k}_{\omit{y}}=0$.
Thus then crosssection becomes
\begin{eqnarray}
d\sigma^{\rm{s}}_{\rm B \neq 0} \ems&=&\frac{ \overline{\left|\mathfrak{M}_{\rm s}\right|^2}}{(2\pi)^24F} \delta(E_P+E_K-E_p-E_k)\delta^3(\vec{K}_{\omit{y}}+\vec{P}_{\omit{y}})\frac{d^3\vec{P}}{E_P}\frac{d^3\vec{K}}{E_K}\\
&=&\frac{ \overline{\left|\mathfrak{M}_{\rm s}\right|^2}_{\vec{K}_{\omit{y}}=-\vec{P}_{\omit{y}}}}{(2\pi)^24F} \delta(E_P+E_K-E_p-E_k)\frac{d^3\vec{P}}{E_KE_P}.
\end{eqnarray}
In the presence of magnetic field, the momentum integration gets
factorized into parallel and perpendicular components
with respect to the direction of magnetic field ($z$-direction),
where the integral over $d^2P_\perp$ in strong
magnetic field limit becomes ($\int_0^{|eB|}d^2P_\perp
=\pi |eB|$). Thus the total crosssection is obtained
by integrating over the parallel component of momentum($P_z$)
\begin{eqnarray}
\sigma^{\rm{s}}_{\rm B \neq 0} \ems &=&\int_{-\infty}^\infty \frac{\pi|eB| \overline{\left|\mathfrak{M}_{\rm s}\right|^2}_{K_z=-P_z}}{(2\pi)^24F} \delta(E_P+E_K-E_p-E_k)\frac{dP_z}{E_KE_P}.\label{d-sig}
\end{eqnarray}
In a strong magnetic field limit ($eB>>m^2; n=0$), the perpendicular
component ($\perp$) of the momentum is zero~\cite{Gusynin:1998nh} so
the particles can only move in $z$ direction. They can either move
in $+ve$ $ z$ direction or in $-ve$ $ z$ direction. Accordingly
the four momentum dot product can be written as
\begin{equation}
p.k=
\begin{cases}
E_pE_k-|\vec{p}||\vec{k}|,& \mbox{if $\vec{p}$ and $ \vec{k}$ are in
the same direction} \\
E_pE_k+|\vec{p}||\vec{k}|,& \mbox{if $\vec{p}$ and $ \vec{k}$ are in
the opposite direction.}
\end{cases}
\end{equation}
In extreme relativistic limit we can neglect the dot product where the momenta
are in the same direction compared to them in opposite direction.
The diagram for the process in the center-of-mass frame (for
$\theta=0$ degree, where
$\theta$ is the angle between $p$ and $P$) is
drawn in the Figure \ref{ricf}.
\begin{figure}[h]
\begin{center}
\includegraphics[height=12cm,width=8cm,angle=-90]{ricf.ps}
\vspace{-1cm}
\caption{Reaction in center-of-mass frame}\label{ricf}
\end{center}
\end{figure}
For our case $\theta$ (scattering angle) can have two values $0$
and $180$ degree, for
$\theta=0$ the $t$ variables in eq~\eqref{matrix-element-square-4} are
negligible as compared to all $u$ variables whereas for $\theta=180$
the $u$ variables are negligible as compared to all $t$ variables.
Thus the squared matrix element eq~\eqref{matrix-element-square-4}
at high energy for $\theta=0$ degree
gets simplified
\begin{eqnarray}
{\overline{\left|\mathfrak{M}_{\rm s}\right|^2}}_{\rm B \neq 0} (e^- e^+
\rightarrow \mu^- \mu^+)
&=&\frac{e^4}{4s^2}\Big[u^2+u_pu_k\Big],
\end{eqnarray}
and for $\theta=180$ degree, it becomes
\begin{eqnarray}
{\overline{\left|\mathfrak{M}_{\rm s}\right|^2}}_{\rm B \neq 0} (e^- e^+
\rightarrow \mu^- \mu^+) &=&\frac{e^4}{4s^2}\Big[t^2+t_pt_k\Big].
\end{eqnarray}
Let us denote $|\vec{p}|=|\vec{k}|=p_i$ and
$|\vec{P}|=|\vec{K}|=p_f$ so
$E_p$ and $E_k$ become the same (say, $E_i$) whereas $E_P$ and $E_K$
become equal (say, $E_f$). As a consequence the square of the matrix
element for $\theta=0$ degree
comes out to be same as for the $\theta=180$ degree. Hence
\begin{equation}
\overline{\left|\mathfrak{M}_{\rm s}\right|^2}{\ems}
=\frac{2e^4}{s^2}E_f^2\left[E_i+p_i \right]^2,\label{ep_in_cm}
\end{equation}
and the flux factor for collinear collision is
\begin{eqnarray}
F&=&|\vec{v}_p-\vec{v}_k|2E_p2E_k\nonumber\\
&=&4p_i\sqrt{s},
\end{eqnarray}
where $\sqrt{s}=E_p+E_k$.
Using the expression of the flux factor $F$ and the squared matrix
element
$\overline{\left|\mathfrak{M}_{\rm s}\right|^2}$, the crosssection
\eqref{d-sig} can be rewritten as
\begin{eqnarray}
\sigma^{\rm{s}}_{\rm B \neq 0} {\ems}=\frac{e^4|eB|}{16\pi} \int_{-\infty}^{\infty}
\frac{\pi e^4|eB|\left[E_i+p_i \right]^2}{s^2 p_i\sqrt{s}}
\delta(W-\sqrt{s})\frac{E_f^2}{p_fW}dW,
\end{eqnarray}
where $W=E_P+E_K$.
With the further approximation: $p_f \approx E_f$
and $p_i \approx E_i$, the crosssection for the $e$-$\mu$
scattering in
$s$-channel\footnote{Detailed calculation of crosssection for the
process $e^-e^+\rightarrow \mu^-\mu^+$ is given in Appendix C},
{\em i.e.}
for the annihilation process ($e^-e^+\rightarrow \mu^-\mu^+$) in the
lowest order takes the final form as a function of the
center-of-mass energy
\begin{eqnarray}
\sigma^{\rm{s}}_{\rm B \neq 0} (e^-e^+\rightarrow \mu^-\mu^+)&=&\frac{\pi \alpha^2 |eB|}{s^2}.
\end{eqnarray}
The approximations, $p_f \approx E_f$
and $p_i \approx E_i$ only hold good at extremely high energies,
where the masses (order of MeV) can be neglected.
For the sake of comparison, the crosssection for the same
in vacuum in the lowest order is~\cite{halzen_martin}
\begin{equation}
\sigma^{\rm{s}}_{\rm B=0} (e^-e^+\rightarrow \mu^-\mu^+)=
\frac{4\pi \alpha^2}{3s}.
\end{equation}
One thus immediately infer that in the
presence of strong magnetic field, the crosssection
is inversely proportional to the fourth power
of the center-of-mass energy while in the absence of magnetic
field, $\sigma_{\rm B =0}$
is inversely proportional to the square of the center-of-mass energy.
To see the effect of strong magnetic field on the annihilation process, we
have plotted a variation of the crosssection with the
center-of-mass energy in the
presence and absence of magnetic field in Fig-\ref{sig-plot}, where $M$ is
the mass of muon.
\begin{figure}[h]
\begin{center}
\includegraphics[height=16cm,width=10cm,angle=-90]{s-graph.ps}
\vspace{-2cm}
\caption{$\sigma_B$ vs center-of-mass Energy at different strength of magnetic field.}\label{sig-plot}
\end{center}
\end{figure}
From the above graph, it is evident that the strong magnetic field
suppresses the scattering, which can be understood by the
fact that the availability of phase space in the presence of
magnetic field
is reduced drastically. Moreover we have also found that for a
fixed center-of-mass energy, as we increase the magnetic field,
$\sigma$ increases
linearly.
\subsection{Electron-Muon scattering in $t$ channel: $e^-\mu^-\rightarrow e^-
\mu^-$ process}
The crosssection for the electron-muon scattering in $t$
channel, {\em i.e.} for the $e^-\mu^-\rightarrow e^-\mu^-$ process
diverges at $t=0$, since the matrix element in $t$-channel diagram has
a pole at
$t=0$. Let us now examine below how the pole $t=0$ translates into the
momentum variable in the final state. The Mandelstam variables, $t$ and $u$
for the Figure~\ref{t_e_mu} are defined as
\begin{eqnarray}
t^2&=&4|\vec{p}|^2|\vec{P}|^2(\cos \theta -1)^2=4|\vec{p}|^2|\vec{P}|^2\left(\frac{P_z}{|\vec{P}|}-1\right)^2=4p_i^2p_f^2\left(\frac{P_z}{p_f}-1\right)^2,\label{tt}\\
u^2&=&4|\vec{p}|^2|\vec{K}|^2(\cos \theta +1)^2=4|\vec{p}|^2|\vec{K}|^2
\left(\frac{P_z}{|\vec{P}|}+1\right)^2=4p_i^2p_f^2\left(\frac{P_z}{p_f}
+1\right)^2\label{uu}~,
\end{eqnarray}
where we denote the momenta $|\vec{p}|$ and $|\vec{k}|$ by $p_i$ and
the momenta $|\vec{P}|$, $|\vec{K}|$ by $p_f$.
As mentioned earlier, in the strong magnetic field limit,
electrons occupy only the lowest Landau levels ($n=0$), so the lower
limit of the transverse momentum becomes vanishing small, i.e. $P_\perp
\sim 0$. Therefore, the momentum, $p_f$ (=$|\bf P|=\sqrt{P_\perp^2+P_z^2}$) simply
becomes $P_z$, hence the pole $t=0$ appears.
Therefore, using the matrix element for $e$-$\mu$ scattering for $t$-channel
\eqref{matrix-element-squared-t} as well a lower cut-off, $\epsilon_B$ to
the transverse
momentum, $P_\perp$ the crosssection \eqref{liff} looks as
\begin{eqnarray}
\sigma^{\rm{t}}_{\rm B \neq 0}(e^-\mu^-\rightarrow e^-\mu^-)&=&
\int_{\epsilon_B}^{\sqrt{eB}} d P_\perp \int_{-\infty}^\infty dP_z
\frac{e^4}{4t^2}\Big[s^2+u^2+s_ks_K+u_Ku_k\Big]
\frac{\delta(W-\sqrt{s})}{(2\pi)4F}\frac{P_\perp}{E_f^2},\qquad
\end{eqnarray}
with the notations: $W=E_P+E_K$, $\sqrt{s}=E_p+E_k$, $E_p=E_k=E_i$ and
$E_K=E_P=E_f$. Thus after integrating the momentum
integrations over $dP_\perp$ and $dP_z$, the crosssection
for the $e^-\mu^-\rightarrow e^-\mu^-$ process in the lowest order is
\footnote{Calculated in Appendix D}
\begin{eqnarray}
\sigma^{\rm{t}}_{\rm B \neq 0}(e^-\mu^-\rightarrow e^-\mu^-)&=&
\frac{\pi \alpha^2}{2s^2}|eB|-\frac{2\pi \alpha^2}{|eB|}+2\pi \alpha^2
\lim_{\epsilon_B \to 0}\left[\frac{1}{\epsilon_B^2}\right]\label{t-sig},
\end{eqnarray}
which implies a divergence in $\epsilon_B \to 0$ limit.
For the sake of comparison, we calculate the same crosssection
in the extreme relativistic limit in the absence of strong magnetic
field using matrix element from~\eqref{t-vacuum}
for the $e$-$\mu$ scattering in the $t$-channel
\begin{equation}
\sigma^{\rm{t}}_{\rm B =0} (e^-\mu^-\rightarrow e^-\mu^-)=
-\frac{4\pi \alpha^2}{s}+\frac{4\pi \alpha^2}{s}\lim_{\epsilon_V \to 0}
\left[\frac{2}{\epsilon_V}+\ln{\Big(\frac{\epsilon_V}{2}\Big)}\right].
\end{equation}
This also shows a divergence in the lower limit of $\epsilon_V$, {\em i.e.}
$\epsilon_V \to 0$, which, in turn, arises due to the
lower limit of scattering angle (between ${\bf p}$ and ${\bf P}$) as
\begin{equation}
\epsilon_V=1-\cos \theta_0.
\end{equation}
However, the above divergences in the presence and the absence of strong
magnetic field are related to each other and we can geometrically
derive\footnote{Calculated at the end of Appendix D} an equation
using the fact that $P_\perp=P_z \tan\theta$,
which is
\begin{equation}
\frac{1}{\epsilon_V}=\frac{3}{2}+\frac{s}{2\epsilon_B^2},
\label{magnetic-vacuum}
\end{equation}
Hence, the above relation helps us to compare the crosssection in vacuum
with the crosssection in the presence of magnetic field
\begin{equation}
\sigma^{\rm{t}}_{\rm B =0} (e^-\mu^-\rightarrow e^-\mu^-)
=\frac{8\pi \alpha^2}{s}+\frac{4\pi \alpha^2}{s}\lim_{\epsilon_V \to 0}
\left[\ln{\Big(\frac{\epsilon_V}{2}\Big)}\right]+4\pi\alpha^2\lim_{\epsilon_B \to 0}\left[\frac{1}{\epsilon_B^2}\right].
\end{equation}
One thus finds that the logarithmic divergence in vacuum disappears due to
the presence of external magnetic field and apart from that in
magnetic field there is an another finite term which is independent
of $s$ but decreases with the increasing magnetic field.
\subsection{Bhabha Scattering: $e^-e^+\rightarrow e^-e^+$}
The crosssection for the Bhabha Scattering, {\em i.e.} for
the $e^-e^+\rightarrow e^-e^+$ processes in the lowest-order can be obtained
from the definition \eqref{liff}, with the matrix element \eqref{Bhabha_M}.
In the presence of strong magnetic field, $\sigma (e^-e^+\rightarrow
e^-e^+$) can be decomposed into $s$- and $t$-channel
contribution due to {\em the vanishing interference term }
\begin{eqnarray}
\sigma_{\rm B \neq 0} (e^-e^+\rightarrow e^- e^+) = \sigma^s_{\rm B \neq 0} (e^-e^+\rightarrow
e^- e^+) + \sigma^t_{\rm B \neq 0} (e^-e^+\rightarrow e^- e^+),\label{Bhabha_sig_tot}
\end{eqnarray}
where the $s$- and $t$-channel contribution
are given by
\begin{eqnarray}
\sigma_{\rm B \neq 0}^{s}\bhabha&=&\frac{e^4}{16\pi s^2}|eB|,\\
\sigma_{\rm B \neq 0}^{t}\bhabha&=&\frac{e^4}{32\pi s^2}|eB|+\frac{e^4}{8\pi}
\lim_{\epsilon_B \to 0}\left[\frac{1}{\epsilon_B^2}-\frac{1}{|eB|}
\right],
\end{eqnarray}
respectively. Therefore the total crosssection
\eqref{Bhabha_sig_tot} for the Bhabha
scattering yields
\begin{eqnarray}
\sigma_{\rm B \neq 0} (e^-e^+\rightarrow e^-e^+)
=\frac{3 \pi \alpha^2}{2 s^2}|eB|-\frac{2\pi \alpha^2}{|eB|}+2 \pi \alpha^2 \lim_{\epsilon_B \to 0}\left[\frac{1}{\epsilon_B^2}\right].\label{bhabha-cross}
\end{eqnarray}
To isolate the effect of strong magnetic field, we have also
calculated
the same at the lowest-order in vacuum only using the matrix
element from~\eqref{bhabha-vacuum-m}, where the interference term
is
nonzero unlike the former case and is given by
\begin{eqnarray}
\sigma_{\rm B =0} (e^-e^+\rightarrow e^-e^+)&=&\sigma^{\rm s}_{\rm B =0}+\sigma^{\rm interference}_{\rm B =0}+\sigma^{\rm t}_{\rm B =0}\nonumber\\
&=&\left[\frac{4\pi \alpha^2}{3s}\right]-\left[\frac{4\pi \alpha^2}{s}\lim_{\epsilon_V \to 0}\ln{\frac{\epsilon_V}{2}}\right]+\left[\frac{4\pi \alpha^2}{s}\lim_{\epsilon_V \to 0}\left(\frac{2}{\epsilon_V}-1+\ln{\frac{\epsilon_V}{2}}\right)\right]\nonumber\\
&=&-\frac{8\pi \alpha^2}{3s}+\frac{8\pi \alpha^2}{s}\lim_{\epsilon_V \to 0}\left[\frac{1}{\epsilon_V}\right].
\end{eqnarray}
Again using the relation \eqref{magnetic-vacuum}, we can write the above
crosssection in terms of the parameter ($\epsilon_B$) in the presence of
strong magnetic field, which causes the divergence
\begin{eqnarray}
\sigma_{\rm B =0} (e^-e^+\rightarrow e^-e^+)&=&\frac{4\pi \alpha^2}{3s}+
4\pi\alpha^2\lim_{\epsilon_B \to 0}\left[\frac{1}{\epsilon_B^2}\right].
\label{bhabha-vacuum}
\end{eqnarray}
If we leave the divergent part and compare only with the finite part
then we find that apart from a constant magnetic field dependent
term,
the crosssection in vacuum at the lowest order decreases
with the center-of-mass energy ($\sqrt{s}$) slower than the same in
the presence of
strong magnetic field. However, the presence of strong magnetic
field does not
alter the degree of divergence.
\subsection{M\o ller Scattering: $e^-e^-\rightarrow e^-e^-$}
The crosssection for the M\o ller scattering at the lowest order
can be obtained from its matrix element \eqref{Moller_M}, which is
factorizable into $u$ and $t$-channel contributions due to the
vanishing interference term. The crosssection for the $t$-channel
matrix element in extreme relativistic limit is same as the
crosssection for electron-muon scattering \eqref{t-sig} in $t$-channel
$\emt$ process. So
we are left with the crosssection due to the $u$-channel and can be
calculated from its matrix element \eqref{moller-u} in a strong magnetic field
\begin{eqnarray}
\sigma^{\rm{u}}_{\rm B \neq 0} &=& \int_{-\infty}^\infty dP_z
\int_0^{\sqrt{eB}} d P_\perp~ \overline{\left|\mathfrak{M}_{\rm u}
\right|^2}\moller \frac{\delta(W-\sqrt{s})}{(2\pi)4F}\frac{P_\perp}{E_f^2},
\end{eqnarray}
with the notations: $W=E_P+E_K$, $\sqrt{s}=E_p+E_k$, $E_p=E_k=E_i$ and
$E_K=E_P=E_f$. In the above integral, the matrix element has a
pole at $u=0$,
which could be translated in terms of momentum variable as follows:
In the presence of strong magnetic field, the lower limit of the
transverse
momentum ($P_\perp$) becomes vanishingly small. As a result, the momentum,
$\bf P$ (=$\sqrt{P_\perp^2+P_z^2}$) comes out to be $ \pm P_z$, which give rise
$t=0$ and $u=0$ poles, defined in \eqref{tt} and \eqref{uu},
respectively and
can be circumvented by taking a lower cut-off to the lower limit of
transverse momentum ($P_\perp$) integration.
As an artifact of the above observation, the $u$-channel matrix element
squared in M\o ller scattering \eqref{moller-u} in the momentum interval,
$P_z \in (-\infty,0]$ is mapped into $t$-channel matrix element \eqref{Moller_M}
in the momentum interval $P_z \in [0,\infty)$ and {\em vice versa}
\begin{eqnarray}
\overline{\left|\mathfrak{M}_{\rm u}\right|^2}_{P_z \in (-\infty,0]}&=&\overline{\left|\mathfrak{M}_{\rm t}\right|^2}_{P_z \in [0,\infty)},\\
\overline{\left|\mathfrak{M}_{\rm u}\right|^2}_{P_z \in [0,\infty)}&=&\overline{\left|\mathfrak{M}_{\rm t}\right|^2}_{P_z \in (-\infty,0]}.
\end{eqnarray}
As a consequence the crosssections from both channels come out
to be same except for the fact that in $t$-channel, the
crosssection peaks
at the forward angle ($\theta=0$) while the $u$-channel peaks
at $\theta=180$. Thus the crosssection for the $\moller$ scattering
in the lowest-order is obtained by doubling the crosssection in
$t$ channel \eqref{t-sig}
\begin{eqnarray}
\sigma_{\rm B \neq 0} (e^-e^-\rightarrow e^-e^-)&=&\frac{\pi \alpha^2}{s^2}|eB|-\frac{4 \pi \alpha^2}{|eB|}+4 \pi \alpha^2 \lim_{\epsilon_B \to 0}\left[\frac{1}{\epsilon_B^2}\right].
\end{eqnarray}
The crosssection for the M\o ller scattering in vacuum in the
lowest order using the matrix element from~\eqref{moler-vacuum-m}
can be easily calculated as
\begin{eqnarray}
\sigma_{\rm B=0} (e^-e^-\rightarrow e^-e^-)&=&\sigma_{\rm B =0}^{\rm u}+\sigma_{\rm B =0}^{\rm interference}+\sigma_{\rm B =0}^{\rm t}\nonumber\\
&=&\left[\frac{4\pi \alpha^2}{s}\lim_{\epsilon_V \to 0}\left(\frac{2}{\epsilon_V}-1+\ln{\frac{\epsilon_V}{2}}\right)\right]-\left[\frac{8\pi \alpha^2}{s}\lim_{\epsilon_V \to 0}\ln{\frac{\epsilon_V}{2}}\right]\nonumber\\
&+&\left[\frac{4\pi \alpha^2}{s}\lim_{\epsilon_V \to 0}\left(\frac{2}{\epsilon_V}-1+\ln{\frac{\epsilon_V}{2}}\right)\right]\\
&=&-\frac{8\pi \alpha^2}{s}+\frac{16\pi \alpha^2}{s}\lim_{\epsilon_V \to 0}
\left[\frac{1}{\epsilon_V}\right],
\end{eqnarray}
which can be compared with the result in a strong magnetic field by
replacing
$\epsilon_V$ in terms of $\epsilon_B$ through the relation \eqref{magnetic-vacuum}
\begin{eqnarray}
\sigma_{\rm B =0} (e^-e^-\rightarrow e^-e^-)&=&\frac{16\pi \alpha^2}{s}+8\pi \alpha^2\lim_{\epsilon_B \to 0}\left[\frac{1}{\epsilon_B^2}\right].
\end{eqnarray}
Like other processes the $s$-dependence in vacuum is being modified due to
the presence of strong magnetic field whereas the
diverging component in vacuum ($1\over \epsilon_B^2$) becomes
halved due to the magnetic field.
\section{Results and Discussions}
We have revisited the lepton-lepton scattering at the lowest order
in an additional presence of strong magnetic
field. The recent observations at the ultra relativistic
heavy ion collisions at Relativistic Heavy-ion Collider and Large
Hadron Collider, where
a very strong magnetic field up to ${10}^{18}$ - ${10}^{20}$ Gauss
is expected to be produced for the noncentral events, motivates
us to revisit the above processes in a strong magnetic field.
In particular, we have calculated the crosssection for
electron-muon ($e$-$\mu$) scattering in both $s$- and $t$-channel, Bhabha
scattering, and M\o ller scattering in the presence of a strong
magnetic field
($|eB|>>m^2$, $m$ is the mass of electron or muon). For that purpose,
using the Dirac spinor in strong magnetic field, we
have first calculated the square of the matrix element and then
summed
over the final spin states. We have found that unlike in vacuum, the
interference term in Bhabha scattering between $s$- and $t$-channel
and in M\o ller scattering between $t$- and $u$-channel
contribution
in the presence of strong magnetic field vanishes. Secondly we have
illustrated the usual procedure to compute the
crosssection from the transition amplitude for the
above mentioned processes in the presence of strong magnetic field.
We have noticed that the $e$-$\mu$ scattering in $s$-channel,
{\em i.e.} $e^-e^+\rightarrow \mu^-\mu^+$ process gets suppressed
due to the presence of strong magnetic field. More precisely
the crosssection at a fixed magnetic field is inversely
proportional
to the fourth power of the center-of-mass energy compared to the
vacuum alone,
where $\sigma$ is inversely proportional to the square of the
center-of-mass energy.
However, for a fixed center-of-mass energy, the crosssection
increases with
the magnetic field.
On the other hand the $e$-$\mu$ scattering in lowest order for $t$-channel,
{\em i.e.} $e^- \mu^- \rightarrow e^- \mu^-$ process also diverges
like in vacuum but in the presence of strong magnetic field, the logarithmic
divergence in vacuum disappears and the infrared divergence remains.
As far as finite terms are concerned, the crosssection decreases
faster like the $s$-channel. However, there is a negative
term, which is independent of the center-of-mass energy and
decreases with
the magnetic field. The above observation is also found
in Bhabha and M\o ller scattering in the strong magnetic field.
The divergence in Bhabha scattering in vacuum arises at the lower
limit of the incident angle whereas the same in the presence of
strong magnetic field arises due to the lower limit of transverse
momentum. However the above divergences in the presence and in the absence
of strong magnetic field inter-related to each other
and we have derived the relation among these two
divergences geometrically so that we can
compare the crosssections in both cases at the same footing.
However, the above mentioned processes have also been studied extensively
up to higher-order in vacuum ({\em i.e.} in the absence of magnetic
field), as a result the divergences appeared have been controlled
by the higher-order corrections. For an example,
the collinear divergence in the Bhabha scattering
had been cured by adding the corrections
to the tree-level vacuum result by the radiative corrections
in Ref.~\cite[and references therein]{G_montagna}, the $\mathcal{O}
(\alpha^3)$ corrections~\cite{M_consoli}, the $\mathcal{O}(\alpha^4)$
corrections with the full mass dependence~\cite{bonciani_ferroglia}
etc., where the contributions from the higher-order diagrams have been
calculated and the IR and UV divergences were regularized by the dimensional
regularization scheme.
\section {Acknowledgement}
We thank to Mr. Mujeeb Hasan, Mr. Bhaswar Chatterjee, Ms. Shubhalaxmi Rath and Mr.
Jitendra Pal for their help from time to time.
|
\section{Introduction}
\label{sec:intro}
Neutron stars are some of the most extreme objects in the Universe,
and thus, they serve as a unique laboratory to probe fundamental physics.
Their large masses ($m \approx 1.4\, M_{\odot}$) combined with their small radii
($R \approx 12$ km) result in supranuclear densities at their cores,
whose description challenges our current understanding of matter. The
latter is encoded in the star's equation of state, whose determination is an
outstanding problem in nuclear astrophysics~\cite{Lattimer:2015nhk}.
Moreover, the strong gravitational fields produced by neutron stars result in gravitational potentials
$\sim G m /(R c^2)$ that are nine orders of magnitude
larger than what we can probe on Earth's surface~\cite{Will:2014kxa,Baker:2014zba}.
Therefore, to correctly describe these stars, we must rely on a relativistic
theory of gravity~\cite{Bonolis:2017fdf}.
The leading theory is of course Einstein's theory of general relativity. During its
centennial existence, the theory has shown a remarkable predictive power, being
consistent within all experimental tests carried out so far, ranging from local,
Solar System experiments~\cite{Will:2014kxa}, to the spectacular
detection of gravitational waves by the LIGO/Virgo
collaboration~\cite{Abbott:2016blz,TheLIGOScientific:2016src}.
This consistency with observation is even
more striking when one notes that the theory (unlike most of its alternatives)
does not possess any free parameters that can be tuned to make its predictions
agree with Nature.
Given the success of general relativity, why should we even consider
modifications to it and examine their observational consequences?
The reasons are many, but they can be organized in
two main classes~\cite{Clifton:2011jh,Berti:2015itd,Alexander:2009tp}.
On the observational front, the late time expansion of the
Universe~\cite{Riess:1998cb,Perlmutter:1998np}, the rotation
curve of galaxies~\cite{Sofue:2000jx,Bertone:2016nfn}, the baryon-antibaryon
asymmetry~\cite{Spergel:2003cb,Canetti:2012zc},
and other cosmological observations seem to point at either exotic dark fluids
or modifications to general relativity.
On the theoretical front, the incompatibility of general relativity with quantum
mechanics has prompted many attempts at extensions, from string theory to
loop quantum gravity and other variations.
Can neutron star observations\footnote{We here focus
on properties of {\it individual} stars. Neutron stars have already proven
invaluable tools to constrain deviations of general relativity when either in
binaries~\cite{Damour:2007uf,Wex:2014nva,Kramer:2016kwa} (or
triple~\cite{Archibald:2018oxs}) systems or
more recently binary neutron stars mergers~\cite{TheLIGOScientific:2017qsa}
(see e.g.~\cite{Sakstein:2017xjx,Baker:2017hug,Ezquiaga:2017ekz,Creminelli:2017sry,Bettoni:2016mij}).}
be used to learn about gravity in extreme environments?
A general prediction of modified theories of
gravity is that the bulk properties of the star (e.g. its radius and mass)
are different from those predicted by general relativity.
Tests of gravity in this direction are however
limited by a strong degeneracy problem between the equation of state and the gravity theory:
the modifications of the bulk properties of these stars caused by changes
in the gravitational theory are (often) degenerate with modifications due to
different equations of state~\cite{Glampedakis:2015sua}.
One option to bypass this issue is to focus on electromagnetic phenomena
in the vicinity of neutron stars~\cite{Psaltis:2008bb}. These phenomena include
e.g.~atomic spectral lines~\cite{DeDeo:2003ju},
burst~\cite{Psaltis:2007rv,Glampedakis:2016pes} and quasiperiodic
oscillations~\cite{DeDeo:2004kk,Doneva:2014uma,Pappas:2015npa,Glampedakis:2016pes}.
One can argue that in these scenarios one can in principle probe the {\it exterior spacetime} of the star,
offering a glimpse on possible deviations from general relativity, without
worrying about the intricacies of the stellar interior.
The observation of x-ray waveforms or pulse profiles
from rotating neutron stars is another potentially interesting
phenomenon to consider~\cite{Arzoumanian:2009qn}.
In this scenario, a region of of the neutron star's surface becomes hot
(relative to the rest of the star's surface) generating an x-ray flux modulated
by the star's rotation.
This {\it hot spot} can be formed in a number of situations
(see~\cite{Poutanen:2008pg,Ozel:2012wu,Watts:2016uzu} for reviews).
In accretion-powered pulsars, material is channeled through the magnetic field
lines and heats the star up when it reaches its magnetic poles.
In burst oscillations, a thermonuclear explosion caused by infalling
matter results in a hot spot on the surface of the accreting neutron
star.
In all these cases, the modeling of the resulting waveform, combined with
x-ray observations, allows for the extraction of a number of properties of the
source, including the neutron star's mass and
radius, see e.g.~\cite{Miller:1997if,Weinberg:2000ip,Muno:2002es,Bogdanov:2012md,Poutanen:2003yd,Lo:2013ava,Miller:2014mca,Stevens:2016xiw,Lo:2018hes,Salmi:2018gsn}.
The ongoing NICER (Neutron star Interior Composition Explorer) mission~\cite{2012SPIE.8443E..13G,2014SPIE.9144E..20A,2017NatAs...1..895G} offers a
substantial improvement over the preceding x-ray observatory,
the Rossi X-ray Timing Explorer (RXTE), opening the path to measurements of
stellar masses and radii with unprecedented accuracy -- with immediate implications
to our understanding on the neutron star equation of star.
Given, the scientific potential of NICER, it is natural to ask:
{\it can we use its observations to probe the strong-field regime of gravity}?
In this paper, we take the first necessary steps to find an answer to this question.
We find hints that NICER can indeed be used to probe the strong-field
regime of gravity, but a definite answer will require a plethora of theoretical and
data analysis work that this paper now enables
(see~\cite{Sotani:2017bho,Sotani:2017rrt} for independent recent work in this direction).
\subsection{Executive summary}
We present a complete toolkit to model the x-ray flux from radiating
neutron stars in scalar-tensor gravity, one of the most well-studied and
well-motivated extensions of general relativity~\cite{Chiba:1997ms,Sotiriou:2015lxa}.
This class of theories extends general relativity by introducing a scalar field that
couples to the metric nonminimally, thus violating the strong-equivalence principle.
Our formalism and the resulting toolkit are completely model-independent
within this class of scalar-tensor theories. Moreover, the resulting waveforms
include Doppler shifts, relativistic aberration and
time-delay effects, thus extending~\cite{Sotani:2017rrt}. All of these effects
are critical ingredients that are necessary to develop an accurate pulse-profile model,
and thus, our toolkit now enables a complete data analysis study that will be carried
out in the future.
Figure~\ref{fig:pp_example} shows a sample waveform that takes all the
previously mentioned effects into account, illustrating the difference
between the predictions of both theories. As will discuss in
Sec~\ref{sec:pulseprofile}, we find that the presence of the scalar field influences the
exterior spacetime of the neutron star, altering the bending of light and the time-delay
experienced by photons emitted by the star.
The net result of these contributions is a waveform that can be considerably
different in comparison to general relativity's predictions. Such a model now enables,
for the first time, a serious data analysis investigation of whether such waveform differences
can be detected/constrained with data or whether they are degenerate with other
waveform parameters.
\begin{figure}[t]
\includegraphics[width=0.48\textwidth]{fig_illustrative_pulseprofile.pdf}
\caption{An illustrative bolometric flux for a single hot spot over one
revolution of the rotating neutron star in general relativity and scalar-tensor
gravity. The star has $u \equiv 2 G_{\ast} m / (R c^2) = 0.5$ and mass
$m = 1.4\, M_{\odot}$ in both cases. The only difference between the two models
is the presence
of a nonzero scalar charge for the star in scalar-tensor gravity. The magnitude
of the scalar charge is quantified by the scalar charge-to-mass ratio $Q$ (described
in Sec.~\ref{sec:review}) which is zero in general relativity and controls by how
much the spacetime is different from the usual Schwarzschild spacetime.
This particular examples a sample star labeled $c_3$ as discussed in
Sec.~\ref{sec:review} (cf. Table~\ref{tab:stellar_models}).}
\label{fig:pp_example}
\end{figure}
The remainder of this paper is organized as follows.
In Sec.~\ref{sec:review}, we briefly review the basic of scalar-tensor gravity,
discussing the properties of the exterior spacetime of neutron stars in this theory.
Next, in Sec.~\ref{sec:pulseprofile}, we present in detail how we construct
our pulse profile model.
Finally, in Sec.~\ref{sec:results} we show a few examples of the pulse
profile of burst oscillations and discuss the modifications introduced by
scalar-tensor gravity. We discuss in details the roles played
by different effects on the overall shape of the resulting waveform.
We close with Sec.~\ref{sec:conclusion}, summarizing our main findings and
discussing some possible extensions and applications of this work.
\section{Overview of scalar-tensor theory}
\label{sec:review}
\subsection{Action and field equation}
Scalar-tensor theories are one of the most widely studied
and well-motivated extensions to general relativity.
In this class of metric theories of gravity, the gravitational interaction
is mediated by an additional scalar field $\varphi$. Assuming this
scalar field to be long-ranged (i.e.~massless), the theory
can be described by an action $S_{\ast}$ in the so-called Einstein
frame given by~\cite{Damour:1992we}:
\begin{align}
S_{\ast} &= \frac{c^4}{4 \pi G_{\ast}} \int \frac{{\rm d}^4 x}{c}\sqrt{-g_{\ast}}
\left[ \frac{1}{4}R_{\ast}
-\frac{1}{2}g^{\mu\nu}_{\ast}\nabla_{\mu}\varphi\nabla_{\nu}\varphi
\right]
\nonumber \\
&\quad+ S_{\rm m}[\psi_{\rm m}, g_{\mu\nu}]\,,
\label{eq:action}
\end{align}
where $c$ is the speed of light in vacuum, and $G_{\ast}$ is the
bare Newtonian gravitational constant.
The action is written in terms
of the Einstein frame metric $g_{\ast\mu\nu}$, with metric determinant
$g_{\ast} \equiv \textrm{det}\,[g_{\ast\mu\nu}]$ and Ricci curvature
scalar $R_{\ast}$.
Matter fields, collectively represented by $\psi_{\rm m}$, are
minimally coupled to the Jordan frame metric
$g_{\mu\nu} \equiv A^2(\varphi)g_{\ast\mu\nu}$, with $A(\varphi)$
being a conformal factor. As such, clocks and rods in the real world
measure time intervals and distances of $g_{\mu\nu}$ and not of $g_{\ast\mu \nu}$.
The advantage of working in the Einstein frame is that the field equations
are simpler, making it technically easier to derive predictions
for the theory.
However, in the end, we will express the relevant observables
in the Jordan frame.
The field equations, obtained by varying Eq.~\eqref{eq:action} with
respect to $g^{\mu\nu}_{\ast}$ and $\varphi$, are
\begin{align}
R_{\ast\mu\nu} &= 2 \partial_{\mu}\varphi\partial_{\nu}\varphi
+ \frac{8\pi G_{\ast}}{c^4}
\left( T_{\ast\mu\nu} - \frac{1}{2}T_{\ast}g_{\ast\mu\nu} \right)\,,
\label{eq:b}
\\
\Box_{\ast} \varphi &= -\frac{4\pi G_{\ast}}{c^4} \alpha(\varphi) T_{\ast}\,,
\label{eq:a}
\end{align}
where $T_{\ast}^{\mu\nu} \equiv (2 c / \sqrt{-g_{\ast}}) (\delta S_{\rm m} / \delta g_{\ast\mu\nu})$
is the Einstein frame energy-momentum tensor and $T_{\ast} \equiv g_{\ast}^{\mu\nu}T_{\ast\mu\nu}$
is its trace. The energy-momentum tensor acts as a source for the scalar field, through the
coupling $\alpha(\varphi) \equiv {\rm d} \log A(\varphi) / {\rm d} \varphi$. The choice of $A(\varphi)$
determines a specific scalar-tensor theory and a number of models have been studied in
the literature.
From the definition of $T_{\ast\mu\nu}$, one can derive the following expressions that connect
the Einstein and Jordan frame energy-momentum tensors (and their traces):
\begin{equation}
T_{\mu\nu} = A^2(\varphi) T_{\ast\mu\nu}\,,\,\,
T^{\mu\nu} = A^{-6}(\varphi)T^{\ast\mu\nu}\,,\,\,
T = A^{-4}(\varphi) T_{\ast}\,.
\end{equation}
Finally, the covariant divergences of $T_{\ast}^{\mu\nu}$
and $T^{\mu\nu}$ read
\begin{equation}
\nabla_{\ast\mu} T_{\ast}^{\mu\nu} = \alpha(\varphi) T_{\ast} \nabla^{\nu}_{\ast}\varphi\,,
\qquad
\nabla_{\mu} T^{\mu\nu} = 0\,.
\end{equation}
\subsection{Exact exterior neutron star spacetimes}
\label{sec:exact}
The pulse profile of a radiating neutron star depends on its
{\it vacuum exterior} spacetime. In general relativity, the exterior spacetime
of a static, spherically symmetric star is described by the Schwarzschild metric
(by Birkhoff's theorem), whose line element is
\begin{align}
{\rm d} s^2_{\rm Sch} &= -\, f_{\rm Sch} \; {\rm d} (ct)^2
+ f_{\rm Sch}^{-1} \; {\rm d} r^2 + r^2 {\rm d}\Omega\,
\label{eq:schw_metric}
\end{align}
in Schwarzschild coordinates $(t,r,\theta,\phi)$.
In this equation, ${\rm d}\Omega = {\rm d}\theta^2 + \sin^2\theta \; {\rm d} \phi$ is the line
element on a unit-sphere and the Schwarzschild factor is defined via
\begin{align}
f_{\rm Sch} \equiv 1 - 2 G m /(c^{2} r)\,,
\end{align}
with $G$ Newton's gravitational constant, and $m$ the gravitational mass of
the star. Thus, the metric is entirely determined by the mass of
the star.
In scalar-tensor gravity, the exterior spacetime of a static, spherically
symmetric star sourcing a nontrivial scalar field in the Einstein frame
is given by the Just spacetime~\cite{Buchdahl:1959nk,Just1959,Coquereaux:1990qs,Damour:1992we}:
\begin{equation}
\label{eq:just-metric}
{\rm d} s_{\ast}^2 =
- f^{b/a} {\rm d} (ct)^2
+ f^{-b/a} {\rm d} \rho^2
+ \rho^2 f^{1 - b / a} {\rm d} \Omega
\,
\end{equation}
in Just coordinates $(t,\rho,\theta,\phi)$~\cite{Just1959}.
In this equation, ${\rm d}\Omega$ is still the line element on an
unit-sphere, $b \equiv 2 G_{\ast} m / c^2$, but now
\begin{equation}
f \equiv 1 - \bar{a}\,.
\end{equation}
with $\bar{a} \equiv {a}/{\rho}$ and $a$ a real constant
with units of length that is related to both the mass of the star $m$
and the strength of the scalar field. The transformation relating
Just and Schwarzschild coordinates is given by~\cite{Damour:1992we}
\begin{equation}
r = \rho (1 - \bar{a})^{(1 - b/a)/2}\,.
\label{eq:sch_just_radiu_rel}
\end{equation}
but this expression is not analytically invertible. Therefore, the Just line
element in Just coordinates cannot in general be transformed to Schwarzschild
coordinates analytically, although it can be done numerically. The Just
line element in the Jordan frame is related with~\eqref{eq:just-metric} as
${\rm d} s^2 = A^2(\varphi) {\rm d} s^2_{\ast}$.
The scalar field configuration sourced by the star has the form
\begin{equation}
\varphi = \varphi_\infty + (q / a) \log(1 - \bar{a})\,.
\label{eq:scalar_exact}
\end{equation}
Far from the star ($\bar{a} \ll 1$), $\varphi \simeq \varphi_\infty - q / \rho$
and therefore, $q$ is the scalar charge of the star, while $\varphi_\infty$ is
the cosmological background value of $\varphi$, which we assume to be
zero\footnote{Rigorously, the present day value of $\varphi_\infty$ has to be
determined from the cosmological evolution of the
theory~\cite{Damour:1992kf,Damour:1993id,Anderson:2016aoi,Anderson:2017phb}.}.
We also assume that the conformal factor is such that $A(\varphi_\infty) = 1$ asymptotically.
As a consequence, the Einstein and Jordan-frame masses are the same
$m = m_{\ast}$~\cite{Pani:2014jra,Minamitsuji:2016hkk}. Additionally
the gravitational constant measured in the weak-field regime in a
Cavendish experiment is $G = A^2(\varphi_\infty)\, G_{\ast} [1 - \alpha(\varphi_\infty)^2]$.
The exterior spacetime of a neutron star in scalar-tensor theories is
therefore determined not just by the star's mass $m$ (or equivalently $b$),
but also by the parameters $q$ and $a$. These constants, however, are
not all independent. Instead, they obey the constraint
\begin{equation}
a^2 - b^2 - (2 q)^2 = 0\,.
\label{eq:constraint}
\end{equation}
We can use this relation to elucidate the meaning of the
ratio $a/b$ present in Eq.~\eqref{eq:just-metric}:
\begin{equation}
a/b = \sqrt{(2q/b)^2 + 1} = \sqrt{1 + Q^2}
\label{eq:ab_ratio}
\end{equation}
where we used $b = 2 G_{\ast} m / c^2$ to define
a dimensionless scalar charge-to-mass ratio
\begin{equation}
Q \equiv q c^2 / (G_{\ast} m).
\label{eq:def_Q}
\end{equation}
The exterior spacetime is therefore fully determined by the pair $(m,Q)$, with $Q$
controlling the departure from the Schwarzschild spacetime, which is recovered
when $Q = 0$ (and thus $a/b = 1$).
\subsection{Constraints on scalar-tensor theories and the parameter space of neutron star models}
In practice, few direct observational constraints on the scalar
charge-to-mass ratio $Q$ exist. In~\cite{Horbatsch:2011nh}, Horbatsch
and Burgess developed a conformal factor $A(\varphi)$ and
equation-of-state-independent framework which can be used
to directly constrain $Q$ if a sufficient number of
post-Keplerian parameters are known from a binary system.
When applied to the binaries PSR J0737-3039A/B and
PSR B1534+12 they obtained the upper bounds $|Q| = 0.21$ and $|Q| = 0.44$ (at 68$\%$
confidence level) respectively.
We use the latter value to guide an estimate of the largest magnitude of $Q$
allowed from observations; for concreteness we use $Q = 0.5$.
The parameter $Q$, is also constrained by
theoretical considerations. Clearly, for any
given equation of state, if the compactness is large enough, the
neutron star will collapse and form a black hole
(see~\cite{Palenzuela:2015ima,Mendes:2016fby} for studies in particular
scalar-tensor gravity models). An estimate of
this compactness was found by Buchdahl~\cite{Buchdahl:1959zz,Wald:1984rg}
in general relativity [$G m / (R c^2) < 4/9$].
Extending this work to scalar-tensor theories, Tsuchida et al.~\cite{Tsuchida:1998jw}
found that in the Einstein frame, $a$, $b$ and $q$ have their values bounded under
a minimal set of assumption akin to those of Buchdahl~\cite{Wald:1984rg}:
that the star is a perfect fluid, that the density is positive (and monotonically
decreasing) and that the solution matches the Just metric at the star's surface.
We can write this constraint on $Q$ differently by working directly
with the parameters $a$ and $b$ from Eq.~\eqref{eq:ab_ratio}. With this
reparametrization, the result from~\cite{Tsuchida:1998jw}
delimits a domain $\mathcal{D}$ (in a plane spanned by
${\bar a}_{\rm s} = a / \rho_{\rm s}$ and
${\bar b}_{\rm s} \equiv b / \rho_{\rm s}$, where $\rho_{\rm s}$ is the
star's radius in Just coordinates) given by:
\begin{align}
{\bar b}_{\rm s} \leq {\bar a}_{\rm s} \leq 2\sqrt{{\bar b}_{\rm s}} - {\bar b}_{\rm s}\,,
\,\,\, &\textrm{for} \,\,\, 0 \leq {\bar b}_{\rm s} \leq 4\left(3-2\sqrt{2}\right)
\nonumber \\
{\bar b}_{\rm s} \leq {\bar a}_{\rm s} \leq 2\left(\sqrt{2{\bar b}_{\rm s}} - {\bar b}_{\rm s}\right)\,,
\,\,\, &\textrm{for} \,\,\, 4\left(3-2\sqrt{2}\right) \leq {\bar b}_{\rm s} \leq 8/9
\nonumber \\
\textrm{no stars exist} \,, \,\,\, &\textrm{for} \,\,\, {\bar b}_{\rm s} > 8/9
\label{eq:tsuchida}
\end{align}
in which the theory [for any given $A(\varphi)$] admits stellar solutions.
We can thus use these inequalities as a guide to select neutron star models
in scalar-tensor gravity, parametrized by ${\bar a}_{\rm s}$,
${\bar b}_{\rm s}$ and $A_{\rm s}$,
{\it irrespective} of the equation of state of the stellar interior and
of the conformal factor $A(\varphi)$.
This makes our analysis {\it as model independent as possible}.
\begin{figure*}
\includegraphics[clip=true,width=0.95\textwidth]{existence_domain.pdf}
\caption{The existence domain $\mathcal{D}$ of stars
in scalar-tensor theories parameterized in terms of ${\bar a}_{\rm s}$
and ${\bar b}_{\rm s} [= 2 G_{\ast} m / (\rho_{\rm s} c^2)]$ (left panel)
and in terms of $Q$ and ${\bar b}_{\rm s}$ (right panel).
Stellar solutions exist within the shaded regions of both panels.
General relativistic stars lay on the
solid line ${\bar a}_{\rm s} = {\bar b}_{\rm s}$ and $Q = 0$, with
Newtonian configurations located near the origin of the plane in the left panel
and near the left end in the right panel.
Buchdahl's limit terminates these lines at the point $\mathcal{B} = (8/9, 8/9)$ in the left panel
and $\mathcal{B} = (1,8/9)$ in the right panel, while
black holes are located at the point $\mathcal{S} = (1, 1)$ in both panels.
Moving away from general relativistic line, we enter the realm of stars in scalar-tensor
gravity ($Q \neq 0$).
Their existence is bound from above by the inequalities in
Eq.~\eqref{eq:tsuchida} (dashed curve). The dashed-dotted line (labeled ``LR'')
correspond to the light ring location given by Eqs.~\eqref{eq:lr_bound}.
}
\label{fig:existence}
\end{figure*}
{}
The domain $\cal{D}$ is shown by the shaded region in Fig.~\ref{fig:existence}.
The left panel shows the inequalities in Eq.~\eqref{eq:tsuchida}, while in the
right panel we re-express them in terms of $a/b = {\bar a}_{s} / {\bar b}_{s}
= \sqrt{1 + Q^2}$ instead of ${\bar a}_{s}$.
For reference, we also shows lines of constant $Q$, with the solid
line corresponding to general relativity with $Q=0$.
Black holes are shown by the point labeled $\mathcal{S}$, while ``Newtonian''
solutions (with very small compactness) are at the origin of the plane (left-panel)
or toward the left-end (right-panel). In virtue of
the no-hair theorem of Sotiriou and Faraoni~\cite{Sotiriou:2011dz}, black holes have
zero scalar charge $q = 0$ (thus $Q = 0$, $a = b$) and compactness
$u / 2 = G_{\ast} m / (R c^2) = 0.5$
in theories described by the action in Eq.~\eqref{eq:action}.
The hatched portion of the plane {\it qualitatively} represents the region in
which neutron stars could exist. Assuming small deviations from general relativity
(i.e. small $Q$ as justified previously), we expect
the conformal factor evaluated at the star's surface to be of order unity
$A_{\rm s} \equiv A[\varphi(\rho_{\rm s)}] \approx 1$,
hence $R \approx \rho_{\rm s}$, where $R$ is the circumferential, Jordan-frame
radius of the star. Astronomical observations show that neutron stars have
masses approximately in the $m \in [0.9, 2.0]$ $M_{\odot}$ and radii
$R \in [8, 14]$ km ranges, which translate into
$u \approx b_{\rm s} \in [0.19, 0.74]$~\cite{Ozel:2016oaf,Miller:2016pom}.
Therefore, the region between the general relativistic solutions and the constant
$Q = 0.5$ line delimits the region where neutron stars could exist.
An additional region of interest arises from studying when the photon
light ring (or light rings) of the Just metric is inside the radius of the star.
This allows a photon emitted tangentially from the surface of the star
to reach the observer~\cite{Pechenick:1983}. Otherwise, if the light ring is outside of
the star, then tangential photons would come back to hit the surface, and thus, they
would not reach the observer. In Appendix \ref{app:lr}, we show that
this occurs when
\begin{equation}
{\bar a}_{\rm s} \leq 2(1 - {\bar b}_{\rm s})\,,
\quad \textrm{or} \quad
({\bar b}_{\rm s}/2) (2 + \sqrt{1+Q^2}) \leq 1\,,
\label{eq:lr_bound}
\end{equation}
which is shown by dash-dotted lines in Fig.~\ref{fig:existence}.
Due to the large parameter space available for stellar models in
scalar-tensor gravity we must make a few sensible requirements for choosing
our illustrative stellar models. All stars in our catalog:
(i) obey the constraints~\eqref{eq:tsuchida};
(ii) have a canonical mass of $m = 1.4$ $M_{\odot}$;
(iii) have twice the compactness of $u = \{1/3, 1/2\}$~\footnote{These are
illustrative values used by Poutanen et al.~\cite{Poutanen:2006hw}
which we will use to validate the numerical implementation of our pulse profile
code in Sec.~\ref{sec:results}}.
In this way, our stellar models in scalar-tensor gravity are
``doppelg\"angers'' of their general relativistic models counterparts:
they have the same gravitational mass $m$ and (Jordan-frame) areal radius
$R$ -- the only difference being the presence of a nonzero scalar charge.
Although we expect $A_{\rm s} \approx 1$, there is no particular
reason to chose it either larger or smaller than unity -- we thus
consider both possibilities.
Under these choices we group them in three classes according to their
value of $Q$.
\begin{itemize}
\item Models $a_{i}$ represent general relativistic stars ($Q=0$).
\item Models $b_i$ represent stars with $Q = 0.5$. To determine
$\rho_{\rm s}$ we assume that the conformal factor evaluated at
the surface $A_{\rm s}$ is $1.0 \pm 0.05$.
\item Models $c_i$ represent stars with $Q = 1.0$. To determine
$\rho_{\rm s}$ we assume that $A_{\rm s}$ is $1.0 \pm 0.1$.
\end{itemize}
The parameters of these stellar models are listed in Table~\ref{tab:stellar_models}.
Given the constraint on $\vert Q \vert$~\cite{Horbatsch:2011nh},
why should we consider values as large as unity? The reasons are twofold. First,
regardless of the constraint, it is of theoretical interest to investigate how
much (and precisely how) this new parameter affects the pulse profile. This question cannot
be answered if we restrict ourselves to $Q \approx 0$. Second, if it were the case
that even for a maximal value of $Q$ in the range of reasonable values (say even for $Q=1$),
the impact of the scalar charge on the pulse profile is minimal, then there would little motivation
to attempt to constraint this theory using pulse profile observations.
\begin{table}
\begin{tabular}{c c c c c c c c c}
\hline
\hline
Name & $u$ & $m/M_{\odot}$ & $R/{\rm km}$ & $\rho_{\rm s}/{\rm km}$
& $A_{\rm s}$ & ${\bar a}_{\rm s}$ & ${\bar b}_{\rm s}$ & $Q$ \\
\hline
$a_{1}$ & 0.333 & 1.4 & 12.40 & 12.40 & 1.00 & 0.333 & 0.333 & 0.0 \\
$a_{2}$ & 0.500 & 1.4 & 8.269 & 8.269 & 1.00 & 0.500 & 0.500 & 0.0 \\
\hline
$b_{1}$ & 0.333 & 1.4 & 12.40 & 13.35 & 0.95 & 0.346 & 0.309 & 0.5 \\
$b_{2}$ & 0.333 & 1.4 & 12.40 & 12.12 & 1.05 & 0.381 & 0.341 & 0.5 \\
$b_{3}$ & 0.500 & 1.4 & 8.269 & 9.040 & 0.95 & 0.511 & 0.457 & 0.5 \\
$b_{4}$ & 0.500 & 1.4 & 8.269 & 8.227 & 1.05 & 0.562 & 0.503 & 0.5 \\
\hline
$c_{1}$ & 0.333 & 1.4 & 12.40 & 14.83 & 0.90 & 0.394 & 0.279 & 1.0 \\
$c_{2}$ & 0.333 & 1.4 & 12.40 & 12.38 & 1.10 & 0.472 & 0.334 & 1.0 \\
$c_{3}$ & 0.500 & 1.4 & 8.269 & 10.37 & 0.90 & 0.564 & 0.398 & 1.0 \\
$c_{4}$ & 0.500 & 1.4 & 8.269 & 8.817 & 1.10 & 0.663 & 0.469 & 1.0 \\
\hline
\hline
\end{tabular}
\caption{{\it Stellar models.} We list the properties of the stellar
models which we used to compute the pulse profiles. The parameters are:
$u$ (twice the compactness),
$m$ (the gravitational mass), $R$ (Jordan-frame radius in
Schwarzschild coordinates), $\rho_{\rm s}$ (Einstein-frame radius in
Just coordinates), $A_{\rm s}$ (conformal factor evaluated at Einstein-frame
radius in Just coordinate $\rho_{\rm s}$),
$\bar{a}_{\rm s} = a / \rho_{\rm s}$,
$\bar{b}_{\rm s} = b / \rho_{\rm s}$,
$Q$ (scalar charge-to-mass ratio). The conformal factor $A_{\rm s}$
is necessary to translate radii between Einstein to Jordan frames.}
\label{tab:stellar_models}
\end{table}
\subsection{Connection to specific scalar-tensor gravity models}
Although our discussion so far has been fully model independent,
one can easily determine the parameters $(m, Q)$
within a specific scalar-tensor model. We here briefly explain how this
can be done once a specific functional form of $A(\varphi)$ is chosen.
The first step consists of integrating the stellar structure equations of a static,
spherically symmetric neutron star in scalar-tensor theory, which generalize
the Tolman-Oppenheimer-Volkoff equations from general relativity.
These equations are derived from the field equations
[Eqs.~\eqref{eq:b} and~\eqref{eq:a}] assuming matter to be
described by a perfect fluid.
The explicit form of these equations and a description of the
numerical algorithm used to integrate them can be found, e.g.
in~\cite{Damour:1993hw,Harada:1997mr,Silva:2014fca}.
Once a stellar model has been constructed, the
quantities $q$, $m$, $R$ and $A_{\rm s}$ are all known.
From $m$, we immediately determine $b$. Using Eq.~\eqref{eq:constraint}
we find $a$, while from Eq.~\eqref{eq:sch_just_radiu_rel} we obtain
$\rho_{\rm s}$. At last, $Q$ is obtained from Eq.~\eqref{eq:def_Q}.
With these values, the Just spacetime is completely determined.
\section{Pulse profile modeling}
\label{sec:pulseprofile}
Having presented an overview of scalar-tensor gravity, the properties
of the exterior neutron star spacetime and some generic properties of
neutron stars in this theory, we are now ready to develop a pulse profile model.
We start with a description of our assumptions, followed by an analysis
of geodesic motion in the Just spacetime and we close by constructing
the pulse profile model.
\subsection{Assumptions and model parameters}
Several models have been developed to study the pulse profile
of radiating, rotating neutron stars, since the pioneering
work by Pechenick et al.~\cite{Pechenick:1983}
(see~\cite{Poutanen:2008pg,Ozel:2012wu,Watts:2016uzu} for reviews).
Given that this is one of the first detailed study of pulse profiles in
non-GR theories, we will make a series of simplifying physical assumptions,
to be relaxed in future work. In this section, we lay down and
justify these assumptions.
\begin{itemize}
\item [(i)] {\it The stellar model, its exterior spacetime and photon
geodesics.} We assume a spherically symmetric neutron star
described by a solution of the field equations of scalar-tensor
theory with a perfect fluid.
The exterior spacetime (Sec.~\ref{sec:exact}) is described by two
parameters: $Q$ and $m$. The ratio of $a/b$
determines the scalar charge-to-mass ratio $Q$
[cf.~Eqs~\eqref{eq:ab_ratio} and~\eqref{eq:def_Q}].
Photons move on geodesics of the Just spacetime in the Jordan frame~\eqref{eq:just-metric}.
\item [(ii)] {\it Hot spot.} We assume that the star has a small (relative to
the size of the star), uniformly radiating, hot spot on its surface,
located at a polar angle $\theta_{\rm s}$.
Analytical estimates~\cite{Baubock:2015ixa} and numerical
simulations~\cite{Bai:2009wf} in general relativity provide evidence that the
hot spot size has a small impact on the pulse profile observed, provided the spot
is small enough.
In reality, the spot size is likely not sufficiently small for this effect to be
ignorable, but here, for simplicity, we will assume it to be infinitesimal in size and
that the rest of the star is dark, assumptions that can be lifted in the future.
We assume that the scalar field does not affect either the process of
generation of radiation nor its spectral properties.
\item [(iii)] {\it Rotation and special relativistic effects.} The star
rotates with a constant frequency $\nu$ and we neglect the
effects of frame-dragging on the emitted radiation. These effects
have been shown to be small (in general relativity)
and we expect them to be equally small in scalar-tensor
gravity~\cite{Sotani:2012eb}.
At high rotation frequencies ($\nu \gtrsim 300$ Hz), the pulse profile is
mostly affected by the quadrupolar deformation of the
star~\cite{Braje:2000qb,Cadeau:2004gm,Cadeau:2006dc,Morsink:2007tv,Psaltis:2013zja,Nattila:2017hdb}
which, however, we do not include.
However, we do include special relativistic effects of
Doppler boost and aberration due to the rapid motion of the hot spot.
That is, we work in a ``Just-Doppler'' approximation, in
analogy with the ``Schwarzschild plus Doppler'' approximation
introduced by Miller and Lamb~\cite{Miller:1997if}
who worked with the Schwarzschild spacetime
(see also Refs.~\cite{Poutanen:2003yd,Poutanen:2006hw}).
\item [(iv)] {\it Photon time delay.} Photons emitted at different rotation phases
$\phi_{\rm s} \equiv 2 \pi \nu t$ take different paths to reach the
observer, resulting in a time-delay $\Delta t$ (Sec.~\ref{sec:timedelay}). This
time-delay forces the observed phase of the pulse $\phi_{\rm obs}$ to be shifted
with respect to the star's rotational phase as
$\phi_{\rm obs} = \phi_{\rm s} + 2 \pi \nu \Delta t$.
This effect is only appreciable for fast rotators: for a rotational period
of $P = 1.5$ ms, this induces at most a $5$\% correction to the
photon arrival phase in general relativity~\cite{Poutanen:2006hw}.
We do include this effect for completeness.
\item [(v)] {\it Observer.} The observer is located a distance $D$ from the
star, at an inclination $\iota_{\rm o}$ relative to the star's rotation axis.
We express the physical observables in the Jordan frame, which are the
quantities measured by the observer's rods and clocks.
\end{itemize}
The geometry of the problem is depicted in Fig.~\ref{fig:diagram}.
The hot spot is located at the stellar surface at polar angle $\theta_{\rm s}$
and its instantaneous position is described by a radial unit vector ${\bm n}$.
The unit vector ${\bm n}$ and the line of sight make an angle $\psi$.
The trajectory of emitted photons can be described by a unit vector
${\bm k}_0$, which makes an angle $\alpha$ with respect to the normal ${\bm n}$.
We are interested in the photons whose trajectories (after
experiencing gravitational light bending) are along the line of sight of
the observer. These photon trajectories are depicted by the unit vector
${\bm k}$.
Finally, ${\bm \beta}$ is a vector, perpendicular to ${\bm n}$ and
tangential to the (instantaneous) direction in which the hot spot moves.
We denote by $\xi$ its angle with respect to ${\bm k}_0$.
When present, an antipodal hot spot is located at
$\theta = \pi - \theta_{\rm s}$ and
$\phi = \pi + \phi_{\rm s}$.
With this at hand, the plan for the remainder of this section is as follows:
(i) to study geodesic motion in the Just spacetime to obtain a relationship
between the angles $\psi$ and $\alpha$, which will tell us which photons
reach the observer, and (ii) to obtain relations for the angles
$\iota_{\rm o}$, $\theta_{\rm s}$, $\phi_{\rm s}$, $\alpha$ and $\xi$ which
we will use to construct the flux measured by the observer, including
special relativistic effects caused by the rapid motion of the hot spot.
\begin{figure}[t]
\includegraphics[width=0.45\textwidth]{pp_diagram_phase_alt.pdf}
\caption{Geometry of a rotating neutron star with hot spots.
The star is here depicted as rotating about the positive $z$ axis, with two
antipodal hot spots in red, and a variety of angles defined in the text.
The line of sight and the vector ${\bm k}$ lay on the same plane which
passes by the origin on the figure.
}
\label{fig:diagram}
\end{figure}
\subsection{Null geodesic motion in the Just spacetime}
\label{sec:geodesic}
To construct the bolometric flux measured by an observer, the first
step is to analyze the photon's geodesic motion in the Just metric.
In particular, the goal of this section is to obtain an expression
for the angle $\psi$, shown in Fig.~\ref{fig:diagram}.
We define the Lagrangian
\begin{equation}
L
= g_{\mu\nu} p^{\mu} p^{\nu}\,,
\end{equation}
where the photon's Jordan-frame four-momenta is defined as
$p^{\alpha} \equiv {\rm d} x^{\alpha} / {\rm d} \lambda$
[with $x^{\mu} = (ct,\rho,\theta,\phi)$ the four-trajectory of the photon],
where $\lambda$ is an affine parameter.
Because of spherical symmetry, the photon geodesic is confined
to a constant $\theta$-plane, which we take to be the equatorial plane
($\theta = \pi / 2$).
Using the line element in Eq.~\eqref{eq:just-metric}
to construct $L$, the equations of motion can be obtained
using the Euler-Lagrange equations and the null constraint
$g_{\mu\nu} p^{\mu} p^{\nu} = 0$:
\begin{subequations}
\begin{align}
{\rm d} t / {\rm d} \lambda &= A^{-2} \varepsilon f^{-b/a}\,,
\label{eq:tdot}
\\
({\rm d} \rho / {\rm d} \lambda)^2 &= A^{-4}
\left[c^2\varepsilon^2
- (h/\rho)^2 f^{2b/a - 1}\right]\,,
\label{eq:rhodot}
\\
{\rm d} \theta / {\rm d} \lambda &= 0\,,
\label{eq:thetadot}
\\
{\rm d} \psi / {\rm d} \lambda &= A^{-2} (h / \rho^2) f^{b/a - 1}\,.
\label{eq:psidot}
\end{align}
\end{subequations}
where $\varepsilon$ and $h$ are constants of motion,
associated with energy and angular momentum respectively.
Combining Eqs.~\eqref{eq:tdot} and~\eqref{eq:rhodot} we obtain:
\begin{equation}
{\rm d}\psi / {\rm d} \rho
= \rho^{-2} f^{b/a - 1}
\left[\sigma^{-2} - \rho^{-2} f^{2b/a - 1} \right]^{-1/2}
\end{equation}
where $\sigma \equiv h / (\varepsilon c)$
is the impact parameter.
Let us now solve for the angle $\psi$ in integral form.
The impact parameter can be eliminated in
favor of $\alpha$ by noticing that the angle between $p^{\psi}$ and $p^{\,\rho}$
at $\rho = \rho_{\rm s}$ is $\tan \alpha = [p^{\psi} p_{\psi} /
(p^{\,\rho} p_{\rho})]^{1/2}$~\cite{Beloborodov:2002mr} (see Fig.~\ref{fig:diagram_planar}).
Using Eqs.~\eqref{eq:rhodot} and~\eqref{eq:psidot}, we find
\begin{equation}
\sin \alpha = (\sigma / \rho_{\rm s}) (1 - {\bar a}_{\rm s})^{b/a - 1/2}\,,
\label{eq:angle_alpha}
\end{equation}
which in turn results in
\begin{align}
\psi &= \sin\alpha \int_0^1 {\rm d} y \, (1 - {\bar a}_{\rm s} y)^{b/a - 1}
\nonumber \\
&\quad \times
\left[(1 - {\bar a}_{\rm s})^{2b/a - 1}
- y^2 \sin^2\alpha \; (1 - {\bar a}_{\rm s} y)^{2b/a - 1}
\right]^{-1/2}\,.
\end{align}
where we have defined $y \equiv \rho_{\rm s} / \rho$.
\begin{figure}[t]
\includegraphics[width=0.45\textwidth]{pp_diagram_planar.pdf}
\caption{Bird's eye view of the rotating neutron star. The photon is
emitted along a unit vector ${\bm k}_0$ with an angle $\alpha$ with
respect to a normal unit vector ${\bm n}$, from a
small hot spot located at $\psi$ with respect to the line of sight.
The star's gravitational field changes the photon trajectory by a
bending angle $\beta \equiv \psi - \alpha$, seen to arrive
with an impact parameter $\sigma$ along the unit vector ${\bm k}$.}
\label{fig:diagram_planar}
\end{figure}
This integral has a singularity whenever $\sin \alpha = 1$
and $y \to 1$. The singularity is power-law integrable and
can be removed introducing the new variable
$x = \sqrt{1 - y}$~\cite{Press:2007:NRE:1403886,Lo:2013ava}.
In terms of $x$, we obtain
\begin{align}
\psi &= 2\sin\alpha \int_0^1 {\rm d} x\,
x\, [1 - {\bar a}_{\rm s}(1-x^2)]^{b/a - 1}
\nonumber \\
&\quad \times
\left\{
(1 - {\bar a}_{\rm s})^{2b/a - 1}
\right.
\nonumber \\
&\quad \left.
-\, (1-x^2)^2 \left[1 - {\bar a}_{\rm s}(1 - x^2) \right]^{2b/a - 1}
\sin^2\alpha \right\}^{-1/2}\,.
\label{eq:psi_final}
\end{align}
In the general relativity limit ($a/b = 1$), this result agrees
with the literature~\cite{Poutanen:2003yd,Poutanen:2006hw,Lo:2013ava}.
In the Newtonian limit ($a_{\rm s} = 0$), $\psi = \alpha$
and the bending angle $\beta \equiv \psi - \alpha$
(cf. Fig.~\ref{fig:diagram_planar}) is zero.
Figure~\ref{fig:cospsi} shows $\psi$ calculated in the range
$\alpha \in [0, \pi/2]$, for the stellar models in Table~\ref{tab:stellar_models}.
In the general relativistic limit ($Q=0$ or $a/b=1$) our results agree with~\cite{Poutanen:2006hw}.
A nonzero scalar charge has the effect of increasing/decreasing $\cos\psi$ relative to general relativity
depending on whether the conformal factor at the stellar surface is smaller/larger than unity.
This enhancement/suppression is a manifestation of spacetime becoming compressed/stretched in the star's vicinity in scalar-tensor gravity because of the conformal factor.
This effect is more salient as $\alpha \to \pi/2$ and is virtually
negligible for small $\alpha$.
\subsection{Visible fraction of the star and visibility conditions}
The light ray that defines the star's visible area
is the one emitted at an angle $\alpha = \pi / 2$ (i.e.
tangent to a vector normal to the surface) from a
position specified by an angle $\psi$, measured
with respect to the line of sight of the observer (cf. Fig.~\ref{fig:diagram_planar}).
In flat spacetime, this means that the visibility condition is
\begin{equation}
\cos\psi > 0. \qquad \textrm{flat spacetime}
\end{equation}
Due to light bending, however, we can see more than half
of a spherically symmetric neutron star, i.e.~a photon emitted with
$\cos\psi < 0$ can reach the observer at spatial infinity.
However, there is a critical value
\begin{equation}
\cos\psi_{\rm c} < 0, \qquad \textrm{curved spacetime}
\label{eq:visibility}
\end{equation}
which determines the last visible ring of the star. The hot spot
is visible if
\begin{equation}
\cos\psi > \cos\psi_{\rm c}. \qquad \textrm{visibility condition}
\end{equation}
This value can be calculated numerically using Eq.~\eqref{eq:psi_final}
with $\alpha = \pi/2$, i.e.
$\psi_{\rm c} \equiv \psi(\alpha = \pi/2)$.
We define the visible fraction of the surface as the ratio:
\begin{align}
\delta f &\equiv
\frac{A_{\rm s}^2\, \rho_{\rm s}^2 (1 - {\bar a}_{\rm s})^{1-b/a}\int_0^{\psi_{\rm c}}\int_0^{2\pi}
{\rm d}\psi'\, {\rm d}\phi \sin\psi' }{4\pi A_{\rm s}^2\,\rho_{\rm s}^2 (1 - {\bar a}_{\rm s})^{1-b/a}}
\nonumber \\
&=
\frac{1-\cos\psi_{\rm c}}{2}.
\end{align}
In the Newtonian limit, $\psi_{\rm c} = \pi/2$ and
$\delta f = 1/2$, i.e. half of the star is visible
as expected.
For relativistic stars, this number increases since
$\cos\psi_{\rm c} < 0$ in general,
as shown in Fig.~\ref{fig:cospsi}. As we have seen in scalar-tensor gravity, this
effect can become larger or smaller in comparison to general relativity depending
on whether the conformal factor at the stellar surface is smaller or larger than unity. For
instance, when $u = 0.5$, $\delta f \approx 0.94$ in general relativity (model $a_{2}$ in Table~\ref{tab:stellar_models}),
while it becomes $\delta f \approx 0.85,\, 0.95$ in scalar-tensor gravity for $A_{s} = 0.9,\, 1.1$
and large $Q=1$ (models $c_3$ and $c_4$ in Table~\ref{tab:stellar_models}).
The above considerations implicitly assume that the light ring
of the star is inside the stellar radius. When this is the case,
tangential photons (ie.~those emitted with $\alpha = \pi/2$) do escape
to spatial infinity. However, if the light ring is outside of the stellar surface,
then tangential photons will not escape to spatial infinity. Instead,
there will be some maximum $\alpha$ for which photons do reach the observer,
and this maximum angle will now need to be used to define the visible fraction
of the star. In this paper, however, for all the stellar models we consider,
the light ring is inside of the stellar surface, and thus, the analysis of
the visible fraction presented above apply.
\begin{figure}[t]
\includegraphics[width=0.48\textwidth]{fig_cospsi_cosalpha.pdf}
\qquad
\caption{Effect of the scalar charge on $\psi$ as a function of $\alpha$.
The scalar charge affects $\psi$ mostly at $\alpha \approx \pi/2$, while
for $\alpha \approx 0$ its influence is negligible.
}
\label{fig:cospsi}
\end{figure}
\subsection{Photon time-delay}
\label{sec:timedelay}
Just as curvature deflects photons around the star, it also slows them down
through the Shapiro time delay effect. The magnitude of this effect can
be calculated using Eqs.~\eqref{eq:tdot} and~\eqref{eq:rhodot}, which gives
\begin{equation}
t = (1/c)\int^{\infty}_{\rho_{\rm s}}
{\rm d}\rho\,
f^{-b/a}\left[1 - (\sigma/\rho)^2 f^{2b/a - 1} \right]^{-1/2}\,,
\end{equation}
and from which we define the time delay as
\begin{equation}
\Delta t \equiv t(\sigma) - t(\sigma=0),
\label{eq:def_deltat}
\end{equation}
defined with respect to a photon emitted directly towards the observer
(i.e. zero impact parameter $\sigma = 0$).
As in the case for $\psi$, we can write the time delay integral
in terms of the emission angle $\sin\alpha$ and introduce the same
integration variables $\rho_{\rm s}/\rho = y = 1 - x^2$, with the
latter to remove (again) a singularity.
The final expression becomes:
\begin{align}
\Delta t &=({2\rho_{\rm s}}/{c})
\int^{1}_0 {\rm d} x \, x \,[1 - {\bar a}_{\rm s}(1- x^2) ]^{-b/a}\,(1-x^2)^{-2}
\nonumber \\
&\quad \times \left\llbracket
\left\{
1 - (1-x^2)^2 (1- {\bar a}_{\rm s} )^{1-2b/a}
\right. \right.
\nonumber \\
&\quad \left. \left. \times \left[ 1 - {\bar a}_{\rm s} (1-x^2)\right]^{2b/a - 1} \sin^2\alpha
\right\}^{-1/2}
- 1 \right\rrbracket\,,
\label{eq:time_delay_final}
\end{align}
which reduces to the general relativistic expression when
$a/b = 1$ or $Q=0$~\cite{Lo:2013ava}.
Figure~\ref{fig:timedelay} shows the time-delay $\Delta t$
as a function of angle $\psi$. Our results agree with~\cite{Poutanen:2006hw}
in the limit $a/b = 1$ (or $Q=0$). In scalar-tensor theory,
this delay is weakly dependent on the mass, as also observed in general relativity by~\cite{Poutanen:2006hw},
having a maximum value of $\Delta t \approx 0.07$ ms irrespective of the
$Q$. The deviations from general relativity increase for larger values of $Q$.
Similarly to the calculation of $\psi$, this increase is largest when
$\psi \to \pi / 2$.
\begin{figure}[t]
\includegraphics[width=0.48\textwidth]{fig_timedelay.pdf}
\caption{Effect of the scalar charge on the time-delay of
photons as they escape the neutron star. As in the case of the angle $\psi$, the effect is
largest when $\alpha$ approaches $\pi / 2$, while it is negligible as
$a \to 0$.
}
\label{fig:timedelay}
\end{figure}
\subsection{The bolometric flux in scalar-tensor gravity}
\label{sec:boloflux}
We now combine the results from the previous sections and
construct the pulse profile equation in the Just spacetime,
following closely~\cite{Poutanen:2003yd,Poutanen:2006hw,Lo:2013ava,Lo:2018hes}.
Before jumping into the calculation of the bolometric flux, let us
define some more geometrical quantities that will be useful in the calculation.
Consider the radial unit vector ${\bm n}$ at the location of the
hot spot and the unit vector ${\bm k}$ along the line of sight (cf. Fig~\ref{fig:diagram}).
Since the angle $\psi$ is defined between ${\bm n}$ and the line of sight, we have
\begin{equation}
\cos\psi = {\bm k} \cdot {\bm n}\,.
\end{equation}
Due to the rotation, $({\bm k} \cdot {\bm n})$ varies periodically
\begin{equation}
\cos\psi = \cos\iota_{\rm o} \cos\theta_{\rm s}
+ \sin\iota_{\rm o}\sin\theta_{\rm s} \cos\phi_{\rm s}\,,
\label{eq:cospsi}
\end{equation}
where $\iota_{\rm o}$ is the angle between the line of
sight with respect to the spin axis, $\theta_{\rm s}$
is the spot's colatitude and $\phi_{\rm s} = 2 \pi \nu t$
is the rotational phase of the spot. We choose $t=0$ to correspond
to when the spot is closest to the observer.
The angle $\psi$ measures the {\it apparent} inclination
of the spot relative to the line of sight, but due to light-bending
this is different from the true inclination. To define the latter, recall that
${\bm k}_0$ is the unit vector in the direction in which the photon is emitted,
such that the emission angle $\alpha$ is then
\begin{equation}
\cos\alpha = {\bm k}_0 \cdot {\bm n}\,.
\end{equation}
Due to special relativistic aberration, an observer comoving with
the hot spot measures an angle $\alpha'$. The angles $\alpha$ and $\alpha'$
are related by~\cite{Poutanen:2003yd}
\begin{equation}
\cos\alpha' = \delta\, \cos\alpha\,,
\label{eq:trans_alpha}
\end{equation}
where $\delta$ is the Doppler factor
\begin{equation}
\delta = [\gamma(\theta_{\rm s}) (1 - \beta \cos\xi)]^{-1}\,.
\label{eq:doppler}
\end{equation}
Here, $\gamma(\theta_{\rm s}) = (1 - \beta^2)^{-1/2}$ is the Lorentz factor and
$\beta = v / c$ is the spot's velocity, which satisfies
\begin{equation}
\beta = \frac{2 \pi \rho_{\rm s}}{c}
\frac{\nu}{(1 - {\bar a}_{\rm s})^{b/(2a)}}
\sin\theta_{\rm s}\,.
\end{equation}
where we corrected the rotation frequency by the redshift factor
\begin{equation}
\nu / \nu_0 = \sqrt{g_{tt}(\rho_{\rm s}) / g_{tt}(\infty)}
= A_{\rm s} (1 - {\bar a}_{\rm s})^{b/(2a)}\,.
\end{equation}
The angle between ${\bm k}_0$ and the spot velocity vector ${\bm \beta}$
is denoted by $\xi$ (see Fig.~\ref{fig:diagram}), and it is related to the other angles
via~\cite{Poutanen:2006hw}:
\begin{equation}
\cos \xi =
({\bm \beta} \cdot {\bm k}_0) / \beta
= - (\sin\alpha / \sin\psi_{\rm s}) \sin\iota_{\rm o} \sin\phi_{\rm s}
\end{equation}
With this at hand, let us now return to the observed flux from the spot at energy $E$.
This quantity can be defined as
\begin{equation}
{\rm d} F_{E} = I(E,\alpha) {\rm d} \Omega\,,
\label{eq:flux_general}
\end{equation}
where $I(E,\alpha)$ is the specific intensity of radiation at
infinity and ${\rm d} \Omega$ is the solid angle on the observer's sky
occupied by the spot with differential area ${\rm d} S'$ measured
in a comoving reference frame.
Let us determine $I(E,\alpha)$ and ${\rm d} \Omega$
separately, starting with the latter.
The solid angle can be written in terms of the impact parameter
$\sigma$ as
\begin{equation}
{\rm d} \Omega = (\sigma\, {\rm d} \sigma\, {\rm d} \varphi)/D^2\,,
\end{equation}
where $D$ is the distance to the source and $\varphi$ is the azimuthal
angle around the vector ${\bm k}$ (not to be confused with the
scalar field). In terms of the surface area [cf. Eq.~\eqref{eq:just-metric}]
\begin{equation}
{\rm d} A =
A_{\rm s}^2\rho_{\rm s}^2 (1 - {\bar a}_{\rm s})^{1-b/a} \sin\psi {\rm d}\psi {\rm d}\varphi\,,
\end{equation}
we have
\begin{equation}
{\rm d} \Omega = \frac{\sigma}{A_{\rm s}^2\rho_{\rm s}^2}
\frac{(1 - {\bar a}_{\rm s})^{b/a - 1}}{\sin\psi}
\frac{{\rm d} \sigma}{{\rm d} \psi} \frac{{\rm d} A}{D^2}\,.
\end{equation}
Using that the spot area projected onto the plane perpendicular to the
photon propagation direction is a Lorentz invariant, we have
\begin{equation}
{\rm d} A' \cos\alpha' = {\rm d} A \cos\alpha\,,
\end{equation}
where ${\rm d} A'$ (${\rm d} A$) is the differential spot area measured by a comoving
(static) reference frame.
We can now write the solid angle as
\begin{equation}
{\rm d} \Omega = \frac{\sigma}{A_{\rm s}^2 \rho_{\rm s}^2}\frac{(1 - {\bar a}_{\rm s})^{b/a -1}}{\sin\psi} \frac{{\rm d} \sigma}{{\rm d} \psi}
\frac{\cos\alpha'}{\cos\alpha} \frac{{\rm d} A'}{D^2}\,,
\label{eq:solid_angle}
\end{equation}
where $dA'$ [written $(\theta,\phi)$ coordinates] is
\begin{equation}
{\rm d} A' =
\gamma(\theta) A_{\rm s}^2\rho_{\rm s}^2 (1 - {\bar a}_{\rm s})^{1-b/a} \sin\theta {\rm d}\theta {\rm d}\phi\,,
\label{eq:area_comoving}
\end{equation}
where we took into account a factor of $\gamma(\theta)$ relating ${\rm d} A$
and ${\rm d} A'$. This term is due to Lorentz contraction of the linear
differential interval correspondent to an angular interval ${\rm d} \phi$ measured
in the static reference frame~\cite{Lo:2018hes,Nattila:2017hdb}.
Finally, using Eq.~\eqref{eq:angle_alpha} for the impact parameter
and remembering that $\alpha$ depends implicitly on $\psi$
[cf. Eq.~\eqref{eq:psi_final}] we find:
\begin{equation}
{\rm d} \sigma / {\rm d} \psi = \rho_{\rm s} (1 - {\bar a}_{\rm s})^{1/2 - b/a}
({\rm d} \sin\alpha / {\rm d} \psi)\,,
\end{equation}
which substituted back into Eq.~\eqref{eq:solid_angle} gives our final
expression for ${\rm d} \Omega$:
\begin{equation}
{\rm d} \Omega = (1 - {\bar a}_{\rm s})^{-b/a}
\cos\alpha'
\frac{{\rm d} \cos\alpha}{{\rm d} \cos\psi}
\frac{{\rm d} A'}{A_{\rm s}^2 D^2} \,.
\label{eq:solid_angle_final}
\end{equation}
Now we turn our attention to the specific intensity.
We assume it can be decomposed (in the comoving reference frame)
as~\cite{Lo:2013ava}
\begin{equation}
I'_0(E'_0,\alpha') = g'(\alpha') f'(E'_0)\,,
\end{equation}
where $g'(\alpha')$ and $f'(E'_0)$ are respectively a beaming and
a spectral function, and $E'_0$ is the photon energy measured by
the comoving observer at the stellar surface.
In general, the specific intensity has a dependence on $\alpha'$ due
to Thompson scattering as the photon propagates through
the star's atmosphere~\cite{ChandraRTBook,Lo:2013ava}.
In order to express the specific intensity in terms of quantities
measured by an observer far away from the star, we proceed in two
steps. First, we Lorentz transform the energy $E'_0$ measured by
a comoving reference frame at the stellar surface to the energy
$E_0$ measured by a static reference frame near the stellar surface via
\begin{equation}
E_0 = \delta \, E'_0\,.
\label{eq:trans_mov_static}
\end{equation}
The ratio ($\textrm{specific intensity} / \textrm{energy}^3$)
is a Lorentz invariant allowing us to write
\begin{equation}
I_0(E_0, \alpha) / E^3_0
= I'_0(E'_0, \alpha') / {E'}^{3}_0
\,.
\end{equation}
We now transform $E_0$ to the energy $E$ measured by the
distant observer. These energies are related by a gravitational redshift
factor
\begin{equation}
E = A_{\rm s} (1 - {\bar a}_{\rm s})^{b/(2a)} E_0\,.
\label{eq:trans_redshift}
\end{equation}
Consequently, our final result becomes:
\begin{equation}
I(E,\alpha) = A_{\rm s}^3 \delta^3
(1 - {\bar a}_{\rm s})^{3b/(2a)}
I'_0(E'_0, \alpha')\,,
\label{eq:intens_transform}
\end{equation}
which relates the specific intensity as measured by a comoving
observer at the stellar surface to that measured by a distant observer,
including both special and general relativistic effects.
Substituting Eqs.~\eqref{eq:solid_angle_final} and~\eqref{eq:intens_transform}
into Eq.~\eqref{eq:flux_general} we obtain
\begin{equation}
{\rm d} F_E = A_{\rm s}(1 - {\bar a}_{\rm s} )^{b/(2a)} \delta^3
I'_0(E'_0,\alpha') \cos\alpha'
\frac{{\rm d} \cos\alpha}{{\rm d} \cos \psi}
\frac{{\rm d} A'}{D^2}\,.
\label{eq:spectral_flux}
\end{equation}
To obtain the bolometric flux, we integrate Eq.~\eqref{eq:spectral_flux}
over energies $E$. Using one last time the relation between $E$ and
$E'_0$ [i.e. using Eqs.~\eqref{eq:trans_mov_static}
and~\eqref{eq:trans_redshift}] we find
\begin{align}
{\rm d} F
&\equiv \int {\rm d} E\, {\rm d} F_{E}\,,
\nonumber \\
&=
A_{\rm s}^2
(1 - {\bar a}_{\rm s})^{b/a}\, \delta^4\, \cos\alpha' \frac{{\rm d} \cos\alpha}{{\rm d} \psi} \frac{{\rm d} A'}{D^2}
\int {\rm d} E'_0 \,I'_0(E'_0,\alpha')\,,
\nonumber \\
&=
A_{\rm s}^2
(1 - {\bar a}_{\rm s})^{b/a}\, \delta^4\, \cos\alpha'
\frac{{\rm d} \cos\alpha}{{\rm d} \cos\psi} \frac{{\rm d} A'}{D^2}I'_0(\alpha')\,.
\end{align}
As a final step, we can reexpress the angle $\alpha'$ (measured by a
comoving reference frame at the stellar surface) in terms of $\alpha$
(measured by a reference frame static near the star) using
Eq.~\eqref{eq:trans_alpha}.
Our final expression for the bolometric flux measured by an observer
due to radiation emission by a small element ${\rm d} S'$ of the hot spot
is
\begin{equation}
{\rm d} F =
A_{\rm s}^2 (1 - {\bar a}_{\rm s})^{b/a}\,
\delta^5\, \cos\alpha \frac{{\rm d} \cos\alpha}{{\rm d} \cos\psi}
\frac{{\rm d} A'}{D^2}I'_0(\alpha')\,.
\label{eq:main_result}
\end{equation}
This formula is the main result of this paper.
It is sufficiently general to describe the Jordan-frame bolometric flux of a
static, spherically symmetric star whose exterior spacetime is described by the
Just metric, including special relativistic effects (aberration and Doppler boosts)
and gravitational redshift.
For simplicity, let us assume that the specific intensity is isotropic
(independent of $\alpha'$) and that the hot spot is small in size.
Writing $I_0(\alpha') = I'_0$, normalizing
the flux ${\rm d} F$ by $I'_0 {\rm d} A' / D^2$
we have\footnote{
For hot spots of large angular radius, the flux must be calculated by a discretization of
the area occupied by the spot on a grid ($\theta_i, \phi_i$), with each cell
having a corresponding Lorentz factor $\gamma(\theta_i)$, see
Eq.~\eqref{eq:area_comoving}~\cite{Lo:2013ava,Lo:2018hes}}.
\begin{equation}
F=
A_{\rm s}^2 (1 - {\bar a}_{\rm s})^{b/a}\,
\delta^5\, \cos\alpha\,
\frac{{\rm d} \cos\alpha}{{\rm d} \cos\psi}\,,
\label{eq:flux_normalized}
\end{equation}
which is our final result, used in Sec.~\ref{sec:results}.
When two antipodal hot spots are present, the total flux
is obtained as (see Fig.~\ref{fig:diagram})
\begin{equation}
F = F(\iota_{\rm o}, \pi - \theta_{\rm s}, \pi + \phi_{\rm s}, {\bar a}_{\rm s},
{\bar b}_{\rm s})
+ F(\iota_{\rm o}, \theta_{\rm s}, \phi_{\rm s}, {\bar a}_{\rm s},
{\bar b}_{\rm s})\,.
\label{eq:flux_normalized_antipo}
\end{equation}
The hot spot is visible for the distant observer when
$\cos\psi > \cos\psi_{\rm c}$. For the antipodal hot spot the visibility
condition is when $\cos\psi > - \cos\psi_{\rm c}$~\cite{Poutanen:2006hw}.
Let us comment on some of the features of Eq.~\eqref{eq:flux_normalized}.
First, the general form of the expression is analogous to the one found
in~\cite{Poutanen:2006hw} and agrees with it in the general relativistic limit
$a/b = 1$. The prefactor $1 - {\bar a}_{\rm s}$
reduces the peak-to-peak amplitude of the pulse's profile in comparison
to the Newtonian result (in which ${\bar a}_{\rm s} \to 0$).
Second, the Doppler factor $\delta^5$ oscillates with time, skewing the
otherwise sinusoidal pulse profile.
\section{Numerical Modeling and Pulse Profiles in Scalar-Tensor Theory}
\label{sec:results}
Having determined the expression for the radiation flux, let us
outline how the pulse profiles are numerically calculated. The steps are:
\begin{itemize}
\item [(i)] From Eq.~\eqref{eq:cospsi} we determine $\psi$ at a
given phase $\phi_{\rm s}$.
\item [(ii)] With $\psi$ at hand, we check whether or not the spot is visible. If
$\cos\psi > \cos\psi_{\rm c}$, the spot is visible and we proceed with the remaining steps.
Otherwise, the flux is zero.
\item [(iii)] With $\psi$ at hand, we then determine $\alpha$, inverting
Eq.~\eqref{eq:psi_final}. This is done numerically with the shooting method.
\item [(iv)] Knowing $\alpha$, we can calculate the derivative
$({\rm d} \cos \alpha / {\rm d} \cos\psi)$ for the current value of $\psi$.
In practice, it is easier to determine
$({\rm d} \cos \psi / {\rm d} \cos\alpha)^{-1}$ and evaluate it at the value of $\alpha$ found
in step (iii). To do so, we first calculate $\cos\psi$
on a fine grid $\cos\alpha \in [0, 1]$, interpolate the data using a spline
interpolation, and from this we calculate the derivative numerically.
\item [(v)] We calculate the Lorentz and Doppler factors in Eq.~\eqref{eq:doppler},
and the time-delay in Eq.~\eqref{eq:time_delay_final}.
\item [(vi)] We combine all these ingredients in Eq.~\eqref{eq:flux_normalized} to obtain
the flux. Finally, we correct the phase due to time-delay by adding $(2\pi\nu\Delta t)$
to $\phi_{\rm s}$.
\end{itemize}
Let us now show some numerical results for one hot spot with our catalog of stellar
models (see Table~\ref{tab:stellar_models}). As described in Sec.~\ref{sec:intro},
these profiles could represent observations from burst oscillations of an
accreting neutron star.
We consider two illustrative situations, one in
which $\theta_{\rm s} = \iota_{\rm o} = 45^{\circ}$ (as in~\cite{Poutanen:2006hw}),
and another with $\theta_{\rm s} = \iota_{\rm o} = 90^{\circ}$
(as in~\cite{Lo:2013ava}) to illustrate the general effects of a nonzero scalar
charge on the pulse profile. In the former case, our results agree with those presented
in~\cite{Poutanen:2006hw} in the limit of general relativity (modulo a different normalization).
To make the effects of scalar-tensor gravity more clearly visible, we consider
extreme stellar models, with $u = 0.5$ and rotation frequency
of $\nu = 600$ Hz. With these choices the strong-gravity effects on the waveform
become more pronounced in both figures.
To avoid excessive clutter in the panels and complement the sample
pulse profile shown in Fig.~\ref{fig:pp_example} we use only the models with
$A_{\rm s} > 1$ (i.e. $b_4$ and $c_4$). The waveforms for
$b_3$ and $c_3$ are similar to that shown in Fig.~\ref{fig:pp_example}.
same compactness.
\begin{figure*}[htb]
\includegraphics[width=\textwidth]{fig_pulseprofiles_4545.pdf}
\caption{We illustrate the effects of a nonzero scalar charge on the
pulse profile for one hot spot.
We choose $\theta_{\rm s} = \iota_{\rm o} = 45^{\circ}$ and stars
with $u = 0.5$ and increasing values of scalar charge-to-mass ration
$Q$ (models $a_2$, $b_4$ and $c_4$ in Table~\ref{tab:stellar_models}, where recall
$a_{2}$ is the general relativistic limit).
We show slowly-rotating stars ($\delta = 1$, $\Delta t = 0$) in the top-left
panel, rapidly rotating stars neglecting time-delay effects ($\delta \neq 1$, $\Delta t = 0$)
in the top-right panel, and including them ($\delta \neq 1$, $\Delta t \neq 0$)
in the bottom-left panel. In the last panel (bottom-right), we compare the
pulse profile in these three situations for a star with $Q = 1$ and $u = 0.5$
(model $c_4$ in Table~\ref{tab:stellar_models}).
}
\label{fig:illustrative_45}
\end{figure*}
Figure~\ref{fig:illustrative_45} studies the
$\theta_{\rm s} = \iota_{\rm o} = 45^{\circ}$ case, showing the pulse profile
in a number of special cases. The top-left panel considers slowly-rotating stars,
whose pulse profile is calculated setting the Doppler and Lorentz factors
to unity and neglecting the travel time delay of photons
[cf. Eq.~\eqref{eq:flux_normalized}].
The top-right (bottom-left) panel shows how the pulse profile changes when we
include Doppler effects (and time-delay). The Doppler factor changes the amplitude of
the waveform and its overall shape by skewing it. A time-delay has the effect of
shifting the arrival time of the pulse (nonuniformly over the course of a revolution),
resulting in an additional small deformation of the waveform.
Overall, the impact of each of these effects in scalar-tensor gravity is identical to
that in general relativity~\cite{Poutanen:2006hw}, the difference being
the magnitude of these effects. This is not surprising given that the flux
formula in Eq.~\eqref{eq:flux_normalized} has the same functional form in scalar-tensor
gravity as in general relativity.
Figure~\ref{fig:illustrative_90} studies the
$\theta_{\rm s} = \iota_{\rm o} = 90^{\circ}$ case.
The same conclusions drawn from Fig.~\ref{fig:illustrative_45} are applicable here.
Because of the geometrical arrangement of the hot spot's location and the line
of sight of the observer, the hot spot becomes invisible to the observer during
certain phase intervals. These intervals when the hot spot is not visible depend
on the values of $Q$ in scalar-tensor theory, and thus, they could provide yet another
telltale sign of a deviation from general relativity. Of course, other stellar
parameters, like the inclination angle, also affect the occultation period; thus,
a full covariance analysis is necessary to determine the detectability of this effect.
\begin{figure*}[htb]
\includegraphics[width=\textwidth]{fig_pulseprofiles_9090.pdf}
\caption{Similar to Fig.~\ref{fig:illustrative_45}, but for
$\theta_{\rm s} = \iota_{\rm o} = 90^{\circ}$. Due to the geometrical configuration
of the hot spot and the observer, the flux hot spot becomes invisible momentarily
as the star rotates. Observe that where the flux disappears depends on the value
of $Q$.}
\label{fig:illustrative_90}
\end{figure*}
\section{Conclusions and outlook}
\label{sec:conclusion}
We introduced a Just-Doppler approximation for calculating
pulse profiles of rotating neutron stars in scalar-tensor
gravity. The main result, encapsulated in Eq.~\eqref{eq:main_result},
allows one to calculate the waveforms, including effects of the strong-gravity
and special relativistic effects, generalizing the Schwarzschild plus Doppler
approximation to scalar-tensor theories of gravity. We presented a selection of
sample results for burst oscillation waveforms of an infinitesimal hot spot and
discussed the implications of the presence of a nonzero scalar charge on it.
The model independent character of the formalism opens the possibility
of constraining a large class of scalar-tensor gravity models with upcoming x-ray timing
data releases from NICER. In this regard, our work is close in spirit
to~\cite{Horbatsch:2011nh}, except that that work focused on binary pulsars. In
a forthcoming paper, we will present a statistical likelihood analysis
discussing the strength of potential future constraints on scalar-tensor gravity with
pulse profile observations. We emphasize, however, that whether future constraints
can be placed on scalar-tensor gravity with NICER data will require a full data analysis
study that varies over all model parameters in the presence of noise; such a study
is now possible thanks to the pulse-profile model presented here, and it will be carried
out in the future.
Despite the generality of the present formalism, it is necessary
to discuss its limitations and signal in which directions it
can be improved further.
At considerably large rotational frequencies ($\nu \gtrsim 300$ Hz),
the Schwarzschild-Doppler approximation becomes inappropriate
as discussed, e.g.~in~\cite{Cadeau:2006dc}, because of the rotation-induced
quadrupolar deformation of star. To include this effect in the pulse profiles, one
must first extend the Just metric to rotating stars.
To leading-order in rotation, that is, including only frame-dragging effects
while keeping a spherical geometry for the star, this calculation was done
in~\cite{Damour:1996ke}.
Using the Hartle-Thorne perturbative expansion,
Berti and Pani~\cite{Pani:2014jra} extended this work to second-order in rotation,
which includes the quadrupolar deformation of the star.
Alternatively, one could also work entirely numerically and carry out ray-tracing calculations (as
in~\cite{Cadeau:2004gm,Cadeau:2006dc,Nattila:2017hdb,Psaltis:2013zja}),
evolving the photon geodesics in a numerically constructed spacetime for rotating
neutron stars in scalar-tensor gravity~\cite{Doneva:2013qva}.
A caveat of this numerical approach is that the spacetime can only be
determined numerically {\it after} choosing a particular conformal factor $A(\varphi)$
\emph{and} the equation of state, therefore inevitably making it
model-dependent.
Another approach would be to follow the work
in~\cite{Pappas:2014gca,Pappas:2015npa,Pappas:2016sye} and
use the Ernst formalism to obtain an axisymmetric spacetime
in Weyl-Papapetrou coordinates
in terms of an multipolar expansions of the multipole moments of the rotating
neutron star and the scalar field. The multipole moments appear as free constants
in the spacetime metric, thus allowing one to develop an extension
of the model-independent pulse profile model introduced here.
In this approach, and as an intermediate step, one would first have to {\it estimate} numerically
the values of these moments in order to have an analytic spacetime that
describes accurately the one constructed numerically, for example in~\cite{Doneva:2013qva}.
We emphasize, however, that Cadeau et al.~\cite{Cadeau:2006dc}
have shown that even in the presence of stellar oblateness,
the Schwarzschild-Doppler approximation works quite well at these frequencies when
$\theta_{\rm s}$ and $\iota_{\rm o}$ are {\it near the equator} as in the case
of Fig.~\ref{fig:illustrative_90}.
Geometrically, this is due to the small difference between the normal vector ${\bm n}$
on a spherical and a oblate surface near the equator, which suppresses the
effect of the star's oblateness. Because of the purely geometrical nature of
this argument, we expect the Just-Doppler approximation to accurately describe
the pulse profile when ${\theta}_{\rm s} \approx \iota_{\rm o} \approx 90^{\circ}$
even for rapidly rotating neutron stars in scalar-tensor gravity.
Finally, it could also be interesting to consider other theories of gravity or
to use a parametrized framework -- to consider model-independent deformations of
the Tolman-Oppenheimer-Volkoff equations and the Schwarzschild spacetime -- as
introduced in~\cite{Glampedakis:2015sua,Glampedakis:2016pes}.
Work in all of these directions is currently underway and will be reported in
several future publications.
\acknowledgments
We thank George Pappas, Cole Miller, Sharon Morsink and Kent Yagi
for useful discussions and important comments on this work.
HOS thanks Lu\'is~C.~B.~Crispino, Caio~F.~B.~Macedo,
Carolina L.~Benone, Leandro~A.~Oliveira and the Universidade Federal do Par\'a for the
hospitality while part of this work was undertaken. This work was
supported by NSF Grant No. PHY-1607130 and NASA grants NNX16AB98G and
80NSSC17M0041.
|
\section{Introduction}\label{sec:intro}
People navigate over 1 billion kilometers a day using mobile routing services~\citep{recode}, and many of these services provide travelers with information on transit networks. In most of such applications, given an origin-destination (OD) pair and a desired departure or arrival time, the route associated with the minimum expected trip time is provided to the user. However, the uncertainty associated with these recommendations, either due to variability in travel time or the transit service headway, is rarely accounted for. In contrast, a number of surveys and numerical studies have highlighted that the degree of risk aversion of transit passengers highly affects their route choices~\citep{szeto2011reliability}, even more so than in the road networks. In this work, we attempt to bridge this gap by formulating and solving the Stochastic On-time Arrival (SOTA) problem for transit networks, which provides a transit routing policy that maximizes the probability of reaching the destination within a given travel time budget in stochastic transit networks. \par
A vast majority of previous studies on routing problems in stochastic transportation networks aim to identify the route or adaptive policy with the least expected travel time (LET) \citep{loui1983optimal,hall1986fastest, polychronopoulos1996stochastic,miller2000least,waller2002online,fan2005shortest,gao2006optimal,huang2012optimal, yang2014constraint, chen2014reliable}. While the expected travel time is a natural optimality metric, there exists a variety of situations where it is not sufficient, and tail statistics must be considered; for instance, travelers who want to catch a flight are more concerned with arriving on time with a high probability rather than with minimizing their expected travel time \citep{yang2017optimizing}. To account for these contexts, other formulations that consider a reliable optimal path have been proposed, starting with~\cite{frank1969shortest}. In this formulation, the goal is to find the path that maximizes the probability of realizing a travel time smaller than some desired time budget. This definition is subsequently extended to the online context in~\cite{fan2005arriving}, and is commonly referred to as the Stochastic On-Time Arrival (SOTA) problem. In the SOTA problem, the weight of each network link is a random variable with a known probability density function that represents the travel time of the link. The solution to the SOTA problem is a routing policy that maximizes the probability of arriving at the destination within a specified time budget. Here the routing policy is an adaptive solution that determines the routing decision at each node based on the realization of the travel time experienced en-route up to that point. \par
Given the complexity of the SOTA problem, significant efforts have been made to design efficient solution algorithms. \cite{fan2006optimal} formulate it as a dynamic programming problem and use a standard successive approximation (SA) procedure. This approach, however, has no finite bound on the maximum number of iterations needed for convergence in networks with loops. In~\cite{samaranayake2012speedup,samaranayake2012tractable}, the authors propose a label-setting algorithm to resolve this issue, exploiting the fact that there is always a non-zero minimum realizable travel time on each link in road networks. To further reduce the computation time, \cite{sabran2014precomputation} modify two deterministic shortest path preprocessing techniques: reach~\citep{gutman2004reach} and arc-flags~\citep{bauer2009sharc, hilger2009fast}, and apply them to the SOTA problem. Other studies \citep{nikolova2006stochastic, nie2009shortest, parmentier2014stochastic, niknami2016tractable} explore computationally efficient solution strategies for providing reliability guarantees in stochastic shortest path problems, where a fixed route (as opposed to a policy) is desired. \par
In transit networks, the problem of determining a set of reliable travel decisions is significantly more complex than in road networks. In this context, the traveler is not in control of the transit line(s) that they may travel on to reach the destination, and may have to make a number of complex decisions regarding which transit line(s) to take. This complexity arises from the following characteristics of transit systems.
\begin{itemize}
\item \textbf{Uncertainty of arrival times/headways:} While the uncertainty in road networks is limited to the travel time, in transit networks one also needs to consider the uncertainty of arrival times and headways. Whether to take a transit line that arrives at a station or wait for a potentially faster service that is yet to arrive depends on the probabilistic trade-off between the extra waiting time and potential travel time savings. Therefore, computing the solution also requires modeling and solving for the headway distributions.
\item \textbf{Combinatorial choice set:}
In road networks, it never pays off to idle at a node and delay the departure from the node in hopes of improving the probability of arriving at the destination on time. However, in transit networks, it may be advantageous to not take the first transit line that arrives at the station (that can get you to your destination) and wait for a better option (e.g. wait for an express train or a more direct bus line). Therefore, a passenger needs to choose between multiple transit lines sharing segments of routes, and the decision regarding which transit line to board at a station depends on the unknown future arrival order of the candidate transit lines, which is an exponentially large set in the number of candidate (feasible) lines at each station, and leads to a combinatorial choice set.
\end{itemize}
Previous research on transit routing problems typically simplifies the problem by assuming that passengers board the first arriving transit service from an attractive line set that is precomputed to minimize the expected total travel time \citep{spiess1989optimal,cominetti2001common, nonner2014shortest, li2015finding}. Under this assumption, the passenger is committed to a transit line set. However, the traveler does not need to make a boarding decision until a transit vehicle is about to depart (\cite{hickman1997transit}), and it can be meaningful to adapt decisions based on information learned (uncertainties that are realized) during the trip. Fortunately, the role of online information has been noted recently in multiple studies. \cite{gentile2005route} propose a frequency-based assignment model under the assumption that the arrival time of the next transit vehicle is available once the passenger arrives at a transit station and demonstrate the potential benefits of utilizing such information. Instead of assuming full information, \cite{chen2015optimal} propose a routing strategy in a transit network with partial online information at the stations. The partial online information represents that the arrival time of the incoming transit vehicles is available only for a subset of the candidate transit lines. \cite{oliker2018frequency} develop a transit assignment model which considers two types of available information (partial information and full information), and demonstrate the impact of online information on assignment results. In our work, we consider the routing problem in transit networks as a fully online problem. The online information setting mentioned above can be easily adopted into our framework by changing the travel time and headway distributions accordingly. \par
The aim of this article is to formulate the SOTA problem for transit networks and develop efficient algorithms to solve it in this setting. The specific contributions of the work include: \par
\begin{itemize}
\item The formulation of the SOTA problem for transit networks, including a general network structure for stochastic transit networks and a decision-making model. Both the waiting time for each transit line and the travel time on each link are assumed to be random variables with known probability density functions. The solution is an adaptive policy that fully considers the transit specific characteristics of the problem mentioned above. \par
\item The design of a dynamic programming (DP) algorithm, which is pseudo-polynomial in the number of transit stations and time budget, and exponential in the number of transit lines at each station, which is practically a small number. To reduce the search space, we develop a definition of transit line dominance and present methods to identify this transit line dominance, which significantly decreases the computation time in our numerical experiments. \par
\item Extensive experiments in a synthetic network and in the Chicago transit network, which show the potential for solving this problem in a real-time route planning application setting. We also propose a general procedure to generate travel time, headway, waiting time distributions from General Transit Feed Specification (GTFS) data. \par
\end{itemize}
The rest of this article is organized as follows. In Section~\ref{sec:formulation}, we formulate the mathematical problem, in particular, we describe the network model and the underlying decision-making framework. In Section~\ref{sec: solve}, we introduce a dynamic programming based approach to solve the problem, as well as the complexity analysis of the algorithm. In Section~\ref{sec: space}, we provide a series of algorithmic techniques to reduce the search space. Section~\ref{sec:experiments} consists of numerical results on the computational performance and practical efficiency of the model and algorithms introduced in this work, both in a synthetic network and in the Chicago transit network. Finally, in Section~\ref{sec:conclusion}, we provide closing remarks and discuss possible directions for future research. \par
\section{Problem formulation.}\label{sec:formulation}
\subsection{Problem description}
In this section, we extend the SOTA problem definition to the context of transit networks. In this setting, as mentioned previously, two types of random variables need to be considered: the travel time between two transit stations and the waiting time for each transit line arrival at each transit station. The objective of the SOTA problem for transit networks is to find the routing (boarding, alighting and transferring) policy that maximizes the probability of arriving at the destination within a time budget. For conciseness, in the remainder of the article, we use \emph{utility} to represent \emph{the probability of arriving at the destination within the remaining time budget}, but note that the framework can be extended to other functions, and to robust settings \citep{flajolet2017robust}. The routing policy here includes the boarding, alighting and transferring decisions at each transit station in different situations (time already spent waiting at a station, arrival order of transit services, etc). \\
\noindent The following modeling assumptions are made.
\begin{myAsum}
Passengers' arrivals are independent of the transit schedule.
\end{myAsum}
This assumption models a situation where the transit service is frequency-based (not schedule-based) and passengers do not adjust their specific departure times based on the transit schedule. This is a standard assumption made in the literature \citep{gentile2005route,li2015finding}. In modern services, transit vehicles' positions may be published in real-time (even for frequency-based services), and thus passenger arrivals may be correlated with the bus schedule. The formulation can be adapted to account for this by changing the travel time distribution and waiting time distribution accordingly. \par
\begin{myAsum}\label{asum: once}
Only the first arrival from each transit line (after the passenger enters the station) is considered as a candidate in the choice set.
\end{myAsum}
This assumption is made for both modeling and algorithmic reasons:
\begin{itemize}
\item Guiding the passenger to ignore the first arrival from a particular line, but board the second arrival from that same transit line makes the routing direction confusing;
\item Without this assumption, it is possible for the passenger to have to make an infinite number of decisions in the theoretical worst case (albeit with vanishing probability).
\end{itemize}
The passenger may not be able to board certain transit services (especially during rush hour or after a major event), if the transit service is full and has no remaining capacity. However, since we do not have data about the occupancy of each transit vehicle, the capacity of each transit service is not considered in our study. Note that this is not a disadvantage of our framework because given the data for the occupancy of transit vehicles, we should be able to define the waiting time to be the waiting time for the first arrival transit service that the passenger can board, and our framework can then be adapted to use these distributions without any changes.
\begin{myAsum}\label{asum: nosame}
No two transit services can arrive at the transit station at the same time.
\end{myAsum}
Without loss of generality, two events never occur at exactly the same time in the continuous setting. In the subsequent discretized form of the problem, this assumption translates to no two transit services arriving at a transit station during the same discretized time interval\footnote{While the optimality of the solution relies on this assumption, if two lines arrive in the same time interval in practice, the algorithm can closely approximate the solution by considering the two cases of one arriving before the other.}.
\subsection{Transit network representation} \label{network_structure}
We consider a directed graph $G(V, E)$, in which $V$ is the set of nodes and $E$ is the set of links. We model the transit stations with three types of nodes:
\begin{itemize}
\item \textbf{Station nodes:} A station node denoted by $S_y^X$ represents a passenger waiting at station $y$ for the set of candidate transit lines $X$.
\item \textbf{Arrival nodes:} An arrival node denoted by $A_y^{i,X}$ represents transit line $i$ arriving first at station $y$ among all the candidate lines in $X \cup \{i\}$, and transit lines $j \in X$ having not arrived yet, since the passenger arrives at the station.
\item \textbf{Line nodes:} A line node denoted by $L_y^i$ represents the passenger boarding transit line $i$ at station $y$.
\end{itemize}
\begin{figure}[htb]
\centering
\includegraphics[width=0.8\textwidth]{Network.png}
\caption{\label{fig:network} Network representation of a transit station served by two transit lines. All the links and nodes within the dashed line rectangle are used to model the decision-making at a single physical station. }
\end{figure}
The set of station nodes, arrival nodes and line nodes are $V_S$, $V_A$ and $V_L$ respectively. Without loss of generality, the OD is selected from the station nodes to model a passenger starting their trip from a station node in the waiting state. Give a passenger in a waiting state at a station node with $m$ candidate transit lines, it is possible for any of the $m$ lines to arrive first. In each of these situations, the passenger either boards the transit line that arrives or continues waiting for other candidate lines. Passengers make decisions that maximize their utility. \\
We categorize links as follows:
\begin{itemize}
\item \textbf{Arrival links:} Links from station nodes to arrival nodes.
\item \textbf{Riding links:} Links from line nodes to arrival nodes.
\item \textbf{Boarding links:} Links from arrival nodes to line nodes.
\item \textbf{Alighting links:} Links from arrival nodes to station nodes.
\end{itemize}
Figure~\ref{fig:network} illustrates the network representation of the decision-making process at a transit station with two transit lines. A passenger starting the trip physically at station 1 starts the trip at station node $S_1^{1,2}$ in the model. If transit line $1$ arrives first, the passenger moves to the corresponding arrival node $A_1^{1,2}$. The passenger has to decide to either board this transit line or continue to wait for line $2$. If the passenger continues waiting, the passenger moves to station node $S_1^2$. Otherwise, the passenger moves to line node $L_1^1$. All the links and nodes within the dashed line rectangle are used to model the decision-making at a single physical station. This network model allows for transfers. For instance, a passenger having boarded line 1 at station $S_0$ (the station preceding station $S_1$) is able to get to station node $S_1^2$ and wait for the next arrival of line $2$. \par
The link costs in the network are defined as follows. A riding link $(i,j)$ is associated with a travel time distribution $p_{i, j}(\cdot)$. A boarding link has no cost, since it refers to an instantaneous event. An arrival link is associated with a waiting time distribution $w_{y}^j(\theta,r)$, characterizing the probability density for the waiting time being $\theta$ for transit line $j$ at station $y$, given that the passenger has already waited at the station for $r$ units of time. Note that $\theta$ is the waiting time on top of $r$, and models transit line $j$ arriving at station $y$ after a total waiting time of $(r+\theta)$ since the passenger arrives at the station. Therefore, $w_{y}^j (\theta, 0)$ is the original waiting time distribution when the passenger arrives at the station. Research has shown that empirical headway data fit better with Loglogistic, Gamma and Erlang distributions than the exponential distribution \citep{li2015finding}. Therefore, we do not assume that the waiting time distribution is memoryless (as some other studies do), and use the following rule to normalize the waiting time distribution given $r$:
\begin{equation}\label{normalize}
w_{y}^j (\theta,r)=\frac{w_{y}^j (r+\theta, 0)}{1-\int_0^{r} w_{y}^j (\alpha, 0) d\alpha}\quad 0\leq \theta\leq T-r, 0< r\leq T
\end{equation}
with $T$ being the total time budget when the passenger begins the trip. \par
In Section~\ref{SOTAtransit}, we discretize the time-space to solve the problem numerically. In the discretized space, we force the waiting time to be at least one unit of the discretized time interval. Therefore, the waiting time equation reads as follows.
\begin{equation}
w_{y}^j (\theta,r)=\frac{w_{y}^j (r+\theta, 0)}{1-\sum_0^{r} w_{y}^j (\alpha, 0) d\alpha}\quad 1\leq \theta\leq T-r, 0< r\leq T
\end{equation}
Although the waiting time distribution is a 2-D array, as explained in Section 3.2, we can save computation time in practice by precomputing and storing $1-\int_0^{r} w_{y}^j (\alpha, 0) d\alpha$ for each $r$ after we discretize the time-space. \par
\subsection{The SOTA problem for transit networks}\label{SOTAtransit}
In this section, we describe how to compute the utility functions of the passenger at the different types of nodes in the model. The utility at a node $i$ is a function of the remaining time budget $t$ when the passenger arrives at the node, denoted by $u_i(t)$. The utility function at the destination node $D$ is:
$$u_D (t)=1 \quad 0\leq t\leq T$$ since the passenger has completed the trip when at node $D$.
\subsubsection{Line nodes}
Recall from Figure~\ref{fig:network} that a line node only contains an outgoing edge to an arrival node. Assume that the line node we are considering is $i$ and the following arrival node is $j$. The travel time on link $(i, j)$ is a random variable $\theta$, and the remaining time budget at node $j$ is $t - \theta$. Therefore, the utility at line node $i$ is a function of the remaining time budget $t$ when the passenger arrives at the node, denoted by $u_i(t)$.
\begin{myDef}\label{def:line_node}
The utility function at a line node $i$ can be computed as follows.
$$u_i (t)= \mathop{\mathbb{E}}_{\theta} (u_{j}(t - \theta)) = \int_0^t p_{i, j}(\theta)\cdot u_j (t-\theta) d\theta, \forall i\in V_L, (i,j)\in E, 0\leq t\leq T$$ where $j\in V_A$ is the subsequent arrival node of line node $i$.
\end{myDef}
\noindent This utility function is analogous to the standard utility function of the SOTA problem for road networks.
\subsubsection{Arrival nodes}
All the routing decisions occur at arrival nodes, since the passenger is faced with the decision of either boarding the transit line that arrives first (selecting the subsequent line node) or continuing to wait for the other candidate transit lines (selecting the corresponding station node). Let $A_{y}^{i, X}$ be the arrival node of interest, and $u_{A_{y}^{i, X}} (t, r)$ be the utility when the passenger has a remaining time budget $t$ and has already waited at the station for $r$ units of time. We call the tuple $(t, r)$ the passenger state. Since the passengers aim to maximize the utility, the utility at an arrival node is the maximum of the utilities at the subsequent line node and station nodes.
\begin{myDef}\label{def:arrivalnode}
The utility function at an arrival node $A_y^{i,X}$ with passenger state $(t, r)$ is:
$$u_{A_y^{i,X}} (t,r)=\max_{ j\in V_L \mid (A_y^{i,X},j)\in E;\ S_y^X\in V_S \mid (A_y^{i,X},S_y^X)\in E} \{u_j (t),u_{S_y^X} (t,r)\}$$
\end{myDef}
\subsubsection{Station nodes}\label{station_nodes}
Station nodes are followed only by arrival nodes, and the utility at a station node is the expectation of the utility at the following arrival nodes. An arrival node is defined by the first arriving transit line and the corresponding passenger state. Let $(i, \theta)$ be a random event which represents that transit line $i$ is the first arriving transit line, and the waiting time for it is $\theta$. Assume that $S_y^{X}$ is the station node of interest, and that we want to compute $u_{S_y^X} (t,r)$. For each $0\leq \theta\leq t$, the probability of the transit line $i$ arriving the first after waiting $\theta$ units of time is $w_{y}^i(\theta,r)\cdot \prod\limits_{j\in X\backslash i}(1-\int_0^\theta w_{y}^j(\alpha,r)~d \alpha)$. In this case, the passenger has waited $\theta+r$ units of time in total, and the utility at this arrival node is $u_{A_y^{i,X\backslash i}} (t-\theta,\theta+r)$.
\begin{myDef}
The utility function at a station node is:
\begin{equation*}
\begin{split}
u_{S_y^X} (t,r) &= \mathop{\mathbb{E}}_{(i, \theta)} (u_{A_y^{i,X\backslash i}} (t-\theta,\theta+r)) \\
&= \sum_{i\in X} \left(\int_0^t w_{y}^i(\theta,r)\cdot\prod_{j\in X\backslash i}(1-\int_0^\theta w_{y}^j(\alpha,r)d\alpha) \cdot u_{A_y^{i,X\backslash i}} (t-\theta,\theta+r)d\theta\right)
\end{split}
\end{equation*}
\end{myDef}
\noindent This utility is the weighted average of the utility at the corresponding arrival nodes. \par
The integrals in the above equations do not have closed form expressions and cannot be solved analytically. Therefore, they need to be integrated numerically by discretizing the time horizon into small intervals. In the discretized model, we assume that no two transit lines can arrive at the same transit station at the same discretized time interval. As a consequence, the solution not guaranteed to be optimal, since there is a small (non-zero) probability that two lines might arrive at the same time interval regardless of how small the discretization is. However, we can set the length of time intervals to be a small number in practice\footnote{In the experiments, we set the time interval length to be 15 seconds. Using this time interval length, the average probability of multiple buses arriving at the same time interval for each station is about 0.6\% in the Chicago transit network we use in numerical experiments. For a particular OD, this probability is usually lower since the candidate bus line set is a subset.}. \par
Given a fixed time discretization, the discrete form of the utility function at station node reads as follows.
\begin{myDef}\label{def:stationnode}
The discrete form of the utility function at station nodes is:
$$u_{S_y^X} (t,r)=\sum_{i\in X} \left(\sum_{\theta=1}^t w_{y}^i(\theta,r)\prod_{j\in X\backslash i}(1-\sum_{\alpha=0}^\theta w_{y}^j(\alpha,r)) \cdot u_{A_y^{i,X\backslash i}} (t-\theta, \theta+r)\right)$$.
\end{myDef}
\noindent The utility function for other types of nodes can be discretized similarly. \par
\begin{myEx}
Here we show a simple example that shows the sophisticated decision-making process that the model allows. Assume that the passenger is waiting at a station with three candidate transit lines. Table~\ref{tab: ex1-distribution} is the waiting time distribution for each transit line, and the utility of taking the transit line corresponding to the waiting time. Figure~\ref{boarding} shows the optimal decision based on the arrival order. We can observe that: (1) The transit line that first arrives is not always the optimal choice. When transit line 3 arrives at time 2, it is the first transit service that arrives. However, it is optimal for the passenger to not board, and continue to wait; (2) Even when it is optimal to not board a transit line at an arrival event, the event may impact the future optimal decisions, as illustrated by the differing policies following the arrival or non-arrival of transit line 3 at time 2.
\end{myEx}
\begin{table}[H]
\centering
\caption{The waiting time and utility distribution for example 1. }
\label{tab: ex1-distribution}
\resizebox{.55\textwidth}{!}{
\begin{tabular}{c|c|c|c}
\hline
Transit line ID & Waiting time & Probability & Utility \\ \hline
\multirow{3}{*}{1} & 1 & 0.05 & 0.90 \\ \cline{2-4}
& 3 & 0.05 & 0.80 \\ \cline{2-4}
& 10 & 0.90 & 0.00 \\ \hline
\multirow{2}{*}{2} & 5 & 0.90 & 0.85 \\ \cline{2-4}
& 15 & 0.10 & 0.00 \\ \hline
\multirow{2}{*}{3} & 2 & 0.50 & 0.70 \\ \cline{2-4}
& 6 & 0.50 & 0.60 \\ \hline
\end{tabular}
}
\end{table}
\begin{figure}[H]
\centering
\includegraphics[width=0.9\textwidth]{boarding.png}
\caption{\label{boarding} The optimal boarding decision corresponding to the transit line arriving order for example 1. The number in each decision node (board and wait) represents the decision's utility. Note that arrival events for which boarding is never optimal are ignored in the figure (e.g. the arrival of transit line 1 at time 3 following the non-arrival of transit line 3 at time 2).}
\end{figure}
\begin{comment}
\begin{table}[H]
\centering
\caption{The optimal decision corresponding to the transit line arriving order for example 1.}
\label{tab: ex1-decision}
\begin{tabular}{c|c|c}
\hline
Case & The transit line arrival & The optimal decisions \\ \hline
\multirow{2}{*}{1} & \multirow{2}{*}{Transit line 1 arrives at time 1} & \multirow{2}{*}{Board transit line 1} \\
& & \\ \hline
\multirow{2}{*}{2} & Transit line 3 arrives at time 2 & continue to wait \\ \cline{2-3}
& Transit line 1 arrives at time 3 & Board transit line 1 \\ \hline
\multirow{2}{*}{3} & Transit line 1 arrives at time 3 & continue to wait \\ \cline{2-3}
& Transit line 2 arrives at time 5 & Board transit line 2 \\ \hline
\multirow{2}{*}{4} & Transit line 1 arrives at time 3 & continue to wait \\ \cline{2-3}
& Transit line 3 arrives at time 6 & Board transit line 3 \\ \hline
5 & Transit line 3 arrives at time 6 & Board transit line 3 \\ \hline
\end{tabular}
\end{table}
\end{comment}
\section{Solving the SOTA problem for transit networks}\label{sec: solve}
In this section, we describe how to solve the discrete time version of the SOTA problem for transit networks. We first show how the problem can still (even in the transit setting) be solved using a dynamic programming approach, then present a complexity analysis of the algorithm, and finally in Section~\ref{sec: space} describe a number of search space pruning methods designed to make the problem tractable in practice.
\subsection{Dynamic programming approach}
In \cite{samaranayake2012tractable}, a label-setting algorithm for the SOTA problem is developed by exploiting the fact that each link in a road network has a positive minimum realizable travel time. This fact guarantees that there will be no loops in the network with zero travel time, and thus allows for a dynamic programming approach for solving the problem. However, in the transit network representation proposed in the previous section, there are links with zero minimum realizable travel time.
\begin{myCla} \label{c1}
In the transit network representation introduced in Section~\ref{sec:formulation}, there is no loop with a zero realizable travel time.
\end{myCla}
\begin{proof}
Only the links starting from the arrival nodes can have zero minimum realizable time. To form a loop, the passenger has to first get to an arrival node from some other type of node, which is not an arrival node. The links originating from all other types of nodes have a positive minimum realizable travel time, so a zero travel time loop is not possible.
\end{proof}
Therefore, we can still use a dynamic programming approach to solve the SOTA problem for transit networks. Each sub-problem in the dynamic program can be defined by $(i, t, r)$ where $i$ is the node that we consider, $t$ and $r$ have the same definition as in Section 2. Let $\text{OPT}(i, t, r)$ denote the utility at node $i$ when the passenger has a time budget of $t$ and has waited at the station for $r$ units of time. Note that the $\text{OPT}(i, t, r)$ satisfies the Bellman's Principle of Optimality, i.e., the remaining decisions only depend on the current state and are not related to the past states and decisions. According to the definitions in Section 2, we claim that $\text{OPT}(i, t, r)$ satisfies the following recurrence relation: \par
\begin{equation}\label{equ: recurrence}
\resizebox{\textwidth}{!}{%
$\text{OPT}(i, t, r)=\left\{
\begin{aligned}
& 1 & \quad \text{if } i \text{ is the destination} \\
& 0 & \quad \text{if } t < 0 \\
& \sum\limits_0^t p_{i, j}(\theta)\cdot \text{OPT}(j, t-\theta, 0) & \quad \text{if } i \text{ is a line node and } (i, j) \in E \\
& \max_{ j\in V_L \mid (A_y^{i,X},j)\in E;\ S_y^X\in V_S \mid (A_y^{i,X},S_y^X)\in E} \{\text{OPT}(j, t, 0), \text{OPT}(S_{y}^{X}, t, r)\} & \quad \text{if } i \text{ is arrival node } A_{y}^{i, X} \\
& \sum_{i\in X} \left(\sum_{\theta=1}^t w_{y}^i(\theta,r)\prod_{j\in X\backslash i}(1-\sum_{\alpha=0}^\theta w_{y}^j(\alpha,r)) \cdot \text{OPT}(A_y^{i,X\backslash i}, t-\theta, \theta+r)\right) & \quad \text{if } i \text{ is station node } S_{y}^{X} \\
\end{aligned}
\right.$}
\end{equation}
The initial problem we wish to solve is given by $(O, T, 0)$, i.e., the passenger is at the origin node $O$ with time budget $T$ and has waited for zero units of time. The algorithm first initializes the utility function corresponding to the destination node to 1 and computes all
the normalized waiting time distributions that are needed in the subsequent computation. Then, the utility at each node is updated in a dynamic programming fashion according to the recurrence relation given in Equation~\ref{equ: recurrence}.\par \par
\subsection{Complexity analysis}
The runtime of the algorithm is pseudo-polynomial in the number of stations in the transit network and time budget, and exponential in the number of transit lines at any station. We first provide the number of nodes of all three types. Then, we analyze the time complexity for computing the utility on each type of node.
\begin{myCla} \label{c: n_nodes}
For a station with $m$ transit lines, there are i) $m$ line nodes, ii) $\sum\limits_{i=1}^{m} C_{m}^i$ station nodes, and iii) $\sum\limits_{i=1}^{m} i\cdot C_{m}^i$ arrival nodes.
\end{myCla}
\begin{proof}
i) For every transit line, there is a corresponding line node. Therefore, there are $m$ line nodes in total. ii) For each combination of transit lines representing the non-empty set of lines yet to arrive, there is a station node. Therefore, there are $\sum\limits_{i=1}^{m} C_{m}^i$ station nodes in total. iii) For each station node, any transit line can be the line arriving the first and yield an associated arrival node. Therefore, there are $\sum\limits_{i=1}^{m} i\cdot C_{m}^i$ arrival nodes.
\end{proof}
We now analyze the time complexity of the algorithm for a transit network with a station set $Y$ and each station has no more than $m$ transit lines. Let $M_y$ be the set of transit lines at station $y$. The algorithm first computes all the normalized waiting time distributions that are needed in the subsequent computation, which takes $O(|Y|\cdot m\cdot T^3)$ time based on Equation~\ref{normalize}. In our implementation, we use $O(|Y|\cdot m \cdot T)$ memory to store $1 - \sum_{\alpha=0}^{\theta} w_{y}^j(\alpha,0), \forall \theta \leq T, y \in Y, j \in M_y$, which decreases the time complexity to $O(|Y|\cdot m\cdot T^2)$.
\begin{myCla} \label{c: u_complexity}
For a station set $Y$ in which each station has no more than $m$ transit lines, the time complexity of computing the utility functions for a time budget of $T$ is:
\begin{itemize}
\item $O(|Y|\cdot m\cdot 2^{m-1}\cdot T^2)$ for all arrival nodes,
\item $O(|Y|\cdot (m^{2} - m) \cdot 2^{m-2}\cdot T^3)$ for all station nodes.
\end{itemize}
\end{myCla}
\begin{proof}
Based on Definition~\ref{def:arrivalnode}, we need to update the utility functions at each arrival node for each possible remaining time budget $t$ and each waiting time $r$. In addition, based on Claim \ref{c: n_nodes}, there are $\sum\limits_{i=1}^{m} i\cdot C_{m}^i$ arrival nodes for a station with $m$ transit lines, so it takes $O(|Y|\cdot \sum\limits_{i=1}^{m} i\cdot C_{m}^i\cdot T^2)=O(|Y|\cdot m\cdot 2^{m-1}\cdot T^2)$ time to compute the utility at all the arrival nodes.
Assume that the station node of interest is $S_y^{X}$. To compute $U_{S_y^{X}}(t, r)$, for each transit line $i\in X$, we need to compute the probability of line $i$ arriving among all lines in $X$, which is $\sum_{\theta=1}^t w_{y}^i(\theta,r)\prod_{j\in X\backslash i}(1-\sum_{\alpha=0}^\theta w_{y}^j(\alpha,r))$ according to Definition~\ref{def:stationnode}. Therefore, for each $i\in X$ and passenger state $(t, r)$, we need $O((|X| - 1)\cdot T^2)$ to compute the above probabilities. Based on Definition~\ref{def:stationnode}, we need to update the utility functions at each station node for each possible remaining time budget $t$ and each waiting time $r$. Consequently, we need $O(|X|\cdot (|X|-1) \cdot T^{4})$ to update the utility functions for each station node in total. According to Claim \ref{c: n_nodes}, there are $\sum\limits_{i=1}^{m} C_{m}^i$ station nodes for a station with $m$ transit lines. Therefore, it takes $O(|Y|\cdot \sum\limits_{i=1}^{m} (C_{m}^i \cdot i \cdot (i-1) )\cdot T^4)=O(|Y|\cdot (m^{2} - m) \cdot 2^{m-2}\cdot T^4)$ time to compute the utility at all the station nodes, without re-using any information. We find that the probability of transit line $j$ arriving after $\alpha$ time given that the passenger has already waited $r$ time at station $y$, which appears in Definition \ref{def:stationnode} reads: $1-\sum_{\alpha=0}^\theta w_{y}^j(\alpha,r) = \frac{1 - \sum_{\alpha=0}^{r + \theta} w_{y}^j(\alpha,0)}{1 - \sum_{\alpha=0}^{r} w_{y}^j(\alpha,0)}$. In our implementation, we use $O(|Y|\cdot m \cdot T)$ memory to store $1 - \sum_{\alpha=0}^{\theta} w_{y}^j(\alpha,0), \forall \theta \leq T, y \in Y, j \in M_y$. Then, the time complexity for computing utility for all station nodes becomes $O(|Y|\cdot m^{2} \cdot 2^m\cdot T^3)$.
\end{proof}
In summary, the time complexity of the algorithm is $O(|Y|\cdot m^2 \cdot 2^m\cdot T^3)$. Considering that there is a constant maximum number of transit lines at a station, the time complexity is $O(|Y|\cdot T^3)$, which leads to a pseudo-polynomial time algorithm in $|Y|$ and $T$.
\section{Search space reduction}\label{sec: space}
Although the DP approach is a pseudo-polynomial time algorithm in $|Y|$ and $T$, the computation time can still be high in large-scale networks when $T$ is large. Therefore, we propose some search space reduction techniques to further decrease the computation time in practice. \par
\subsection{Eliminate the infeasible paths}
The simplest pruning technique we employ is to eliminate infeasible paths, i.e., paths that have zero probability of being used based on the minimum realizable travel time on each link. This pruning can be performed by simply running a shortest path search on a modified graph where the link weight is the minimum realizable travel time, to find the shortest path distance from each station to the destination. Assume that the minimum realizable travel time from station $y$ to the destination is $\alpha_y$. The utility for any node at station $y$ given a time budget smaller than $\alpha_y$ will be zero.
\par
This method is similar to the pruning method in \cite{samaranayake2012tractable} for road networks. Since the complexity of the shortest path algorithm is dominated by the complexity of the SOTA problem, the cost of the pruning method is negligible compared to the total computation time. \par
\subsection{Search space reduction using transit line dominance}
To compute the utility at an arrival node, we need to compare the utility of boarding the transit service (line node) and continuing to wait (station node). In this section, we propose a definition of transit line dominance, and a set of computationally efficient conditions to check for such dominance. This allows us to save the computation for the utility of the corresponding station node in the case of a dominating line node. \par
\subsubsection{Dominance definition and properties} \label{sec: relation}
\begin{myDef}\label{def:dom}
Assume that the passenger is at node $A_{y}^{i, X}$ with passenger state $(t, r)$. Transit line $i$ dominates $X$ if $u_{L_{y}^{i}}(t) \geq u_{S_{y}^{X\backslash i}} (t, r)$. We use $i\succeq X$ to represent that $i$ dominates $X$, and $i\prec X$ to represent that $i$ does not dominate $X$.
\end{myDef}
According to the definition, if $i \succeq X$, the passenger should board the transit line $i$; If $i \prec X$, the passenger should continue to wait for the transit lines in $X$. We now propose a useful claim that will be used in the proof of the dominance properties that we propose later. \par
\begin{myCla}\label{c: addline}
Assume that the passenger is at station $y$ with passenger state $(t, r)$. $u_{S_y^{X}}(t, r) \leq u_{S_y^{X^\prime}}(t, r), \forall X^\prime \supset X$.
\end{myCla}
\begin{proof}
The passengers in our system are utility maximizers. Therefore, more transit line choices will make the utility of the station node increase or remain the same.
\end{proof}
\begin{myPro}\label{cor: domsub}
Assume that the passenger is at station $y$ with passenger state $(t, r)$. If $i \succeq X$, then $i \succeq X^\prime, \forall X^\prime \subset X$. Correspondingly, if $i \prec X$, then $i \prec X^\prime, \forall X^\prime \supset X$.
\end{myPro}
\begin{proof}
We prove the first part of the claim. The second part can be proved using similar reasoning. Assume that the passenger is at station $y$. According to Definition~\ref{def:dom}, $u_{L_y^{i}}(t) \geq u_{S_y^{X}}(t, r)$ if $i \succeq X$. In addition, for any $X^\prime \subset X$, $u_{S_y^{X^\prime}}(t, r) \leq u_{S_y^{X}}(t, r)$ based on Claim~\ref{c: addline}. Therefore, $u_{L_y^{i}}(t) \geq u_{S_y^{X^\prime}}(t, r)$, and thus $i\succeq X \implies i \succeq X^\prime, \forall X^\prime \subset X$.
\end{proof}
\begin{myPro}
Assume that the passenger is at station $y$ with passenger state $(t, r)$. If $u_{L_y^{i}}(t) \geq u_{L_y^{j}}(t)$, then: (1) $i \succeq X$ for any $X$ such that $j\succeq X$; (2) $j \prec X^\prime$ for any $X^\prime$ such that $i \prec X^\prime$.
\end{myPro}
\begin{proof}
For (1), if $j \succeq X$, then $u_{L_y^{j}}(t) \geq u_{S_y^{X}}(t, r)$ according to the definition of transit line dominance. Since $u_{L_y^{i}}(t) \geq u_{L_y^{j}}(t)$, we derive $u_{L_y^{i}}(t) \geq u_{S_y^{X}}(t, r)$, which proves that $i \succeq X$. \par
For (2), as $i \prec X^\prime$, $u_{L_y^{i}}(t) \leq u_{S_y^{X}}(t, r)$. Since $u_{L_y^{i}}(t) \geq u_{L_y^{j}}(t)$, $u_{L_y^{j}}(t) \leq u_{S_y^{X}}(t, r)$, which represents $j \prec X^\prime$.
\end{proof}
According to the two properties above, we can infer transit line dominance based on the transit line dominance that we already know. Therefore, at an arrival node $A_y^{i, X}$ with passenger state $(t, r)$, if we can infer $i \succeq X$ from the known dominance, we can reduce the computation for $u_{S_y^{X}}(t, r)$. \par
\subsubsection{Dominance conditions}\label{sec: dom}
In this section, we present a series of conditions that can be assessed efficiently, and are practically useful to reduce the number of utility computations at station nodes. The conditions are considered at arrival nodes in order. \par
The first dominance condition is a subproblem pruning technique. If the condition is satisfied, we do not need to compute the utility at the corresponding station node at this specific passenger state. \par
\begin{myProp} \label{p: dom}
(Subproblem pruning) Assume that the passenger is at node $A_y^{i, X}$ with passenger state $(t, r)$. If $u_{L_y^i} (t)\geq \max\limits_{j\in X} \{ u_{L_y^j} (t-1)\}$, then $i\succeq X$.
\end{myProp}
\begin{proof}
The utility function is a nondecreasing function with respect to the time budget, i.e., $u_{L_y^j} (t-1)\geq u_{L_y^j} (t^\prime), \forall t^\prime\leq t-1, \forall j \in X$. In addition, $u_{S_y^{X}}(t, r)$ is a weighted average of $u_{L_y^{j}}(t^\prime)$ for $t^\prime \leq t$ and $j \in X$ where the sum of the weight is not larger than 1. Therefore, if the condition is satisfied, $u_{L_y^i} (t) \geq u_{S_y^{X}}(t, r) \implies i \succeq X$.
\end{proof}
Proposition \ref{p: dom} needs $O(m)$ time to be checked. The intuition is as follows: Assume that we know all the transit lines in $X$ will arrive right after $i$ arrives. If boarding the transit line $i$ under this assumption has a larger utility than waiting for the rest, then boarding the transit line $i$ also has a larger utility than waiting for the rest without the assumption (i.e. the transit lines in $X$ might arrive later). \par
If Proposition~\ref{p: dom} cannot prove that $i \succeq X$, then we need to compute $u_{S_y^{X}}(t, r)$ explicitly. We show next that, in some cases, we can safely remove some transit lines in $X$ and only consider a subset of the candidate transit lines. Before we propose this candidate set pruning technique, we first provide the following definition of \textit{improving lines} that will be used later on. \par
\begin{myDef} \label{def: improve}
Assume that the passenger is at station $y$ with passenger state $(t, r)$. Transit line $k$ improves line $j$ at station $y$ if $u_{S_y^{\{k, j\}}}(t, r) > u_{S_y^{\{j\}}}(t, r)$.
\end{myDef}
A sufficient and necessary condition for transit line $k$ improving line $j$ at station $y$ is given as follows. \par
\begin{myCla} \label{c6}
Assume that the passenger is at node $A_y^{i, X}$ with passenger state $(t, r)$. Line $k$ improves line $j$ if there exists some state $(t-\gamma,r+\gamma)$ such that $u_{L_y^k} (t-\gamma)>\sum\limits_{\theta=1}^{t-\gamma} (w_{y}^j (\theta,r+\gamma)\cdot u_{L_y^j} (t-\gamma-\theta))$.
\end{myCla}
\begin{proof}
If there exists a passenger state $(t-\gamma,r+\gamma)$ satisfying the condition above in Claim~\ref{c6}, the passenger should board $k$ when it arrives at state $(t-\gamma, r+\gamma)$ instead of continuing to wait for $j$. Therefore, the utility at the station node with $\{k,j\}$ is larger than the utility at station node with only line $j$.
\end{proof}
Now, we propose the candidate set pruning technique as follows. \par
\begin{myProp}
(Candidate set pruning) Assume that the passenger is at node $A_y^{i, X}$ with passenger state $(t, r)$. Let $Z = \{j \in X\mid u_{L_y^{i}(t)} < u_{L_y^j}(t - 1)\}$. When we compute $u_{S_y^{X}}(t, r)$, we can instead compute $u_{S_y^{X^\prime}}(t, r)$ where $X^\prime \subseteq X$ and $X^\prime = Z \cup \{j \in X \mid j \text{ improves at least one line in } Z\}$.
\end{myProp}
\begin{proof}
Let $K = X \backslash X^\prime$. Since no line in $K$ improves the lines in $Z$, it is never better to take a line in $K$ than waiting for the lines in $Z$, i.e., $u_{S_y^{Z \cup K}}(t, r) = u_{S_y^{Z}}(t, r)$. Since $Z \subset X^\prime$, it is also never better to take the line in $K$ than waiting for the lines in $X^\prime$, which implies $u_{S_y^{X^\prime}}(t, r) = u_{S_y^{X}}(t, r)$.
\end{proof}
We can consider the candidate set pruning technique as an extension of Proposition~\ref{p: dom} since $Z$ is obtained when we check the condition in Proposition~\ref{p: dom}. If $|Z| = 0$, the condition in Proposition~\ref{p: dom} is satisfied, $i \succeq X$. Otherwise, we prune the transit lines in $\{j \in X\backslash Z \mid j \text{ cannot improve any line in } Z\}$ before computing for the utility at the station node. \par
According to Definition~\ref{def:stationnode}, we iterate through the time intervals $\{z \mid 1 \leq z \leq t\}$ when we compute $u_{S_y^{X}}(t, r)$. Here $z$ is the waiting time for the first transit line arrival. In some cases, we can infer $i \succeq X$ when we finish the computation for an iteration $z < t$ and thus reduce the computation for the rest of the iterations. \par
\begin{myProp} \label{p: early_stop}
(Time interval pruning) Assume that the passenger is at $A_y^{i,X}$ with passenger state $(t,r)$. Let $u_{\text{max}}$ be the maximum utility among all subsequent arrival nodes of $S_{y}^{X}$ at passenger state $(t - z, r + z)$, i.e., $u_{\text{max}}(z)=\max\limits_{j\in X}\{{u_{A_y^{j,X\backslash j}}}(t-z,z+r)\}$. Let $p_{j, X}(\theta)$ be the probability that transit service $j$ is the first transit service arrives at the station among all transit lines in $X$, and it arrives at the station at passenger state $(t - \theta, r + \theta)$, i.e., $p_{j, X}(\theta) = w_{y}^j(\theta,r)\prod\limits_{k\in X\backslash j}(1-\sum\limits_{\alpha=1}^\theta w_{y}^k(\alpha,r))$. Let $u_{\text{sum}}(z)$ be the expected utility for the cases where there is one transit service in $X$ arriving at the station from passenger state $(t, r)$ to $(t - z, r + z)$, i.e., $u_{sum}(z)=\sum\limits_{j\in X} \sum\limits_{\theta=1}^z (p_{j, X}(\theta) \cdot u_{A_y^{j,X\backslash j}} (t-\theta,\theta+r))$. If there exists $z\leq t$ such that $u_{L_y^i} (t)\geq u_{sum}(z)+(1-\sum\limits_{j\in X} \sum\limits_{\theta=1}^z p_{j, X}(\theta))\cdot u_{max}(z)$, then $i\succeq X$.
\end{myProp}
\begin{proof}
If we can show that $u_{sum}(z)+(1-\sum\limits_{j\in X} \sum\limits_{\theta=1}^z p_{j, X}(\theta))\cdot u_{max}(z) \geq u_{S_y^{X}}(t, r)$, then the proposition is proved. According to Definition~\ref{def:stationnode}, $u_{S_y^{X}}(t, r) = u_{sum}(t)$. We can infer that $u_{S_y^{X}}(t, r) = u_{sum}(z) + \sum\limits_{j\in X} \sum\limits_{\theta=z + 1}^t (p_{j, X}(\theta) \cdot u_{A_y^{j,X\backslash j}} (t-\theta,\theta+r))$. Since the utility function is non decreasing, we know that $u_{\text{max}}(z) = \max\limits_{j \in X, z \leq \theta \leq t} \{{u_{A_y^{j,X\backslash j}}}(t-\theta,\theta+r)\}$. In addition, $1 - \sum\limits_{j\in X} \sum\limits_{\theta=1}^z p_{j, X}(\theta) \geq \sum\limits_{j\in X} \sum\limits_{\theta=z + 1}^t p_{j, X}(\theta)$. Therefore, we can conclude that the proposition is correct since $u_{S_y^{X}}(t, r) \leq u_{\text{sum}}(z)+(1-\sum\limits_{j\in X} \sum\limits_{\theta=1}^z p_{j, X}(\theta) \cdot u_{\text{max}}(z)$.
\end{proof}
Whenever we finish computing for a specific time interval $z = i$, we can check the condition in Proposition \ref{p: early_stop}. If the condition is not satisfied, we let $z = i + 1$ and continue the procedure. However, if the condition is satisfied, we know that $i\succeq X$ and thus can reduce the computation for $i + 1 \leq z \leq t$. \par
To sum up all the techniques discussed in this section and Section \ref{sec: relation}, we modify the procedure of computing the utility at arrival nodes as follows. Two dictionaries $dom$ and $nondom$ are used to record the transit line dominance and non-dominance. $update\_dom$ and $update\_nondom$ are two functions to update $dom$ and $nondom$ according to dominance properties in Section \ref{sec: relation}. Specifically, for any arrival node $A_y^{i,X}$, the pruning is done via the following set of steps. \par
\noindent\fbox{%
\parbox{\textwidth}{%
Step 1: Check if $X$ is in $dom[i,t,r]$: if yes, return $u_{L_y^i}(t)$; Otherwise, go to Step 2.
Step 2: Check if $X$ is in $nondom[i,t,r]$: if yes, compute $u_{S_y^X}(t,r)$ and return it; Otherwise, go to Step 3.
Step 3: Check if $i\succeq X$ or $i\prec X$ using the pruning techniques. If $i\succeq X$, run $update\_dom$ and return $u_{L_y^i}(t)$; Otherwise, run function $update\_nondom$ and return $u_{S_y^X}(t,r)$.
}%
}
We call the algorithm DP with dominance. Note that all the methods discussed retain the optimality of the solution.\par
\subsection{Heuristic rules}\label{sec:heur}
In this section, we propose three heuristic rules, which can be employed in the procedure of checking dominance, to further decrease the computation time. We will show in the numerical experiments that the heuristics can significantly reduce the computation time without much loss of the results accuracy. \par
\begin{myHeu}\label{h: dom}
Assume that the passenger is at node $A_y^{i,X}$ with passenger state $(t,r)$. Let $Z = \{j \in X\mid u_{L_y^{i}(t)} < u_{L_y^j}(t - 1)\}$, and $p = \prod\limits_{j \in Z} \sum\limits_{\theta = 1}^{t} I_{u_{L_y^{i}(t)} \geq u_{L_y^{j}(t - \theta)}} \cdot w_{y}^{j}(\theta, r)$ where $I_{u_{L_y^{i}(t)} \geq u_{L_y^{j}(t - \theta)}}$ is an indicator function. If $p \geq \epsilon$, we say that the passenger should board the transit line $i$.
\end{myHeu}
This is an extension of Proposition~\ref{p: dom}. $\epsilon$ is a constant coefficient. $p$ is the probability of the case that boarding transit line $i$ has a larger utility than boarding any other single transit line in $X$. If $p$ is large, then only with a small probability, the realization of the waiting time for a transit line in $X$ can make its utility larger than boarding transit line $i$. \par
\begin{myHeu}\label{h: anyother}
Assume that the passenger is at node $A_y^{i, X}$ with passenger state $(t, r)$. If $u_{L_y^i} (t) \geq \sum\limits_{\theta=1}^{t} w_{y}^j (\theta,r)\cdot u_{L_y^j} (t-\theta), \forall j \in X$, we say that the passenger should board the transit line $i$.
\end{myHeu}
In other words, the passenger should board transit line $i$ if boarding it has a larger utility than waiting for any single line in $X$. It is an approximation since $u_{S_y^{X}}(t, r) \geq u_{S_y^{\{j\}}}(t, r), \forall j\in X$ according to Claim~\ref{c: addline}. Therefore, it is possible that $u_{S_y^{X}}(t, r) > u_{L_y^i} (t)$. \par
\begin{myHeu}\label{h: timeinterval}
Assume that the passenger is at $A_y^{i,X}$ with passenger state $(t,r)$. Let $u_{max}(z)=\max\limits_{j\in X}\{{u_{A_y^{j,X\backslash j}}}(t-z,z+r)\}$, and $u_{sum}(z)=\sum\limits_{j\in X} (\sum\limits_{\theta=1}^z w_{y}^j(\theta,r)\prod\limits_{k\in X\backslash j}(1-\sum\limits_{\alpha=1}^\theta w_{y}^k(\alpha,r)) \cdot u_{A_y^{j,X\backslash j}} (t-\theta,\theta+r))$. If there exists $z\leq t$ such that $\beta \cdot u_{L_y^i} (t)\geqslant u_{sum}(z)+(1-\sum\limits_{j\in X} (\sum\limits_{\theta=1}^z w_{y}^j(\theta,r)\prod\limits_{k\in X\backslash j}(1-\sum\limits_{\alpha=1}^\theta w_{y}^k(\alpha,r))))\cdot u_{max}(z)$, then we say that the passenger should board transit line $i$.
\end{myHeu}
This is an extension of Proposition~\ref{p: early_stop}. The difference is that we add a constant relaxation coefficient $\beta$ ($\beta > 1$) in the condition. This represents that when we compute the utility of a station node, if $\beta$ times the utility of the line node is larger than the largest possible utility of the station node, we say that the passenger should board the transit line that arrives. $\beta$ in Heuristic~\ref{h: timeinterval} and $\epsilon$ in Heuristic~\ref{h: dom} are used to control the tradeoff between the results accuracy and computation performance. In the numerical experiments, we let $\beta = 1.25$ and $\epsilon = 0.75$. \par
In the following section, we present numerical results and compare the three versions of the solution algorithm introduced in this work: (1) DP; (2) DP with dominance; (3) DP with dominance and heuristics. \par
\section{Numerical experiments}\label{sec:experiments}
In this section, we present numerical experiments focused on illustrating the runtime performance of the various versions of the solution, as well as the practical value of using a SOTA policy in transit networks. All the algorithms are coded in Python $3.6$, and all tests are conducted on an AMD Ryzen processor computer (3.4 gigahertz, 16 gigabytes RAM). We first validate the algorithms and conduct a sensitivity analysis in a controlled setting using a synthetic transit network, and then test them on the Chicago transit network.
\subsection{Estimating travel time and headway distributions}
To obtain the required network and transit line information, we develop a procedure for taking the input data (GTFS data) and generate all the processed data we need in the experiments, including travel time distributions, headway distribution, and waiting time distributions. The first step is to generate the travel time distributions for each transit line. Studies have shown that travel time distributions on road networks can be well approximated by a lognormal distribution (see \cite{emam2006using,hunter2009path,zou2014space}). Therefore, in our experiments, we model the travel time distributions using appropriately parameterized lognormal distributions. A lognormal distributed random variable $X$ is parameterized by two parameters $\mu$ and $\sigma$ that are, respectively, the mean and standard deviation of the variable's natural logarithm. For every two adjacent stations $S_1$ and $S_2$ in a transit line, we calibrate $\mu$ and $\sigma$ for the travel time between the two stations based on the following steps: \par
(1) Compute the minimum realizable travel time $t_m$ by dividing the distance between $S_1$ and $S_2$ by the corresponding speed limit. \par
(2) Let the mode of the lognormal distributed variable, i.e., the point of the maximum of the probability density function, be the difference between the scheduled departure time at $S_2$ and $S_1$. The mode of a lognormal distribued variable is $e^{\mu - \sigma^2}$. Therefore, if the schedule departure time at $S_1$ and $S_2$ are 8:45 and 9:00 and $t_m = 5$ min, then we have $e^{\mu - \sigma^2} = 10$. \par
(3) In the absence of more information from GTFS, $\sigma$ of the lognormal distribution is sampled uniformly from $0.25 \leq \sigma \leq 0.5$ so that the variance is neither too large or too small. \par
(4) Compute $\mu$ according to the mode and $\sigma$. Shift the distribution rightwards by $t_m$. \par
As $\sigma$ is randomly sampled from a given range, a sensitivity analysis is conducted in Section~\ref{sec: sigma}. If one can get access to the transit vehicle's real-time location, $\sigma$ can be obtained from the real realized travel time data. After computing the travel time distribution between every two adjacent stations, we can obtain the travel time distribution between the origin and all other stations on each transit line by computing the appropriate summation of the random variables due to the independence assumption. \par
The headway at the origin station of each transit line is set to the mean of the difference between the scheduled departure time for every two consecutive trips in our experiment time span. For instance, if the scheduled departure times for a transit line $i$ at the origin station are 8:45, 8:55, 9:10, then we let the headway be $\frac{10 + 15}{2} = 12.5$. The headway for other stations is computed as follows. Let $X_1$ and $X_2$ be the arrival time of the first transit vehicle and the second transit vehicle of the transit line of interest at a station $S$. $O$ is the origin, and $h$ is the deterministic headway at $O$. The cumulative distribution function of the headway is computed as follows: \par
$$P(X_2-X_1\leq t)=P(X_2\leq X_1+t)=\int_0^\infty P(X_2\leq x+t)\cdot P(X_1=x)dx$$
where $P(x_2\leq x+t)$ and $P(X_1 = x)$ are the cumulative distribution function of $X_2$ and the probability density function of $X_1$. Let $t_{O, S}$ be the travel time between $O$ and $S$. The distribution of $X_1$ is the same as the distribution for $t_{O, S}$, and the distribution of $X_2$ is to shift the distribution for $t_{O, S}$ rightwards by $h$. \par
Finally, we estimate the waiting time distribution from the headway distributions following~\cite{larson1981urban}. Let $\mathbb{E}_{i,y}[h_{i,y}]$ be the expected headway of transit line $i$ at station $y$, and $H_{i, y}(\cdot)$ be the cumulative distribution function of headway of transit line $i$ at station $y$. Then the waiting time reads:
$$w_y^i(t,0)=\frac{1}{\mathbb{E}_{i,y}[h_{i,y}]}\cdot (1-H_{i,y}(t))$$
\subsection{Synthetic network experiments}
The synthetic network we use is a 3-line 3-station transit network, as shown in Figure~\ref{fig:small_network}. The three transit lines pass through each of the three stations. The line attributes are presented in Table~\ref{tab: line_attributes}. The headway and travel time are in minutes. \par
\newsavebox{\tempbox}
\newlength{\tempwidth}
\begin{figure*}[htb]
\savebox{\tempbox}{\includegraphics[scale=0.295,clip=true,draft=false,]{small_network.png}}%
\settowidth{\tempwidth}{\usebox{\tempbox}}%
\hfil\begin{minipage}[b]{\tempwidth}%
\raisebox{-\height}{\usebox{\tempbox}}%
\captionof{figure}{A 3-line synthetic transit network.}%
\label{fig:small_network}%
\end{minipage}%
\savebox{\tempbox}
\scriptsize{
\begin{tabular}{@{}cccc@{}}
\toprule
Line i & \begin{tabular}[c]{@{}c@{}}Headway\\ at station A\end{tabular} & \begin{tabular}[c]{@{}c@{}}Travel time\\ from A to B\end{tabular} & \begin{tabular}[c]{@{}c@{}}Travel time\\ from B to C\end{tabular} \\ \midrule
1 & 10 & 4 & 5 \\
2 & 15 & 4 & 3 \\
3 & 12 & 7 & 4 \\ \bottomrule
\end{tabular}
}}%
\settowidth{\tempwidth}{\usebox{\tempbox}}%
\hfil\begin{minipage}[b]{\tempwidth}%
\raisebox{-\height}{\usebox{\tempbox}}%
\captionof{table}{Line attributes of the 3-line synthetic transit network. All times are provided in minutes. }%
\label{tab: line_attributes}%
\end{minipage}%
\end{figure*}
\subsubsection{Performance analysis}\label{sec: Ae}
We first illustrate the performance of the algorithms, specifically the relationship between the computation time and the time budget. Let station $A$ be the origin, and station $C$ be the destination, and the time discretization of the algorithm to be $15$ seconds. The three algorithms introduced in Section~\ref{sec: solve} are tested on time budgets ranging from 10 min to 45 min with a step size of 2.5 min. \par
\begin{figure}
\begin{minipage}[t]{0.5\linewidth}
\centering
\includegraphics[width=3.5in]{run_time_performance_small.png}
\end{minipage}%
\begin{minipage}[t]{0.5\linewidth}
\centering
\includegraphics[width=3.5in]{run_time_performance_small_domandheu.png}
\end{minipage}
\caption{Computation time of three algorithms on the synthetic network. \textit{Left:} All three algorithms. \textit{Right:} Only algorithms with the dominance based pruning to showcase the relative improvement. }
\label{fig:computation_time}
\end{figure}
As shown in Figure~\ref{fig:computation_time}, the algorithms with search space reduction techniques perform significantly better than the basic DP approach. The DP with the dominance test can provide an average computation time reduction of $77.8\%$ compared to the basic DP approach. The heuristics further decrease the computation time by $68.9\%$ on average relative to the DP with dominance. The time reduction percentage ($\frac{\text{Computation time of DP - Computation of DP with dominance}}{\text{Computation time of DP}}$) is typically higher when the time budget is large. When the time budget is 45 min, the time reduction can be $97.3\%$ for the DP with dominance and heuristics. The average relative error for the heuristic method is only $2.8\%$ in this experiment. Therefore, the heuristic (at least in this case) provides a satisfactory trade-off between accuracy and computation time. \par
\begin{figure}
\centering
\includegraphics[width=3.6in]{utility_diff_small.png}
\caption{Comparison of the utility between the SOTA policy and LET path.}
\label{fig:u_diff_small}
\end{figure}
We also quantify the benefits of utilizing a SOTA policy instead of using the least expected travel time (LET) path. This is done by using the expected value of both the travel time and waiting time distributions to compute the LET path between station $A$ and station $C$. The utility of the LET solution is then computed based on the probability of success when using the LET path for different time budgets. Figure~\ref{fig:u_diff_small} illustrates the utility difference between the LET path and the SOTA policy we obtain from our algorithm. When the time budget is very small, the utility difference is zero because the passenger cannot reach the destination on time regardless of the policy the passenger uses. Similarly, when the time budget is very large, the utility difference is also 0 because the passenger reaches the destination on time with high probability even with a sub-optimal route. The SOTA policy will always be no worse than the LET solution by definition. The maximum utility difference observed in this experiment is 23 percent (when the time budget is 22.5 min in this case). \par
\subsubsection{Sensitivity to the modes of the travel time distributions. }\label{sec: modes}
In this section, we analyze the effect of the modes\footnote{Note that \textit{mode} here is the statistical definition.} of travel time distributions on computation time. The modes of the travel time distributions on each link are sampled uniformly from a range instead of being predefined. Two sets of experiments are conducted. In the first set of experiments, we let the width of the range be 1. More specifically, we first generate a random number $i \in \{1,2,3,4,5,6,7,8,9\}$ uniformly, and then set the range to be $[i,i+1)$. Assume that $X_1$ and $X_2$ are two random variables sampled uniformly from this range. Then, $E(|X_1-X_2|) = \frac{1}{3}$ since $|X_1-X_2|$ follows a triangle distribution with the parameters $a=0$, $c=0$ and $b=1$. In the second set of experiments, the range is set to $[1,10)$, and $E(|X_1-X_2|)=3$ since $|X_1-X_2|$ follows a triangle distribution where $a=0$, $c=0$ and $b=9$. We call the first set of experiments \emph{low diff}, and the second set of experiments \emph{high diff}. Each set of experiments are conducted for 100 times. \par
\begin{figure}[htb!]
\centering
\includegraphics[width=0.65\textwidth]{high_low_diff.png}
\caption{\label{fig:highlow} Computation time reduction of speed-up techniques on two cases. \emph{high diff} represent the case where the travel time distributions' modes are highly different, while \emph{low diff} represents the case where the travel time distributions' modes are relatively close. }
\end{figure}
The computation time reduction with respect to the time budget is shown in Figure~\ref{fig:highlow}. The search space reduction techniques work well in both sets of experiments. The average computation time reductions for the DP with dominance are $88.9\%$ and $67.1\%$ for the \emph{low diff} and \emph{high diff} cases respectively relative to standard DP, and the time reductions are $95.7\%$ and $89.8\%$ for the DP with dominance and heuristics. The search space reduction techniques work better in the \emph{low diff} cases because the transit services that arrive the first will intuitively dominate the other transit lines more often with similar travel time distributions, since waiting is not likely to increase the utility. Another interesting result is that the heuristics will provide a higher time reduction for the \emph{high diff} cases compared to the DP with dominance algorithm. A potential reason is that a line will have a higher probability of being better than any other single candidate line at the same station in this case, which indicates dominance according to the Heuristic~\ref{h: anyother}. \par
\subsubsection{Sensitivity to the $\sigma$ parameter} \label{sec: sigma}
We now analyze the sensitivity of the algorithm performance to the parameter $\sigma$ in the lognormal travel time distributions. Assume that the CDF of a lognormal distribution is $F(x)$, and $\gamma$ is the mode of the distribution. In Section~\ref{sec: Ae}, we use $\sigma=0.25$, which makes $F(1.5\cdot \gamma)\approx 0.9$ when $\gamma$ is in the range shown in Table \ref{tab: line_attributes}. To show the effects of $\sigma$ on the distribution, we compare the distribution with $\sigma = 0.25$ and $\sigma = 0.5$ in Figure~\ref{fig:lognormal}. When $\sigma=0.5$, $F(1.5\cdot \gamma)\approx 0.6$. \par
\begin{figure}[htb]
\centering
\includegraphics[width=0.6\linewidth]{sigma_pdf.png}
\captionof{figure}{The PDF of lognormal distributions with different $\sigma$.}
\label{fig:lognormal}
\end{figure}
To analyze the sensitivity of $\sigma$ on the algorithm's performance, two sets of experiments are conducted. In the first set of experiments, the algorithms are tested under $\sigma=0.5$. In the second set of experiments, $\sigma$ is uniformly sampled from $[0.25,0.5]$ for each link of each transit line. Again, each set of experiments are conducted for 100 times. \par
\begin{table}[htb]
\centering
\caption{Algorithms' performance under different $\sigma$ settings.}
\label{diffsigma}
\begin{tabular}{@{}lll@{}}
\toprule
Parameter setting & \begin{tabular}[c]{@{}l@{}}Computation time reduction\\ for DP with dominance\end{tabular} & \begin{tabular}[c]{@{}l@{}}Computation time reduction\\ for DP with dominance and heuristics\end{tabular} \\ \midrule
$\sigma=0.25$ & $77.8\%$ & $92.6\%$ \\
$\sigma=0.5$ & $80.3\%$ & $94.1\%$ \\
\begin{tabular}[c]{@{}l@{}}$\sigma$ randomly\\ chosen from $[0.25,0.5]$\end{tabular} & $79.8\%$ & $93.9\%$ \\ \bottomrule
\end{tabular}
\end{table}
As shown in Table~\ref{diffsigma}, the search space reduction techniques perform similarly under different $\sigma$ values. The computation time is reduced slightly more when the variance of the travel time distribution is larger. In the following experiments for the Chicago transit network, we use $\sigma$ randomly sampled from $[0.25, 0.5]$. \par
\subsection{Chicago network experiments}
We now use the Central/South region (including downtown) of the Chicago transit network as a case study. The network is created according to the GTFS data published by the Chicago Transit Authority (CTA) through Google’s GTFS project. The GTFS data contains schedules and associated geographic information. In the experiments, only active bus lines in the morning (6 am to 10 am) are considered. The transit network, which contains 49 transit lines and 1565 stations, is shown in Figure~\ref{fig:chicago}. \par
\begin{figure}[htb]
\centering
\includegraphics[width=0.4\textwidth]{network_map.png}
\caption{\label{fig:chicago} Chicago transit network (Central/South region).}
\end{figure}
\subsubsection{Runtime performance in the Chicago network} \label{sec: Ae-chi}
The average travel time to commute in the United States is 26.1 min according to the U.S. Census Bureau \citep{www.census.gov}. Therefore, we select 100 ODs randomly from the station set under the constraints: (1) The expected trip duration is from 15 min to 45 min; (2) At least one of the origin and the destination is in the downtown area, to simulate the case of commuting. The experiments are conducted for the time budgets which also span from 10 min to 45 min. For each time budget, we use the same three algorithms described previously to compute the utility. The computation time for each algorithm is shown in Figure~\ref{fig:chicago_efficiency}. \par
\begin{figure}
\begin{minipage}[t]{0.5\linewidth}
\centering
\includegraphics[width=3.5in]{run_time_performance_chicago.png}
\end{minipage}%
\begin{minipage}[t]{0.5\linewidth}
\centering
\includegraphics[width=3.5in]{run_time_performance_chicago_domandheu.png}
\end{minipage}
\caption{Computation time of three algorithms on the Chicago transit network. \textit{Left:} All three algorithms. \textit{Right:} Only algorithms with the dominance based pruning to showcase the relative improvement. }
\label{fig:chicago_efficiency}
\end{figure}
The DP with dominance reduces the computation time by $89.3\%$ on average. The computation time reductions are significant, but not as high as in the synthetic network shown in Figure~\ref{fig:small_network} especially when $T$ is large. In the synthetic network, there are only three stations which limit the search space. However, in real-size networks, the search space will be much larger since there are more stations and transit lines to be considered. The heuristic rules further decrease the computation time by $34.0\%$ on average relative to the DP with dominance. The average relative error for the heuristic rules is about $1.6\%$. \par
\begin{figure}[htb]
\centering
\includegraphics[width=0.65\textwidth]{num_dom_reduction.png}
\caption{\label{fig:station_reduction} The computation reduction for computing the utility on station nodes. The vertical axis shows the percentage of function calls reduction to compute the utility at stations. }
\end{figure}
Recall from Section~\ref{sec: space}, if $i\succeq X$ at an arrival node $A_y^{i, X}$ given passenger state $(t, r)$, we save the computation for $u_{S_y^{X}}(t, r)$. In the experiments, we also record the number of function calls to compute the utility at station nodes. Figure~\ref{fig:station_reduction} shows the percentage of the function calls reduction by DP with dominance and DP with dominance and heuristics. Both algorithms provide more than $90\%$ function calls reduction for all the cases. However, the reduction decreases as the time budget increases. When the time budget is higher, the set of feasible arrival nodes that followed by a station node $S_y^{X}$ at passenger state $(t, r)$ is larger. Here an arrival node $A_y^{i, X}$ is \textit{feasible} at passenger state $(t, r)$ if it is not pruned by the techniques in Section~\ref{sec: space}. Let $Z_y^X$ be the set of feasible arrival nodes followed by station node $S_y^{X}$. To save the computation for $u_{S_y^{X}} (t, r)$, every line in $\{i \mid A_y^{i, X} \in Z_y^{X}\}$ needs to dominate $X$, which is of less probability when $|Z|$ is higher. \par
The average computation time of the DP with dominance and heuristics is only about 9.2 seconds on a personal computer, which indicates the potential of running the algorithm in real-time applications. \par
\subsubsection{Compare the utility of the SOTA policy and LET path}
\begin{figure}[htb]
\centering
\includegraphics[width=0.65\textwidth]{utility_diff.png}
\caption{\label{fig:utility_diff} The histogram of the utility difference between the SOTA policy and LET path on the Chicago transit network.}
\end{figure}
Recall from Section~\ref{sec: Ae}, the utility difference between the SOTA policy and LET path can be as large as 0.23 in the synthetic network. In this section, we test the utility difference in the Chicago transit network. Intuitively, if there is one and only one transit line that can directly take the passenger to the destination without transferring, the SOTA policy and LET path will likely be the same due to the time cost of transferring. Therefore, we compare the utility difference of all the ODs from the station set under the constraints: (1) The expected trip duration is from 15 min to 45 min; (2) At least one of the origin and the destination is in the downtown area; (3) There is no transit line or there are more than one transit line that can directly take the passenger to the destination. The number of ODs satisfying the above conditions is 9290. Again, the experiments are conducted for the time budgets from 10 min to 45 min. For each OD, we record the maximum utility difference between the SOTA policy and LET path among all the time budgets we test. Figure~\ref{fig:utility_diff} shows the histogram of the maximum utility difference. The utility difference for $19.6\%$ of the ODs is larger than 0.05, and the utility difference for $7.6\%$ of the ODs is larger than 0.1. The utility difference is close to 0 for most of the ODs, which are the cases that either there is only one route choice between the origin and the destination or there is one obvious better line than the other choices. It should be noted that the results highly depend on the network and the travel time distribution we use. \par
\begin{figure}
\begin{minipage}[t]{0.5\linewidth}
\centering
\includegraphics[width=3.6in]{n_start_stations.png}
\end{minipage}%
\begin{minipage}[t]{0.5\linewidth}
\centering
\includegraphics[width=3.6in]{leastetime.png}
\end{minipage}
\caption{The utility improvement of the SOTA policy over the LET path under different circumstances. \textit{Left:} With respect to the number of transit lines passing by the origin. \textit{Right:} With respect to the ratio of the travel time of the LET path and the time budget. }
\label{fig:n_start_stations}
\end{figure}
To give insights of what trips will likely benefit from the SOTA policy, we show the utility increases of the SOTA policy over the LET path on different circumstances with respect to: i) The number of transit lines passing by the origin; ii) The ratio of the least expected time (LET) path travel time to the time budget. As shown in Figure~\ref{fig:n_start_stations} (Left), a SOTA policy outperforms the LET path more when the number of transit lines at the origin increases, which demonstrates that a SOTA policy is more advantageous when there are more states and options to consider. In addition, even though there might be more transit lines to transfer on the way, it may not be beneficial to conduct a transfer in most of the cases due to the transfer time. Therefore, when there is only one transit line passing by the origin, the average utility difference is about 0.025, which is close to 0. The ratio of the LET path travel time to the time budget represents how sufficient the time budget is relative to the trip length. Figure~\ref{fig:n_start_stations} (Right) shows the utility increase versus the ratio. If the ratio is relatively large (larger than 2), passengers cannot increase the utility by using the adaptive routing policy since it is difficult to get to the destination in a very limited time budget; On the other hand, if the ratio is very small (smaller than 0.5), the utility increase is also close to 0 since passengers will likely reach the destination within the given time budget whatever transit line they board. The SOTA policy performs the best when the ratio is around 1. In summary, the SOTA policy in transit networks shines the most when there are more states to consider and when the time budget is neither too high or too low relative to the trip length. \par
\section{Conclusion.}\label{sec:conclusion}
In this article, we extend the SOTA formulation to the case of transit networks. A network representation comprised of three types of nodes is proposed to model the common line problem in the SOTA framework. Instead of assuming that the passenger will board the first arriving transit service in an attractive line set that is precomputed, we give a routing policy which outperforms a-priori solutions for all practical purposes. From the perspective of computation, we show that computing the utility in transit networks is significantly more difficult than in road networks, and design an dynamic programming based algorithm to solve the problem, which is pseudo-polynomial in the number of stations and time budget, and exponential in the number of transit lines at a station, which is practically a small number. \par
To reduce search space, we propose a definition of transit line dominance, and a series of techniques to check if a transit line dominates. Experiments are conducted on both a synthetic network and the Chicago transit network. The results show that the search space reduction techniques can reduce the computation time by up to around $90\%$, and the heuristic rules can further decrease the computation time by about $34\%$. Two sensitivity analyses on the travel time distribution show that the search space reduction techniques can significantly decrease the computation time in all the cases we test. Finally, the algorithms are applied in the Chicago transit network, and similar conclusions are obtained. \par
In future research, we will focus on designing proper preprocessing techniques to further reduce the computation time in practice. In reality, the spatial and temporal correlation of link travel time usually exists in transit networks. In addition, the travel time and waiting time can be time-varying at the different time of a day (e.g., at peak hours or off-peak hours). Therefore, one possible future research direction is to explore the ways to generate more realistic travel time and waiting time distributions and design algorithms to deal with them. \par
\bibliographystyle{elsarticle-harv}
|
\section{Introduction}
With the increasing popularity of online music streaming services, the task of selecting relevant and personalized content in large music catalogs becomes important to avoid choice overload \cite{bollen2010understanding}.
An open challenge in the area of personalization for music recommender systems is known as \emph{automatic playlist continuation} (APC), where the task is to recommend tracks that are likely to be selected as additional tracks for an existing playlist. In APC it is important to recommend relevant content while, at the same time, respecting the characteristics of the original playlist \cite{schedl2018current}. For example, the recommended songs for the continuation of a playlist that consists of Christmas songs should be other Christmas songs.
To promote progress in the area of APC, the RecSys Challenge 2018 focuses on this task. The challenge was organized by Spotify, The University of Massachusetts, Amherst, and Johannes Kepler University, Linz, and was open for submissions from January to July 2018.
In the competition, participants were challenged to create a recommendation system for APC using a dataset of one million playlists that have been created by Spotify users in North America.
The competition task was to generate a list of 500 tracks as playlist continuation for each of the 10000 playlists in the challenge dataset. The playlists in the challenge set were divided into ten \emph{challenge categories}, based on the number of seed tracks (the tracks that are already known to be present in the playlist) and the availability of the playlist title.
This manuscript describes the solution proposed by Team Latte. The main idea behind the proposed approach is to construct several collaborative filters, based on the co-occurrence of tracks with other tracks, artists, albums, and words in the playlist title. Furthermore, the solution adopts a specialized optimization strategy, where the weights of each collaborative filter are optimized locally within each challenge category using an optimization method called Tree-structured Parzen Estimator (TPE)~\cite{bergstra2011algorithms}. The final recommendation scores are produced as a weighted sum of the individual collaborative filtering components, and then post-processed using several heuristic strategies.
The remainder of this paper is organized as follows:
Section~\ref{sec:dataset} describes the data provided during the competition.
Section~\ref{sec:model} describes the proposed framework based on several collaborative filters and their combination.
Section~\ref{sec:setup} further details the model optimization and selection procedures to combine the collaborative filers, and discusses some results on internal validation sets.
Finally, in Section~\ref{sec:conclusion} we conclude the paper.
\section{Dataset}
\label{sec:dataset}
The dataset consists of one million playlists created by Spotify users and distributed as the Million Playlist Dataset (MPD) for exclusive use in the competition. The MPD dataset includes information about the playlist (title, identification, number of artists, playlist duration) and track information (album name, identification, artist, track duration, track name) for every playlist. A complete description of the dataset can be found in \cite{spotify2018mpd}.
In addition to the MPD dataset, the organizers provided the \emph{challenge dataset}, i.e. an official test dataset that contains partial information about 10000 playlists: the playlist title and/or a number of \emph{seed tracks} (a subset of tracks present in the playlist). The aim of the challenge was to generate a list of 500 \emph{recommended tracks} for each of these playlists, based on the partial information available. Additionally, each playlist contained a number of \emph{holdout tracks} that were known only to the organizers of the challenge. The submissions of the participants were evaluated and ranked based on the correspondence between the recommended tracks and the holdout tracks~\cite{spotify2018challengerules}.
The playlists in the challenge dataset can be divided into ten distinct challenge categories based on the type of the provided partial information \cite{spotify2018challengerules}, namely:
\begin{itemize}
\item G1: Playlists with a title only
\item G2: Playlists with a title and the first track
\item G3: Playlists with a title and the first five tracks
\item G4: Playlists with first five tracks (no title)
\item G5: Playlists with a title and the first ten tracks
\item G6: Playlists with first ten tracks (no title)
\item G7: Playlists with a title and the first 25 tracks
\item G8: Playlists with a title and 25 random tracks
\item G9: Playlists with a title and the first 100 tracks
\item G10: Playlists with a title and 100 random tracks
\end{itemize}
\begin{comment}
Playlists in $D_{test}$ are grouped accordingly to the total number of seed tracks $K$, namely:
$K=0$ for empty playlists;
$K=1$ for playlists containing only one track;
$K=5$ for playlists containing five tracks;
$K=10$ for playlists containing ten tracks;
$K=25$ for playlists containing twenty five tracks;
and $K=100$ for playlists containing one hundred tracks;
\end{comment}
Table~\ref{tab:challenge_dataset} shows the distribution of playlists, the average number of holdout tracks ($H_\mathit{avg}$), the number of unique seed tracks ($N_\mathit{Track}$), and the number of unique artists ($N_\mathit{Artist}$) present the challenge dataset, grouped according to the number of seed tracks $K$.
\begin{table}[tb]
\caption{Challenge Dataset Statistics}
\label{tab:challenge_dataset}
\begin{tabular}{ccccc}
\toprule
Group & $N_\mathit{Playlist}$ & $H_\mathit{avg}$ & $N_\mathit{Track}$ & $N_\mathit{Artist}$ \\
\midrule
K=0 & 1000 & 29 & 0 & 0 \\
K=1 & 1000 & 23 & 932 & 715 \\
K=5 & 2000 & 55 & 6790 & 2762 \\
K=10 & 2000 & 53 & 11877 & 4096 \\
K=25 & 2000 & 126 & 22507 & 6253 \\
K=100 & 2000 & 88 & 53552 & 11517 \\
\bottomrule
\end{tabular}
\end{table}
\section{Framework}
\label{sec:model}
In this section, we describe the framework for automatic playlist continuation (APC). The framework consists of three steps: first a collection of multiple different collaborative filtering models are extracted in the \emph{collaborative filtering stage}, then, the predictions of the collaborative filtering models are combined into a single relevance prediction per playlist-track combination in the \emph{composition stage}, and finally, the recommendations are generated in the \emph{playlist continuation stage}. We now continue with describing these stages in detail.
\subsection{Collaborative Filtering Stage}
The task of collaborative filtering is to predict the utility of items (tracks) to a particular \emph{context} (playlist) based on vector similarities between these entities extracted from data \cite{breese1998empirical}. This \emph{context} can be based on different aspects of a playlist. For example, in item-item collaborative filtering, the context is based on the tracks that are already present in the playlist.
A total of four collaborative models were built in order to capture different contexts:
\begin{description}
\item[track-track model ($M_u$)]{models the relevance of a given track for a given playlist based on the set of tracks that are currently present in the playlist. This model is a traditional item-item collaborative filtering model.}
\item[word-track model ($M_w$)]{models the relevance of a given track for a given playlist based on the name of the playlist. This collaborative filter that models the relation between words in the playlist name and the occurrence of tracks when a playlist contains this word in the playlist title. The words are extracted from the playlist names by splitting the playlist name on the space character (i.e. ' '), transforming the results to lowercase, and removing punctuation marks.}
\item[album-track model ($M_\mathit{al}$)]{models the relevance of a given track for a given playlist based on the albums from the tracks that are currently present in the playlist. This collaborative filter models the relation between the set of albums of the tracks that are currently in the playlist and the occurrence of tracks when these albums are in the playlist.}
\item[artist-track model ($M_\mathit{ar}$)]{models the relevance of a given track for a given playlist based on the artists the created the tracks that are currently present in the playlist. This collaborative filter models the relation between the set of albums of the tracks that are currently in the playlist and the occurrence of tracks when these albums are in the playlist.}
\end{description}
\subsection{Composition Stage}
The output of every collaborative filter is combined in a final ranking model ($M_{c}$) using a weighted sum given by:
\begin{equation}
M_{c} = W_u * M_u + W_w * M_w + W_\mathit{al} * M_\mathit{al} + W_\mathit{ar} * M_\mathit{ar}
\end{equation}
where $W_u$, $W_w$, $W_\mathit{al}$ and $W_\mathit{ar}$ are real-valued weights in range $[0,1]$.
The best configuration of weights is found using an optimization procedure, such as Tree-structured Parzen Estimator (TPE)~\cite{bergstra2011algorithms}. We experiment with two types of weighting schemes: 1) global weights (optimized over all instances) and 2) local weights (optimized separately for each challenge category). We describe the procedure to determine the weights $W_u$, $W_w$, $W_\mathit{al}$, and $W_\mathit{ar}$ in detail in Section~\ref{sec:setup}.
\subsection{Playlist Continuation Stage}
To determine the recommended tracks for a given playlist, we filter the tracks on $M_c>0$ and then sort the tracks in descending order based on their $M_c$ value, using the $M_c$ value that uses the weights that we found in the composition stage. However, it can be the case that fewer than 500 tracks have a value of $M_c$ that is larger than zero, in which case the requirement of recommending 500 songs would not be satisfied.
To improve the order of tracks in the recommendations ranking and to guarantee a total 500 recommended tracks for every playlist, we apply two post-processing steps.
The first post-processing step aims at completing the albums that are currently already present in the playlists. This is motivated by the fact that a reasonable number of playlists in the dataset contained exactly all the tracks of a single album, and we found the $M_c$ to be insufficient to properly detect this scenario and complete the album for playlists that contain a high number of tracks from the same album. When the ratio of the number of tracks from the number of distinct albums that are currently in the playlist exceeds a threshold $m$ (where $m$ is a tunable parameter), we first recommend all the tracks from that remaining album before recommending the tracks based on $M_c$.
As a second post-processing step, to fulfill the requirement of recommending exactly 500 tracks, we append the list of recommended tracks with the most popular tracks in the dataset in decreasing order of overall frequency until the list of recommended tracks contains exactly 500 tracks.
\section{Model selection}
\label{sec:setup}
We evaluate the model instantiations using a combination of three measures R-precision, NDCG, and CLICKS, which are the same three measures that are used by the RecSys challenge organizers to score the submissions. We select the best performing model instantiation for submission.
In this section we present the evaluation measures, the procedure for optimizing the models' parameters, the procedure and the results for selecting the best model. The framework was implemented in Python and can be found at \cite{latte2018recsys} under open source license.
\subsection{Evaluation measures}
\label{sec:evaluation}
The R-precision, defined as:
\begin{equation}
\text{R-precision} = \frac{|G \cap R|}{|G|}
\end{equation}
where $G$ is the set of ground truth (holdout) tracks, and $R$ is the set of recommended tracks. The notation $|.|$ denotes the number of elements in the set.
The Normalized discounted cumulative gain (NDCG), defined as:
\begin{equation}
\label{equ:ndcg}
\text{NDCG} = \frac{DCG}{IDCG}
\end{equation}
where:
\begin{equation}
\text{DCG} = rel_1 + \sum_{i=2}^{|R|} \frac{rel_i}{log_2(i+1)}
\end{equation}
\begin{equation}
\text{IDCG} = 1 + \sum_{i=2}^{|G|} \frac{1}{log_2(i+1)}
\end{equation}
The Recommended Songs CLICKS metric, that mimics a Spotify feature for track recommendation where ten tracks are presented at a certain time to the user as the suggestion to complete the playlist. This metric captures the number of refreshes needed before a relevant track is encountered, and is defined as:
\begin{equation}
\text{CLICKS} = \frac{arg min_i \{R_i:R_i \in G\} - 1}{10}
\end{equation}
While NDCG and CLICKS were calculated based on the track-level agreement between the holdout tracks and the recommended tracks, R-precision was calculated on the artist-level agreement. In other words, it was considered sufficient if the artist of a recommended track matched the artist of a holdout track.
\subsection{Approaches}
We tested three instantiations of the proposed framework, namely:
\begin{itemize}
\item composition via global weights;
\item composition via local weights, without album completion (i.e., $m=\infty$);
\item composition via local weights, where the album completion threshold $m$ is optimized through the same procedure as optimizing the weights.
\end{itemize}
As a baseline, we compared the results to a simple popularity-based model, where the recommendation list is created based on the overall popularity of songs in a non-personalized manner.
\subsection{Model Optimization}
For each of the tested approaches, the weights for combining the collaborative filters needed to be optimized. Furthermore, in the variant with album completion, the song to album ratio $m$ was optimized. To this end, we extracted a optimization dataset ($D_\mathit{opt}$) containing 10k playlists (playlists $980001-990000$ from the MPD). Similarly to the original challenge dataset, we divided these playlists into 10 distinct categories that match the challenge categories (see Section~\ref{sec:dataset}) via random sampling. The statistics of the $D_\mathit{opt}$ dataset can be seen in Table~\ref{tab:opt_dataset}.
\begin{table}[tb]
\caption{Optimization Dataset Statistics}
\label{tab:opt_dataset}
\begin{tabular}{lrrrr}
\toprule
Group & $N_\mathit{Playlist}$ & $H_\mathit{avg}$ & $N_\mathit{Track}$ & $N_\mathit{Artist}$ \\
\midrule
K=0 & 1000 & 38 & 0 & 0 \\
K=1 & 1000 & 37 & 942 & 758 \\
K=5 & 2000 & 33 & 7548 & 3496 \\
K=10 & 2000 & 33 & 13487 & 5127 \\
K=25 & 2000 & 32 & 27185 & 8789 \\
K=100 & 2000 & 53 & 76648 & 18242 \\
\bottomrule
\end{tabular}
\end{table}
The optimization process was set maximize the NDCG metric (Equation~\ref{equ:ndcg}) and was executed using Tree-structured Parzen Estimator (TPE)~\cite{bergstra2011algorithms}, which is a type of a Sequential Model-Based Global Optimization (SMBO)~\cite{Hutter2011} algorithm. We use the TPE implementation that is available in the Python library Hyperopt~\cite{bergstra2013hyperopt}.
The TPE optimization process was set to run for 100 iterations and the search space of weights defined as a uniform random variable ranging from 0 to 1.
Table~\ref{tab:optimal_weights} shows the optimized best sets of weights separately for each category (used in the local weights composition) and global weights (used in the global weights composition).
\begin{table}[!htbp] \centering
\caption{Optimized weights}
\label{tab:optimal_weights}
\begin{tabular}{lrrrrr}
\toprule
Category & $W_u$ & $W_w$ & $W_\mathit{al}$ & $W_\mathit{ar}$ & $m$ \\
\midrule
title\_only & $0.000$ & $1.000$ & $0.000$ & $0.000$ & - \\
1\_with\_title & $1.000$ & $0.423$ & $0.001$ & $0.011$ & 1 \\
5\_no\_title & $1.000$ & $0.000$ & $0.040$ & $0.040$ & 2 \\
5\_with\_title & $1.000$ & $0.337$ & $0.006$ & $0.010$ & 2 \\
10\_no\_title & $1.000$ & $0.000$ & $0.003$ & $0.009$ & 2 \\
10\_with\_title & $1.000$ & $0.964$ & $0.002$ & $0.001$ & 2 \\
25\_first & $1.000$ & $0.795$ & $0.001$ & $0.037$ & 2 \\
25\_random & $1.000$ & $0.437$ & $0.145$ & $0.022$ & 2 \\
100\_first & $1.000$ & $0.828$ & $0.028$ & $0.045$ & 2 \\
100\_random & $0.915$ & $1.000$ & $0.169$ & $0.133$ & 3 \\
global & $1.000$ & $0.517$ & $0.084$ & $0.056$ & - \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Model Selection}
In order to select the best model from the proposed framework in an offline manner (without making an official submission), we extracted a validation dataset ($D_\mathit{val}$) containing 10k playlists (playlists $990001-1000000$ from the MPD). Again, we divided these playlists into ten distinct challenge categories via random sampling. The statistics of the $D_\mathit{val}$ dataset can be seen in Table~\ref{tab:val_dataset}. The validation set was used as a proxy to the challenge leaderboard, guiding model selection and improvements.
\begin{table}[tb]
\caption{Validation Dataset Statistics}
\label{tab:val_dataset}
\begin{tabular}{lrrrr}
\toprule
Group & $N_\mathit{Playlist}$ & $H_\mathit{avg}$ & $N_\mathit{Track}$ & $N_\mathit{Artist}$ \\
\midrule
K=0 & 1000 & 38 & 0 & 0 \\
K=1 & 1000 & 38 & 943 & 737 \\
K=5 & 2000 & 34 & 7451 & 3427 \\
K=10 & 2000 & 33 & 13455 & 5197 \\
K=25 & 2000 & 33 & 26864 & 8512 \\
K=100 & 2000 & 53 & 80171 & 18570 \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Results}
\label{sec:results}
This subsection presents and discusses the results of our experiments.
Figure~\ref{fig:models_performance} shows the performance (in terms of NDCG, CLICKS, and RPREC) for the tested instantiations of the framework and the baseline popularity model. Note that in all cases, the composed collaborative model performs better than the popularity model. The model with local weights and album completion is the best performing model and was the selected strategy for our final submission.
\begin{figure}[tbh]
\centering
\includegraphics[width=1\linewidth]{model_performance}
\caption{Performance of different models on validation set}
\label{fig:models_performance}
\end{figure}
The results of the final model on both the validation set ($D_\mathit{val}$) and the challenge set are presented in Table~\ref{tab:leaderboard}.
The Leaderboard score is the score given by the submission website, calculated by the organizers based on the recommended tracks and the holdout tracks (the ground truth values not available to participants) in the challenge dataset.
\begin{table}[tbh]
\caption{Results of Composed Model}
\label{tab:leaderboard}
\begin{tabular}{lrr}
\toprule
Metric & Validation & Leaderboard\\
\midrule
RPREC & 0.150587 & 0.203652 \\
NDCG & 0.288921 & 0.361175 \\
CLICKS & 5.6156 & 2.0240 \\
\bottomrule
\end{tabular}
\end{table}
To further analyse the performance of the final model within different challenge categories, Table~\ref{tab:results_testgroup} presents the results for the composed model in each of these categories in the validation set\footnote{Note that the overall scores are slightly different than in the above, since this detailed evaluation was executed with training on 400k playlists and a total of 100k tracks only, to reduce the computations.}. We can see in this table that the model is doing considerably better in the groups where the seed tracks were selected randomly from the playlist. The performance is lowest in the category where only the playlist title was provided as input.
\begin{table}[tbh] \centering
\caption{Results by challenge category (model trained on 400k playlists, 100k tracks)}
\label{tab:results_testgroup}
\begin{tabular}{lrrrr}
\toprule
Category & NDCG & CLICKS & RPREC \\
\midrule
title\_only & $0.179$ & $13.990$ & $0.092$ \\
1\_with\_title & $0.266$ & $7.252$ & $0.141$ \\
5\_no\_title & $0.270$ & $6.373$ & $0.135$ \\
5\_with\_title & $0.271$ & $5.943$ & $0.135$ \\
10\_no\_title & $0.259$ & $6.740$ & $0.128$ \\
10\_with\_title & $0.263$ & $6.565$ & $0.130$ \\
25\_first & $0.258$ & $6.353$ & $0.122$ \\
25\_random & $0.353$ & $3.307$ & $0.199$ \\
100\_first & $0.219$ & $6.093$ & $0.108$ \\
100\_random & $0.372$ & $2.782$ & $0.223$ \\
\midrule
Overall & $0.271$ & $6.540$ & $0.141$ \\
\bottomrule
\end{tabular}
\end{table}
\section{Conclusion}
\label{sec:conclusion}
In the 2018 RecSys challenge, teams competed in the task of automatic playlist competition. To simulate different challenges in the playlist completion task, a challenge dataset was provided with ten different types of seed information (called challenge categories). Our solution was based on combining multiple different collaborative filters that each capture different aspects of a playlist, and we combined them using a Tree-structured Parzen Estimator optimization approach where we optimized the weights locally for each of the challenge categories. The solution strategy shows promising results, ranking our team in position 12 out of 112 teams in the final competition leaderboard.
\bibliographystyle{ACM-Reference-Format}
|
\section{INTRODUCTION}
In 1950, Alan Turing developed a test to assess a machine's ability to display behavior equivalent to that of a human being~\cite{turing2009computing}. The evaluator (human) will be exposed to a blind conversation with a machine and another human, and by formulating questions (s)he will try to identify his interlocutor (i.e., whether it is the human or the machine). If the evaluator cannot distinguish between human and machine, the latter one is said to have passed the test. In this paper, we conduct an experiment inspired by the Turing test. We want to check whether RePOR our automatic software refactoring tool~\cite{RePOR_JSS_symred} can be as effective as human developers at least from the refactoring structure point of view.
Indeed, one major limitation of RePOR, and existing refactoring tools, is the lacking of semantic and contextual information of identifiers. To have a fair comparison in our study, for a subset of refactoring actions, we used synthetic identifiers also for human refactored code and asked developers to judge the soundness of refactoring actions and code structure. It is important to underline, RePOR evaluation goal was not to verify if the refactoring results were deemed useful or necessary by the original developers or if original developers would have performed the same code changes. The overarching goal was the verify if RePOR was capable to produce refactoring actions and code comparable to human changes quality wise.
To this aim (1) we presented practitioners with refactorings performed manually by developers and refactorings performed by RePOR, and asked them to identify the origin of the refactorings. We also asked them to judge the design quality of the refactorings. Next, (2) we asked practitioners to perform a series of comprehension tasks on (both manually and automatically) refactored code and we assessed their performance.
\parrafo{Context} Software systems age as the result of deviations of their original design due to the implementation of new features~\cite{Parnas94-ICSE-SoftwareAging}, changes in the business logic, etc. One way to combat software system deterioration is to perform maintenance activities continuously; correcting
poor design choices ({\textit{a.k.a.,}}~anti-patterns)~\cite{Brown98-AntiPatterns} for example through refactorings~\cite{fowler1999refactoring}. Refactoring is a software maintenance activity that aims to reorganize code structure without altering system's behavior~\cite{fowler1999refactoring}. In fact, agile methodologies like \emph{eXtreme Programming (XP)} encourages developers to interleave refactoring with their code tasks to ease software evolution. However, refactoring is a time-consuming activity because (1) developers have to identify software components that contain anti-patterns; (2) select the adequate refactorings to clean-up the code; and (3) select the best order to apply the refactorings determined in the previous step. Yet, the benefits of code refactoring to combat software aging can only be observed in the long term.
Previous works have studied refactoring practice in academic and industrial settings, and found that refactoring is a common practice, that the refactorings manually performed by developers differ from those applied using tool support, and that tool support is underused due to a lack of awareness~\cite{griswold1993automated,murphyreftools2008,vakilian2012}. In the last decade, tools and frameworks supporting refactorings have been proposed~\cite{moghadamAutoRefDesig2012,Ouni201518,morales2016Tarf,Morales2016Recon}. However, to the best of our knowledge, little effort has been taken to evaluate the impact on code understandability of automated-refactoring compared to manual refactoring. As a consequence, many developers express natural reluctance about \reviewed{applying automated refactoring to their code bases}~\cite{le2012systematic}. To accept or reject the notion that automated-refactoring can replace manual one, we propose to submit the refactorings applied by human and machine to the scrutiny of software developers, who will act as judges, \reviewed{similar to a Turing test}, to assess the quality, pertinence, and understandability of automatically generated code structure. Our goal is to verify if RePOR challenges the idea that automated refactorings can't compete with refactorings performed by humans, in terms of quality, according to human judges.
\parrafo{Premise}
Refactoring is conjectured in the literature to improve the design quality of systems. Despite the large number of studies on refactoring summarized in Section~\ref{sec:relatedwork}, few studies have empirically investigated the impact of automated refactoring tools on program comprehension. Yet, program comprehension is crucial to practitioners responsible of performing software maintenance. Hence, a better understanding of the conditions in which automated- is as suitable as manual-refactoring can promote the development of more human-like tools that perform automation efficiently, reducing maintenance costs.\\
\parrafo{Goal}
We want to gather quantitative and qualitative evidences that automated-refactoring can succeed in improving the design quality of a system at least as well as a human being would do in similar conditions. \\
\parrafo{Study}
We perform two experiments: (1) a refactoring survey ($E_1$) where we invited developers from Java mailing lists, and technical groups on social networks, to differentiate refactoring code changes (that aimed to remove anti-pattern's instances) generated by a software tool (machine) from those generated by freelance developers (human). For simplicity, we called this test in the consecutive "Turing test", acknowledging the differences on implementation, as a Turing tests typically require human interaction between human judges and machine in the form of natural language. The freelancers that refactored the code were hired from two well-known crowdsourcing marketplace websites\footnote{Freelancer.com and Guru.com}. We also asked the participants of the refactoring survey to rank the refactorings according to their quality. (2) We conducted a series of code comprehension experiments ($E_2$) where we studied whether the refactoring code changes generated by tool are more difficult to understand than those generated by developers. In $E_1$, we surveyed 80 developers using code from 10 different Java systems \reviewed{(two study groups, manual and automated code changes)}. In $E_2$, we hired 30 more additional developers, from the two aforementioned crowdsourcing marketplaces, to perform two comprehension tasks covering three out of the four categories of comprehension questions identified by~Sillito~{et al.}~\cite{sillito2008asking}. We measured the subjects' performance using: (1) the NASA task load index for their effort; (2) the time that they spent performing the tasks; and, (3) their percentages of correct answers.\\
\parrafo{Results}
From the collected data, we observe that: (1) the ability of developers to recognize automatically-generated refactoring changes depends on the types of anti-patterns removed. For example, automatically generated refactoring changes to remove Blob, Spaghetti Code and Lazy Class anti-patterns are hard to distinguish, while those generated for correcting Long Parameter List and Speculative Generality anti-patterns are not. (2) Developers do not have preference between refactoring changes generated by humans or by machine (i.e., automated tools), except for those that correct Blob classes. (3) There is no statistically significant difference between the impact of automated- and manual- refactorings on program comprehension.
\parrafo{Relevance}
Understanding the context in which automated-refactoring is beneficial to developers and the cases where a human supervision is required is crucial from the points of view of both researchers and practitioners. For researchers, our results debunk the myth that automated-refactoring is not reliable and-or cannot be safely performed without human intervention. At the same time, they also highlight the need to develop approaches able to create identifiers that are semantically and contextual dependent. For practitioners, our results build confidence in the adoption of automated-refactoring tools. We hope that our results serve as inspiration to both groups, and help them develop new tools and approaches to support software maintenance and evolution.\\
\parrafo{Organization}
Section~\ref{sec:relatedwork} relates our study with previous works. Section~\ref{sec:experimentalDesign} describes the design of our empirical study. Section~\ref{sec:results} presents the study results, while Section~\ref{sec:discussion} discuss further the obtained results and their implications. Section~\ref{sec:threats} discusses threats to validity. Finally, Section~\ref{sec:conclusions} concludes the paper and discusses avenues for future work.
\section{Related Work}
\label{sec:relatedwork}
Fowler~\cite{fowler1999refactoring} popularized the term refactoring, by defining some heuristics to improve the design of existing code while preserving its functionality. Brown~\cite{Brown98-AntiPatterns} introduced the notion of anti-patterns as poor-design choices that hinders code evolution. Opdyke~\cite{opdyke1992refactoring} is the first to formulate a set of pre- and post-conditions to automatize the refactoring of object-oriented systems. Mens~{et al.}~\cite{mens2004survey} published an extensive overview of existing research work of software refactoring, and there are still new works published every year.
In this section, we summarize some of the works related to refactoring and its impact on design quality and code comprehension, and relevant studies that compared manual and automatic refactoring.
\subsection{Impact of Refactoring anti-patterns on design quality and code comprehension}
Deligiannis~{et al.}~\cite{DELIGIANNIS2004129} proposed the first quantitative study of the impact of anti-patterns on software development and refactoring. Through a controlled experiment involving 20 students and two systems, they found that Blob classes hinder the evolution of software design and the subject's use of inheritance. This work did not evaluate the understandability of the code, neither the subjects' ability to successfully perform comprehension tasks on the systems studied.
Stroulia and Kapoor~\cite{xing2006refactoringPractice} investigated the impact of refactoring on metrics associated to size and coupling and found that those metrics improved after refactoring.
In another academic setting, Du Bois~{et al.}~\cite{Bois2006DoesGC} found that the refactoring of God Classes into a number of collaborating classes can improve code understandability. The participants were asked to perform simple refactoring to decompose God classes. They found that participants exhibited less difficulties to understand the refactored code.
Murphy~{et al.}~\cite{murphy2012HowWeRefactoring} performed an empirical study on refactoring practice and the use of refactoring tool support. By analyzing the commit history of more than $39,000$ Eclipse developers, with their interaction traces (using Mylyn plug-in), they found that: (1) refactoring tooling support is rarely used; (2) that 40\% of refactorings assisted by tools occur on batch; (3) developers prefer interleaving refactoring with other code activities; (4) the kind of refactorings performed with tools differs from the kind of refactorings performed manually. However, this work did not study the impact on code comprehension of refactorings performed by machine and human, and what is the take of developers on automatically refactored code. These questions that were left unanswered served as motivation to perform our study.
Abbes~{et al.}~\cite{5741260} performed an empirical study to determine the impact of Blob and Spaghetti Code anti-patterns on program comprehension. They performed three experiments, each of them with 24 subjects and three different Java systems. They measured the subjects performance using three metrics: NASA TLX, tasks' completion times, and percentage of correct answers. They found that the occurrence of one single instance of anti-pattern does not impact significantly the comprehensibility of the code, while the combination of two anti-patterns does impact comprehension negatively.
Moser~{et al.}~\cite{moser2008case} analyzed the impact of refactoring on productivity in agile environments. They measured developer's productivity by dividing lines of code and effort (measured in time). They reported a statistically significant increase in productivity after refactoring along with some improvements in complexity and coupling.
Kim~{et al.}~\cite{6802406} performed a case study in an industrial setting at Microsoft on the benefits of refactoring. They found that developers perceive manual refactoring as expensive and sometimes risky, and that refactoring changes loosely match the literature definition of semantics-preserving code transformations. They also reported that the top 5 percent of preferentially refactored modules in Windows 7 reported higher reduction in the number of inter-module dependencies and several complexity metrics, but as a consequence larger increase in module's size compared to the remaining 95\% of the refactored modules.
All these studies corroborate the idea that refactoring anti-patterns improve code design quality in several dimensions. However, all of them focused on manual refactoring.
\subsection{Manual compared to automatic refactoring studies}
Negara~{et al.}~\cite{10.1007/978-3-642-39038-8_23} performed an empirical study to compare semi-automatic (IDE tool-assisted) refactoring and manual refactoring operations of 23 participants from the academy and industry in a controlled experiment. They reported that developers performed 11\% more refactorings manually than using tool support; less experienced developers use less tool support than experienced ones; that developers perform large refactoring changes ({\textit{e.g.,}}~extract method) with and without tool support, and that the size of a refactoring is not a decisive factor.
Sz\H{o}ke~{et al.}~\cite{7332494} performed a case study with a R\&D company to study the effects of semi-automatic refactoring on code maintainability. They corroborate the statement by Murphy~{et al.}~\cite{murphy2012HowWeRefactoring} that the quality of machine-generated code might differ from manual refactoring when developers are prompted to provide input to the tool to decide between a list of refactorings, and when they blindly select the default machine propositions. They conclude that companies could achieve a considerable increment of code maintainability by only applying automatic refactorings. Note that they did not compare an equal number of manual and automatic refactorings, but commanded developers to perform refactoring manually and then developed a tool to reproduce manual refactoring behavior followed by developers. The level of automation of their proposed tool is not disclosed.
Note that none of the studies in this category compared the understandability of manual and automatic refactoring on the same source code, and--or developers' perceived quality of the refactored code.
\subsection{Evaluation of Refactoring process}
Kataoka~{et al.}~\cite{kataoka2002quantitative} proposed an approach to measure the effect of refactorings on the maintainability of the code, using
coupling metrics. Although, they find that coupling metrics can be useful to asses the degree of maintainability improvement of a refactoring change, they also recognized that the type of refactorings that can be evaluated using coupling metrics is limited,
In a recent preliminary work by Arima~{et al.}~\cite{ArimaERANaturalness}, a novel technique is proposed to assess the \emph{naturalness} of refactored code automatically, using probabilistic language models.
\emph{Code naturalness} is a numerical value that measures how natural a given word sequence is for the model. As oracle they used a curated dataset of commits belonging to JUnit, where the refactorings were clearly identified by the authors of the commits. Their approach achieved 68\% of accuracy on 28 refactoring operations. This study does not evaluate the impact of refactoring, but tried to propose a new metric to quantitatively measure how well does a refactoring matches the text content of a software system. In the future, this technique could serve for further comparisons of automatically and manually refactored code in a quantitative way, and to improve existing automatic refactoring tools.
\section{Experimental Design}
\label{sec:experimentalDesign}
We perform two experiments to assess the comprehension of source code by developers in the presence of five types of object-oriented anti-patterns: Blob (BL), Lazy Class (LC), Long Parameter List (LP), Spaghetti Code (SC) and Speculative Generality (SG). We chose these anti-patterns because they are are representative of poor design choices . In fact, Palomba et al.~\cite{palomba2014they} found in a case study with developers from both industry and academia, that Blob class and Spaghetti Code are highly recognized and considered high severity design problems; Speculative Generality was perceived as problem of medium severity; Long-parameter list and Lazy class were considered low severity problem. The rational to study Long-parameter list is that this anti-pattern type is deemed to affect code readability, and Lazy class to bloat code design unnecessarily. Additionally, previous works have proposed approaches to detect the anti-patterns studied in this work~\cite{moha2010decor,marinescuDetection}. Experiment 1 is about (1) recognizing if a refactoring code change applied to remove an instance of the aforementioned anti-pattern types was performed by a developer, or an automatic tool. We also asked participants to rank the refactoring code changes according to their perceived quality.
In Experiment 2, \reviewed{using a different pool of participants, we present them with code refactored by either developers or an automated tool}, and ask them to complete some comprehension tasks on the refactored code entities. The aim is to assess the comprehensibility of code after refactoring.
In each experiment, we selected two instances of each anti-pattern type studied: one \textit{easy} and one \textit{difficult}. The level of difficulty is decided based on several object-oriented metrics, based on each anti-pattern type.
\subsection{Research Questions}
Our research questions stem from our goal of understanding the impact of automated refactoring on developer's comprehension. We state them as follows.\\ \emph{RQ1: \RQone} \\
\emph{RQ2: \RQtwo} \\ \emph{RQ3: \RQthree}
\subsection{Hypotheses}
For RQ1, we test the following null hypothesis when subjects review refactored code changes.
\begin{myenumi}
\item [$H_{0identify:}$] Developers correctly differentiate the refactoring changes generated by an automated tool from the ones generated by another developer.
\end{myenumi}
For RQ2, to examine differences in the perceived quality of the refactoring changes, we test the following null hypothesis.
\begin{myenumi}
\item[$H_{0rank:}$] Developers prefer manual refactoring over automated refactoring.
\end{myenumi}
RQ3 is related to subjects performing code comprehension tasks on refactored code. We study the performance of the subjects along three dimensions: time to execute the comprehension task, effort measured using NASA task load index (TLX)~\cite{sharek2011useable}, and percentage of correct answers.
We test $H_{0performance:}$ There is no difference between the performance of developers performing comprehension tasks on code refactored by developers compared to automatically refactored code.
\subsection{Objects}
\reviewed{The selection criteria of the anti-pattern instances refactored in this study is based on the detection results of our own tool for detecting and refactoring anti-patterns RePOR~\cite{RePOR_JSS_symred} applied to the SF110 corpus~\cite{TOSEM_evaluation}. SF110 is a representative sample of 110 Java projects from \emph{SourceForge} Website\footnote{https://sourceforge.net/}, which is a popular open source software repository, with more than 500,000 projects and with more than 33 million registered users. The projects in the SF110 corpus are packed into a common build infrastructure, including developers tests suites and automated-generated test suites previously validated by the authors of the corpus. This allows us to validate, to some extent, that a regression is not introduced in the code after refactoring either by our approach or by developers, when generating the objects of our experiment, and to have a common framework to import and build the systems (SF110 uses Ant build system to automate the building process).}
\reviewed{Once we collected the results of the anti-patterns detection, the first author of this work and a Master's student in our research lab selected two instances, for each anti-pattern type studied, that were deemed representative examples of anti-patterns based on the works of Brown~{et al.}~and Fowler~{et al.}~\cite{Brown98-AntiPatterns, fowler1999refactoring}. In case of disagreement, we came to the second author of this work for solving differences, and followed a conservative approach of discarding instances where we could not reach consensus.
We are also interested to know if different refactoring types affect the perception of developers or not. However, study only one single instance of each type would be insufficient as one system could be intrinsically easier or more difficult to refactor. Hence, we opted for selecting one easy and one hard system for each anti-pattern type. In total we study 10 Java systems (two for each anti-pattern type) from different domains and sizes.
We select the anti-pattern examples from different systems for both treatments (automated and manual), to control for possible learning effect on the respondents, who might get familiar with the system's code design. }
We decide on the level of difficulty of the systems based on different metrics that reflect effort required to perform the refactoring. In the case of the Blob, we measure the size of the Blob class (lines of code) and the number of \emph{data classes} associated to the Blob class; in this work, a \emph{data class} is a class composed mainly by attributes, and the only methods declared are accessors (\emph{getters} and \emph{setters}). With respect to Collapse Hierarchy and Lazy Class, we measure the \emph{number of incoming invocations (NII)}, as we suggest that removing these classes requires to update all method calls to the inlined classes, which in turn require extra effort when values of NII are large. For Spaghetti Code, we measured \emph{Mccabe Complexity (CC)}. Finally, for Long-parameter list classes, we consider the number of parameters.
The anti-patterns types studied are briefly introduced in~\Cref{table:defects}. We provide the name, a brief description and the refactoring strategies used to remove them according to the literature of anti-patterns~\cite{fowler1999refactoring,Brown98-AntiPatterns}.
\begin{table}[t]
\centering
\renewcommand{\arraystretch}{1.0}
\renewcommand{\tabcolsep}{.5mm}
\caption{List of studied Anti-patterns types and the refactorings used to correct them.}
\label{table:defects}
\scriptsize
\begin{tabular}{|p{0.2\columnwidth}|p{0.43\columnwidth}|p{0.34\columnwidth}|}
\hline
Name &\ Description & Refactoring(s) strategy\tabularnewline
\hline
Blob (BL)~\cite{Brown98-AntiPatterns} & A large class that absorbs most of the functionality of the system with very low cohesion between its constituents. In addition, Blob Classes are surrounded by classes who serve mainly as data holders, and that does not implement any functionality ({\textit{a.k.a.,}}~data classes). & \textit{Move Method (MM)}. Move the methods that does not seem to fit in the Blob class abstraction to more appropriate classes~\cite{sengSB06}. Another strategy, when there are not suitable classes to move methods, is the creation of new classes with methods and attributes that have high cohesion, and that are semantically related ({\textit{a.k.a.,}}~extract class refactoring).\tabularnewline
\hline
Lazy Class (LC)~\cite{fowler1999refactoring} & Small classes with low complexity that do not justify their existence in the system. & \textit{Inline Class (IC)}. Move the attributes and methods of the LC to another class in the system.\tabularnewline
\hline
Long Parameter List (LP)~\cite{fowler1999refactoring} & A class with one or more methods having a long list of parameters, specially when two or more methods are sharing a long list of parameters that are semantically connected. & \textit{Introduce Parameter Object (IPO)}. Extract a new class with the long list of parameters and replace the method signature by a reference to the new object created. Then access to this parameters through the parameter object. \tabularnewline
\hline
Spaghetti Code (SC)~\cite{Brown98-AntiPatterns} & A class without structure that declares long methods without parameters.
& \textit{Replace Method With method Object (RMWO)}. Extract long methods into new classes so that all local variables become attributes on that object.\tabularnewline
\hline
Speculative Generality (SG)~\cite{fowler1999refactoring} & There is an abstract class created to anticipate further features, but it is only extended by one class adding extra complexity to the design. & \textit{Collapse Hierarchy (CH)}. Move the attributes and methods of the child class to the parent and remove the \textit{abstract} modifier. \tabularnewline
\hline
\end{tabular}
\end{table}
\reviewed{In~\Cref{table:systems} we present information about the systems studied. The ID column contains the ID assigned by SF110 (we provided it as reference) to a system, system's name, description, use of synthetic identifiers for new code entities (S.N.), number of classes in the system (N.C.), anti-pattern type (Ap. Type), Lines of Code of source class containing the anti-pattern instance (LOC), level of difficulty according to the authors assessment (i.e., the Difficulty column), and the normalized static entropy (N.E.)~\cite{5070510} of the refactoring change for machine (N.E.M) and human patches (N.E.H). S.N. column can take any of this values: A (all refactoring examples), H (just human example), and N (none of the examples). The normalized static entropy is computed using Equation 1.}
\begin{align}
N.E.(code \ change) = -\sum_{k=1}^{n}\left(p_{k} {\ast} \log_{n}p_{k}\right),
\end{align}
where $p$ is the probability of changing a file $k$, with $p_k\geq 0, \forall k \in 1,2,\ldots,n$ and $\sum_{k=1}^{n} p_{k}=1$.
\reviewed{We measure $N.E.$ to quantify the effort required to refactor an anti-pattern based on the number of changes made by a developer (respectively a tool). From \Cref{table:systems} we observe that the refactoring effort as measured with $N.E.$ is very similar between machine and human treatments for most cases.}
\begin{table}[t]
\centering
\renewcommand{\arraystretch}{1.0}
\renewcommand{\tabcolsep}{.5mm}
\caption{List of systems from where we extract the anti-pattern's instances studied }
\scriptsize
\label{table:systems}
\begin{tabular}{|l|p{0.1\textwidth}|p{0.25\textwidth}|l|r|p{0.1\textwidth}|r|l|r|r|}
\hline
ID & System & Description &S.N. & N.C. & Ap. Type & LOC & Difficulty & N.E.M. & N.E.H.\\ \hline
110 & FireBird & A relational database manager &H &158 & Blob & 27 & Easy & 0.94 & 0.94 \\ \hline
83 & Xbus & It is a central Enterprise Application Integration (EAI) &H & 159 & Blob & 595 & Hard & 0.89 & 0.93 \\ \hline
52 & Lagoon & A XML-based framework for web site maintenance & N &53 & Collapse hierarchy & 32 & Easy & 0.97 & 0.97 \\ \hline
86 & Advanced T-Robots & Fighting robots Arena &N &177 & Collapse hierarchy & 25 & Hard & 0.92 & 0.92 \\ \hline
47 & dvd-homevideo & Application to burn DVDs &N &2 & Lazy class & 24 & Hard & 0.95 & 0.94 \\ \hline
101 & SAP NetWeaver & Server Adapter for Eclipse &N & 154 & Lazy class & 46 & Easy & 0.88 & 0.89 \\ \hline
103 & Sweet Home 3D & An interior design application &A & 2 & Long-parameter List & 712 & Hard & 0.86 & 0.89 \\ \hline
104 & Vuze & BitTorrent Client &A &1740 & Long-parameter List &187 & Easy & 0.85 & 0.88 \\ \hline
81 & JavAthena & Online role playing game & A &31 & Spaghetti Code & 808 & Easy & 0.44 & 0.99 \\ \hline
70 & EchoDep & Digital preservation application &A &64 & Spaghetti Code & 645 & Hard & 0.59 & 0.62 \\ \hline
\end{tabular}
\end{table}
Our benchmark of refactoring changes is comprised of two sets: 10 \emph{high-level} refactorings applied automatically by our approach (\emph{M}); and 10 \emph{high-level} refactorings applied by real developers (\emph{D}). A {high-level} refactoring is a refactoring composed of more than one {low-level} refactoring~\cite{opdyke1992refactoring}. For example: Collapse Hierarchy, which is a {high-level} refactoring, is composed of the following {low-level} refactorings: one or more pull-up method/attributes; delete class; remove abstract modifier; update class references.
For set \emph{M}, we refactored the selected anti-patterns' instances, using the Eclipse plug-in implementation of RePOR\footnote{\url{https://github.com/moar82/RefGen}}. For set \emph{D}, we hired five experienced Java developer freelancers from two well-known marketplace websites (Guru.com and Freelancers.com). We provide them with the definition of the studied anti-patterns and the Fowler's catalog of refactorings~\cite{fowler1999refactoring}, to allow them to select the most adequate refactorings, according to their experience. \reviewed{One can argue that it would be more natural to collect refactorings from the original developers of the studied systems, as they are familiar with the code. However, it would be hard to first locate, and then convince the original developers from open-source systems, as most of contributors of open-source projects are volunteers, without any contractual relationship with the project. Instead, we assume an hypothetical scenario where a new developer is integrating to a project, and is assigned the task of refactoring the existing code.}
\reviewed{To ensure that the refactoring changes written by freelancers does not alter the original behavior of the code in question, we asked them to validate that the project's unit tests provided in the benchmark pass and, if necessary, that they perform any pertinent update the unit tests to be valid. For example, if the refactoring change implies moving a method, it is necessary to update the unit test method to point to the new location of the refactored method.
We checked that unit tests pass after refactoring to have certain confidence that no code regression was introduced on the systems studied.}
\reviewed{Each freelancer performed two refactorings, and they were excluded from experiment 1 and 2 to avoid introducing bias in our results.}
\subsection{Subjects}
\label{subsec:subjects}
We invited participants to participate in experiment $E_1$
through Java developers' mailing lists, and through the social network of the authors of this paper. In total we received 87 answers from which we filtered out responses that were incomplete, and those where participants who did not provide justifications for their choices. We ended up with 80 responses for this study. All participants were volunteers and could withdraw at any time from the study, for any reason. From the 80 responses, 57 participants declared software development as their main occupation, 7 participants declared their main occupation to be research, 6 work as software architects, 4 scrum masters, 4 students and 2 fall in the \emph{Other} category. Among these 80 participants, 37 declared to work on both open-source and proprietary systems, 34 declared to work on proprietary systems and only 9 on open-source. Fifty percent (40 out of 80) of participants declared to have more than 5 years of experience as developers. Eighteen participants have more than 2 and up to 5 years of experience, 14 participants have between 1 to 2 years of experience, and 8 participants have less than one year of experience.
For $E_2$, we hired 30 more additional Java developers from aforementioned two marketplace websites. To select a candidate, we perform an informal interview with the candidates, and we set a minimum of experience of one year developing Java software for the industry. To control for possible confounding factors, we registered the number of years of experience of each participant as professional developers, and asked them to provide their Java's level of confidence (using a Likert scale from 1 to 5).
\subsection{Independent variable}
The independent variable for the two experiments is related to who refactored an anti-pattern instance, i.e., its origin, and it is a binary variable stating whether the refactoring was performed by an automatic tool or by a developer.
\subsection{Dependent Variables}
In $E_1$, the first dependent variable is a binary variable stating whether or not the participant correctly identified the origin of the refactoring change. The second dependent variable is a categorical variable capturing the perceived quality of the refactoring solution proposed, on a scale of 1 to 5, where 1 is for ``Poor'' quality, and 5 is for ``Outstanding'' quality.
In $E_2$, the dependent variables measure the subjects' performance, in terms of effort, time spent, and percentage of correct answers.
We measure a subject's performance using the NASA Task Load Index (TLX). TLX evaluates the subjective workload of subjects. It is a multidimensional measure that provides an overall workload index based on a weighted average of ratings on six sub-scales: mental demands, physical demands, temporal demands, own performance, effort and frustration. We combine weights and ratings provided by the subjects into an overall weighted workload index by multiplying ratings and weights; the sum of the weighted ratings divided by fifteen (sum of the weights) represents the effort~\cite{sharek2011useable}. To measure the time that participants spent on each task, we recorded the participants' remote session on the virtual machine where participants performed their assignment. The time reported only considers the time spent on the comprehension task, from the moment they open the project in the IDE, until they close it. We compute the percentage of correct answers for each question by dividing the number of correct elements found by a participant by the total number of correct elements (s)he has found. For example, if the question requires to find the total number of code references for a given object, and there are ten references but the subject finds only four, the percentage of correct answers is forty for that question.
\subsection{Questions}
For $E_1$, we use an online survey system (Jotform\footnote{http:www.jotform.com}) where we present the refactoring change, the online link to repository of the system that contained the anti-pattern's instance, and the type of anti-pattern. To summarize the refactoring changes applied to the studied systems, we generate for each anti-pattern's instance refactored, a patch file using the diff command from the control version system (Git). A patch file contains the description of the changes made using diff notation, which is a unified format that developers and control version systems can understand. The link to the repository of the systems studied provide respondents with a complete reference of the refactoring changes applied. We also provide a link to the repository containing the original source code; for example, if they want to study deeper the impact of the refactorings applied, they could clone the repository and apply the patch. In Listing~\ref{lst:refactoring-patch} we show a fragment of a refactoring change from our online survey.
\lstinputlisting[,language=Java, caption=Fragment of a refactoring change from the online survey.,label={lst:refactoring-patch},escapeinside={(*@}{@*)},basicstyle=\scriptsize,numbers=left,xleftmargin=15pt,breaklines=true]{ec81.tex}
In diff notation, a patch does not show the complete file, but only shows the code fragments that were modified. Those code fragments are called chunk. The lines starting with "+" indicate new lines added, and "-" lines removed. "@@" is the chunk header, where Git indicates which lines were affected. For example, line 5 in Listing~\ref{lst:refactoring-patch} indicates that from file A (original source code represented by a "-"), 6 lines are extracted starting from line 6. From file B (refactored source code represented by a "+" ), 7 lines are displayed, starting also from line 6 . The text after "@@" serves to clarify the context. Git tries to display a method name or other contextual information of where this chunk was taken from in the file.
We asked developers to answer whether the refactoring change was generated by a developer or a software tool. We also had an option ``Unknown'' that participants could select if they were unable to tell whether the refactoring change was performed by a developer or generated by a tool. To control for possible randomness in the answers, we asked participants to provide their level of confidence in their answers (on a scale of 1 to 5), and a brief explanation to support each answer. Since the quality of developers' solutions may be different from that of our automated approach, we asked participants to rate the quality of the refactoring changes (according to their perception of quality) and mark their preferred solutions. We also asked them to provide
any additional comment about the refactoring change's quality, if they considered it appropriate.
For $E_2$, we used comprehension questions to elicit comprehension tasks and collect data on the subject's performance. As in~\cite{5741260}, we consider questions in three of the four categories of questions regularly asked and answered by developers~\cite{sillito2008asking}: (1) finding a focus point in some subset of the classes and interfaces of some source code, relevant to a comprehension task; (2) focusing on a particular class believed to be related to some task and on directly-related classes; (3) understanding a number of classes and their relations in some subset of the source code; and, (4) understanding the relations between different subsets of the source code. Each category contains several questions of the same type.
We only selected questions in the first three categories, since the last category concerns different subsets of the source code, while in our experiments, we focus exclusively on one or two packages at most, that are affected by a particular anti-pattern. For each category, we choose the two most relevant questions through discussions between the first author and a Master's student intern who collaborated on this work. The decisions were validated by the second author. Selecting two questions for each category of question, which provided us with two data points from each participant. The six selected questions are the followings. The text in bold is a placeholder that we replace with the appropriate behaviors, concepts, elements, methods and types depending on the system on which the subjects performed their tasks.
\begin{itemize}
\item Category 1: Finding focus points:
\begin{itemize}
\item Question 1: Where is the code involved in the implementation of \textbf{this behavior}?
\item Question 2: Which type represents \textbf{this domain concept} or \textbf{this UI element or action}?
\end{itemize}
\item Category 2: Expanding focus points:
\begin{itemize}
\item Question 1: Where is \textbf{this method} called or \textbf{this type} referenced?
\item Question 2: What data can we access from \textbf{this object}?
\end{itemize}
\item Category 3: Understanding a subset of classes:
\begin{itemize}
\item Question 1: How are \textbf{these types or objects} related?
\item Question 2: What is the behavior that \textbf{these types} provide together and how is it distributed over \textbf{these types}
\end{itemize}
\end{itemize}
For example, with system \reviewed{\emph{83 Xbus} ({\textit{cf.,}}~\Cref{table:systems}), we replace ``\textbf{this behavior}'' in question 1, category 1, by ``manipulating and store the email received by class \texttt{POP3XMLReceiver}''.}
For category 2, we acknowledge that the questions might be answered by developers using the IDE search functionality. However, developers still must identify and understand the classes or methods that they consider relevant to the task. Additionally, discovering classes and relationships that capture incoming connections prepare the developers for the questions of the third category. \reviewed{Below we present the comprehension task for system \emph{83 Xbus}. } \\
\fbox{
\noindent\begin{minipage}[t]{\textwidth}
Answer the following comprehension questions related to the system Xbus.
We focus on the package \textbf{net.sf.xbus.technical.mail}.
\begin{enumerate}
\item Where is the code involved in manipulating and storing the email received by class \texttt{POP3XMLReceiver}?
\item Which type (class) defines a method for reading an email after registering its receiver in the Transaction manager?
\item Where is the type \texttt{POP3XMLReceiver} referenced?
\item What data can we access from the object \texttt{mEmailMessage}?
\item How are \texttt{Email} and \texttt{POP3XMLReceiver} related?
\item What is the behavior that \texttt{POP3XMLReceiver} and \texttt{Email} provide together and how is it distributed over these types?
\end{enumerate}
\end{minipage} }
\subsection{Anonymization of new code lexicon}
\label{subsec:anonymization}
Identifier names and comments (code lexicon), when thoughtfully assigned, can support developers to better understand a software system. However, as already stated, current existing automatic refactoring tools and frameworks are not equipped with mechanisms allowing them to generate a human-like name for a new class, or a new method introduced when refactoring. On the other hand, developers generate names that reflect the roles of the entities and--or follow projects guidelines. To ensure a fair experiment using the refactoring changes generated by humans and by RePOR, we decided to post-process the changes by removing any code comment, and renaming any new class or method added to the code base with an artificial name (for both origins), to avoid providing any hint that can lead the human evaluators to discover the origin of the changes. Note that renaming new classes and methods was only necessary for the following refactoring types: Introduce parameter-object, Replace method with object, and Extract class, while Inline Class, Collapse Hierarchy, and Move method did not require it.
\subsection{Design}
For $E_1$, we divided the 20 refactoring patches (10 changes for each origin) into 4 groups. So each respondent will answer a survey containing 5 refactoring changes (each change corresponds to a different system) for removing 5 different types of anti-patterns. We also interleave the origin of the changes (machine, human) and the level of difficulty (easy and hard) in a way that any group has more than 3 patches from the same origin or level of difficulty. As previously stated, the level of difficulty is assessed using well known object oriented metrics. To assess feasibility, that time given to respond the survey was enough, and anticipate adverse events, we performed a pilot study with two Post-doctoral fellows and one Ph.D. student from our lab, with more than 5 years of experience developing with Java. The pilot study also helped us to refine the questions, and improve the visual design of our survey in terms of readability. Note that none of the people that participated in the pilot study took part to the final study.
In addition to asking respondents to identify the origin of the patch, we also asked them to justify their response in a free-text box, and to indicate how confident they felt when answering the questions, using a scale from 1 to 5, to control for possible randomness in the answers.
We present our design in~\Cref{table:e1design}. First column is the number of question and the rest of the columns corresponds to the different groups. Each cell contains the type of anti-pattern, using the abbreviations of Table~\ref{table:defects}, the ID of the system, and a letter indicating the origin of the patch (H:human, M: Machine).
\begin{table}[t]
\centering
\renewcommand{\arraystretch}{1.0}
\renewcommand{\tabcolsep}{.5mm}
\caption{$E_1$ Experimental Design}
\label{table:e1design}
\begin{tabular}{|l|l|l|l|l|}
\hline
Question & Group 1 & Group 2 & Group 3 & Group 4 \\ \hline
1 & SC-83-M & BL-83-H & BL-110\_M & LP-103-M \\ \hline
2 & CH-86-H & SC-70-H & SC-70-M & SC-81-H \\ \hline
3 & LC-101-M & CH-52-M & LP-104-H & LC-101-H \\ \hline
4 & BL-83-M & LP-104-M & LC-47-0 & BL-110-H \\ \hline
5 & LP-103-H & LC-47-H & CH-52-H & CH-86-M \\ \hline
\end{tabular}
\end{table}
For $E_2$, we use the same collection of refactoring examples from $E_1$, 5 anti-pattern types, and two instances for each type; each instance with two possibilities: being generated by human or by machine. That totals 10 refactoring changes for each origin.
We hired 30 freelancers, who completed two tasks each of them, leading to 60 comprehension tasks.
\subsection{Procedure}
All the data collected is anonymous. The subjects could drop the experiment at any time, for any reason and without penalty of any kind. For the freelancers hired for refactoring the systems, and the second group hired for answering the comprehension tasks, we set milestones which clearly specify work deliverables, so we pay for each task completed. All the freelancers hired for refactoring the anti-patterns studied passed an interview, where they stated their experience as Java developers, and refactoring, and correcting anti-patterns. In addition to the anti-patterns definitions and refactoring strategies, we provided them with references from the literature and web sites related to refactoring, but we did not persuade them to blindly follow any of these materials, but encouraged them to base their actions on their work experience and own reasoning.
For $E_1$, we first briefly introduced the description of the anti-patterns using~\Cref{table:defects}. Next, we asked them, to not base their judgment on code lexicon, and to accept that they will only focus on code structure and its quality, and not on indentations, naming conventions, space, etc.
For $E_2$, the subjects knew that they would perform comprehension tasks, but did not know the goal of the experiment nor whether the system was refactored by a developer or by an automated approach. We informed them of the goal of the study after they finished the experiment.
\subsection{Analysis method}
In RQ1, to attempt rejecting $H_{0identify:}$, we test whether the proportion of refactoring changes correctly identified (or not) by participants, significantly varies between changes generated by human or by machine. We use Fisher's exact test~\cite{sheskin2003handbook}, which checks whether a proportion vary between two samples. We also compute the odds ratio ($OR$)~\cite{sheskin2003handbook} that indicates the likelihood for an event to occur. The odds ratio is defined as the ratio of the odds $p$ of an event occurring in one sample, {\textit{i.e.,}}~the odds that changes generated by machine were correctly identified, to the odds $q$ of the same event occurring in the other sample, {\textit{i.e.,}}~the odds that changes generated by humans were correctly identified: $OR=\frac{p/(1-p)}{q/(1-q)} $. An odds ratio of 1 indicates that the event is equally likely in both samples. An $OR$ greater than 1 indicates that the event is more likely in the first sample (machine), while an $OR$ less than 1 indicates that it is more likely in the second sample (human).
In RQ2, we use a (non-parametric) Mann-Whitney test to compare the perceived quality ({\textit{i.e.,}}{} the rates assigned by participants) of refactoring changes generated by machine with the rates of refactoring changes generated by humans. Non-parametric tests do not require any assumption on the underlying distributions. Other than testing the hypothesis, it is of practical interest to estimate the magnitude of the difference between the rates assigned to changes generated by machine and humans. Therefore, we compute the non-parametric effect size measure Cliff's $\delta$ ($ES$)~\cite{cliff2014ordinal}, which indicates the magnitude of the effect of the treatment on the dependent variable. The effect size is considered negligible if $< 0.147$, small if between $0.147$ and $0.33$, medium if between $0.33$ and $0.474$, and large if $> 0.474$~\cite{romano2006exploring}.
For RQ3 and RQ4 we use Mann-Whitney test to compare two sets of dependent variables and assesses whether their difference is statistically significant. The two sets are the subject's data collected when they answer the comprehension questions on the systems refactored by either machine or humans. For example, we compute the Mann-Whitney tests to compare the set of times measured for each subject on the systems refactored by either machine or humans. We also compute the Cliff's $\delta$ ($ES$).
\section{Study Results}
\label{sec:results}
We now describe the collected data and present the results of our study, answering the four RQs formulated in~\Cref{sec:experimentalDesign}.
\subsection{RQ1: \RQone}
In~\Cref{table:rq1general} we summarize the results of $E_1$. In the first column we use the following abbreviations: \emph{all} for all the anti-pattern types studied, or the abbreviation of each type; columns 2 to 4 are the percentage of correct, wrong, or unknown answers; columns 5 to 7 are the corresponding percentages for machine changes (m), while
columns 8 to 10 are the corresponding percentages for human changes (h).
When considering all anti-pattern types, we observe that the proportion of correct and wrong answers are the same (45\%). In the remaining 10\% of cases, respondents were not able to discern the origin of the changes, {\textit{i.e.,}}{} \emph{I do not know} answer was selected, first row, columns 2-3. With respect to the anti-pattern's type fixed, Long-parameter and Spaghetti code have the highest percentage of machine changes correctly identified (more than 70\%), and therefore we consider that they could not pass by manual changes. Conversely, Speculative generality, Blob and lazy class were correctly identified in less than 50\% of cases, indicating that to some extent they mimic human behavior on refactoring tasks.
If we study the percentage of correct answers by refactoring origin, we observe that respondents found it more difficult to identify human patches (38\%), whereas machine patches were little easier to identify (52\%). This trend holds for all anti-patterns studied, except for Speculative Generality (SG), where the percentage of correctly identified machine changes (23\%) is lower than the correctly identified human ones (55\%). This result is surprising since the refactoring type applied for removing the two different SG instances, which is Collapse Hierarchy (CH), was applied for both treatments. To find a plausible explanation for this result, we manually examined the cases where respondents failed to distinguish refactoring changes corresponding to SG anti-pattern type, and observed that they belong to the easy instance, which corresponds to 14 different respondents who failed to identify the origin of the changes generated by the automatic tool to remove SG anti-pattern type. In general, human judges doubted the ability of software tools to perform this refactoring type. For example, one responded commented: \emph{``The abstract class is useless (the change is better). I don't think a tool can detect it.''}.
\begin{table*}[t]
\centering
\renewcommand{\arraystretch}{1.0}
\renewcommand{\tabcolsep}{.5mm}
\caption{$E_1$ RQ1 overall results}
\label{table:rq1general}
\begin{tabular}{|l|r|r|r|r|r|r|r|r|r|}
\hline
ap-type & correct & wrong & unknown & correct-m & wrong-m & unknown-m & correct-h & wrong-h & unknown-h \\ \hline
all & 45.00\% & 45.00\% & 10.00\% & 52.00\% & 38.50\% & 9.50\% & 38.00\% & 45.50\% & 16.50\% \\ \hline
bl & 30.00\% & 61.25\% & 8.75\% & 37.50\% & 57.50\% & 5.00\% & 22.50\% & 65.00\% & 12.50\% \\ \hline
lc & 47.50\% & 45.00\% & 7.50\% & 47.50\% & 47.50\% & 5.00\% & 35.00\% & 42.50\% & 22.50\% \\ \hline
lp & 57.50\% & 33.75\% & 8.75\% & 80.00\% & 12.50\% & 7.50\% & 35.00\% & 55.00\% & 10.00\% \\ \hline
sc & 50.00\% & 45.00\% & 5.00\% & 72.50\% & 22.50\% & 5.00\% & 42.50\% & 37.50\% & 20.00\% \\ \hline
sg & 45.00\% & 45.00\% & 10.00\% & 22.50\% & 52.50\% & 25.00\% & 55.00\% & 27.50\% & 17.50\% \\ \hline
\end{tabular}
\end{table*}
To control for some confounding factors, like development experience, or confidence when answering the questionnaire, we cluster the results based on developer's experience (Table~\ref{table:e1byexperience}) and developer's confidence for each patch reviewed (Table~\ref{table:e1byconfidence}).
We observed that the largest number of respondents declared to have more than five years of experience as Java developers, while the smallest group declared to have less than one year. Contrary to what one would expect, the group with more correct answers (51\%) was the group with one to two years of experience, followed closely by the group with more than five years of experience (48\%), indicating that the experience factor was not decisive to correctly identify the origin of the refactoring changes. The same situation occurs with the other groups, as the group with less than one year of experience correctly identified more refactoring changes than the group with more than two, up to five years of experience. \reviewed{We suggest that perhaps developers surveyed with less experience, are more skilled in Java and more versed in object-oriented design, and beside their little experience, they have worked on more challenging projects than their more experienced peers, or they simply have more experience on code refactoring.}
Concerning the number of correctly identified refactoring changes by origin (columns 5-6, Table~\ref{table:e1byexperience}), we corroborate what we observed in the overall results, respondents had more difficulty to identify refactoring changes generated by human than by machine. With one remarkable exception: developers with one up to two years of experience, who correctly identified more refactoring changes applied by human than by machine.
\begin{table}[t]
\centering
\renewcommand{\arraystretch}{1.0}
\renewcommand{\tabcolsep}{.5mm}
\caption{$E_1$ RQ1 results by respondent's expertise}
\label{table:e1byexperience}
\begin{tabular}{|l|r|r|r|r|r|}
\hline
Expertise (years) & Total & Correct & Correct \% & Correct-h& Correct-m \\ \hline
\textgreater{}5 & 200 & 96 & 48.00\% & 40 & 56 \\ \hline
\textgreater{}2 to 5 & 90 & 32 & 36.00\% & 12 & 20 \\ \hline
1 to 2 & 70 & 36 & 51.00\% & 20 & 16 \\ \hline
\textless{}1 & 40 & 16 & 40.00\% & 4 & 12 \\ \hline
\end{tabular}
\end{table}
\begin{table}[t]
\centering
\renewcommand{\arraystretch}{1.0}
\renewcommand{\tabcolsep}{.5mm}
\caption{$E_1$ RQ1 results by respondent's confidence per question}
\label{table:e1byconfidence}
\begin{tabular}{|l|r|r|r|r|r|}
\hline
Confidence & Total & Correct & Correct \% & Correct-h & Correct-m \\ \hline
5 & 109 & 48 & 44.04\% & 24 & 24 \\ \hline
4 & 142 & 69 & 48.59\% & 29 & 40 \\ \hline
3 & 105 & 46 & 43.81\% & 18 & 28 \\ \hline
2 & 30 & 12 & 40.00\% & 3 & 9 \\ \hline
1 & 14 & 5 & 35.71\% & 2 & 3 \\ \hline
\end{tabular}
\end{table}
With respect to confidence level declared per question, we observed that respondents felt confident enough, as in 60\% of the questions they selected a confidence level between 4 and 5. We also observed that the largest proportion of correct answers (48.59\%) corresponds to the confidence level of 4, followed by the ones identified with a confidence level of 5 (44.04\%); the next group are respondents with a confidence level of 3 (43.81\%); refactoring changes identified with a confidence level of 2, reached 40\% of correctly identified patches; the last group, the one with the lowest level of confidence, achieved the smallest proportion of correct answers (35.71\%), which makes sense.
If we analyze results based on the origin of the refactoring changes (columns 5-6), we observe that the machine's ones are more easily identified than the human ones, with one exception: those changes identified with a confidence level of 5 have the same proportion of correct answers. Contrary to the analysis based on respondents' expertise, the analysis based on respondents' level of confidence is more linear,
as higher level of confidence led to higher number of correctly identified refactoring changes, in almost sequential order (the only exception being between levels 4 and 5, where respondents with confidence level of 4 achieved better results than those with confidence level of 5). \reviewed{Respondents that selected confidence level of 4 may have been more modest than those who selected 5. Still both intervals are on the top of the confidence scale.}
In Table~\ref{table:fisher} we present the contingency table for the number of correctly/incorrectly identified refactoring changes with respect to their origin: machine (m), human (h) and the results of the Fisher's exact test and odds ratio $OR$ between changes generated by human or machine. The contingency table shows the frequency distribution of the refactoring changes identified by the respondents with respect to the refactoring change's origin. Fisher's exact test indicates whether a significant difference of proportions between automatically- and manually-generated refactoring changes exists. Odds ratio indicates the probability of a respondent to correctly identify a change according to the origin of the refactoring change analyzed.
The first column of Table~\ref{table:fisher} corresponds to the anti-pattern's type; columns 3 to 4 correspond to the number of correctly/incorrectly identified changes by origin; and columns 5 to 6 reports the result of Fisher's exact test and $OR$s when testing $H_{0identify}$. In the first row, when aggregating all refactoring types together, the $p-value$ of the Fisher's exact test is statistically significant, with an $OR$ greater than one, indicating that changes generated by machine were easier to identify than the ones generated by humans. Aggregating all anti-pattern types might not be very descriptive, specially considering that the refactoring types studied cover a wide range of design problems (coupling, lack of cohesion, design size, readability, etc).
To make a more precise analysis of the results, we expand our analysis to consider the refactorings of each anti-pattern type separately (rows 2 to 6, Table~\ref{table:fisher}). We observe that only in two anti-pattern types, the results are statistically significant. Refactoring changes to remove long-parameter list have higher probability to be correctly identified by respondents when they are automatically generated (7 times more according to $OR$). Conversely, refactoring changes to correct Speculative Generality classes are more likely to be correctly identified when they are refactored manually ($OR<1$). Based on some respondent's comments, we suggest that the results obtained for Long-parameter list reflect respondent's view that the refactorings performed did not improve the quality of the systems. Because of this perception of poor quality, the respondents may have considered that such refactorings could have only been produced by an automatic tool.
This results suggests that the human judges considered the introduce parameter-object applied by both, developers and machine, artificial. For example, Table~\ref{table:rq1general} shows that the largest proportion of correct answers for an anti-pattern type correspond to long-parameter list with 80\% of correct answers; conversely, 55\% of respondents incorrectly chose machine origin when the origin was human; the remaining 10\% of respondents declared themselves unable to decide (they did not know).
For Speculative Generality anti-pattern, respondents tend to attribute the refactoring changes proposed for removing this type to humans, and some of them even mentioned that the refactoring proposed was too complex to be generated by an automatic tool. Note that existing popular Java IDEs like Eclipse and IntelliJ IDEA provide refactoring support for Long-parameter list (introduce-parameter object refactoring) but not for Speculative Generality ({\textit{e.g.,}}~Collapse hierarchy refactoring). Hence, they could have been misled by the wide availability of automatic refactoring tool support for refactoring long-parameter list anti-pattern when making their choices.
For the remaining anti-patterns types, the $p-values$ are $> 0.01$. Hence, we cannot reject $H_{0identify}$ for those types of anti-patterns. We therefore conclude that developers cannot differentiate between refactorings changes made by human and machine for those types of anti-patterns.
\npdecimalsign{.}
\nprounddigits{2}
\begin{table}[t]
\centering
\renewcommand{\arraystretch}{1.0}
\renewcommand{\tabcolsep}{.5mm}
\caption{Contingency table and Fisher's exact test results for developers' survey on refactoring changes}
\label{table:fisher}
\begin{tabular}{|l|r|r|r|r|r|r|}
\hline
ap\_type & correct-m & wrong-m & correct-h & wrong-h & $p-value$ & $OR$ \\ \hline
all & 104 & 96 & 76 & 124 & \textbf{<0.01} & 1.77 \\ \hline
bl & 15 & 25 & 9 & 31 & 0.22 & 2.05 \\ \hline
lc & 19 & 21 & 14 & 26 & 0.36 & 1.67 \\ \hline
lp & 32 & 8 & 14 & 26 & \textbf{<0.01} & 7.22 \\ \hline
sc & 29 & 11 & 17 & 23 & 0.01 & 3.51 \\ \hline
sg & 9 & 31 & 22 & 18 & \textbf{<0.01} & 0.24 \\ \hline
\end{tabular}
\end{table}
\hypobox{In general, automatically generated refactorings and refactoring changes made by humans were equally difficult to identify. In 10\% of cases, developers couldn't even make a decision and opted for the \textit{``I do not know''} option. Based on anti-pattern types, the results vary depending on the type. Refactorings generated for removing Long-parameter list and Speculative Generality anti-patterns failed the Turing test, whereas Blob, Lazy class and Spaghetti Code anti-patterns passed it.}
\subsection{RQ2: \RQtwo}
In this research question, we examine developers' perception of automated refactorings. We want to know whether they have the same appreciation for automatically generated refactorings and manual refactorings. The absence of statistically significant difference between the two treatments could serve as evidence that automated refactoring can be interleaved with manual changes without affecting the quality of a system.
\begin{table}[t]
\centering
\renewcommand{\arraystretch}{1.0}
\renewcommand{\tabcolsep}{.5mm}
\caption{$E_1$, RQ2 Median rates by anti-patterns' type}
\label{table:user_rates}
\begin{tabular}{|l|r|r|r|l|}
\hline
Antipattern Type & rate\_machine & rate\_human & $p-value$ & $ES$ \\ \hline
all & 3/5 & 3/5 & 0.02 & negligible \\ \hline
bl & 3/5 & 4/5 & \textbf{\textless{}0.01} & medium \\ \hline
lc & 3/5 & 4/5 & 0.13 & small \\ \hline
lp & 3/5 & 3/5 & 0.65 & negligible \\ \hline
sc & 3/5 & 3.5/5 & 0.10 & small \\ \hline
sg & 4/5 & 3/5 & 0.38 & negligible \\ \hline
\end{tabular}\end{table}
In~\Cref{table:user_rates} we present the respondent’s median rates for the refactoring changes presented in $E_1$, \reviewed{and in~\Cref{fig:user_rates} the boxplots of user's rates distribution. }
\begin{figure*}[tp]
\centering
\captionsetup{justification=centering}
\includegraphics[scale=0.45]{rates_by_ap_type.pdf}
\caption{User's rates distribution. }
\label{fig:user_rates}
\end{figure*}
Respondents input their rate for each refactoring change using a Likert-scale from 1 to 5. When considering all anti-patterns types, both sources of change (human, machine) attained the same rate. With respect to anti-pattern's type, the median rates for human changes are higher (0.5 to 1 point) than the median rates of machine changes (Blob, Lazy class and Spaghetti code types). Conversely, the machine refactoring changes to remove Speculative Generality are rated higher than the changes generated by humans to remove the same anti-pattern type. Automatic and manual refactoring changes for removing Long-parameter list anti-patterns achieved the same rate.
We apply the Mann-Whitney test to determine if the results are statistically significant, and we find that only refactoring changes that remove Blob instances achieve statistical significance, with a $p-value$ less than 0.01, and a medium Cliff's $\delta \ ES$. Hence, we reject $H_0rank$ for all anti-patterns types except the Blob.
\hypobox{We conclude that for the anti-patterns types studied, developers do not find manual refactoring changes to be of better quality than automated refactoring changes. The exception being the Blob type.
\subsection{RQ3: \RQthree}
In this research question, we are interested to know the impact on understandability of refactoring changes generated by human and machine.
We now present the results of $E_2$. \Cref{table:comprehension} summarizes the median of collected data, \reviewed{and \Cref{fig:comprehension} presents the overall performance of participants on the comprehension tasks, in terms of: time, effort, and percentage of correct answers. In addition, we divide the refactoring changes in two groups (see rows 4 to 9 of \Cref{table:comprehension}), i.e., those who deceived human judges in $E_1$ (T. passed), and those who failed the test (T. failed).} In columns 5, and 6, we report the $p-value$ of the Mann-Whitney tests and the Cliff's $\delta \ ES$ obtained for each studied dimensions.
\begin{figure*}[tp]
\centering
\captionsetup{justification=centering}
\includegraphics[scale=0.45]{comprehension_task_results_by_dimension_crop.pdf}
\caption{Comprehension tasks' performance results distribution.}
\label{fig:comprehension}
\end{figure*}
We observe that when considering all refactoring changes, the median time for completing the comprehension task on code automatically refactored is slightly higher than the median time spent on code manually refactored (approximately 5 minutes more). However, the difference is not statistically significant and the effect size is small.
In the case of effort, the effort perceived by developers is almost the same, with a negligible effect size.
In the case of percentage of correct answers, the median values are the same for both human and machine.
We repeated the analysis along the same three dimensions while
dividing the refactoring changes in two groups (passed and failed the Turing tests). With respect to time, the difference of the medians for anti-patterns that passed the Turing test ({\textit{i.e.,}}~BL, SC, LC) is higher (medium $ES$) for automatically generated changes, though the difference is not statistically significant. In the case of effort, the effect size is negligible and not statistically significant, same for percentage of correct answers.
For anti-patterns that failed the Turing test, the results are not statistically significant either.
\npdecimalsign{.}
\nprounddigits{2}
\begin{table}[t]
\centering
\renewcommand{\arraystretch}{1.0}
\renewcommand{\tabcolsep}{.5mm}
\caption{RQ3 Comprehension task results $E_2$: median values for the three dimensions studied}
\label{table:comprehension}
\begin{tabular}{|l|l|n{5}{2}|n{5}{2}|n{5}{2}|l|}
\hline
Grouping & Dimension & Human & {Machine} & {$p-value$} & {$ES$} \\ \hline
ALL & Time (seconds) & 1328 & 1657.5 & 0.069845 & small \\ \hline
ALL & Effort (TLX index) & 50.895 & 49.705 & 0.822843 & negligible \\ \hline
ALL & \% of correct answers & 67 & 67 & 0.760319 & negligible \\ \hline
T. passed & Time (seconds) & 1100 & 1606 & 0.087628 & medium \\ \hline
T. passed & Effort (TLX index) & 51.365 & 49.705 & 0.720635 & negligible \\ \hline
T. passed & \% of correct answers & 67 & 75 & 0.829502 & negligible \\ \hline
T. failed & Times (seconds) & 1176 & 1606 & 0.159852 & small \\ \hline
T. failed & Effort (TLX index) & 51.365 & 48 & 0.363437 & small \\ \hline
T. failed & \% of correct answers & 83 & 67 & 0.011397 & large \\ \hline
\end{tabular}
\end{table}
Since we could not find statistically significant differences in the performance of developers when performing code comprehension tasks on systems that were manually and automatically refactored, we reject $H_{0performance}$. Hence, we conclude that for the anti-patterns studied in this work there is no difference between the performance of developers performing comprehension tasks on code refactored by developers compared to automatically refactored code. This is good news for toolsmiths working on automatic tools for removing anti-patterns. For developers, this result will likely increase their trust in automatic refactoring tools, and for researchers it extends the existing body of knowledge on the benefits of automatic refactoring.
\hypobox{We conclude that for the set of refactoring strategies studied, developers can safely use automated tools since the impact of the automated changes on the comprehension of the code is not significantly different from the impact of changes performed by human developers.}
\section{Discussion}
\label{sec:discussion}
In this section, we provide further information about the obtained results and discuss their implications.
The results in $E_1$ show that in general, automatically generated refactorings and refactoring changes made by humans were equally difficult to identify. In 10\% of cases, developers couldn't even make a decision and opted for the \textit{``I do not know''} option.
\subsection{\reviewed{Refactoring change differences by anti-pattern type}}
When analysing results based on anti-patterns types, the picture is a bit different.
\subsubsection{Long-parameter-list anti-patterns automatically refactored were easier to identify than their manual counterpart}
For the two instances of Long-Parameter list studied, the refactoring strategy selected by both human and machine was to introduce a parameter-object. In fact, the solution provided for the \emph{easy} instance is conceptually the same for both the automatic tool and the developer (the only difference is the name selected for the new class). However, since we replaced the name in both solutions by an artificial name ({\textit{e.g.,}}~Clazz09465) to avoid introducing bias in the experiment (as explained in Section~\ref{subsec:anonymization}), the two refactoring changes are semantically equivalent. In the \emph{hard} instance, there was a small difference between the two solutions; \reviewed{the freelancer} declared the new class in the same file, while our automated approach created a new file; besides that minor change, we can consider both solutions to be semantically equivalent too.
As we mentioned in Section~\ref{sec:results}, we suggest that the quality of the refactoring changes applied to correct Long-parameter list did not influence the choice of the participants in the Turing test, but their expectations about refactoring. Some respondents mentioned that there is no reason to extract a new parameter object, given that the parameter object only stores data. Hence, we suggest that the developers surveyed did not give importance to the understandability aspect of having a long-list of parameters over having an extra class that clusters the common parameters together, and that is why they attributed the change to a machine. Note that the number of parameters extracted for the Long-parameter list instance identified by the authors as \emph{hard} is considerably long (8 parameters) compared to the rest of the classes (4 parameters) in the system where the instances of Long-parameter list anti-pattern resides.
\subsubsection{Refactoring changes of speculative generality and lazy class are semantically equivalent}
With respect to Speculative Generality and Lazy class, the refactoring strategies employed by both the automated tool and the developers are the same for the first anti-pattern's type ({\textit{i.e.,}}{} \textit{Collapse hierarchy}), while for the second type ({\textit{i.e.,}}{} Lazy class) the refactoring applied ({\textit{i.e.,}}{} \textit{inline class}) exhibits a small variation on the selected target class. The Lazy classes instances are composed of a unique static method, so they could be placed in any class that makes use of this method, so we consider the refactoring solutions for these two anti-pattern types to be equivalent, despite their origin. Given that in the solutions for both anti-pattern's types no classes or methods are added to the system, these refactorings can be completely automated, as they do not require to provide names to entities, according to some code conventions set for the systems.
\subsubsection{The hard instance of Spaghetti code that was manually refactored was easier to identify than the automatically refactored one}
\reviewed{The machine and human refactoring changes applied to remove spaghetti code, for the easy and hard instances, used the same strategy: \emph{replace long method with method object}. This strategy consists of extracting a long method inside the spaghetti code class into a new class.
For the easy instance, both solutions differ just in the name selected for the new class, and since we normalized the names, we consider them semantically equivalent. On the other hand, the refactorings applied for the hard instance of spaghetti code class exhibit some design differences that may have help some survey respondents to identify their origin.}
\reviewed{The long method extracted from the spaghetti code class makes use of a private attribute and a static method inside the spaghetti code class, and it does not require access to other external attributes or methods. In the automated refactoring change, two private attributes where moved to the new extracted class, while in the manual solution the freelancer added public getter to access the required private attribute outside the spaghetti code class. Hence, the use of the private attributes, in the automatically refactored code, are delegated to the extracted class inside the spaghetti code class. Since both machine and human set the access modifier of the static method to public, there was no need to pass a reference to the spaghetti code class in the automated refactoring solution. But in the manual solution the extracted long method requires a reference to the spaghetti code class to use the required private attribute. These differences were noticed by one of the survey respondent, who provided the following explanation to justify why he thought that the refactoring was automatically generated:\emph{``I don't think that a developer would extract the long method from the spaghetti code class to the new extracted class, and add as a parameter to the long method a reference to the spaghetti code class, because that reference is not used''}. With respect to the manual solution, one respondent commented that the addition of a getter, to the private attribute, and the way the extracted class was instantiated to call the long method does not seem to be automatically generated. To provide a complete picture of what we discussed, we show in Listing \ref{lst:refactoring-patch-sc-m} and Listing \ref{lst:refactoring-patch-sc-h}, the actual changes performed to instantiate the new class, and execute the extracted long method for both automatic and manual solutions. We use the acronyms ADD, DEL, and CHG to indicate added, deleted, and changed line.}
\noindent\begin{minipage}{.45\textwidth}
\lstinputlisting[language=Java, caption=Extract from refactoring change for \emph{hard} instance of spaghetti code by machine.,label={lst:refactoring-patch-sc-m},escapeinside={(*@}{@*)},basicstyle=\scriptsize,numbers=none,xleftmargin=15pt,breaklines=true,frame=tlrb]{ec_70_0_workflowmanager.tex}
\end{minipage}\hfill
\begin{minipage}{.45\textwidth}
\lstinputlisting[language=Java, caption=Extract from refactoring change for \emph{hard} instance of spaghetti code by human.,label={lst:refactoring-patch-sc-h},escapeinside={(*@}{@*)},basicstyle=\scriptsize,numbers=none,xleftmargin=15pt,breaklines=true,frame=tlrb]{ec_70_1_workflowmanager.tex}
\end{minipage}
\reviewed{The differences between the automatic solution for the \emph{hard} instance of Spaghetti code with respect to the manual one could be easily matched by adding a validation to prevent passing a reference of the refactored spaghetti class as parameter to the long method call, when none of its attributes and methods are used. Toolsmiths should pay attention to corner cases like this, which can only be revealed by thoroughly testing their tools with several scenarios and systems.}
\subsection{Perceived quality of refactoring changes}
Concerning quality perceived by developers, we observed that it is only for the Blob anti-pattern type (which passed our Turing test) that the difference of rates achieved between manual and automatic refactoring changes is statistically significant, favoring manual refactoring.
We \reviewed{identify} the following differences in the refactoring strategies \reviewed{used} by human and the automatic approach that are worth to discuss.
First, the automated approach selected \emph{move method} as a mean to decompose Blob classes and to redistribute functionality to other classes in the system, while the human approach relied on extracting new classes to reduce the size of Blob classes.
With respect to the refactoring strategy applied for the \emph{easy} instance of Blob, the automated solution consists of delegating 12 methods to a class where an association relationship exists. One respondent complained that the coupling between these two classes increased dramatically, which was not the case. \reviewed{In contrast, the manual solution for the \emph{hard} Blob's instance took a different path}. It consists of extracting two attributes and their corresponding getters and setters to a new class. This solution, though semantically correct, does not reduce the size of the Blob class significantly, but introduce coupling to a new data class.
For the \emph{hard} instance, one respondent commented about the manual solution as follows: \emph{``it seems good and needs to be corrected with minor changes''}. We studied the manual solution and found that it performs a clear separation of concerns by extracting methods and attributes to new classes according to their names. That would be hard for a machine to achieve, unless they gather some knowledge about the code lexicon. The refactoring strategy followed by the automated approach consists of delegating three methods to a related class that is member attribute of the Blob class. \reviewed{ Although, the strategy taken by the tool prevents adding extra coupling to a new extracted class while reducing Blob's class size, respondents of $E_1$, did not like it. They highlighted that when moving methods, the automated approach changed the visibility of the required attributes to public, which implied several changes with no clear benefit.}
\reviewed{Respondents rated the automatically-generated refactorings for Blob type} with average scores of 2.6 and 2.7, for the easy and hard instances respectively. While the human ones obtained 4.4 and 2.5 respectively. It is interesting to note that although the automatically-generated refactorings of Blob did not achieved the best rate, the perceived quality showed little variation between the two instances despite their different levels of complexity of each system. On the other hand, the rate achieved by the two manual refactoring changes for Blob type varied considerably \reviewed{based on the human effort of each freelancer, and the complexity of the code that enable (or not) to abstract functionality to new classes. From a practical point of view, developers will not only reduce maintenance costs by using automated approaches, but can ensure a standardized quality gain after the refactoring process.}
\subsection{Improving automated refactoring of Blob classes}
To improve the performance of automated refactoring on Blob anti-patterns, it would be necessary to control for code semantics targeting two aspects: (1) the generation of refactoring candidates should be guided by code lexicon; (2) an automated refactoring approach should incorporate a mechanism for naming new code entities based on code lexicon. The first aspect has been already discussed by Ouni~{et al.}~\cite{Ouni201518}, while the second point remains unexplored. We suggest that by addressing these two aspects we could improve the performance of automated refactoring of Blob classes, making it as good as the human refactorings, or even better.
\subsection{Refactoring changes impact on code comprehension}
Concerning the impact on code comprehension, we could not make an analysis by anti-pattern type as we only have 6 points for each anti-pattern type. However, $E_2$ reveals no significant difference between subject’s efforts, times, and percentages of correct answers on systems refactored by human and machine. We investigated whether the two mitigating variables : expertise and Java's level of confidence (not to be confused with respondent's level of confidence per question in $E_1$) declared by participants impacted our results. We set 4 levels for the years of experience and 5 levels for the degree of Java's confidence, using a Likert scale, from \emph{no confidence} to \emph{highly confident}. \\
Table~\ref{table:mitigatingComprehension} presents some descriptive statistics of the data collected for these two mitigating variables. Since for each mitigating variable we have multiple levels (more than two) corresponding to multiple categories, we used the Kruskal-Wallis Test, which is a non-parametric test for comparing multiple medians, to assess the impact of the mitigating variables on the three performance dimensions (time, effort, and \% of correct answers). We observed that the mitigating variables do not impact our results, as shown by the high $p-values$ in Table~\ref{table:mitigatingComprehension}; {\textit{i.e.,}}{} participants mostly had the same performance on refactorings from the same origin, no matter their level of expertise/Java confidence, which reinforces our assessment that refactoring changes are what did matter.
\begin{table}[t]
\centering
\renewcommand{\arraystretch}{1.0}
\renewcommand{\tabcolsep}{.5mm}
\caption{$p-values$ of the impact of the mitigating variables on the performance of participants for $E_2$}
\label{table:mitigatingComprehension}
\begin{tabular}{|l|l|r|r|r|}
\hline
Variable & origin & \multicolumn{1}{p{0.2\columnwidth}|}{effort: $p-values$} & \multicolumn{1}{p{0.2\columnwidth}|}{time: $p-values$} & \multicolumn{1}{p{0.2\columnwidth}|}{\% of correct answers: $p-values$} \\ \hline
\multirow{2}{*}{Expertise} & human & 0.3658 & 0.9516 & 0.857 \\ \cline{2-5}
& machine & 0.2172 & 0.6145 & 0.9109 \\ \hline
\multirow{2}{*}{Java confidence} & human & 0.4522 & 0.01657 & 0.346 \\ \cline{2-5}
& machine & 0.8893 & 0.8166 & 0.1339 \\ \hline
\end{tabular}
\end{table}
\section{Threats to Validity}
\label{sec:threats}
There are threats that limit the validity of this study. We discuss these threats and how we alleviate or accept them following common guidelines provided in~\cite{wohlin2012experimentation}.
\subsection{Construct Validity}
Construct validity threats concern the relation between theory and observations.
\reviewed{This work relies on good developers practices ({\textit{i.e.,}}~extreme programming~\cite{Beck:2004:EPE:1076267}) where developers are advised to perform refactoring to remove anti-patterns, in order to maintain the design quality at an acceptable levels, and hence ease software evolution. However, we cannot claim that removing anti-patterns is the prime reason for developers to perform refactoring, specifically the ones that we studied. However, relying on the notion of refactoring anti-patterns allowed us to objectively evaluate two methods (manual and automated refactoring), and to control from a large space of possibilities. That is why we had to constraint our study of the refactoring practice to existing well-known refactorings~\cite{fowler1999refactoring,Brown98-AntiPatterns}.}
In $E_2$, we use time and percentage of correct answers to measure the subjects' performance. The measured time was extracted from the video recording of the comprehension tasks sessions, while the percentage of correct answers was evaluated by one author of the paper and one Master's student. We believe that these measurements are objective, even if they can be affected by external factors, such as fatigue. We also use NASA TLX score to measure the subjects' effort. The TLX is by its own nature subjective and, thus, it is possible that our subjects declared effort values that do not perfectly reflect their effort.
The degree of severity of the anti-patterns is also a threat to construct validity. The anti-patterns instances selected in each system were validated through a voting process for decisions. The first author and a Master's student voted for the anti-patterns, and the second author reviewed the decisions. We based our selections on the definitions and examples provided by Brown and Fowler~\cite{Brown98-AntiPatterns,fowler1999refactoring}. To validate the solutions proposed by the automatic tool and the developers ({\textit{i.e.,}}{} the freelancers), we check that the solution preserves code's behavior based on the unit tests included in the SF110 corpus. We also control for the level of complexity of the refactoring changes proposed by freelancers and our tool by computing normalized change entropy metric, which showed that the changes judged by human evaluators are fair for both treatments~{\textit{cf.,}}{\ref{table:systems}}. Yet, it is possible that some of the refactoring changes proposed would have a different effect if applied to other systems in different contexts.
Construct validity threats could be the result of a mistaken relation between (automated) refactoring and program comprehension. We believe that this threat is mitigated by the fact that this relation seems rational. The results of our analysis suggest that certain anti-patterns' type can be automatically refactored with the same level of quality as the refactorings performed by human developers.
\subsection{Internal Validity}
We identify 4 threats to the internal validity of our study: learning, selection, instrumentation, and diffusion.
\subsubsection{Learning} Learning threats do not affect our study for the two experiments because we used a between-subject design. A between-subject design uses different groups of subjects, to whom different treatments are assigned. Additionally, we took each anti-pattern instance from different systems. For $E_1$, we balanced the groups (alternating difficulty level, and origin); then we randomized the appearance order of the refactoring changes for each group. For $E_2$, the freelancers had to perform two comprehension tasks. To mitigate the learning effect, the systems were presented in the same order for both treatments (manual and automatic refactoring). For example, consider a comprehension task for systems 47, and 52. We anticipated that developers will spent more time for the first system ({\textit{i.e.,}}~47), while getting familiar with the instructions, developer environment, etc., than with the second system (52). Hence, the extra time spent as a consequence of the learning process is considered in the same system for both treatments.
\subsubsection{Participant's selection} Participant's selection threats could impact our study due to the natural difference among the participants' skills. In $E_1$, we tried to mitigate this threat by inviting developers through technical Java developers mailing list, ({\textit{e.g.,}}~openJDK project), developers groups on social networks ({\textit{e.g.,}}~LinkedIn, Reddit, and Facebook). In $E_2$, we studied the possible impact of their expertise in Java through two mitigating variables and found no significant
impact on the obtained results.
\subsubsection{Instrumentation} Instrumentation threats were minimized by using objectives measures like time and percentage of correct answers. We observed some subjectivity in measuring refactoring changes quality, developer's confidence, and developer's experience in $E_1$; and developers' effort measured using the TLX score. For example, 5 years of experience of one developer, could be the equivalent of 3 years for another one. However, this subjectivity is inevitable in self-evaluations.
\reviewed{Another instrumentation threat to our study is the anonymization of new code lexicon. This only affects Blob, Long-parameter list and Introduce-parameter object anti-pattern types. Automated refactorings that introduce new elements are likely to be distinguished from their manual equivalents. But naming code lexicon is just one part of the semantic context. By anonymizing the names of newly created entities, we wanted to steer the focus of the respondents toward the structure of the code changes.
Recent works~\cite{hu2018deep} in automated code comment generation have shown promising results on generating human-like comments using deep learning.
Hence, we believe that there is a reasonable possibility to overcome the current code lexicon limitations of automated refactoring tools in the future, and this study can serve as base for performing further studies when the technology is mature enough.}
\subsubsection{Diffusion} Diffusion threats do not impact our study because (1) we recruit participants through web platforms and mailing lists, and they do not have physical interactions, and (2) we asked them to not disclose any information about the content of the surveys and the systems.
\subsection {Conclusion Validity}
Conclusion validity threats concern the relation between the treatment and the outcome. We paid attention not to violate the assumptions of the statistical tests that were performed. Indeed, we used non-parametric tests that do not require to make assumptions about the distribution of the data set.
\subsection {Reliability Validity}
Reliability validity threats concern the possibility of replicating this study. We provide all the necessary details to replicate our study in our lab's Web page~\cite{refturingReplication}, including a sample of the questionnaire and the comprehension tasks, and raw data to compute the statistics. The systems analyzed from SF110 are open-source and can be downloaded from the author’s web site. Our automated refactoring tool is also available on-line at \url{https://github.com/moar82/RefGen}.
\subsection{External Validity}
We performed our study on 10 different real-world systems belonging to different domains and with different sizes (see~\Cref{table:systems}). Our experimental design, providing a few classes of each system to each participant, is reasonable because, in real maintenance projects, developers perform their tasks on small parts of whole systems and probably limit themselves as much as possible to avoid getting lost in a large code base. In $E_1$ we summarize the refactoring changes using the diff notation, and provide it along with the original source code; allowing developers to spot changes fast, in a readable and standard way, and in case of doubt, to clone the repository to explore the code and--or apply the changes (by applying the diff file as a patch). To mitigate the impact that a \reviewed{lack of} familiarity to diff notations could have on the responses of our study, we explained the notation to participants prior to \reviewed{participating} in the study and provided them multiple examples in a guidelines document.
\reviewed{We cannot claim that our results can be generalized to other refactoring tools, and to other subjects. To generalize our results, we should implement other approaches different from RePOR. However, this was not the main objective of this work. Rather, we want to empirically evaluate if automatically refactored code can be interleaved with human code on development and maintenance activities of real world systems.}
Our future work includes replicating this study in other contexts, with other subjects, tools, questions, anti-patterns, and software systems.
\section{Conclusions}
\label{sec:conclusions}
Refactoring tool support is conjectured in the literature to be underused due to lack of awareness, and developers reluctance to incorporate machine-generated-code into their code base. To debunk this myth, and foster awareness of automated refactoring, we performed two experiments to evaluate the perceived quality of automated refactorings and their impact on code comprehension. Our results show that developers could not distinguished between automatically generated refactorings and refactorings created by humans, for 3 out of the 5 anti-pattern types studied. Moreover, in general, developers did not prefer refactorings generated by humans over automatically generated refactorings, the only exception being for removing Blob classes. We found no significant difference between the performance of developers performing comprehension tasks on code refactored by developers or by an automatic tool, to remove Blob, Lazy class, and Speculative generality.
Hence, we conclude that automated refactoring can be as effective as manual refactoring. However, for complex anti-patterns' types like Blob, we suggest that developer's expertise be included in the refactoring process as much as possible.
In the future work, we plan to enhance automated refactoring with code semantics by leveraging the code lexicon of systems when determining the best candidates to receive functionalities extracted
from Blob classes, and for automatically naming the new classes, and--or methods introduced during a refactoring process. By doing this, we could generate more natural refactoring solutions, and close the gap between human and machine generated refactorings.
\balance
\bibliographystyle{spmpsci}
|
\section*{Introduction}
Let $X$ be a smooth complex projective variety of dimension $n \ge 2$. We define the \textit{Konno invariant} of $X$ to be the minimal geometric genus of
a pencil of connected divisors on $X$:
\[
\textnormal{Konno}(X) \ = \ \min \Big\{ \, g \, \Big | \
\begin{minipage}{3in}{ $\exists$ a connected rational pencil
\ $ \pi : X \dashrightarrow \mathbf{P}^1 $ \ whose general fibre $F$ has $p_g(F) = g$ }\end{minipage}
\ \Big \}.\]
(The geometric genus of an irreducible projective variety is understood to be the $p_g$ of any desingularization.) This invariant was introduced and studied by Konno \cite{Konno}, who computed it for smooth surfaces in $\mathbf{P}^3$: he proves that in this case pencils of minimal genus are given by projection from a line. In general, one should view $\textnormal{Konno}(X)$ as one of many possible measures of the ``complexity" of $X$. The purpose of this note is to estimate this invariant for some natural classes of varieties.
Our first result involves the Konno invariant of varieties such as general complete intersections whose Picard groups are generated by a very ample divisor.
\begin{propositionalpha} \label{ThmA}
Assume that $\textnormal{Pic}(X)= \mathbf{Z} \cdot [H]$ where $H$ is a very ample divisor on $X$. Then
\[
h^0(K_X) \, - \, h^0(K_X - H) \ \le \ \textnormal{Konno}(X) \ \le \ h^0(K_X + H) - h^0(K_X) \, + \, h^1(K_X).
\]
\end{propositionalpha}
\noindent So for example if
\[ X_d \ \subseteq \ \mathbf{P}^{n+1}\]
is a hypersurface of degree $d$ (very general if $n = 2$), then as a function of $d$
\[ \textnormal{Konno}(X_d) \sim \frac{d^n}{n!}. \]
Observe that at least when $H^1(K_X) = 0$, the upper bound in Proposition A is the geometric genus of a general pencil in $\linser{H}$. However if $h^0(X, H) \gg n$, then one can construct special pencils of highly singular hypersurfaces of somewhat smaller geometric genus.
Our second result deals with polarized K3 surfaces of large degree.
\begin{theoremalpha} \label{K3ThmIntro}
Let $(S_d, L_d)$ be a polarized K3 surface of genus $d\ge 3$, and assume that
\[ \textnormal{Pic}(S_d) \ = \ \mathbf{Z} \cdot [L_d]. \]
Then
\[ \textnormal{Konno}(S_d) \, \in \, \Theta\big(\sqrt{d}\big), \]
i.e. there are constants $C_1, C_2> 0$ such that
\[ C_1 \cdot \sqrt{d}\ < \ \textnormal{Konno}(S_d) \ < \ C_2 \cdot \sqrt{d} \]
for all such surfaces $S_d$ and all large $d$.
\end{theoremalpha}
\noindent It is conjectured by Stapleton \cite{Stapleton} that the same statement holds for the degree of irrationality $\textnormal{irr}(S_d)$ of $S_d$, but this remains an intriguing open problem. An analogue of Theorem B is also valid for general polarized abelian surfaces.
The proof of Proposition \ref{ThmA} occupies \S 1. It arises as a special case of a somewhat more general (but very elementary) result dealing with one-dimensional families of hypersurfaces. Section 2 is devoted to a more refined lower bound for surfaces, from which we deduce Theorem \ref{K3ThmIntro}; following \cite{Konno}, the key point here is to use some classical statements of Noether computing the invariants of a linear series in terms of its multiplicities at finite and infinitely near points. We conclude with an appendix in which we review Noether's formulae, and show in particular how they lead to quick proofs of theorems of Deligne-Hoskin and Lech concerning finite colength ideals on a surface.
We are grateful to Francesco Bastianelli, Craig Huneke, David Stapleton and Ruijie Yang for useful discussions.
\section{Geometric genera of covering families of divisors}
Let $X$ be a smooth complex projective variety of dimension $n$.
\begin{theorem} \label{Sect1Thm}
Let $\{ F_t\}_{t \in T}$ be a family of divisors on $X$ parametrized by a smooth curve $T$. Assume that the $F_t$ are generically irreducible and that they cover $X$, and denote by $F \subseteq X$ a general element in the family.
Then
\begin{equation} \label{Thmeqn} p_g(F) \ \ge \ \hh{0}{X}{\mathcal{O}_X( K_X)} \, - \, \hh{0}{X}{\mathcal{O}_X( K_X - F)}.\end{equation}
\end{theorem}
\noindent Note that although we don't assume that the $\{F_t\}$ are all linearly equivalent, the expression on the right is independent of the choice of a generic element of the family. Observe also that it can happen that equality holds in \eqref{Thmeqn}: for example one can take $X = C \times F$ where $C$ is an elliptic curve.
\begin{proof} We adapt the elementary argument proving \cite[Theorem 1.10]{BDELU}. One can construct a diagram:
\[
\xymatrix{Y\ar[d]_\pi \ar[r]^\mu &X\\
T}
\]
where $Y$ is smooth, and almost all fibres
\[ E_t \, =_\text{def} \, \pi^{-1}(t) \ \subseteq \ Y \]
are smooth irreducible divisors mapping birationally to their images $F_t \subseteq X$. Denote by $E$ a general fibre of $\pi$, with $F = \mu(E) \subseteq X$. So by definition $p_g(F) = \hh{0}{E}{\mathcal{O}_E(K_E)}$. Now
\[ K_Y \ \equiv_{\text{lin}} \ \mu^* K_X \, + \, R, \]
where $R$ is effective, and $K_E = K_Y \mid E$. Therefore
\[ p_g(F) \ \ge \ \hh{0}{E}{\mu^*\mathcal{O}_X(K_X) \mid E}. \] On the other hand, $\mu^*$ gives rise to a natural injection
\[ \HH{0}{F}{\mathcal{O}_F(K_X)} \hookrightarrow \HH{0}{E}{\mathcal{O}_E(\mu^*
K_X)},\]
so we arrive finally at the inequality
\[ p_g(F) \ \ge \ \hh{0}{F}{\mathcal{O}_X(K_X)\mid F}. \]
The statement then follows by using the exact sequence
\[ 0 \longrightarrow \mathcal{O}_X(K_X - F) \longrightarrow \mathcal{O}_X(K_X) \longrightarrow \mathcal{O}_F(K_X) \longrightarrow 0
\] to estimate $\hh{0}{F}{\mathcal{O}_F(K_X)}$.
\end{proof}
\begin{proof}[Proof of Proposition \ref{ThmA}]
We apply the previous result with $F \in \linser{rH}$ for some $r \ge 1$. The right hand side of \eqref{Thmeqn} is mimimized when $r = 1$, and the lower bound follows. The upper bound follows by considering a general pencil in $\linser{H}$.
\end{proof}
\begin{remark} \textbf{(Covering families of curves)}. By a similar argument, if $\{ C_t \}_{t\in T}$ is a family of irreducible curves of geometric genus $g$ that covers a Zariski-open subset of $X$, then
\begin{equation} (2g - 2) \ \ge \ \big ( K_X \cdot C \big ), \label{Cov.Fam.Curves} \end{equation}
where $C$ is a general curve in the family.
\end{remark}
\section{The Konno invariant of an algebraic surface}
The inequality of Theorem \ref{Sect1Thm} says nothing for varieties with trivial canonical bundle. In the case of surfaces we prove here a variant that does yield non-trivial information in this case. The approach is inspired by the arguments of Konno in \cite{Konno}.
\begin{theorem} \label{SfBoundThm} Let $S$ be a smooth complex projective surface, and let $L$ be an ample line bundle on $S$. Fix a two-dimensional subspace $V \subseteq \HH{0}{S}{L}$ with only isolated base-points defining a rational pencil
\[ \phi_{|V|} : S \dashrightarrow \mathbf{P}^1\]
with generically irreducible fibres. If $g$ denotes the geometric genus of the general fibre, then
\begin{equation} \left ( 2g-2 \right) \ge \ \big( K_S \cdot L \big) \, + \, \sqrt{(L^2)}\end{equation} \label{Thm2.1Ineq}
\end{theorem}
\begin{remark} Compare the bound appearing above in equation \eqref{Cov.Fam.Curves}. \end{remark}
\begin{proof}
By a sequence of blowings up at points, we construct a resolution of the indeterminacies of $\linser{V}$:
\[
\xymatrix{S^\prime\ar[d]_\pi \ar[r]^\mu &S\\
\mathbf{P}^1 }.
\]
We can suppose that the centers of the blowings-up are the (actual and infinitely near) base-points of $\linser{V}$. Let $m_i$ denote the multiplicity of the proper transform of a general curve $C \in \linser{V}$ at the $i^{\text{th}}$ base-point, and denote by $C^\prime$ proper transform of $C$ in $S^\prime$, so that $C^\prime$ is a general fibre of $\pi$ and $g = g(C^\prime)$. Then by a classical theorem of Noether, which we recall in the Appendix (Proposition \ref{Noether.Prop}), one has
\begin{align*}
\big( C \cdot C \big)_S\ = \ (C^\prime \cdot C^\prime)_{S^\prime}+ \sum m_i^2,
\end{align*}i.e.
\[ \big(L^2 \big) \ = \ \sum m_i^2 \ . \tag{*} \]
Furthermore,
\[ \left(2 p_a(C) - 2\right) \ = \ (2 g - 2) \, + \, \sum m_i(m_i-1). \]
But $ (2p_a(C) - 2)= (K_X+ L) \cdot L$, so we find that
\[ \left( 2g - 2 \right) \ = \ \big( L \cdot K_X \big) \, + \, \sum m_i . \]
The stated inequality \eqref{Thm2.1Ineq} then follows from (*) and the fact that $\sum x_i \ge \sqrt{\sum x_i^2}$ for any non-negative real numbers $x_i$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{K3ThmIntro}]
Let $(S,L) = (S_d, L_d)$ be a polarized K3 surface of genus $d$, so that
\[
\big( L^2 \big) \ = \ 2d - 2 \ \ , \ \ \hh{0}{S}{L} \ = \ d +1.
\] We assume that Pic$(S) = \mathbf{Z} \cdot [L]$, and consider a rational pencil
\[ \phi : S \dashrightarrow \mathbf{P}^1 \]
of curves of geometric genus $g$. Then for some $r \ge 1$, $\phi$ is defined by a two-dimensional subspace $V \subseteq \HH{0}{S}{rL}$ with isolated base points. Theorem \ref{SfBoundThm} implies that
\[ 2g - 2 \ \ge \ r \cdot \sqrt{2d - 2} \ \ge \ \sqrt{2d -2}, \]
so $\textnormal{Konno}(S) \ge C_1 \cdot \sqrt{d} $ for suitable $C_1 > 0$.
It remains to construct a pencil of small genus, for which we use an argument of Stapleton \cite{Stapleton}. Specifically, fix a point $x \in S$, and choose an integer $m \ge 1 $ so that
\[ (m+2)^2 \ \ge \ 2d \ \ge \ (m+1)^2 . \tag{*}\]
It follows from (*) that
\[ (d+1) \, - \, \frac{m(m+1)}{2} \ \ge \ 2, \]
and therefore \[\hh{0}{S}{L \otimes I_x^m} \ge 2 \]
All the curves in $\linser{L \otimes I_x^m}$ are reduced and irreducible, so we get a pencil of curves of geometric genus $g$ with
\[ (2g - 2) \ \le \ (2d-2) \, - \, m(m-1). \] But $2d - m^2 \le 4m + 4$ thanks to (*), and one then finds that $(2g - 2) \le 3 \cdot \sqrt{2d}$. Thus we have constructed a pencil of geometric genus $\le C_2 \cdot \sqrt{d}$ for suitable $C_2$, as required.
\end{proof}
\begin{remark} (\textbf{Abelian surfaces}). Let $A$ be an abelian surface with a polarization of type $(1,d)$ that generates the N\'eron-Severi group of $A$. Then essentially the same argument shows that
\[ \textnormal{Konno}(A) \ \in \ \Theta(d).\]
\end{remark}
\begin{remark} (\textbf{Non-linear families of curves on K3 surfaces)}.
One can view Theorem \ref{K3ThmIntro} as asserting there are no \textit{lines}
\[ \mathbf{P}^1 \, \subseteq \, \linser{L_d} \]
contained in the locus of curves having small geometric genus. It would be interesting to know whether one can also rule out the presence of rational curves of higher degree. For example, a general polarized K3 surface $(S, L_d)$ contains a (non-compact) two-dimensional family of nodal curves of geometric genus $p_g = 2$. Does this surface contain any rational curves? More generally, do the Severi varieties parametrizing nodal curves of small geometric genus in $\linser{L_d}$ exhibit hyperbolic tendencies?
\end{remark}
\begin{remark} (\textbf{Calabi-Yau or hyper-K\"ahler manifolds}). Can one establish non-trivial lower bounds on the Konno invariant of a Calabi-Yau or hyper-K\"ahler manifold?
\end{remark}
|
\section{Introduction}
\label{secmot}
Controller's loss of authority over parts of an autonomous system may happen in many scenarios:
\begin{enumerate}[(a)]
\item \textit{System damage and component degradation.} An autonomous system operating for substantial periods of time in remote, unknown, or hostile environment will inevitably sustain damage or experience partial system failures over time due to malfunctions. Examples include unmanned aerial vehicles (UAVs) operating over contested territory \cite{RatSen04}, search-and-rescue robots \cite{Chaetal18}, and rovers performing missions on extraterrestrial surfaces \cite{Wasetal99}.
\item \textit{Hostile takeover.}
In a number of adversarial settings, the adversary will attempt to take over elements of the system and disturb its regular functions. A typical setting is that of attacks on computer networks \cite{Vat01} and power systems \cite{AmiGia12,Zhuetal14}, where, because of the vastness of the network and heterogeneity and physical distance between system elements, an adversarial agent may be able to penetrate a part of the system. Hostile takeover scenarios also include recent successful attacks resulting in loss of control over UAVs; see, e.g., \cite{Woletal14,HarGil16}.
\item \textit{User-responsive systems.} Settings where an automated controller is required to respond to (a priori unknown) user inputs in a particular way necessarily yield a part of the control authority to the user. Such scenarios include resource distribution in parallel computing \cite{Feietal90}, semantic web service composition \cite{Rodetal12}, and communication protocols \cite{Sha48}.
\end{enumerate}
In all of the above settings, it is critical to ensure that the autonomous system can perform its tasks regardless of external inputs that may affect the system. A standard method of ensuring continued functioning of the system is through imposing redundancy or near-redundancy in design. For instance, critical components in commercial airplanes are duplicated \cite{Dow09}, and military UAVs use a combination of different sensing systems for navigation \cite{HarGil16}. In the latter example, while these different sensing systems do not work in the same way and, in regular flight regime, serve to complement each other, each system is able to ensure that the UAV can achieve basic control objectives even if complementary systems are not functioning.
Motivated by the above scenarios, our work seeks to investigate how to guarantee continued safe operation of an abstract control system in which some components are no longer under the controller's authority. We focus on systems with linear driftless discrete-time dynamics, and interpret the partial loss of control authority as limitations on the controller's choice of actions, based on adversarial inputs. The control objective that we investigate is {\em safety}: the system state is required to remain within a particular set throughout the system run. We are interested in (i) developing a safe control policy, if one exists, and (ii) determining a {\em resilient system design} --- i.e., a partition of the set of all control inputs that the system can apply --- which ensures that the system will be able to remain safe even if the adversary limits the available actions to a single element of the partition at any given time.
The work in this paper is closely related to previous research on control of safety-critical systems \cite{TomLyg00,Tab09} and safety games \cite{Beretal02, DoyRas11, NelTop16}. In particular, as we will show, given a system design, i.e., possible control inputs given an adversarial input, a safe control policy can be interpreted as a winning strategy for a turn-based safety game. This interpretation leads to a computationally efficient algorithm for designing a safe control policy. However, such an algorithm does not directly provide for a computationally feasible procedure of determining whether there {\em exists} a resilient system design, as each design corresponds to a different safety game, and searching through all possible games is computationally prohibitive. We address this challenge through a method based on a graph-theoretical interpretation of system design.
The outline of the remainder of this paper is as follows. In Section \ref{probsta} we provide a motivation for theoretical framework used in the paper, and formally describe the problems of safe control design and resilient system design under adversarial action. We then interpret such problems within the context of safety games in Section \ref{gam}, resulting in a simple solution to the problem of safe control design. We interpret the problem of resilient system design in a graph-theoretical setting in Section \ref{gra}, and --- using the probabilistic method, as described in \cite{AloSpe08} --- provide a sufficient condition and a necessary condition for its solvability in Section \ref{condit}. Based on the previous section, we provide a computationally efficient algorithm for resilient system design and construction of a safe control policy in Section \ref{ela}. Section \ref{exams} illustrates our techniques on two examples: an autonomous vehicle experiencing partial loss of control authority, and design of codes for communication over a channel with a bounded running digital sum.
\textbf{Notation.} The symbol ${\mathbb N}$ denotes all strictly positive integers, ${\mathbb N}_0$ denotes all nonnegative integers, and ${\mathbb Z}$ denotes all integers. For $m\in{\mathbb N}$, $[m]$ denotes the set $\{1,\ldots,m\}$. For a set ${\mathcal{X}}$, $|{\mathcal{X}}|$ denotes its cardinality, and $2^{{\mathcal{X}}}$ the set of all its subsets. For an event $B$ within a particular probability distribution, $\Pr(B)$ denotes the probability of $B$ occurring. For a graph $G=(V,E)$ and vertex $v\in V$, $\deg_G(v)$ denotes the (outgoing, if the graph is directed) degree of $v$, and $\mindeg(G)$ denotes the minimal (outgoing) degree of any vertex in $V$. If $G, H$ are graphs, $G\subseteq H$ signifies that $G$ is an induced subgraph of $H$. Vector $e_i$ denotes the standard basis vector consisting solely of zeros, except for a $1$ in the $i$-th position. Symbol $\|v\|_\infty$ denotes the max-norm of a vector $v\in{\mathbb R}^n$, and $\|v\|_1$ denotes the $1$-norm of a vector $v$.
\section{Problem Statement}
\label{probsta}
Consider a system operating with discrete-time dynamics \begin{equation}
\label{dtd}
x(t+1)=x(t)+u(t)
\end{equation}
for all times $t\in{\mathbb N}_0$, where $x(0)\in{\mathbb Z}^n$ and $u\in{\mathcal{U}}\subseteq{\mathbb Z}^n$, with a finite ${\mathcal{U}}$.
While model \eqref{dtd} is simple, our use of it is motivated by its wide presence in robotic exploration (see, e.g., \cite{Yam97, Megetal12, Oswetal16}, and the references therein) as well as its use in communication over a channel \cite{CohLit91}. As we will discuss in subsequent sections, \eqref{dtd} yields a straightforward graph-theoretical interpretation of system motion which may lead to generalizations for more complex models.
To provide motivation for the problems that we will pose, let us assume that dynamics \eqref{dtd} represent an autonomous system controlled by actuators $A_1$, $A_2$, \ldots, $A_p$. The control effort $u(t)$ is then given as $u(a_1(t),\ldots,a_p(t))$, where $a_i(t)\in{\mathcal{A}}_i$ is the setting of actuator $A_i$ at time $t$, and ${\mathcal{U}}=\{u(a_1,\ldots,a_p)~|~a_i\in{\mathcal{A}}_i, i=1,\ldots,p\}$.
We are interested in the scenario where the controller experiences loss of authority over some of the actuators, say $A_1, \ldots, A_r$. Thus, the choice of $a_1(t), \ldots, a_r(t)$ is not made by the controller, and any control actuation $u(t)$ needs to chosen in the set
$U(a_1(t),\ldots,a_r(t))= \{u(a_1(t),\ldots,a_r(t),a_{r+1},\ldots, a_p)~|~a_i\in {\mathcal{A}}_i, i\geq r+1\}$.
We assume that we do not possess any prior knowledge about the inputs $a_1(t), \ldots, a_r(t)$; these may be subjects to adversarial choices.
The control objective that we consider is {\em safety}. That is, we want to ensure that $x(t)\in S$ for all $t\geq 0$, where $S\subseteq{\mathbb Z}^n$ is a predetermined set with $x(0)\in S$. We are interested in two questions:
\begin{enumerate}[(i)]
\item For given sets $U(a_1,\ldots, a_r)$, determine, if it exists, a control policy that guarantees system safety regardless of choices $a_1(t), \ldots, a_r(t)$.
\item Design sets $U(a_1, \ldots, a_r)$ so that the above control policy exists.
\end{enumerate}
The latter question corresponds to designing the abilities and role of each actuator in such a way that the system is resilient to loss of authority over some of the actuators.
If the system can exhibit perfect redundancy, i.e., $U(a_1, \ldots, a_r)={\mathcal{U}}$ for every $a_1\in{\mathcal{A}}_1$, \ldots, $a_r\in{\mathcal{A}}_r$, questions (i) and (ii) are simple. However, redundancy is often undesirable due to cost, weight, or resource consumption \cite{Sghetal08}. Thus, we assume that it is impossible to execute exactly the same control with two different actuations. Under this assumption, $\{U(a_1,\ldots,a_r)~|~ a_1\in{\mathcal{A}}_1, \ldots, a_r\in{\mathcal{A}}_r\}$ is a partition of ${\mathcal{U}}$.
For the sake of simpler notation, we denote ${\mathcal{A}}_1\times{\mathcal{A}}_2\times\cdots\times{\mathcal{A}}_r=[m]$ for some $m\in{\mathbb N}$.
Questions (i) and (ii) are now formulated as follows.
{\em Restricted partition control problem (RPCP):}
Let $S\subseteq{\mathbb Z}^n$ and $x(0)\in S$. Let ${\mathcal{U}}\subseteq{\mathbb Z}^n$ be finite, and $U:[m]\to 2^{\mathcal{U}}$ such that $\{U(1),\ldots,U(m)\}$ is a partition of ${\mathcal{U}}$. Does there exist a function $\hat{u}:\cup_{i=1}^\infty [m]^i\to{\mathcal{U}}$ such that
\begin{enumerate}[(i)]
\item $\hat{u}(d_1,\ldots,d_k)\in U(d_k)$ for all $d_1,\ldots,d_k\in [m]$, and
\item for every $d:{\mathbb N}_0\to[m]$, if $x(t)$ is the solution of \eqref{dtd} with $u(t)=\hat{u}(d(0),\ldots,d(t))$, then $x(t)\in S$ for all $t\in{\mathbb N}_0$?
\end{enumerate}
{\em Free partition control problem (FPCP):}
Let $S\subseteq{\mathbb Z}^n$ and $x(0)\in S$. Let ${\mathcal{U}}\subseteq{\mathbb Z}^n$ be finite. Does there exist a partition $\{U(1),\ldots,U(m)\}$ for which the RPCP admits a solution?
We note that in practice the available choices of partitions in the FPCP may be subject to constraints, e.g., physical limitations in design of actuators. We use the unconstrained version to provide an elegant illustration of a general approach to solving the above problems. Before moving towards solutions of the RPCP and the FPCP, let us introduce a running example.
\begin{example}[Damaged vehicle]
\label{exrun}
Consider an autonomous vehicle moving on ${\mathbb Z}^2$ according to dynamics \eqref{dtd}. At every instance in time, the vehicle can perform one of five actions: go one position to the north, south, east or west, or remain in the same position. In other words, ${\mathcal{U}}=\{e_1,-e_1,e_2,-e_2,(0,0)\}$. The vehicle's initial position is given by $x(0)=(0,0)$, and the safe set $S$ is given by $S=\{x\in{\mathbb Z}^2~|~\|x\|_\infty\leq 1\}$. The setup is graphically illustrated in Fig.~\ref{runn}.
Let us first consider the RPCP with $m=2$ and $U(1)=\{e_1,e_2\}$, $U(2)=\{-e_1,-e_2,(0,0)\}$. In such a case, the RPCP does not admit a solution. For instance, if the adversary continually chooses $d=1$, the vehicle will have to keep moving north or east. Hence, after no more than $3$ steps, it will be forced to leave $S$. This situation is shown on the left side of Fig.~\ref{runn}.
\begin{figure}[ht]
\center
\begin{tikzpicture}[scale=0.95]
\clip (-1.65,-1.65) rectangle (2.65,2.65);
\tikzstyle{point}=[thick,fill=black, draw=black,circle,inner sep=0pt,minimum width=7pt,minimum height=7pt]
\tikzstyle{darkstyle}=[circle,draw,fill=gray!40,inner sep=0pt,minimum width=15pt,minimum height=15pt]
\tikzstyle{greenstyle}=[circle,draw,fill=green!20,inner sep=0pt,minimum width=15pt,minimum height=15pt]
\tikzset{->-/.style={decoration={
markings,
mark=at position .5 with {\arrow{>}}},postaction={decorate}}}
\tikzset{-<-/.style={decoration={
markings,
mark=at position .5 with {\arrow{Computer Modern Rightarrow[slant=.6]}}},postaction={decorate}}}
\foreach \x in {-3,...,3}
\foreach \y in {-3,...,3}
\node (\x\y) at (\x,\y) {};
\foreach \x in {-3,...,3}
\foreach \y in {-3,...,3} {
\draw[red,->-] (\x\y) arc(-45:45:0.71);
\draw[blue,->-] (\x\y) arc(45:135:0.71);
\draw[blue,->-] (\x\y) arc(135:225:0.71);
\draw[red,->-] (\x\y) arc(225:315:0.71);
}
\foreach \x in {-1,...,2}
\foreach \y in {-1,...,2} { \draw[blue,-<-] (\x,\y+0.2) arc(-90:270:0.13);
}
\foreach \x in {0,...,1}
\foreach \y in {0,...,1} {
\draw[ultra thick, red, ->-] (\x\y) arc(225:315:0.71);
\draw[ultra thick, red, ->-] (\x\y) arc(-45:45:0.71);
}
\draw[ultra thick, red, ->-] (0,2) arc(225:315:0.71);
\draw[ultra thick, red, ->-] (0,2) arc(-45:45:0.71);
\draw[ultra thick, red, ->-] (2,0) arc(225:315:0.71);
\draw[ultra thick, red, ->-] (2,0) arc(-45:45:0.71);
\foreach \x in {-2,...,2}
\foreach \y in {-2,...,2}
\node[darkstyle] (\x\y) at (\x,\y) {};
\foreach \x in {-1,...,1}
\foreach \y in {-1,...,1}
\node[greenstyle] (\x\y) at (\x,\y) {};
\node[point] at (0,0) {};
\end{tikzpicture}
\hspace*{-2.5pt}
\unskip\ \vrule width 1pt\
\begin{tikzpicture}[scale=0.95]
\clip (-1.65,-1.65) rectangle (2.65,2.65);
\tikzstyle{point}=[thick,fill=black, draw=black,circle,inner sep=0pt,minimum width=7pt,minimum height=7pt]
\tikzstyle{darkstyle}=[circle,draw,fill=gray!40,inner sep=0pt,minimum width=15pt,minimum height=15pt]
\tikzstyle{greenstyle}=[circle,draw,fill=green!20,inner sep=0pt,minimum width=15pt,minimum height=15pt]
\tikzstyle{darkgreen}=[circle,draw,fill=green!70!black,inner sep=0pt,minimum width=15pt,minimum height=15pt]
\tikzset{->-/.style={decoration={
markings,
mark=at position .5 with {\arrow{>}}},postaction={decorate}}}
\tikzset{-<-/.style={decoration={
markings,
mark=at position .5 with {\arrow{Computer Modern Rightarrow[slant=.6]}}},postaction={decorate}}}
\foreach \x in {-3,...,3}
\foreach \y in {-3,...,3}
\node (\x\y) at (\x,\y) {};
\foreach \x in {-3,...,3}
\foreach \y in {-3,...,3} {
\draw[blue,->-] (\x\y) arc(-45:45:0.71);
\draw[red,->-] (\x\y) arc(45:135:0.71);
\draw[blue,->-] (\x\y) arc(135:225:0.71);
\draw[red,->-] (\x\y) arc(225:315:0.71);
}
\foreach \x in {-1,...,2}
\foreach \y in {-1,...,2} {
\draw[blue,-<-] (\x,\y+0.2) arc(-90:270:0.13);
}
\draw[thick, blue,-<-] (0,0.2) arc(-90:270:0.13);
\draw[thick, blue,-<-] (1,0.2) arc(-90:270:0.13);
\draw[ultra thick, blue] (0,0.2) arc(-90:270:0.13);
\draw[ultra thick, blue] (1,0.2) arc(-90:270:0.13);
\draw[ultra thick, red,->-] (0,0) arc(225:315:0.71);
\draw[ultra thick, red,->-] (1,0) arc(45:135:0.71);
\foreach \x in {-2,...,2}
\foreach \y in {-2,...,2}
\node[darkstyle] (\x\y) at (\x,\y) {};
\foreach \x in {-1,...,1}
\foreach \y in {-1,...,1}
\node[greenstyle] (\x\y) at (\x,\y) {};
\node[darkgreen] at (0,0) {};
\node[point] at (0,0) {};
\node[darkgreen] at (1,0) {};
\end{tikzpicture}
\caption{The picture on the left illustrates a counterexample to solvability of the RPCP for $U(1)=\{e_1,e_2\}$, $U(2)=\{-e_1,-e_2,(0,0)\}$ in Example \ref{exrun}. The safe set $S$ is denoted in light green. The vehicle's initial position $x(0)=(0,0)$ is denoted by a black circle. Possible vehicle movements from each $x\in{\mathbb Z}^n$ are denoted by an arrow. Red arrows denote available movements when the adversary chooses $d=1$, and blue arrows denote available movements when $d=2$. Possible movements in the case when the adversary chooses $d(0)=d(1)=d(2)=1$ are drawn thickly. The picture on the right illustrates of solvability of the FPCP in Example \ref{exrun}. Same notation as in the left is used. The partition $\{U(1),U(2)\}$ is chosen in such a way that, regardless of the choice of $d(t)$, the vehicle can always remain in the dark green subset of the safe set.}
\label{runn}
\end{figure}
On the other hand, the FPCP admits a solution for $m=2$. Let $U(1)=\{e_1,-e_1\}$ and $U(2)=\{e_2,-e_2,(0,0)\}$. Then, when the adversary chooses $d=1$ for the first time, the vehicle can choose to move east, then west the next time, then east again, etc. If the adversary chooses $d=2$, the vehicle can remain in place. Hence, the vehicle will always remain within $S$. Such a strategy is depicted on the right side of Fig.~\ref{runn}.
\hfill $\square$
\end{example}
We now continue towards providing a solution for the RPCP and the FPCP.
\section{Game Formulation}
\label{gam}
The RPCP can be easily formulated as the question of existence of a winning strategy in the following two-player game.
\begin{game}
\label{gm1}
Let $S\subseteq{\mathbb Z}^n$ and $x(0)\in S$. Let ${\mathcal{U}}\subseteq{\mathbb Z}^n$ be finite, and $\{U(1),\ldots,U(m)\}$ be a partition of ${\mathcal{U}}$. Let $G=(V,E)$ be a graph with $V={\mathbb Z}^n$ and $E=\{(x,y)~|~y-x\in{\mathcal{U}}\}$, and $l:E\to[m]$ a labeling given by
\begin{equation}
\label{labe}
l(x,y)=d \qquad\textrm{if }y-x\in U(d)\textrm{.}
\end{equation}
The game proceeds as follows. Before time $t=0$, a token is placed at $x(0)$. At every time step $t$, Player 1 first chooses an element $d\in [m]$. Then, Player 2 chooses an element $x(t+1)\in V$ such that $(x(t),x(t+1))\in E$ and $l(x(t),x(t+1))=d$, if such an element exists, and moves the token to $x(t+1)$. The game now proceeds to the next time step. Player 2 wins the game if it can always move the token, and the token remains within $S$ for all $t\in{\mathbb N}_0$. Otherwise, Player 1 wins.
\end{game}
\begin{proposition}
\label{propo0}
The RPCP admits a solution if and only if there exists a winning strategy for Player 2 in Game \ref{gm1}.
\end{proposition}
\begin{proof}
By taking $u(t)=x(t+1)-x(t)$, it is clear that the movement of the token in Game \ref{gm1} corresponds to \eqref{dtd}. The requirement that $(x(t),x(t+1))\in E$ and $l(x(t),x(t+1))=d$ corresponds to the requirement that $u(t)\in U(d(t))$. Thus, Player 2 has a winning strategy in Game \ref{gm1} if and only if there exists $u(t)\in U(d(t))$, possibly dependent on all previous inputs $d(0),\ldots,d(t)$, such that $x(t)\in S$. The latter statement is exactly the statement of the RPCP.
\end{proof}
Game \ref{gm1} is a turn-based safety/reachability game with complete information as described \cite{DoyRas11}. Thus, for finite $S$, the RPCP can be solved in linear time with respect to the size of $S$ \cite{DoyRas11}. In the remainder of this paper, we focus on the FPCP. In a game-theoretical setting, the FPCP can be posed as follows.
\begin{game}
\label{gm2}
Let $S$, $x(0)$, ${\mathcal{U}}$, and $G=(V,E)$ be as in Game~\ref{gm1}. Let $m\in{\mathbb N}$. At time $t=-1$, Player 2 chooses $U(d)\subseteq{\mathcal{U}}$ for all $d\in [m]$ in such a way that $\{U(d)~|~d\in [m]\}$ is a partition of ${\mathcal{U}}$. Then, each edge $(x,y)\in E$ is labeled as in \eqref{labe}. After this step, the game proceeds the same as Game \ref{gm1}.
\end{game}
Analogously to Proposition \ref{propo0}, it can be easily shown that the FPCP admits a solution if and only if Player 2 has a winning strategy in Game \ref{gm2}.
The problem of the existence of a winning strategy in Game \ref{gm2} can nominally be solved by reducing it to the problem of existence of a winning strategy in Game \ref{gm1}. Namely, every choice of a partition $\{U(d)~|~d\in [m]\}$ at time $t=-1$ generates a different instance of Game \ref{gm1}, so Player 2 has a winning strategy in Game \ref{gm2} if and only if there exists a partition $\{U(d)~|~d\in [m]\}$ for which Player 2 has a winning strategy in Game \ref{gm1}. However, an algorithm that determines a winning strategy for Game \ref{gm2} by considering all partitions $\{U(d)~|~d\in [m]\}$ is infeasible for large ${\mathcal{U}}$, as the number of those partitions is not less than $m^{|{\mathcal{U}}|-m}$ \cite{RenDob69}.
In the following section, we propose a graph-theoretical approach to the problem of determining the existence of winning strategies for Player 2 in the above games, resulting in easily computable conditions for the existence of a partition and a controller in the FPCP.
\section{Graph Labeling Problem}
\label{gra}
The previous section interprets system motion as a game on a labeled graph. By building upon this approach, we can convert the problem of finding a partition of the set of control inputs that admits a safe control policy --- the FPCP --- to an equivalent problem of labeling of graph edges
\begin{theorem}
\label{propo1}
Let $S$, $x(0)$, ${\mathcal{U}}$, and $G$ be as in Game~\ref{gm1}. The FPCP admits a solution if and only if there exist an induced subgraph $G_{\hat{S}}=(\hat{S},E_{\hat{S}})\subseteq G$ with $\hat{S}\subseteq S$ and a labeling $l~:~E_{\hat{S}}\to [m]$ such that the following properties hold:
\begin{enumerate}[leftmargin=23pt]
\item[(C1)] $x(0)\in \hat{S}$,
\item[(C2)] for all $x\in \hat{S}$, $$l\left(\{(x,x')\in E_{\hat{S}}~|~x'\in\hat{S}\}\right)=[m]\textrm{,}$$ and
\item[(C3)] if $(x,y),(x',y')\in E_{\hat{S}}$ satisfy $y-x=y'-x'$, then $l(x,y)=l(x',y')$.
\end{enumerate}
\end{theorem}
\begin{proof}
As previously noted, the FPCP admits a solution if and only if there exists a winning strategy for Player 2 in Game \ref{gm2}. Assume first that such a winning strategy exists, with the corresponding partition $\{U(d)~|~d\in [m]\}$ and a labeling $l:E\to [m]$ that satisfies \eqref{labe}. Let us now define $G_{\hat{S}}=(\hat{S},E_{\hat{S}})$ as the induced subgraph of $G$ with its vertex set $\hat{S}$ consisting of {\em all} the values that the system state $x(t)$ can possibly assume under the chosen winning strategy, for {\em all} potential input sequences $d:{\mathbb N}_0\to [m]$. We claim that $G_{\hat{S}}$, with the labeling $l$ restricted to $E_{\hat{S}}$, satisfies (C1)--(C3).
First, since $\hat{S}$ is constructed from the winning strategy of Player 2, $x(0)\in\hat{S}\subseteq S$. Thus, (C1) holds. Property (C2) holds because, by definition of $\hat{S}$, for each $x\in \hat{S}$ there exists a $t\geq 0$ and a sequence $d(0),\ldots,d(t-1)$ such that $x(t)=x$, and for each $d'\in [m]$, setting $d(t)=d'$ requires that $l(x(t),x(t+1))=d'$. Property (C3) holds by \eqref{labe}.
In the other direction, assume that there exist an induced subgraph $G_{\hat{S}}$, $\hat{S}\subseteq S$, and a labeling function $l:E_{\hat{S}}\to[m]$ that satisfies (C1)--(C3). We will prove that the FPCP admits a solution.
Define
\begin{equation}
\label{defur1}
\tilde{U}(d)=\left\{y-x~|~(x,y)\in E_{\hat{S}},\, l(x,y)=d\right\}
\end{equation}
for all $d\in \{1,\ldots,m\}$, and
\begin{equation}
\label{defur2}
\begin{split}
U(d)& =\tilde{U}(d) \textrm{ for all } d\leq m-1\textrm{,}\\
U(m)& =\tilde{U}(m)\bigcup\left({\mathcal{U}}\backslash\bigcup_{d=1}^{m-1}\hat{U}(d)\right)\textrm{.}
\end{split}
\end{equation}
Clearly, $\{U(1),\ldots,U(m)\}$ is a partition of ${\mathcal{U}}$. We define $\tilde{l}:E\to [m]$ by \eqref{labe}, with $U(d)$ defined as in \eqref{defur1}--\eqref{defur2}. For any $(x,y)\in E_{\hat{S}}$, $\tilde{l}(x,y)=d$ if and only if $y-x\in U(d)$ by \eqref{labe}, which by \eqref{defur1}--\eqref{defur2} implies $l(x,y)=d$. Thus, $\tilde{l}$ and $l$ are the same on $E_{\hat{S}}$, so with a standard abuse of notation, we will refer to $\tilde{l}$ as $l$ in the remainder of the proof.
Let us now define $\hat{u}:\hat{S}\times [m]\to{\mathcal{U}}$ as any function with a following property:
\begin{equation}
\label{defu}
\hat{u}(x,d)\in \{y-x~|~y\in \hat{S},\, (x,y)\in E_{\hat{S}},\, l(x,y)=d\}\textrm{.}
\end{equation}
We note that the existence of a function $\hat{u}$ that satisfies \eqref{defu} follows from (C2), although uniqueness is not guaranteed.
We claim that any system run given by $x(t+1)=x(t)+\hat{u}(x(t),d(t))$ results in the system state remaining within $\hat{S}\subseteq S$, and that $\hat{u}(x(t),d(t))\in U(d(t))$ for all $t\in{\mathbb N}_0$. For the claim that $x(t)\in \hat{S}$ for all $t$, we proceed by induction. By (C1), $x(0)\in \hat{S}$. Assume now that $x(t)\in \hat{S}$. Then, $x(t+1)=x(t)+\hat{u}(x(t),d(t))\in \hat{S}$ by \eqref{defu}.
For the claim that $\hat{u}(x(t),d(t))\in U(d(t))$ for all $t$, we note that $l(x(t),x(t)+\hat{u}(x(t),d(t)))=d(t)$ by \eqref{defu}, so $\hat{u}(x(t),d(t))\in U(d(t))$ by \eqref{defur1}--\eqref{defur2}.
Thus, $\hat{u}$ is a solution to the RPCP for the partition $\{U(1),\ldots,U(m)\}$. Hence, the FPCP admits a solution.
\end{proof}
\begin{remark}
In the latter direction in the proof of Theorem \ref{propo1}, technically we constructed a memoryless policy $\hat{u}:\hat{S}\times [m]\to{\mathcal{U}}$ instead of a memory-conscious policy $\hat{u}:\cup_{i=1}^\infty [m]^i\to{\mathcal{U}}$ as required in the RPCP. Thus, Theorem \ref{propo1} also shows that Game \ref{gm1} and Game \ref{gm2} admit a winning strategy for Player 2 if and only if they admit a {\em memoryless} winning strategy, which was also discussed in \cite{DoyRas11}.
\end{remark}
With Theorem \ref{propo1} in mind, the FPCP can be transformed into the following problem.
{\em Invariant subgraph labeling problem (ISLP):}
Let $S$, $x(0)$, ${\mathcal{U}}$, $m$, and $G$ be as in Game \ref{gm1}. Let $m\in{\mathbb N}$. Determine whether there exist an induced subgraph $G_{\hat{S}}=(\hat{S},E_{\hat{S}})\subseteq G$ with $\hat{S}\subseteq S$ and a labeling $l:E_{\hat{S}}\to [m]$ which satisfy (C1)--(C3).
Let us briefly note that if one was to omit requiring (C3) from the ISLP, such a problem reduces to finding an induced subgraph $G_{\hat{S}}\subseteq G$ with $x(0)\in\hat{S}\subseteq S$ and $\mindeg(G_{\hat{S}})\geq m$. This problem is a variant of the {\em minimum subgraph of minimum degree} problem; see, e.g., \cite{Amietal12} and the references therein.
We now proceed to determine sufficient and necessary conditions for the ISLP to admit a solution.
\section{Conditions for a Good Labeling}
\label{condit}
As discussed above, property (C2) in Theorem \ref{propo1} trivially imposes a simple necessary condition for the ISLP to admit a solution.
\begin{proposition}
\label{lem1}
If there exist an induced subgraph $G_{\hat{S}}$ and a labeling $l$ satisfying the conditions of ISLP, then $$\mindeg(G_{\hat{S}})\geq m\textrm{.}$$
\end{proposition}
The condition given in Proposition \ref{lem1} is not sufficient for existence of a labeling satisfying the conditions of the ISLP. The following example gives an induced subgraph $G_{\hat{S}}\subseteq G$ with $\mindeg(G_{\hat{S}})\geq m$ such that no labeling $l:E_{\hat{S}}\to [m]$ satisfies (C2)--(C3).
\begin{example}
\label{exn}
Consider $n=2$, $m=3$, $x(0)=0$, $S=\{x\in{\mathbb Z}^2~|~\|x\|_1\leq 1\}$, and $U=\{u\in{\mathbb Z}^2~|~\|u\|_\infty=1\}$. Let $\hat{S}=S$. Clearly, $x(0)\in\hat{S}$, and, as illustrated in Fig.~\ref{fig2}, $\mindeg(G_{\hat{S}})\geq m$. Nonetheless, $S$ does not admit a labeling satisfying both (C2) and (C3). Assume otherwise. Let $l:~E_{\hat{S}}\to [m]$ be such a labeling. By (C3), $l$ is translation-invariant. Thus, we denote by $\hat{l}(1)$ the label of all edges that point north (i.e., $(x,y)\in E_{\hat{S}}$ such that $y-x=(0,1)$), $\hat{l}(2)$ the label of NE edges ($(x,y)\in E_{\hat{S}}$ such that $y-x=(1,1)$), $\hat{l}(3)$ for E edges, etc. By applying (C2) to
\begin{enumerate}
\item[(i)] vertices $(0,-1)$, $(-1,0)$, $(0,1)$, and $(1,0)$, respectively, we can conclude that, for each $k\in\{0,1,2,3\}$, $\hat{l}(2k)$, $\hat{l}(2k+1)$, and $\hat{l}(2k+2)$ need to be all different (for ease of notation, we identify $\hat{l}(0)$ with $\hat{l}(8)$),
\item[(ii)] vertex $(0,0)$, we note that $\hat{l}(1)$, $\hat{l}(3)$, $\hat{l}(5)$, and $\hat{l}(7)$ need to have three different values.
\end{enumerate}
\begin{figure}[ht]
\centering
\begin{tikzpicture}[scale=2]
\tikzstyle{point}=[thick,fill=black, draw=black,circle,inner sep=0pt,minimum width=4pt,minimum height=4pt]
\tikzset{->-/.style={decoration={
markings,
mark=at position .5 with {\arrow{Straight Barb[]}}},postaction={decorate}}}
\tikzset{-<-/.style={decoration={
markings,
mark=at position .5 with {\arrow{Straight Barb[reversed]}}},postaction={decorate}}}
\node[point, label=0:{$x(0)$}] at (0,0) {};
\node at (0,1.15) {};
\node[point] at (0,1) {};
\node[point] at (-1,0) {};
\node[point] at (1,0) {};
\node[point] at (0,-1) {};
\draw[->-] (0,0) arc (45:135:0.71);
\draw[->-] (-1,0) arc (225:315:0.71);
\draw[->-] (1,0) arc (45:135:0.71);
\draw[->-] (0,0) arc (225:315:0.71);
\draw[->-] (0,1) arc (135:225:0.71);
\draw[->-] (0,0) arc (-45:45:0.71);
\draw[->-] (0,0) arc (135:225:0.71);
\draw[->-] (0,-1) arc (-45:45:0.71);
\draw[-<-] (0,1) -- (-1,0);
\draw[->-] (0,1) arc (45+70:225-70:2.07);
\draw[-<-] (-1,0) -- (0,-1);
\draw[->-] (-1,0) arc (135+70:315-70:2.07);
\draw[-<-] (0,-1) -- (1,0);
\draw[->-] (0,-1) arc (-135+70:45-70:2.07);
\draw[-<-] (1,0) -- (0,1);
\draw[->-] (1,0) arc (-45+70:135-70:2.07);
\end{tikzpicture}
\caption{An illustration of Example \ref{exn}. The vertices of $\hat{S}=S$ and corresponding directed edges of $E_{\hat{S}}$ are denoted in black.}
\label{fig2}
\end{figure}
Now, from (ii), assume without loss of generality that $\hat{l}(1)=1$, $\hat{l}(3)=2$, and $\hat{l}(5)=3$. Then, by (i) for $k=0$ and $k=1$, $\{\hat{l}(8),\hat{l}(2)\}=\{2,3\}$ and $\{\hat{l}(2),\hat{l}(4)\}=\{1,3\}$. Hence, $\hat{l}(2)=3$, $\hat{l}(8)=2$, and $\hat{l}(4)=1$. Since $\{\hat{l}(4),\hat{l}(6)\}=\{1,2\}$ by (i) for $k=2$, we have $\hat{l}(6)=2$. Thus, $\hat{l}(8)=\hat{l}(6)$, which is in contradiction with (i) for $k=3$.
\hfill $\square$
\end{example}
Even though Proposition \ref{lem1} only gives a necessary condition for the ISLP to admit a solution, there does exist a related sufficient condition. Namely, if there exists an induced subgraph $G_{\hat{S}}$ with large enough $\mindeg(G_{\hat{S}})$, then there exists a labeling of $E_{\hat{S}}$ which solves the FPCP. We prove such a result using the probabilistic method (see, e.g., \cite{Erd59, Erd61, AloSpe08} for more details).
\begin{theorem}
\label{thmbi}
Let $S$, $x(0)$, ${\mathcal{U}}$, and $G=(V,E)$ be as in Game~\ref{gm1}. If there exists a finite induced subgraph $G_{\hat{S}}=(\hat{S},E_{\hat{S}})\subseteq G$ with $x(0)\in\hat{S}\subseteq S$ and \begin{equation}
\label{bou}
\mindeg(G_{\hat{S}})\geq m\ln\left(m|\hat{S}|\right)\textrm{,}
\end{equation}
then there exists a labeling $l:E_{\hat{S}}\to [m]$ such that $G_{\hat{S}}$ and $l$ satisfy properties (C1)--(C3).
\end{theorem}
\begin{proof}
Let us label each element $u\in{\mathcal{U}}$ by $\hat{l}(u)\in [m]$, where each label is chosen independently and uniformly. We define $l:E_{\hat{S}}\to [m]$ by $l(x,y)=\hat{l}(y-x)$.
By definition of $l$, (C3) is satisfied. Property (C1) is also satisfied by the theorem assumptions.
Let $B$ be the event that the label $l$ does not satisfy (C2), i.e., that there exists a vertex $x\in\hat{S}$ such that
\begin{equation}
\label{aux1}
l\left(\{(x,x')\in E_{\hat{S}}~|~x'\in\hat{S}\}\right)\neq [m]\textrm{.}
\end{equation} Define $B_x$ as the event that $l$ satisfies \eqref{aux1} for the particular $x\in\hat{S}$. In particular, define $B_x^i$ as the event that $i\notin l(\{(x,x')\in E_{\hat{S}}~|~x'\in\hat{S}\})$.
If we can show that $\Pr(B)<1$, this will mean that there exists at least one labeling $l$ such that $B$ does {\em not} occur, i.e., that (C1)--(C3) are all satisfied.
By the definitions of $B_x$ and $B_x^i$ and the union bound \cite{Ven12}, we obtain $$\Pr(B)\geq\sum_{x\in\hat{S}}\Pr(B_x)\geq \sum_{\substack{x\in\hat{S} \\
i \in [m]}}\Pr(B_x^i)\textrm{.}$$ Hence, if we show that
\begin{equation}
\label{aux2}
\Pr(B_x^i)<1/(m|\hat{S}|)
\end{equation}
holds for all $x\in\hat{S}$ and $i\in [m]$, we are done.
Consider the event $B_x^i$ for fixed $x\in\hat{S}$ and $i\in [m]$. For each edge $(x,x')\in E_{\hat{S}}$, $x'-x\in{\mathcal{U}}$ is different. Thus, labels $l(x,x')$ have been chosen uniformly and independently. Hence, \begin{equation*}
\begin{split}\Pr(B_x^i)& =\Pr\left(l(x,x')\neq i \textrm{ for all } (x,x')\in E_{\hat{S}}\right) \\ & =\prod_{(x,x')\in E_{\hat{S}}}\Pr\left(l(x,x')\neq i\right)=\prod_{(x,x')\in E_{\hat{S}}} (1-1/m)\textrm{.}
\end{split}
\end{equation*}
Thus, $\Pr(B_x^i)=(1-1/m)^{\deg_{G_{\hat{S}}}(x)}\leq (1-1/m)^{\mindeg(G_{\hat{S}})}$. By simply noting that $(1-1/m)^m<e^{-1}$ (see, e.g., \cite{Mol12}), we obtain $\Pr(B_x^i)\leq (1-1/m)^{\mindeg(G_{\hat{S}})}<e^{-\mindeg(G_{\hat{S}})/m}$. We now obtain \eqref{aux2} from \eqref{bou}.
\end{proof}
Theorem \ref{thmbi} gives a condition for solving the ISLP, i.e., the FPCP, based on finding a suitable subset $\hat{S}$ of the safe set. One way of producing such a subset is by finding a sufficiently dense subgraph of $S$, with a suitable definition of density. In the interest of brevity, we omit further details. We provide two illustrative examples of determining $\hat{S}$ in Section \ref{exams}.
Returning to the running example, construction on the right side of Fig.~\ref{runn}, where $\mindeg(G_{\hat{S}})=m<m\ln (m|\hat{S}|)$, shows that the condition expressed in Theorem \ref{thmbi} is not necessary for the solvability of the FPCP. We will return to this example in Section \ref{exams}, where we provide some intuition for the ``reason'' that it yields a solution to the FPCP, even though it does not satisfy the sufficient condition expressed in Theorem~\ref{thmbi}.
\section{Efficient Labeling Algorithm}
\label{ela}
The proof of Theorem \ref{thmbi} does not provide a mechanism for constructing a good labeling. Instead, it merely states that a uniformly chosen labeling will solve the FPCP with probability $1-\Pr(B)\geq 1-m|\hat{S}|(1-1/m)^{\mindeg(G_{\hat{S}})}$. Thus, an algorithm that randomly chooses labelings until it reaches one that solves the FPCP is going to have expected computational complexity no greater than $$O\left(\frac{|E_{\hat{S}}|}{1-m|\hat{S}|(1-1/m)^{\mindeg(G_{\hat{S}})}}\right)\textrm{,}$$ assuming that a random draw is performed in $O(1)$ time, and including the time to verify whether a labeling satisfies (C2). Thus, if $m|\hat{S}|(1-1/m)^{\mindeg(G_{\hat{S}})}\approx 1$, a randomized algorithm might take a substantial amount of time to finish.
We now present an alternative deterministic algorithm that produces a correct labeling in $O(|E_{\hat{S}}|+|{\mathcal{U}}|m|\hat{S}|)$ operations.
\begin{algorithm}
\label{alg}
Let ${\mathcal{U}}=\{u_1,\ldots,u_{|{\mathcal{U}}|}\}$. Define a labeling $\hat{l}$ on ${\mathcal{U}}$ inductively as follows. Let \begin{equation}
\begin{split}
\label{defli}
l_i\in\argmin_{l'\in [m]}\sum_{\substack{x\in\hat{S} \\
j \in [m]}}& \Pr(B_x^j~|~\hat{l}(u_1)=l_1,\ldots, \\
&\ldots,\hat{l}(u_{i-1})=l_{i-1},\hat{l}(u_i)=l')
\end{split}
\end{equation}
and define $\hat{l}(u_i)=l_i$ for $i=1,2,\ldots,|{\mathcal{U}}|$, where labeling $l~:~E_{\hat{S}}~\to~[m]$ is given by $l(x,y)=\hat{l}(y-x)$ for all $(x,y)\in E_{\hat{S}}$.
\end{algorithm}
\begin{theorem}
Assume that \eqref{bou} holds for a finite induced subgraph $G_{\hat{S}}$ with $x(0)\in\hat{S}\subseteq S$. Let $\hat{l}$ and $l$ be defined as in Algorithm \ref{alg}. Then, $G_{\hat{S}}$ and $l$ satisfy properties (C1)--(C3).
\end{theorem}
\begin{proof}
By \eqref{defli}, for each $i\in[|{\mathcal{U}}|]$, \begin{equation*}
\begin{split}& \sum_{\substack{x\in\hat{S} \\
j \in [m]}}\Pr(B_x^j~|~\hat{l}(u_1)=l_1, \ldots, \hat{l}(u_i)=l_i)\leq \\ & \sum_{k\in[m]}\frac{1}{m}\sum_{\substack{x\in\hat{S} \\
j \in [m]}}\Pr(B_x^j~|~\hat{l}(u_1)=l_1, \ldots, \hat{l}(u_i)=k)= \\
& \sum_{\substack{x\in\hat{S} \\
j \in [m]}}\sum_{k\in[m]}\frac{1}{m}\Pr(B_x^j~|~\hat{l}(u_1)=l_1, \ldots, \hat{l}(u_i)=k)\leq
\\
& \sum_{\substack{x\in\hat{S} \\
j \in [m]}}\Pr(B_x^j~|~\hat{l}(u_1)=l_1, \ldots, \hat{l}(u_{i-1})=l_{i-1})\textrm{.}
\end{split}
\end{equation*}
Hence, inductively,
\begin{equation}
\label{bom}
\begin{split}
\sum_{\substack{x\in\hat{S} \\
j \in [m]}}\Pr(B_x^j~|~\hat{l}(u_1)=l_1,& \ldots, \hat{l}(u_{|{\mathcal{U}}|})=l_{|{\mathcal{U}}|}) \\ \leq\sum_{\substack{x\in\hat{S} \\
j \in [m]}}\Pr(B_x^j)<1\textrm{,}
\end{split}
\end{equation} where the last inequality holds by the proof of Theorem \ref{thmbi}. On the other hand, $l$ is entirely defined by $\hat{l}(u_1), \ldots, \hat{l}(u_{|{\mathcal{U}}|})$. Hence, $\Pr(B_x^j~|~\hat{l}(u_1)=l_1, \ldots, \hat{l}(u_{|{\mathcal{U}}|})=l_{|{\mathcal{U}}|})$ equals either $0$ or $1$ for each $x\in\hat{S}$, $j\in[m]$. By \eqref{bom}, we thus have $\Pr(B_x^j~|~\hat{l}(u_1)=l_1, \ldots, \hat{l}(u_{|{\mathcal{U}}|})=l_{|{\mathcal{U}}|})=0$ for all $x\in\hat{S}$, $j\in[m]$, i.e., $G_{\hat{S}}$ and $l$ satisfy the conditions of the ISLP.
\end{proof}
\begin{proposition}
Algorithm \ref{alg} can be performed in $O(|E_{\hat{S}}|+|{\mathcal{U}}|m|\hat{S}|)$ operations.
\end{proposition}
\begin{proof}
Clearly, the computational complexity of Algorithm \ref{alg} depends on the complexity of solving the optimization problem in \eqref{defli} for each $i\in[|{\mathcal{U}}|]$. For each $i\in[|{\mathcal{U}}|]$, $x\in\hat{S}$, and $j\in[m]$, if $j=l_k$ for some $k\in\{1,\ldots,i\}$ and $x+u_k\in\hat{S}$, then $\Pr(B_x^j~|~\hat{l}(u_1)=l_1, \ldots, \hat{l}(u_i)=l_i)=0$. Otherwise, \begin{equation*}
\begin{split}\Pr(B_x^j~|~\hat{l}(u_1)=l_1, & \ldots, \hat{l}(u_i)=l_i)= \\
& (1-1/m)^{|\{i<k\leq |{\mathcal{U}}|~|~x+u_k\in\hat{S}\}|}\textrm{.}
\end{split}
\end{equation*}
Thus, if we precompute whether $x+u_k\in\hat{S}$ for each $x\in\hat{S}$ and $k\in[|{\mathcal{U}}|]$, and all values $|\{i<k\leq |{\mathcal{U}}|~|~x+u_k\in\hat{S}\}|$, which can be performed in $O(E_{\hat{S}})$ operations, computing \eqref{defli} can be performed in $m|\hat{S}|$ time for each $i\in\{1,\ldots,|{\mathcal{U}}|\}$, by merely updating all $\Pr(B_x^j~|~\hat{l}(u_1)=l_1, \ldots, \hat{l}(u_i)=l_i)$ at the end of step $i$. Hence, Algorithm \ref{alg} indeed operates in $O(E_{\hat{S}}+|{\mathcal{U}}|m|\hat{S}|)$ time.
\end{proof}
Provided with a labeling $l:E_{\tilde{S}}\to[m]$, given in Algorithm \ref{alg}, the system design and control policy which solve the FPCP are given by \eqref{defur1}--\eqref{defur2} and \eqref{defu}. We now proceed to illustrate the obtained results on two practical scenarios.
\section{Examples}
\label{exams}
\subsection{Damaged Vehicle}
Having given conditions for solvability of the RPCP and the FPCP, we return to our running example. Let us consider a vehicle operating on $V={\mathbb Z}^n$, with the ability to either move along the coordinate axes or stay in place, i.e., ${\mathcal{U}}=\{0,\pm e_1,\ldots,\pm e_n\}$. Naturally, only $n\leq 3$ makes direct physical sense. A similar example has been considered in the context of safety games in \cite{NelTop16}. However, in that paper the agent and the adversary alternate in taking control of the vehicle, and the focus of the paper was on efficient computation of safe control policies for a given system design, and not on determining a good system design.
The safety objective that we consider is that the vehicle remains close to its initial position $x(0)=0$, i.e., $S=\{x~|~\|x\|_\infty\leq k\}$ for some $k\in{\mathbb N}_0$. As we showed in Example \ref{exrun}, there exists a safe system design for $n=2$, $m=2$, and $k=1$. In this section, we are interested in discussing the maximal loss of control that still enables a safe system design, i.e., for a given $n$ and $k$, the maximal $m$ such that the FPCP admits a solution.
It is clear that if $k=0$, the agent cannot afford any loss of authority, i.e., the only acceptable $m$ equals $1$. If $k\geq 1$, we claim that the maximal $m$ equals $n+1$.
Let us first show that the FPCP has a solution for $m=n+1$. A partition $\{U_1,\ldots,U_{n+1}\}$ that admits a solution to the RPCP is given by $U_i=\{e_i,-e_i\}$ for $i\leq n$, and $U_{n+1}=\{0\}$. Indeed, analogously to the construction on the right side of Fig.~\ref{runn}, a control policy which alternately chooses $e_d$ and $-e_d$ every time the adversary chooses input $d\in [n]$, and $0$ if the adversary chooses $d=n+1$, results in the agent's state always remaining in $\hat{S}=\{x~|~\|x\|_\infty\leq 1\}$.
On the other hand, if $m\geq n+2$, since there is a total of $2n$ non-zero elements in ${\mathcal{U}}$, for any partition $\{U_1,\ldots,U_m\}$, some partition element $U_j$ will equal $\{e_i\}$ or $\{-e_i\}$ for some $i$. However, by then repeatedly choosing $d(t)=j$, the adversary can be assured that $\|x(t)\|_\infty=t$, i.e., $\|x\|_\infty>k$ after finitely many steps. Thus, the maximal value of $m$ for which the FPCP admits a solution is indeed $n+1$.
If $m=n+1$ and $\hat{S}=S=\{x~|~\|x\|_\infty\leq 1\}$, sufficient condition \eqref{bou} from Theorem \ref{thmbi} does not hold, as $\mindeg(G_{\hat{S}})=m<m\ln(m|\hat{S}|)$. Nonetheless, the solution to the FPCP exists. Let us briefly discuss this gap between sufficiency and necessity of condition \eqref{bou}. The proof of Theorem \ref{thmbi} relies on some degree of genericity of a correct labeling, i.e., a positive probability that a randomly chosen labeling will be correct. On the other hand, the solution to the FPCP when $m=n+1$ is highly structured. Namely, each element of $\{U_1,\ldots,U_m\}$ needs to equal $\{0\}$ or $\{e_i,-e_i\}$ for some $i$. Otherwise, there will exist $U_j$ that equals $\{e_i\}$ or $\{-e_i\}$ for some $i$, and by repeating $d(t)=j$, the adversary will be able to force the system state to move arbitrarily far away from $x(0)$. Hence, the partition that yields a solution to the RPCP is in fact unique up to a permutation: $U(i)=\{e_{i},-e_{i}\}$ for all $i\leq n$, and $U(n+1)=\{0\}$. Thus, as $n$ increases, the probability of a uniformly chosen partition yielding a solution to the RPCP tends to $0$.
\subsection{Communication over a Channel}
We now move from the setting of damaged autonomous systems to that of user-responsive systems. Consider the framework --- originally introduced in \cite{Sha48} --- where, at every time $t$, a message chosen from some finite message set ${\mathcal{M}}$, $|{\mathcal{M}}|=m$, is sent over a communication channel. Each message is encoded as a bit-string (i.e., {\em codeword}) of some fixed length $n$. This codeword does not need to be the same every time the same message is sent; there could be multiple ways to communicate the same message. However, two different messages cannot be encoded in the same way.
The {\em running digital sum} (RDS) $x(t)$ is defined as the vector consisting of differences in the number of $1$'s and $0$'s that were sent in each coordinate of the bit-string until time $t$. Thus, $x(t)$ satisfies \eqref{dtd} for $x(0)=0$, where $u(t)$ is an encoding of the message passed at time $t$, with zeros in the bit-string replaced by $-1$'s, and ${\mathcal{U}}=\{-1,1\}^n$ \cite{CohLit91}. An illustration of such a system for $n=2$ is given in Fig.~\ref{fig}.
\begin{figure}[ht]
\centering
\begin{tikzpicture}[scale=1]
\clip (-2.4,-2.4) rectangle (2.4,2.4);
\tikzstyle{point}=[thick,fill=black, draw=black,circle,inner sep=0pt,minimum width=4pt,minimum height=4pt]
\tikzset{->-/.style={decoration={
markings,
mark=at position .35 with {\arrow{Straight Barb[]}}},postaction={decorate}}}
\tikzset{-<-/.style={decoration={
markings,
mark=at position .35 with {\arrow{Straight Barb[reversed]}}},postaction={decorate}}}
\foreach \x in {-3,...,3}
\foreach \y in {-3,...,3}
{
\node[point] at (\x,\y) {};
\draw[->-] (\x,\y) arc(-45-20:-45+20:2.07);
\draw[->-] (\x,\y) arc(45-20:45+20:2.07);
\draw[->-] (\x,\y) arc(135-20:135+20:2.07);
\draw[->-] (\x,\y) arc(225-20:225+20:2.07);
}
\end{tikzpicture}
\caption{An illustration of the dynamical system that describes the RDS. The vertices of $G$ and the corresponding directed edges of $E$ are denoted in black.}
\label{fig}
\end{figure}
Encoding policies for which the RDS in a channel remains small regardless of the passed messages naturally reduce the effects of various categories of noise \cite{CohLit91}, \cite{Sch04}. Since encodings of different messages are pairwise disjoint, the problem of constructing encoding policies with bounded RDS can be naturally interpreted as the FPCP, with the safe set $S=\{x~|~\|x\|_\infty\leq k\}$. In this section, we are primarily interested in finding the smallest codeword length $n$ such that there exists an encoding policy for which the RDS remains within $S$.
For $k=0$, there clearly does not exist $n$ which yields a solution for the RPCP. For $k=1$, the only $m$ for which there exists an $n$ which yields a solution for the RPCP is $m=1$, and in that case $n=1$ suffices. For $k\geq 2$, a bound on $n$ can be obtained from Theorem \ref{thmbi} as follows.
\begin{proposition}
\label{line}
Let $m,n\in{\mathbb N}$, ${\mathcal{U}}=\{-1,1\}^n$, and $x(0)=0$. Then, if $n\geq 3\max(\log_2 m,11)$, the FPCP admits a solution for $S=\{x\in{\mathbb Z}^n~|~\|x\|_\infty\leq 2\}$.
\end{proposition}
\begin{proof}
Let us define $\hat{S}=V_1\cup V_2$, where $V_1=\{-1,1\}^n$, and $V_2=\{(x_1,\ldots,x_n)\in\{-2,0,2\}^n~|~x_i=0 \textrm{ for at least } n/2\textrm{ } i\textrm{'s}\}$. We note that $x(0)\in\hat{S}\subseteq S$.
Let us examine the outgoing degree $\deg_{G_{\hat{S}}}(v)$ of every vertex $v\in\hat{S}$ in the induced subgraph $G_{\hat{S}}\subseteq G$. If $v=(v_1,v_2,\ldots,v_n)\in V_1$, then
\begin{equation}
\label{deg1}
\deg_{G_{\hat{S}}}(v)=\sum_{i\geq n/2}\binom{n}{i}\geq 2^{n-1}\textrm{,}
\end{equation}
as the set of neighbors of $v$ is given by all vertices $\overline{v}=(\overline{v}_1,\ldots,\overline{v}_n)\in{\mathbb Z}^n$ that satisfy (i) $\overline{v}_i\in\{0,2v_i\}$ for all $i\in[n]$, and (ii) $\overline{v}_i=0$ for at least $n/2$ $i$'s. If $v\in V_2$, then
\begin{equation}
\label{deg2}\deg_{G_{\hat{S}}}(v)\geq 2^{n/2}\textrm{,}
\end{equation}
as the set of neighbors of $v$ is given by all $\overline{v}\in{\mathbb Z}^n$ that satisfy $\overline{v}_i\in\{-1,1\}$ if $v_i=0$, and $\overline{v}_i=v_i/2$ otherwise. Thus, from \eqref{deg1} and \eqref{deg2}, we obtain $\mindeg(G_{\hat{S}})\geq 2^{n/2}$.
We note that $|\hat{S}|=|V_1|+|V_2|\leq 2^n+3^n\leq 3^{n+1}$. Thus, $m\ln(m|\hat{S}|)\leq m\ln m+m(n+1)\ln 3\leq n2^{n/3}\ln(2)/3+(n+1)2^{n/3}\ln (3)$. It can be shown that $n2^{n/3}\ln(2)/3+(n+1)2^{n/3}\ln (3)\leq 2^{n/2}$ for all $n\geq 33$. Thus, the conditions of Theorem \ref{thmbi} are satisfied.
\end{proof}
An illustration of the construction of $\tilde{S}$ used in the proof of Proposition \ref{line} is given in Fig.~\ref{figlin}, for $m=n=2$. We note that Fig.~\ref{figlin} shows that it is possible to construct a labeling (i.e., partition $\{U_1,U_2\}$) even for $n=2$, indicating that the bound in Proposition \ref{line} is very liberal.
\begin{figure}[ht]
\centering
\begin{tikzpicture}[scale=1]
\clip (-2.4,-2.4) rectangle (2.4,2.4);
\tikzstyle{point}=[thick,fill=black, draw=black,circle,inner sep=0pt,minimum width=8pt,minimum height=8pt]
\tikzset{->-/.style={decoration={
markings,
mark=at position .45 with {\arrow{Straight Barb[]}}},postaction={decorate}}}
\tikzset{-<-/.style={decoration={
markings,
mark=at position .45 with {\arrow{Straight Barb[reversed]}}},postaction={decorate}}}
\foreach \x in {-3,...,3}
\foreach \y in {-3,...,3}
{
\node[point,minimum width=4pt,minimum height=4pt] at (\x,\y) {};
\draw[->-,blue,dashed] (\x,\y) arc(-45-20:-45+20:2.07);
\draw[->-,red,dashed] (\x,\y) arc(45-20:45+20:2.07);
\draw[->-,blue,dashed] (\x,\y) arc(135-20:135+20:2.07);
\draw[->-,red,dashed] (\x,\y) arc(225-20:225+20:2.07);
}
\foreach \x in {0,...,2}
{
\node[point,green!70!black] at (\x,2-\x) {};
\node[point,green!70!black] at (-\x,2-\x) {};
\node[point,green!70!black] at (\x,-2+\x) {};
\node[point,green!70!black] at (-\x,-2+\x) {};
}
\node[point,green!70!black] at (0,0) {};
\draw[->-,red, ultra thick] (0,2) arc(225-20:225+20:2.07);
\draw[->-,red, ultra thick] (1,1) arc(225-20:225+20:2.07);
\draw[->-,red, ultra thick] (-1,1) arc(225-20:225+20:2.07);
\draw[->-,red, ultra thick] (0,0) arc(225-20:225+20:2.07);
\draw[->-,red, ultra thick] (-2,0) arc(225-20:225+20:2.07);
\draw[->-,red, ultra thick] (-1,-1) arc(225-20:225+20:2.07);
\draw[->-,red, ultra thick] (0,-2) arc(45-20:45+20:2.07);
\draw[->-,red, ultra thick] (1,-1) arc(45-20:45+20:2.07);
\draw[->-,red, ultra thick] (2,0) arc(45-20:45+20:2.07);
\draw[->-,blue, ultra thick] (0,2) arc(135-20:135+20:2.07);
\draw[->-,blue, ultra thick] (-1,1) arc(135-20:135+20:2.07);
\draw[->-,blue, ultra thick] (1,1) arc(135-20:135+20:2.07);
\draw[->-,blue, ultra thick] (0,0) arc(135-20:135+20:2.07);
\draw[->-,blue, ultra thick] (2,0) arc(135-20:135+20:2.07);
\draw[->-,blue, ultra thick] (1,-1) arc(135-20:135+20:2.07);
\draw[->-,blue, ultra thick] (-2,0) arc(315-20:315+20:2.07);
\draw[->-,blue, ultra thick] (-1,-1) arc(315-20:315+20:2.07);
\draw[->-,blue, ultra thick] (0,-2) arc(315-20:315+20:2.07);
\end{tikzpicture}
\caption{An illustration of a safe labeling for the RDS with $m=2$, $n=2$. Set $\hat{S}$ is denoted in green. A control policy on $\hat{S}$ that ensures safety is described by thicker arrows. We note that each element in $\hat{S}$ has an outgoing thick arrow in each color pointing into $\hat{S}$. Hence, the system state controlled by such a law will always remain within $\hat{S}$, for any $x(0)\in\hat{S}$.}
\label{figlin}
\end{figure}
As it is necessary to use codewords (i.e., bit-strings) of length at least $\lceil\log_2 m\rceil$ to distinguish between $m$ different messages, Proposition \ref{line} states that, if we use three times as many bits as necessary, we can ensure that the RDS stays within the smallest possible bounds. We remark that from the proof of Proposition \ref{line} it is clear that $n\geq 3\max(\log_2 m, 11)$ can be replaced by $n\geq (2+\varepsilon)\max(\log_2 m, n_\varepsilon)$ for any $\varepsilon\geq 0$, where $n_\varepsilon\to\infty$ as $\varepsilon\to 0$.
\section{Conclusion and Future Work}
\label{conc}
This paper presents a preliminary discussion on control, design, and motion planning abilities of an autonomous system where the controller experienced a partial loss of control authority. The paper is primarily interested in developing sufficient and necessary conditions for existence of a safe control policy in such a partly controlled system. In order to obtain these conditions, we interpreted the system motion as a variant of an adversarial safety game on a graph, where one of the player's moves is to label the edges of the game graph. We showed that the safety objective in the original control system is attainable if and only if such a game has a winning strategy, and showed that the game has a winning strategy if and only if there exists a labeling of the game graph that satisfies particular properties. We found a sufficient condition and a necessary condition for the existence of such a labeling in terms of minimal degrees of a subgraph of the original graph, and discussed how those conditions apply to the motion of an autonomous vehicle operating on an $n$-dimensional surface and to communication using a set of codewords of length $n$ with a bounded running digital sum.
The primary avenue of future work is in broadening the scope of the considered framework. In addition to discussing system dynamics more general than \eqref{dtd} --- which may be achieved by considering two-stage motions on a graph, one stage being involuntary ("drift"), and the other resulting from the performed actions --- it is meaningful to consider a broader class of control specifications, rather than solely safety. In general, tasks for autonomous systems are often expressed by a temporal logic specification (e.g., ``visit area $A$ infinitely many times, never go into area $B$, and eventually reach area $C$''). Previous work on designing provably correct control policies --- i.e., policies that are guaranteed to result in the system behavior satisfying a temporal logic specification --- primarily deals with systems whose control abilities are not compromised; see \cite{BaiKat08} for a thorough study. While there is a substantial body of work (see, e.g., \cite{Kupetal01} and the references therein) on systems whose control originally introduced in \cite{Sha48}, may depend on the environment, procedures for determining provably correct control policies for such systems are computationally complex. Providing simple graph-based criteria for existence of a system design that admits a correct control policy would present a significant next step towards ensuring system resilience under partial loss of control authority.
\bibliographystyle{IEEEtran}
|
\section{Introduction\label{sec:Introduction}}
\paragraph*{Background}
Deep learning has achieved great success as in many fields (\cite{lecun2015deep}).
Recent studies have focused on understanding why DNNs, trained by
(stochastic) gradient-based methods, can generalize well, that is,
DNNs often fit the test data well which are not used for training
in practice. Counter-intuitively, although DNNs have many more parameters
than training data, they can rarely overfit the training data in practice.
Several studies have focused on the local property (sharpness/flatness)
of loss function at minima (\cite{hochreiter1995simplifying}) to
explore the DNN's generalization ability. \cite{keskar2016large}
empirically demonstrated that with small batch in each training step,
DNNs consistently converge to flat minima, and lead to a better generalization.
However, \cite{dinh2017sharp} argued that most notions of flatness
are problematic. To this end, \cite{dinh2017sharp} used deep networks
with rectifier linear units (ReLUs) to theoretically show that any
minimum can be arbitrarily sharp or flat without specifying parameterization.
With the constraint of small weights in parameterization, \cite{wu2017towards}
proved that for two-layer ReLU networks, low-complexity solutions
lie in the areas with small Hessian, that is, flat and large basins
of attractor (\cite{wu2017towards}). They then concluded that a random
initialization tends to produce starting parameters located in the
basin of flat minima with a high probability, using gradient-based
methods.
Several studies rely on the concept of stochastic approximation or
uniform stability (\cite{bousquet2002stability,hardt2015train}).
To ensure the stability of a training algorithm, \cite{hardt2015train}
assumed loss function with good properties, such as Lipschitz or smooth
conditions. However, the loss function of a DNN is often very complicated
(\cite{zhang2016understanding}).
Another approach to understanding the DNN's generalization ability
is to find general principles during the DNN's training. Empirically,
\cite{arpit2017closer} suggested that DNNs may learn simple patterns
first in real data, before memorizing. \cite{xu_training_2018} found
a similar phenomenon empirically, which is referred to \emph{Frequency
Principle (F-Principle)}, that is, for a low-frequency dominant function,
DNNs with common settings first quickly capture the dominant low-frequency
components while keeping the amplitudes of high-frequency components
small, and then relatively slowly captures those high-frequency components.
F-Principle can explain how the training can lead DNNs to a good generalization
empirically by showing that the DNN prefers to fit the target function
by a low-complexity function \cite{xu_training_2018}. \cite{rahaman2018spectral}
found a similar result that “Lower Frequencies are Learned First”.
To understand this phenomenon, \cite{rahaman2018spectral} estimated
an inequality that the amplitude of each frequency component of the
DNN output is controlled by the spectral norm of DNN weights theoretically.
\cite{rahaman2018spectral} then show that the spectral norm\footnote{In \cite{rahaman2018spectral}, ``for matrix-valued weights, their
spectral norm was computed by evaluating the eigenvalue of the eigenvector
obtained with 10 power iterations. For vector-valued weights, we simply
use the $L_{2}$ norm''.} increases gradually during training with a small-size DNN empirically.
Therefore, the inequality implies that ``longer training allows the
network to represent more complex functions by allowing it to also
fit higher frequencies''. However, for a large-size DNN, the spectral
norm almost does not change during the training (See an example in
Fig.\ref{fig:SpectralNorm} in Appendix.), that is, the bound of the
amplitude of each frequency component of the DNN output almost does
not change during the training. Then, the inequality in \cite{rahaman2018spectral}
cannot explain why the F-Principle still holds for a large-size DNN.
\paragraph*{Contribution}
In this work, we develop a theoretical framework by Fourier analysis
aiming to understand the training process and the generalization ability
of DNNs with sufficient neurons and hidden layers. We show that for
any parameter, the gradient descent magnitude in each frequency component
of the loss function is proportional to the product of two factors:
one is a decay term with respect to (w.r.t.) frequency; the other
is the amplitude of the difference between the DNN output and the
target function. This theoretical framework shows that DNNs trained
by gradient-based methods endow low-frequency components with higher
priority during the training process. Since the power spectrum of
the tanh function exponentially decays w.r.t. frequency, in which
the exponential decay rate is proportional to the inverse of weight.
We then show that small (large) initialization would result in small
(large) amplitude of high-frequency components, thus leading the DNN
output to a low (high) complexity function with good (bad) generalization
ability. Therefore, with small initialization, sufficient large DNNs
can fit any function (\cite{cybenko1989approximation}) while keeping
good generalization.
We demonstrate that the analysis in this work can be qualitatively
extended to general DNNs. We exemplified our theoretical results through
DNNs fitting natural images, 1-d functions and MNIST dataset (\cite{lecun1998mnist}).
The paper is established as follows. The common settings of DNNs in
this work are presented in Section \ref{sec:Methods}. The theoretical
framework is given in Section \ref{sec:Theoretical-framework}. We
then study the evolution of the mean magnitude of DNN parameters during
the DNN training empirically in Section \ref{sec:The-magnitude-of}.
The theoretical framework is validated by experiments in Section \ref{sec:Understanding-deep-learning}.
The conclusions and discussions are followed in Section \ref{sec:Discussions}.
\section{Methods\label{sec:Methods}}
The activation function for each neuron is tanh. We use DNNs of multiple
hidden layers with no activation function for the output layer. The
DNN is trained by Adam optimizer (\cite{kingma2014adam}). Parameters
of the Adam optimizer are set to the default values (\cite{kingma2014adam}).
The loss function is the mean squared error of the difference between
the DNN output and the target function in the training set.
\section{Theoretical framework\emph{\label{sec:Theoretical-framework}}}
In this section, we will develop a theoretical framework in the Fourier
domain to understand the training process of DNN. For illustration,
we first use a DNN with one hidden layer with tanh function $\sigma(x)$
as the activation function:
\[
\sigma(x)=\tanh(x)=\frac{e^{x}-e^{-x}}{e^{-x}+e^{x}},\quad x\in{\rm \mathbb{R}}.
\]
The Fourier transform of $\sigma(wx+b)$ with $w,b\in{\rm \mathbb{R}}$
is,
\begin{equation}
F[\sigma(wx+b)](k)=\sqrt{\frac{\pi}{2}}\delta(k)+\sqrt{\frac{\pi}{2}}\frac{i}{|w|}\exp\left(-ibk/w\right)\frac{1}{\exp(\pi k/2w)-\exp(-\pi k/2w)},\label{eq:FSigOri}
\end{equation}
where
\[
F[\sigma(x)](k)=\int_{-\infty}^{\infty}\sigma(x)\exp(-ikx){\rm d}x.
\]
Consider a DNN with one hidden layer with $N$ nodes, 1-d input $x$
and 1-d output:
\begin{equation}
\Upsilon(x)=\sum_{j=1}^{N}a_{j}\sigma(w_{j}x+b_{j}),\quad a_{j},w_{j},b_{j}\in{\rm \mathbb{R}}.\label{eq: DNNmath}
\end{equation}
Note that we call all $w_{j}$, $a_{j}$ and $b_{j}$ as \emph{parameters},
in which $w_{j}$ and $a_{j}$ are \emph{weights}, and $b_{j}$ is
a \emph{bias} \emph{term}. When $|\pi k/w_{j}|$ is large, without
loss of generality, we assume $\pi k/w_{j}\gg0$ and $w_{j}>0$,
\begin{equation}
F[\Upsilon](k)\approx\sum_{j=1}^{N}a_{j}\left[\sqrt{\frac{\pi}{2}}\frac{i}{w_{j}}\exp\left(-(ib_{j}+\pi/2)k/w_{j}\right)\right].\label{eq:FTW}
\end{equation}
We define the difference between DNN output and the \emph{target function}
$f(x)$ at each $k$ as
\[
D(k)\triangleq F[\Upsilon](k)-F[f](k).
\]
Write $D(k)$ as
\begin{equation}
D(k)=A(k)e^{i\theta(k)},\label{eq:DAT}
\end{equation}
where $A(k)$ and $\theta(k)\in[-\pi,\pi]$, indicate the amplitude
and phase of $D(k)$, respectively. The loss at frequency $k$ is
$L(k)=\frac{1}{2}\left|D(k)\right|^{2}$, where $|\cdot|$ denotes
the norm of a complex number. The total loss function is defined as:
\[
L=\sum_{k}L(k).
\]
Note that according to the Parseval's theorem\footnote{Without loss of generality, the constant factor that is related to
the definition of corresponding Fourier Transform is ignored here.}, this loss function in the Fourier domain is equal to the commonly
used loss of mean squared error, that is,
\begin{equation}
L=\frac{1}{2}\sum_{x}(\Upsilon(x)-f(x))^{2},\label{eq:Loss1}
\end{equation}
At frequency $k$, the amplitude of the gradient with respect to
each parameter can be obtained (see Appendix for details),
\begin{align}
\frac{\partial L(k)}{\partial a_{j}}=i\left(C_{1}-1\right)\frac{C_{0}}{w_{j}}A(k)\exp\left(-\pi k/2w_{j}\right)\label{eq:DLalpha}
\end{align}
\begin{equation}
\frac{\partial L(k)}{\partial w_{j}}=C_{0}C_{2}\frac{a_{j}}{2w_{j}^{3}}A(k)\exp\left(-\pi k/2w_{j}\right)\label{eq:paritialLA}
\end{equation}
\begin{equation}
\frac{\partial L(k)}{\partial b_{j}}=\left(C_{1}-1\right)C_{0}\frac{a_{j}k}{w_{j}^{2}}A(k)\exp\left(-\pi k/2w_{j}\right),\label{eq:paritialLB}
\end{equation}
where
\begin{equation}
C_{0}=\sqrt{\frac{\pi}{2}}e^{i\left[\theta(k)+b_{j}k/w_{j}\right]}
\end{equation}
\begin{equation}
C_{1}=\exp\left(-2i\left(b_{j}k/w_{j}+\theta(k)\right)\right),\label{eq:C1}
\end{equation}
\begin{equation}
C_{2}=\left[C_{1}\left(i(\pi k-2w_{j})-2b_{j}k\right)+\left(-i(\pi k-2w_{j})-2b_{j}k\right)\right],\label{eq:C2}
\end{equation}
The descent amount at any direction, say, with respect to parameter
$\Theta_{jl}$, is
\begin{equation}
\frac{\partial L}{\partial\Theta_{jl}}=\sum_{l}\frac{\partial L(k)}{\partial\Theta_{jl}}.\label{eq:GDfreq}
\end{equation}
The absolute contribution from frequency $k$ to this total amount
at $\Theta_{jl}$ is
\begin{equation}
\left|\frac{\partial L(k)}{\partial\Theta_{jl}}\right|=A(k)\exp\left(-|\pi k/2w_{j}|\right)G_{jl}(\Theta_{j},k),\label{eq:DL2}
\end{equation}
where $\Theta_{j}\triangleq\{w_{j},b_{j},a_{j}\}$, $\Theta_{jl}\in\Theta_{j}$,
$G_{jl}(\Theta_{j},k)$ is a function with respect to $\Theta_{j}$
and $k$, which can be found in one of Eqs. (\ref{eq:DLalpha}, \ref{eq:paritialLA},
\ref{eq:paritialLB}).
When the component at frequency $k$ does not converge yet, $\exp\left(-|\pi k/2w_{j}|\right)$
would dominate $G_{jl}(\Theta_{j},k)$ for a small $w_{j}$. Therefore,
the behavior of Eq. (\ref{eq:DL2}) is dominated by $A(k)\exp\left(-|\pi k/2w_{j}|\right)$.
This dominant term also indicates that weights are much more important
than bias terms, which will be verified by MNIST dataset later.
To examine the convergence behavior of different frequency components
during the training, we compute the relative difference of the DNN
output and $f(x)$ in the frequency domain at \emph{each recording
step}, that is,
\begin{equation}
\Delta_{F}(k)=\frac{|F[f](k)-F[\Upsilon](k)|}{|F[f](k)|}.\label{eq:dfreq}
\end{equation}
Through the above analysis framework, we have the following theorems
(The proofs can be found at Appendix.)
\begin{thm}
\label{thm:Priority}Consider a DNN with one hidden layer using tanh
function $\sigma(x)$ as the activation function. For any frequencies
$k_{1}$ and $k_{2}$ such that $k_{2}>k_{1}>0$ and there exist $c_{1},c_{2},$
such that $A(k_{1})>c_{1}>0$, $A(k_{2})<c_{2}<\infty$, we have
\begin{equation}
\lim_{\delta\rightarrow0}\frac{\mu\left(\left\{ w_{j}:\left|\frac{\partial L(k_{1})}{\partial\Theta_{jl}}\right|>\left|\frac{\partial L(k_{2})}{\partial\Theta_{jl}}\right|\quad\text{for all}\quad j,l\right\} \cap B_{\delta}\right)}{\mu(B_{\delta})}=1,\label{eq:thm1proof-1}
\end{equation}
where $B_{\delta}$ is a ball with radius $\delta$ centered at the
origin and $\mu(\cdot)$ is the Lebesgue measure of a set.
\end{thm}
Thm \ref{thm:Priority} shows that for any two non-converged frequencies,
almost all sufficiently small weights satisfy that a lower-frequency
component has a higher priority during the gradient-based training.
\begin{thm}
\label{thm:equalprioritySmall}Consider a DNN with one hidden layer
with tanh function $\sigma(x)$ as the activation function. For any
frequencies $k_{1}$ and $k_{2}$ such that $k_{2}>k_{1}>0$, we consider
non-degenerate situation, that is, $\left|F[f](k_{1})\right|>0$,
and there exist positive constants $C_{a}$, $\xi$, $\xi_{1}$, and
$\xi_{2}$, such that $A(k_{2})<C_{a}$, $\left|1-C_{1}(k_{1})\right|>\xi$,
and $\xi_{2}>|C_{2}(k_{1})|>\xi_{1}$. $\forall\epsilon>0$. Then,
for any $\epsilon>0$, there exists $M>0$, for any $w_{j}\in[-M,M]\backslash\{0\}$
such that there exists $\Theta_{jl}\in\{w_{j},b_{j},a_{j}\}$ satisfying
\begin{equation}
\left|\frac{\partial L(k_{1})}{\partial\Theta_{jl}}\right|=\left|\frac{\partial L(k_{2})}{\partial\Theta_{jl}}\right|,\label{eq:neq0-1}
\end{equation}
we have $\Delta_{F}(k_{1})<\epsilon$.
\end{thm}
The case that the gradient amplitude of the low-frequency component
equals to that of the high-frequency component indicates that low
frequency does not dominate. Thm \ref{thm:equalprioritySmall} implies
that when the deviation of the high-frequency component ($A(k_{2})$)
is bounded, we can always find small weights such that the relative
error of the low-frequency component stays very small when the high-frequency
component has the similar priority as the low-frequency component.
In another word, when the DNN starts to fit the high-frequency component,
the low-frequency one stays at the converged state.
\begin{thm}
\label{thm:equalpriorityHigh} Consider a DNN with one hidden layer
with tanh function $\sigma(x)$ as the activation function. For any
frequencies $k_{1}$ and $k_{2}$ such that $k_{2}>k_{1}>0$, we consider
non-degenerate situation, that is, $\left|F[f](k_{1})\right|>0$,
and there exist positive constants $\xi$, $\xi_{1}$, and $\xi_{2}$,
such that $\left|1-C_{1}(k_{1})\right|>\xi$ and $\xi_{2}>|C_{2}(k_{1})|>\xi_{1}$,
for any $B>0$, there exists $M>0$, for any $A(k_{2})>M$, such that
there exists $\Theta_{jl}\in\{w_{j},b_{j},a_{j}\}$ satisfying
\begin{equation}
\left|\frac{\partial L(k_{1})}{\partial\Theta_{jl}}\right|=\left|\frac{\partial L(k_{2})}{\partial\Theta_{jl}}\right|\neq0,\label{eq:neq02-1}
\end{equation}
we have $\Delta_{F}(k_{1})>B$.
\end{thm}
Thm \ref{thm:equalpriorityHigh} implies that for a large $A(k_{2})$,
when the low-frequency component does not converge yet, it already
can be significantly affected by the high-frequency component. A large
$A(k_{2})$ can exist when the target function is high-frequency dominate,
that is, a higher-frequency component has a larger amplitude (See
an example in Appendix \ref{subsec:Fitting-high-frequency-dominant}).
Next, we demonstrate that the analysis of Eq. (\ref{eq:DL2}) can
be qualitatively extended to general DNNs. $A(k)$ in Eq. (\ref{eq:DL2})
comes from the square operation in the loss function, thus, is irrelevant
with DNN structure. $\exp\left(-|\pi k/2w_{j}|\right)$ in Eq. (\ref{eq:DL2})
comes from the exponential decay of the activation function in the
Fourier domain. The analysis of $\exp\left(-|\pi k/2w_{j}|\right)$
is insensitive to the following factors. i) Activation function. The
power spectrum of most activation functions decreases as the frequency
increases; ii) Neuron number. The summation in Eq. (\ref{eq: DNNmath})
does not affect the exponential decay; iii) Multiple hidden layers.
If there are multiple hidden layers, the composition of continuous
activation functions is still a continuous function. The power spectrum
of the continuous function still decays as the frequency increases;
iv) High-dimensional input. We need to consider the vector form of
$k$ in Eq. (\ref{eq:FSigOri}); v) High-dimensional output. The total
loss function is the summation of the loss of each output node. Therefore,
the analysis of $A(k)\exp\left(-|\pi k/2w_{j}|\right)$ of a single
hidden layer qualitatively applies to different activation functions,
neuron numbers, multiple hidden layers, and high-dimensional functions.
\section{The magnitude of DNN parameters during training\label{sec:The-magnitude-of}}
Since the magnitude of DNN parameters is important to the analysis
of the gradients, such as $w_{j}$ in Eq. (\ref{eq:DL2}), we study
the evolution of the magnitude of DNN parameters during training.
Through training DNNs by MNIST dataset, empirically, we show that
for a network with sufficient neurons and layers, the mean magnitude
of the absolute values of DNN parameters only changes slightly during
the training. For example, we train a DNN by MNIST dataset with different
initialization. In Fig.\ref{fig:MNISTnorm}, DNN parameters are initialized
by Gaussian distribution with mean $0$ and standard deviation 0.06,
0.2, 0.6 for (a, b, c), respectively. We compute the mean magnitude
of absolute weights and bias terms. As shown in Fig.\ref{fig:MNISTnorm},
the mean magnitude of the absolute value of DNN parameters only changes
slightly during the training. Thus, empirically, the initialization
almost determines the magnitude of DNN parameters. Note that the magnitude
of DNN parameters can have significant change during training when
the network size is small (We have more discussion in \emph{Discussion}).
\begin{center}
\begin{figure}
\begin{centering}
\includegraphics[scale=0.6]{pic/MNIST/NORM/MNISTNorm}
\par\end{centering}
\caption{Magnitude of DNN parameters during fitting MNIST dataset. DNN parameters
are initialized by Gaussian distribution with mean $0$ and standard
deviation 0.06, 0.2, 0.6 for (a, b, c), respectively. Solids lines
show the mean magnitude of the absolute weights (red) and the absolute
bias (green) at each training epoch. The dashed lines are the mean$\pm$std
for the corresponding color. Note that the green and the red lines
almost overlap. We use a tanh DNN with width: 800-400-200-100. The
learning rate is $10^{-5}$ with batch size $400$. \label{fig:MNISTnorm}}
\end{figure}
\par\end{center}
\section{Experiments\label{sec:Understanding-deep-learning}}
Here, we consider only DNNs with sufficient neurons and layers. Since
initialization is very important to deep learning, we will first discuss
small initialization and then discuss large initialization.
\subsection{Small initialization\label{subsec:Fitting-low-frequency-dominant} }
To show that the F-Principle holds in real data, we train a DNN to
fit a natural image, as shown in Fig.\ref{fig:LowDominate}a---a
mapping from position $(x,y)$ to gray scale strength, which is subtracted
by its mean and then normalized by the maximal absolute value. As
an illustration of F-Principle, we study the Fourier transform of
the image with respect to $x$ for a fixed $y$ (red dashed line in
Fig.\ref{fig:LowDominate}a, denoted as the \emph{target function}
$f(x)$ in the spatial domain). In a finite interval, the frequency
components of the target function can be quantified by Fourier coefficients
computed from Discrete Fourier Transform (DFT). Note that the frequency
in DFT is discrete. The Fourier coefficient of $f(x)$ for the frequency
component $k$ is denoted by $F[f](k)$ (a complex number in general).
$|F[f](k)|$ is the corresponding amplitude. Fig.\ref{fig:LowDominate}b
displays the first $40$ frequency components of $|F[f](k)|$. As
the relative error shown in Fig.\ref{fig:LowDominate}c, the first
four frequency peaks converge from low to high in order. This is consistent
with Thm \ref{thm:Priority}. In addition, the relative difference
of low-frequency components stays small when the DNN is capturing
high-frequency components. This is consistent with Thm \ref{thm:equalprioritySmall}.
Due to the small initialization, as an example in Fig.\ref{fig:LowDominate}d,
when the DNN is fitting low-frequency components, high-frequency components
stay relatively small.
Viewing from the snapshots during the training process, we can see
that DNN captures the image from coarse-grained low-frequency components
to detailed high-frequency components (Fig.\ref{fig:LowDominate}e-\ref{fig:LowDominate}g).
With small weights, the DNN output generalizes well (as shown in Fig.\ref{fig:LowDominate}h),
that is, the test error is very close to the training error.
In addition, this theoretical framework can also explain the more
complicated training behavior of DNNs' fitting of high-frequency dominant
functions (Thm \ref{thm:equalpriorityHigh} and an example in Appendix
\ref{subsec:Fitting-high-frequency-dominant}).
\begin{center}
\begin{figure}
\begin{centering}
\includegraphics[scale=0.7]{pic/Image/42693/Ink/SmallIniImage}
\par\end{centering}
\caption{Convergence from low to high frequency for a natural image. The training
data are all pixels whose horizontal indexes are odd. (a) True image.
(b) $|F[f]|$ of the red dashed pixels in (a) as a function of frequency
index---Note that for DFT, we can refer to a frequency component
by the \emph{frequency index} instead of its physical frequency---with
selected peaks marked by black dots. (c) $\Delta_{F}$ at different
training epochs for different selected frequency peaks in (b). (d)
$|F[f]|$ (red) and $|F[\Upsilon]|$ (green) at epoch 1369. (e-g)
DNN outputs of all pixels at different training steps. (h) Loss functions.
We use a DNN with width 500-400-300-200-200-100-100. We train the
DNN with the full batch and learning rate $2\times10^{-5}$. We initialize
DNN parameters by Gaussian distribution with mean $0$ and standard
deviation $0.08$. \label{fig:LowDominate}}
\end{figure}
\par\end{center}
\subsection{Large initialization}
When the DNN is fitting the natural image in Fig.\ref{fig:LowDominate}a
with small initialization, it generalizes well as shown in Fig.\ref{fig:LowDominate}h.
Here, we show the case of large initialization. Except for a larger
standard deviation of Gaussian distribution for initialization, other
parameters are the same as in Fig.\ref{fig:LowDominate}. After training,
the DNN can well capture the training data, as shown in the left in
Fig.\ref{fig:largeInitSlow}a. However, the DNN output at the test
pixels are very noisy, as shown in the right in Fig.\ref{fig:largeInitSlow}a.
The loss functions of the training data and the test data in Fig.\ref{fig:largeInitSlow}b
also show that the DNN generalizes poorly. To visualize the high-frequency
components of DNN output after training, we study the pixels at the
red dashed lines in Fig.\ref{fig:largeInitSlow}a. As shown in the
left in Fig.\ref{fig:largeInitSlow}c, the DNN accurately fits the
training data. However, for the test data, the DNN output fluctuates
a lot, as shown in Fig.\ref{fig:largeInitSlow}d. Compared with the
case of small initialization, as shown in Fig.\ref{fig:largeInitSlow}e,
the convergence order of the first four frequency peaks are not very
clean. This is consistent with Thm \ref{thm:equalpriorityHigh}.
\begin{center}
\begin{figure}
\begin{centering}
\includegraphics[scale=0.6]{pic/Image/45519/ink/LargeIniImage}
\par\end{centering}
\caption{Analysis of the training process of DNNs with large initialization
while fitting the image in Fig.\ref{fig:LowDominate}a. The weights
of DNNs are initialized by a Gaussian distribution with mean $0$
and standard deviation $0.5$. (a) The DNN outputs at the training
pixels (left) and all pixels (right). (b) Loss functions. (c) DNN
outputs of training data at the red dashed position in (a). (d) DNN
outputs including test data at the red dashed position in (a). (e)
$\Delta_{F}$ at different training epochs for different selected
frequency peaks in Fig.\ref{fig:LowDominate}b. \label{fig:largeInitSlow}}
\end{figure}
\par\end{center}
\subsection{Schematic analysis}
Different initialization can result in very different generalization
ability of DNN's fitting and very different training courses. Here,
we schematically analyze DNN's output after training. With finite
training data points, there is an effective frequency range for this
training set, which is defined as the range in frequency domain bounded
by Nyquist-Shannon sampling theorem (\cite{shannon1949communication})
when the sampling is evenly spaced, or its extensions (\cite{yen1956nonuniform,mishali2009blind})
otherwise. Based on the effective frequency range, we can decompose
the Fourier transform of DNN's output into two parts, that is, effective
frequency range and \emph{extra-higher} frequency range. For different
initialization, since the DNN can well fit the training data, the
frequency components in the effective frequency range are the same.
Then, we consider the frequency components in the extra-higher frequency
range. The amplitude at each frequency for node $j$ is controlled
by $\exp\left(-|\pi k/2w_{j}|\right)$ with a decay rate $|\pi/2w_{j}|$.
This decay rate is large for small initialization. Since the gradient
descent does not drive these extra-higher frequency components towards
any specific direction, the amplitudes of the DNN output in the extra-higher
frequency range stay small. For large initialization, the decay rate
$|\pi/2w_{j}|$ is relatively small. Then, extra-higher frequency
components of the DNN output could have large amplitudes and much
fluctuate compared with small initialization.
Higher-frequency function is of more complexity (for example, \cite{wu2017towards}
uses $\mathbb{E}\left\Vert \nabla_{x}f\right\Vert _{2}^{2}$ to characterize
complexity). With small (large) initialization, the DNN's output is
a low-complexity (high-complexity) function after training. When the
training data captures all important frequency components of the target
function, a low-complexity DNN output can generalize much better than
a high-complexity DNN output.
We next use MNIST dataset to verify the effect of initialization.
We use Gaussian distribution with mean $0$ to initialize DNN parameters.
For simplicity, we use a \emph{two-dimensional vector} $(\cdot,\cdot)$
to denote standard deviations of weights and bias terms, respectively.
Fix the standard deviation for bias terms, we consider the effect
of different standard deviations of weights, that is, $(0.01,0.01)$
in Fig.\ref{fig:MNIST-largeInitSlow}a and $(0.3,0.01)$ in Fig.\ref{fig:MNIST-largeInitSlow}b.
As shown in Fig.\ref{fig:MNIST-largeInitSlow}, in both cases, DNNs
have high accuracy for the training data. However, compared the red
dashed line in Fig.\ref{fig:MNIST-largeInitSlow}a with the yellow
dashed line in Fig.\ref{fig:MNIST-largeInitSlow}b, for the small
standard deviation of weight, the prediction accuracy of the test
data is much higher than that of the large one.
Note that the effect of initialization is governed by weights rather
than bias terms. To verify this, we initialize bias terms with standard
deviation $2.5$. As shown by black curves in Fig.\ref{fig:MNIST-largeInitSlow}a,
the DNN with standard deviation $(0.01,2.5)$ has a bit slower training
course and a bit lower prediction accuracy.
\begin{center}
\begin{figure}
\begin{centering}
\includegraphics{pic/MNIST/Ini/a0d01b2d5_100085/MNISTIni}
\par\end{centering}
\caption{Analysis of the training process of DNNs with different initialization
while fitting MNIST dataset. Illustrations are the prediction accuracy
on the training data and the test data at different training epochs.
We use a tanh DNN with width: 800-400-200-100. The learning rate is
$10^{-5}$ with batch size $400$. DNN parameters are initialized
by Gaussian distribution with mean $0$. The legend $(\cdot,\cdot)$
denotes standard deviations of weights and bias terms, respectively.
\label{fig:MNIST-largeInitSlow}}
\end{figure}
\par\end{center}
\section{Discussions\label{sec:Discussions}}
In this work, we have theoretically analyzed the DNN training process
with (stochastic) gradient-based methods through Fourier analysis.
Our theoretical framework explains the training process when the DNN
is fitting low-frequency dominant functions or high-frequency dominant
functions (See Appendix \ref{subsec:Fitting-high-frequency-dominant}).
Based on the understanding of the training process, we explain why
DNN with small initialization can achieve a good generalization ability.
Our theoretical result shows that the initialization of weights rather
than bias terms mainly determines the DNN training and generalization.
We exemplify our results through natural images and MNIST dataset.
These analyses are not constrained to low-dimensional functions. Next,
we will discuss the relation of our results with other studies and
some limitations of these analyses.
\paragraph{Weight norm}
In this work, we analyze the DNN with sufficient neurons and layers
such that the mean magnitude of DNN parameters keeps almost constant
throughout the training. However, if we use a small-scale network,
the mean magnitude of DNN parameters often increases to a stable value.
Empirically, we found that the training increases the magnitude of
DNN parameters when the DNN is fitting high-frequency components,
for which our theory can provide some insight. If the weights are
too small, the decay rate of $\exp\left(-|\pi k/2w_{j}|\right)$ in
Eq. (\ref{eq:FTW}) will be too large. The training will have to increase
the weights fit the high-frequency components of the target function
because the number of neurons is fixed. By imposing regularization
on the norm of weights in a small-scale network, we can prevent the
training from fitting high-frequency components. Since high-frequency
components usually have small power and are easily affected by noise,
without fitting high-frequency components, the norm regularization
will improve the DNN generalization ability. This discussion is consistent
with other studies (\cite{poggio2018theory,nagarajan2017generalization}).
We will address this topic about the evolution of the mean magnitude
of DNN parameters of small-scale networks in our the future work.
\paragraph{Loss function and activation function}
In this work, we use the mean square error as loss function and tanh
as activation function for the training. Empirically, we found that
by using the mean absolute error as loss function, we can still observe
the convergence order from low to high frequency when the DNN is fitting
low-frequency dominant functions. The term $A(k)$ can be replaced
by other forms that can characterize the difference between DNN outputs
and the target function, by which the analysis of $A(k)$ can be extended.
The key exponential term $\exp\left(-|\pi k/2w_{j}|\right)$ comes
from the property of the activation function. Note that when computing
the Fourier transform of the activation function, we have ignored
the windowing's effect -- which would not change the property of
the activation function in the Fourier domain whose power decays as
the frequency increases. Therefore, for any activation function where
power decreases as the frequency increases and any loss function which
characterizes the difference of DNN outputs and the target function,
the analysis in this work can be qualitatively extended. The exact
mathematical forms of different activation functions and loss functions
can be different. We leave this analysis to our future work.
\paragraph{Sharp/flat minima and generalization}
Since $\exp\left(-|\pi k/2w_{j}|\right)$ exists in all gradient forms
in Eqs. (\ref{eq:DLalpha}, \ref{eq:paritialLA}, \ref{eq:paritialLB}),
$\exp\left(-|\pi k/2w_{j}|\right)$ will also exist in the second-order
derivative of the loss function with respect to any parameter. Here,
we only consider DNN parameters with similar magnitude. With smaller
weights at a minima, the DNN has a good generalization ability along
with that the second-order derivative at the minima is smaller, that
is, a flatter minima. When the weights are very large, the minima
is very sharp. When the DNN is close to a very sharp minima, one training
step can cause the loss function to deviate from the minima significantly
(See a conceptual sketch in Figure 1 of \cite{keskar2016large}).
We also observe that in Fig.\ref{fig:largeInitSlow}, for large initialization,
the loss fluctuates significantly when it is small. Our theoretical
analysis qualitatively shows that a flatter minima is associated with
a better DNN generalization, which resembles the results of other
studies (\cite{keskar2016large,wu2017towards}) (see \emph{Introduction}).
\paragraph{Early stopping }
Our theoretical framework through Fourier analysis can well explain
F-Principle, in which DNN gives low-frequency components with higher
priority as observed in \cite{xu_training_2018}. Thus, our theoretical
framework can provide insight into early stopping. High-frequency
components often have low power (\cite{dong1995statistics}) and are
noisy, but with early stopping, we can prevent DNN from fitting high-frequency
components to achieve a better generalization.
\paragraph{Noise and real data}
Empirical studies found the qualitative differences in gradient-based
optimization of DNNs on pure noise vs. real data (\cite{zhang2016understanding,arpit2017closer,xu_training_2018}).
We would discuss the mechanism underlying these qualitative differences.
\cite{zhang2016understanding} found that the convergence time of
a DNN increases as the label corruption ratio increases. \cite{arpit2017closer}
concluded that DNNs do not use brute-force memorization to fit real
datasets but extract patterns in the data based on experimental findings
in the dataset of MNIST and CIFAR-10. \cite{arpit2017closer} suggests
that DNNs may learn simple and general patterns first in the real
data. \cite{xu_training_2018} found similar results as the following.
With the simple visualization of low-dimensional functions on Fourier
domain, \cite{xu_training_2018} found F-Principle for low-frequency
dominant functions empirically. However, F-Principle does not apply
to pure noise data (\cite{xu_training_2018}).
To theoretically understand the above empirically findings, we first
note that the real data in Fourier domain is usually low-frequency
dominant (\cite{dong1995statistics}) while pure noise data often
do not have clear dominant frequency components, for example, white
noise. For low-frequency dominant functions, DNNs can quickly capture
the low-frequency components. However, for pure noise data, since
the high-frequency components are also important, that is, $A(k)$
in Eq. (\ref{eq:DL2}) could be large for a large $k$, the priority
of low-frequency during the training can be relatively small. During
the training, the gradients of low frequency and high frequency can
affect each other significantly. Thus, the training processes for
real data and pure noise data are often very different. Therefore,
along with the analysis in \emph{Results}, F-Principle thus can well
apply to low-frequency dominant functions but not pure noise data
(\cite{arpit2017closer,xu_training_2018}). Since the DNN needs to
capture more large-amplitude higher-frequency components when it is
fitting pure noise data, it often requires longer convergence time
(\cite{zhang2016understanding}).
\paragraph{Memorization vs. generalization}
Traditional learning theory---which restricts capacity (such as VC
dimension (\cite{vapnik1999overview})) to achieve good generalization---cannot
explain how DNNs can have large capacity to memorize randomly labeled
dataset (\cite{zhang2016understanding}), but still possess good generalization
in real dataset. Another study has shown that sufficient large-scale
DNNs can potentially approximate any function (\cite{cybenko1989approximation}).
In this work, our theoretical framework further resolves this puzzle
by showing that both the DNN structure and the training dataset affect
the effective capacity of DNNs. In the Fourier domain, high-frequency
components can increase the DNN capacity and complexity. With small
initialization, DNNs invoke frequency components from low to high
to fit training data and keep other extra-higher frequency components,
which are beyond the effective frequency range of training data, small.
Thus, DNN's learned capacity and complexity are determined by the
training data, however, they still have the potential of large capacity.
This effect raised from individual dataset is consistent with the
speculation from \cite{kawaguchi2017generalization} that suggests
the generalization ability of deep learning is affected by different
datasets.
\paragraph{Limitations}
i) The theoretical framework in its current form could not analyze
how different DNN structures, such as convolutional neural networks,
affect the detailed training and generalization ability of DNNs. We
believe that to consider the effect of DNN structure, we need to consider
more properties of dataset in addition to that its power decays as
the frequency increases. ii) This qualitative framework cannot analyze
the difference between the depth of layers and the width of layers.
To this end, we need an exact mathematical form for the DFT of the
output of the DNN with multiple hidden layers, which will be left
for our future work. iii) This theoretical framework in its current
form cannot analyze the DNN behavior around local minima/saddle points
during training.
\section*{Acknowledgments}
The author wants to thank Tao Luo for helpful discussion on formalizing
Theorem 1, Wei Dai, Qiu Yang, Shixiao Jiang for critically proofreading
the manuscript. This work was funded by the NYU Abu Dhabi Institute
G1301.
|
\section{Introduction} \label{sec.intro}
In this review I would like to summarize some exciting recent developments in the area
of applying gauge-theory ideas, and in particular Seiberg dualities, to integrable models.\footnote{For presentation slides, see \newline \url{http://www.tfc.tohoku.ac.jp/wp-content/uploads/2018/06/05_Yamazaki_2018SRM-E02.pdf}.}
In the literature there are several different (although related) characterizations of integrable models.
In this lecture integrable models are defined to be the solutions of
the Yang-Baxter equation \cite{Yang:1967bm,Baxter:1972hz} (YBE) {\it with spectral parameters}.
\begin{figure}[htbp]
\centering{\includegraphics[scale=0.18]{Fig1.eps}}
\caption{Graphical representation of the YBE.}
\label{Fig1}
\end{figure}
The graphical representation of YBE is given in Figure \ref{Fig1}.
Here we have three intersecting lines (often called rapidity lines) intersecting. We label them by $1,2,3$,
and we associate three vector spaces $V_1, V_2, V_3$ for each line.
At each crossing of the two lines (say $i$ and $j$)
we associate the so-called R-matrix (see Figure \ref{Fig1_2}):
\be
R_{ij} \in \textrm{End}(V_i\otimes V_j) \;.
\ee
\begin{figure}[htbp]
\centering{\includegraphics[scale=0.18]{Fig1_2.eps}}
\caption{We associate an R-matrix at the crossing of two rapidity lines.}
\label{Fig1_2}
\end{figure}
The YBE states that the product of three R-matrices, where the product is taken in two different orders,
coincide:
\be
&R_{23}R_{13} R_{12} = R_{12} R_{13} R_{23} \in \textrm{End}(V_1\otimes V_2\otimes V_3) \;.
\ee
Here for example $R_{12}\in \textrm{End}(V_1\otimes V_2)$ is extended to a operator
$R_{12} \otimes 1_{V_3} \in \textrm{End}(V_1\otimes V_2\otimes V_3)$, which for simplicity we denoted by the same symbol $R_{12}$.
In this lecture we are interested in YBE with spectral parameters. This means that we have three parameters $z_1, z_2, z_3$
for the three lines $1,2,3$, and the R-matrix at the intersection of rapidity lines $i$ and $j$ depends on the associated spectral parameters $z_i$ and $z_j$:
\be
R_{ij}=R_{ij}(z_i, z_j) \in \textrm{End}(V_i\otimes V_j) \;.
\ee
In almost all of the known integrable models,
we have the property that the R-matrix depends only on the differences between the spectral parameters:
\be
R_{ij}(z_i, z_j)=R_{ij}(z_i-z_j) \;.
\label{rapidity_difference}
\ee
This is known as the rapidity difference property of the R-matrix.
This property holds for almost all the models discussed in this lecture,
and in the following we will assume this property unless stated otherwise.
The YBE, now with spectral parameters, reads
\be
&R_{23}(z_2-z_3) R_{13}(z_1-z_3) R_{12}(z_1-z_2) \\
& \quad = R_{12}(z_1-z_2) R_{13}(z_1-z_3) R_{23}(z_2-z_3) \in \textrm{End}(V_1\otimes V_2\otimes V_3) \;.
\label{YBE2}
\ee
Once we have the YBE, then the textbook-construction of the integrable models (see e.g.\ \cite{Baxter:1982zz}) ensures
that transfer matrices commute with each other, and hence by expanding with respect to the spectral parameters $z$
we learn that the model has infinitely-many commuting conserved charges.
This is another definition of integrability. For this reason YBE has been studied rather intensively in integrable model literature.
While there are many important questions in the field of integrable models,
one of the most fundamental questions is the following:
\begin{equation*}
\textit{Why integrable models exist?}
\end{equation*}
This is a highly non-trivial question. For example, if one naively tries to solve the YBE \eqref{YBE2}
one finds that the equations are over-constrained in general. To see this, let us consider the situation
where all the vector spaces $V_i$ are $N$-dimensional. Then the R-matrix, associated with the crossing of two lines,
has four boundary lines and hence has $O(N^4)$ components. By contrast, the YBE, associated with a figure with six
boundary lines, has $O(N^6)$ components. This over-counting problem is especially severe when $N$ is large, say when $N$ is infinity
(namely when we consider an infinite-dimensional representation),
which is indeed the case in many of the models discussed in this lecture.
It is therefore a fundamental question to understand why integrable models can exist at all.
Once we have a good understanding of this one can then try to understand the pattern of existing integrable models
in the literature, and moreover try to find new integrable models based on such understanding.
I myself has been fascinated by this question over the past years, and have worked mainly on two different approaches, both based on quantum-field-theory ideas.
One approach is based on the ``4d Chern-Simons theory''. This approach was pursued first by Costello
in 2013 \cite{Costello:2013zra,Costello:2013sla}, and more recently developed further in by Costello, Witten and myself \cite{Witten:2016spx,Costello:2017dso,Costello:2018gyb,Costello:2019tri,PartIV}.
Another approach is based on supersymmetric quiver gauge theories.
This approach is called {\it the Gauge/YBE correspondence}, since this correspondence claims rather direct relation between the YBE and the partition functions of gauge theories. This approach
was initiated in 2012-2013 in my papers (partly in collaboration with Terashima) \cite{Yamazaki:2012cp,Terashima:2012cx,Yamazaki:2013nra}.
Our work is based on several previous important ideas in the literature, most notably the integrable models (known as the ``master solution'')
constructed by Bazhanov and Sergeev \cite{Bazhanov:2010kz,Bazhanov:2011mz}.\footnote{There are by now substantial literature on this subject, see e.g.\ \cite{Spiridonov:2010em,Xie:2012mr,Yagi:2015lha,Yamazaki:2015voa,Gahramanov:2015cva,Kels:2015bda,Gahramanov:2016ilb,Bazhanov:2016ajm,Maruyoshi:2016caf,Spiridonov:2016uae,Yamazaki:2016wnu,Gahramanov:2016sen,Yagi:2017hmj,Kels:2017toi,Jafarzade:2017fsc,Gahramanov:2017idz,Gahramanov:2017ysd}.}
By now I have succeeded in reproducing their construction from scratch, and generalized the master solution in a number of different directions,
based on quantum-field-theory techniques.
In the rest of this lecture I will discuss the second approach to this subject.
\section{Basic Ingredients}
While there are many technical complications needed in the actual discussions of the subject,
the fundamental ideas behind the Gauge/YBE correspondence are rather simple,
and can be listed as follows:
\begin{itemize}
\item a statistical lattice as a quiver diagram
\item supersymmetric localization
\item Seiberg(-like) duality
\end{itemize}
\subsection{Statistical Lattice as Quiver Diagram}
\subsubsection{Statistical Mechanical Model}
In textbook presentation of statistical mechanics, one starts with a statistical lattice,
such as the one shown in Figure \ref{Fig2_1}. Let us denote the set of vertices by $V$,
and edges by $E$.
\begin{figure}[htbp]
\centering{\includegraphics[scale=0.4]{Fig2_1.eps}}
\caption{Statistical lattice where a statistical-mechanical model is defined. We can also regard this as a quiver diagram, a defining data for the quiver gauge theory.}
\label{Fig2_1}
\end{figure}
In the statistical-mechanical model,
the dynamical degrees of freedom of the theory,
namely `spins', are placed at the vertices of the lattice: $s_v$ at vertex $v\in V$.
The spins here can take values in any set and can be discrete or continuous.
For example $s_v=\pm 1$ for the Ising model. In this lecture we will encounter
more complicated examples where $s_v$ takes values in e.g.\ $(\mathbb{R}\times \mathbb{Z}_r)^{N-1}$
for integers $r, N$ (with $N>1$) \cite{Yamazaki:2013nra}: in this example we $s_v$ has multiplet components,
and has both discrete and continuous components.
The partition function of the statistical-mechanical model is defined as
\begin{align}
Z:= \sumint \left( \prod_{v \in V} ds_v \right) e^{-\mathcal{E}\left( \{ s_v \}_{v\in V}\right)} \;.
\label{Z_stat}
\end{align}
Here the symbol $\sumint \prod_{v \in V} ds_v$
denotes the summation over the spin configuration $\{s_v \}_{v\in V}$.
When $s_v$ has only discrete (continuous) components this should be understood as
$\sum_{\{s_v \}_{v\in V}}$ ($\int \prod_{v \in V} ds_v$), and in general we sum (integrate)
over the discrete (continuous) part of the spins.
Inside the integrand of \eqref{Z_stat} we have the so-called Boltzmann weight
$\mathcal{E}\left( \{ s_v \}\right)$, a function of spins
given as a sum of contributions from edges and vertices:
\be
\mathcal{E}\left( \{ s_v \}_{v\in V} \right)=\sum_{v \in V} \mathcal{E}^v (s_v)
+\sum_{e \in E} \mathcal{E}^e\left( \{ s_v \}_{v\in e} \right) \;,
\label{L_stat}
\ee
where $v\in e$ means that the vertex $v$ is one of the two endpoints of the edge $e$.
The function $\mathcal{E}^e$ specifies the nearest-neighbor interaction of the spins,
while $\mathcal{E}^v (s_v)$ specifies the local interaction at the vertex (such as the magnetic field dependence in the Ising model,
or the self-interaction between different components of the spin).
\subsubsection{Quiver Gauge Theory}\label{sec.2.1}
Instead of starting with a statistical-mechanical model, we would like to start with a
quantum field theory defined from the statistical lattice.
In high energy physics there are several ideas along these lines.
For example, the lattice gauge theory indeed is defined on a discrete lattice.
We here instead discuss the concept of quiver gauge theory.\footnote{The discussion of the quiver gauge theory here is
somewhat more general than the more standard usage of the word---for example, the matter fields are not necessarily in the bifundamental or adjoint representation.
We will eventually specialize to the more standard construction, see section \ref{sec.details}.}
In quiver gauge theory we regard the statistical lattice
as the ``quiver diagram'', which specifies the gauge/matter content of the theory.
Namely,
\begin{itemize}
\item For each vertex $v\in V$ we associate a gauge field $A_v$ valued in the gauge group $G_v$.
This means that the total gauge group of the theory is
\be
G_{\rm total}=\prod_{v\in V} G_v \;.
\ee
\item For each edge $e \in E$ with two endpoints given by $v, v'\in V$
we associate a matter field $\phi_e$ which transforms non-trivially under
some representation of $G_v\times G_{v'}$, but otherwise
trivially under $G_{v''}$ for $v''\ne v, v'$.
\end{itemize}
In this discussion, the statistical lattice is simply a mnemonic for the
gauge/matter content of the theory,
and does not have any direct relation with the actual spacetime
where the quantum field theory is defined.\footnote{For this reason it is sometimes said that the
quiver diagram lives in the ``theory space''.}
The partition function of this quantum field theory is defined by the path-integral:
\be
Z=\int \prod_v \mathcal{D} A_v \prod_e \mathcal{D} \phi_e \,\, e^{-L\left(\{A_v \}_{v\in V}, \{\phi_e \}_{e\in E}\right)} \;,
\label{Z_QFT}
\ee
where $\mathcal{D} A_v$ and $\mathcal{D} \phi_e$ denote the path-integral measures
for the quantum fields $A_v$ and $\phi_e$.
The integrand, namely the Lagrangian $L$, is a functional of the
field configurations $\{A_v \}_{v\in V}, \{\phi_e \}_{e\in E}$, and is given by
\be
L\left(\{A_v \}_{v\in V}, \{\phi_e \}_{e\in E}\right)
=\sum_{v\in V} L^v\left(A_v \right)
+
\sum_{e\in E} L^e\left(\{A_v \}_{v\in e}, \phi_e \right) \;.
\label{L_QFT}
\ee
Here the term $L^v$ is the kinetic term for the gauge field $A_v$,
and the term $L^e$ is the Lagrangian for the field $\phi_e$,
which couples non-trivially to the gauge fields $\{A_v \}_{v\in e}$.
\subsubsection{Comparison}
There is a striking parallel between the two stories above, between statistical-mechanical models
and quiver gauge theories. For example, the role of the spins $s_v$ are played by the gauge fields
$A_v$, and in the partition functions \eqref{Z_stat} and \eqref{L_stat} are replaced by \eqref{Z_QFT} and \eqref{L_QFT}, respectively.
Encouraged by this parallel, one can try to construct the statistical-mechanical model from quantum field theory.
There are still important differences between the two stories, however.
\begin{itemize}
\item In statistical mechanics all the degrees of freedom (namely spins $s_v$) live at the vertices.
Quiver gauge theory, by contrast, has matter fields $\phi_e$ located at the edges, in addition to the gauge fields $A_v$ located at the vertices. In other words, we have a non-standard
statistical-mechanical model where we have degrees of freedom both at the edges and the vertices.
\item In quiver gauge theory the partition function \eqref{Z_QFT} is defined by the path integral,
and does not necessarily fit with the more conventional discussion of statistical-mechanical models \eqref{Z_stat},
where the partition function is defined by a finite integral/sum.
\end{itemize}
In the next subsection we shall explain how these differences are removed with the help
of supersymmetry.
\subsection{Supersymmetric Localization}\label{sec.2.2}
The second ingredient of the Gauge/YBE correspondence is supersymmetric localization.
Suppose that we supersymmetrize our quiver gauge theory.
The details of how this works depends of course on the
dimensionality and the number of supersymmetries.
While the story below will work in general, let us
for concreteness consider the case of four-dimensional
quiver gauge theory with ${\cal N}=1$ supersymmetry (namely four supercharges).
Then we have
\begin{itemize}
\item For each vertex $v\in V$ we associate an ${\cal N}=1$ vector multiplet $V_v=(A_v, \lambda_v, F_v)$,
whose on-shell degrees of freedom contains a gauge field $A_v$, a gaugino $\lambda_v$, and an auxiliary field $F_v$,
all in the adjoint representation of the
gauge group $G_v$.
\item For each edge $e \in E$ with two endpoints given by $v, v'\in V$
we associate an ${\cal N}=1$ chiral multiplet $\Phi_e=(\phi_e, \psi_e, D_e)$ in a non-trivial representation of $G_v\times G_{v'}$:
this multiplet contains a complex scalar $\phi_e$, a fermion $\psi_e$, and an auxiliary field $D_e$.
\end{itemize}
The definition of the partition function is of course then supersymmetrized:
\be
Z& =\int \prod_v \mathcal{D} V_v \prod_e \mathcal{D} \Phi_e \,\, e^{-\mathcal{L}\left(\{V_v \}_{v\in V}, \{\Phi_e \}_{e\in E}\right)} \;.
\label{Z_SUSY}
\ee
where we denoted the supersymmetrized path-integral measure as
\be
\mathcal{D} V_v = \mathcal{D} A_v \mathcal{D} \lambda_v \mathcal{D} F_v \;, \quad
\mathcal{D} \Phi_e = \mathcal{D} \phi_e \mathcal{D} \psi_e \mathcal{D} D_e \;.
\ee
Then partition function \eqref{Z_SUSY}, when considered in the flat spacetime, is in general a divergent quantity.
We can however regularize the IR divergence by placing the theory on a certain compact manifold $M$,
and after suitably regularizing the UV divergence we can
define a well-defined partition function $Z[M]$;\footnote{We need to keep track of the counterterms to see if $Z[M]$ are indeed regularization independent.
Typically supersymmetric further constrains the possible counterterms, thus making the partition function well-defined.} we use the same formula \eqref{Z_SUSY} but now everything is defined on $M$.
Now the supersymmetric partition function states that we can compute the partition function $Z[M]$,
which is a priori defined as an infinite-dimensional path-integral, reduces to
a finite-dimensional integral/sum:
\be
Z& = \sumint \prod_{v\in V} d\sigma_v \,\, e^{-\mathcal{L}(\sigma_v)} \;.
\label{Z_localize}
\ee
In \eqref{Z_localize} $\sigma_v$ is a finite-dimensional variable associated at the vertex $v\in V$ (namely associated with the gauge group $G_v$),
and as we will see can be continuous or both continuous and discrete variables depending on the choice of the manifold $M$. In the examples below
$\sigma_v$ arises from the holonomies of the gauge field along the non-trivial cycles of $M$.
Note that there is no integral/sum variable associated with the edges of the graph, i.e.\ the matter fields $\phi_e$ vanish at the saddle point locus.
While this is not guaranteed in general supersymmetric localization,
this is indeed the case in all the examples we study in this lecture, and in any case in the following we will assume this.\footnote{In supersymmetric localization the
path-integral reduces to an integral/sum over the saddle point configurations. If all the scalars in the matter multiplets $\phi_e$ are trivial on this saddle point then
there is edge variables in \eqref{Z_localize}.}
Finally, the integrand is $\mathcal{L}(\sigma_v)$ takes the form
\be
\mathcal{L}\left(\{\sigma_v \}_{v\in V} \right)
=\sum_{v\in V} \mathcal{L}^v\left( \sigma_v \right)
+
\sum_{e\in E} \mathcal{L}^e\left(\{\sigma_v \}_{v\in e} \right) \;.
\label{L_localize}
\ee
Here $\mathcal{L}^v$ ($\mathcal{L}^e$) denotes the
classical plus one-loop contributions from the
gauge fields (matter fields). This formula should be compared with the Boltzmann weight for the
statistical-mechanical model \eqref{L_stat}.
\subsubsection{Some More Details}\label{sec.details}
The discussion of the quiver gauge theory above is rather general,
and for practical purposes one needs to specify the theory more in detail.
First, we need to specify the spacetime dimension $D$ for the
the quiver gauge theories are defined---for example, the gauge field $A_v$ introduced above
is $A_v^{\mu}(x)$, where $\mu$ runs from $0, \ldots, D-1$, and $x$ is a point of the $D$-dimensional geometry, such as $\mathbb{R}^{1,D-1}$.
In practice we will consider the case of $D=4,3,2,1$.
Second, we did not specify the representation of the matter field $\phi_e$ under the symmetry $G_v\times G_{v'}$
(where two endpoints of $e$ are denoted by $v, v' \in e$). In this lecture we choose
the gauge group $G_v$ to be an $SU(N_v)$ gauge group (with rank specified by an integer $N_v$),
and the matter field to be in the so-called bifundamental representation,
namely in the representation $(\square, \overline{\square})$ under $G_v\times G_{v'}$. We then need to distinguish the two endpoints,\footnote{This is because the fundamental $\square$ and the anti-fundamental representation $\overline{\square}$ are different representation for $\mathrm{SU}(N)$ groups.
The exception happens for $N=2$, where the fundamental and anti-fundamental representations are isomorphic.}
and hence combinatorially we now need to consider oriented edges, as in Figure \ref{Fig2_2}.
\begin{figure}[htbp]
\centering{\includegraphics[scale=0.4]{Fig2_2.eps}}
\caption{Statistical lattice where a statistical-mechanical model is defined.}
\label{Fig2_2}
\end{figure}
That we need to specify the orientation of the arrow means that the our quiver gauge theory is now chiral.
This is indeed the case for almost all the theories we discuss in this paper (except in section \ref{sec.SU2}). In terms of statistical-mechanical models,
this chirality means that the Boltzmann weight for an edge $e$
is asymmetric with respect to the exchange of the endpoints of the edge:
\begin{align}
\mathcal{E}^e(s_{v}, s_{v'}) \ne \mathcal{E}^e(s_{v'}, s_{v}) \;.
\end{align}
In Figure \ref{Fig2_2} we have chosen the orientation of the edges such that
for each face of the lattice (e.g.\ a square in Figure \ref{Fig2_2}) all the edges around the face are all in the same direction,
either all clockwise or all counterclockwise. This also means that we have a corresponding gauge-invariant product of the
bifundamentals around the face, which we can use as a term in the superpotential:
\be
W=\textrm{Tr}\left[\overset{\curvearrowright}{\prod}_{e\in \textrm{ clockwise face}} \Phi_e \right]-\textrm{Tr}\left[\overset{\curvearrowleft}{\prod}_{e\in \textrm{ counterclockwise face}} \Phi_e \right] \;.
\label{W_rule}
\ee
This will be our general rule in the rest of this lecture.\footnote{This type of rules has been most systematically studied in the context of ``brane tilings'' \cite{Hanany:2005ve,Franco:2005rj}, see e.g.\ \cite{Kennaway:2007tq,Yamazaki:2008bt} for review.}\footnote{We have set the coefficients of the superpotential terms to be $1$. See \cite{Imamura:2007dc} for detailed discussion of the coefficients.}
Finally, we have not the boundary condition for the statistical-mechanical model:
this can be say either periodic boundary condition or fixed boundary condition.
In quiver gauge theory language, the difference is whether to consider the quiver diagram on the torus \cite{Hanany:2005ve,Franco:2005rj}
or on a disc \cite{Franco:2012mm,Xie:2012mr}. For the latter case we have external quiver vertices, whose associated symmetry
we regard as a non-dynamical flavor symmetry.
\subsection{Seiberg(-like) Duality}
The final ingredient is the Seiberg(-like) duality.
This is represented graphically in Figure \ref{Fig3_1}.
\begin{figure}[htbp]
\centering{\includegraphics[scale=0.32]{Fig3_1.eps}}
\caption{Graphical representation of the Seiberg(-like) duality for quiver gauge theories.
The same picture can also represent a quiver mutation in cluster algebras, or
the star--star relation in integrable models.}
\label{Fig3_1}
\end{figure}
The two quivers in Figure \ref{Fig3_1} can be regarded as the defining data for the quiver gauge theory.
As for the vertices, the circles represent the gauge symmetry, where as the squares represent the flavor symmetry.
Here and in the following we choose symmetry group to be $G_v=\mathrm{SU}(N)$ at all the vertices $v\in V$, either gauge or flavor,
and fix this integer $N$ throughout.\footnote{Seiberg duality in itself is known for other gauge groups, however its relevance for the Gauge/YBE correspondence is unknown as of this writing.}
The theory is known as the $\mathrm{SU}(N)$ SQCD with $2N$ flavors (since two arrows comes into the middle circle vertex,
and two arrows out).
The non-Abelian part of the flavor symmetry is $\mathrm{SU}(2N)_L\times \mathrm{SU}(2N)_R$,\footnote{There is also a $U(1)$ R-symmetry, whose role will be discussed in section \ref{sec.spectral}.}
which in this case is broken to its subgroup $\mathrm{SU}(N)^4$.
Such a breaking is caused by the presence of ``mesons'', which are represented as two
arrow connecting a square to another, and hence transforms as bifundamental representation
with respect to the $\mathrm{SU}(N)\times \mathrm{SU}(N) \subset \mathrm{SU}(N)^4$ flavor symmetry.
For the case of 4d ${\cal N}=1$ theories,
the duality between the two theories, which holds in the long distance limit,
is then the famous Seiberg duality \cite{Seiberg:1994pq},
where the case of $N_f=2N$ flavors is right in the middle of the conformal window.
Note that following the previous rule \eqref{W_rule} we here include a cubic superpotential term for each triangle in Figure \ref{Fig3_1}, and such a superpotential term (meson--quark--quark superpotential term) is needed for
the Seiberg duality.
There are similar versions in 3d ${\cal N}=2$ \cite{Aharony:1997gp} and 2d ${\cal N}=(2,2)$ \cite{Gadde:2013dda,Benini:2014mia},\footnote{In two dimensions
the gauge group at the vertex should be $\mathrm{U}(N)$, not $\mathrm{SU}(N)$.} and we will collectively refer to all of them as Seiberg-like dualities, or
simply Seiberg dualities.
For mathematically-inclined readers Figure \ref{Fig3_1} can be thought of as a special example of the
quiver mutation \cite{FominZelevinsky1,FominZelevinsky4}.
We will not really take advantage of this interpretation, partly because Seiberg duality in four dimensions corresponds only to a special subset of the
more general quiver mutation (in two dimensions there is a closer relation between Seiberg duality and quiver mutation \cite{Benini:2014mia}, a fact which we will comment again later in this paper).
It is fair to say that these two interpretations mentioned above of Figure \ref{Fig3_1}, in terms of Seiberg duality and quiver mutation,
are well-known in the literature.
What is less known is that the same graph has
yet another interpretation in statistical mechanics---the star--star relation
(SSR) \cite{Baxter:1986,Bazhanov:1992jqa}.\footnote{Interestingly, Baxter and Bazhanov recognized this relation years before the
discovery of Seiberg duality. The connection to Seiberg duality, however, was noticed only recently.}
The motivation for SSR in the context of integrable models is that
SSR implies YBE, and hence once we solve the SSR we automatically
obtain an integrable model. This fact has far-reaching implications, and is one of the cornerstones of the Gauge/YBE correspondence.
We can now put together all the ingredients discussed in this section.
We can start with Figure \ref{Fig3_1}, which we can regard as the quiver diagram
and associate quiver gauge theories.
Thanks to Seiberg duality, we know that the two theories are dual,
and hence their partition functions coincide in the IR. Since Figure \ref{Fig3_1} is simultaneously the SSR,
we have solved the SSR by computing the supersymmetric partition function of the
quiver gauge theories: that such a partition function has statistical-mechanical interpretation was explained in
section \ref{sec.2.1} and \ref{sec.2.2}. This means that we have solved SSR, and hence have solved the YBE!
\section{YBE as Duality}
While the explanation given in the previous section is enough to understand minimally
how we obtain integrable models, it is instructive to
look into the construction in more detail. Interestingly,
almost all the ingredients in integrable models have counterparts in
quiver gauge theories, and we find a step-by-step parallel.
\subsection{Theory for R-matrix}
For our later purposes it is useful to represent the SSR in Figure \ref{Fig3_1} more symmetrically
by adding half-arrows \cite{Yamazaki:2015voa}. This is shown in Figure \ref{Fig3_2},
where half-arrows are shown as dotted lines.
The rule is that when we have a full arrow in one direction and a half arrow
in the opposite direction, we can then cancel them into a half-arrow (as in Figure \ref{Fig3_2}).
After this manipulation
we obtain an expression as in the bottom of Figure \ref{Fig3_2}.
\begin{figure}[htbp]
\centering{\includegraphics[scale=0.28]{Fig3_2.eps}}
\caption{More symmetric representation of the SSR. Here a dotted line is a half-line, which when combined with a full-line in an opposite orientation
creates a half-line.}
\label{Fig3_2}
\end{figure}
One should note that there is no such thing as a half-matter (say a half-chiral multiplet) in quiver gauge theories.
Such a half-line, however, helps to explain the integrability structure in a more symmetric and unified manner.
Note that in our following discussions half-arrows always connect flavor symmetry gauge groups (squares),
and hence non-dynamical. At the level of the supersymmetric partition functions the only change from the full matter is that
we take a square root of the corresponding classical/one-loop contribution.
In any case, whenever we discuss dualities among quiver gauge theories we can always cancels the half arrows
so that the duality can be stated in terms of full arrows only (we will see this in Figure \ref{Fig8_4}.).
The claim now is that the quivers in Figure \ref{Fig3_2} should be regarded as
the quiver for the R-matrix (Figure \ref{Fig8_1}):
\begin{figure}[htbp]
\centering{\includegraphics[scale=0.33]{Fig8_1.eps}}
\caption{We associate he quiver in the Figure \ref{Fig3_2} to the R-matrix of the integrable model:
the R-matrix is now promoted to a ``theory'' $\mathcal{T}[R]$!}
\label{Fig8_1}
\end{figure}
In other words, the theory $\mathcal{T}[R]$ specified by the quiver\footnote{Since the quiver contains a half-arrow,
the ``theory'' $\mathcal{T}[R]$ is strictly speaking is not an authentic theory, as explained above. We need to add half-arrows appropriately to
make it into an actual theory. It is nevertheless in practice convenient to refer to $\mathcal{T}[R]$ as a theory.} is the
``gauge-theory uplift'' of the R-matrix.
To better understand this statement we need to discuss the YBE for this R-matrix.
\subsection{Gluing as Gauging}
In order to discuss YBE we need to combine the three R-matrices. Let us see how this works
at the level of the quiver diagram, and hence of the quiver gauge theory.
Let us first glue two R-matrices. In the left figure of Figure \ref{Fig8_2} we need to cancel the half-lines in the opposite directions,
to obtain the quiver as in the right of Figure \ref{Fig8_2}.
\begin{figure}[htbp]
\centering{\includegraphics[scale=0.33]{Fig8_2.eps}}
\caption{Gluing of two R-matrices. }
\label{Fig8_2}
\end{figure}
We can next combine three R-matrices as in Figure \ref{Fig8_3}.
We again cancel the two half-arrows in opposite directions; in addition we combine
the two half-arrows into a one full arrow, when the two are in the same direction.
Moreover the vertex in the center of the figure at the contact point of three rhombi
is now turned into a circle, namely the corresponding $\mathrm{SU}(N)$ symmetry is gauged.
In general, the rule is that an internal vertex is gauged, whereas an external vertex is not gauged.\footnote{This rule in natural string theory, where the
$\mathrm{SU}(N)$ symmetry arises from $N$ D-branes. The gauge coupling is finite if the areas wrapped by that D-branes is finite,
otherwise when the D-branes run off to infinity then the area is infinite and the gauge coupling constant is zero, thus making the symmetry non-dynamical.
See e.g.\ \cite{Heckman:2012jh} for a discussion for the type for theories discussed in this paper.}
In this sense we are gluing the three copies of the theory $\mathcal{T}[R]$ by appropriate gauging.
\begin{figure}[htbp]
\centering{\includegraphics[scale=0.33]{Fig8_3.eps}}
\caption{We combine the three R-matrices. Depending on the orientations the two half-arrows
either cancels or combines into a full arrow. The vertex in the center, now an internal vertex, is promoted to the gauge symmetry.}
\label{Fig8_3}
\end{figure}
Note that the idea of ``gluing the theories by gauging'' is has appeared in many physics problems in the past, and has been popular after the discovery of the so-called class $S$ theories \cite{Gaiotto:2009we}. Indeed, many of our discussions here is parallel to the discussion there. For example the role of the trinion theory (the so-called $T_N$ theory) there is played by our R-matrix theory $\mathcal{T}[R]$. The choice of the Riemann surface obtained by gluing trinion, is here played either by the choice of the toric diagram for quivers on the torus \cite{Hanany:2005ve,Franco:2005rj} and the choice of the decorated permutation \cite{Xie:2012mr,Franco:2012mm}\footnote{The decorated permutation labels the positroid cell of the positive Grassmannian,
which combinatorial structure was used intensively in the discussion of scattering amplitudes of 4d ${\cal N}=4$ supersymmetric Yang-Mills theory \cite{ArkaniHamed:2012nw}. One can take this point more seriously and introduce a spectral parameter deformation to the scattering amplitude \cite{Ferro:2012xw,Ferro:2013dga,Bargheer:2014mxa,Ferro:2014gca}. From integrable model viewpoint, this gives a Grassmannian formula for a R-matrix of the Yangians for $\mathfrak{psu}(4|4)$ and $\mathfrak{osp}(6|4)$.}
for quivers on the disc.
\subsection{Yang-Baxter Duality}
We are now ready to discus the YBE. Since the combination of the three R-matrices shown in Figure \ref{Fig8_3}
is the left hand side of the YBE,\footnote{Here we are using the formulation of the integrable model as an IRF (interaction-around-a-face) model.
Our integrable model can also be formulated as a vertex model.} we can do the same for the right hand side and immediately write down the quiver counterpart of the YBE, to obtain Figure \ref{Fig8_4}(b). We can again add half-arrows as in Figure \ref{Fig8_4}(c), and then cancel/combine the half-arrows, to
obtain Figure \ref{Fig8_4}(d).
\begin{figure}[htbp]
\centering{\includegraphics[scale=0.29]{Fig8_4.eps}}
\caption{The quiver counterpart of the YBE. After combining the three R-matrices as in (a)(b). After adding the half-arrows for the external global symmetries as in (c) and canceling/combining the half-arrows on both sides of the identity, we obtain an equality in (d). The gauge theory interpretation of this is that the two quivers in (c) are dual, namely they flow to the same IR fixed point.}
\label{Fig8_4}
\end{figure}
The final equality in Figure \ref{Fig8_4}(d) contains a precise gauge theory statement.
Namely, the quiver gauge theories, described by two quivers in Figure \ref{Fig8_4}(d), are dual, namely they flow to the same IR fixed point.
Since this duality is the precise counterpart of the YBE, let us call this duality {\it the Yang-Baxter duality}.\footnote{When I had the chance to talk with Prof.\ Rodney J.\ Baxter
in 2015, I introduced the terminology ``Yang-Baxter duality''. His immediate reaction was ``I'm not a duality!''. The terminology ``Yang-Baxter duality'' is not meant to be a personal duality between C.N.\ Yang and R.J.\ Baxter; as is hopefully clear the terminology was chosen since this is a duality underlying YBE.}
Of course, the question is then if this duality holds. Fortunately for us, we can show that the Yang-Baxter duality follows
by repeated use of the Seiberg duality, as shown in Figure \ref{Fig8_5}.
Of course, this is essentially the explanation that the SSR implies the YBE, a fact
which was known already in the old literature of integrable models \cite{Baxter:1986,Bazhanov:1992jqa}.
What is new here is that we are now discussing the quiver gauge theories attached to the same figures,
and as we will see this understanding is remarkably powerful in constructing new integrable models.
\begin{figure}[htbp]
\centering{\includegraphics[scale=0.43]{Fig8_5.eps}}
\caption{The Yang-Baxter duality follows by repeated use of the Seiberg duality.
In this figure the blue vertices denote the gauge groups where we perform the Seiberg duality (quiver mutation)
in the next step.}
\label{Fig8_5}
\end{figure}
\section{Spectral Parameter as R-charge}\label{sec.spectral}
\subsection{Zig-Zag Paths and R-charge}
There is still one missing ingredient in our discussion so far: the spectral parameter
in integrable models. As commented before, this parameter is essential in obtaining an infinitely-many conserved charges.
In the Gauge/YBE correspondence, the gauge-theory counterpart of the spectral parameter in integrable models
is the R-charge. Simply put, in both cases the parameters are associated with the so-called ``zig-zag paths''.\footnote{
The relevance of zig-zag paths for quiver gauge theories was pointed out in \cite{Hanany:2005ss}. Zig-zag paths are also discussed in the mathematical literature on bipartite graphs, see e.g. \cite{ThurstonDomino,Goncharov:2011hp}.}
Since a picture is worth a thousand words, the best way to explain the zig-zag path is simply to show Figure \ref{Fig5},
and I hope that the rule is self-explanatory. In Figure \ref{Fig5} the zig-zag paths (colored red) is a path coming in from infinity
and eventually goes off to infinity. They intersect precisely once for each edge of the original quiver diagram (colored black).
\begin{figure}[htbp]
\centering{\includegraphics[scale=0.22]{Fig5.eps}}
\caption{Zig-zag paths, represented by red lines. We have shown the zig-zag paths as dotted or undotted lines,
to make it easier to distinguish between different lines. We can recover the original quiver diagram from the zig-zag paths only.
For this reason the zig-zag paths contain the same data as the original quiver diagram.}
\label{Fig5}
\end{figure}
The zig-zag paths give a nice parametrizations of the R-charges.
Let us label the zig-zag paths by $i=1,2,\ldots$, and let us associate
an angle $\theta_i$ for each path. Since $\theta_i$ is an angle, we have the identification $\theta_i\sim \theta_i+2\pi$.
Suppose that one wants to know the R-charge of the matter multiplet.
In quiver gauge theory such a matter is located at an edge of the quiver diagram,
where two zig-zag paths, say $i$ and $j$, cross.
Then the R-charge of the matter multiplet is
given by $\theta_{ij} / \pi$, where $\theta_{ij}$ is the relative angle between $\theta_i$ and $\theta_j$.
Readers are referred to \cite{Hanany:2005ss,Yamazaki:2012cp,Xie:2012mr}
for more details of this parametrization. Note that the condition that the sum of R-charges is $2$ is analogous to the statement that the
sum of the R-charges is $2\pi$, and the point is that one can promote this analogy into a precise correspondence.\footnote{The idea that the R-charge can be identified with an angle
also appeared in the context of the 3d--3d correspondence \cite{Terashima:2011qi,Terashima:2011xe,Dimofte:2011jd,Dimofte:2011ju,Dimofte:2011py,Cecotti:2011iy}. The relation between Gauge/YBE and 3d--3d deserves further exploration, see \cite{Yamazaki:2012cp,Terashima:2012cx,Yagi:2015lha} for
preliminary discussion.}\footnote{The R-charge in itself is a real parameter. However, we can naturally complexify it into a complex parameter,
see \cite{Festuccia:2011ws} for explanation from supergravity.}
Apart from parametrizing R-charges, the zig-zag paths contain exactly the same data as the original quiver diagram.
In fact, by looking at Figure \ref{Fig5} one can convince oneself that we can recover the original quiver diagram from the zig-zag paths only.
For this reason, we are free to disregard the quiver diagram and keep only the zig-zag paths.\footnote{This is a variant of the Baxter's Z-invariant lattice \cite{Baxter:1978xr}.}
If we do this for the left figure of Figure \ref{Fig6} and deform the paths (while keeping the topology of the graph),
we arrive at the right figure of Figure \ref{Fig6}, which is very similar to the picture for the YBE (recall Figure \ref{Fig1}).
We can also represent the SSR and the YBE in terms of the zig-zag paths, as in Figures \ref{Fig10} and \ref{Fig7}.
The previous statement that SSR implies YBE is now an exercise in graphical calculus: we need to show the equality of the zig-zag paths in Figure \ref{Fig7}
by using the moves in Figure \ref{Fig10}. I would like to invite the readers to check this statement directly to one's satisfaction.
\begin{figure}[htbp]
\centering{\includegraphics[scale=0.17]{Fig6.eps}}
\caption{We can disregard the quiver diagram from Figure \ref{Fig5}, to obtain a set of zig-zag paths as in this figure.
By appropriately deforming the paths (while keeping the topology of the graph) we arrive at the picture on the right,
which is very similar to the picture for the YBE (recall Figure \ref{Fig1}). }
\label{Fig6}
\end{figure}
\begin{figure}[htbp]
\centering{\includegraphics[scale=0.23]{Fig10.eps}}
\caption{Graphical representation of SSR, now in terms of the zig-zag paths.
SSR is here represented by a move involving four zig-zag paths.}
\label{Fig10}
\end{figure}
\begin{figure}[htbp]
\centering{\includegraphics[scale=0.17]{Fig7.eps}}
\caption{YBE as implied by SSR, now reformulated in terms of zig-zag paths. The resulting picture has doubled rapidity lines running parallel in the opposite directions, and has `twists' in several places.}
\label{Fig7}
\end{figure}
There are still some differences from Figure \ref{Fig1}. First, the role of a single rapidity line is now
played by a pair of zig-zag paths which run parallel with each other in the opposite direction.\footnote{By using the parametrization of the zig-zag paths
we actually find a $2$-parameter extension of the R-matrix (in addition to the spectral parameter),
and the Gauge/YBE correspondence generates a solution to a certain generalization of the YBE with these extra parameters, see \cite{Yamazaki:2015voa}.} Second, in the bottom figure of Figure \ref{Fig7} such a pair of zig-zag paths are
twisted in four locations. One can therefore say what we obtain here is a ``doubled twisted'' YBE.
The existence of the `twist' of course does not mean that
we have failed to solve the YBE. Recall that whole construction arises from
solving the YBE as in Figure \ref{Fig8_4}, and we have already seen in Figure \ref{Fig8_5}
that the Seiberg-like duality (solution to SSR) solves the YBE on the nose.
The `twist' in Figure \ref{Fig7} arises only when we try to eliminate the half-arrows and represent everything in terms of zig-zag paths.
The origin of this `twist' can be traced back to the existence of the mesons in the Seiberg duality in Figure \ref{Fig3_2}.
It is interesting to see that the integrable model experts, at least in their own context, knew that mesons are needed for a
proper discussion of SSR (Seiberg-like duality), years before the discovery of Seiberg duality in 1994.
\subsection{Zig-Zag Paths from String Theory}
We have so far introduced the zig-zag paths as purely combinatorial objects. This is often the case in the literature,
either in integrable models or in quiver gauge theories. Fortunately we can do better in string theory
and identify the zig-zag paths as the ``shape of the branes''.
Let us discuss this point briefly for the case of 4d $\mathcal{N}=1$ quiver gauge theories, following the author's Master's thesis \cite{Yamazaki:2008bt}.\footnote{See also \cite{Imamura:2006ie,Imamura:2007dc} for related earlier work. The type IIB brane configuration here is mirror to type IIA description of \cite{Feng:2005gw}. Our discussion here is when the quiver diagram is written on the torus; see e.g.\ \cite{Heckman:2012jh} for quiver diagrams on the disc.}
Let us consider type IIB string theory on
$\mathbb{R}^{3,1}_{0123}\times (\mathbb{C}^{\times})^2_{4567}\times \mathbb{C}_{89}$.
$N$ D5-brane spread in $012357$-directions,
with $57$ directions compactified (namely gives $T^2$), realizing $\mathrm{SU}(N)$ gauge groups.
We in addition have one NS5-brane,
filling the $0123$ directions as well as
a holomorphic curve $\Sigma(x,y)=0$ (with $x=e^{x_4+i x_5}, y=e^{x_6+i x_7}$)
inside $(\mathbb{C}^{\times})^2$ ($4567$-directions), see Table \ref{tbl:D5NS5}.
\begin{table}[htpb]
\caption{The type IIB configuration realizing 4d $\mathcal{N}=1$ quiver gauge theories.}
\begin{center}
\begin{tabular}{c|cccc|cccc|cc}
\hline
\hline
&0&1&2&3&4&5&6&7&8&9 \\
\hline
D5&$\circ$ &$\circ$ &$\circ$ &$\circ$ & & $\circ$ & &$\circ$& & \\
NS5&$\circ$ &$\circ$ &$\circ$ &$\circ$ &\multicolumn{4}{c}{$\Sigma$ (2-dim surface)} & \multicolumn{1}{|c}{} \\
\hline
\end{tabular}
\label{tbl:D5NS5}
\end{center}
\end{table}
The NS5-brane and D5-brane intersect along paths (1-cycles) on the $57$-torus,
and these paths are identified with the zig-zag paths introduced before. The zig-zag paths divide the torus into various regions, realizing the quiver structure of the gauge group (see \cite{Yamazaki:2008bt} for details).\footnote{A similar setup has been discussed more recently in the mathematical work of \cite{Shende:2015mzx}.}
The Seiberg duality (as in Figure \ref{Fig10}) can be understood as a reordering of the five-branes reminiscent of the Hanany--Witten transition \cite{Hanany:1996ie}. Note that in this brane setup
the R-symmetry (which we associate with the spectral parameters) can be identified with the rotation symmetry in the transverse $89$-plane.
Other than providing a natural explanation of the zig-zag paths, the five-brane configuration
could be a useful starting point for exploring other approaches to integrable models,
for example with the ``4d Chern-Simons'' approach mentioned in introduction. This requires further study, see e.g. \cite{Yagi:2015lha,Ashwinkumar:2018tmm,Costello:2018txb,Yamazaki:2019prm}.
\section{\texorpdfstring{$\mathrm{SU}(2)$ and Star-Triangle Relation}{\mathrm{SU}(2) and Star-Triangle Relation}}\label{sec.SU2}
Let us next comment on the special case where
the gauge groups at the quiver vertex is $\mathrm{SU}(2)$ (as oppose to $\mathrm{SU}(N)$ with $N\ge 2$).
This is special since in this case there is no distinction between the fundamental representation and the anti-fundamental representation.
Correspondingly the theory is non-chiral, and there is no need to specify the orientation of the edges of the quiver diagram,
and of the zig-zag paths.
Of course, everything we said so far applies to the special case of $\mathrm{SU}(2)$. The notable difference for the $\mathrm{SU}(2)$ case
is that the Seiberg duality as represented in Figure \ref{Fig3_2} follows in turn from another
Seiberg duality, as shown in Figure \ref{Fig9}.
\begin{figure}[htbp]
\centering{\includegraphics[scale=0.25]{Fig9.eps}}
\caption{For $\mathrm{SU}(2)$ gauge groups we can consider the duality for the star-triangle relation.
This in turn implies the SSR, and hence is more fundamental as a duality. In terms of zig-zag paths
this represents the reshuffling of the positions of the three zig-zag paths. Note that the zig-zag paths are here not oriented, and the theory is non-chiral.}
\label{Fig9}
\end{figure}
In the left figure of Figure \ref{Fig9} the gauge vertex in the middle
has three arrows coming in, and hence we have $N_f=3$ (namely $6$ doublets),
with $\mathrm{SU}(6)$ symmetry broken into $\mathrm{SU}(2)^3$.
This theory is known (again by the Seiberg duality \cite{Seiberg:1994pq}) to a non-gauge theory with $\binom{6}{2}=15$ mesons.
In the right figure of Figure \ref{Fig9}, $12$ of these mesons are shown as edges connecting two flavor $\mathrm{SU}(2)$ symmetries, where the remaining $3$ mesons are in the adjoint representation with respect to the
one of the $\mathrm{SU}(2)$ flavor symmetries.
Figure \ref{Fig9} is known in integrable model literature as the star-triangle relation (STR), one of the expressions for the YBE \cite{Baxter:1982zz}.
We therefore conclude that the duality in Figure \ref{Fig9} generates a solution to the STR \cite{Spiridonov:2010em}.\footnote{The three adjoints are not charged under the gauge symmetry and hence appear as the normalization factor outside the integral. Such a normalization factor has been considered in the
general discussion of the STR.}
In terms of zig-zag paths
Figure \ref{Fig9} means that we can move the relative positions of any three zig-zag paths. This implies the SSR and then in turn the YBE (representation of the Yang-Baxter duality) for our R-matrix; in terms of zig-zag paths both are represented as certain moves of the paths (see Figures \ref{Fig10} and \ref{Fig7}), and these easily follow since we can change the position of any three paths, and hence of any paths.
Summarizing, we find
\begin{equation}
\begin{tabular}{c}
\textrm{STR: ($\mathrm{SU}(2)$ Seiberg duality with $N_f=3$ ($6$ doublets))}\\
$\downarrow$ \\
\textrm{SSR: ($\mathrm{SU}(2)$ Seiberg duality with $N_f=4$ ($8$ doublets))}\\
$\downarrow$\\
\textrm{YBE: (Yang-Baxter duality)} \;.
\end{tabular}
\end{equation}
\section{New Integrable Models}
Having explained the general theory,
let us now turn to the question: which concrete integrable model
arises from the general construction of the previous section?
\subsection{General Lessons}
Before answering this question,
let us first summarize our overall logic as in Figure \ref{Fig.overall}.
\begin{figure}[htbp]
\begin{equation*}
\begin{tikzcd}
& & \textrm{YBE}_1 \arrow[dll,bend right,shift left=1ex,"\textrm{``categoficiation''}"]
\arrow[dddd, dotted,"\textrm{degeneration}"]& \\
\textrm{Yang-Baxter duality}_a \arrow[dddd,dotted, "\textrm{dimensional reduction etc.}"] \arrow[urr,bend left,"Z {[ M_1]}"] \arrow[drrr,bend right,"Z {[ N_1]}"] & & & \\
&& & \textrm{YBE}_2 \arrow[dddd, dotted,"\textrm{degeneration}"]\\
& & & \\
& & \textrm{YBE}_3\arrow[dddd, dotted,"\textrm{degeneration}"] & \\
\textrm{Yang-Baxter duality}_b \arrow[dddd,dotted, "\textrm{dimensional reduction etc.}"]\arrow[urr,bend left, "Z {[ M_2]}"] \arrow[drrr,bend right,"Z {[ N_2]}"] & & & \\
& & & \textrm{YBE}_4 \arrow[dddd, dotted,"\textrm{degeneration}"]\\
& & & \\
& & \textrm{YBE}_5 & \\
\textrm{Yang-Baxter duality}_c \arrow[urr,bend left, "Z {[ M_3]}"] \arrow[drrr,bend right,"Z {[ N_3]}"] & & & \\
& & & \textrm{YBE}_6 \\
\end{tikzcd}
\end{equation*}
\caption{The Yang-Baxter duality ``uplifts'' integrable models into gauge theories and
unifies many integrable models which are unrelated otherwise. Namely, starting with a single duality we can
compute various different supersymmetric partition functions and obtain
many different dualities. In addition we can change the gauge theory side by for example
dimensionally reducing the theory. Such a gauge theory operation, as long as it preserves the duality,
has counterparts in the integrable model side.}
\label{Fig.overall}
\end{figure}
In a typical discussion of integrable models we choose a specific integrable model (a solution of YBE),
for example the one denoted as YBE$_1$ in Figure \ref{Fig.overall}. Rather, here we
are interested in the underlying Yang-Baxter duality (denoted Yang-Baxter duality$_a$ in Figure \ref{Fig.overall}):
this can be thought of as a ``gauge-theory uplift''\footnote{Or a ``categorification'' in a rather loose sense.} of YBE$_1$, and conversely YBE$_1$ is obtained by computing a
supersymmetric partition function $Z[M_1]$ on a manifold $M_1$.
Once we have the Yang-Baxter duality we can systematically construct a variety of integrable models,
by computing a different supersymmetric partition function (for example $Z[N_1]$ in Figure \ref{Fig.overall}).
In this sense the Yang-Baxter duality unifies many different integrable models, which as integrable models
look completely unrelated.
Once we establish the correspondence we can further generalize the setup by
applying various operations on the field theory side. For example, we can dimensionally reduce the theory
along some directions of the compactification manifold, or we can integrate out some matters by giving mass.
As long as these operations preserve the duality\footnote{Since the Yang-Baxter duality is an IR duality, it is a non-trivial question
whether flowing to the IR and dimensional reduction commute. See \cite{Aharony:2013dha} for a related discussion. For our practical purposes we do not necessarily have to go through all the intricacies of the reduction, and could start with the known lower-dimensional version of the Seiberg and Yang-Baxter dualities. The degeneration at the level of the integrable models can then be verified independently.}, one should land on another
version of the Yang-Baxter duality, which we called Yang-Baxter duality$_b$ in Figure \ref{Fig.overall}.
The relation between two different versions of the Yang-Baxter duality (Yang-Baxter duality$_a$ and Yang-Baxter duality$_b$ in Figure \ref{Fig.overall}) on the field theory side can now be translated into the
integrable model side. For example, suppose that we dimensionally reduced the field theory along a circle, and suppose that the manifold $M_1$ contains a circle as a fiber, with the based manifold $M_2$. We then expect that the partition function on $M_1$ should reduce to the partition function on $M_2$,
and this process looks like a degeneration of the associated integrable models.
We can repeat the procedure above, and obtain a whole zoo of integrable models.
\subsection{New Integrable Models}
Let us now make the setup more concrete. We first need to specify which
spacetime dimension we are in (in all the dimensions we consider a supersymmetric theory with four supercharges).
We then need to choose a supersymmetric partition function, which we represent by the choice of the manifold $M$.\footnote{In general even if we specify the geometry $M$ there are still choices to made as to which supergravity backgrounds we consider, and in general different such choices might lead to different answers of supersymmetric partition functions. For our purposes at hand it is in practice sufficient to specify the geometry only,
and hence we use the manifold $M$ itself to represent the choice of the supersymmetric partition function.}
Once we fix the choice of the supersymmetric partition function, the resulting partition function is
in general a function of the certain data of the manifold $M$. For example,
for $S^1\times S^3$ the partition function depends on two continuous parameter ${\mathsf p}$ and ${\mathsf q}$,
which parametrizes the complex structure on $S^1\times S^3$ \cite{Kodaira,Closset:2013vra}.
The partition function is (up to the overall supersymmetric Casimir energy contribution)
the same as the superconformal index of \cite{Kinney:2005ej}, where the parameters ${\mathsf p}, {\mathsf q}$ are the fugacities in the definitions of the index.
\begin{figure}[htbp]
\begin{equation*}
\begin{tikzcd}
\textrm{4d ${\cal N}=1$} \arrow[rr,dashed,dash]&& S^1\times S^3/\mathbb{Z}_r, \arrow[dl,"r\to\infty"]\arrow[rr,dashed,dash] \arrow[dr,"S^1\to 0"]& & S^2\times T^2\arrow[ddll,bend left,"T^2\to\infty"] \arrow[dd,"S^2\to 0"] \\
\textrm{3d ${\cal N}=2$} \arrow[r,dashed,dash]&S^1\times S^2\arrow[dr,"S^1\to 0"] \arrow[rr,dashed,dash] & & S^3/\mathbb{Z}_r \arrow[dl,"r\to\infty"] &\\
\textrm{2d ${\cal N}=(2,2)$} \arrow[rr,dashed,dash]& & S^2 \arrow[rr,dashed,dash]& &T^2\arrow[d, "S^1\to 0"]\\
\textrm{1d ${\cal N}=4$} \arrow[rrrr,dashed,dash]& & && S^1 \\
\end{tikzcd}
\end{equation*}
\caption{A zoo of supersymmetric partition functions (and of associated integrable models). The arrows represents a dimensional reduction into one
dimension lower. The case of $S^1\times S^3/\mathbb{Z}_r$ in four dimensions, whose associated integrable model is the super-master solution of YBE, is particularly general, and contains many other known solutions in the literature.}
\label{Fig.zoo}
\end{figure}
The actual integrable model has been discussed in many papers in the literature. While it is difficult to precisely summarize the rich literature,
it might be useful for readers to loosely classify the literature into: 4d $\mathcal{N}=1$ on $S^1\times S^3$ \cite{Bazhanov:2010kz,Bazhanov:2011mz,Spiridonov:2010em,Yamazaki:2012cp,Terashima:2012cx,Xie:2012mr,Maruyoshi:2016caf,Bazhanov:2016ajm,Yagi:2017hmj}, on $S^1\times S^3/\mathbb{Z}_r$ \cite{Yamazaki:2013nra,Spiridonov:2016uae,Kels:2015bda,Gahramanov:2016ilb,Kels:2017toi},
3d $\mathcal{N}=2$ on $S^1\times S^2$ \cite{Yagi:2015lha,Gahramanov:2015cva,Gahramanov:2016sen,Gahramanov:2017idz}, on $S^3$ \cite{Bazhanov:2007mh,Yamazaki:2012cp,Terashima:2012cx} and $S^3/\mathbb{Z}_r$ \cite{Gahramanov:2016ilb},
2d $\mathcal{N}=(2,2)$ on $S^2$ \cite{Yamazaki:2016wnu}, on $T^2$ \cite{Yagi:2015lha,Yamazaki:2015voa,Jafarzade:2017fsc}, and 1d $\mathcal{N}=4$ on $S^1$ \cite{Yamazaki:2016wnu}.
Rather than listing all the solutions, let us here comment on some notable features of the
resulting integrable models.
The Boltzmann weights are written in terms of some special functions.
For example,
\begin{itemize}
\item For 4d $S^1\times S^3/\mathbb{Z}_r$ we find the elliptic gamma function (as noticed first in \cite{Dolan:2008qi})\footnote{For $r>1$ it is natural to define a certain combination of the elliptic gamma function, known as the lens elliptic gamma function \cite{Yamazaki:2013nra,Kels:2015bda,Kels:2017toi,Spiridonov:2016uae}. It would be interesting to further study the properties of this special function.}
\be
\Gamma(x;{\mathsf p},{\mathsf q})=\prod_{j,k\ge 0} \frac{1-x^{-1} {\mathsf p}^{j+1} {\mathsf q}^{k+1}} {1-x {\mathsf p}^j {\mathsf q}^j} \;, \quad |{\mathsf p}|, |{\mathsf q}|<1 \;.
\label{ell_gamma}
\ee
\item For 3d $\mathcal{N}=2$ on $S^1\times S^2$, on $S^3/\mathbb{Z}_r$
we have the ${\mathsf q}$-Pochhammer symbol\footnote{For $S^3/\mathbb{Z}_r$ we also have the quantum dilogarithm function,
which is written as a certain ratio of the Pochhammer symbols.}
\be
(x;{\mathsf q})=\prod_{j=0}^{\infty} (1-x{\mathsf q}^j) \;, \quad |{\mathsf q}|<1 \;.
\ee
\item For 2d ${\cal N}=(2,2)$ theory on $S^2$ we have the gamma function
\be
\Gamma(x)=\frac{e^{-\gamma x}}{x}\prod_{n=1}^{\infty}
\frac{e^{\frac{x}{n}}}{\left( 1+\frac{x}{n} \right)} \;,
\ee
where $\gamma$ is the Euler--Mascheroni constant.
\end{itemize}
That the theories in 4d, 3d and 2d are related by dimensional reduction is reflected in the following hierarchy of degenerations
in the special functions:\cite{Gadde:2011ia,Dolan:2011rp,Imamura:2011uw,Benini:2012ui,Yamazaki:2013fva}
\be
\Gamma(x;{\mathsf p}, {\mathsf q}) \longrightarrow (x;{\mathsf q}) \longrightarrow \Gamma(x) \;.
\ee
This can be regarded as the elliptic--trigonometric--rational hierarchy of integrable models,
somewhat reminiscent of the Belavin--Drinfeld classification \cite{BelavinDrinfeld} of integrable models.
One should note, however, almost all the integrable models constructed from the Gauge/YBE correspondence does not
fit into the classification scheme of \cite{BelavinDrinfeld}. In \cite{BelavinDrinfeld} the R-matrix is quasi-classical, namely
has a perturbative expansion starting with identity. Our models, however, do not seem to have such a property.\footnote{One can of course divide our R-matrix by some normalization factor, to find a perturbative expansion starting with identity. However, the normalization in itself will then be written as an complicated integral, which looks highly unnatural. Moreover, \cite{BelavinDrinfeld} assumes a particular Ansatz for the subleading piece,
which does not fit well with our R-matrix. Somewhat relatedly, the Hamiltonians of our integrable models, as defined from the expansion of the transfer matrix, seems to have complicated expressions.}
Most of the integrable models in Figure \ref{Fig.zoo} gives rise to solutions to the standard YBE.
The exception is the case of 2d ${\cal N}=(2,2)$ theory on $S^2$, where we obtain a generalization of the
YBE called the ``cluster-enriched YBE'' in \cite{Yamazaki:2016wnu}.
Here we have a hybrid of the YBE with the transformation properties of the
``cluster $y$-variable'' \cite{FominZelevinsky4} in the cluster algebra \cite{FominZelevinsky1}.
Here the $S^2$ partition function is a non-trivial function on the set of FI parameters
associated to the vertices of the quiver diagram,
and the Seiberg-like duality holds only when we change the FI parameter,
whose transformation property coincides with the that in the cluster algebra \cite{Benini:2014mia}.
For 2d ${\cal N}=(2,2)$ theory we obtain the standard YBE.
However, when we write the partition function \eqref{Z_localize}
there is a subtlety in the choice of the integration contour, and the correct prescription \cite{Benini:2013xpa}
is given by the Jeffrey--Kirwan residue \cite{MR1318878}. That we need to use such a residue prescription in the summation over `spins' in integrable models is a new feature of the integrable model.
\subsection{Super-Master Solution}
The case of 4d ${\cal N}=1$ theory on $S^1\times S^3/\mathbb{Z}_r$ \cite{Benini:2011nc,Razamat:2013opa} is particularly interesting.
The associated solution to the YBE \cite{Yamazaki:2013nra}, which we might call the ``super-master solution'' of the SSR (and hence YBE),
is one of the most general solutions to YBE ever known in the literature.
We can consider quiver gauge theories where the gauge group at each node is $\mathrm{SU}(N)$.
Then the model depends on two integers $N$ and $r$. The spins at each vertex takes values in
\be
(\mathbb{R}\times \mathbb{Z}_r)^{N-1} \;.
\ee
This arises from the holonomy of the gauge field, and the two factors $\mathbb{R}$ and $\mathbb{Z}_r$
correspond to the two generators of the fundamental group of $S^1\times S^3/\mathbb{Z}_r$:
\be
\pi_1(S^1\times S^3/\mathbb{Z}_r)= \mathbb{Z}\oplus \mathbb{Z}_r \;.
\ee
The partition function in this case is known as the lens index and was computed in \cite{Benini:2011nc} (see also \cite{Razamat:2013opa}).
In addition to the two integers $N$ and $r$ the partition function depends on
two parameters $p$ and $q$, which parametrizes the complex structure of
the geometry $S^1\times S^3/\mathbb{Z}_r$ \cite{Kodaira}.
While integrability is a consequence of gauge
theory duality, integrability was recently proven mathematically
directly by \cite{Kels:2017toi}, by generalizing several earlier works
(\cite{Kels:2015bda} for $N=2, r>1$, \cite{RainsTransf} for $N\ge 2, r=1$ and $N=2, r=1$ \cite{SpiridonovBeta,MR2281166}). This is arguably the most impressive test of the 4d ${\cal N}=1$ Seiberg duality in the literature.
One might imagine that
the R-matrices for the super-master solution can be understood as
intertwiners for some quantum-group like structure:\footnote{There is a possibility that the integer $r$ is a label for the
representation of a single algebra $\mathcal{U}_{p,q}(\mathfrak{sl}_N)$. There are several indications, however, that this is not the case.}
\be
\mathcal{U}_{p,q;r}(\mathfrak{sl}_N) \;.
\label{Upqr}
\ee
For the special case of $r=1$ and $N=2$, it is known \cite{Zabrodin:2010qm} that this algebra is identified with the
Sklyanin algebra $\mathcal{U}_{p,q}(\mathfrak{sl}_2)$, an algebra defined by Sklyanin \cite{Sklyanin:1982tf}
from the eight-vertex R-matrix \cite{Baxter:1972hz} via the so-called RLL=LLR relation.\footnote{The L-operator for the case $r=1$ is identified as a BPS surface defect inside the 4d ${\cal N}=1$ theory \cite{Maruyoshi:2016caf}.} It is natural to conjecture
that the algebra for $r=1, N>2$ is given by the $\mathfrak{sl}_N$-generalization of the
Sklyanin algebra constructed by Cherednik \cite{CherednikGeneralized}. The algebra for the case of general $r>1$,
even for $N=2$, seems to be unknown. It is a fascinating question to
identify the algebra \eqref{Upqr} for general $r$ and $N$.
\subsubsection{Root-of-Unity Limit}
Sine YBE is an equality, one can take appropriate limit of the parameters
and one could still hope that we have a solution to YBE.
The elliptic--trigonometric--rational degeneration we discussed above,
whose counterpart in quiver gauge theory is the dimensional reduction,
is a good example for this.
There is another particularly interesting limit
worth mentioning. This is the root of unity limit.
Such a limit has been discussed for example in the representation theory of quantum groups.
In our context, this limit is special since the Boltzmann weight diverges.
Let us first consider the case $N=2, r=1$ of the super-master solution.
The Boltzmann weights
are given in terms of the elliptic gamma function (in the notation slightly different from \eqref{ell_gamma})
\be
\Phi(z;{\mathsf p},{\mathsf q}) = \prod_{j,k=0}^{\infty} \frac{1-e^{2i z} {\mathsf p}^{2j+1} {\mathsf q}^{2k+1}}{1-e^{-2i z} {\mathsf p}^{2j+1} {\mathsf q}^{2k+1}} \;.
\ee
Let us consider the limit where ${\mathsf q}$ approaches a $2M$-th root of unity, while ${\mathsf p}$ stays finite:\footnote{Note that this integer $M$ is different from $N$,
which specifies the rank of the gauge groups.}
\be
{\mathsf p}=e^{i\pi \tau} \;,
\quad
{\mathsf q}=e^{-\frac{\epsilon}{2 M^2}} \zeta \;,
\quad
\zeta^{2M}=1 \;.
\label{pq_limit}
\ee
Then $\Phi(z;{\mathsf p},{\mathsf q})$ diverges as \cite{Bazhanov:2010kz,Kels:2017vbc}\footnote{
See \cite{BazhanovReshetikhin,Garoufalidis:2014ifa,Ip:2014pva} for similar discussion for the quantum dilogarithm function.}
\be
\Phi(z;{\mathsf p},{\mathsf q})=\exp\left(
\frac{i}{\epsilon} 2M \int_0^z du \ln \overline{\vartheta}_3 (Nu| N\tau)
\right) \times \textrm{(subleading finite piece)} \;,
\label{Phi_limit}
\ee
where $\overline{\vartheta}_3(x|\tau)$ is the Jacobi theta function
\be
\overline{\vartheta}_3(x\,|\,\tau):= \prod_{n=1}^{\infty} (1+e^{2i x} e^{\pi i \tau (2n-1)}) (1+e^{-2 i x} e^{\pi i \tau (2n-1)}) \;.
\ee
This means that the Boltzmann weight diverges, which schematically reads
\be
Z \longrightarrow \sumint \prod_v d\sigma_v \,\, e^{\frac{1}{\epsilon} W^{(0)}(\sigma) + W^{(1)} (\sigma)+\mathcal{O}(\epsilon) } \;.
\ee
In the limit $\epsilon\to 0$ we need to do the the saddle point analysis,
where the saddle point equation is given by
\be
\frac{\partial W^{(0)}(\sigma)}{\partial \sigma_v}=0 \;.
\ee
for each $v$.
Interestingly, for the case $N=2$ this saddle point equation in itself is identified with a discrete classical integrable equation
known as (Q4) in the list of Adler, Bobenko and Suris \cite{ABS}. In this paper the authors classified discrete integrable equations under several assumptions, and found that all their solutions arise from a degeneration of the most general equation (Q4).
The saddle point equation is also the vacuum equation for the 2d ${\cal N}=(2,2)$ theory,
which is obtained from the 4d theory reduced on $T^2$. In the Gauge/Bethe correspondence of \cite{Nekrasov:2009uh}
this equation should be associated with the Bethe Ansatz equation for the integrable model.
One this note, however, this integrable model is different from ours. See \cite{Kels:2017vbc}
for further discussion on this point.\footnote{Note that the integrable models discussed in the work of Nekrasov and Shatashvili \cite{Nekrasov:2009uh}
and also of Costello \cite{Costello:2013zra} is the six-vertex model (Heisenberg XXZ spin chain) and its generalizations, which are themselves of different class from the chiral Potts models and their generalizations discussed in this lecture. However, there is a known connection between
chiral Potts model and the six-vertex model due to Bazhanov and Stroganov \cite{Bazhanov:1989nc},
and one expects that this will be the key for finding appropriate relations with the Gauge/YBE correspondence (see \cite{Yagi:2017hmj} for recent related
discussion).}\footnote{See \cite{Sadri:2005gi} for yet another relation connection between quiver gauge theories and integrable spin chains.}
Now, once we have a solution to the saddle point equation we can then
evaluate the subleading corrections, to obtain the YBE. By following this strategy
Bazhanov and Sergeev \cite{Bazhanov:2010kz} reproduced the Kashiwara--Miwa model \cite{Kashiwara:1986tu}
as well as the chiral Potts model \cite{vonGehlen:1984bi,AuYang:1987zc,Baxter:1987eq}.
In both cases the spins take values in $\mathbb{Z}_M$, where $M$ was introduced before in \eqref{pq_limit}.
This analysis was recently generalized to the case $r>1$ \cite{Kels:2017vbc} (still $N=2$).
Since the expression for the Boltzmann weight is much more complicated than the $r=1$ case (in particular we already have $\mathbb{Z}_r$ discrete spins even before taking the limit)
we might expect to obtain a new solution of the YBE in the root of unity limit. We found a surprise in the root-of-unity limit of \eqref{pq_limit}:
after suitable change of variables, the $\mathbb{Z}_r$ spin, which was present before taking the limit, combines with the $\mathbb{Z}_M$ spin into a
single $\mathbb{Z}_{Mr}$ spin. We moreover reproduced the same class of models as in the case of $r=1$ (namely Kashiwara--Miwa and chiral Potts models), with the only difference
being $M$ is replaced by $Mr$. We hope that this observation would be of help in the search for the unknown algebraic structure $\mathcal{U}_{p,q;r}(\mathfrak{sl}_N)$ discussed in \eqref{Upqr}.
It is rather interesting that the chiral Potts model arises from the root-of-unity limit of the partition function of supersymmetric quiver gauge theories.
The chiral Potts models is special in that the spectral parameters take values in the higher-genus spectral curve.
Another peculiar feature of the model is that it does not have the rapidity difference property \eqref{rapidity_difference}.
Partly due to these peculiarities the model has long defied understanding from quantum field theory. For example, in the 1989 paper \cite{Witten:1989wf} Witten wrote,
in the paper where he discussed his approach to integrable models from three-dimensional Chern-Simons theory:
\begin{quotation}
There are several obvious areas for further investigation. \dots Another question, which may or may not be related, is to understand the spin models formulated only rather recently in [24] in which the spectral parameter is not an abelian variable (as in previous constructions),
but is a point on a Riemann surface of genus greater than one.
\end{quotation}
In this quote, reference [24] points to the literature on chiral Potts models. One might be able to say that the Gauge/YBE correspondence finally provided an answer
to this question. Interestingly, this paragraph also raises a few more open questions, which reads (again quoting from \cite{Witten:1989wf})
\begin{quotation}
In terms of statistical mechanics, one compelling question is to understand the origin of the spectral parameter
(and the elliptic modulus) in IRF and vertex models; this is essential for explaining the origin of integrability. \dots If possible, one would also like to understand vertex models and quantum groups directly at physical values of $q$, without the less than appealing analytic continuation that is used in this paper.
\end{quotation}
Both these questions are answered in our framework; the origin of the spectral parameter is the R-charge in supersymmetric gauge theories (recall section \ref{sec.spectral}),\footnote{One should add that the Costello's approach \cite{Costello:2013zra,Costello:2013sla} also answers this question, in the setup much closer than us to the original 3d Chern-Simons setup of \cite{Witten:1989wf}.}
and the value of the parameter ${\mathsf q}$ (and one more parameter ${\mathsf p}$) are parameters for the complex structure moduli and hence are continuous parameters from the outset in our context .
\section{Conclusion}
In this lecture we reviewed the subject of the Gauge/YBE correspondence.
For the convenience of the readers we can summarize the
dictionary between integrable models and quiver gauge theories in Table \ref{tab.summary}.
\begin{table}[htbp]
\caption{Dictionary for the Gauge/YBE correspondence.}
\begin{center}
{\renewcommand\arraystretch{1.2}
\begin{tabular}{c|c}
integrable model & quiver gauge theory \\
\hline
spin lattice & quiver diagram \\
chiral & chiral \\
rapidity line & zig-zag path \\
spectral parameter & R-charge \\
statistical partition function & supersymmetric partition function \\
temperature-like parameters & quantum parameters (such as ${\mathsf p}, {\mathsf q}$) \\
spin variables & gauge holonomies along non-contractible cycles \\
self-interaction & gauge multiplet \\
nearest-neighbor interaction & bifundamental matter multiplet \\
star--star relation & Seiberg(-like) duality \\
R-matrix & theory $\mathcal{T}[R]$ (Figure \ref{Fig8_1}) \\
composition of R-matrices & gluing $\mathcal{T}[R]$'s by gauging\\
Yang-Baxter equation & Yang-Baxter duality (Figure \ref{Fig8_4})\\
degeneration & dimensional reduction
\end{tabular}
}
\end{center}
\label{tab.summary}
\end{table}
I began this lecture by asking the question: why integrable models exist?
The Gauge/YBE correspondence provides an elegant answer for this question:
integrable models (solutions to Yang-Baxter equation) exist because
there are underlying gauge theory dualities, namely the Yang-Baxter dualities.
One can still try to go deeper, and ask the question of why gauge theory dualities exist.
One possible answer is that the gauge `symmetry' is really not a symmetry but rather a
redundancy of the description, introduced to make manifest many of the fundamental principles of quantum field theory,
such as locality and unitarity. This means that integrability is ultimately tied with
locality and unitarity. Such an understanding might contain deep insight into the true origin of integrable models.
One novel aspect of the Gauge/YBE correspondence is that
the integrability resides not in each individual quiver gauge theory, but in the ``theory space'' spanned by a {\it class} of quiver gauge theories
glued together by gauging---the conserved charges of integrable models maps one theory to another.
This might be the indication that there are many unknown structures yet to be discovered inside the theory space,
see \cite{YamazakiBook} for exposition of the relevant idea and \cite{Yamazaki:2013xva,Benini:2014mia,Terashima:2013fg} for some related attempts.
In conclusion, I hope that in these lectures I have convinced the reader of the
richness of the subject of the Gauge/YBE correspondence. The correspondence
has lead us to new physics and new mathematics in the past,
and I am convinced that it will continue to do so in the years to come.
\section*{Acknowledgements}
The author would like to take this opportunity to thank
my collaborators in various stages of this research, and to the organizers of the String-Math 2018 conference for
providing a stimulating environment.
The underlying material for this lecture has been presented in several workshops,\footnote{For example,
``Workshop on Geometric Correspondences of Gauge Theories'', ICTP, Sep.\ 2013;
``Integrability in Gauge and String Theory (IGST 2014)'', DESY, Jul.\ 2014;
``IGST 2015'', King's College, Jul.\ 2015;
``Baxter 2015: Exactly Solved Models and Beyond'', Palm Cove, Australia, Jul.\ 2015.}
and he would like to thank the audience for feedback.
This research was supported in part by World Premier International Research Center Initiative (WPI Initiative), MEXT, Japan.,
and by JSPS Grant No.\ 17KK0087.
\bibliographystyle{nb}
|
\section*{Acknowledgment}
This research was funded by TCS Research Scholarship Program, EU FP7 Marie Curie Actions Cleansky Project (Contract No. 607584) and Young Faculty Research Fellowship (Visvesvaraya Ph.D. scheme) from MeitY, Govt. of India.
\section{Conclusion}\label{sec:conclusion}
We proposed, QAware, a novel cross-layer MPTCP scheduler that combines hardware device queue occupancy and TCP RTT for efficient scheduling decisions. We detailed its design and implementation. We evaluated QAware using an extensive set of simulations and real network experiments for various network configurations and applications such as bulk data transfers, web browsing, web file downloads, and video streaming. Comparisons with various state-of-the-art schedulers such as DAPS, BLEST, and ECF were used to demonstrate the efficacy of QAware. It outperformed other schedulers in all network configurations and workloads we tested.
Further, we show that QAware quickly adapts to co-existing applications and sudden variations in network conditions. We have open-sourced QAware's implementation as a modular scheduler for latest stable MPTCP Linux release.
\section{MPTCP real-world performance analysis [SKK: Plan to edit]}
\section{Motivating Use of Cross-Layer Information} \label{sec:incognizance}
\begin{figure}[!t]
\centering
\includegraphics[width=0.4\textwidth]{Figs/network_topology}
\caption{Topology used in experiments and simulations.}
\label{fig:netw_topology}
\end{figure}
Figures~\ref{fig:emc_throughput} and~\ref{fig:emc_rtt} respectively show loading (bits offered per second) and the corresponding estimates of round-trip times (RTT) of two available subflows by the default MPTCP scheduler, minSRTT. They were obtained from controlled testbed experiments and show how the scheduler optimizes over two available TCP subflows that use non-interfering end-to-end paths. The network topology used in the experiment is shown in Figure~\ref{fig:netw_topology}. The last-mile links were WiFi using 802.11g and the rest were $1$ Gbps Ethernet. Neither flow dropped any packets during the length of the experiment.
In the experiment, the default scheduler only utilizes $\approx 60\%$ of available aggregated bandwidth. Observe (Figure~\ref{fig:emc_throughput}) that the default scheduler, more often than not, prefers to send packets on one flow over the other. However, this by itself is not responsible for the low utilization of the available bandwidth. The reason, we argue, is that the default scheduler loads a flow deemed to be the best amongst available flows for \textit{undesirably long} intervals. This is because the scheduler uses only the SRTT of the flows, which is a delayed end-to-end transport layer measurement, for its scheduling decisions.
Consider the RTT of flow $1$ in Figure~\ref{fig:emc_rtt}. The RTT captures in a lagged manner the impact of scheduling decision on the subflow. The consistently high values (see interval $12$s to $14$s in the figure) correspond to an earlier interval of time when the subflow was being assigned packets by the scheduler while it was heavily loaded. That is, the device queue corresponding to the subflow had previously many packets queued at the NIC.
The sharp dip in values (around time $14$s in the figure) captures the transition from when the flow stopped being assigned packets due to high RTT to when it was again assigned packets. These assigned packets arrive at a rather lightly loaded flow and see much smaller RTT, which causes the dip. The small RTT that follows the dip corresponds to packets being assigned to the flow while it was still lightly loaded. As the subflow continues to be assigned packets, the same is reflected, albeit in a delayed manner, in increasing RTT (seconds $16$ to $18$ in Figure~\ref{fig:emc_rtt}) that eventually peaks as it did during $12-14$ seconds. By the time the resulting large RTT makes the scheduler switch to the other flow, the scheduler has already spent an undesirably long time injecting packets to a loaded subflow.
In summary, the scheduling decisions that lead to high device queue occupancy and increase in RTT were made using values of RTT that corresponded to an earlier interval when the flow was less loaded. So while a device queue (local to the MPTCP sender and used by the MPTCP flow) is loaded with packets, MPTCP scheduler remains oblivious to the same. Instead, it waits to be informed via a delayed end-to-end RTT based feedback mechanism. In the process, it loses out on many opportunities of scheduling packets to the other better flow; one that is lightly loaded.
The above observations motivate QAware. It uses the occupancy of the device queues together with RTT estimates to use all available flows more efficiently.
\section{Discussion and Future Work}\label{sec:discussion}
The efficiency of MPTCP is mainly dependent on its choice of path scheduler. In this paper, we argue that the current way of treating the individual subflows by the default and several state-of-the-art schedulers, i.e., practically separately relying only on their TCP level information, is not sufficient. A scheduler's reliance on TCP-level information disables it to fully leverage the available capacity and a more holistic approach to MPTCP and how it manages the connections is needed.
In this paper, we show via controlled experiments that the MPTCP's default minSRTT scheduler essentially forces the protocol to use only one of the many available flows and thus leads to lower performance. We develop a novel MPTCP path scheduler, QueueAware, which exploits hardware device queue occupancy along with TCP RTT for robust and effective scheduling decisions. We implement QueueAware as a modular scheduler in latest stable MPTCP Linux implementation release. We evaluate QueueAware in simulated and real networks for various network configurations and applications such as bulk data transfers, web browsing, web file downloads, video streaming, etc. By comparing QueueAware with minSRTT and several other state-of-the-art schedulers such as DAPS, BLEST, and ECF, we show that our approach outperforms all other schedulers and can adapt to changes in network conditions and background traffic more effectively and efficiently. This results in increased throughput, efficient capacity utilization and better user Quality-of-Experience.
For future work, we intend to investigate the effects of MPTCP path schedulers on several other network factors such as congestion at an access point, over-utilization of channel capacity, queueing delay at receiver etc. We believe that several such network aberrations can be overcome by utilizing network-specific aspects, such as congestion, and receiver-specific parameters (such as window size) etc. Explicitly monitoring, predicting and incorporating these parameters with local network variables and TCP-level information is an interesting research question of its own.
\section{Evaluation Methodology}\label{sec:evaluation}
In following sections, we evaluate QAware's performance through an extensive set of simulations and real-world experiments. We model our evaluation methodology to mimic real MPTCP network configurations and application use-cases. In majority of our evaluation, we model a realistic network scenario (as illustrated in Figure \ref{fig:netw_topology}) wherein a client leverages two distinct network paths to connect to a distant server.
For simulations, we implement QAware on ns-3 network simulator. We compare QAware with default minSRTT and Earliest Completion First (ECF)~\cite{ecf} scheduler for constant bit rate (CBR), file downloads, and web browsing workloads. The simulations help us zoom into the workings of the schedulers and allow us to evaluate QAware over a variety of workloads and network path configurations. Our evaluation setup and results are described in Section \ref{sec:simulation}.
We further examine and validate the performance gains obtained by QAware in simulated environments via real network experiments. We utilize our Kernel implementation summarized in Section \ref{sec:implementation}. The experiments were performed in a university data center and consider a variety of workloads such as video streaming, web file downloads, etc. We compare QAware with several state-of-the-art schedulers such as minSRTT, Delay Aware Packet Scheduler (DAPS)~\cite{daps}, Blocking Estimation based scheduler (BLEST)~\cite{blest}, and ECF~\cite{ecf}. The details of our experiments and consequent results are discussed in Section \ref{sec:realexperiments}. All our results throughout evaluation are averaged over multiple runs.
\section{Implementation}
\label{sec:implementation}
We implement QAware as a modular scheduler using MPTCP v0.93 based on Linux kernel v4.9.60 \cite{mptcplinuximpl}. The code is available at \cite{queueawarecode}.
As shown in Section~\ref{sec:scheduler}, QAware's functioning depends on the current estimate of network interface (NIC) queue occupancy. Conventionally, the NIC queues were either implemented within the hardware itself or as part of the driver; which made NIC queues invisible to the Kernel and its occupancy extremely hard to estimate. However, since Linux Kernel $>$ v3.3.0, several NIC queue management protocols, known as Byte Queue Limits (BQL), have been introduced as part of the Kernel code to resolve starvation and latency at the NIC \cite{ietfbql}. The BQL algorithms push queueing abstractions from hardware drivers to specific data structures which can be accessed from within the Kernel \footnote{Currently, only PCIe-based ethernet drivers support BQL \cite{bqldrivers}. However, a significant effort is being made from the Linux developer community to support broader list of NICs, including wireless NIC's \cite{makewififast}.}.
Our implementation closely follows the Algorithm \ref{alg:queueaware}. We first tap the network device address mapped to MPTCP socket via \texttt{struct dst\_entry} to access \textit{DQL}\footnote{In Linux, BQL is implemented as Dynamic Queue Limit (DQL).} as follows:
\begin{align*}
\texttt{
dql = netdev\_get\_tx\_queue(dst->dev)->dql
}
\end{align*}
We further utilize \textit{DQL} entry to estimate current NIC (\textit{netdevice}) queue occupancy of each MPTCP subflow.
\begin{equation*}
\begin{split}
\texttt{qSize} = \{\texttt{dql->num\_queued -} \\
\hfill \texttt{dql->num\_completed}\}
\end{split}
\end{equation*}
Here, \texttt{num\_queued} and \texttt{num\_completed} refer to the total number of bytes queued in the network device and number of bytes successfully transmitted by the device respectively.
Apart from NIC queue estimates, we utilize the smoothed mean RTT estimates in microseconds via \texttt{srtt\_us} accessible through \texttt{struct tcp\_sock}. We ensure that our implementation is in line with guidelines mentioned in RFC 6182 \cite{mptcprfc}.
\section{Introduction}\label{sec:introduction}
Multipath TCP (MPTCP) is a recently-standardized extension to TCP that allows devices with multiple network interfaces, e.g., smartphones with WiFi and LTE, to seamlessly form multiple parallel connections to exploit the full network capacity. MPTCP offers increased robustness and resilience, as well as seamless handovers and it has been proposed to be also used in datacenters~\cite{datacentermptcp}, opportunistic networks~\cite{opportunisticmptcp}, etc. There is both an open source implementation for Linux~\cite{mptcplinux}, and companies, such as Apple, have incorporated MPTCP into their products and have made the APIs open to application developers~\cite{applemptcp}.
Figure~\ref{fig:mptcp_device_queue} shows the network stack of MPTCP-compliant machine. Applications utilizing MPTCP can send their data over multiple TCP subflows, where each subflow is associated with a unique network interface. TCP packets scheduled over a subflow wait in the device driver queue of the corresponding network interface before they are transmitted by the network interface card (NIC). The choice of network path for sending application data is made by the MPTCP scheduler block and depends on the scheduling policy.
\begin{figure}[!t]
\centering
\includegraphics[width=0.35\textwidth]{Figs/mptcp_device_queue}
\caption{\label{fig:mptcp_device_queue} An illustration of MPTCP-compliant machine and how its subflows interact with their corresponding network interface queues.}
\end{figure}
Scheduling between the multiple connections is an obvious research problem and recently multiple proposals~\cite{blest},~\cite{dems},~\cite{daps},~\cite{ecf} have emerged to improve the default MPTCP scheduler~\cite{mptcpscheduler}. Typically, these schedulers use a transport layer estimate of the end-to-end bandwidth/delay (for example, the smoothed round-trip time) for each TCP subflow as an input to the scheduling policy that decides on how application data must be assigned to the multiple subflows.
In this paper, we propose a novel scheduler for MPTCP, QAware, which departs from the previous proposals in a fundamental way. While we also use the end-to-end delay estimates, like current schedulers, QAware additionally considers the number of packets in the device driver queue of the sender. This modification is motivated by our findings, which we discuss further in Section~\ref{sec:incognizance}. The key motivation, as we will demonstrate, is that as a particular flow is used more, its end-to-end delay increases gradually, making it less attractive to use. However, the traditional, purely end-to-end-based estimation, reacts very slowly to these changes.
Additionally, utilizing queue occupancy information allows QAware to use all available subflows optimally, especially when their properties are highly heterogeneous. Existing proposals like~\cite{blest, ecf, otias}, treat the flows as separate entities and typically do not fully use all the flows.
QAware optimizes transmission over all the flows and gets a significantly higher aggregate throughput, with no loss of performance in any situation.
The contributions of this paper are:
\begin{enumerate}[(a)]
\item We propose QAware, which is a novel cross-layer approach to scheduling packets across all available MPTCP subflows. The design is motivated by our experimental findings that combining local device driver queue occupancy with the traditional end-to-end delay measurements yield far superior performance.
\item We model available MPTCP subflows as multiple parallel service facilities that can service data provided by an application. This enables us to leverage queueing theoretical insights to create a scheduling policy that combines end-to-end delays and device driver queue occupancy.
\item Our simulations and real-world experimentation over a wide range of applications compare QAware with the default MPTCP scheduler~\cite{mptcpscheduler}, ECF~\cite{ecf}, DAPS~\cite{daps}, and BLEST~\cite{blest}.
\end{enumerate}
Rest of the paper is organized as follows. We discuss the relevant background and related works in Section~\ref{sec:relatedwork}. Section~\ref{sec:incognizance} motivates the need for a cross-layer approach to scheduling. In Section~\ref{sec:scheduler}, we describe the scheduling policy used by QAware. Section~\ref{sec:implementation} provides implementation details of QAware in latest MPTCP v0.93. Section~\ref{sec:evaluation} provides an overview of our evaluation methodology. Sections~\ref{sec:simulation} and~\ref{sec:realexperiments} quantify the performance of QAware using extensive simulations and real-world experiments, respectively. We conclude in Section~\ref{sec:conclusion}.
\section{Real-World Setup And Experiments}
\label{sec:realexperiments}
\begin{figure}[!t]
\begin{center}
\includegraphics[width=0.38\textwidth]{Figs/mptcp_testbed}
\caption{Real network experiment setup.}
\label{fig:realsetup}
\end{center}
\end{figure}
We assess QueueAware in a real setting using the Linux kernel implementation~\cite{queueawarecode} summarized in Section~\ref{sec:implementation}. Figure \ref{fig:realsetup} shows the test network topology in the University of Helsinki data center. We set up two similar machines with 16 core AMD Opteron processor, 8 GB DDR2 RAM running Ubuntu 16.04 LTS with latest stable MPTCP implementation (version 0.93, based on Linux kernel v4.9.33 \cite{mptcpv93}) as client and server. Both client and server machines are interconnected via two separate Gigabit Ethernet interfaces. One Ethernet connection is routed through the internal University of Helsinki network and therefore sees background traffic from University staff. It has an end-to-end RTT of $>$1ms. The other connection is over a Top-of-Rack (ToR) switch with RTT $<$1 ms.
We compare QueueAware with the following schedulers: (a) minSRTT, (b) Delay Aware Packet Scheduler (DAPS)~\cite{daps}, (c) Blocking Estimation based Scheduler (BLEST)~\cite{blest}, and (iv) Earliest Completion First (ECF)~\cite{ecf}\footnote{For DAPS and BLEST, we use the implementation at \url{https://bitbucket.org/blest_mptcp/nicta_mptcp}. For ECF, we use the implementation at \url{http://cs.umass.edu/~ylim/mptcp_ecf}} \footnote{It must be noted that DAPS, BLEST, and ECF are implemented on MPTCP v0.89 whereas the default minSRTT and our QueueAware scheduler is based on current MPTCP v0.93}. We examine the schedulers on four different real-network workloads: bulk traffic transfer, video streaming, web-file download and adaptability to heterogeneous paths under realistic network conditions for each experiment. We use the Linux Traffic Control system (tc) in combination with a Hierarchical Token Bucket (HTB) packet scheduler using Statistical Fair Queuing (SFQ) for network shaping. All our results are averaged over 10 runs unless otherwise mentioned. In between runs, we flush out the TCP cache to ensure that each run is independent of the next.
\subsection{Bulk Traffic}
\begin{figure*}[!t]
\hspace*{\fill
\begin{center}
\begin{minipage}[t]{0.243\linewidth}
\vspace{0pt}
\includegraphics[width=0.89\linewidth]{Figs/IFIPGraphs/iPerf-Linux}
\vfill
\caption{Average Throughput Comparison for Bulk Traffic}
\label{fig:iperf_linux}
\end{minipage}
\hfill
\begin{minipage}[t]{0.72\linewidth}
\vspace{0pt}
\begin{subfigure}{0.31\linewidth}
\centering
\includegraphics[height=4cm, width=4.2cm]{Figs/IFIPGraphs/video-2+2}
\caption{\label{fig:video-2+2}\textbf{2Mbps+2Mbps}}
\end{subfigure}
\hfill
\begin{subfigure}{0.31\linewidth}
\centering
\includegraphics[height=4cm, width=4.2cm]{Figs/IFIPGraphs/video-24+16}
\caption{\label{fig:video-24+16}\textbf{2.4Mbps+1.6Mbps}}
\end{subfigure}
\hfill
\begin{subfigure}{0.31\linewidth}
\centering
\includegraphics[height=4cm, width=4.2cm]{Figs/IFIPGraphs/video-2+16}
\caption{\label{fig:video-2+16}\textbf{2Mbps+1.6Mbps}}
\end{subfigure}
\caption{Measured Average Bitrate in Video Streaming for Different Bandwidth Regulations (\textbf{higher is better})}
\label{fig:videoexp}
\end{minipage}
\end{center}
\hspace*{\fill
\vspace*{-2em}
\end{figure*}
In this section, we compare QueueAware's performance with other schedulers for high application transfer rate over somewhat-homogeneous subflows. The purpose of this experiment is to validate QueueAware's performance improvement for heavily network-dependent applications from simulations. Figure \ref{fig:iperf_linux} shows the average throughput of all schedulers over ten runs.
From the results, it can be observed that QueueAware achieves more than 45\% increase in throughput for bulk traffic over homogeneous paths than DAPS, BLEST and ECF schedulers. QueueAware's performance also supersedes that of default minSRTT scheduler by 37\%. Interestingly, the minSRTT scheduler outperforms DAPS, BLEST, and ECF in the experiment. We attribute minSRTT's efficiency to two reasons. Firstly, DAPS, BLEST and ECF schedulers have been designed to improve MPTCP performance for heterogeneous network path delays wherein BLEST and ECF even goes as far as not sending an available packet on slower subflow and wait for the faster path to become available. For homogeneous subflows, the default scheduler places more number of packets on each path as opposed to BLEST and ECF. Secondly, based on latest MPTCP version minSRTT enjoys several code improvements and optimizations in latest MPTCP kernel unlike DAPS, BLEST and ECF.
\subsection{Video Streaming}
\begin{table}[!t]
\centering
\begin{subtable}{\linewidth}
\centering
\begin{tabular}{llll}
\toprule
Bandwidth (Mbps) & 2.4 & 2& 1.6 \\
Delay (ms) & 10 & 20& 30 \\
\bottomrule
\end{tabular}
\caption{Configuration for Video Streaming Experiment}
\label{table:netwvideo}
\end{subtable}
\vspace*{1em}
\begin{subtable}{\linewidth}
\centering
\begin{tabular}{lllll}
\toprule
Delay (ms) & 1 & 40& 80& 160 \\
Bandwidth (Mbps) & 950 & 600& 300& 300 \\
\bottomrule
\end{tabular}
\caption{Configuration for Heterogeneous Paths Experiment}
\label{table:netwhetro}
\end{subtable}
\caption{Bandwidth and RTT regulations for experiment setup. Table \ref{table:netwvideo} presents values after both bandwidth and delay shaping whereas \ref{table:netwhetro} shows bandwidth achieved by delay throttling on a 1Gbps ethernet interface}
\end{table}
In this section, we investigate whether QueueAware improves the performance of video streaming applications when compared to other schedulers. We choose streaming as one of our evaluation scenario as it is dominant Internet use case and is widely adopted by content providers such as Netflix, YouTube etc. \cite{streamingapp}. We set up a DASH server and host \textit{Big Buck Bunny} available from a public dataset on it \cite{dashDataset}. We configure the streaming server to provide five representations of the video from 240p to 1080p (same as most content providers). We re-encode each representation in at least three different bitrates with overall available bit rates from 128Kbps to 3.8Mbps using H.264/MPEG-4 AVC codec. The streaming client employs an Adaptive Bit Rate (ABR) algorithm to download video segments according to available network bandwidth. We throttle our testbed bandwidth to match the bitrates of DASH encodings. Table \ref{table:netwvideo} shows the average delay measured at client-side for each bandwidth configuration. We evaluate scheduler performance when the two subflows are \textit{homogeneous} (2Mbps+2Mbps), \textit{semi-heterogeneous} (2Mbps+1.6Mbps) and \textit{heterogeneous} (2.4Mbps+1.6Mbps). Figure \ref{fig:videoexp} shows the results.
We observe that QueueAware improves the performance of streaming applications in all network conditions and achieves a closer to ideal bitrate. The performance improvement is more significant in homogeneous and semi-heterogeneous scenarios (8\% and 5\% with default and 10\% and 6\% with ECF respectively) as QueueAware utilizes available paths more efficiently than other schedulers. DAPS yields the worst performance out of all schedulers due to its dependence on RTT ratio of two subflows which provides incorrect results for RTT estimations in Linux. BLEST performance degrades when the paths are heterogeneous whereas QueueAware still achieves the best bitrate in this setting.
\subsection{Web File Download}
\begin{figure}[!t]
\begin{center}
\includegraphics[width=0.46\textwidth]{Figs/IFIPGraphs/linux-fileDownload}
\caption{Normalized Download Completion Time for Different File Sizes (\textbf{lower is better})}
\label{fig:linux-file-download}
\end{center}
\end{figure}
We now evaluate QueueAware's performance for simple web downloads using \emph{curl}. We set up an Apache 2.2.22 as HTTP server and host varying file sizes ranging from 128KB to 500MB. In order to eliminate application connection time, we only consider the transport-level time to completion in overall completion time. Figure \ref{fig:linux-file-download} presents the average completion time normalized to the maximum achieved value by a scheduler in each file size experiment.
For small web transfers ($<$1MB) all schedulers perform quite similar to each other (it took 0.002s to download a 128KB file by QueueAware vs. 0.003s by minSRTT). This is because for small data transfers, the bandwidth of the primary subflow is more than capable of single shot transmission and thus MPTCP rarely switches to the secondary subflow. Therefore, until the performance of primary subflow degrades during transfer, the choice of the scheduler does not affect the performance of small file downloads. As BLEST and ECF add additional delays by waiting for faster subflow to become available, the default and DAPS scheduler achieve lower completion time for medium file sizes ($\approx$10/100MB). For large files (500MB), BLEST and ECF utilize faster subflow more efficiently than default and DAPS thus achieving a lower completion time. QueueAware always outperforms other schedulers and realizes up to 30\% decrease for medium file sizes and 10\% for large file downloads.
\subsection{Heterogeneous Path Delays}
\begin{figure}[!t]
\centering
\begin{subfigure}{0.22\textwidth}
\centering
\includegraphics[height=4cm]{Figs/IFIPGraphs/delay-1_2}
\caption{\label{fig:delay-1_2}\textbf{40ms+80ms}}
\end{subfigure}%
\hspace{0.7em}
\begin{subfigure}{0.22\textwidth}
\centering
\includegraphics[height=4cm]{Figs/IFIPGraphs/delay-1_4}
\caption{\label{fig:delay-1_4}\textbf{40ms+160ms}}
\end{subfigure}
\caption{Average Throughput Comparison for Heterogeneous Path Delays}
\label{fig:delay_linux}
\end{figure}
We now examine QueueAware's performance for network scenario when subflows have high delay differences such as offered by cellular and WiFi paths. For this experiment, we shape the path delays to emulate two realistic network settings:
\begin{inparaenum}[i)]
\item when delay on one subflow is twice of the other and,
\item when delay on one subflow is four times of the other subflow.
\end{inparaenum}
The per-path bandwidth for different delay throttling is presented in Table \ref{table:netwhetro}.
Figure \ref{fig:delay_linux} shows the average throughput achieved by all schedulers. For path delay ratio of 2, QueueAware yields an average throughput of 310Mbps which is an improvement of $\approx10\%$ over default scheduler and DAPS and $\approx5\%$ over ECF and BLEST. For network delay ratio of 4, all schedulers perform quite similar to each other as all try to fully utilize the lower delay path to avoid \textit{Head-of-Line blocking} at receiver. In this case, QueueAware still manages to achieve an improvement of $\approx7\%$ over default scheduler and DAPS, and $\approx4\%$ over BLEST and ECF.
\section{Real-World Setup And Experiments}
\label{sec:realexperiments}
\begin{figure}[!t]
\begin{center}
\includegraphics[width=0.38\textwidth]{Figs/mptcp_testbed}
\caption{Real network testbed in university datacenter.}
\label{fig:realsetup}
\end{center}
\end{figure}
We next examine QAware's performance in real network environments. Figure \ref{fig:realsetup} shows our test network topology in University of Helsinki data center. We assign two similar machines with 16 core AMD Opteron processor, 8 GB DDR2 RAM running Ubuntu 16.04 LTS with latest stable MPTCP implementation (version 0.93, based on Linux kernel v4.9.60 \cite{mptcplinuximpl}) as client and server. The implementation uses default congestion control algorithm (coupled OLIA). Both machines are interconnected via two separate Gigabit Ethernet interfaces. One Ethernet connection is routed through internal University of Helsinki network and therefore encounters background traffic from University staff. It has an end-to-end RTT of $>$1ms. The other connection is over Top-of-Rack (ToR) switch with RTT $<$1ms.
We compare QAware with the following schedulers:
\begin{inparaenum}[i)]
\item minSRTT,
\item Delay Aware Packet Scheduler (DAPS)~\cite{daps}
\item Blocking Estimation based Scheduler (BLEST)~\cite{blest}, and
\item Earliest Completion First (ECF)~\cite{ecf}\footnote{For DAPS and BLEST, we use the implementation at \url{https://bitbucket.org/blest_mptcp/nicta_mptcp}. For ECF, we use the implementation at \url{http://cs.umass.edu/~ylim/mptcp_ecf}} \footnote{DAPS, BLEST, and ECF are implemented on MPTCP v0.89 whereas the default minSRTT and QAware are based on MPTCP v0.93. We could not implement QAware on MPTCP v0.89 as it is based on Linux v3.18 which does not support BQL. Please see \cite{mptcplinuximpl} for exact changes between the two versions.}.
\end{inparaenum}
We first compare scheduler performance for application generating bulk traffic. This workload provides a qualitative validation of the results we obtained in Section~\ref{sec:simulation}. We further present scheduler performance for DASH video streaming and web file downloads. We used the Linux Traffic Control system (\emph{tc}) in combination with a Hierarchical Token Bucket (HTB) packet scheduler using Statistical Fair Queuing (SFQ) for network shaping. In between runs, we flushed out the TCP cache to ensure that each run is independent of the next. All our results are averaged over ten runs.
\subsection{Bulk Traffic}
In this section, we compare QAware's performance with other schedulers for high application transfer rate over both subflows. We performed experiments with different settings of delays along the two paths. The setting includes
\begin{inparaenum}[i)]
\item default path delays ($< 1$ms and $> 1$ms),
\item delay shaping to introduce $40$ms of delay along one path and $80$ms along the other, and
\item $40$ms along one path and $160$ms along the other.
\end{inparaenum}
Path bandwidths corresponding to the different delays are stated in Table~\ref{table:netwhetro}.
Figure~\ref{fig:iperf_linux} compares average throughput obtained by different schedulers for default path delays. QAware achieves more than 45\% increase in throughput compared to DAPS, BLEST and ECF. QAware also provides an improvement of $37\%$ over the default minSRTT scheduler. Interestingly, the minSRTT scheduler outperforms DAPS, BLEST, and ECF in the experiment. We attribute minSRTT's efficiency to two reasons. Firstly, DAPS, BLEST and ECF schedulers have been designed to improve MPTCP performance for heterogeneous delays along available network paths. In fact, BLEST and ECF even go as far as not sending an available packet on a slower subflow and wait for the faster subflow to become available. When subflows witness similar delays (as in the current case), the default scheduler places more packets on each path as opposed to DAPS, BLEST, and ECF. Secondly, based on latest MPTCP kernel, minSRTT enjoys several code improvements and optimizations
\begin{figure}[!t]
\centering
\begin{subfigure}{0.165\textwidth}
\centering
\includegraphics[width=\textwidth]{Figs/IFIPGraphs/iPerf-Linux1}
\caption{\label{fig:iperf_linux}Default}
\end{subfigure}%
\hfill
\begin{subfigure}{0.16\textwidth}
\centering
\includegraphics[width=\textwidth]{Figs/IFIPGraphs/delay-1_2}
\caption{\label{fig:delay-1_2}40+80ms}
\end{subfigure}%
\hfill
\begin{subfigure}{0.16\textwidth}
\centering
\includegraphics[width=\textwidth]{Figs/IFIPGraphs/delay-1_4}
\caption{\label{fig:delay-1_4}40+160ms}
\end{subfigure}
\caption{Bulk Traffic throughputs for different access path delays.}
\label{fig:delay_linux}
\end{figure}
For when the path delays are $40$ and $80$ms, QAware yields an average throughput of 310 Mbps which is an improvement of about $10\%$ over the default scheduler and DAPS and $5\%$ over ECF and BLEST (shown in Figure \ref{fig:delay-1_2}). As presented in Figure \ref{fig:delay-1_4}, all schedulers perform quite similar to each other as all try to fully utilize the lower delay subflow when path delays are $40$ and $160$ms. In this case, QAware still manages to achieve an improvement of about $7\%$ over the default scheduler and DAPS, and about $4\%$ over BLEST and ECF.
\subsection{Video Streaming}
\begin{figure}[!t]
\centering
\begin{subfigure}{0.158\textwidth}
\centering
\includegraphics[width=\textwidth]{Figs/IFIPGraphs/video-2+2}
\caption{\label{fig:video-2+2}2+2 Mbps}
\end{subfigure}
\hfill
\begin{subfigure}{0.158\textwidth}
\centering
\includegraphics[width=\textwidth]{Figs/IFIPGraphs/video-2+16}
\caption{\label{fig:video-2+16}2+1.6 Mbps}
\end{subfigure}
\hfill
\begin{subfigure}{0.158\textwidth}
\centering
\includegraphics[width=\textwidth]{Figs/IFIPGraphs/video-24+16}
\caption{\label{fig:video-24+16}2.4 +1.6 Mbps}
\end{subfigure}
\caption{Average bitrate in video streaming for different path bandwidths.
\label{fig:videoexp}
\end{figure}
\begin{table}[!ht]
\centering
\begin{subtable}{\linewidth}
\centering
\begin{tabular}{lllll}
\toprule
Delay (ms) & 1 & 40& 80& 160 \\
Bandwidth (Mbps) & 950 & 600& 300& 200 \\
\bottomrule
\end{tabular}
\caption{Configurations for Bulk Traffic Experiments}
\label{table:netwhetro}
\end{subtable}
\vspace*{1em}
\begin{subtable}{\linewidth}
\centering
\begin{tabular}{llll}
\toprule
Bandwidth (Mbps) & 2.4 & 2& 1.6 \\
Delay (ms) & 10 & 20& 30 \\
\bottomrule
\end{tabular}
\caption{Configurations for Video Streaming Experiments}
\label{table:netwvideo}
\end{subtable}
\caption{\ref{table:netwhetro} shows bandwidth achieved by delay throttling on a 1Gbps Ethernet interface whereas \ref{table:netwvideo} presents values after both bandwidth and delay shaping}
\end{table
Streaming is a dominant Internet use case and is widely adopted by content providers such as Netflix and YouTube~\cite{streamingapp}. We set up a DASH server and host \textit{Big Buck Bunny}, available from a public dataset, on it \cite{dashDataset}. We configured the streaming server to provide five representations of the video from 240p to 1080p (same as most content providers). We re-encoded each representation in at least three different bitrates with overall available bit rates from 128Kbps to 3.8Mbps using H.264/MPEG-4 AVC codec. The streaming client employs an Adaptive Bit Rate (ABR) algorithm to download video segments according to the available network bandwidth. We throttled our testbed bandwidth to match the bitrates of DASH encodings. Table \ref{table:netwvideo} shows the average delay measured at client-side for each bandwidth configuration. We evaluate and compare QAware's performance with other schedulers for when the two subflows
\begin{inparaenum}[i)]
\item have bandwidths of 2 Mbps,
\item have bandwidths of 2 Mbps and 1.6 Mbps, and
\item have bandwidths of 2.4 Mbps and 1.6 Mbps.
\end{inparaenum}
From Figure \ref{fig:videoexp}, we observe that QAware improves the performance of streaming applications in all network conditions. The performance improvement is more significant in scenarios where the path bandwidths are similar (8\% and 5\% with respect to default and 10\% and 6\% with respect to ECF, in Figures~\ref{fig:video-2+2} and \ref{fig:video-2+16} respectively) as QAware utilizes available paths more efficiently than other schedulers. DAPS consistently gives the worst performance out of all schedulers due to its strong dependence on RTT ratio of two subflows.
\subsection{Web File Download}
\begin{figure}[!b]
\begin{center}
\includegraphics[width=0.42\textwidth]{Figs/IFIPGraphs/linux-fileDownload}
\caption{Normalized download completion time for different file sizes ({smaller is better}).}
\label{fig:linux-file-download}
\end{center}
\end{figure}
We now evaluate QAware's performance for simple web downloads using \emph{curl}. We set up an HTTP server using Apache 2.2.22 and hosted varying file sizes of range 128KB to 500MB. We eliminate application connection time by only considering the transport-level time in overall download completion time observed at the client. Figure \ref{fig:linux-file-download} presents the average completion time normalized to the maximum achieved value by scheduler for a given file size.
For small web transfers ($<$1MB) all schedulers perform quite similar to each other (it took 0.002s to download a 128 KB file by QAware vs. 0.003s by minSRTT). This is because for small data transfers, the bandwidth of the primary subflow is more than capable of single shot transmission and thus MPTCP rarely switches to the secondary subflow. Therefore, until the performance of primary subflow degrades during transfer, the choice of the scheduler does not affect the performance for small files. The default and DAPS scheduler achieve lower completion time for medium file sizes ($\approx$10/100 MB) in comparison to BLEST and ECF. This is likely because BLEST and ECF add additional delays by waiting for the faster subflow to become available. For large files (500 MB), BLEST and ECF utilize faster subflow more efficiently than default and DAPS, thus achieving a lower completion time. QAware always outperforms other schedulers and realizes up to 20\% decrease in completion time for medium file sizes (0.709s by QAware vs. 0.895s by ECF for 100 MB file) and 30\% for large file downloads (3.46s by QAware vs. 4.93s by minSRTT for 500 MB).
\section{Background and Related Work}\label{sec:relatedwork}
\begin{figure}[!t]
\centering
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[width=\linewidth]{Figs/real_throughput}
\caption{\label{fig:emc_throughput}}
\end{subfigure}%
\hspace{0.005\textwidth}
\begin{subfigure}{.22\textwidth}
\centering
\includegraphics[width=\linewidth]{Figs/real_rtt}
\caption{\label{fig:emc_rtt}}
\end{subfigure}
\caption{(a) Loading (Mbps) at the subflows and (b) their RTT(s). The paths taken by the subflows and the network are shown in Figure~\ref{fig:netw_topology}.}
\label{fig:emc_results}
\end{figure}
The default MPTCP scheduler (minSRTT) allocates traffic on the fastest subflow (one that has the smallest smoothed RTT) with available congestion window at each packet arrival. Several researchers have proposed improvements to the default minSRTT scheduler. Most approaches leverage the difference in RTT of the subflows \cite{slow-path-adaptation, packet-scheduling}. Others have also considered additional TCP-layer parameters such as SSThresh, congestion window, selective ACK and receiver buffer size along with RTT \cite{receiver-scheduler, ocps, f2dpds}.
In \cite{otias}, the authors introduce an additional sender queue to schedule packets on a subflow even when it is unavailable. Delay Aware Packet Scheduler (DAPS) \cite{daps} generates a schedule for sending future segments over subflows based on their RTT ratios. However, this makes DAPS unable to react promptly to network changes due to pre-computed long schedules.
Blocking-Estimation-based MPTCP Scheduler (BLEST) \cite{blest} aims to reduce head-of-line blocking by waiting for the faster subflow despite the availability of space in congestion window of the slower subflow. ECF \cite{ecf} follows a similar principle as that of BLEST, but while BLEST aims to reduce out-of-order delivery assuming that the send buffer is a bottleneck, ECF aims to minimize completion time.
Researchers have also proposed schedulers that improve MPTCP performance for specific application use-cases. Decoupled Multipath Scheduler (DEMS) \cite{dems} aims to reduce fixed-size file's delivery time over MPTCP by estimating available bandwidth on subflows. However, the authors rely on exact knowledge of data chunk boundary for efficient scheduling.
In \cite{crosslayer_video} authors leverage application layer information for flow scheduling decisions to provide delay-resilient video streaming in MPTCP. MP-DASH \cite{mp-dash} exploits path information from streaming client to improve DASH video delivery.
\cite{crosslayer_infocomm} labels WiFi subflow as active/inactive for data transmission based on a minimum desired signal strength. However, unlike other cross-layer approaches which optimize specific application performance over MPTCP, QAware taps into lower layer information to improve performance for \emph{all} MPTCP traffic. Furthermore, as shown later in the paper, QAware's unique design of leveraging hardware queue occupancy enables it to swiftly adapt to varying network conditions and co-existent network applications sharing bottleneck paths.
\section{Simulation Setup and Results}
\label{sec:simulation}
\begin{figure}[!t]
\centering
\begin{subfigure}{0.47\textwidth}
\centering
\includegraphics[width=\textwidth]{Figs/IFIPGraphs/experiment1a_homo_pattern_new}
\caption{\label{fig:experiment1a_homo} Subflows F1 and F2 use links with PHY rates of $6$ Mbps.}
\end{subfigure}%
\vspace{0.005\textwidth}
\begin{subfigure}{0.47\textwidth}
\centering
\includegraphics[width=\textwidth]{Figs/IFIPGraphs/experiment1a_hetro_pattern_new}
\caption{\label{fig:experiment1a_hetro} Subflow F1 and F2 use links with PHY rate of $12$ Mbps and $6$ Mbps respectively.}
\end{subfigure}
\caption{Throughput achieved by minSRTT, ECF and QAware schedulers for different CBR rates.}
\label{fig:throughput_no_error}
\end{figure}
We simulated network topologies of the kind shown in Figure~\ref{fig:netw_topology}. For all simulations, the links between the access points and the backbone switch and between the backbone switch and the server were modeled as wired links with rate $30$ Mbps and $50$ Mbps respectively. The client is connected to the two access points over wireless links with physical layer (PHY) rates in the range $6-12$ Mbps. These two wireless links provided the two network paths over which application data was sent. Both subflows use independent congestion control.
We simulated the following \emph{applications}:
\begin{inparaenum}[i)]
\item constant bit rate (CBR) data from low to high rates,
\item file transfer for sizes of $10-30$ MB,
\item web browsing of top 10 out of the US Alexa-100 websites, and
\item CBR with one of the paths being shared by UDP traffic.
\end{inparaenum}
For the applications, we simulated the following \emph{network configurations}:
\begin{inparaenum}[i)]
\item both wireless links have the same rate,
\item one link is much faster than the other, and
\item one link drops TCP packets.
\end{inparaenum}
Comparisons of QAware with ECF and minSRTT\footnote{In simulation, the scheduler assigns packets over independent TCP streams. We do not incorporate other MPTCP functionality such as re-transmission handler and path manager.}
demonstrate the benefits that are accrued by QAware because it optimally utilizes both network paths.
\begin{figure}[!t]
\centering
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{Figs/IFIPGraphs/behavior_srtt}
\caption{\label{fig:behavior_srtt}minSRTT Scheduler}
\end{subfigure}%
\vspace{0.005\textwidth}
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{Figs/IFIPGraphs/behavior_ecf}
\caption{\label{fig:behavior_ecf}ECF Scheduler}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{Figs/IFIPGraphs/behavior_queueaware}
\caption{\label{fig:behavior_qa}QAware Scheduler}
\end{subfigure}%
\caption{Per-flow throughput, device driver queue occupancy, and SRTT behavior as a function of time. These correspond to the throughputs in Figure~\ref{fig:experiment1a_homo} and a CBR rate of $12$ Mbps.}
\label{fig:explanation}
\end{figure}
\subsection{Constant Bit Rate Traffic} \label{sec:cbrsimulation}
\subsubsection*{Access paths with no packet errors}
Figure~\ref{fig:experiment1a_homo} shows the TCP throughputs obtained by the schedulers for increasing CBR rates. Each wireless link was configured with a PHY rate of $6$ Mbps. This results in homogeneous network paths. On average, QAware achieves percentage throughput gains of about $40\%$ over the rest. Further, note that all schedulers use both subflows. However, unlike the others, QAware utilizes both the subflows almost equally for the entire simulation time for all the CBR loads. To better understand their behaviors, consider Figure~\ref{fig:explanation}, which shows for each scheduler and subflow, the variation of throughput, device driver queue occupancy, and smoothed RTT, as a function of time, for a $2$ second interval. The CBR rate was set at $12$ Mbps. From the subflow throughputs and queue occupancy, it is clear that QAware uses both subflows almost simultaneously. ECF uses just one subflow for most of the interval, and while minSRTT uses both flows during the interval, it switches between them very infrequently. Both minSRTT and ECF rely on the delayed feedback provided by SRTT and so end up scheduling packets to one subflow for longer intervals than QAware. Essentially, they switch flows when SRTT of the subflow in use exceeds that of the other subflow. In addition, ECF, by design, declines scheduling opportunities to a subflow with a larger RTT and prefers to wait for faster subflows. This explains the reason for using one flow for a longer duration than minSRTT scheduler. In minSRTT and ECF, subflows experience swings in SRTT. The SRTT increases linearly while it is the subflow of choice. This increase eventually makes the subflow less desirable than the other and the scheduler switches to the other flow, which, due to the current low occupancy in the corresponding device queue, experiences low SRTT.\footnote{Our observations with respect to QAware and minSRTT for three homogeneous paths are similar. We skip them due to lack of space.}
\begin{figure}[!t]
\centering
\includegraphics[width=0.47\textwidth]{Figs/IFIPGraphs/experiment1b_homo_pattern_new}
\caption{Per-flow throughput comparison for different CBR rates where subflow F1 experiences a packet drop rate of $10^{-2}$.}
\label{fig:experiment1b_homo}
\end{figure}
Figure~\ref{fig:experiment1a_hetro} shows throughputs obtained by the CBR application when the PHY rate of one of the wireless links is $6$ Mbps and the other is $12$ Mbps. While all schedulers utilize the subflow using the $12$ Mbps link equally, QAware also utilizes the subflow mapped on the $6$ Mbps link. On average, QAware achieves throughput gains of about $50\%$ over the rest.
\subsubsection*{Access paths with packet errors}
Figure~\ref{fig:experiment1b_homo} shows the throughput obtained when one subflow suffers a packet loss rate of about $10^{-2}$. Both wireless links have PHY rates of $6$ Mbps. Upon detecting packet loss, the congestion window of the subflow decreases based on \textit{TCP congestion avoidance algorithm}, which limits the number of packets that can be sent on that subflow. Even in this situation, QAware is able to exploit both subflows better and achieves about $32\%$ and $15\%$ improvement over minSRTT and ECF respectively. For the case when the wireless links are $12$ Mbps and $6$ Mbps with an error on the slower link, the corresponding gains are $53\%$ and $6\%$ (figure not shown due to space limitations). Note that since ECF is biased toward using the faster path, it performs almost as well as QAware when the error-free path has a faster wireless link. On the other hand, while minSRTT uses the error-prone path better than ECF, it is unable to make good use of the error-free path as the other two schedulers.
\subsection{Fixed Size File Transfer}
\begin{figure}[!t]
\centering
\includegraphics[width=0.4\textwidth]{Figs/IFIPGraphs/large-file-simulation}
\caption{File download completion times when both subflows use wireless link with PHY rate of $6$ Mbps.}
\label{fig:download_simulator}
\end{figure}
Figure~\ref{fig:download_simulator} shows the download completion time achieved by the three schedulers for five different file sizes ranging from 10MB to 30MB. Both wireless links were set to a PHY rate of $6$ Mbps. Observe that QAware obtains the least download time for all the file sizes. This is explained by its ability to effectively utilize both the subflows for data transfer. The performance gap increases proportionally with file size. Overall, QAware achieves $35\%$ and $30\%$ reduction in average download time over minSRTT and ECF respectively.
\subsection{Web-browsing}
\begin{table}[!b]
\centering
\begin{tabular}{c|ccccc}
\toprule
\textbf{Website} & News & Tech & Radio & Shopping & Finance \\
\midrule
\textbf{\#Objects} & 202 & 67 & 66.2 & 52.2 & 39.7 \\
\textbf{Size (KB)} & 3821.2 & 2152.2 & 2453 & 1000.7 & 1988.1 \\
\hline\midrule
\textbf{Website} & Wiki & Market & Social & Movie & Travel \\
\midrule
\textbf{\#Object} & 28 & 49 & 69 & 39 & 21 \\
\textbf{Size (KB)} & 601.2 & 2032.8 & 1700.2 & 845.7 & 2000.4 \\
\bottomrule
\end{tabular}
\caption{Web objects for traffic generation}
\label{table:webbrowsing}
\end{table}
\begin{figure}[!t]
\centering
\includegraphics[width=0.48\textwidth]{Figs/IFIPGraphs/web_homo}
\caption{Download completion time for 10 websites from top U.S. Alexa-100 websites.}
\label{fig:web_homo}
\end{figure}
To simulate web browsing, we deployed objects of 10 out of top U.S. Alexa-100 websites, which are summarized in Table \ref{table:webbrowsing}, in our simulated server. The client consecutively downloaded relevant objects of each website from the server at a variable rate between 10Mbps to 30Mbps chosen in a probabilistic manner. We compared scheduler performance for when both wireless links are $6$ Mbps and when one of the links is $12$ Mbps. QAware achieves a significant reduction in download completion time for both configurations, specifically up to $35\%$ for the former (see Figure~\ref{fig:web_homo}) and up to $28\%$ for the latter (figure not shown due to space limitations). On the other hand, ECF and minSRTT perform similarly.
\subsection{Multiple Applications}
In current computing environments, end hosts typically run multiple applications which must share the interfaces available at the host for network transfers. An ideal MPTCP scheduler must be able to efficiently adapt to bandwidth competition on bottleneck links in such coexisting environment. To evaluate the impact of such sharing on the schedulers, we used the following setup. The PHY rates of the wireless links were set to $9$ and $6$ Mbps. A CBR application generated data for a 10 second interval and used both the MPTCP subflows. The results are shown in Figure \ref{fig:mixed_traffic}.
Starting at 4 seconds, we introduced traffic from a UDP application that used the network path with the $9$ Mbps wireless link. The greyed area in the figure denotes the time duration when both MPTCP and UDP applications were active at the client. The UDP traffic lasted for 4 seconds. Before the start of the UDP traffic, only QAware scheduler was utilizing both available subflows. Once the UDP application starts, the device queue of the $9$ Mbps wireless link saturates. QAware, however, quickly adapts to it and reduces the traffic being sent on the corresponding subflow. All the while, it keeps utilizing the subflow over the slower wireless link. On the other hand, both minSRTT and ECF need to wait for several RTT updates for the impact of UDP traffic on queue wait times to get reflected in the SRTT of the subflow. Lastly, unlike the other schedulers, QAware is also quick to detect the availability of the subflow after the 8 second mark, which is when the UDP application stops its transfer. Overall, QAware leads to gains of about $40\%$ over minSRTT and about $50\%$ over ECF.
\begin{figure}[!t]
\centering
\includegraphics[width=0.42\textwidth]{Figs/IFIPGraphs/multipleApps}
\caption{Per-flow throughputs when the interface used by subflow F1 sees UDP traffic for $4$ seconds (greyed).}
\label{fig:mixed_traffic}
\end{figure}
\section{QAware Scheduler} \label{sec:scheduler}
\begin{figure}[!t]
\begin{center}
\includegraphics[width=0.4\textwidth]{Figs/queue_decision}
\caption{Queueing abstraction of an \emph{end-to-end} MPTCP connection with two subflows.}
\label{fig:networkqueue}
\end{center}
\end{figure}
We consider a simplified queue-theoretic abstraction to capture the essentials of the scheduling problem, with the goal of maximizing \emph{end-to-end} throughput. Specifically, we model each subflow by a service facility. Figure~\ref{fig:networkqueue} illustrates the abstraction for an MPTCP \emph{end-to-end} connection that uses two TCP flows. The abstraction allows us to apply results from analysis of multi-queue systems\cite{rosberg}.
In our queueing abstraction, packets generated by an application arrive into a queue that models the TCP send buffer (Figure~\ref{fig:mptcp_device_queue}). Packets in this queue are assigned to one of the available service facilities in a first-come-first-serve (FCFS) manner. Each facility consists of a finite queue and a server. Packets inside a facility are serviced in an FCFS manner.
The queue in a service facility is the device driver queue (Figure~\ref{fig:mptcp_device_queue}) that is used by the TCP subflow corresponding to the facility. The server includes the source host NIC, access network used by the subflow, intermittent nodes in the core and the destination host (all layers of the TCP/IP stack).
When a packet is assigned to a service facility, it may find other packets waiting for service in the facility's queue. This packet must wait for all the other waiting packets to finish service before it enters the server of the facility. The total time a packet spends in a facility, often referred to as its \emph{system time}, includes the time it waits in the facility's queue and the time it spends getting service.
\emph{Origins of the QAware scheduler:} Many analytical works on queueing systems have looked at scheduling customer/packet arrivals to parallel service facilities\cite{rosberg,weber,whitt,winston}. For many general arrival processes and service time distributions, when all servers are stochastically identical, the optimal policy is to choose a service facility with a minimum number of packets in its queue~\cite{rosberg,weber,winston}, that is it minimizes the average packet system time.
For the case of non-identical servers, a scheduling policy that assigns a packet to a service facility that minimizes the conditional expected system time of the packet, conditioned on the knowledge of the number of packets waiting for service in the facility, shows good performance~\cite{rosberg}. Our QAware scheduler uses the policy in an MPTCP setting.
Consider $K$ service facilities indexed $1,\ldots,K$. Let facility $k$ have a service rate of $\mu_k$. The two facilities in Figure~\ref{fig:networkqueue} have service rates of $\mu_1$ and $\mu_2$. Let $n_k(t)$ be the number of packets waiting for service in facility $k$ at time $t$. The policy assigns a packet to a service facility $k^*$ given by
\begin{align}
k^* = \arg \min_k \frac{n_k(t)+1}{\mu_k}.
\label{eqn:policy1}
\end{align}
Note that $1/\mu_k$ is the expected service time of a packet in facility $k$. Thus, the conditional waiting time of a packet that enters such a facility is ${n_k(t)}/{\mu_k}$, which is the sum of the expected service times of the $n_k(t)$ packets currently waiting for service in the facility. In addition, we add the term $1/\mu_k$ to ${n_k(t)}/{\mu_k}$, to include the expected service time of the packet to be scheduled. Thus, the expression being minimized in~(\ref{eqn:policy1}) is the conditional expected \emph{system time} of a packet if it were to be assigned to facility $k$.
\emph{Adapting scheduling policy~(\ref{eqn:policy1}) to multiple end-to-end TCP subflows:} The number $n_k(t)$ of packets in the queue of service facility $k$ is the number of packets waiting in the device driver queue of the corresponding subflow $k$ and can be obtained. However, we must estimate the average service time $1/\mu_k$ of subflow $k$.
Consider the $i$\textsuperscript{th} packet arrival. Let $t_i^s$ be the time the packet is assigned to a subflow. Let $t_i^a$ be the time that a TCP ACK acknowledges receipt of the packet. The round-trip time of the packet is $\text{RTT}_i = t_i^a - t_i^s$. Note that this includes the time packet waits in the device driver queue of its assigned subflow before it starts service and the time it spends in service. This is the \emph{system time} of the packet. Let $W_i$\footnote{For simplicity of exposition we ignore the time a TCP ACK may have to wait in a queue before being sent to the TCP layer.} be the time the packet $i$ waits in the queue. This time can be calculated locally at the MPTCP sender. The time $X_i$ that the packet spends in service begins when the packet enters the NIC for transmission and ends when a TCP ACK for the packet is received. Given $W_i$ and $\text{RTT}_i$, we have $X_i = \text{RTT}_i - W_i$. The estimate of the service time is updated on receipt of a TCP ACK. Let $\hat{S}_k$ be the current estimate of the average service time of facility $k$. On receipt of a TCP ACK for packet $i$, we update
\begin{align}
\hat{S}_k = \alpha \hat{S}_k + (1-\alpha) X_i,
\label{eqn:hatSk}
\end{align}
where $0 < \alpha < 1$ applies appropriate weights to the last estimate of the average and the current service time. We use $\alpha=0.8$ in this work which is also the smoothing factor for TCP congestion control \footnote{We examined for other values of $\alpha$ which did not impact the overall performance of QAware.} .
The corresponding estimate of the service rate is $1/\hat{S}_k$. At time $t$, QAware schedules to the TCP subflow $k^*$ that satisfies
\begin{align}
k^* = \arg \min_k {(n_k(t) + 1)} \hat{S}_k.
\label{eqn:policy3}
\end{align}
Finally, note that since $X_i = \text{RTT}_i - W_i$, we have $\hat{S}_k = \text{RTT} - \widehat{W}$, where RTT and $\widehat{W}$ are the exponentially weighted moving averages, with coefficient $\alpha$, of packet round-trip times and device driver queue waiting times, respectively, for the subflow $k$. In our real implementation, summarized in (Algorithm~\ref{alg:queueaware}), we use RTT estimates that are readily available for each subflow and we calculate an approximation of $\widehat{W}$ based on information available from device driver queues.
\begin{algorithm}[!t]\small
\caption{QAware Algorithm}\label{alg:queueaware}
\begin{algorithmic}[1]
\Inputs{Available Subflows \texttt{SF$\in \{1,\ldots,n\}$}}
\Initialize{\strut\texttt{minService} $\gets$ \texttt{0xFFFFFFFF} \\
\texttt{selectedSubflow} $\gets$ NONE}
\item[]
\LState //The function below will return best subflow for packet $P_k$
\For{each \texttt{$\text{subflow} \in$SF}}
\LState $n_k \gets$ \texttt{queueSize(subflow)}
\If{\texttt{$n_k \not= 0$}}
\LState $\Delta t$ $\gets$ sampling time
\LState $\Delta $packets $\gets$ packets dequeued in $\Delta t$
\LState $W_k \gets [1/(\frac{\text{$\Delta $packets}}{\text{$\Delta t$}})] n_k$
\Else
\LState $W_k \gets 0$
\EndIf
\LState $\widehat{W} \gets \alpha \widehat{W} + (1-\alpha) W_k$
\LState $\hat{S}_k = [\text{RTT} - \widehat{W}]$
\LState $TS_k = (n_k + 1) \hat{S}_k$
\If {\texttt{$ TS_k < \text{minService}$}}
\LState $\texttt{minService} \gets TS_k$
\LState \texttt{selectedSubflow $\gets$ subflow}
\EndIf
\EndFor
\end{algorithmic}
\end{algorithm}
|
\section{Introduction}
\label{js:sec:introduction}
We envision future mixed traffic scenarios~\cite{mbBRZ+17}, involving automated cars, trucks, sensor-equipped infrastructure, and other road users equipped with smart devices and other wearables which are interconnected by means of an ad hoc network and cooperate on different levels, e.g., on the level detected intentions. Vulnerable road users (VRUs) will continue to take part in future urban traffic, recognizing them and detecting their intentions poses an important goal towards a safer traffic environment. Many vehicles already incorporate a range of sensors, but they still rely on an unhindered line of sight. The adoption of smart devices implements an alternative considering almost everyone has one with them while taking part/participating in traffic. It comes pre-equipped with the necessary hardware like accelerometer, gyroscope and magnetic field sensors and in near future will be capable to connect with its surrounding via ad hoc networks. GPS sensors enable only a rough localization and do not offer the sampling rates
to detect fast movement changes. System based on inertial data allow for a fast detection of changes in motion. In this article we focus on a reliable detection of starting movement transitions, i.e., initial movement detection.
On the one hand side, the detector has to be robust, i.e., avoid false positive detections, and on the other hand side it has to be fast.
As we have shown in~\cite{mbBZD+17}, the initial movement detection can greatly support the VRU's trajectory forecast.
In terms of organic computing, the system requires self-reflection capabilities in order to adapt subsequent algorithms, e.g.,
for trajectory forecasting.
In this article an approach to smart devices initial movement
detection is presented, aiming to bring self-reflection capabilities and context awareness to smart devices in domain of VRU intention detection.
\subsection{Main Contributions and Outline}
\label{js:sec:outline}
The approach presented in this article aims to detect starting movements of cyclists fast and yet robust, i.e., with few false positive detections. The approach is based on an convolutional neural networks trained on the inertial data (accelerometer and gyroscope) obtained from the smart device worn by the cyclist.
\\
In Section~\ref{js:sec:related_work} the related work is reviewed. Section~\ref{js:sec:methodology} presents our approach including the preprocessing pipeline and the applied deep learning methods. Following a description of the data acquisition and evaluation methodology in Section~\ref{js:sec:data}, Section~\ref{js:sec:experimental_results} demonstrates the experimental results. Section \ref{js:sec:Conclusion} concludes the results and provides further insight into future work.
\section{Related Work}
\label{js:sec:related_work}
Previous work focused on the detection of human activities for various applications ranging from elder care, motion types like standing, walking up or down stairs etc., or even footstep detection \cite{singh2017}. Employing Deep Convolutional Neural Networks (CNNs), rather than handcraft features, appears to pose a viable option based on the success of this emerging technology. Applying machine learning to smart phone data was used for example by Jiang and Yin \cite{jiang2015} in order to perform human activity recognition. Their data is given by accelerometer and gyroscope of wearable devices. Deep Convolutional Neural Networks enable them to automatically learn the optimal features from activity images for the activity recognition task, providing state of the art results on data from wearable sensors. The key feature of their success relies on activity images created out of the Fourier Transform and permutation to enable the model to learn features connecting information contained in multiple tracks of data. The learned filters in a convolutional network thereby substitute handcrafted features which were in the past usually extracted independently from multiple time series sensor signals. This indicates the possibilities to detect further motion primitives like starting and stopping motions in a traffic environment.
Bieshaar et al.~\cite{mbBZH+18} envision a cooperative traffic environment where each individual is connected to the surrounding traffic participants and infrastructure, while maintaining a model of the present situation. They propose a system involving video data from cameras at intersections and smartphone data via a boosted stacking ensemble. This combination leads to a robust classification, although being a complex procedure. Their approach was able to detect 99\% of the starting movements within 0.24 ms after the first movement of the bicycle's back wheel. The system is comprised of a 3D-CNN for the image processing and a smart device-based starting movement detector consisting of an extreme gradient boosting classifier (XGBoost).
Working on a similar topic, Zernetsch et al. \cite{zernetsch2018} also operate on camera data but this time evaluate based on Motion History Images input into a Residual Network (ResNet) and a Support Vector Machine (SVM). Both are capable of a correct classification, whereby the ResNet approach is more robust and accurate than the SVM. In another article \cite{zernetsch2018_2}, they built a forecasting model predicting the location of vulnerable road users for the next 2.5\,s. The experiments include two methods: a physical model of a cyclist and a polynomial leas-squares approximation in combination with a multilayer perceptron. Both were compared to a Kahlman Filter approach and improved on its results. The physical model achieved a 27\% more accurate positional prediction, while the least-squares approximation attained a 34\% increase in accuracy for the starting and stopping motion of a cyclist with otherwise similar outcome as the first approach.
\section{Methodology}
\label{js:sec:methodology}
Our approach aims to detect the movement transition between waiting and moving (i.e., starting) of cyclists as early as possible using a smart device. The starting movement detector is realized by means of a HAR pipeline~\cite{Bulling2014THA}. A schematic of the starting movement detector consisting of three stages is depicted in Fig. \ref{fig:pipeline}. The cyclists starting movement is modeled as a classification problem of the classes \textit{waiting}, \textit{starting} and \textit{moving}. Inertial data, i.e., accelerometer and gyroscope, is used as input for the starting movement detector. The data is preprocessed (i.e., transformed into a device attitude invariant representation) followed by a features extraction. Finally, the detection is realized by means of a deep learning based classifier taking the form of a residual neural network.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{pipeline.png}
\caption[Processing pipeline consisting of preprocessing and feature extraction, feature selection, and classification.]
{Processing pipeline consisting of three stages: preprocessing and feature extraction, feature selection, and classification using convolutional neural network taking the form of a residual neural network.}
\label{fig:pipeline}
\end{figure}
\subsection{Preprocessing and Feature Extraction}
\label{sec:preprocessing}
The first modules of the starting movement detector consist of the data preprocessing and feature extraction. We compare four different sets of features.
The raw data is provided by the accelerometer and gyroscope. Before feeding it into the model, the x, y and z axis are rotated to lie flat on the earth, i.e., the z-axis pointing towards the sky. This is done with the help of rotational information provided by a magnetometer. To avoid the tedious estimation of device to VRU transformation, we consider the magnitude of the x and y axis. This leaves four features to be used: the gravity adjusted magnitude and z values of the accelerometer and gyroscope. This comprises the first set of features. A schematic of this procedure is depicted in Fig. \ref{fig:preprocess_pipeline}.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{preprocess_pipeline.png}
\caption[Two-step pre-processing of the raw sensor signal; rotation and magnitude calculation.]
{Two-step pre-processing of the raw sensor signal. First attitude information is used to rotate x, y and z value of the accelerometer and gyroscope, then the magnitude of x and y is calculated.}
\label{fig:preprocess_pipeline}
\end{figure}
Additionally, the Fourier Transform up to the fifth degree was constructed based on a sliding window segmentation of the preprocessed data. It provides coefficients and phase over windows of size 640\,ms and 5120\,ms. This comprises the second set of features.
The preprocessed data was further handled to generate over 150 features, including the mean, variance, energy, min, max and polynomial characteristics over window sizes of 100\,ms, 500\,ms, 1000\,ms and 2000\,ms. Those features (including the Fourier coefficients) were sieved through a filter approach consisting of four procedures depicted in Fig. \ref{fig:feature_selection} that selects the ten most indicative features of each cross-validation fold separately. To obtain a large diversity concerning the selected features four complementary types of filters were picked. Two feature selection techniques: minimum redundancy maximum relevance (MrMR), mutual information (MIFS); and two model based methods applying an elastic net (ElasticNet) and extreme gradient boosted trees (XGBoost). An union set of those selected features was then utilized, amounting to roughly 19 features per fold. Throughout this text, this third set of features is referred to as filter features.
\begin{figure}
\centering
\includegraphics{feature_selection.png}
\caption[The filter pipeline reducing the number of selected features.]{The filter approach pipeline reducing the selected features. Four classifiers select the 10 best features and thereafter a union produces the final features set.}
\label{fig:feature_selection}
\end{figure}
Moreover, the Fourier coefficients and phases were evaluated for each sensor (accelerometer, gyroscope) and window (5120\,ms, 640\,ms) separately. Also a combination of the filter features and the Fourier phase was tested to see if further improvements could be achieved. They were therefore simply concatenated creating the largest of input spaces containing 31 features and comprising the fourth set of features.
\subsection{Classification using Residual Networks}
\label{sec:classification}
The last stage of the starting movement detector realized by mean of deep learning, i.e., a residual neural network, which is a special kind of convolutional neural network. Deep learning has been recently proven to be extremely successful in various domains and has lead to an increased usage of machine learning techniques in a multitude of fields.
Convolutional neural networks (CNN) have already been applied to many practical tasks and have risen in popularity after achieving superhuman accuracy on image classification tasks \cite{ILSVRC15}. They work by sequentially applying convolutions over an input space and are trained via back-propagation of error gradients. Their input space typically takes the form of a 2D or 3D image with depth greater or equal to one. Stacking multiple layers of convolution helps to identify broader features and covers more of the input space the deeper the network grows, enabling the model to learn abstract characteristics regarding the data.
\subsubsection{Residual Networks}
\label{js:sec:Residual_Networks}
In standard convolutional networks, a layer is only connected to the preceding and proceeding layers limiting the information visible to deeper layers to activations of the preceding layer, i.e. only features capture by previous layer are attainable. An architecture getting around this constraint is the Residual Network (ResNet) proposed by He et al. \cite{he2015}. A network comprised of this ResNet architecture won the 1st place of the ImageNet Large Scale Visual Recognition Competition \cite{ILSVRC15} in 2015. It was originally designed to conquer the problem of degradation in very deep networks. When deeper networks start converging accuracy gets saturated and then degrades rapidly with increasing network depth. To resolve this issue, the authors introduced shortcut connections skipping convolutional layers and therefore providing deeper layers with the identity of the original input.
Instead of learning a direct mapping of $H(x): x \rightarrow y$, a residual function $F \left(x \right) = H\left(x\right) - x$ is defined.
It can be reformed into $H(x) = F(x) + x$, where $F(x)$ and $x$ represent the non-linear layers, e.g., convolution, activation or regularization layers, and the identity function respectively.
According to the findings of He et al. the residual function $F(x)$ is easier to optimize than $H(x)$. For example when trying to learn the identity function, learning to push $F(x)$ to zero can be easily achieved using a stack of non-linear convolutional layers.
\paragraph{Residual Unit}
\label{js:sec:Residual_Unit}
The residual unit is a computational unit composing the identity shortcut $x$ and the residual function $F(x)$, which itself is composed of convolutional, activation and batch normalization layers \cite{ioffe2015}. He et al. strongly suggest to apply them in the order of batch normalization, activation function followed by the convolution. The best position of the shortcut is around two sets of layers ordered as just mentioned \cite{he2016}. A schematic of a residual unit is depicted in Fig \ref{fig:resunit}.
\begin{figure}[h]
\centering
\includegraphics[scale=0.55]{resunit.png}
\caption[Schematic of a residual unit composed of the residual function $F(x)$ and the identity shortcut $x$.]
{Schematic of a residual unit composed of the identity shortcut $x$ (left) and the residual function $F(x)$ (right) transforming the input from the previous layer $x_l$ to the input of the next layer $x_{l+1}$.}
\label{fig:resunit}
\end{figure}
\subsubsection{Regularization}
\label{js:sec:Regularization}
Regularization methods aim at decreasing the validation and test error, thus increasing the generalization abilities of machine learning models. In the following the regularization schemes adopted in this work will be described.
\paragraph{Spatial Dropout}
\label{js:sec:Spatial_Dropout}
Spatial dropout is a special form of dropout being used in convolutional networks, motivating the learning of more independent features. In contrast to normal dropout, where single neurons are excluded, whole filters are set to zero. This mimics the original intent of omitting complete and randomly selected features during training more closely than just discarding parts of it. Although spatial dropout increases the temporal requirements while training, it encourages the learning of more robust features, consequently leading to a smaller validation and test error, i.e., better generalization ability.
\paragraph{Batch Normalization}
\label{js:sec:Batch_Normalization}
To speed up the learning process and increase the generalization ability, batch normalization can be applied. It is implemented by adding special layers before the convolutions, which normalizes the input signal of those layers according to the batch statistics, namely the mean and variance of the current mini-batch \cite{ioffe2015}.
This works because it minimizes the amount of covariance shift of the hidden units. As an example, consider a network trained on black cat images: When applying this network on colored cats, it will perform worse, because while it still depicts a cat, the distribution of color values is different. In other words, if an algorithm learned some X to Y mapping, and if the distribution of X changes, then we might need to retrain the learning algorithm by trying to align the distribution of X with the distribution of Y. Batch normalization is favoring a more stable learning of features by normalizing the inputs to deep hidden layers reducing the effect of covariance shift, because the mean and variance remain similar during training \cite{ng2017}.
Additional benefits of batch normalization include being able to use higher learning rates, since the input values to the convolutional layers are restricted from being very high or very low. It also adds a regularization effect by adding noise to the activations (due to the statistics being calculated only on the current mini-batch).
\subsubsection{Architecture}
\label{js:sec:Architecture}
Two basic architectures of the ResNet were examined during the experiments. They are depicted in Fig \ref{fig:architecture}. The smaller one (Fig. \ref{fig:architecture}, top left) consists of two arrays of residual units, preceded by a first simple convolutional layer and followed by a fully connected layer to reduce the spatial dimensions to the desired output size of three classes. The first array holds two residual units and the seconds holds one. Softmax is applied to obtain values in the interval of $[0, 1]$. The network starts off with eight filters and doubles the amount for the second array of residual units, in this instance also the spatial dimension are halved. In this case, an additional convolution takes place in the shortcut to fit the passing identity to the new dimensions.
The larger network follows (Fig. \ref{fig:architecture}, right) the same architectural approach but with three arrays holding two residual units each and resulting in 14 layers also counting the first convolutional and last fully connected layers. The first layer and array begin with 16 filters, doubling the number of filters and reducing the spatial dimensions by one half every array, just as in the smaller model.
Both networks were trained to optimize the weighted cross-entropy applying the Adam-Algorithm. As indicated earlier, spatial dropout and batch normalization were applied. In this project the residual unit was structured as explained in Section \ref{sec:Residual_Unit} with the addition of a dropout layer right before each convolution (excluding the first convolutional layer), as can be seen in the lower left of Fig. \ref{fig:architecture}. The convolutions were carried out with 3x3 filters, stride set to one and zero padding.
\begin{figure}[h]
\centering
\includegraphics[scale=0.5]{architecture.png}
\caption[Architecture of the small eight layer and large 14 layer networks, with close up of a residual unit.] {Architecture of the networks used: (top-left) the small network consisting of eight layers and starting with eight filters. (right) the larger network consisting of 14 layers starting with 16 filters. The shortcut connection each encompasses two blocks of batch normalization, activation function (Rectified Linear Unit), spatial dropout and convolution as depicted in the bottom-left.}
\label{fig:architecture}
\end{figure}
\section{Data Acquisition and Evaluation}
\label{js:sec:data}
\subsection{Data Acquisition}
\label{js:sec:data_acqu}
The data of 49 test persons was gathered during a week long experiment at a public intersection incorporating bicycle lanes as well as walkways. It consists of 84 start scenes and the ground truth was generated through a wide angle stereo camera system and visual inspection~\cite{mbBZH+18}. The cyclist were instructed to cross the road or drive alongside it, adhering to the traffic rules. Each carried a smart phone (Samsung Galaxy S6) in the right trouser pocked.
\subsection{Evaluation}
We model the starting movement detection problem as a classification problem. It consists of three classes indicating \textit{waiting}, \textit{starting} and \textit{moving}, which split a time series in three parts. This separation is depicted in Fig. \ref{fig:classes}. The most common label is \textit{waiting} ($I$) and describes an interval of low movement and acceleration, while the cyclist is waiting, e.g., at a traffic light. An initial movement that leads to a start, including the starting motion up to the \textit{moving} phase is labeled as \textit{starting} ($II$). Thereby, \textit{moving} ($III$) is defined as the first movement of the back wheel followed by an acceleration and driving.
\begin{figure}
\centering
\includegraphics[scale=0.2]{ex_netout_scene.png}
\caption[The three classes making up the labeling and separation of a time series with example probabilties for the third class.]
{In red: example probabilities for the \textit{moving} class of a starting prediction. Area $I$ is the \textit{waiting} class of low activity; $II$ denotes the \textit{starting} class meaning the initial motion leading to the start, beginning at time $s$; $III$ is the \textit{moving} class labeled after the first movement of the bicycles back wheel, marked as $m$.}
\label{fig:classes}
\end{figure}
In order to calculate scores on the validation and test data the $F_1$ score was used. It is calculated from the precision and recall, which in turn are calculated from the true-positive, true-negative, false-positive and false-negative values defined as follows: if the prediction of \textit{moving} is after the true \textit{starting} label, it is classified as a true-positive. This means that a classification of a start is still classified as a true positive even if it is detected before the \textit{moving} label but during the initial movement of the test person intending to start. Else if it is too early, it is classified as a false-positive. True-negatives normally do not appear since every time series contains \textit{starting} label. When the predictor would not output the \textit{starting} class at any time step, it is classified as a false-negative.
The delay of the starting prediction is determined only for true positive classifications, as otherwise false positives would greatly reduce the average measured delay. It is calculated relative to the first true \textit{moving} label. Negative values are possible, because a classification during the \textit{starting} phase is still considered a true-positive.
\section{Experimental Results}
\label{js:sec:experimental_results}
The creation and evaluation of models for the starting intention detection is separated into two parts: selecting the input data and corresponding features before optimizing the model architecture.
\subsection{Feature Selection}
\label{js:sec:Feature_Selection}
First, the feature selection was conducted. The results turned out as expected: a smaller input window leads to smaller delays, a larger input window leads to larger delays. The convolutions themselves add to the delay as they cover multiple time steps on higher level learned features. The filter features lead to higher delays compared the raw features, due to the mean, variance etc. being calculated over windows as well. Nonetheless, they are more robust and provide a more stable score over different thresholds applied on the final softmax output.
The overall best scores were achieved by the filter features. On an input window of 100 time steps it reaches an $F_1$-score of 95\% and has a mean delay of 861\,ms. With an input window of 50 time steps it has a slightly faster mean delay of 783\,ms.
The raw features end up in the middle field, yet, not as good as the filter features with an input window of 10 time steps. With an input window of 100 time steps their highest $F_1$-score is 76\% with a very low mean delay of 257\,ms.
Combining the filter features with the phase of a discrete Fourier Transformation slightly reduces the scores compared to only filter features. Supposedly because the network is too small to deal with so many features. This assumption is strengthened considering that the same features are contained with addition of the phase.
Using the discrete Fourier Transform coefficients and phases thereof do not provide sufficient information for a starting intention detection. The highest $F_1$-score of 63\% was reached by the Fourier coefficients of the Accelerometer over a window of 5120\,ms with an input window of 100 time steps. All other Fourier features are worse, especially the phase reaching only a maximum $F_1$-score of 32\%.
\subsection{Architecture Optimization}
\label{js:sec:Architecture_Optimization}
For the architecture optimization, the filter features with an input window of 50 time steps was picked as a compromise between the classification score and delay. Also an input window of 10 time steps is quite prone to false positives, since there is often movement while waiting, e.g., switching the leg on which the cyclist rests, preparing the pedals or simply looking around. Due to the same reasons filter features are preferred to the raw data , because spikes of activity are smoothed out over the feature windows. An overview of the delay distribution can be seen in Fig. \ref{fig:filter_06_results}. Most starting detection occur between 0.5\,s and 1\,s after the \textit{moving} label. It is noticeable though, that some detections happen almost 1\,s before the bicycle moves.
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{filter_06_results.png}
\caption[Classification results and Kernel Density Estimation of the delay of the best performing model]
{Classification results (left) and Kernel Density Estimation of the delay (right) of the best performing model, which processes the filter features, has an input window of 50 time steps and keep probability of 0.6. Most classification occur between 0.5\,s and 1\,s after the \textit{moving} label. Noticeably, some starting detection happen almost 1\,s before the bicycle moves.}
\label{fig:filter_06_results}
\end{figure}
The bigger models did not reach the same performance on the test set, having a maximal $F_1$-score of 93\% with a keep probability of 0.8. This is supposedly because of over-fitting since the training loss was comparable to the smaller models. Although, the mean delay times improved with the model size, reaching a minimal mean delay of 592\,ms with a keep probability of 0.7. A decisive conclusion on the keep probability could not be drawn due to no apparent trend for it being high or low.
Choosing a minimum desired $F_1$-score of 95\%, the best model was achieved by the small network and a keep probability of 0.6. It reached a mean $F_1$-score of 97\% and delay of 783\,ms.
\section{Conclusion}
\label{js:sec:Conclusion}
The high scores on the selected models show, that the use of convolutional neural network on sensory data from smart phones is justified. Our approach reaches an $F_1$-score of 97\% with an average delay of 783\,ms. Deep Residual Neural Networks proved to be applicable for the task of starting detection. They achieve a reliable classification, although with a comparatively high delay of over half a second.
Further improvements can be achieved with more training data.
Cutting the data to not include stopping motion greatly improved scores, since very high excitations are basically only encountered at the starting motion. This could be implemented with a separate waiting detector, which activates the starting detector only when the stopping motion was finished and switching back to it once the driving.
\input{sections/acknowledgment}
\bibliographystyle{splncs}
\section*{\large Acknowledgment}
This work results from the project DeCoInt$^2$, supported by the German Research Foundation (DFG) within the priority program SPP 1835: "Kooperativ interagierende Automobile", grant number SI 674/11-1.
|
\section{Introduction}
The self-avoiding walk (SAW) is a combinatorial
model of lattice paths without self-intersections.
In addition to its intrinsic mathematical interest, it arises in
polymer science as a model of linear polymers, and in statistical mechanics
as a model that exhibits critical behaviour. The mathematical problems
associated to the SAW are notoriously difficult and there remain longstanding
unsolved problems that are central to the subject. A closely related model
is the weakly self-avoiding walk (WSAW),
which is predicted to exhibit the same critical behaviour as
the SAW.
The critical behaviour of the SAW or WSAW is expressed in terms of
critical exponents, which have a qualitative and quantitative relationship
with the critical exponents in models of ferromagnetism including
the Ising and $|\varphi|^4$ spin models.
Within physics, the critical exponents are well understood,
but they nevertheless present deep mathematical problems.
This article is a review of recent mathematical results about critical
exponents for the WSAW and $|\varphi|^4$ models,
with focus on the critical
behaviour of the susceptibility.
Some background on the SAW
and Ising models is provided for motivation and context.
The results we present
involve a unified treatment of the WSAW and $|\varphi|^4$
models, via an exact relation between the WSAW and a ``zero-component''
$|\varphi|^4$ model.
The proofs are based on a rigorous version of Wilson's renormalization group
(RG) approach. We provide an introduction to the RG method from the
perspective of a mathematician.
\section{Self-avoiding walk}
\label{sec:SAW}
We discuss the SAW and WSAW models, as well as their long-range versions.
The emphasis is on the critical behaviour, particularly for the susceptibility.
\subsection{Strictly self-avoiding walk (SAW)}
\label{sec:SSAW}
\subsubsection{Universality and scale invariance}
An $n$-step SAW on the integer lattice $\Zbold^d$ is
a map $\omega :\{0,1,\ldots,n\} \to \Zbold^d$, such that the Euclidean distance
between $\omega(i)$ and $\omega(i+1)$ equals $1$ (nearest-neighbour
steps), and such that $\omega(i) \neq \omega(j)$ for all $i \neq j$
(self-avoidance).
Let $\Scal_n$ denote the finite set of $n$-step SAWs
with $\omega(0)=0$ (walk starts at origin of $\Zbold^d$), and let $c_n$ be its
cardinality. We declare each element of $\Scal_n$ to have equal probability,
which must therefore be $c_n^{-1}$.
Random $n$-step SAWs on the square lattice
$\Zbold^2$, with $n=10^2$ and $10^8$, are depicted in Figure~\ref{fig:saws}.
\begin{figure}[h]
\centering{
\includegraphics[scale=0.13]{d2saw100a.png}
\hspace{18mm}
\includegraphics[scale = 0.013]{d2saw100000000a.jpg}
}
\caption{Random SAWs on the square lattice
$\Zbold^2$ with $n=10^2$ and $n=10^8$; image Nathan Clisby.}
\label{fig:saws}
\end{figure}
The $10^8$-step SAW in Figure~\ref{fig:saws}
would not be statistically
distinguishable from a SAW instead on the hexagonal lattice,
or on the triangular lattice, or indeed on any one of a wide
variety of 2-dimensional lattices. This feature is called \emph{universality}.
It is similar to the invariance principle for Brownian motion,
which is the generalization of the central limit theorem that asserts that
(ordinary) random walk with any finite-variance step distribution
converges to Brownian motion. The search for a corresponding statement
for SAW, i,e., the identification of a limiting probability law for SAW---a
\emph{scaling limit}---is one of the subject's big problems.
A related and in general unproven
feature is \emph{scale invariance}: a $10^{10}$-step SAW,
rescaled to the same size as the $10^8$-step SAW in Figure~\ref{fig:saws},
would be statistically indistinguishable from the $10^8$-step SAW.
The scale invariance is quantified in terms of a
universal \emph{critical exponent} whose existence has not been proven in general.
\subsubsection{The SAW connective constant}
Since $c_nc_m$ counts the number of ways that an $n$-step and an $m$-step SAW
can be concatenated, with the two subwalks possibly intersecting each other,
we have $c_{n+m} \le c_n c_m$. From this, it readily follows
that there exists $\mu=\mu(d)$, the \emph{connective constant}, such that
$\lim_{n\to\infty}c_n^{1/n} = \mu$ and $c_n \ge \mu^n$ (see, e.g., \cite{MS93}).
Good numerical estimates and rigorous
bounds on the connective constant are known, but the exact value for $\Zbold^d$ is not
known for any $d \ge 2$. For SAWs defined instead
on the hexagonal lattice, it has been proved that
$\mu=\sqrt{2+\sqrt{2}}$ \cite{D-CS12h}.
As the dimension $d$ goes to infinity, there is an asymptotic expansion
$\mu \sim 2d-1 + \sum_{n=1}^\infty a_n(2d)^{-n}$ with integer coefficients
$a_n$ whose values are known up to and including $a_{11}$ \cite{CLS07}.
The connective constant for SAWs in more general settings than $\Zbold^d$
is a topic of current research \cite{GL18}.
Our focus here is on
the asymptotic behaviour of the ratio $c_n/\mu^n$, which,
unlike the connective constant, is predicted to have a \emph{universal}
asymptotic behaviour.
\subsubsection{SAW critical exponents}
\label{sec:SAWexponents}
There is considerable evidence from numerical studies and from arguments
from theoretical physics that there exists $\gamma$ (depending on the
dimension $d$) such that
\begin{equation}
\label{e:cnasy}
c_n \sim A \mu^n n^{\gamma-1} \qquad (n \to \infty).
\end{equation}
(The symbol $\sim$ denotes that the ratio of left-hand and right-hand sides has
limit $1$.)
Since $c_n \ge \mu^n$, necessarily $\gamma \ge 1$.
The \emph{susceptibility} $\chi$ is the generating function of $c_n$:
\begin{equation}
\label{e:SAWsusceptibility}
\chi(z) = \sum_{n=0}^\infty c_n z^n
\qquad (z \in \mathbb{C}).
\end{equation}
It has radius of convergence $z_c = \mu^{-1}$ since $c_n^{1/n}\to \mu$.
In view of \eqref{e:cnasy}, $\chi$ can be expected to obey
\begin{equation}
\chi(z) \sim \frac{A'}{(z_c - z)^{\gamma}}
\qquad
(z \uparrow z_c).
\end{equation}
Let $R_n^2$ be the average over $\Scal_n$ of $\|\omega(n)\|_2^2$.
Then $R_n$ is the root-mean-square displacement of $n$-step SAWs, which
is a measure of the average end-to-end distance of an $n$-step SAW.
There is again considerable evidence that there exists $\nu$
(depending on $d$) such that
\begin{equation}
R_n \sim D n^{\nu} \qquad (n \to \infty).
\end{equation}
The exponent $\nu$ quantifies scale invariance.
This article is about \emph{critical exponents} such as $\gamma,\nu$
for certain SAW and lattice spin models, with
the emphasis on $\gamma$. There are other critical exponents that we
do not discuss.
The exponents are predicted to be universal, depending
essentially only on the dimension of the lattice. For example, $\gamma$ and $\nu$
should have the same values on the square lattice as on the hexagonal
or triangular lattices,
unlike the connective constant.
A central problem in the subject is to prove the existence of
the critical exponents and to show that they have the values listed in
Table~\ref{table:sawexponents}.
\begin{table}[!h]
\caption{SAW critical exponents.}
\label{table:sawexponents}
\begin{tabular}{@{}rll}
\hline
$d$ & $\gamma$ & $\nu$ \\
\hline
$2$ & $\frac{43}{32}$ & $\frac 34$\\
$3$ & $1.15695300(95)$ & $0.58759700(40)$\\
$4$ & $1$ \, \text{with $\log^{1/4}$} & $\frac 12$ \, \text{with $\log^{1/8}$}\\
$\ge 5$ & $1$ & $\frac 12$ \\
\hline
\end{tabular}
\end{table}
The rational exponents for $d=2$ in Table~\ref{table:sawexponents}
were computed by Nienhuis \cite{Nien82} using nonrigorous arguments
based on spin systems like the ones we discuss later.
An important breakthrough came with the identification of the stochastic
process
${\rm SLE}_{8/3}$ (Schramm--Loewner Evolution with parameter $\frac 83$)
as the only plausible candidate for the scaling
limit \cite{LSW04}, which additionally provided an alternate explanation for the
exponents $\frac{43}{32}$ and $\frac 34$. However it remains an open problem to prove
the existence of the critical exponents for $d=2$,
to prove that they have the rational values
in Table~\ref{table:sawexponents},
and to prove that ${\rm SLE}_{8/3}$ truly is the scaling limit.
For $d=3$, there is no currently known stochastic process to serve as a scaling
limit for SAWs, and the best estimates for critical exponents come from
numerical work.
To compute the exponents, it is natural to attempt to enumerate SAWs
for small $n$ and then extrapolate. Fisher and Gaunt \cite{FG64}
found $c_n$ by hand for $n \le 11$, in all dimensions. More than half a century
later, for $d=3$ the enumeration has reached only $n=36$,
which is insufficient
for reliable high-precision estimation of the exponents. See
Table~\ref{table:sawcount}.
It is a challenging problem in enumerative combinatorics
to produce a good algorithm to extend Table~\ref{table:sawcount} significantly.
The Monte Carlo method known as the
\emph{pivot algorithm} gives more accurate estimates for
critical exponents, and those appearing for $d=3$ in Table~\ref{table:sawexponents}
are estimates using this method \cite{CD16,Clis17}.
\begin{table}[!h]
\caption{$c_n$ for $d=3$, $n \le 36$. The most recent values
are for $31\le n\le 36$ \cite{SBB11}.}
\label{table:sawcount}
\begin{tabular}{@{}rlrlrl}
\hline
$n$ & $c_n$ & $n$ & $c_n$ & $n$ & $c_n$\\
\hline
1 & 6
&13 & 943\,974\,510
& 25 & 116\,618\,841\,700\,433\,358\\
2 & 30
&14 & 4\,468\,911\,678
& 26 & 549\,493\,796\,867\,100\,942\\
3 & 150
&15 & 21\,175\,146\,054
& 27 & 2\,589\,874\,864\,863\,200\,574\\
4 & 726
&16 & 100\,121\,875\,974
& 28 & 12\,198\,184\,788\,179\,866\,902\\
5 & 3\,534
&17 & 473\,730\,252\,102
& 29 & 57\,466\,913\,094\,951\,837\,030\\
6 & 16\,926
&18 & 2\,237\,723\,684\,094
& 30 & 270\,569\,905\,525\,454\,674\,614\\
7 & 81\,390
& 19 & 10\,576\,033\,219\,614
& 31 & 1\,274\,191\,064\,726\,416\,905\,966\\
8 & 387\,966
& 20 & 49\,917\,327\,838\,734
& 32 & 5\,997\,359\,460\,809\,616\,886\,494\\
9 & 1\,853\,886
& 21 & 235\,710\,090\,502\,158
& 33 & 28\,233\,744\,272\,563\,685\,150\,118\\
10 & 8\,809\,878
& 22 & 1\,111\,781\,983\,442\,406
& 34 & 132\,853\,629\,626\,823\,234\,210\,582\\
11 & 41\,934\,150
& 23 & 5\,245\,988\,215\,191\,414
& 35 & 625\,248\,129\,452\,557\,974\,777\,990\\
12 & 198\,842\,742
& 24 & 24\,730\,180\,885\,580\,790
& 36 & 2\,941\,370\,856\,334\,701\,726\,560\,670\\
\hline
\end{tabular}
\end{table}
The exponents $\gamma=1$ and $\nu=\frac 12$
for $d\ge 5$ in Table~\ref{table:sawexponents} are the same as
those of simple random walk.
This is summarized by the statement that the \emph{upper critical dimension} is 4.
Here is an argument to guess this: Brownian paths
are 2-dimensional, and since two 2-dimensional objects generically do not intersect
in dimensions $d>4$, SAW should behave like simple random walk when $d>4$.
There is a full rigorous understanding of dimensions $d \ge 5$. The
following theorem from \cite{HS92a} is an example of this.
\begin{theorem}
\label{thm:HS92a}
For $d \ge 5$, the scaling limit of SAW is Brownian motion, and $\gamma=1$ and $\nu= \frac 12$ in the sense that as $n \to \infty$,
\[
c_n \sim A \mu^n ,
\quad\quad
R_n \sim Dn^{{1/2}} .
\]
\end{theorem}
Theorem~\ref{thm:HS92a} is proved using the lace expansion,
which was originally introduced in \cite{BS85}
and has subsequently been extended to many other high-dimensional models
including percolation \cite{Slad06,HH17book}.
For $d=4$, the logarithms in Table~\ref{table:sawexponents}
reflect the prediction that the two asymptotic formulas in
Theorem~\ref{thm:HS92a} must be modified by an additional factor $(\log n)^{1/4}$
for $c_n$ and $(\log n)^{1/8}$ for $R_n$.
We will return to such logarithmic factors
later.
For $d=2,3,4$, none of the entries in Table~\ref{table:sawexponents}
have been proved.
In 1962, Hammersley and Welsh \cite{HW62}
proved the following upper bound on $c_n$.
\begin{theorem}
\label{thm:HW}
For any $B>\pi(2/3)^{1/2}$ there
exists an $N$ such that for all $d \ge 2$,
\[
\mu^n \le c_n \le \mu^{n} e^{B\sqrt{n}} \quad (n \ge N).
\]
\end{theorem}
Shortly thereafter, for $d \ge 3$, Kesten improved the $\sqrt{n}$ in the exponent
to $n^{2/(d+2)}\log n$ \cite{Kest64,MS93}.
For $d=2$ the best improvements since 1962 are the replacement of $B$ by $o(1)$
\cite{Hutc18}, and a proof that the upper bound holds for infinitely many $n$ when
$B\sqrt{n}$ is replaced by $n^{0.4979}$ \cite{DGHM18}.
This is slow progress in over half a century.
For $R_n$, the best results are in the following theorem. The lower bound
was proved in \cite{Madr14} and the upper bound in \cite{D-CH13}.
\begin{theorem}
\label{thm:Rn}
For $d \ge 2$,
\[
\tfrac{1}{6}n^{\frac{2}{3d}} \le R_n \le o(n).
\]
\end{theorem}
The lower bound fails to prove that on average the endpoint of a SAW
is at least as far away as it is for simple random walk, namely $n^{1/2}$,
even though it
appears obvious that the self-avoidance constraint must push the SAW
farther than a walk without the constraint. The upper bound states
that $R_n/n \to 0$ but there is no bound on the rate. In particular,
it is not proved that there is a constant $C$ such that
$R_n \le C n^{0.99999}$.
The large gap for $d=2,3,4$ between the predicted results
in Table~\ref{table:sawexponents} and those proven in
Theorems~\ref{thm:HW}--\ref{thm:Rn} is
an invitation to look for more tractable models that ought to be in
the same universality class as SAW. The weakly
self-avoiding walk is such a model.
\subsection{Weakly self-avoiding walk (WSAW)}
\label{sec:WSAW}
There are two versions of the WSAW: one based on discrete time
(also known as the \emph{Domb--Joyce model}) and one based on
continuous time (also known as the \emph{lattice Edwards model}).
Our focus is on the latter. It differs from the SAW in two respects:
(i) the underlying simple random walk model takes its steps at random
times rather than after a fixed unit of time, and (ii) walks are allowed
to have self-intersections but are weighted as less likely according
to how much self-intersection occurs.
More precisely, let $\sigma_i$ be a sequence of independent exponential
random variables with mean $\frac{1}{2d}$. Let $(X(t))_{t \ge 0}$ denote the
random walk on $\Zbold^d$ which starts at the origin at time $t=0$, waits until time
$\sigma_1$ and then steps immediately to a randomly chosen one of the $2d$ neighbours of
the origin, then waits an amount of time $\sigma_2$ until stepping to
an independently randomly chosen neighbour of its current position,
and so on.
The \emph{self-intersection local time} up to time $T$ is the random variable
\begin{equation}
I(T)
= \int_0^T \int_0^T \mathbbm{1}_{X(s)=X(t)}ds \, dt
,
\end{equation}
which provides a measure of how much time the random walk has spent intersecting itself
by time $T$.
For fixed $g > 0$, let $c_{T,g}= E(e^{-gI(T)})$, where
$E$ denotes expectation for the random walk. As a function of $T$,
$c_{T,g}$ is analogous to the sequence $c_n$ for SAW.
Every walk contributes to $c_{T,g}$, but an exponential weight diminishes the
role of walks with large self-intersection local time.
The elementary argument which led to the existence of the connective constant
generalizes to $c_{T,g}$, and yields the conclusion that there exists
$\nu_c(g) \le 0$ such that $\lim_{T\to\infty}c_{T,g}^{1/T} = e^{\nu_c}$
and $c_{T,g} \ge e^{\nu_c T}$. Thus the \emph{susceptibility}
\begin{equation}
\chi(g,\nu) = \int_0^\infty
c_{T,g} e^{-\nu T}
dT \qquad (\nu \in \mathbb{R})
\end{equation}
is finite if and only if $\nu > \nu_c$.
WSAW is predicted to be in the same universality class as SAW for all $g>0$,
meaning that it has the same critical exponents and scaling limits as SAW.
The following theorem is an example of this, for the upper critical
dimension $d=4$ and for sufficiently
small $g>0$ \cite{BBS-saw4-log}.
It reveals that $\gamma=1$ with a modification by a
logarithmic correction as indicated in Table~\ref{table:sawexponents}.
In the physics literature, the computation of logarithmic corrections for $d=4$ goes
back half a century \cite{LK69,BGZ73,WR73,Dupl86}.
A number of related results have been proved for the 4-dimensional WSAW
\cite{BBS-saw4,ST-phi4,BSTW-clp},
all of which are consistent with the predictions for SAW.
\begin{theorem}
\label{thm:saw4-log}
For $d = 4$ and small $g>0$, as $t=\nu-\nu_c \downarrow 0$,
\begin{align*}
\chi(g,\nu)
& \sim A_g\frac {1}{t} |\log t|^{ \frac 14}
.
\end{align*}
\end{theorem}
The proof of Theorem~\ref{thm:saw4-log} is
based on a rigorous and nonperturbative implementation of the renormalization
group approach \cite{WK74}.
The renormalization group has for decades been one of the basic tools of theoretical
physics, for which Wilson
was awarded the Nobel Prize in Physics in 1982.
Its reach extends across critical phenomena, many-body theory,
and quantum field theory. We make no attempt to refer to the vast
physics literature, e.g., \cite{Card96}.
In a 1972 paper with the intriguing title
``Critical exponents in $3.99$ dimensions'' \cite{WF72},
Wilson and Fisher considered the dimension $d$ as a
\emph{continuous} variable $d=4-\epsilon$,
and applied the renormalization group approach to compute critical exponents in dimension $4-\epsilon$ for small $\epsilon>0$. This captures the idea that
the critical behaviour can be expected
to vary in a continuous manner as the dimension varies,
so dimensions below $d=4$ can be regarded as a perturbation of $d=4$.
Within physics, this has become well developed and it is found that, e.g.,
${ \gamma = 1 + \frac 18 \epsilon + \cdots + ({\rm known})\epsilon^6 + \cdots}$.
Although presumably a divergent asymptotic expansion,
such $\epsilon$-\emph{expansions} have been used to obtain numerical estimates
of critical exponents for $d=3$.
However, from the perspective of mathematics, the dimension is not a continuous
variable and this raises more questions than it answers.
\subsection{Long-range walks}
A different idea to move slightly below the upper critical dimension
was also proposed in 1972 \cite{FMN72,SYI72}.
In this framework, the upper critical dimension (formerly $d=4$) assumes
a continuous value $d_c \in (0,4)$. In particular,
for $d=1,2,3$ we can choose
$d_c=d+\epsilon$, and thereby study integer dimension
$d$ below $d_c$ without the need to define the WSAW in any non-integer dimension.
In our present context, this idea can be formulated in terms of walks
taking long-range steps, as follows.
The long-range steps are defined in terms of a parameter
$\alpha \in (0,2)$.
Let $d=1,2,3$, and
consider the random walk on ${\mathbb Z}^d$ that takes independent steps of length $r$ (in any direction) with probability proportional to
$r^{-(d+\alpha)}$.
This step distribution has infinite variance, a \emph{heavy tail}.
A convenient choice of such a step distribution is the
fractional power $-(-\Delta)^{\alpha/2}$ of the discrete Laplace operator
\begin{equation}
\label{e:Deltadef}
(\Delta f)_x = \sum_{e\in \Zbold^d:\|e\|_2=1} (f_{x+e}-f_x).
\end{equation}
Thus we consider the random walk on $\Zbold^d$ with transition probabilities
\begin{equation}
\label{e:pxy}
p_{x,y}= {\mathbb P}(\text{{ next step to $y$}}|\text{{ now at $x$}})
\propto -((-\Delta)^{\alpha/2})_{x,y} \asymp \frac{1}{\|x-y\|_2^{d+\alpha}}
\end{equation}
(the notation $f\asymp g$ means $cg \le f \le Cg$ for some constants $c,C$).
This heavy-tailed random walk converges to an $\alpha$-stable
process. The paths of an $\alpha$-stable process have dimension $\alpha$
\cite{BG60}, so two such paths
generically do not intersect in dimensions $d>2\alpha$. This
suggests that in dimensions $d>d_c=2\alpha$, long-range SAW or WSAW
should behave like the $\alpha$-stable process. Figure~\ref{fig:long-range-srw}
shows a 2-dimensional long-range simple random walk with $\alpha = 1.1$,
next to a nearest-neighbour walk for comparison.
The heavy tail of the long-range
walk produces big jumps, which in turn create fewer self intersections,
thereby making it easier for a walk to be self-avoiding and
lowering the upper critical dimension.
\vspace*{-15pt}
\begin{figure}[!h]
\hspace{10mm}
\centering
\includegraphics[scale = 0.18]{d2srw100000-min.pdf}
\hspace{10mm}
\includegraphics[width=50mm]{path_alpha_1point1_steps_100000.png}
\caption{$2$-dimensional $10^5$-step nearest-neighbour (left) and long-range
(right, $\alpha = 1.1$) walks (not to scale: the diameter of the left walk is
about 100 times smaller than that of the right walk); image Nathan Clisby. }
\label{fig:long-range-srw}
\end{figure}
We can define a long-range model of SAW as follows. A
long-range $n$-step SAW is any sequence
$\omega=(\omega(0),\ldots, \omega(n))$ with $\omega(i) \in \Zbold^d$
and $\omega(i) \neq \omega(j)$ for $i \neq j$. Let $\omega(0)=0$.
The probability of $\omega$ is the product
$\prod_{i=1}^n p_{\omega(i-1),\omega(i)}$, with $p_{x,y}$ given
by \eqref{e:pxy}.
The following theorem \cite{Heyd11} proves that
this SAW does behave like the unconstrained long-range random walk in dimensions $d\ge 1$
as long as $\alpha < \frac d2$. This is a long-range version
of Theorem~\ref{thm:HS92a}; its proof is also based on the lace expansion.
A technical point is that the theorem actually applies to a so-called
\emph{spread-out} version of the long-range SAW, a small modification.
\begin{theorem}
\label{thm:SAWlr}
For $\alpha \in (0,2)$ and $d>2\alpha$, the scaling limit of
spread-out long-range SAW is an $\alpha$-stable process, and the
critical exponents are $\gamma=1$, $\nu = \frac 1\alpha$.
\end{theorem}
However, our primary interest here is to go below $d_c$ to observe scaling
behaviour that is different from that of the $\alpha$-stable process.
For this, we consider WSAW and its susceptibility $\chi$ defined as in
Section~\ref{sec:SAW}\ref{sec:WSAW}
but with the expectation $E$ now with respect to the
continuous-time long-range random walk.
We choose $\alpha = \frac 12 (d+\epsilon)$, so that $d=d_c-\epsilon$ is
below the critical dimension $d_c=2\alpha$.
The following theorem \cite{Slad18} gives an example of an $\epsilon$-expansion. It is proved using a rigorous renormalization group method.
The restriction on $g$ in the hypothesis of the theorem is
used in the proof,
but the statement is expected to be true for all $g>0$.
Further results are obtained in \cite{LSW17}.
A related paper which is focused on renormalization rather than critical
exponents is \cite{MS08}.
\begin{theorem}
\label{thm:WSAWlr}
Let $d = 1,2,3$.
For small $\epsilon>0$, for $\alpha = \frac 12 (d+\epsilon)$,
and
for $g\in [c\epsilon,c'\epsilon]$ for some $c<c'$,
there is a constant $C$ such that
as $t=\nu-\nu_c \downarrow 0$,
\begin{align*}
&
C^{-1}\frac{1}{t^{ 1+ \frac{1}{4} \frac\epsilon\alpha -C\epsilon^2}}
\le
\chi(g,\nu) \le C\frac{1}{t^{ 1+ \frac{1}{4}\frac\epsilon\alpha +C\epsilon^2}}
,
\quad \text{i.e.,} \;\;
{ \gamma = 1 + \frac{1}{4} \frac\epsilon\alpha + O(\epsilon^2)}.
\end{align*}
\end{theorem}
\section{Spin systems}
Spin systems are basic models in statistical mechanics.
We discuss two examples here: the Ising and $|\varphi|^4$ models.
At first sight, spin
systems appear to be unrelated to SAW, but a connection will
be made in Section~\ref{sec:n0}.
\subsection{Ising model}
The most fundamental spin system is the Ising model
of ferromagnetism, which is defined
as follows. Let $\Lambda \subset \Zbold^d$ be a finite box.
An \emph{Ising spin configuration} on $\Lambda$ is an assignment of $+1$ or
$-1$ to each site in $\Lambda$, i.e., $\sigma = (\sigma_x)_{x\in\Lambda}$
with $\sigma_x \in \{-1,1\}$. An example for $d=2$ is depicted in
Figure~\ref{fig:ising-arrows}.
\begin{figure}[h]
\begin{center}
\includegraphics[scale = 0.35]{Ising-arrows}
\end{center}
\caption{
An Ising spin configuration.}
\label{fig:ising-arrows}
\end{figure}
Spin configurations are random, with a probability distribution parametrized
by \emph{temperature} $T$ and determined by the
energy of $\sigma$ which is defined to be
\begin{equation}
H_{\Lambda}(\sigma)
=
\frac 12 \sum_{x\in\Lambda} \sigma_x (-\Delta \sigma)_x .
\end{equation}
Here $\Delta$ is the discrete Laplace operator \eqref{e:Deltadef}, restricted to $\Lambda$.
Apart from an unimportant constant, $H_{\Lambda}(\sigma)$ is equal to
$- \sum_{x\sim y} \sigma_x\sigma_y$ where the sum is over all pairs of neighbouring
sites in $\Lambda$.
At temperature $T$, the
probability of $\sigma$ is given by the \emph{Boltzmann weight}
\begin{equation}
P_{T,\Lambda}(\sigma)
=
\frac{e^{-\frac 1T H_{\Lambda}(\sigma) }}
{\sum_{\sigma}e^{-\frac 1T H_{\Lambda}(\sigma) }}
\propto
e^{\frac{1}{T} \sum_{x\sim y} \sigma_x\sigma_y}
.
\end{equation}
Thus spin configurations with more alignment between neighbouring spins
are more likely than those with less alignment, and this effect is magnified
for small $T$ compared to large $T$.
\begin{figure}[h]
\begin{center}
\includegraphics[width = .20 \textwidth]{t_sub}%
\includegraphics[width = .20 \textwidth]{tc-interface}%
\includegraphics[width = .20 \textwidth]{t_sup}%
\end{center}
\begin{center}
{\scriptsize Low temperature $T<T_c$ \hspace{2mm} Critical temperature $T=T_c$ \hspace{3mm}
High temperature $T>T_c$}
\end{center}
\caption{Ising configurations on a $200 \times 400$ box,
with boundary spins fixed white on top half and dark on bottom half.
}
\label{fig:ising-2D}%
\end{figure}
For dimensions $d \ge 2$, there is a
\emph{critical} temperature $T_c$ such that when $T>T_c$ typical spin
configurations are disordered, whereas for $T<T_c$ there is long-range order.
This is depicted in Figure~\ref{fig:ising-2D} where the $+/-$ symmetry
is broken by a boundary condition. The behaviour at $T_c$, and
as $T$ approaches $T_c$, is of great current interest and there
is a vast literature, particularly for $d=2$ where the model is exactly solvable
and exciting connections with ${\rm SLE}$ have been discovered, e.g.,
\cite{CDHKS14}. At the critical temperature, the rich geometric structure
apparent in Figure~\ref{fig:ising-2D} is scale invariant.
Critical exponents are rigorously known for $d=2$ and for $d>4$ but not for
$d=3$, although in the physics literature the conformal
bootstrap has been used to compute exponents to high accuracy for $d=3$
\cite{EPPRSV14}.
A recent survey of mathematical work on the Ising and related models can be
found in \cite{Dumi19}.
\subsection{$|\varphi|^4$ model}
The $|\varphi|^4$ model is an extension of the Ising model in which the
Ising spin $\sigma_x \in \{-1,+1\}$ is replaced by an $n$-component vector
spin $\varphi_x \in \Rbold^n$ ($n\ge 1$).
To preserve translation invariance we replace the box used for the Ising
model by one with periodic boundary conditions, i.e., $\Lambda$ is a discrete
$d$-dimensional torus.
The {\it a priori} or \emph{single-spin} distribution
of $\varphi_x$ is set to be proportional to $e^{-V(\varphi_x)} d\varphi_x$, where
$d\varphi_x$ is Lebesgue measure on $\Rbold^n$ and
\begin{equation}
V(\varphi_x)
= \tfrac{1}{4}g |\varphi_x|^4 + \tfrac{1}{2} \nu |\varphi_x|^2
\end{equation}
with $g>0$, $\nu\in \Rbold$, and with $|\varphi_x|$ the Euclidean norm of $\varphi_x\in
\Rbold^n$. We are primarily interested in
$\nu<0$, in which case for $n=1$ the potential $V$ has the double-well shape
of Figure~\ref{fig:V}.
The probability density of a spin configuration
$(\varphi_x)_{x\in \Lambda}\in (\Rbold^n)^{|\Lambda|}$ is then proportional to
the Boltzmann weight
\begin{align}
\label{e:phi4}
dP_{g,\nu,\Lambda}(\varphi)
& \propto
e^{
-
\sum_{x \in \Lambda} (
V(\varphi_x)
+
\tfrac{1}{2} \varphi_x\cdot (-\Delta \varphi)_x
)
}
\prod_{x\in \Lambda}
d\varphi_x
.
\end{align}
\begin{figure}[h]
\begin{center}
\scalebox{0.75}
{\input{phi4-doublewell.pspdftex}}
\end{center}
\caption{\label{fig:V}
For $n=1$, the double-well potential $V$ (left) and single-spin density
$e^{-V}$ (right).}
\end{figure}
For $n=1$, spins are more likely to assume values near one of the two minima
of the double well. For $n >1$, there is a continuous set of minima.
The Laplacian term in \eqref{e:phi4} discourages large differences between
neighbouring spins and is thus a ferromagnetic interaction.
For example, for $n=1$ it encourages
the spins to break the symmetry and primarily prefer one minimum over the other.
Now $\nu$ plays the role played by the temperature $T$ in the Ising model, and there
is a phase transition and corresponding critical exponents associated with
a critical value $\nu_c(g)<0$.
For $\nu<\nu_c$, spins are typically aligned, whereas they are disordered
for $\nu>\nu_c$.
The existence of a phase transition is proved for $d \ge 3$ for
general $n \ge 1$ in \cite{FSS76}, and for $d=2$ and $n=1$ in \cite{GJS75};
the Mermin--Wagner theorem states that there is no phase transition for $n >1$
when $d=2$.
The \emph{susceptibility} is defined by
\begin{equation}
\label{e:phi4susceptibility}
\chi(g,\nu) = \lim_{\Lambda \uparrow \Zbold^d} \frac 1n \sum_{x \in \Lambda}
\int_{(\Rbold^n)^{|\Lambda|}} \varphi_0 \cdot \varphi_x
\;
dP_{g,\nu,\Lambda}(\varphi),
\end{equation}
assuming the limit exists.
It represents the sum over all $x$ of the correlation of the spin at $0$
with the spin at $x$. If $\nu$ is above the critical point $\nu_c$ then
correlations remain summable, but there is divergence at $\nu=\nu_c$.
The predicted behaviour of the susceptibility,
as $t =\nu - \nu_c \downarrow 0$, is
\begin{equation}
\chi(g,\nu) \sim A_{g,n} \frac{1}{t^{ \gamma}}
,
\end{equation}
with a universal critical exponent $\gamma$ (depending on $d,n$, but not $g$), and
with a logarithmic correction for $d=4$. It was proven in 1982 that $\gamma=1$
for $d>4$ \cite{Aize82,Froh82}.
The concept of universality was discussed in
Section~\ref{sec:SAW}\ref{sec:SSAW}.
It is predicted that the 1-component $|\varphi|^4$ model is in the
same universality class as the Ising model, and that more generally the
$n$-component $|\varphi|^4$ model lies in the universality class of the
model in which the single spin distribution $e^{-V(\varphi_x)}d\varphi_x$
is replaced by the uniform distribution on the sphere
of radius $\sqrt{n}$ in $\Rbold^n$.
In addition, if the nearest-neighbour interaction
given by the Laplacian is replaced by any other finite-range interaction
respecting the symmetries of $\Zbold^d$, then the resulting model is predicted
to be in the
same universality class as the nearest-neighbour model.
The following theorem from \cite{BBS-phi4-log} determines the
asymptotic form of the susceptibility for $d=4$.
Its proof is via a rigorous renormalization group method.
It and related work \cite{ST-phi4,BSTW-clp} give
extensions of mathematical work from the 1980s \cite{GK85,HT87,FMRS87}.
(A caveat for Theorems~\ref{thm:phi4-log}--\ref{thm:phi4lr}
is that the susceptibility is defined with the
infinite volume limit taken through a sequence of tori of period $L^N$
with fixed large $L$, as $N \to \infty$.)
\begin{theorem}
\label{thm:phi4-log}
For $d=4$, $n \ge 1$, small $g>0$, as $t=\nu-\nu_c \downarrow 0$,
\begin{align*}
\chi(g,\nu) &\sim A_{g,n} \frac{1}{t}|\log t|^{\frac{n+2}{n+8}}.
\end{align*}
\end{theorem}
To go below the upper critical dimension, we again consider a long-range
version of the model, by replacing the Laplacian term
$\varphi_x\cdot (-\Delta \varphi)_x$ in \eqref{e:phi4} by a term
$\varphi_x\cdot ((-\Delta)^{\alpha/2} \varphi)_x$ with
fractional Laplacian and $\alpha \in (0,2)$.
The upper critical dimension is again $d_c=2\alpha$.
Several rigorous results use the lace expansion to
prove mean-field behaviour for various long-range models when $d>d_c$,
e.g., \cite{HHS08} for the Ising model.
The following theorem from \cite{Slad18}
concerns dimensions $d=d_c-\epsilon$ which lie slightly below $d_c=2\alpha$.
Related results are proved in \cite{LSW17}, and earlier mathematical papers for
long-range models are \cite{BMS03,Abde07,ACG13}.
\begin{theorem}
\label{thm:phi4lr}
Let $d=1,2,3$.
For $n \ge 1$, for small $\epsilon>0$, for $\alpha = \frac 12 (d+\epsilon)$,
and for $g\in [c\epsilon,c'\epsilon]$ for some $c<c'$,
there is a constant $C$ such that
as $t=\nu-\nu_c \downarrow 0$,
\begin{align*}
&
C^{-1}
\frac{ 1}{t^{ 1+ \frac{n+2}{n+8} \frac\epsilon\alpha -C\epsilon^2}}
\le
\chi(g,\nu) \le
C\frac{1}{t^{ 1+ \frac{n+2}{n+8} \frac\epsilon\alpha +C\epsilon^2}},
\quad i.e.,\;
{ \gamma = 1 + \frac{n+2}{n+8} \frac\epsilon\alpha + O(\epsilon^2)}.
\end{align*}
\end{theorem}
\section{Supersymmetry and $n=0$}
\label{sec:n0}
Comparison of Theorems~\ref{thm:phi4-log}--\ref{thm:phi4lr} with
Theorems~\ref{thm:saw4-log} and \ref{thm:WSAWlr} reveals that when
the $|\varphi|^4$ theorems have the number $n$ of components replaced
by $n=0$ then the WSAW statements result.
This apparently curious observation is not an accident.
Indeed, de Gennes argued in 1972 \cite{Genn72}
that the
self-avoiding walk model \emph{is} the $n=0$ version of the
$n$-component spin model.
Roughly speaking, his reasoning was that
the susceptibility of an $n$-component spin model
has a geometric representation involving a self-avoiding walk and loops, with
the loops weighted by $n$. When $n$ is set equal to zero, only the SAW remains.
This has been a very productive observation in physics. For example, once
it has been established that the $|\varphi|^4$ exponent is
$\gamma = 1 + \frac{n+2}{n+8} \frac{\epsilon}{\alpha}+\cdots$,
then the inference is made that the SAW exponent is $\gamma =
1 + \frac{1}{4}\frac{\epsilon}{\alpha} +\cdots$.
However, from a mathematical perspective, just as it halted progress to
consider non-integer dimensions $d=4-\epsilon$, it is also
problematic to contemplate
the notion of a $0$-component spin, or
of a limit $n \downarrow 0$ when the dimension $n$ of the spin
is a natural number.
An alternate idea from physics with a similar conclusion to de Gennes's
was proposed independently in 1980 by
Parisi and Sourlas \cite{PS80} and by McKane \cite{McKa80}.
Their idea was that while
an $n${-component boson field} $\varphi$ (usual spin)
contributes a factor $n$ for
every loop in the geometric representation of the susceptibility,
an $n$-component \emph{fermion} field contributes $-n$.
When combined, all loops cancel, leaving the self-avoiding walk.
From a mathematical point of view,
this realization of zero components as $n-n$ is less problematic
than setting $n=0$ or considering the limit $n \downarrow 0$,
and it leads to a theorem. Some history of the mathematical work
in this direction can be found in \cite{BBS-brief}. An important
early step was \cite{BM91}, which was inspired by \cite{Lutt83}.
Fermion fields are often defined in terms of Grassmann variables, which
multiply with an
anti-commuting product. A fermion field can also be constructed
using differential forms with their anti-commuting wedge product,
and we follow this route in the following.
Given any finite set $\Lambda$ of cardinality $M=|\Lambda|$, we consider
$2M$ real coordinates and corresponding 1-forms:
\begin{equation}
u_1,v_1, \ldots , u_{M},v_{M}
\quad
\text{and}
\quad
du_1,dv_1,\ldots,du_{M},dv_{M}.
\end{equation}
The wedge product $\wedge$ is associative and anti-commuting, e.g.,
$du_x \wedge dv_y = - dv_y \wedge du_x$.
Let $u=(u_1,\ldots,u_M)$ and similarly for $v$.
A \emph{form} is a function of $(u,v)$ times a product
of 1-forms, or a linear combination of these.
A sum of forms which each contains a product of $p$ distinct 1-forms is
called a $p$-form. Due to the anti-commutativity, $p$-forms are zero if
$p>2M$. Also, any $2M$-form $F$ can be written uniquely as
$F=f(u,v) du_1 \wedge dv_1 \wedge \cdots \wedge du_{M} \wedge dv_{M}$. Integration of a $2M$-form $F$ is defined by
the Lebesgue integral
\begin{equation}
\label{e:intdef}
\int F \;\; =
\int_{\Rbold^{2{M}}} f(u,v) du_1dv_1\cdots du_{M}dv_{M}
,
\end{equation}
and the integral of a $p$-form is defined to be zero if $p<2M$.
This definition of integration extends by linearity to arbitrary forms.
We write the $2M$ real coordinates in terms of $M$ complex coordinates:
\begin{equation}
\phi_x = u_x+iv_x, \quad \bar\phi_x = u_x-iv_x, \quad
d\phi_x = du_x+idv_x, \quad d\bar\phi_x = du_x-idv_x.
\end{equation}
Let $\psi_x = \frac{1}{\sqrt{2\pi i}} d\phi_x$ and
$\bar\psi_x = \frac{1}{\sqrt{2\pi i}} d\bar\phi_x$ .
The field $\phi_x$ is a 2-component \emph{boson} field on $\Lambda$, and $\psi_x$ is
a 2-component \emph{fermion} field.
We define the differential forms
\begin{equation}
\label{e:taudef}
\tau_x = \phi_x\bar\phi_x +
\psi_x \wedge \bar\psi_x,
\quad
\tau_{\Delta,x} = \phi_x(-\Delta \bar\phi)_x
+ \psi_x \wedge (-\Delta \bar\psi)_x.
\end{equation}
Smooth functions of forms are defined by Taylor expansion in $\psi,\bar\psi$,
which terminates as a Taylor polynomial due to the anti-commutativity.
For example,
\begin{equation}
e^{-\sum_{x\in \Lambda} \tau_x} = e^{-\sum_{x\in \Lambda} \phi_x\bar\phi_x}\sum_{m=0}^{M} \frac{(-1)^m}{m!}
\Big(\sum_{x\in \Lambda} \psi_x \wedge \bar\psi_x\Big)^m.
\end{equation}
The susceptibility of the WSAW on a finite subset $\Lambda \subset \Zbold^d$
is then given by the remarkable identity
\begin{equation}
\label{e:intrep}
\chi_\Lambda(g,\nu) =
\sum_{x\in \Lambda}
\int_{\Rbold^{2|\Lambda|}}
e^{-\sum_{z\in\Lambda}(g\tau_z^2 + \nu \tau_z + \tau_{\Delta,z})}\bar\phi_0\phi_x
,
\end{equation}
with the integral on the right-hand side evaluated according to the definition
of the integral as in \eqref{e:intdef} after conversion of the complex coordinates
to real coordinates \cite{BM91}.
The identity \eqref{e:intrep} is discussed in
detail in \cite{BBS-brief}, where a proof is given based on
\emph{supersymmetry}, which is a symmetry
that relates the boson and fermion fields. Replacement of
$-\Delta$ by $(-\Delta)^{\alpha/2}$ in the definition of $\tau_{\Delta,x}$
in \eqref{e:taudef} and \eqref{e:intrep}
gives a corresponding identity for the long-range model.
The right-hand side of \eqref{e:intrep} is reminiscent of the
right-hand side of the definition of the susceptibility in
\eqref{e:phi4susceptibility}. For example, the bosonic part of the exponent
on the right-hand side of \eqref{e:intrep}
matches the exponent on the
right-hand side of the Boltzmann weight \eqref{e:phi4} for the $|\varphi|^4$ model.
The renormalization group method discussed in Section~\ref{sec:RG}
applies equally well
with or without the presence of the fermion field.
This allows a treatment of
WSAW simultaneously with the $n$-component $|\varphi|^4$ model,
as the $n=2-2 = 0$ case, and provides a mathematically
rigorous implementation of de Gennes's idea,
via the supersymmetric formulation introduced by Parisi and Sourlas and by McKane.
\section{Renormalization group (RG) method}
\label{sec:RG}
Theorems~\ref{thm:saw4-log}, \ref{thm:WSAWlr}
and \ref{thm:phi4-log}--\ref{thm:phi4lr} are proved via a rigorous
RG method. Aspects of the RG method are described in this section.
\subsection{RG Strategy}
Scaling limits in critical phenomena have the feature of scale invariance
visible in the long SAW in Figure~\ref{fig:saws}
and in the simulation of the critical Ising model
in Figure~\ref{fig:ising-2D}.
Wilson's brilliant strategy to exploit the scale invariance to simultaneously
explain universality and provide a practical tool for the computation of
universal quantities such as critical exponents
can be outlined schematically as follows:
\begin{enumerate}
\item
Introduce a mapping, the RG map, that maps a model at one scale to a model
at a larger scale.
Scale invariance corresponds to a fixed point of the RG map.
\item
A stable fixed point has a domain of attraction under iteration of the RG map.
The domain of attraction is a universality class of models.
\item
The universal
properties of a scale invariant model can be calculated from
the behaviour of
the RG map in the vicinity of the fixed point.
\end{enumerate}
The ``group'' operation in the term ``renormalization group'' is the
operation of composition of maps. The maps are generally not invertible,
so this is a semigroup with identity rather than a group. The terminology
renormalization ``group'' has nevertheless become commonplace.
The proofs of Theorems~\ref{thm:phi4-log}--\ref{thm:phi4lr}
for the $|\varphi|^4$ model are based on
the above strategy, with some adaptation due to lattice effects.
As discussed in Section~\ref{sec:n0}, the proofs of
Theorems~\ref{thm:saw4-log} and \ref{thm:WSAWlr} for the WSAW require relatively
minor modifications of the proofs for $|\varphi|^4$.
In the remainder of the paper, we flesh out the above strategy
as it is employed in our context.
To focus on the main ideas, we consider only the
long-range $\varphi^4$ model with $n=1$ component in dimensions $d=1,2,3$.
The essential problem is one of a certain Gaussian integration.
\subsection{Multi-scale Gaussian integration}
Let $d=1,2,3$.
Let $\Lambda$ be the discrete $d$-dimensional torus of period $L^N$,
where $L>1$ is a fixed integer.
The infinite-volume limit is achieved by $N \to \infty$.
Let $C$ be a positive-definite $|\Lambda| \times |\Lambda|$ matrix.
The Gaussian expectation $\mathbb{E}_C$ with covariance $C$ of a function
$F:\Rbold^{|\Lambda|} \to \Rbold$ is defined by
\begin{equation}
\mathbb{E}_C F
=
\frac{
\int_{{\mathbb R}^{|\Lambda|}} F(\zeta) e^{-\frac 12 (\zeta,C^{-1}\zeta)}
\prod_{x\in\Lambda} d\zeta_x
}{{\int_{{\mathbb R}^{|\Lambda|}} e^{-\frac 12 (\zeta,C^{-1}\zeta)}
\prod_{x\in\Lambda} d\zeta_x}}.
\end{equation}
Fix $\alpha \in (0,2)$.
Let $m^2>0$ and let $C$ be the
positive-definite $|\Lambda| \times |\Lambda|$ matrix
\begin{equation}
\label{e:Calpha}
C = ((-\Delta_\Lambda)^{\alpha/2} + m^2)^{-1}.
\end{equation}
For $g_0>0$ and $\nu_0 \in \mathbb{R}$, let
\begin{equation}
Z_0=e^{-V_0(\Lambda)},
\quad
V_0(\Lambda)
=
V_0(\Lambda,\varphi)
=
\sum_{x\in\Lambda}
(\tfrac{1}{4}g_0\varphi_x^4 + \tfrac{1}{2}\nu_0 \varphi_x^2 ).
\end{equation}
The essential problem is to compute the convolution of the Gaussian
expectation $\mathbb{E}_C$ with $Z_0$, namely
\begin{equation}
\label{e:ZNdef}
Z_N(\varphi)=\mathbb{E}_CZ_0(\varphi+\zeta) = \mathbb{E}_C e^{-V_0(\Lambda,\varphi+\zeta)},
\end{equation}
uniformly as $m^2 \downarrow 0$ and $N \to \infty$.
For example, it is an exercise in calculus to see that the finite-volume
susceptibility is given by
\begin{equation}
\label{e:chiZN}
\chi_N(g,\nu_0+m^2) = \frac{1}{m^2} + \frac{1}{m^4}\frac{1}{|\Lambda|}
\frac{D^2Z_N(0;\mathbbm{1},\mathbbm{1})}{Z_N(0)},
\end{equation}
where the directions $\mathbbm{1}$ in the directional derivative are the constant
function $\mathbbm{1}_x=1$.
To evaluate \eqref{e:ZNdef}, the Gaussian integration is carried out incrementally, or
\emph{progressively}, with each increment effecting integration over a single
length scale. For this, we use the elementary
property of Gaussian integration that
if $C=C'+C''$ then
\begin{equation}\label{e:progressive2}
\mathbb{E}_C F(\varphi+\zeta) = \mathbb{E}_{C''} \mathbb{E}_{C'} F(\varphi+\zeta''+\zeta'),
\end{equation}
where on the right-hand side the inner Gaussian integral integrates
with respect to $\zeta'$ (holding $\varphi+\zeta''$ fixed), and the outer
Gaussian integral then integrates with respect to $\zeta''$.
The choice of $L^N$ as the period of the torus allows for the partition of
the torus into disjoint $j$-\emph{blocks} of side $L^j$, for $j=0,1,\ldots,N$.
The $0$-blocks are simply the points of $\Lambda$, and the unique $N$-block
is $\Lambda$ itself. In general, the set ${\cal B}_j$ of $j$-blocks
has $L^{(N-j)d}$ elements.
Small values of $j$ are depicted in Figure~\ref{fig:RG_hierarchy1}.
The \emph{scales} $j=0,1,2,\ldots,N$ for the progressive integration
correspond to the block side lengths $L^0,L^1,L^2,\ldots,L^N$.
\begin{figure}[h]
\begin{center}
\scalebox{0.75}
{\input{hier1.pspdftex}}
\end{center}
\caption{\label{fig:reblock}
Some blocks in ${\cal B}_j$ for
$j=0,1,2,3$, with $d=2$ and $L=2$.}
\label{fig:RG_hierarchy1}
\end{figure}
Given a covariance decomposition
\begin{equation}
C=((-\Delta_\Lambda)^{\alpha/2} + m^2)^{-1}=\sum_{j=1}^N C_j,
\end{equation}
it follows from \eqref{e:ZNdef} and \eqref{e:progressive2} that
\begin{equation}
\label{e:progressive}
Z_N(\varphi) =
\mathbb{E}_{C_N+\cdots+C_1}(Z_0(\varphi+\zeta))
= \mathbb{E}_{C_N} \cdots \mathbb{E}_{C_2} \mathbb{E}_{C_1}(Z_0(\varphi + \zeta_N + \cdots \zeta_1)) .
\end{equation}
We use a carefully constructed covariance decomposition, such that
in the corresponding decomposition $\zeta=\zeta_1+\cdots+\zeta_N$ of the field,
the \emph{fluctuation field} $\zeta_j$ captures the fluctuations of the
field $\zeta$ on scale $j-1$. This is quantified by estimates on the covariances
in the decomposition,
which express the fact that a typical Gaussian field with covariance
$C_{j+1}$ is roughly constant on $j$-blocks and has size
of order $L^{-j(d-\alpha)/2}$.
These estimates hold until the \emph{mass scale} $j_m$, which is the
smallest value of $j$ for which $L^{\alpha j}m^2 \ge 1$; for scales $j > j_m$
the covariance is smaller and the integrations for such covariances is subject
to a simpler analysis.
An additional finite-range property
of the covariances plays an important
simplifying role by making the field values in
non-contiguous blocks independent
\cite{BGM04,Baue13a,Mitt16,Slad18}.
In view of \eqref{e:progressive}, we define a sequence iteratively by
\begin{equation}
Z_{j+1}(\varphi)= \mathbb{E}_{C_{j+1}} Z_j (\varphi+\zeta),
\quad\quad Z_0(\varphi)=e^{-V_0(\Lambda,\varphi)}.
\end{equation}
Each step in the sequence performs integration of a fluctuation field
on a single scale.
Then $Z_N$ is the final element of the sequence, and we are interested in
the limit $N \to \infty$. We wish to start the
sequence with $Z_0$ defined in terms of $V_0$ with $\nu_0$ slightly above
the critical value $\nu_c$. However, we do not have a useful
{\it a priori} description of
$\nu_c$; its identification is part of the problem.
To deal with this issue, we enlarge the focus, and consider a Gaussian
convolution as a mapping on a space of functions of the field, defined
on a suitable domain.
In other words, given a covariance $C_+=C_{j+1}$, we write $\mathbb{E}_+=\mathbb{E}_{C_+}$
and define
a scale-dependent map $Z \mapsto Z_+$ by
\begin{equation}\label{e:ZZ}
Z_+(\varphi) = \mathbb{E}_+ Z(\varphi+\zeta),
\end{equation}
for integrable $Z$.
Given a function $F$ of the field, and given a field $\varphi$,
we define a new function
$\theta_\varphi F$ by $(\theta_\varphi F)(\zeta) = F(\varphi+\zeta)$.
Then we can rewrite \eqref{e:ZZ} compactly as
\begin{equation}\label{e:ZZtheta}
Z_+ = \mathbb{E}_+ \theta Z.
\end{equation}
We wish to capture the scale invariance at the critical point as
a ``fixed point'' of the mapping $Z \mapsto Z_{+}$. We do not achieve this
literally, because of lattice effects. Indeed, the mapping is between
\emph{different} spaces, with different norms that implement rescaling.
Nevertheless the notion of a fixed point provides vital guidance.
\subsection{Relevant and irrelevant monomials}
The mapping $Z \mapsto Z_{+}$ is a transformation of one function of the
field to another, and we wish to
identify which are the important aspects of the map to track carefully, and
which parts can be regarded as remainders.
For small $\varphi$, an approximation of $Z(\varphi)$ involves monomials
$\varphi_x^p$. The relative importance of such monomials is assessed
by calculating their size when summed over a block $B \in {\cal B}_j$, when
$\varphi_x$ is a typical Gaussian field for the covariance $C_+$.
For the specific choice $\alpha = \frac 12 (d+\epsilon)$ in the
covariance \eqref{e:Calpha}, this leads to
\begin{equation}
\label{e:monomials}
\sum_{x \in B} \varphi_x^p \approx L^{dj}(L^{-j(d-\alpha)/2})^p
=
\begin{cases}
L^{dj} & (p=0)
\\
L^{\alpha j} & (p=2)
\\
L^{\epsilon j} & (p=4).
\end{cases}
\end{equation}
For powers $p>4$, a negative power of $L^j$ instead occurs, so such monomials
scale down as the scale is advanced. The monomials $1,\varphi^2,\varphi^4$
are said to be \emph{relevant} or \emph{expanding},
while $\varphi^6, \varphi^8, \ldots$
are \emph{irrelevant}. The relevant monomials $\varphi^2,\varphi^4$ appear
already in $V_0$. The monomial $1$ plays a relatively insignificant
role for the analysis of the susceptibility.
Monomials containing
spatial gradients need also to be considered, in general, but for the long-range
model such monomials are irrelevant.
\subsection{Perturbation theory}
With the classification of monomials as relevant or
irrelevant in mind, we treat $Z$ as approximately equal to
$e^{-V(\Lambda)}$ with $V$ given by a local
polynomial $V(\Lambda) = \sum_{x\in \Lambda}(\frac 14 g\varphi_x^4 +
\frac 12 \nu \varphi_x^2 + u)$
with \emph{coupling constants} $g,\nu,u$.
We seek to find $V_+$, defined with new coupling constants
$g_+,\nu_+,u_+$, such that $Z_+$ is well approximated by $e^{-V_+(\Lambda)}$.
Then the map $Z \mapsto Z_+$ is approximately captured by the map $V \mapsto V_+$.
We refer to $V$ as the \emph{perturbative coordinate}.
The term ``perturbation theory'' refers to the evaluation of the map
$V \mapsto V_+$ to some specific order in $V$, together with
the analysis of this approximate map to compute critical exponents. We
consider second-order perturbation theory here.
It is straightforward to compute $\mathbb{E}_+e^{-\theta V(\Lambda)}$ as a formal power
series in $V$ to within an error of order $V^3$.
Details of a way to do this
are laid out in \cite{BBS-rg-pt}. Up to irrelevant
terms, the upshot is that
$\mathbb{E}_+ e^{-\theta V(\Lambda)} \approx e^{-V_{+}(\Lambda)}$ with
$V_+(\Lambda)=\sum_{x\in\Lambda}(g_+\varphi_x^4+\nu_+\varphi_x^2+u_+)$
and with the coupling constants $g_+,\nu_+,u_+$
given by an explicit quadratic polynomial in $g,\nu,u$ with
coefficients determined by the covariance $C_+$.
In order to maintain the approximation of $Z$ by $e^{-V}$ over all scales,
a critical
$m^2$-dependent choice $\nu_0=\nu_0^c(m^2)$ is required.
With the wrong choice, $V$ would grow
exponentially and not remain small as the scale advances.
As $m^2\downarrow 0$, $\nu_0(m^2)$ approaches the critical value $\nu_c$.
If we are able to control the above approximations over all scales,
then we finally arrive at $Z_N(\varphi) \approx e^{-V_N(\Lambda,\varphi)}$.
Substitution of this approximation into the right-hand side of \eqref{e:chiZN}
leads, after a small calculation, to
\begin{equation}\label{e:chinu}
\chi_N(\nu_0^c(m^2)+m^2) \approx \frac{1}{m^2} - \frac{\nu_N(m^2)}{m^4}.
\end{equation}
The proof of Theorem~\ref{thm:phi4lr} in \cite{Slad18}
verifies that the approximation
\eqref{e:chinu} is indeed valid, and that, moreover,
for small $m^2>0$ the following limits hold for general $n$:
\begin{equation}\label{e:nuinfty}
\lim_{N\to\infty}\nu_N(m^2) = O(m^2\epsilon),
\quad
\lim_{N\to\infty} \frac{\partial \nu_N(m^2)}{\partial \nu_0} \Big|_{\nu_0=\nu_0^c(m^2)}
\asymp m^{2\frac{n+2}{n+8}\frac{\epsilon}{\alpha} + O(\epsilon^2)}.
\end{equation}
Together, \eqref{e:chinu}--\eqref{e:nuinfty} imply the differential inequalities
\begin{equation}\label{e:diffineq}
\frac{\partial \chi}{d\nu}
\asymp
- \chi^{2-\frac{n+2}{n+8}\frac{\epsilon}{\alpha} + O(\epsilon^2)},
\end{equation}
and integration then yields the statement of Theorem~\ref{thm:phi4lr}.
Thus the proof of Theorem~\ref{thm:phi4lr} reduces to the validation
of the approximation \eqref{e:chinu} with careful choice of $\nu_0^c(m^2)$,
and the computation of the limits in \eqref{e:nuinfty}.
The first of these two problems is significantly more difficult than the
second.
A change of variables is helpful to understand the flow
of coupling constants under the RG map.
To incorporate the effect of the growth of relevant monomials
in \eqref{e:monomials}, it is natural
to rescale the coupling constants at scale $j$ as $\hat{g}_j=L^{\epsilon j}g_j$ and
$\hat{\nu}_j=L^{\alpha j}\nu_j$.
A further explicit change of variables $(\hat g, \hat \nu) \mapsto (s,\mu)$
creates a simpler triangular system.
In terms of the new variables, the map $V\mapsto V_+$ is described by
\begin{align}
\label{e:sflow}
s_{+}& = L^\epsilon s(1-\beta s) ,
\\
\label{e:muflow}
\mu_{+}
& =
L^\alpha \left( 1-{ \frac{n+2}{n+8}}\beta s\right) \mu + \cdots,
\end{align}
with a remainder that does not play an important role. The coefficient $\beta$ is
given in terms of the accumulated covariance $w_{k} = \sum_{i=1}^k C_i$ by
\begin{equation}
\beta =\beta_j (m^2)
=
(n+8)L^{-\epsilon j} \sum_{x \in \Lambda}(w_{j+1;0,x}^2-w_{j;0,x}^2).
\end{equation}
Properties of the covariance decomposition imply that, for $m^2=0$, the limit
$a=\lim_{j\to\infty}\beta_j(0)$ exists; this permits $\beta$ to be replaced by $a$
in \eqref{e:sflow}--\eqref{e:muflow} up to a controlled error.
The equation $s = L^\epsilon s(1-a s)$ has two fixed points: an unstable
fixed point $s=0$ and a stable fixed point $\bar s = \frac{1}{a}(1-L^{-\epsilon})$
which is order $\epsilon$.
The perturbative equations \eqref{e:sflow}--\eqref{e:muflow} can be analysed
to conclude that if $s$ is initially close to $\bar s$
then there is an initial choice of $\mu$, which determines $\nu_0^c(m^2)$, from
which equations \eqref{e:sflow}--\eqref{e:muflow} can be iterated indefinitely
and \eqref{e:nuinfty} holds. The requirement that $s$ be chosen close
to $\bar s$ is a requirement to be initially near the stable fixed point,
and is responsible for the restriction on $g$ in
Theorems~\ref{thm:WSAWlr} and \ref{thm:phi4lr}.
The above analysis is based on the supposition that the approximation
$\mathbb{E}_+ e^{-V(\Lambda,\varphi +\zeta)} \approx e^{-V_{+}(\Lambda,\varphi)}$ remains
valid over all scales and over the entire volume $\Lambda$.
This approximation has uncontrolled nonperturbative
errors as
the volume parameter $N$ goes to infinity or as the field $\varphi$ becomes
large.
\subsection{Nonperturbative RG coordinate}
The perturbative coordinate $V$ is supplemented by
a nonperturbative coordinate $K$ which controls all errors in the
above approximations. A description of $K$ requires the
introduction of the following concepts.
We fix a scale $j$ which we drop from the notation; scale $j+1$ is denoted by
$+$.
A \emph{polymer} is a union (possibly empty) of blocks from ${\cal B}$. We write ${\cal P}$
for the set of polymers, and ${\cal B}(X)$ and ${\cal P}(X)$ for the
sets of blocks and polymers contained in the polymer $X\in {\cal P}$.
Let ${\cal N}$ denote the algebra of smooth
functions of the field $\varphi$. We consider maps $F:{\cal P}\to{\cal N}$,
e.g., $F(X)=e^{-V(X)}$ with $V(X,\varphi) =\sum_{x\in X}V(\varphi_x)$.
Given $F, G: {\cal P} \to {\cal N}$,
we define the \emph{circle product} $F \circ G: {\cal P} \to {\cal N}$ by
\begin{equation}
\label{e:FcircG}
(F \circ G)(X) = \sum_{Y \in {\cal P}(X)} F(Y)G(X\setminus Y)
\quad\quad (X \in {\cal P}).
\end{equation}
The circle product depends on the scale, since ${\cal P}$
does. It is commutative and
associative, with unit $\mathbbm{1}$ which takes the value $1$ on the empty polymer
and the value $0$ on any nonempty polymer.
We say that $F: {\cal P} \to {\cal N}$
\emph{factorizes over blocks} if $F(X) = \prod_{B\in {\cal B}(X)} F(B)$ for all
$X \in {\cal P}$, e.g., $F(X)=e^{-V(X)}$ factorizes over blocks.
If $F$ and $G$ both factorize over blocks then
\begin{equation}
\label{e:binom}
(F \circ G)(X)
= \prod_{B\in{\cal B}(X)} (F(B)+G(B))
\quad\quad (X \in {\cal P}),
\end{equation}
since in this case expansion of the product on the right-hand side produces
the sum in
\eqref{e:FcircG}.
Instead of the approximation $Z(\Lambda) \approx e^{-V(\Lambda)}$ used in perturbation theory,
we use an exact formula
\begin{equation}
\label{e:ZIcircK}
Z(\Lambda) = e^{-u|\Lambda|}(I\circ K)(\Lambda).
\end{equation}
Here $I=I(V)$ factorizes over blocks;
it may be regarded for present purposes
as $I(X) = e^{-V(X)}$ but in fact an additional term must be included.
The $K$ appearing on the right-hand side of \eqref{e:ZIcircK}
is
a nonperturbative quantity which encapsulates all errors in perturbation
theory, much as the Taylor remainder formula expresses the error in
a Taylor approximation.
Initially, at scale $0$ we have $Z_0(\Lambda) = e^{-V_0(\Lambda)}
= (e^{-V_0}\circ \mathbbm{1})(\Lambda)$, so \eqref{e:ZIcircK} holds with $K_0=\mathbbm{1}$.
We seek to preserve the form
of $Z$ after the Gaussian expectation:
\begin{equation}
\label{e:ZZ+}
Z_+(\Lambda)
= \mathbb{E}_+\theta Z(\Lambda)
= e^{-u_+|\Lambda|}(I_+\circ K_+)(\Lambda)
,
\end{equation}
with a scale-$(j+1)$ circle product, and with
the operator $\theta$ as in \eqref{e:ZZtheta}.
The choice of $I_+$ is determined by perturbation theory.
Given \emph{any} choice of $I_+$, there is a $K_+$ such that
\eqref{e:ZZ+} holds. In fact, there are many, as the representation
$Z_+(\Lambda)= e^{-u_+|\Lambda|}(I_+\circ K_+)(\Lambda)$ does not uniquely
determine $K_+$.
The following proposition is a prototype for an effective choice of $K_+$.
For its statement,
the \emph{closure} of a polymer $X\in {\cal P}$ is defined to be
the smallest polymer
$\overline X \in {\cal P}_+$ such that $X \subset \overline X$.
\begin{prop}
\label{prop:reblock}
Suppose that $I,I_+$ factorize over blocks $B \in {\cal B}_j$.
For $X \in {\cal P}_j$, let $\delta I (X)= \prod_{B \in {\cal B}(X)}
(\theta I (B) - I_+(B))$.
Then
\begin{equation}
\label{e:EIK1}
\mathbb{E}_{+}\theta (I \circ K)(\Lambda)
= ( I_+ \circ \tilde{K}_+)(\Lambda)
\end{equation}
with
\begin{equation}
\label{e:EIK2}
\tilde{K}_+ (U)
=
\sum_{X \in{\cal P}(U)}
I_+(U\setminus X)
\mathbb{E}_{+} \big(
(\delta I
\circ
\theta K
)(X) \big)
\mathbbm{1}_{\overline X =U}
\quad\quad(U \in {\cal P}_{+}).
\end{equation}
\end{prop}
\begin{proof}
By hypothesis, and by \eqref{e:binom}
at scale $j$ with $F=I_+$ and $G=\delta I$,
\begin{equation}
\label{e:chvar1}
\theta(I\circ K)
= (I_+ \circ \delta I) \circ \theta K
=
I_+ \circ (\delta I \circ \theta K).
\end{equation}
Let $J = \delta I \circ \theta K$.
Since $I_+$ does not depend on the integration variable (which is introduced only by the
operation $\theta$),
\begin{align}
\mathbb{E}_{+}\theta
(
I \circ K )(\Lambda)
& =
(
I_+ \circ \mathbb{E}_{+} J )(\Lambda)
=
\sum_{X \in {\cal P}}
I_+(\Lambda \setminus X)
\mathbb{E}_{+}\big(J(X) \big)
.
\end{align}
We reorganize the sum over $X$ by first summing over polymers $U \in {\cal P}_+$
and then summing over all $X\in {\cal P}$ with closure $\overline X = U$. This gives
\begin{align}
\mathbb{E}_{+}\theta
(
I \circ K )(\Lambda)
& =
\sum_{U \in {\cal P}_{+}}
I_+(\Lambda \setminus U)
\sum_{X \in {\cal P}(U) }
I_+(U \setminus X)
\mathbb{E}_{+}\big(J(X) \big)
\mathbbm{1}_{\overline X =U}
.
\end{align}
The right-hand side is \eqref{e:EIK1} with $\tilde{K}_+$ given by \eqref{e:EIK2}, and
the proof is complete.
\end{proof}
The nonperturbative
coordinate $K$ must have two features: it must be $O(V^3)$,
and it must contract as the scale advances. Each of these demands requires
a norm on $K:{\cal P}\to {\cal N}$; we do not describe the delicate
choice of norm here \cite{BS-rg-step}.
The $\tilde K_+$ produced by Proposition~\ref{prop:reblock} is a start, but it is
insufficient as it can be shown to be $O(V^2)$ rather than $O(V^3)$,
and neither is it contractive.
Delicate adjustments are required to achieve these two
goals \cite{BS-rg-step}.
On the other hand, $\tilde K_+$ does preserve a good
factorization property.
We say that polymers
$X,Y\in {\cal P}_j$ are \emph{disconnected} if they are separated
by distance at least $L^j$.
A polymer is \emph{connected} if it is
not the union of two disconnected polymers,
and any polymer $X$ partitions into connected components
${\rm Comp}(X)$ which are separated by distance at least $L^j$.
We say that $F: {\cal P} \to {\cal N}$
\emph{factorizes over connected components}
if $F(X) = \prod_{Y \in {\rm Comp}(X)} F(Y)$.
The finite-range property of the covariance decomposition is the statement
that $C_{j;xy}=0$ if $\|x-y\|_1 \ge \frac 12 L^j$.
This ensures that
\begin{equation} \label{e:factor}
\mathbb{E}_{+}\big(F(X)G(Y)\big) = \mathbb{E}_{+}\big(F(X)\big)\mathbb{E}_{+}\big(G(Y)\big)
\quad
\text{if $X,Y \in {\cal P}_+$ are disconnected,}
\end{equation}
because uncorrelated Gaussian random variables are independent.
Suppose that $K$ factorizes over connected components at scale $j$.
It can be verified that $\tilde{K}_+$ then factorizes over connected components
at scale $j+1$, using \eqref{e:factor}. This factorization
functions in parallel with the norm, which
has the property that the norm of a product is at most the product of the norms.
The geometry of the identity \eqref{e:EIK2} defining $\tilde{K}_+(U)$ is illustrated in Figure~\ref{reblock-euc}, which is helpful for the verification of factorization.
\begin{figure}[h]
\begin{center}
\includegraphics[scale = 0.18]{reblock}
\end{center}
\caption{\label{reblock-euc}
The five large shaded blocks represent $U$, which is the closure of the
polymer $X$ consisting of the four small dark blocks (the support of $\delta I$) and the
small shaded rectangle (the support of $K$).}
\end{figure}
\subsection{RG map and phase portrait}
The RG map is a scale-dependent map
\begin{equation}
\label{e:RGmap}
{\rm RG} :(s,\mu , K)
\mapsto
(s_{+} , \mu_{+} , K_{+}),
\end{equation}
defined on a suitable domain. It is defined in such a way that $K_{+}$ is
third order in $(s,\mu)$ if $K$ is, and the $K$ component is
contractive under change of scale.
The values of $(s_{+},\mu_{+})$ depend on $K$ as well as
$(s,\mu)$, and this dependence is engineered to remove the
relevant parts from $K$. This extraction is responsible
for the contraction of $K$ under change of scale, and is
indispensable
for the iteration of the RG map over all scales. As long as $K$ is
third order, its effect on the flow of the coupling
constants does not change the second order perturbation theory that determines
the asymptotic
behaviour of $\nu_N$ and the critical exponent $\gamma$ for the susceptibility.
The RG map is used to define a map $T$
on a space of sequences $(s_j,\mu_j,K_j)_{j \ge 0}$, such that a fixed point of
$T$ corresponds to a sequence which provides a recursive solution to
\eqref{e:RGmap} for all scales $j$. The $j=0$ value
of this \emph{global RG flow} identifies the
critical point $\nu_c$. The RG flow is depicted schematically
by the phase portrait shown in Figure~\ref{fig:phase}.
For the long-range model with $\alpha=\frac 12(d+\epsilon)$ in dimensions $d=1,2,3$,
the non-Gaussian Wilson--Fisher hyperbolic
fixed point is stable and the Gaussian fixed
point is unstable. The critical point lies on the stable manifold from
which the flow converges to the non-Gaussian fixed point. The 1-dimensional
unstable manifold reflects the growth of $\mu$ for a non-critical choice of
initial condition.
For the 4-dimensional nearest-neighbour model, the two fixed points merge
into a single stable non-hyperbolic Gaussian fixed point.
\begin{figure}[h]
\begin{center}
\includegraphics[scale = 0.405]{WF-fixed-point.pdf}
\end{center}
\caption{\label{fig:phase}
Schematic phase portrait of the dynamical system.}
\end{figure}
\section{Conclusion}
The creation of a comprehensive
theory of phase transitions and critical phenomena is one of
the great achievements of theoretical physics during the second half of the
last century.
The mathematical problems posed by that theory remain a very active topic of
current research. Wilson's RG approach
is a cornerstone of the physical theory. Mathematical theorems based on
the RG approach began to appear decades ago, but a great deal remains to
be done to provide a complete and nonperturbative understanding
of critical phenomena, without
uncontrolled approximations.
This paper concerns some recent contributions in this direction,
for the weakly self-avoiding walk and the $|\varphi|^4$ lattice spin model,
including a rigorous version of the $\epsilon$-expansion.
\section*{Acknowledgements}
I am grateful to David Brydges and Roland Bauerschmidt for all I
have learned from them
during our long collaborations; their influence
runs throughout the paper.
I thank both of them for helpful suggestions,
and Nathan Clisby for providing
Figures~\ref{fig:saws}--\ref{fig:long-range-srw}.
Figures~\ref{fig:ising-arrows}--\ref{reblock-euc}
are taken from \cite{BBS-brief}.
|
\section{The Importance and Impact of Binary Supermassive Black Hole Observations}\label{sec:intro}
The ngVLA could be a powerful probe of dual and binary supermassive black holes (SMBHs), performing detailed studies of their population and evolution, and enabling powerful multi-messenger science when combined with gravitational-wave detection. \fixme{Mention long baselines here?}
Dual (\mbox{$\lesssim$ 10\,kpc separation)} and binary (\mbox{$\lesssim$ 10\,pc} separation) SMBH systems can form during major galaxy mergers. They form and evolve in extended interactions with their environment. At small orbits, their in-spiral is controlled by the emission of some of the brightest gravitational waves in the Universe \citep[Fig.\,\ref{fig:lifecycle}; also][]{begelman80}. When one or both SMBHs power an active galactic nucleus, multi-wavelength emission can directly mark the presence and locations of the SMBHs. Jets produced by these active nuclei will generate radio emission. Because jet cores closely trace the location of a SMBH, and because of the extended lifetime of synchrotron jet observability, radio observations both at high resolution and on large scales provide an excellent tool to trace the current and past dynamics of SMBHs in a merging system.
SMBH pairs confidently mark going galaxy mergers and imminent SMBH coalescences, and are therefore a strong probe of redshift-dependent merger rates, occupation fractions of dual SMBHs in galaxies, and post-merger dynamical evolution. If we determine the rate and environments of systems containing two or more widely separated active galactic nuclei (AGN), we can explore merger-induced activity of nuclei and susbequent SMBH growth. These dual AGN probe the critical sub-10 kpc regime in which both star formation and AGN activity peak during galaxy mergers, according to cosmological simulations \citep{blecha13}. Having a sample of dual AGN is key to testing predictions from simulations, such as whether the most luminous AGN are preferentially triggered in mergers and whether there is a time-lag between star formation triggering and AGN triggering in mergers \citep{hopkins12,hopkins14}.
A systematic survey of dual and binary SMBHs also allows a direct prediction of the projected gravitational wave signals in the low and very low frequency regime of gravitational radiation. Both pulsar timing arrays (PTAs) and future space-based laser interferometers have binary SMBHs as a key target. They may detect both ``discrete'' individual targets, and a stochastic background of small-orbit SMBH binaries.
\begin{figure}
\centering
\includegraphics[width=0.90\textwidth,trim=0mm 0mm 0mm 0mm, clip]{figs/spastic.png}
\vspace{-4mm}
\caption{Figure from \citet{GWAWG-paper}, summarizing the binary SMBH life-cycle. A major unknown in binary evolution theory is the efficiency of inspiral from $\sim$10\,pc down to $\sim$0.1\,pc separations, after which the binary can coalesce efficiently due to gravitational waves.
The ngVLA will discover widely separated binaries, and depending on its long-baseline sensitivity, could discover
targets detectable by PTAs.
PTAs can detect supermassive ($>10^8\,{\rm M_\odot}$) binaries within $\sim$0.1\,pc separation (second panel in the lower figure).
On rare occasions, PTAs may detect the permanent space-time deformation (gravitational wave memory) caused by a binary's coalescence \citep{favata09}.
LISA-like missions can detect lower-mass binaries, up to $\sim10^7\,{\rm M_\odot}$, in the \lb{weeks or months} leading up to coalescence.
Image credits: Galaxies, Hubble/STSci; 4C37.11, \citet{rodriguez}; Simulation visuals, C. Henze/NASA; Circumbinary accretion disk, C. Cuadra.
}\label{fig:lifecycle}
\vspace{-2mm}
\end{figure}
The ngVLA will also be coming online at a time when PTAs like the North American Nanohertz Observatory for Gravitational Waves (NANOGrav) expects to have already detected a number of discrete binary SMBHs \citep[e.\,g.][]{mingarelli-2mass,kelley+submitted}. LISA is expected to be directly constraining SMBH binary coalescence rates soon after its launch
\fixme{reference here?} \lb{Amaro-Seoane et al. 2017 [http://lanl.arxiv.org/abs/1702.00786]}.
Gravitational-wave and multi-messenger astronomy with SMBHs will thus be a leading pursuit of the ngVLA era. Particularly with long baselines, ngVLA will have the potential to identify the hosts of emitting gravitational-wave systems, enabling broad-scoped science that includes:
\begin{itemize}
\vspace{-2mm}\item Precise orbital tracking of binary SMBHs via parsec-scale jet morphology and core tracking \citep{bansal}.
\vspace{-2mm}\item Use of SMBHs as a standard ruler \citep{dorazioloeb}.
\vspace{-2mm}\item Measurement of merger-induced accretion rates and imaging of small-scale AGN feedback. \fixme{reference here?}
\lb{e.g., Muller-Sanchez et al. 2015 [http://adsabs.harvard.edu/abs/2015ApJ...813..103M]}
\vspace{-2mm}\item Detailed multi-wavelength and multi-messenger studies of circumbinary disk evolution \lb{and binary inspiral timescales} \citep[e.\,g.][]{kelley+17}.
\vspace{-2mm}\item Calibration of MBH-host relations up to moderate/high redshifts via direct SMBH binary mass measurements (host identification breaks the distance/mass degeneracy encountered in PTA discrete source detection).
\end{itemize}
\section{Identifying Paired SMBHs with Radio Emission}\label{sec:search}
The ngVLA could be used to conduct surveys for SMBH pairs, follow up candidates identified in other surveys, or both. Either way, in order to reveal dual or binary SMBH candidates with the ngVLA, a few identification methods are available. ngVLA can identify pairs through studies of morphology, spectra, and/or time-dependence of the bulk relativistic flow of particles from the SMBHs.
\begin{figure}
\centering
\includegraphics[width=0.99\textwidth,trim=0mm 0mm 0mm 0mm, clip]{figs/vertical_fig.png}
\vspace{-2mm}
\caption{The observed population of candidate AGN pairs to date compared with the resolution of the ngVLA. Each point represents a dual \hbox{AGN}, with dots, X's, and circles marking optical/near-IR \citep{liu+11}, radio, and X-ray discoveries, respectively. Approximate lines for critical stages in binary formation and evolution are marked as horizontal lines.
The red, green, and purple curves indicate the resolution limit of 10\,GHz, 50\,GHz, and 120\,GHz center observing bands, respectively; for each observing set-up, one cannot resolve a binary orbit below that line. The dashed curves show the resolution limit of 150\,km baselines, while the solid curves show a nominal 1000\,km extended-baseline array. The ngVLA with a VLBI expansion can resolve, and hence directly image and identify, double supermassive black holes at sub-10\,pc separations. Longer baselines and higher frequencies have the potential to resolve multi-messenger SMBH binaries in the nearby Universe (indicated by the curves that cross the dark yellow region).}
\label{fig:vlares}
\end{figure}
\subsection{Direct Core Imaging}\label{sec:spectra}
Radio imaging of multiple self-absorbed synchrotron cores in the $\sim$1--50\,GHz frequency range is one of the most direct routes to identify dual and binary SMBHs. It is exceedingly rare to find bright multiple flat-spectrum objects within close proximity \citep[e.\,g.][]{rodriguez,radiocensus}. The radio ``core'' that is seen in SMBHs represents the shock front near the base of the radio jet, which is compact and in a region of considerable optical depth. These can show a flat spectrum (with spectral index $\alpha\gtrsim-0.5$, where $S\propto f^\alpha$) up to moderate GHz frequencies \citep[e.\,g.][]{blandford+79}. This technique relies on detecting more than one flat-spectrum, compact radio source within a common merger envelope \lb{[host?]}. Confident confirmation of the nature of such detections requires modelling of broad-band continuum spectra, high-resolution morphological modelling, and ideally long-term studies of proper motion. For instance, in the object 4C37.11 (Fig.\,\ref{fig:lifecycle}), there are two cores at a projected separation of 7\,pc that remain compact down to sub-mas resolutions, with larger-scale diffuse emission, that support the interpretation of the two knots as distinct gravitationally interacting, SMBHs. Long-term astrometric observations of this object have led to the first tentative ``orbital tracking'' for a resolved binary, although its orbital period is fairly long at $\sim$$10^4$\,yr \citep{bansal}.
Figure~\ref{fig:vlares} shows a census of known and candidate resolved (imaged) dual SMBHs. Uniquely, radio identifications have probed separations within the radius at which stellar cores are expected to merge \citep{merrittLRR}. Systems within this radius are \emph{direct precursors} to the gravitational-wave emitters detectable by PTAs and space-based laser interferometers, so are of utmost interest to the target science outlined in Section\,\ref{sec:intro}. There is only one confident binary in this range \citep{rodriguez}, and only a scant list of candidates \citep{deane,kharb+17}.
Binary SMBHs are relatively rare, and performing a complete blind search for dual radio cores requires relatively high dynamic range and thorough $u, v$ coverage at high resolution. Present facilities have aspects of these but not all. While the VLA has excellent sensitivity and $u, v$ coverage, it lacks sufficient resolution to discover close pairs.
While current long-baseline facilities have excellent resolution, they lack the snapshot sensitivity and $u, v$ coverage to thoroughly probe cores. The ngVLA has a unique potential to reach a balance between these factors (Bansal et al.\ in prep).
\subsection{Jet Morphology}\label{sec:shapes}
Identification of dual AGN via large-scale jet morphology is somewhat fraught because of the difficulty in discerning external vs.\ internal influences on the jet(s).
Dual AGN emitting on long time scales can produce quadruple jets (Fig.~\ref{fig:dualagn}), and
binary interactions can result in precessing jets that produce large-scale S-shaped radio jet morphologies. Parabolic SMBH-SMBH encounters and binary coalescences can rapidly reorient a jet, potentially producing X-shaped morphologies \citep{merrittekers}. Previous large-scale radio surveys (e.\,g., \hbox{NVSS}, FIRST) have had a strong output of candidate binary or post-merger SMBHs identified through their large-scale jet morphologies. Recent surveys of such radio jet morphologies include 87 X-shaped radio jets found with the VLA \citep{roberts18}. However, past studies have revealed that many such morphologies can be described by backflows rather than originating from a dual system \citep{xnotbinary}.
Still, some (as shown in Fig.\,\ref{fig:dualagn}) have been demonstrated as genuine dual AGN. Broad searches for such features require low-surface-brightness sensitivity coupled with resolution. Large-scale periodic knots along radio jets have also been hypothesized to be periodic accretion episodes fueled by a binary, however these are difficult to distinguish from flow instabilities \citep{godfrey_0637-752}.
\begin{figure
\centering
\includegraphics[width=0.33\textwidth,trim=0mm 7mm 0mm 5mm,clip]{figs/3c75.jpg}
\includegraphics[width=0.33\textwidth,trim=0mm 15mm 0mm 10mm,clip]{figs/ngc326.png}~
\includegraphics[width=0.2\textwidth]{figs/roos-93.png}
\vspace{-2ex}
\caption{Three examples of candidate dual AGN identified through their static or time-dependent radio jet morphology (Sections \ref{sec:shapes}, \ref{sec:time}). \textit{Left:} 3C~75, the dual cores of which have a 7~kpc projected separation and exhibits large-scale jet structures that have misaligned but correlated morphology. (Credit: \hbox{NRAO/AUI}; F.~N.~Owen, C.~P.~O'Dea, M.~Inoue, \& J.~Eilek). \textit{Center:} NGC~326, which has an \texttt{X}-shaped structure that inspired its dual nucleus interpretation. The cores were confirmed at other wavelengths, as shown. (Credit: \hbox{NRAO/AUI}; STScI [inset]). \textit{Right:} The helical jet of 1928+738 shows a 2.9-year cycle \citep{roos+93}.}
\label{fig:dualagn}
\vspace{-3mm}
\end{figure}
\subsection{Variability in Morphology or Flux}\label{sec:time}
Some radio quasars show light curves that appear to be periodic or sinusoidal. At high spatial resolutions, some jets have morphological variability: e.\,g.\ 1928+738 has jets that trace out helical patterns as a function of time \citep[Figure\,\ref{fig:dualagn};][]{roos+93}.
A common interpretation of both of these effects is that an accreting, jet-producing black hole has a jet axis whose angle is being modified by a secondary, orbiting SMBH. Confident identification of periodicity in a light curve requires long-term variability monitoring (over more than a few cycles) to ensure that the observed variability is genuinely cyclical, and not simply the detection of a red noise process \citep{vaughan+16,charisi+18}. Thus, light curve variability programs require a long-term time investment for candidate binary identification.
As with the tracking of 4C37.11 \citep{bansal}, in an ideal case we will be able to track the orbital movements of one or two distinct cores. However, making the simple but suitable assumption that binaries are circular and obey Kepler's third law, their approximate largest physical separation scales with binary period as \mbox{$a\propto P^{2/3}$}. This means that as seen in Figure \ref{fig:lifecycle}, the majority of resolvable binaries will have orbital periods well beyond human lifetimes. Tracking the orbital motion will only be possible for long baselines, and for the massive compact systems at low redshifts.
With long-baseline ($\sim$10,000\,km) arrays, such observations will be possible for $10^9 M_\odot$ radio-emitting binaries with few parsec separations out to $z\sim0.1$ on timescales of a few years. This clearly
underscores the importance of a VLBI option for the ngVLA. If ngVLA baselines are limited to 300 km, then the timescales for resolving orbital properties will increase by a factor of 200, essentially making the ngVLA unsuitable for such studies despite its terrific snapshot sensitivity.
\section{Past Observations, and the Limitations of Current Instruments}\label{sec:past}
\subsection{Radio Searches}
In the past, most radio-identified candidate binary SMBHs have mainly been found serendipitously.
A few concerted efforts have attempted to survey directly for binary SMBHs. Here we discuss these, their successes, and limitations.
A number of blind or semi-blind searches have been made for dual radio cores. \citet{radiocensus} searched all available historical data from the Very Long Baseline Array with dual-frequency imaging; despite the large sample size ($\sim$3200), this search was limited primarily by sensitivity.
\cite{tremblay16} performed multi-frequency imaging of the flux-limited $\sim$1100 AGN included in the VIPS survey \citep{VIPS}, all of which were pre-selected to be bright and flat-spectrum. This search did not identify any new binary systems from the VIPS sample, and concluded that 15 candidates remained uncomfirmed. Finally, \citet{condon-2mass} performed deep observations of 834 radio-bright 2MASS galaxies, under the hypothesis that the most massive galaxies must have undergone a major merger in their history, therefore might contain a binary or recoiling SMBH. So far from these surveys, only one resolved dual core, 4C37.11, was re-detected. On simple arguments of merger rates, inspiral timescales, and radio AGN luminosity functions, at least tens of SMBH pairs must exist within these samples; thus, it is clear that these past experiments have been limited by a combination of $u, v$ coverage, sensitivity, dynamic range, and perhaps sample size.
The multi-wavelength searches for SMBH systems with large separations, corresponding to early stages of galactic mergers, have so far successfully identified a few dozen dual and offset AGN \citep[][and others]{ngc6240, fu11,koss11, koss16, liu13b, comerford15,mullersanchez15,barrows16}.
Observational techniques used to search for binary (sub-$\sim$10\,pc) systems have so far largely relied on direct imaging using VLBI, optical photometry, and spectroscopic measurements \citep[see][for a review]{Bogdanovic2015}. They have recently been complemented by observations with PTAs. We note that in all cases SMBH binaries have been challenging to identify because of their small separation on the sky, the uncertainties related to the uniqueness of their observational signatures, and the limited baseline of monitoring campaigns.
Photometric surveys, like the Catalina Real-Time Transient Survey, the Palomar Transient Factory and others, have uncovered about 150 SMBH binary candidates \citep[e.g.,][]{valtonen08, Graham2015, charisi16, liu16}. In this approach, the quasi-periodic variability in the lightcurves of monitored quasars is interpreted as a manifestation of binary orbital motion. Because of the finite temporal extent of the surveys, which must record at least several orbital cycles of a candidate binary, most photometrically identified SMBH binary candidates have relatively short orbital periods, of order a few years. However, irregularly sampled, stochastically variable lightcurves of ``normal" quasars, powered by a single SMBH, can be mistaken for periodic sources and can lead to false binary identifications in photometric surveys \citep{vaughan+16}. It is also worth noting that one of the most well-known binary SMBH candidates from a quasar light curve, PG1302, is losing traction. Additional data from the All-Sky Automated Survey for Supernovae (ASAS-SN), extending the light curve to present-day, shows that the periodicity does not persist and disfavors the binary SMBH interpretation \citep{liu18}.
The upper limit placed by PTAs on the gravitational wave background largely rules out the amplitude implied by the $\sim150$ photometric binary candidates, implying that some significant fraction of them are unlikely to be SMBH binaries \citep{sesana17}. While this will lead to a downward revision in the number of photometrically identified SMBH binary candidates, it provides a nice example of the effectiveness of multi-messenger techniques, when they can be combined.
Spectroscopic searches for SMBH binaries have so far identified several dozen candidates. They rely on the detection of a velocity shift in the emission-line spectrum of an SMBH binary that arises as a consequence of the binary orbital motion. This approach is similar to that for detection of single- and double-line spectroscopic binary stars, where the lines are expected to oscillate about their local rest-frame wavelength on the orbital time scale of the system. In the context of the binary model, the spectral emission lines are assumed to be associated with gaseous accretion disks that are gravitationally bound to the individual SMBHs \citep{gaskell83,gaskell96,bogdanovic09}. The main complication of this approach is that the velocity-shift signature is not unique to SMBH binaries\lb{; it may also arise from outflowing winds near the SMBH, or even from a gravitational-wave recoil kick} \citep[e.g.,][]{popovic12,Eracleous2012,barth15}. To address this ambiguity, spectroscopic searches have been designed to monitor the offset of broad emission-line profiles over multiple epochs and target sources in which modulations in the offset are consistent with binary orbital motion \citep{bon12, bon16, Eracleous2012, decarli13, Ju2013, liu13, shen13, Runnoe2015, Runnoe2017, li16, wang17}. Two drawbacks of this approach however remain: (a) temporal baselines of spectroscopic campaigns cover only a fraction of the SMBH binary orbital cycle and (b) intrinsic variability of the line profiles may mimic or hide radial velocity variations \citep{Runnoe2017}.
\section{ngVLA Studies and Synergies}\label{sec:ngvla}
\fixme{Without being too detailed about the specific design of the ngVLA, which might change, I'd like to lay out here what ngVLA needs to have to allow the science to get done, and various experiments that could get done with specific capabilities.}
The pair identification techniques outlined above can potentially be realized by a number of specific studies enabled by the ngVLA:
\begin{description}
\item[Blind surveys.] The ngVLA can realize an ability to survey many sources (thousands) with high sensitivity, high dynamic range, and high two-dimensional image fidelity (i.\,e.\ relatively complete $u, v$ coverage) in the frequency range 1--50\,GHz with moderate resolution. These abilities would enable a blind survey for resolved dual cores in the $\gtrsim$few parsec separation regime. \lb{Such a survey could also reveal single, offset cores that may arise from a gravitational-wave recoil kick following a SMBH merger.} Candidates can be cross-matched with future optical/near-IR or X-ray (e.\,g., Athena) surveys to measure spectra, redshifts, and disturbed host galaxy structures. While the multi-wavelength observations might not have the resolution of the \hbox{ngVLA}, they would provide a more complete understanding of a candidate and whether
\lb{an object} is a genuine dual \hbox{AGN}, a gravitational lens,
\lb{a chance projection, or even a recoiling SMBH.}
\item[Targeted surveys to test candidate binaries.] Binary SMBH candidates from a number of upcoming facilities (including LSST and the extremely-large-optical telescopes currently being planned) will be identified at other wavelengths in the coming decades. The ngVLA will be the premier instrument to follow up these candidates in radio, with the goal of understanding their long-term dynamics through large-scale jet morphologies (at low GHz frequencies, given sufficient sensitivity to low surface brightness emission). The ngVLA can also search for resolved dual cores that would give strong evidence to a binary hypothesis. \jc{[Julie: what about using proper motion measurements to really confirm a binary?]}
\item[Core variability.] The ngVLA's higher frequency range lends itself well to observing core-dominated sources, as diffuse radio lobes typically have a relatively steep spectrum and thus preferentially dominate the sky at low radio frequencies. Past work has shown that in blind surveys at $\sim10\,$GHz, radio cores are more dominant \citep{massardi+11}. Thus, the high sensitivity of the ngVLA at 10's of GHz frequencies can permit easier time-resolved radio quasar light curve monitoring, and thus have the potential to identify more examples that might exhibit periodic activity due to variable accretion or orbital precession.
\item[Multi-messenger searches.] Given the rapid progress of gravitational wave (GW) experiments, GW astronomy with binary SMBHs is likely to be a leading pursuit of the ngVLA era. However, GW experiments have poor ability to localize their targets on the sky \citep{ellis13}. In concert with multi-wavelength facilities, the ngVLA can perform efficient follow-up surveys for all galaxies within a given error radius, identifying a galactic host by looking for distinct radio binary signatures as described above. With sufficiently high sensitivity on long baselines, the ngVLA could observe and track the orbital motions of radio cores for the most nearby GW candidates, thus enabling specific multi-messenger studies. \lb{Indeed, by far the loudest GW sources in the PTA band will be massive binaries ($M \ga 10^9 M_{\odot}$) at low redshift ($z \sim 0.1$), making them ideal targets for this analysis (e.g., Kelley et al. 2017a, Mingarelli et al 2017).}
\end{description}
\noindent \lb{The above studies enabled by the ngVLA will provide unique probes of SMBH binary evolution in regimes inaccessible with other methods. This includes crucial data on binaries in the truly gravitationally bound phase, such as their separations, inspiral timescales, luminosities and variability even when not in a rapidly-accreting 'quasar' state. Moreover, studies of radio jet morphology and energetics will provide unprecedented constraints on the mechanisms by which SMBH feedback drives host evolution and contributes to the origin of the observed SMBH-galaxy correlations. In concert with complementary SMBH binary studies by PTAs and LISA,\footnote{See also the chapter on ngVLA/LISA synergies in this volume; Ravi et al.} the ngVLA will be a key cornerstone in the coming era of multimessenger astronomy. The synergy between these electromagnetic and gravitational-wave probes of the binary regime has the potential to unravel the full puzzle of binary SMBH evolution and its close ties to the evolution of galaxies.}
\acknowledgements Part of this research was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration.
SBS is supported by NSF award \#1458952. SBS and TJWL are members of the NANOGrav Physics Frontiers Center, which is supported by NSF award \#1430284. T.B. acknowledges support by the National Aeronautics and Space Administration under Grant No. NNX15AK84G issued through the Astrophysics Theory Program and by the Research Corporation for Science Advancement through a Cottrell Scholar Award. L.B. is supported by NSF award \#1715413.
\section{The Importance and Impact of Binary Supermassive Black Hole Observations}\label{sec:intro}
The ngVLA will be a powerful probe of dual and binary supermassive black holes (SMBHs), performing detailed studies of their population and evolution, and enabling powerful multi-messenger science when combined with gravitational-wave detection. The ngVLA will enable such studies by an unprecedented combination of sensitivity, frequency coverage, and particularly if equipped with long baselines, essentially unparalleled angular resolution.
Dual (\mbox{$\lesssim$ 10\,kpc separation)} and binary (\mbox{$\lesssim$ 10\,pc} separation) SMBH systems can form during major galaxy mergers.
Figure \ref{fig:lifecycle} illustrates the formation and evolution of SMBH systems, from the initial stages of a merger of galaxies to the emission of gravitational waves to an SMBH merger \citep{begelman80}.
When one or both SMBHs power an active galactic nucleus, multi-wavelength emission can directly mark the presence and locations of the SMBHs. Jets produced by these active nuclei will generate radio emission. Because jet cores closely trace the location of a SMBH, and because of the extended lifetime of synchrotron jet observability, radio observations both at high resolution and on large scales provide an excellent tool to trace the current and past dynamics of SMBHs in a merging system.
SMBH pairs confidently mark going galaxy mergers and imminent SMBH coalescences, and are therefore a strong probe of redshift-dependent merger rates, occupation fractions of dual SMBHs in galaxies, and post-merger dynamical evolution. If we determine the rate and environments of systems containing two or more widely separated active galactic nuclei (AGN), we can explore merger-induced activity of nuclei and susbequent SMBH growth. These dual AGN probe the critical sub-10 kpc regime in which both star formation and AGN activity peak during galaxy mergers, according to cosmological simulations \citep{blecha13}. Having a sample of dual AGN is key to testing predictions from simulations, such as whether the most luminous AGN are preferentially triggered in mergers and whether there is a time-lag between star formation triggering and AGN triggering in mergers \citep{hopkins12,hopkins14}.
A systematic survey of dual and binary SMBHs also allows a direct prediction of the projected gravitational wave signals in the low and very low frequency regime of gravitational radiation. Both pulsar timing arrays (PTAs) and future space-based laser interferometers have binary SMBHs as a key target. They may detect both ``discrete'' individual targets, and a stochastic background of small-orbit SMBH binaries.
\begin{figure}
\centering
\includegraphics[width=0.90\textwidth,trim=0mm 0mm 0mm 0mm, clip]{figs/spastic.png}
\vspace{-4mm}
\caption{Figure from \citet{GWAWG-paper}, summarizing the binary SMBH life-cycle. A major unknown in binary evolution theory is the efficiency of inspiral from $\sim$10\,pc down to $\sim$0.1\,pc separations, after which the binary can coalesce efficiently due to gravitational waves.
The ngVLA will discover widely separated binaries, and depending on its long-baseline sensitivity, could discover
targets detectable by pulsar timing arrays (PTAs).
PTAs such as NANOGrav can detect supermassive ($>10^8\,{\rm M_\odot}$) binaries within $\sim$0.1\,pc separation (second panel in the lower figure).
On rare occasion, PTAs may detect the permanent space-time deformation (GW memory) caused by a binary's coalescence \citep{favata09}.
The ngVLA's studies will complement those of LISA, which will detect lower-mass binaries, up to $\sim10^7\,{\rm M_\odot}$, in the weeks or months
leading up to coalescence.
Image credits: Galaxies, Hubble/STSci; 4C37.11, \citet{rodriguez}; Simulation visuals, C. Henze/NASA; Circumbinary accretion disk, C. Cuadra.
}\label{fig:lifecycle}
\vspace{-5mm}
\end{figure}
The ngVLA will also be coming online at a time when PTAs like the North American Nanohertz Observatory for Gravitational Waves (NANOGrav) expects to have already detected a number of discrete binary SMBHs \citep[e.\,g.][]{mingarelli-2mass,kelley+submitted}. LISA is expected to be directly constraining SMBH binary coalescence rates soon after its launch \citep{lisa-science-2,as+17}
Gravitational-wave and multi-messenger astronomy with SMBHs will thus be a leading pursuit of the ngVLA era. Particularly with long baselines, ngVLA will have the potential to identify the hosts of emitting gravitational-wave systems, enabling broad-scoped science that includes:
\begin{itemize}
\vspace{-2mm}\item Precise orbital tracking of binary SMBHs via parsec-scale jet morphology and core tracking \citep{bansal}.
\vspace{-2mm}\item Use of SMBHs as a standard ruler \citep{dorazioloeb}.
\vspace{-2mm}\item Measurement of merger-induced accretion rates and imaging of small-scale AGN feedback \citep{mullersanchez15}
\vspace{-2mm}\item Detailed multi-wavelength and multi-messenger studies of circumbinary disk evolution and binary inspiral timescales \citep[e.\,g.][]{kelley+17}.
\vspace{-2mm}\item Calibration of MBH-host relations up to moderate/high redshifts via direct SMBHB mass measurements (host identification breaks the distance/mass degeneracy encountered in PTA discrete source detection).
\end{itemize}
In the remainder of this chapter, we describe the various approaches that can be used to identify SMBH pairs with the ngVLA \S\ref{sec:search}, compare those with past searches in \S\ref{sec:past}, and discuss synergies with multi-wavelength observations in \S\ref{sec:synergies}. Finally, \S\ref{sec:ngvla} lays out several example ngVLA studies and their general requirements for success.
\section{Identifying Paired SMBHs with Radio Emission}\label{sec:search}
The ngVLA could be used to conduct surveys for SMBH pairs, follow up candidates identified in other surveys, or both. Either way, in order to reveal dual or binary SMBH candidates with the ngVLA, a few identification methods are available. ngVLA can identify pairs through studies of morphology, spectra, and/or time-dependence of the bulk relativistic flow of particles from the SMBHs.
\begin{figure}
\centering
\includegraphics[width=0.99\textwidth,trim=0mm 0mm 0mm 0mm, clip]{figs/vertical_fig.png}
\vspace{-2mm}
\caption{The observed population of candidate AGN pairs to date compared with the resolution of the ngVLA. Each point represents a dual \hbox{AGN}
discovered via radio (X), X-ray (circles), or optical/near-IR (dots; from \citealt{liu+11}).
Approximate lines for critical stages in binary formation and evolution are marked as horizontal lines.
The red, green, and purple curves indicate the resolution limit of 10\,GHz, 50\,GHz, and 120\,GHz center observing bands, respectively; for each observing set-up, one cannot resolve a binary orbit below that line. The dashed curves show the resolution limit of 150\,km baselines, while the solid curves show a nominal 1000\,km extended-baseline array. The ngVLA with a VLBI expansion can resolve, and hence directly image and identify, double supermassive black holes at sub-10\,pc separations. Longer baselines and higher frequencies have the potential to resolve multi-messenger SMBH binaries in the nearby Universe (indicated by the curves that cross the dark yellow region).}
\label{fig:vlares}
\end{figure}
\subsection{Direct Core Imaging}\label{sec:spectra}
Radio imaging of multiple self-absorbed synchrotron cores in the $\sim$1--50\,GHz frequency range is one of the most direct routes to identify dual SMBHs. It is exceedingly rare to find bright multiple flat-spectrum objects within close proximity \citep[e.\,g.][]{rodriguez,radiocensus}. The radio ``core'' that is seen in SMBHs represents the shock front near the base of the radio jet, which is compact and in a region of considerable optical depth. These can show a flat spectrum (with spectral index $\alpha\gtrsim-0.5$, where $S\propto f^\alpha$) up to moderate GHz frequencies \citep[e.\,g.][]{blandford+79}. This technique relies on detecting more than one flat-spectrum, compact radio source within a
common merger host. Confident confirmation of the nature of such detections requires modelling of broad-band continuum spectra, high-resolution morphological modelling, and ideally long-term studies of proper motion. For instance, in the object 4C37.11 (Fig.\,\ref{fig:lifecycle}), there are two cores at a projected separation of 7\,pc that remain compact down to sub-mas resolutions, with larger-scale diffuse emission that support the interpretation of the two knots as distinct, but gravitationally interacting, SMBHs. Long-term astrometric observations of this object have led to the first tentative ``orbital tracking'' for a resolved binary, although its orbital period is fairly long at $\sim$$10^4$\,yr \citep{bansal}.
Figure~\ref{fig:vlares} shows a census of known and candidate resolved (imaged) dual SMBHs. Uniquely, radio identifications have probed separations within the radius at which stellar cores are expected to merge \citep{merrittLRR}. Systems within this radius are \emph{direct precursors} to the gravitational-wave emitters detectable by PTAs and space-based laser interferometers, so are of utmost interest to the target science outlined in Section\,\ref{sec:intro}. There is only one confident binary in this range \citep{rodriguez}, and only a scant list of candidates \citep{deane,kharb+17}.
Binary SMBHs are relatively rare, and performing a complete blind search for dual radio cores requires relatively high dynamic range and thorough $u, v$ coverage at high resolution. Present facilities have aspects of these but not all. While the VLA has excellent sensitivity and $u, v$ coverage, it lacks sufficient resolution to discover close pairs.
While current long-baseline facilities have excellent resolution, they lack the snapshot sensitivity and $u, v$ coverage to thoroughly probe cores \citep[e.\,g.][]{deller+middelberg}. The ngVLA has a unique potential to reach a balance between these factors (Bansal et al.\ in prep).
\subsection{Jet Morphology}\label{sec:shapes}
Identification of dual AGN via large-scale jet morphology is somewhat frought because of the difficulty in discerning external vs.\ internal influences on the jet(s).
Dual AGN emitting on long time scales can produce quadruple jets (Fig.~\ref{fig:dualagn}), and
binary interactions can result in precessing jets that produce large-scale S-shaped radio jet morphologies. Parabolic SMBH-SMBH encounters and binary coalescences can rapidly reorient a jet, potentially producing X-shaped morphologies \citep{merrittekers}. Previous large-scale radio surveys (e.\,g., \hbox{NVSS}, FIRST) have had a strong output of candidate binary or post-merger SMBHs identified through their large-scale jet morphologies. Recent surveys of such radio jet morphologies include 87 X-shaped radio jets found with the VLA \citep{roberts18}. However, past studies have revealed that many such morphologies can be described by backflows rather than originating from a dual system \citep{xnotbinary}.
Still, some (as shown in Fig.\,\ref{fig:dualagn}) have been demonstrated as genuine dual AGN. Broad searches for such features require low-surface-brightness sensitivity coupled with resolution.
Large-scale periodic knots along radio jets have also been hypothesized to be periodic accretion episodes fueled by a binary, however these are difficult to distinguish from flow instabilities \citep{godfrey_0637-752}.
\begin{figure
\centering
\includegraphics[width=0.33\textwidth,trim=0mm 7mm 0mm 5mm,clip]{figs/3c75.jpg}
\includegraphics[width=0.33\textwidth,trim=0mm 15mm 0mm 10mm,clip]{figs/ngc326.png}~
\includegraphics[width=0.2\textwidth]{figs/roos-93.png}
\vspace{-2ex}
\caption{Three examples of candidate dual AGN identified through their static or time-dependent radio jet morphology (Sections \ref{sec:shapes}, \ref{sec:time}). \textit{Left:} 3C~75, the dual cores of which have a 7~kpc projected separation and exhibits large-scale jet structures that have misaligned but correlated morphology. (Credit: \hbox{NRAO/AUI}; F.~N.~Owen, C.~P.~O'Dea, M.~Inoue, \& J.~Eilek). \textit{Center:} NGC~326, which has an \texttt{X}-shaped structure that inspired its dual nucleus interpretation. The cores were confirmed at other wavelengths, as shown. (Credit: \hbox{NRAO/AUI}; STScI [inset]). \textit{Right:} The helical jet of 1928+738 shows a 2.9-year cycle \citep{roos+93}.}
\label{fig:dualagn}
\vspace{-3mm}
\end{figure}
\subsection{Variability in Morphology or Flux}\label{sec:time}
Some radio quasars show light curves that appear to be periodic or sinusoidal. At high spatial resolutions, some jets have morphological variability: e.\,g.\ 1928+738 has jets that trace out helical patterns as a function of time \citep[Figure\,\ref{fig:dualagn};][]{roos+93}.
A common interpretation of both of these effects is that an accreting, jet-producing black hole has a jet axis whose angle is being modified by a secondary, orbiting SMBH. Confident identification of a periodicity in a light curve requires long-term variability monitoring (over more than a few cycles) to ensure that the observed variability is genuinely cyclical, and not simply the detection of a red noise process \citep{vaughan+16,charisi+18}. Thus, light curve variability programs require a long-term time investment for candidate binary identification.
As with the tracking of 4C37.11 \citep{bansal}, in an ideal case we will be able to track the orbital movements of one or two distinct cores. However, making the simple but suitable assumption that binaries are circular and obey Kepler's third law, their approximate largest physical separation scales with binary period as \mbox{$a\propto P^{2/3}$}. This means that as seen in Figure \ref{fig:lifecycle}, the majority of resolvable binaries will have orbital periods well beyond human lifetimes. Tracking the orbital motion will only be possible for long baselines, and for the massive compact systems at low redshifts.
With long-baseline ($\sim$10,000\,km) arrays, such observations will be possible for $10^9 M\odot$ radio-emitting binaries with few parsec separations out to $z\sim0.1$ on timescales of a few years. This clearly
underscores the importance of a VLBI option for the ngVLA. If ngVLA baselines are limited to 300 km, then the timescales for resolving orbital properties will increase by a factor of 200, essentially making the ngVLA unsuitable for such studies despite its terrific snapshot sensitivity.
\subsection{Maser Dynamics}\label{sec:masers}
Long-baseline observations of masers have been instrumental in finding SMBH masses and analyzing accretion disk dynamics (e.g., \citealt{Miyoshi95}, \citealt{Braatz10}, \citealt{Kuo13}, \citealt{Gao16}). Deviations from Keplerian motion in the maser rotation curves can indicate that the potential is not that of a point source (e.g., \citealt{Kuo11}), and they can therefore be sensitive to close SMBH binaries (i.e., orbital separations of roughly $\sim$0.01--0.1\,pc).
For wider SMBH binaries ($\sim$1--100\,pc separation), the maser rotation curve is not expected to be substantially different from Keplerian. In this case, one can compare the barycenter velocity derived from the maser rotation curve (which unambiguously traces the SMBH velocity) to that of the galaxy on large scales ($\sim$10's of kpc, traced using stellar or HI dynamics). Significant differences between these two velocities are then indicative of binary or recoiling SMBH, although a measurement of a velocity offset cannot by itself distinguish between the two possibilities (Pesce et al. 2018, submitted to ApJ).
\section{Current Radio Searches and Limitations}\label{sec:past}
In the past, most radio-identified candidate binary SMBHs have mainly been found serendipitously.
A few concerted efforts have attempted to survey directly for binary SMBHs. Here we discuss these, their successes, and limitations.
A number of blind or semi-blind searches have been made for dual radio cores. \citet{radiocensus} searched all available historical data from the Very Long Baseline Array with dual-frequency imaging; despite the large sample size ($\sim$3200), this search was limited primarily by sensitivity.
The VIPS survey \citep{VIPS} originally precipitated the discovery of 4C37.11 as a binary SMBH \citep{rodriguez}. \cite{tremblay16} performed multi-frequency imaging of the flux-limited $\sim$1100 AGN included in VIPS, all of which were pre-selected to be bright and flat-spectrum. This search did not identify any new binary systems from the VIPS sample, and concluded that 15 candidates remained uncomfirmed. Finally, \citet{condon-2mass} performed deep observations of 834 radio-bright 2MASS galaxies, under the hypothesis that the most massive galaxies must have undergone a major merger in their history, therefore might contain a binary or recoiling SMBH. So far from these surveys, only one resolved dual core, 4C37.11, was re-detected. On simple arguments of merger rates, inspiral timescales, and radio AGN luminosity functions, at least tens of SMBH pairs must exist within these samples; thus, it is clear that these past experiments have been limited by a combination of $u, v$ coverage, sensitivity, dynamic range, and perhaps sample size.
\section{Current Multi-wavelength Synergies and Limitations}\label{sec:synergies}
Dual and binary SMBH signatures have been proposed in a number of wavebands (see, e.\,g., the review in \citealt{sbs-cqg}), and searches pursuing those signatures have revealed a number of binary candidates. We note that in all cases SMBH binaries have been challenging to identify because of their small separation on the sky, the uncertainties related to the uniqueness of their observational signatures, and the limited baseline of monitoring campaigns. Direct radio imaging, however, can provide an excellent follow-up to explore the nature of objects identified with these signatures.
The most prolific binary search method thus far has been the use of photometric and spectroscopic optical signatures to reveal candidate SMBH binaries. We describe that work here given its relevance to upcoming synoptic sky surveys that will be contemporary to ngVLA.
Multi-wavelength searches for SMBH systems with large separations, corresponding to early stages of galactic mergers, have so far successfully identified a few dozen dual and offset AGN \citep[][and others]{ngc6240, fu11,koss11, koss16, liu13b, comerford15, barrows16,mullersanchez15}.
SMBHs with even smaller (parsec and sub-parsec) separations are representative of the later stages of galactic mergers in which the two SMBHs are sufficiently close to form a gravitationally bound pair. Direct imaging using VLBI can confidently demonstrate such a system, however requires deep observations to do so successfully. In this regard, optical surveys can help to identify a sample of SMBH binaries for VLBI followup \citep[see][for a review]{Bogdanovic2015}.
Current studies have been exploring periodic quasar variability as a potential indicator for binary SMBH emissions; these and future surveys of this kind using synoptic instruments like PAN-STARRS and LSST will reveal many candidate binary systems ideal for follow-up with the ngVLA plus a long-baseline extension.
Current photometric surveys
have already uncovered about 150 SMBH binary candidates \citep[e.g.,][]{valtonen08, Graham2015, charisi16, liu16}. In this approach, the quasi-periodic variability in the lightcurves of monitored quasars is interpreted as a manifestation of binary orbital motion. Because of the finite temporal extent of the surveys,
most of these candidates have orbital periods of order a few years, hence would be in the regime of gravitational-radiation-dominated inspiral (Fig.\,\ref{fig:lifecycle}). However, studies of this type suffer the same red-noise issues noted in Sec.\,\ref{sec:time}.
A recent, initially quite promising binary SMBH candidate from a quasar light curve, PG1302, has recently been losing traction. Additional data from the All-Sky Automated Survey for Supernovae (ASAS-SN), extending the light curve to present-day, shows that the periodicity does not persist and disfavors the binary SMBH interpretation \citep{liu18}.
Furthermore, the upper limit placed by PTAs on the gravitational wave background largely rules out the amplitude implied by the $\sim150$ photometric binary candidates, implying that some significant fraction of them are unlikely to be SMBH binaries \citep{sesana17}. While this will lead to a downward revision in the number of photometrically identified SMBH binary candidates, it provides a nice example of the effectiveness of multi-messenger techniques, when they can be combined.
Spectroscopic searches for SMBH binaries have so far identified several dozen candidates. They rely on the detection of a velocity shift in the emission-line spectrum of an SMBH binary that arises as a consequence of the binary orbital motion
\citep{gaskell83,gaskell96,bogdanovic09}. The main complication of this approach is that the velocity-shift signature is not unique to SMBH binaries; it may also arise from outflowing winds near the SMBH, or even from a gravitational-wave recoil kick \citep[e.g.,][]{popovic12,Eracleous2012,barth15}. To address this ambiguity, spectroscopic searches have been designed to monitor the offset of broad emission-line profiles over multiple epochs and target sources in which modulations in the offset are consistent with binary orbital motion \citep{bon12, bon16, Eracleous2012, decarli13, Ju2013, liu13, shen13, Runnoe2015, Runnoe2017, li16, wang17}. Two drawbacks of this approach however remain: (a) temporal baselines of spectroscopic campaigns cover only a fraction of the SMBH binary orbital cycle and (b) intrinsic variability of the line profiles may mimic or hide radial velocity variations, making these measurements more challenging and less deterministic \citep{Runnoe2017}.
\section{Example ngVLA Studies}\label{sec:ngvla}
Dual and binary SMBH studies can be realized by a number of specific studies enabled by the ngVLA:
\begin{description}
\item[Blind surveys.] The ngVLA can survey many sources (thousands) with high sensitivity, high dynamic range, and high two-dimensional image fidelity (i.\,e.\ relatively complete $u, v$ coverage) in the frequency range 1--50\,GHz with moderate resolution. These abilities would enable a blind survey for resolved dual cores in the $\gtrsim$few parsec separation regime. Such a survey could also reveal single, offset cores that may arise from a gravitational-wave recoil kick following a SMBH merger. Candidates can be cross-matched with future optical/near-IR or X-ray (e.\,g., Athena) surveys to measure spectra, redshifts, and disturbed host galaxy structures. While the multi-wavelength observations might not have the resolution of the \hbox{ngVLA}, they would provide a more complete understanding of a candidate and whether it
is a genuine dual \hbox{AGN}, a gravitational lens,
a chance projection, or even a recoiling SMBH.
\item[Targeted surveys to test candidate binaries.]
Many binary SMBH candidates are likely to be identified in the coming decades by numerous instruments (including PTAs, LSST and the extremely large optical telescopes currently being planned). The ngVLA will be the premier instrument to follow up these candidates in radio, with the goal of understanding their long-term dynamics through large-scale jet morphologies (at low GHz frequencies, given sufficient sensitivity to low surface brightness emission). A VLBI component on the ngVLA will enable observation and tracking of resolved dual cores, which will truly confirm a binary and allow in-depth study of SMBH binary evolution.
\item[Core variability.] The frequency range of ngVLA makes it well-suited to observe AGN that are dominated by core emission.
Past work has shown that in blind surveys at $\sim$10\,GHz, radio cores are more dominant than diffuse lobes \citep{massardi+11}. Thus, the high sensitivity of the ngVLA at tens of GHz frequencies can permit easier time-resolved radio quasar light curve monitoring, and thus have the potential to identify more examples that might exhibit periodic activity due to variable accretion or orbital precession.
\item[Multi-messenger searches.]
GW experiments can only poorly localize their detections \citep{ellis13}. In concert with multi-wavelength facilities, the ngVLA could perform efficient follow-up surveys for all galaxies within a given error radius, identifying a host by looking for the signatures described above. With sufficiently high sensitivity on long baselines, the ngVLA could observe and track the orbital motions of radio cores for the most nearby GW candidates, thus enabling specific multi-messenger studies. In particular, PTAs will preferentially detect the most massive ($M_{\rm BH} \ga 10^9$ M$_{\odot}$), nearby ($z \la 0.1$) objects that are accessible to Earth-based VLBI \citep[e.\,g.][]{kelley+17,mingarelli-2mass}.
\item[Long-baseline benefits.]
A VLBI component to the ngVLA has clear benefits for this field of research. A long-baseline extension to the ngVLA, where each long baseline consists of a cluster of $\sim$5 antennas, would yield several major benefits. First, much of the imaging of double black holes is resolution-limited, meaning that going to longer baselines would immediately allow the probing of closer pairs of black holes. Second, long baselines would allow the maser experiments described in Section\,\ref{sec:masers}. Third, the clusters themselves can be used for pulsar timing when the ngVLA core is not being used as part of a VLBI project, giving access to the pulsars too far north for SKA timing.
\end{description}
\noindent The above studies enabled by the ngVLA will provide unique probes of SMBH binary evolution in regimes inaccessible with other methods. This includes crucial data on binaries in the truly gravitationally bound phase, such as information about their separations, inspiral timescales, luminosities and variability even when not in a rapidly-accreting ``quasar" state. Moreover, studies of radio jet morphology and energetics will provide unprecedented constraints on the mechanisms by which SMBH feedback drives host evolution and contributes to the origin of the observed SMBH-galaxy correlations. In concert with complementary SMBH binary studies by PTAs and LISA,\footnote{See also the chapter on ngVLA/LISA synergies in this volume; Ravi et al.} the ngVLA will be a key cornerstone in the coming era of multimessenger astronomy. The synergy between these electromagnetic and gravitational-wave probes of the binary regime has the potential to unravel the full puzzle of binary SMBH evolution and its close ties to the evolution of galaxies.
\acknowledgements Part of this research was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration.
SBS is supported by NSF award \#1458952. SBS and TJWL are members of the NANOGrav Physics Frontiers Center, which is supported by NSF award \#1430284. T.B. acknowledges support by the National Aeronautics and Space Administration under Grant No. NNX15AK84G issued through the Astrophysics Theory Program and by the Research Corporation for Science Advancement through a Cottrell Scholar Award. L.B. is supported by NSF award #1715413.
\section{The Importance and Impact of Binary Supermassive Black Hole Observations}\label{sec:intro}
The ngVLA will be a powerful probe of dual and binary supermassive black holes (SMBHs), performing detailed studies of their population and evolution, and enabling powerful multi-messenger science when combined with gravitational-wave detection. The ngVLA will enable such studies by an unprecedented combination of sensitivity, frequency coverage, and particularly if equipped with long baselines, essentially unparalleled angular resolution.
Dual (\mbox{$\lesssim$ 10\,kpc separation)} and binary (\mbox{$\lesssim$ 10\,pc} separation) SMBH systems can form during major galaxy mergers.
Figure \ref{fig:lifecycle} illustrates the formation and evolution of double SMBH systems, from the initial stages of a merger of galaxies to the emission of gravitational waves to an SMBH merger \citep{begelman80}.
When one or both SMBHs power an active galactic nucleus, multi-wavelength emission can directly mark the presence and locations of the SMBHs. Jets produced by these active nuclei will generate radio emission. Because jet cores closely trace the location of a SMBH, and because of the extended lifetime of synchrotron jet observability, radio observations both at high resolution and of larger-scale jet structures provide excellent tools to trace the current and past dynamics of SMBHs in a merging system.
SMBH pairs confidently mark ongoing galaxy mergers and imminent SMBH coalescences, and are therefore a strong probe of redshift-dependent merger rates, occupation fractions of dual SMBHs in galaxies, and post-merger dynamical evolution. If we determine the rate and environments of systems containing two or more widely separated active galactic nuclei (AGN), we can explore merger-induced activity of nuclei and susbequent SMBH growth. These dual AGN probe the critical regime during which the black holes and their associated galactic-scale stellar cores are virialized (e.\,g.\ within $\sim$10\,kpc black hole separation). During this phase, both star formation and AGN activity peak during galaxy mergers, according to cosmological simulations \citep{blecha13}. Having a sample of dual AGN is key to testing predictions from simulations, such as whether the most luminous AGN are preferentially triggered in mergers and whether there is a time-lag between star formation triggering and AGN triggering in mergers \citep{hopkins12,hopkins14}.
A systematic survey of dual and binary SMBHs also allows a direct prediction of the projected gravitational wave signals in the low and very low frequency regime of gravitational radiation. Both pulsar timing arrays (PTAs) and future space-based laser interferometers have binary SMBHs as a key target. They may detect both ``discrete'' individual targets, and a stochastic background of small-orbit SMBH binaries.
\begin{figure}
\centering
\includegraphics[width=0.90\textwidth,trim=0mm 0mm 0mm 0mm, clip]{figs/spastic.png}
\vspace{-4mm}
\caption{Figure from \citet{GWAWG-paper}, summarizing the binary SMBH life-cycle. A major unknown in binary evolution theory is the efficiency of inspiral from $\sim$10\,pc down to $\sim$0.1\,pc separations, after which the binary can coalesce efficiently due to gravitational waves.
The ngVLA will discover widely separated binaries, and depending on its long-baseline sensitivity, could discover
targets detectable by pulsar timing arrays (PTAs).
PTAs such as NANOGrav can detect supermassive ($>10^8\,{\rm M_\odot}$) binaries within $\sim$0.1\,pc separation (second panel in the lower figure).
On rare occasion, PTAs may detect the permanent space-time deformation (GW memory) caused by a binary's coalescence \citep{favata09}.
LISA will detect lower-mass binaries, up to $\sim10^7\,{\rm M_\odot}$, in the weeks or months
leading up to coalescence.
Image credits: Galaxies, Hubble/STSci; 4C37.11, \citet{rodriguez}; Simulation visuals, C. Henze/NASA; Circumbinary accretion disk, C. Cuadra.
}\label{fig:lifecycle}
\vspace{-5mm}
\end{figure}
The ngVLA will also be coming online at a time when PTAs like the North American Nanohertz Observatory for Gravitational Waves (NANOGrav) expects to have already detected a number of discrete binary SMBHs \citep[e.\,g.][]{mingarelli-2mass,kelley+submitted}. LISA is expected to directly constrain SMBH binary coalescence rates soon after its launch \citep{lisa-science-2,as+17}.
Gravitational-wave and multi-messenger astronomy with SMBHs will thus be a leading pursuit of the ngVLA era. Particularly with long (VLBI) baselines, ngVLA will have the potential to identify the hosts of emitting gravitational-wave systems, enabling broad-scoped science that includes:
\begin{itemize}
\vspace{-2mm}\item Precise orbital tracking of binary SMBHs via parsec-scale jet morphology and core tracking \citep{bansal}.
\vspace{-2mm}\item Use of SMBHs as a standard ruler \citep{dorazioloeb}.
\vspace{-2mm}\item Measurement of merger-induced accretion rates and imaging of small-scale AGN feedback \citep{mullersanchez15}
\vspace{-2mm}\item Detailed multi-wavelength and multi-messenger studies of circumbinary disk evolution and binary inspiral timescales \citep[e.\,g.][]{kelley+17}.
\vspace{-2mm}\item Calibration of SMBH-host relations up to moderate/high redshifts via direct SMBHB mass measurements (host identification breaks the distance/mass degeneracy encountered in PTA discrete source detection).
\end{itemize}
In the remainder of this chapter, we describe the various approaches that can be used to identify SMBH pairs with the ngVLA (Sec.\,\ref{sec:search}), compare those with past searches in Sec.\,\ref{sec:past}, and discuss synergies with multi-wavelength observations in Sec.\,\ref{sec:synergies}. Finally, Sec.\,\ref{sec:ngvla} lays out several example ngVLA studies and their general requirements for success.
\section{Identifying Paired SMBHs with Radio Emission}\label{sec:search}
The ngVLA could be used to conduct surveys for SMBH pairs, follow up candidates identified in other surveys, or both. Either way, in order to reveal dual or binary SMBH candidates with the ngVLA, a few identification methods are available. ngVLA can identify pairs through studies of morphology, spectra, and/or time-dependence of the bulk relativistic flow of particles from the SMBHs.
\begin{figure}
\centering
\includegraphics[width=0.99\textwidth,trim=0mm 0mm 0mm 0mm, clip]{figs/ngvla-colorfig.pdf}
\vspace{-2mm}
\caption{The observed population of candidate AGN pairs to date compared with the resolution of the ngVLA. Each point represents a dual \hbox{AGN}
discovered via radio (X), X-ray (circles), or optical/near-IR (dots; from \citealt{liu+11}).
Approximate lines for critical stages in binary formation and evolution are marked as horizontal lines.
The red, green, and purple curves indicate the resolution limit of 10\,GHz, 50\,GHz, and 120\,GHz center observing bands, respectively; for each observing set-up, one cannot resolve a binary orbit below that line. The dashed curves show the resolution limit of 150\,km baselines, while the solid curves show a nominal 1000\,km extended-baseline array. The ngVLA with a VLBI expansion can resolve, and hence directly image and identify, double supermassive black holes at sub-10\,pc separations. Longer baselines and higher frequencies have the potential to resolve multi-messenger SMBH binaries in the nearby Universe (indicated by the curves that cross the dark yellow region).}
\label{fig:vlares}
\end{figure}
\subsection{Direct Core Imaging}\label{sec:spectra}
Radio imaging of multiple self-absorbed synchrotron cores in the $\sim$1--50\,GHz frequency range is one of the most direct routes to identify dual SMBHs. It is exceedingly rare to find bright multiple flat-spectrum objects within close proximity \citep[e.\,g.][]{rodriguez,radiocensus}. The radio ``core'' that is seen in SMBHs represents the shock front near the base of the radio jet, which is compact and in a region of considerable optical depth. These can show a flat spectrum (with spectral index $\alpha\gtrsim-0.5$, where $S\propto f^\alpha$) up to moderate GHz frequencies \citep[e.\,g.][]{blandford+79}. This technique relies on detecting more than one flat-spectrum, compact radio source within a
common merger host. Confident confirmation of the nature of such detections requires modelling of broad-band continuum spectra, high-resolution morphological modelling, and ideally long-term studies of proper motion. For instance, in the object 4C37.11 (Fig.\,\ref{fig:lifecycle}), there are two cores at a projected separation of 7\,pc that remain compact down to sub-mas resolutions, with larger-scale diffuse emission that support the interpretation of the two knots as distinct, but gravitationally interacting, SMBHs. Long-term astrometric observations of this object have led to the first tentative ``orbital tracking'' for a resolved binary, although its orbital period is fairly long at $\sim$$10^4$\,yr \citep{bansal}.
Figure~\ref{fig:vlares} shows a census of known and candidate resolved (imaged) dual SMBHs. Uniquely, radio identifications have probed separations within a few hundred parsecs: the radius of a typical galactic stellar bulge \citep{merrittLRR}. Double black hole systems with separations smaller than a few tens of parsecs are \emph{direct precursors} to the gravitational-wave emitters detectable by PTAs and space-based laser interferometers, so are of utmost interest to the target science outlined in Section\,\ref{sec:intro}. There is only one confident binary in this range \citep{rodriguez}, and only a scant list of candidates \citep{deane,kharb+17}.
Binary SMBHs are relatively rare, and performing a complete blind search for dual radio cores requires relatively high dynamic range and thorough $u, v$ coverage at high resolution. Present facilities have aspects of these but not all. While the VLA has excellent sensitivity and $u, v$ coverage, it lacks sufficient resolution to discover close pairs.
While current long-baseline facilities have excellent resolution, they lack the snapshot sensitivity and $u, v$ coverage to thoroughly probe cores \citep[e.\,g.][]{deller+middelberg}. The ngVLA has a unique potential to reach a balance between these factors (Bansal et al.\ in prep).
\subsection{Jet Morphology}\label{sec:shapes}
Identification of dual AGN via large-scale jet morphology is somewhat fraught because of the difficulty in discerning external vs.\ internal influences on the jet(s).
Dual AGN emitting on long time scales can produce quadruple jets (Fig.~\ref{fig:dualagn}), and
binary interactions can result in precessing jets that produce large-scale S-shaped radio jet morphologies. Parabolic SMBH-SMBH encounters and binary coalescences can rapidly reorient a jet, potentially producing X-shaped morphologies \citep{merrittekers}. Previous large-scale radio surveys (e.\,g., \hbox{NVSS}, FIRST) have had a strong output of candidate binary or post-merger SMBHs identified through their large-scale jet morphologies. Recent surveys of such radio jet morphologies include 87 X-shaped radio jets found with the VLA \citep{roberts18}. However, past studies have revealed that many such morphologies can be described by backwards-flows of expelled jet material caused by fluid dynamics, rather than originating from a dual system \citep{xnotbinary}.
Still, some (as shown in Fig.\,\ref{fig:dualagn}) have been demonstrated either conclusively or extensively as candidate genuine dual AGN. Broad searches for such features require low-surface-brightness sensitivity coupled with resolution.
Large-scale periodic knots along radio jets have also been hypothesized to be periodic accretion episodes fueled by a binary; however, these are difficult to distinguish from flow instabilities \citep{godfrey_0637-752}.
\begin{figure
\centering
\includegraphics[width=0.33\textwidth,trim=0mm 7mm 0mm 5mm,clip]{figs/3c75.jpg}
\includegraphics[width=0.33\textwidth,trim=0mm 15mm 0mm 10mm,clip]{figs/ngc326.png}~
\includegraphics[width=0.2\textwidth]{figs/roos-93.png}
\vspace{-2ex}
\caption{Three examples of candidate dual AGN identified through their static or time-dependent radio jet morphology (Sections \ref{sec:shapes}, \ref{sec:time}). \textit{Left:} 3C~75, the dual cores of which have a 7~kpc projected separation and exhibits large-scale jet structures that have misaligned but correlated morphology. (Credit: \hbox{NRAO/AUI}; F.~N.~Owen, C.~P.~O'Dea, M.~Inoue, \& J.~Eilek). \textit{Center:} NGC~326, which has an \texttt{X}-shaped structure that inspired its dual nucleus interpretation. The cores were confirmed at other wavelengths, as shown. (Credit: \hbox{NRAO/AUI}; STScI [inset]). \textit{Right:} The helical jet of 1928+738 shows a 2.9-year cycle \citep{roos+93}.}
\label{fig:dualagn}
\vspace{-3mm}
\end{figure}
\subsection{Variability in Morphology or Flux}\label{sec:time}
Some radio quasars show light curves that appear to be periodic or sinusoidal. At high spatial resolutions, some jets have morphological variability: e.\,g.\ 1928+738 has jets that trace out helical patterns as a function of time \citep[Figure\,\ref{fig:dualagn};][]{roos+93}.
A common interpretation of both of these effects is that an accreting, jet-producing black hole has a jet axis whose angle is being modified by a secondary, orbiting SMBH. Confident identification of a periodicity in a light curve requires long-term variability monitoring (over more than a few cycles) to ensure that the observed variability is genuinely cyclical, and not simply the detection of a red noise process \citep{vaughan+16,charisi+18}. Thus, light curve variability programs require a long-term time investment for candidate binary identification.
As with the tracking of 4C37.11 \citep{bansal}, in an ideal case we will be able to track the orbital movements of one or two distinct cores. However, making the simple but suitable assumption that binaries are circular and obey Kepler's third law, their approximate largest physical separation scales with binary period as \mbox{$a\propto P^{2/3}$}. This means that the majority of resolvable binaries will have orbital periods well beyond human lifetimes. Tracking the orbital motion will only be possible for long baselines, and for the massive compact systems at low redshifts.
With long-baseline ($\sim$10,000\,km) arrays, such observations will be possible for $10^9 M\odot$ radio-emitting binaries with few parsec separations out to $z\sim0.1$ on timescales of a few years. This clearly
underscores the importance of a VLBI option for the ngVLA. If ngVLA baselines are limited to 300 km, then the timescales for resolving orbital properties will increase by a factor of 200, essentially making the ngVLA unsuitable for such studies despite its terrific snapshot sensitivity.
\subsection{Maser Dynamics}\label{sec:masers}
Long-baseline observations of masers have been instrumental in finding SMBH masses and analyzing accretion disk dynamics (e.g., \citealt{Miyoshi95}, \citealt{Braatz10}, \citealt{Kuo13}, \citealt{Gao16}). Deviations from Keplerian motion in the maser rotation curves can indicate that the potential is not that of a point source (e.g., \citealt{Kuo11}), and they can therefore be sensitive to close SMBH binaries (i.e., orbital separations of roughly $\sim$0.01--0.1\,pc).
For wider SMBH binaries ($\sim$1--100\,pc separation), the maser rotation curve is not expected to be substantially different from Keplerian. In this case, one can compare the barycenter velocity derived from the maser rotation curve (which unambiguously traces the SMBH velocity) to that of the galaxy on large scales ($\sim$10's of kpc, traced using stellar or HI dynamics). Significant differences between these two velocities are then indicative of binary or recoiling SMBHs, although a measurement of a velocity offset cannot by itself distinguish between the two possibilities (Pesce et al. 2018, submitted to ApJ).
\section{Current Radio Searches and Limitations}\label{sec:past}
In the past, most radio-identified candidate binary SMBHs have mainly been found serendipitously.
A few concerted efforts have attempted to survey directly for binary SMBHs. Here we discuss these, their successes, and limitations.
A number of blind or semi-blind searches have been made for dual radio cores, which as described in Sec.\,\ref{sec:spectra} is most easily realized through spectral imaging to search for multiple compact, flat-spectrum radio components. \citet{radiocensus} searched all available historical data from the Very Long Baseline Array with dual-frequency imaging; despite the large sample size ($\sim$3200 AGN), this search was limited primarily by sensitivity.
The VIPS survey of radio cores \citep{VIPS} originally discovered 4C37.11 as an extended flat-spectrum object; this source was later evidenced to be a binary SMBH \citep{rodriguez}. \cite{tremblay16} performed multi-frequency imaging of the flux-limited $\sim$1100 AGN included in VIPS, all of which were pre-selected to be bright and flat-spectrum. This search did not identify any new binary systems from the VIPS sample, and concluded that 15 candidates remained unconfirmed. Finally, \citet{condon-2mass} performed deep observations of 834 radio-bright 2MASS galaxies, under the hypothesis that the most massive galaxies must have undergone a major merger in their history, and therefore might contain a binary or recoiling SMBH. So far from these surveys, only one resolved dual core, 4C37.11, was re-detected. On simple arguments of merger rates, inspiral timescales, and radio AGN luminosity functions, at least tens of SMBH pairs must exist within these samples; thus, it is clear that these past experiments have been limited by a combination of $u, v$ coverage, sensitivity, dynamic range, and perhaps sample size.
\section{Current Multi-wavelength Synergies and Limitations}\label{sec:synergies}
Dual and binary SMBH signatures have been proposed in a number of wavebands (see, e.\,g., the review in \citealt{sbs-cqg}), and searches pursuing those signatures have revealed a number of binary candidates. We note that in all cases SMBH binaries have been challenging to identify because of their small separation on the sky, the uncertainties related to the uniqueness of their observational signatures, and the limited baseline of monitoring campaigns. Direct radio imaging, however, can provide an excellent follow-up to explore the nature of objects identified with these signatures.
The most prolific binary search method thus far has been the use of photometric and spectroscopic optical signatures to reveal candidate SMBH binaries. We describe that work here given its relevance to upcoming synoptic sky surveys that will be contemporary to ngVLA.
Multi-wavelength (radio, optical, and X-ray) searches for SMBH systems with large separations, corresponding to early stages of galactic mergers, have so far successfully identified a number of dual and offset AGN, as shown in Fig.\,\ref{fig:dualagn} \citep[][and others]{ngc6240, fu11,koss11, koss16, liu13b, comerford15, barrows16,mullersanchez15}.
SMBHs with even smaller (parsec and sub-parsec) separations are representative of the later stages of galactic mergers in which the two SMBHs are sufficiently close to form a gravitationally bound pair. Direct imaging using VLBI can confidently detect such a system, however requires deep observations to do so successfully. In this regard, optical surveys can help to identify a sample of SMBH binaries for VLBI followup \citep[see][for a review]{Bogdanovic2015}.
Current studies have been exploring periodic quasar variability as a potential indicator for binary SMBH emissions; these and future surveys of this kind using synoptic instruments like PAN-STARRS and LSST will reveal many candidate binary systems ideal for follow-up with the ngVLA plus a long-baseline extension.
Current photometric surveys
have already uncovered about 150 SMBH binary candidates \citep[e.g.,][]{valtonen08, Graham2015, charisi16, liu16}. In this approach, the quasi-periodic variability in the lightcurves of monitored quasars is interpreted as a manifestation of binary orbital motion. Because of the finite temporal extent of the surveys,
most of these candidates have orbital periods of order a few years, hence would be in the regime of gravitational-radiation-dominated inspiral (Fig.\,\ref{fig:lifecycle}). However, studies of this type suffer the same red-noise issues noted in Sec.\,\ref{sec:time}.
A recent, initially quite promising binary SMBH candidate from a quasar light curve, PG1302, has recently been losing traction. Additional data from the All-Sky Automated Survey for Supernovae (ASAS-SN), extending the light curve to present-day, shows that the periodicity does not persist and disfavors the binary SMBH interpretation \citep{liu18}.
Furthermore, the upper limit placed by PTAs on the gravitational wave background largely rules out the amplitude implied by the $\sim150$ photometric binary candidates, implying that some significant fraction of them are unlikely to be SMBH binaries \citep{sesana17}. While this will lead to a downward revision in the number of photometrically identified SMBH binary candidates, it provides a nice example of the effectiveness of multi-messenger techniques, when they can be combined.
Spectroscopic searches for SMBH binaries have so far identified several dozen candidates. They rely on the detection of a velocity shift in the emission-line spectrum of an SMBH binary that arises as a consequence of the binary orbital motion
\citep{gaskell83,gaskell96,bogdanovic09}. The main complication of this approach is that the velocity-shift signature is not unique to SMBH binaries; it may also arise from outflowing winds near the SMBH, or even from a gravitational-wave recoil kick \citep[e.g.,][]{popovic12,Eracleous2012,barth15}. To address this ambiguity, spectroscopic searches have been designed to monitor the offset of broad emission-line profiles over multiple epochs and target sources in which modulations in the offset are consistent with binary orbital motion \citep{bon12, bon16, Eracleous2012, decarli13, Ju2013, liu13, shen13, Runnoe2015, Runnoe2017, li16, wang17}. Two drawbacks of this approach however remain: (a) temporal baselines of spectroscopic campaigns cover only a fraction of the SMBH binary orbital cycle and (b) intrinsic variability of the line profiles may mimic or hide radial velocity variations, making these measurements more challenging and less deterministic \citep{Runnoe2017}.
\section{Example ngVLA Studies}\label{sec:ngvla}
Dual and binary SMBH studies can be realized by a number of specific studies enabled by the ngVLA:
\begin{description}
\item[Blind surveys.] The ngVLA can survey many sources (thousands) with high sensitivity, high dynamic range, and high two-dimensional image fidelity (i.\,e.\ relatively complete $u, v$ coverage) in the frequency range 1--50\,GHz with moderate resolution. These abilities would enable a blind survey for resolved dual cores in the $\gtrsim$few parsec separation regime. Such a survey could also reveal single, offset cores that may arise from a gravitational-wave recoil kick following a SMBH merger. Candidates can be cross-matched with future optical/near-IR or X-ray (e.\,g., Athena) surveys to measure spectra, redshifts, and disturbed host galaxy structures. While the multi-wavelength observations might not have the resolution of the \hbox{ngVLA}, they would provide a more complete understanding of a candidate and whether it
is a genuine dual \hbox{AGN}, a gravitational lens,
a chance projection, or even a recoiling SMBH.
\item[Targeted surveys to test candidate binaries.]
Many binary SMBH candidates are likely to be identified in the coming decades by numerous instruments (including PTAs, LSST and the extremely large optical telescopes currently being planned). The ngVLA will be the premier instrument to follow up these candidates in radio, with the goal of understanding their long-term dynamics through large-scale jet morphologies (at low GHz frequencies, given sufficient sensitivity to low surface brightness emission). A VLBI component on the ngVLA will enable observation and tracking of resolved dual cores, which will truly confirm a binary and allow in-depth study of SMBH binary evolution.
\item[Core variability.] The frequency range of ngVLA makes it well-suited to observe AGN that are dominated by core emission.
Past work has shown that in blind surveys at $\sim$10\,GHz, radio cores are more dominant than diffuse lobes \citep{massardi+11}. Thus, the high sensitivity of the ngVLA at tens of GHz frequencies can permit easier time-resolved radio quasar light curve monitoring, and thus have the potential to identify more examples that might exhibit periodic activity due to variable accretion or orbital precession.
\item[Multi-messenger searches.]
GW experiments can only poorly localize their detections \citep{ellis13}. In concert with multi-wavelength facilities, the ngVLA could perform efficient follow-up surveys for all galaxies within a given error radius, identifying a host by looking for the signatures described above. With sufficiently high sensitivity on long baselines, the ngVLA could observe and track the orbital motions of radio cores for the most nearby GW candidates, thus enabling specific multi-messenger studies. In particular, PTAs will preferentially detect the most massive ($M_{\rm BH} \ga 10^9$ M$_{\odot}$), nearby ($z \la 0.1$) objects that are accessible to Earth-based VLBI \citep[e.\,g.][]{kelley+17,mingarelli-2mass}.
\item[Long-baseline benefits.]
A VLBI component to the ngVLA has clear benefits for this field of research. A long-baseline extension to the ngVLA, where each long baseline consists of a cluster of $\sim$5 antennas, would yield several major benefits. First, much of the imaging of double black holes is resolution-limited, meaning that going to longer baselines would immediately allow the probing of closer pairs of black holes. Second, long baselines would allow the maser experiments described in Section\,\ref{sec:masers}. Third, the clusters themselves can be used for pulsar timing when the ngVLA core is not being used as part of a VLBI project, giving access to the pulsars too far north for SKA timing.
\end{description}
\noindent The above studies enabled by the ngVLA will provide unique probes of SMBH binary evolution in regimes inaccessible with other methods. This includes crucial data on binaries in the truly gravitationally bound phase, such as information about their separations, inspiral timescales, luminosities and variability even when not in a rapidly-accreting ``quasar" state. Moreover, studies of radio jet morphology and energetics will provide unprecedented constraints on the mechanisms by which SMBH feedback drives host evolution and contributes to the origin of the observed SMBH-galaxy correlations. In concert with complementary SMBH binary studies by PTAs and LISA,\footnote{See also the chapter on ngVLA/LISA synergies in this volume; Ravi et al.} the ngVLA will be a key cornerstone in the coming era of multimessenger astronomy. The synergy between these electromagnetic and gravitational-wave probes of the binary regime has the potential to unravel the full puzzle of binary SMBH evolution and its close ties to the evolution of galaxies.
\acknowledgements
SBS is supported by NSF award \#1458952. SBS and TJWL are members of the NANOGrav Physics Frontiers Center, which is supported by NSF award \#1430284. T.B. acknowledges support by the National Aeronautics and Space Administration under Grant No. NNX15AK84G issued through the Astrophysics Theory Program and by the Research Corporation for Science Advancement through a Cottrell Scholar Award. L.B. is supported by NSF award \#1715413. D. P. acknowledges support from the NSF through the Grote Reber Fellowship Program administered by Associated Universities, Inc./National Radio Astronomy Observatory. Y.S. acknowledges support from an Alfred P. Sloan Research Fellowship and NSF grant AST-1715579. Part of this research was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration.
|
\section*{\large Acknowledgment}
This work results from the project DeCoInt$^2$, supported by the German Research Foundation (DFG) within the priority program SPP 1835: "Kooperativ interagierende Automobile", grant number SI 674/11-1.
\section{\large Conclusions and Future Work}
\label{sec_conclusion}
\addtolength{\textheight}{-0.0cm}
In this article, we presented an approach to detect the starting movement of cyclists using smart devices.
The approach is based on HAR and a modelling of the starting movement detection as a classification problem.
We introduced an additional auxiliary class modelling the transition between the \textit{waiting} and \textit{moving} classes.
This class allows to integrate early movement indicators, i.e., body movements indicating future behaviour.
In this way we improve the robustness and reduce the detection time of the classifiers.
It is complemented by a novel two-stage feature selection procedure, selecting robust and understandable features.
In our experiment involving 49 test subjects conducted in real-world traffic, we found
that our detector based on the \textit{location-agnostic} XGBoost classifier reaches
an $F_1$-score of 67\% at a mean detection time of \SI{0.33}{\second} evaluated for all wearing locations.
Although, our approach is not able to reach an $F_1$-score of one, our analysis of the detection times
showed that a considerable amount of starting movements are detected
before the first movement of the wheel.
We also showed that the device worn in the trouser pocket provides the best detection results.
Moreover, we demonstrated that training distinct classifiers for specific locations
can further improve detection results, i.e., reaching an $F_1$-score of 94\% with a mean detection time of \SI{0.34}{\second} for the device worn in the trouser pocket.
Our proposed approach is still not fully competitive with
accurate video-based methods (e.g., ~\cite{ZKS+18}) as available in infrastructure and vehicles, yet.
Nevertheless, these methods fail in case of bad visibility or occlusion.
We see the information delivered by smart devices as complementing the cooperative intention detection
process~\cite{decoint2BZH+18}, e.g., allowing to resolve occlusion situations.
For future work, we will focus on extending our presented approach towards a fully smart device based intention detection system including improved localization abilities and trajectory forecasts.
We will generalize our approach to different VRUs such as pedestrians.
Moreover, instead of using only a single smart device, we assume that in near future VRUs will
have various wearable devices at different wearing locations, e.g., smart
watches at their wrist, sensor-equipped helmets and shoes.
We will also investigate the continuous on-line improvement of the detection models.
Therefore, we aim to integrate imprecise labels by means of active and semi-supervised learning techniques to reduce the dependency on the high quality camera system.
Furthermore, we will investigate a more tight integration into a cooperative
setting, i.e., combining information from smart devices and other road users for advanced VRU safety.
We will investigate opportunities to fuse this information to improve our starting movement detection approach.
\section{\large Introduction}
\label{sec_introduction}
\subsection{Motivation}
In our work, we envision future mixed traffic scenarios where
automated cars, trucks, sensor-equipped infrastructure,
and other road users equipped with
smart devices or other wearables are interconnected by means of ad hoc networks.
This allows the traffic participants to cooperate, i.e., determine and
maintain local models of the surrounding traffic situations.
Vulnerable road users (VRUs) will still play an important role in future urban traffic. To avoid accidents, it is not only important to detect VRUs, but also to anticipate their intentions. Although modern vehicles are equipped with forward looking safety systems, dangerous situations for VRUs may still occur as a result of occlusions or sensor malfunctions.
\begin{figure}[t]
\centering
\includegraphics[width=0.45\textwidth, clip, trim= 20 30 40 20]{images/cyclist_behind_bus_rendered.png}
\vskip 4mm
\caption{Typical dangerous situation in urban environment. The cyclist intending to cross the
road is occluded by the bus and cannot be seen by the approaching vehicle.
The smart device worn by the cyclist anticipates the starting movement and transmits (indicated by dashed line) to the approaching vehicle such that a potential collision is avoided.
}
\vskip -6mm
\label{fig:cyclist_crossing}
\end{figure}
Figure~\ref{fig:cyclist_crossing} sketches such a typical dangerous occlusion situation. A cyclist intends to cross the street while a vehicle hidden behind the bus is approaching. A smart device worn by the cyclist can anticipate the starting movement and communicate the detected intention to the approaching vehicle to warn the driver or initiate a braking maneuver.
Smart devices and other wearables are capable of detecting the position using the device's
integrated global navigation satellite system (GNSS)
and detecting movement transitions early using their inertial sensors.
Especially the latter are of great interest, since data from the inertial sensors
is not affected by GNSS outage (a phenomena often encountered in urban areas) and
it is available at high sampling rates. Hence, it is suitable to quickly detect transitions between movement states of VRUs, e.g., waiting and moving.
In this article, the focus is set on detecting cyclists' starting movement fast and yet reliable using inertial sensors. Our approach to detect starting movements is based on human activity recognition (HAR)~\cite{Bulling2014THA} and machine learning techniques.
The difficulty concerning smart device based starting movement detection is not whether we will detect a starting movement but to detect it as early as possible, i.e., in the range of a few hundred milliseconds after the actual starting movement.
For illustration consider the following urban scenario: An automated vehicle is approaching with 50\,km/h and has a braking deceleration of 8\,m/$s^2$. If it initiates the braking 15\,m before a crossing cyclist, then it will stop 3\,m ahead of the cyclist.
After the cyclist has entered the driving corridor, the automated system only has $0.58$\,s to initiate a braking maneuver and to prevent an accident.
Nevertheless, the detector must also be robust, i.e., avoid false positive detections potentially leading to unnecessary emergency braking maneuvers.
\subsection{Main Contributions and Outline of this Paper}
Our main contribution is an approach based on HAR and machine learning using inertial sensors to detect cyclists' starting movements in real world traffic scenarios. We model the detection as a classification problem.
This article considers the following new aspects to detect cyclists' starting movements:
\begin{itemize}
\item An improved starting movement detection process using
a novel training procedure based on an auxiliary class modelling the transition between waiting and moving.
It reduces the false positive detections and leads to earlier starting movement detections.
\item A two-stage feature selection procedure to select robust features. It reduces false positive detections and leads to understandable and interpretable features.
\item
Investigation on different device wearing locations in which
we found that wearing the device in the trouser pocket leads to the best results.
\item Training of distinct classifiers for different wearing location to improve detection results.
\end{itemize}
The remainder of this article is structured as follows:
In Section~\ref{sec_related_work} related work is presented,
Section~\ref{sec_method_overview} describes our methodology for starting movement detection. The data acquisition and evaluation methodology is presented in Section~\ref{sec_evaluation} and the experimental results are reviewed in Section~\ref{sec_ResultsOutline}. Lastly, in Section~\ref{sec_conclusion}, the main conclusions and open issues for future work are reviewed.
\section{Related Work} \label{sec_related_work}
Many dangerous situations involving vehicles and VRUs occur in urban areas.
For better VRU protection, intelligent vehicles are equipped with active safety systems aiming to anticipate the VRU's intentions.
We see the motion of a VRU as a sequence of activities or basic movements (e.g., standing or moving)
and trajectory of certain body points (e.g., center of gravity or head) in
the 3D space. Basic movement detection and trajectory forecasting are part of what we term as intention detection.
In this article, we focus on detecting the starting movement. An early basic movement detection can support the trajectory forecast~\cite{mbBZD+17}.
Most of the conducted research concerning intention detection relies on computer vision based solution.
In~\cite{KG14, Kooij2014}, the authors compare different approaches involving the past VRU positions and additional features derived from camera images. The approaches for pedestrian intention detection are based on recursive Bayesian filter (e.g., Kalman filter) and machine learning techniques, e.g., Gaussian process dynamical models, dynamic Bayesian networks, and probabilistic hierarchical trajectory matching.
In~\cite{ZKS+18}, Zernetsch~et~al. use machine learning techniques, i.e., support-vector machines and convolutional neural networks, to image sequences of an infrastructure-based camera system for early starting intention detection of cyclists. Their approach safely detects starting movements on average 0.14\,s after the cyclist starts moving.
Although all of these techniques show promising detection results they all rely on camera- or laser
scanner-based perception, which might not always be available (e.g., in case of occlusion). Cooperative
approaches involving smart devices can alleviate this.
In~\cite{Liebner2013}, Liebner~et~al. present a bicycle warning system to warn turning vehicles
before approaching cyclists using the smartphone's integrated GPS and 3G HSDPA for smartphone-to-vehicle
communication. The authors mainly focus on the evaluation of GPS accuracy.
They assume that the cyclists move with constant velocity for predicting future cyclists' trajectory (for \SI{3}{\second}). This and additionally the communication delay limits the application of their approach.
The authors in \cite{MSN17} propose a cooperative system using smartphones and vehicles. GPS information originating from the smartphone is used to resolve occlusion situations and to enhance the perceptual horizon of vehicles. The authors solely focus on GPS and cooperative perception. They do not consider any other smartphone-based detectors.
A prototype system for cooperative safety applications involving a cyclist equipped with a smartphone
is proposed by Thielen~et~al. in~\cite{Thielen2012}. The authors show a prototype application that warns a vehicle driver if the collision with a crossing cyclist is likely to occur within the next few seconds. The cyclist's trajectory forecast is based on GPS using a least squares fitting for extrapolation.
Due to the low GPS sample frequency and the extrapolation for prediction, fast movement transitions cannot
be detected. In~\cite{Engel2013}, an approach involving Car2Pedestrian communication for pedestrian tracking is proposed.
It combines GPS data with inertial sensors allowing to transmit position and movement type to
an approaching car. This allows to warn the pedestrian and the driver. Although their approach concerning
pedestrian movement detection is similar to ours, they do not focus on the fast detection of movement
transitions. Moreover, they focused on a single device wearing location using specialized features.
In~\cite{BMD17}, the authors propose a smartphone-based approach to detect additional
context information, e.g., detection of pedestrians leaving the curb, in order to improve the
smart device based collision risk assessment.
In~\cite{mbMLT+17}, the authors present and compare different approaches to pedestrian path prediction for a time horizon of \SI{5}{\second} using a smartphone. They consider approaches based on artificial neural networks and dead reckoning. Their preliminary results on a small evaluation set consisting of only two pedestrians are promising. Nevertheless, their approach does not focus on detecting fast and critical movement changes nor do the authors evaluate different wearing locations.
In our previous work \cite{mbBZD+17, decoint2BZH+18}, we presented a cooperative approach to
cyclists' starting intention detection involving smart devices as an essential component. The approach presented in this article is an extension with special focus solely on smart devices.
\section{Data Acquisition and Evaluation}
\label{sec_evaluation}
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{images/trajectories.png}
\vskip 4mm
\caption{Overview of the intersection with all starting movements of instructed cyclists. The arrow is pointing into the starting direction.}
\label{fig:trajectories}
\vskip -4mm
\end{figure}
\subsection{Data Acquisition} \label{subsec_data_acquisition}
For the evaluation of our approach, we used a dataset consisting of cyclists's starting movements.
It contains 49 female and male test subjects. They were instructed to move between certain points at an intersection with public, uninstructed traffic while adhering to the traffic rules. Since there are two traffic lights at the intersection, we obtained 84 starting motions, with a maximum of two starting motions per test subject.
The trajectories of the cyclists are plotted in Fig. \ref{fig:trajectories}.
In order to label the starting movements as described in Section~\ref{subsec:starting_movement_classifcation},
we recorded the images of two HD cameras, which were installed at the
intersection as part of a wide angle stereo camera system.
All test subjects were equipped with four smart devices worn at different body locations, i.e., wearing locations.
The smart devices used for evaluation are Samsung Galaxy S6 smartphones.
In our opinion these four locations listed in the following
are typical representatives of wearing locations observed in everyday life.
The test subjects were equipped with a smartphone in their front trouser pocket.
Moreover, the test subjects were equipped with a smartphone placed in a backpack.
Another smart device was placed in the front pocket of the test subjects' jackets at the height of their chests.
For those test subjects which did not posses a front pocket at the chest
the smartphone was mounted within a pocket located at the chest belt of the backpack.
The fourth device was mounted at the rack of the bicycle. If no rack was available, then it was put into
a bag mounted beneath the saddle. The motivation behind choosing the placement on the rack is that many cyclist
carry their smartphone in a bag on their rack.
The smart devices were all placed in predefined orientations, e.g., for front trouser pocket: Upright position and
display facing outwards. This makes the experimental setup comprehensible and reproducible. This does not limit the general applicability of the presented approach since it uses features computed on a device orientation invariant representation of the inertial data (cf. Section~\ref{subsec:preprocessing}).
We used different bicycles during the experiments, ranging from mountain and touring over city to racing bikes.
The dataset used in this article for evaluation of our approach is made publicly available\footnote{\url{https://git.ies.uni-kassel.de/intention_detection/starting_detection}}.
\subsection{Evaluation} \label{subsec_evaluation}
We performed the evaluation of our starting movement detection approach off-line using a ten-fold cross-validation over the VRUs.
For detection performance assessment, we propose a scene-wise evaluation.
A scene comprises the timespan after stopping till the cyclist is leaving the field of view of the camera
used for labeling. It only consist of a single \textit{waiting}, \textit{starting}, and \textit{moving} phase (cf.~Fig.~\ref{fig:ex_net_out}),
whereby the \textit{starting} is optional, since some cyclists start moving right away without showing
any early movements. A scene is rated as false positive if the detection time falls into phase I.
If the detection is in phase II or III, then it is rated as true positive.
If the \textit{moving} class is not detected, then it is rated as false negative.
Since every waiting phase ends in a starting phase, we do not consider true negatives.
The overall quality regarding robustness (i.e., avoiding false detection)
of the detectors is evaluated by means of the $F_1$-score calculated over all scenes.
For evaluation we additionally remove \SI{3}{\second} at the beginning of each waiting phase
since we focus on detecting \textit{waiting} to \textit{moving} transitions, i.e., starting movements.
Besides the robustness, also the detection time $t_d$ is crucial.
Therefore, we calculate the mean time difference $\overline{\delta_t}$ between the detection time $t_d$ and the start time of phase III $t_{III}$ of all true positives over all $L$ scenes (Eq. \ref{eq:mean_detection_time}).
Smaller values are better, even negative values are possible.
\begin{equation}
\overline{\delta_t}=\frac{1}{L}\sum_{i=1}^{L}(t_{di}-t_{IIIi})
\label{eq:mean_detection_time}
\end{equation}
The trade-off regarding robustness and a fast detection can be considered a multi-objective
optimization problem~\cite{mie99}. For evaluation we adopt the concept of Pareto optimality. A solution is Pareto optimal
if none of the involved objective functions can be further improved without deteriorating some of the other
involved objectives. The set of Pareto optimal solutions is referred as Pareto frontier. Those solutions
which are not Pareto optimal are referred as dominated solutions.
Without any additional weighting (i.e., rating which goal is more important) all Pareto optimal solutions are considered equally good. We create the Pareto frontiers by means of a parameter sweep, i.e., we evaluate the $F_1$-score and mean detection time for a set of different parameter configurations. Subsequently, we can calculate the Pareto frontier for this set.
\subsection{First-Level Movement Primitive Detection}
\label{subsec:movement_primitive_detection}
Slow movement primitive detection.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth, clip, trim=50 60 100 105]{"images/har_pipeline"}
\caption{Process for movement primitive detection based on smart device and head trajectories.}
\label{fig:recognition_pipeline}
\vskip -4mm
\end{figure}
\subsection{Preprocessing} \label{subsec:preprocessing}
\begin{figure}[b]
\centering
\includegraphics[width=0.5\textwidth, clip, trim=0 735 200 0]{images/preprocess_pipeline_frames.pdf}
\caption{Preprocessing of linear accelerometer and gyroscope involving coordinate transformation and magnitude calculation.}
\vskip -4mm
\label{fig:preprocessing}
\end{figure}
The first stage of the starting movement detector is concerned with the sensor preprocessing.
Our approach uses the accelerometer and gyroscope sensor, which are sampled
with a frequency of \SI{50}{\Hz}. The angular velocity measured by the gyroscope
can be used to detect rotation movements, such as pedalling, while the accelerometer
is better suited to detect linear movements, e.g., forward movement.
We use gravity compensated accelerometer data, referred as linear acceleration data.
The gyroscope values are bias corrected using the calibration supplied by the manufacturer.
A further drift compensation is not performed. Modern smart devices run sensor fusion algorithms. These incorporate data originating from gyroscope, accelerometer, and magnetometer to estimate the smart device orientation with respect to a global reference frame.
The preprocessing described in the following is depicted in Fig.~\ref{fig:preprocessing}.
The three components ($x$, $y$, and $z$) of the linear accelerometer ${}^{b}acc$ and gyroscope ${}^{b}gyro$ are transformed using the orientation estimation supplied by the smart device's operating system. The smart device coordinate system $b$ is referred to as body frame.
The linear acceleration and gyroscope data is now in a coordinate frame which is leveled with
the local earth ground plane, i.e., the $z$-axis is pointing towards the sky. This coordinate
frame is referred to as the local tangential frame $t$.
The compass is not considered due to its sensitivity to a precise calibration~\cite{MGF+17} and possible magnetic perturbations. For this reason, we do not consider the precise transformation from this local frame to a global reference frame (e.g., north-east direction frame) or the difficult estimation of
the device's orientation with respect to the VRU.
We resolve this issue by considering the magnitude of the linear accelerometer $|| {}^{t}acc_{x,y} ||$ and the gyroscope $|| {}^{t}gyro_{x,y} ||$ in the local tangential horizontal $x-y$ plane (also referred as ground plane). This representation is invariant concerning orientation. Hence, we avoid the challenging estimation of the transformation between the device orientation with respect to the VRU~\cite{MGF+17}. In addition, we consider the projection of the linear accelerometer ${}^{t}acc_{z}$ and the gyroscope ${}^{t}gyro_{z}$ on the local vertical $z$-axis, i.e., the gravity axis.
We do not consider the smart device integrated GPS for the following two reasons.
First, GPS is not always available or noisy due to multipath effects especially in urban
areas. Second, the sampling frequencies of \SI{1}{\Hz} provided by modern smart devices is too low to detect fast changes in cyclists' movements.
\subsection{Segmentation and Feature Extraction} \label{subsec:segmentation_feature_extraction}
In the second stage, we perform a sliding window segmentation using different window sizes of the transformed
signals and consecutively extract features. Here, we consider features commonly used in HAR~\cite{Bulling2014THA}, such as the energy, minimum, and maximum, which are computed for
sliding window sizes \SI{0.1}{\second}, \SI{0.5}{\second}, \SI{1.0}{\second}, and \SI{2.0}{\second}.
Features computed for different window sizes allow the detector to
handle different time scales, i.e., features computed with the smaller window sizes capture short term dependencies and larger window sizes capture dependencies on a longer timescale.
Furthermore, we consider features based on orthogonal polynomial approximation up to the $3^{\mathrm{rd}}$ degree
for window lengths of \SI{0.5}{\second}, \SI{1.0}{\second}, and \SI{2}{\second}. These polynomial coefficients
are in a least squares sense best estimators of the mean, slope, and curvature of the input signal~\cite{Fuchs2010}.
In addition, we consider the magnitude of the discrete Fourier transform (DFT) coefficients
for window sizes \SI{0.64}{\second} and \SI{5.12}{\second},
as successfully applied for human walking speed estimation in~\cite{PPC+12}.
We only consider coefficients up to the $5^{\mathrm{th}}$ order since typically human motion
is best represented by the lower frequencies.
In order to make the DFT coefficients independent of the energy, we normalize the coefficients with respect to the
square root of the signal's overall energy within the respective window.
In total, $164$ features are computed.
\subsection{Feature Selection} \label{subsec:feature_selection}
\begin{figure}[h]
\centering
\includegraphics[width=0.48\textwidth]{images/feature_selection.png}
\vskip 4mm
\caption{Two-stage feature selection approach: First we apply four different filters (i.e., MIFS, mRMR XGBoost, and elastic net) and union ten best scoring features of each filter, Second, we apply backward feature selection.}
\vskip -4mm
\label{fig:feature_selection}
\end{figure}
Selecting good features is the key for robust and yet fast detection.
The feature selection is realized in the third stage of the starting movement detector.
As with the final classification, we have to face the problem of highly
imbalanced classes. In order to compensate for this, we randomly
undersample the \textit{waiting} class up to the size of the \textit{moving} class.
Subsequently, we randomly oversample the \textit{starting} class, such that
all three classes have equally many samples.
For the feature selection, we adopt the following approach:
First, we apply filters~\cite{LCW+17} and then we fine-tune the feature selection
by means of a wrapper approach. A schematic of this is depicted
in Fig.~\ref{fig:feature_selection}. The goal of using filters is to
pre-select a set of potentially meaningful features and to
reduce the computational complexity. In order to get a large diversity
concerning pre-selected features, we apply different filters, i.e., mutual information (MIFS),
minimum redundancy maximum relevance (mRMR) feature selection and two model-based selection techniques based on elastic net and XGBoost.
MIFS selects features which generally have a high mutual information score for the
classification target, whereas mRMR selects features with high mutual information which additionally do not overlap much.
The model-based techniques select features suitable for the respective classifier, i.e., the XGBoost and linear discriminative models.
We consider the union of the ten best scoring features of each filter resulting in at most $40$ different features.
Then, in the second stage of the feature selection,
we use these previously selected features in a wrapper approach in
combination with the currently considered detector. This fine-tunes the
feature selection by only selecting those which are relevant for the
optimization criteria.
We use an $F_1$-score defined over starting scenes (described in Section~\ref{subsec_evaluation}) as optimization criteria to
perform a backward feature selection~\cite{GE03} with the respective
classifier. The parameters of both classifiers are fixed during feature selection.
\subsection{Classification}
In the fourth stage of the starting movement detector,
the starting detection is realized by means of a frame-based classification using
a linear SVM and XGBoost~\cite{CG16}. The frame-based classification is performed at discrete points with a frequency of \SI{50}{\Hz}. The linear SVM is chosen since it has proven a good generalization ability and
its linear decision boundary can be efficiently evaluated when applying the classifier in the field.
Especially the latter is appealing for the given evaluation frequency of \SI{50}{\Hz}.
The XGBoost algorithms has won many awards in current machine learning challenges and can be considered
state-of-the-art. Nevertheless, the algorithm is computationally more involved.
The linear SVM was optimized using the Hinge loss. Moreover, due to the large number of
training samples, we trained the linear SVM in the primal space.
The XGBoost algorithm is an ensemble method based on classification
and regression trees. We considered the logistic loss as the objective function underlying the
XGBoost training with a regularization term controlling the model complexity (see~\cite{CG16} for more details). In order to further improve the generalization ability, we considered regularization in form of learning rate shrinkage and random feature subsampling.
The classifier is trained on labeled data consisting of three classes, i.e., \textit{waiting}, \textit{starting}, and \textit{moving}. During run-time only the \textit{waiting} and \textit{moving} classes are considered for starting movement detection.
As before for the feature selection, we randomly undersample the \textit{waiting} class
and oversample the \textit{starting} class. The class ratio, i.e, fraction of samples of each class used for training, is an important factor influencing the design of the resulting starting movement detector. For example,
putting strong emphasis on the \textit{waiting} class results in robust starting movement detection but high detection times.
Instead of directly using the classification returned by the frame-based classifier,
we favour to use class probabilities representing confidence estimates about the classification, i.e., detection.
This probability estimate can be used to design the starting detector,
such that it reacts only upon exceeding a certain threshold applied on the \textit{moving} class probability.
Since neither the linear SVM nor the XGBoost classifier return proper probability estimates
(e.g., the XGBoost tends to predict overconfident probabilities), a probability calibration fitting an additional
sigmoid (i.e., Platt calibration~\cite{Niculescu-Mizil2005PGP}) is performed.
Before determining the final decision using the threshold on the \textit{moving} class probability,
an additional smoothing is performed to reduce the amount of false positive detections.
This is realized by means of a soft voting ensemble approach
i.e., running average of the \textit{moving} class probability.
\subsection{Cooperative Trajectory Forecast}
\label{sec_method_overall}
The second stage of the intention detection process is completed by the cooperative trajectory forecast.
It is realized by a soft gating approach, weighting
the waiting and starting forecast model in an additive fashion. A schematic of this process is depicted in Fig~\ref{fig:stage2}.
The cooperative trajectory forecast for a particular prediction time $t_\mathrm{c}$ is formally
defined in the following equation:
\begin{align}
y \left(t_\mathrm{c} \right) = &g \left( x_{t_\mathrm{c}}^\mathrm{w} \right) \cdot f_\mathrm{wait} \left( t_\mathrm{c}, x_{t_\mathrm{c}} \right) \\
&\left( 1 - g \left( x_{t_\mathrm{c}}^\mathrm{w} \right)\right) \cdot f_\mathrm{start} \left( t_\mathrm{c}, x_{t_\mathrm{c}} \right) \nonumber
\label{eq:Fusion}
\end{align}
The gating function $g$ depends on the movement primitives computed either by the smart device,
camera, or the combination of both.
It implements the model combination by weighting
between the forecasts of the waiting and starting forecast model ($f_\mathrm{wait}$ and $f_\mathrm{start}$,
respectively) at time $t_\mathrm{c}$. Moreover, $x_{t_\mathrm{c}}^\mathrm{w}$ denotes the input feature vector for the model fusion, whereas
$x_{t_\mathrm{c}}$ denotes the input for the forecasting models.
The probability output generated by the movement primitive classifier (cf. Section~\ref{sec:cooperative_mp_pred}) is directly used as a gating function to weight between the predicted forecasting models. The probability is the product of the cooperative movement primitive detection.
\begin{figure}[t]
\includegraphics[width=0.49\textwidth, clip, trim=0 180 370 80]{images/EnsembleForecast.pdf}
\caption{Stage 2: Ensemble forecast approach using two models with adaptive gating function to weight the models predictions.}
\label{fig:stage2}
\vskip -0.3cm
\end{figure}
\section{\large Method}
\label{sec_method_overview}
Our approach aims to detect cyclists' movement transitions between waiting and moving (i.e., starting) as early as
possible using smart devices carried by cyclists. The starting movement detector is realized by means of a HAR
pipeline~\cite{Bulling2014THA}.
A schematic of the starting movement detector consisting of four stages is depicted in Fig.~\ref{fig:recognition_pipeline}.
The cyclists' starting movement detection is modelled as a classification problem, i.e., \textit{waiting}, \textit{moving}.
Additionally, we consider an auxiliary class \textit{starting} modelling the transition between \textit{waiting} and \textit{moving}. The starting movement detector uses inertial data, i.e., accelerometer and gyroscope as input. The data is preprocessed (i.e., transformed in
device attitude invariant representation), segmented, and features are extracted. Subsequently, to increase the generalization ability, feature selection is performed. Finally, the detection is realized by means of machine learning based classifiers, i.e., a support-vector machine with linear kernel (linear SVM) and an extreme gradient boosting classifier (XGBoost)~\cite{CG16}.
\subsection{Trajectory Forecast}
The trajectory forecast of a cyclist belongs to the second stage of the intention detection process.
To forecast the future positions of the cyclist, the measured cyclist positions over the last second are transformed from 2D world coordinates into the ego coordinate system of the cyclist and are derived numerically to obtain the longitudinal and lateral velocities. The velocities are split into different temporal windows (Fig.~\ref{fig:MLP_IN_OUT}: gray segments) and are approximated with orthogonal base polynomials. The polynomial coefficients are used as input of two MLPs, one trained for waiting and one trained for starting motions. The output of the MLPs are the coefficients of polynomials, which represent the forecast longitudinal and lateral cyclist velocities over a certain forecast horizon in different temporal windows (Fig.~\ref{fig:MLP_IN_OUT}: blue segments). The velocities are reconstructed by evaluation of the coefficients, and through numerical integration of the velocities the forecast positions are retrieved. The output of both networks are combined as described in Section \ref{sec_method_overall}.
In~\cite{Zernetsch.2016} a single MLP was trained using cyclist trajectories of all motion types. For this work two networks were trained, one model for waiting ($f_\mathrm{wait}$) and one model for starting ($f_\mathrm{start}$). The reason to separate the forecast of waiting and starting motions is to reduce the forecast error, especially of starting motions. Various movements of cyclists during the waiting phase (e.g. swaying back and forth or leaning over the handlebar) lead to a decreased performance of the network described in~\cite{Zernetsch.2016} when it comes to forecasting starting motions since these movements look similar to starting, but they do not result in a large change of the cyclist velocity.
{
\setlength{\abovecaptionskip}{5pt plus 3pt minus 3pt}
\begin{table}[h]
\begin{center}
\caption{Configuration of polynomials used to train the MLP.}
\begin{tabular}{| c | c | c | c |}
\hline
Polynomial & Degree & Start Time & Length \\ \hline
$P_{in1}$ & $2$ & $-1.0\,$s & $0.5\,$s \\ \hline
$P_{in2}$ & $3$ & $-0.5\,$s & $0.5\,$s \\ \hline
$P_{out1}$ & $3$ & $0.0\,$s & $0.5\,$s \\ \hline
$P_{out2}$ & $3$ & $0.5\,$s & $0.5\,$s \\ \hline
$P_{out3}$ & $3$ & $1.0\,$s & $0.5\,$s \\ \hline
$P_{out4}$ & $3$ & $1.5\,$s & $0.5\,$s \\ \hline
$P_{out5}$ & $3$ & $2.0\,$s & $0.5\,$s \\ \hline
\end{tabular}
\label{tab:polynomials}
\end{center}
\end{table}
}
To train the networks, 124 starting and 124 waiting scenes of uninstructed cyclists were used. The motion types were labeled manually where the beginning of a starting motion $t_\mathrm{s}$ is defined as the first movement and the beginning of waiting $t_\mathrm{w}$ is defined as the last movement of the rear wheel of the bicycle. To train the starting network, trajectories from \mbox{$t_\mathrm{s} - 1\,$s} until the cyclists leave the field of view (FOV) were used and for the waiting network, trajectories between \mbox{$t_\mathrm{w} + 1\,$s} and \mbox{$t_\mathrm{s} - 1\,$s} were used. The networks were trained using the Resilient Backpropagation (RPROP) \cite{Riedmiller.1993} algorithm.
Tab. \ref{tab:polynomials} shows the chosen parameters for the in- and output polynomials of the MLPs. The MLP consists of two hidden layers with 20 neurons in the first and 10 neurons in the second layer. A coarse-to-fine grid search was performed, to find the optimal network structure.
\begin{figure}
\includegraphics[width=0.5\textwidth]{images/mlp_in_output_2.png}
\caption{Example for in- and output of MLP: cyclist head velocity (gray), current time $t_\mathrm{c} = 2.0\,$s (green), input polynomials (black), output polynomials (ochre). }
\label{fig:MLP_IN_OUT}
\end{figure}
\subsection{Second-Level Movement Primitive Detection}
\label{sec:second_level_mp_detection}
Describe movement primitive detection using HAR here.
\paragraph{Human Activity Recognition Techniques}
Classifier trained using HAR goes here.
\paragraph{Convolutional Neural Network}
CNN goes here.
\subsection{Detection of Starting Movements}
\label{subsec:starting_movement_classifcation}
In addition to the \textit{waiting} and \textit{moving} class, we introduce an auxiliary class \textit{starting}, which allows to integrate early movement indicators~\cite{mbHZD+17}, i.e., body movements happening before the beginning of \textit{moving}, indicating future behaviour. The \textit{starting} class is used in the model training and optimization process.
It helps to avoid many false positive \textit{moving} detections, e.g.,
uncertain movements, which may either be classified as \textit{waiting} or \textit{moving} can now be classified as \text{starting}.
Using the \textit{starting} class, we model the movement transition from \textit{waiting} to \textit{moving} as
follows: $P_{waiting}$, $P_{starting}$, and $P_{moving}$ denote the probabilities assigned by the classifiers to the different classes. A phase is labelled as \textit{waiting} if neither the rear wheel of the bicycle is moving, nor is the cyclist performing a movement that leads to a starting movement. The time between the first visible movement of the cyclist that leads to a start and the first movement of the wheel of the bicycle is labeled as \textit{starting}. Finally, the sequence after the first movement of the bicycle wheel is labeled as \textit{moving}.
The \textit{starting} class is optional, since
the labels are defined manually based on the evaluation of camera images.
A sample scene involving the three phases of the starting movement is depicted in Fig.~\ref{fig:ex_net_out}.
The red line represents the probability $P_{moving}$ of the starting movement detection.
In phase I this probability should be close to zero, during phase II the probability should increase, and reach a probability close to one in phase III. A detection is obtained by applying a threshold $s$ on $P_{moving}$. The time of the first \textit{moving} classification is referred as starting
detection time $t_d$.
\begin{figure}[b]
\centering
\includegraphics[width=0.30\textwidth]{images/ex_netout_scene.png}
\vskip 3mm
\caption{Exemplary detection output of one scene with the moving probability $P_{moving}$ (red), the labeled \textit{starting} time (orange), and the labeled \textit{moving} time (purple).}
\vskip -5mm
\label{fig:ex_net_out}
\end{figure}
A starting movement detector has to be robust, i.e., against false positive moving detections, and yet
it has to be fast, i.e., low detection time.
These are two opposing goals resulting in a trade-off which has to be solved.
How this trade-off is solved (i.e., which model parametrization is considered)
depends on the rating of the goal. If the starting movement detection
is used as supplementary information supporting the trajectory forecast~\cite{mbBZD+17}, then allowing
a few false positive detections might be acceptable while in other cases having zero false positive
detections is mandatory.
\subsection{Movement Primitive Detection}
In this section we evaluate and compare different models optimized to detect movement primitives based on smart devices, infrastructure, and cooperation.
For training and evaluation of the classifier a $5$ fold cross-validation over the test subjects is performed. Hence,
training and test subjects are disjunctive. In order to optimize the classifiers hyperparameters, $30\%$ of each training fold are retained and used as the validation set, while the other $70\%$ are used to train the model. Hyperparameters to be optimized are $C$ and $\gamma$ of the SVM as well as the class weight $cl_{w}$ used for training.
\begin{figure}[h]
\includegraphics[clip, trim=0 0.3cm 0 0.3cm, width=0.5\textwidth]{images/start_wait_f1_score_delay.pdf}
\vskip 3mm
\caption{Weighted F1 score and delay of selected models. (Green: smartphone, Magenta: head trajectory
and Yellow: cooperative approach.)}
\label{fig:f1score_delay}
\vskip -6mm
\end{figure}
Fig.~\ref{fig:f1score_delay} shows the weighted F1 score of different models plotted over the delay axis with
various parameter settings on the test set.
The movement primitive classifiers for smart device (S), head trajectory (H), and cooperative movement primitive classifier (C)
are selected ($10$ models each) from the Pareto front of the weighted F1 score and delay evaluated on the validation set.
The reported classification statistics are obtained by evaluation of the selected models on the test set
and without consideration of the probability calibration, i.e. direct class outcome of the SVM.
The elements of the confusion matrix plus the delay of nine selected classifiers ($3$ of each) are denoted in the last five columns of Tab~\ref{tab:motion_primitive_results}. The statistics of the confusion matrix are weighted in the same fashion as the F1 score.
{
\setlength{\abovecaptionskip}{5pt plus 3pt minus 3pt}
\begin{table}[b]
\vskip -4mm
\begin{center}
\caption{Classification results of selected SVM with RBF kernel according to the multi-objective parameter optimization.}
\resizebox{8.4cm}{!}{
\begin{tabular}{| c | c | c | c || c | c | c | c | c | c |}
\hline
ID & $C$ & $\gamma$ & $cl_{w}$ & TP & TN & FP & FN & Delay $[\text{ms}]$\\ \hline \hline
S1 & $2^9$ & $2^{-7}$ & $22$ & $0.994$ & $0.911$ & $0.093$ & $0.006$ & $32$ \\ \hline
S2 & $2^{-1}$ & $2^{-4}$ & $15$ & $0.996$ & $0.918$ & $0.086$ & $0.003$ & $44$ \\ \hline
S3 & $2^{-3}$ & $2^{-4}$ & $7$ & $0.992$ & $0.940$ & $0.063$ & $0.007$ & $106$ \\ \hline \hline
H1 & $2^{-3}$ & $2^{-7}$ & $40$ & $0.996$ & $0.906$ & $0.099$ & $0.004$ & $54$ \\ \hline
H2 & $2^{-5}$ & $2^{-3}$ & $7$ & $0.992$ & $0.969$ & $0.033$ & $0.008$ & $222$ \\ \hline
H3 & $2^{-1}$ & $2^{-5}$ & $5$ & $0.989$ & $0.978$ & $0.023$ & $0.011$ & $246$ \\ \hline \hline
\textbf{C1} & $\boldsymbol{2^{-7}}$ & $\boldsymbol{2^{-3}}$ & $\boldsymbol{7}$ & $0.995$ & $0.964$ & $0.039$ & $0.004$ & $96$ \\ \hline
C2 & $2^{-3}$ & $2^{-1}$ & $27$ & $0.994$ & $0.958$ & $0.045$ & $0.006$ & $94$ \\ \hline
C3 & $2^{1}$ & $2^{-7}$ & $10$ & $0.992$ & $0.978$ & $0.024$ & $0.007$ & $112$ \\ \hline
\end{tabular}}
\label{tab:motion_primitive_results}
\end{center}
\end{table}
}
We observe that classifiers based on the smart device activity data have small delays but
a high false positive rate as compared to the classifiers based on head trajectories.
For the latter the false positive rate is lower, going
along with a higher delay. The three classifiers $S1$ to $S3$ and $H1$ to $H3$ are representatives which range from highly sensitive with
a high false positive rate to rather slow starting detection behavior with a low false positive rate.
A combination of these classifiers in pairs of two are used to create the cooperative classifier as described in Sec.~\ref{sec:cooperative_mp_pred}.
The hyperparameter of the cooperative classifier are again optimized using a grid search over the parameter space.
Selected classifiers are depicted in Fig.~\ref{fig:f1score_delay} and Tab.~\ref{tab:motion_primitive_results}.
The cooperative approach trades off a low false positive rate while retaining a fast starting detection time, i.e. low delay ($<100\,$ms).
The selected cooperative classifier $C1$ for further evaluation of the trajectory forecast
is depicted in bold in Tab.~\ref{tab:motion_primitive_results}. It is a combination of a sensitive smart device classifier $S1$, which
reacts upon the smallest movements (e.g. first instance of beginning pedaling movement), and a conservative head trajectory classifier $H3$. It detects the starting intention after $96\,$ms with an accuracy of $90\%$ and the accuracy reaches $95\,\%$ after $212\,$ms.
\section{Experimental results}
\label{sec_ResultsOutline}
\subsection{Wearing Location} \label{sec:eval_wearing_location}
\begin{figure*}[ht!]
\centering
\vskip -4mm
\begin{subfigure}[t]{0.49\linewidth}
\centering
\includegraphics[width=0.9\columnwidth]{"images/pareto_fronts_xgb_agnostic"}
\vskip -2mm
\caption{Pareto frontiers of \textit{location-agnostic} XGBoost.}
\label{fig:xgb_all_pareto_frontier}
\end{subfigure}
\begin{subfigure}[t]{0.49\linewidth}
\centering
\includegraphics[width=0.85\columnwidth]{"images/pareto_fronts_lin_svm_agnostic"}
\vskip -2mm
\caption{Pareto frontiers of \textit{location-agnostic} linear SVM.}
\label{fig:lin_svm_all_pareto_frontier}
\end{subfigure}
\vskip 4mm
\begin{subfigure}[t]{0.49\linewidth}
\centering
\includegraphics[width=0.9\columnwidth]{"images/pareto_fronts_xgb_device_specific"}
\vskip -2mm
\caption{Pareto frontiers of \textit{location-specific} XGBoost.}
\label{fig:xgb_pareto_frontier}
\end{subfigure}
\begin{subfigure}[t]{0.49\linewidth}
\centering
\includegraphics[width=0.85\columnwidth]{"images/pareto_fronts_svm_device_specific"}
\vskip -2mm
\caption{Pareto frontiers of \textit{location-specific} linear SVM.}
\label{fig:lin_svm_pareto_frontier}
\end{subfigure}
\vskip 4mm
\begin{subfigure}[t]{0.49\linewidth}
\centering
\includegraphics[width=0.9\columnwidth]{"images/pareto_fronts_xgb_device_specific_short"}
\vskip -2mm
\caption{Pareto frontiers of \textit{location-specific} XGBoost for shortened starting scenes.}
\label{fig:short_xgb_all_pareto_frontier}
\end{subfigure}
\begin{subfigure}[t]{0.49\linewidth}
\centering
\includegraphics[width=0.9\columnwidth]{"images/pareto_fronts_lin_svm_device_specific_short"}
\vskip -2mm
\caption{Pareto frontiers of \textit{location-specific} linear SVM for shortened starting scenes.}
\label{fig:short_lin_svm_all_pareto_frontier}
\end{subfigure}
\vskip 5mm
\caption{Pareto frontiers of $F_1$-scores and mean detection times for XGBoost and linear SVM classifiers
evaluated for different wearing locations (Blue dot: trouser pocket, Red square: backpack, Cyan star: jacket pocket/ chest,
Gray cross: Bicycle rack, and Yellow triangle: Average over all wearing locations using the respective \textit{location-agnostic} classifier).
The two upper figures show the evaluation \textit{location-agnostic} classifiers and the two in the middle show the evaluation of the \textit{location-specific} classifiers. The lower two show the evaluation of the \textit{location-specific} classifiers on the shortened starting scenes.
}
\label{fig:res_wearing_location}
\vskip -2mm
\end{figure*}
In this section, we evaluate and compare different models and parametrizations to detect cyclist starting movements.
Moreover, we evaluate the detection performance with respect to four different wearing locations.
We trained separate linear SVM and XGBoost classifiers on data originating from devices worn at the four different
wearing locations. These are referred as \textit{location-specific} classifiers. Additionally, we trained
classifiers incorporating data from all four wearing locations. These are referred to as \textit{location-agnostic} classifiers.
To create the Pareto frontier, we randomly sampled and evaluated 250 different parameter combinations for each \textit{location-specific} and \textit{location-agnostic} classifier. For the XGBoost classifier we considered
the number of trees (50,~100,~200,~300,~500,~and~700), the maximum tree depth (between 3 and 10),
and the learning rate (between 0.01 and 0.2) as parameters. For the linear SVM we only considered the penalty term (between $2^{-8}$ and $2^{8}$).
In order to speed up the evaluation, we performed a random subsampling.
We experimentally determined that at least approximately 7500 (30000) samples are required to achieve reasonable results
for a \textit{location-specific} (\textit{location-agnostic}) classifier.
We considered the class weighting as an additional parameter. We integrated this into our parameter sweep
by means of different class subsample sizes. For a \textit{location-specific} classifier we considered sample sizes of 2500 to 15000 (in steps of 2500) for each class. For a \textit{location-agnostic} classifier we sampled 10000 to 22500 (in steps of 2500) samples from each class.
The Pareto frontiers of $F_1$-score and mean detection time for the linear SVM and the XGBoost classifiers resulting from our random parameter sweep are depicted in Fig.~\ref{fig:res_wearing_location}.
The results of the \textit{location-agnostic} classifiers are depicted in Fig.~\ref{fig:xgb_all_pareto_frontier} and
\ref{fig:lin_svm_all_pareto_frontier}.
We see that the linear SVM and the XGBoost classifiers averaged over all wearing locations perform very similar, i.e., the Pareto frontiers are close. But neither of both model types reaches an $F_1$-score of
one. Furthermore, we observe a strong dependency on the wearing location. We can see that the XGBoost \textit{location-agnostic} classifiers for smart devices located
in the trouser pocket and backpack show the best solutions concerning the $F_1$-score as well as the mean detection time.
For the linear SVM this increase in detection performance is even more pronounced regarding the trouser pocket wearing location (cf. Fig.\ref{fig:lin_svm_all_pareto_frontier}).
The results for the \textit{location-specific} classifiers are depicted in Figs.~\ref{fig:xgb_pareto_frontier} and \ref{fig:lin_svm_pareto_frontier}. For the XGBoost classifiers we observe an improvement for the trouser pocket and backpack wearing locations. The best XGBoost classifier with a mean detection time of under half a second has an $F_1$-score of $94\%$ and a mean detection time of \SI{0.34}{\second}.
The detection results of the device mounted to the rack of the bicycle are slightly better while we do not observe any
noticeable change for the device worn in the jacket pocket.
We measure no improvements for the Pareto optimal solutions of the \textit{location-specific} linear SVM classifier
as can be seen in Fig.~\ref{fig:lin_svm_pareto_frontier}.
Instead, the mean detection times of the solutions on the Pareto frontiers are higher than those of the linear SVM \textit{location-agnostic} classifiers.
Only for the device in the trouser pocket we observe an increased $F_1$-score.
While for the \textit{location-agnostic} classifiers the XGBoost and linear SVM show comparable results.
In the case of \textit{location-specific} classifiers the XGBoost outperforms the linear SVM.
One major challenge that we observed with all classifiers is that many false positives are due to miss-classifications at the
beginning of the waiting phase. Movements, such as getting off the bike or adjustment of pedals, produce a similar sensor pattern resulting in false positive classification.
In order to investigate the starting detection performance more closely, we additionally examined shortened starting scenes.
We cut the scenes to \SI{7}{\second} before the labeled starting movement. This removes the remaining motion originating from the stopping movement, e.g., getting off the bike. The Pareto frontiers of the shortened scenes of the XGBoost and linear SVM \textit{location-specific} classifiers are depicted in Figs.\ref{fig:short_xgb_all_pareto_frontier} and \ref{fig:short_lin_svm_all_pareto_frontier}.
We observe an increased $F_1$-score, i.e., the XGBoost classifier reaches an $F_1$-score of 97.6\% at mean detection time of \SI{0.35}{\second}.
\subsection{Feature Selection}
\begin{figure*}[ht!]
\centering
\begin{subfigure}[t]{0.49\linewidth}
\includegraphics[width=0.9\columnwidth]{"images/dv1_koehler_plot"}
\caption{\textit{Location-specific} XGBoost.}
\end{subfigure}
\begin{subfigure}[t]{0.49\linewidth}
\includegraphics[width=0.9\columnwidth]{"images/multiple_koehler_plot"}
\caption{\textit{Location-agnostic} XGBoost.}
\end{subfigure}
\vskip 5mm
\caption{$F_1$-score (blue), precision (yellow) and mean detection time (green) over the probability threshold of all starting scenes for the selected device \textit{location-specific} (left) and \textit{location-agnostic} (right) classifier evaluated for the smart device worn in the trouser pocket. The dashed lines indicates the reference evaluation involving all wearing positions.}
\label{fig:koehler_plot}
\vskip -2mm
\end{figure*}
\begin{figure*}[h]
\centering
\begin{subfigure}[t]{0.49\linewidth}
\centering
\includegraphics[width=0.8\linewidth]{"images/delay_kde_histogram_dv1_short"}
\vskip 2mm
\caption{\textit{Location-specific} XGBoost.}
\label{fig:location_specific_delay_histogram}
\end{subfigure}
\begin{subfigure}[t]{0.49\linewidth}
\centering
\includegraphics[width=0.8\linewidth]{"images/delay_kde_histogram_dvall_combined"}
\vskip 2mm
\caption{\textit{Location-agnostic} XGBoost.}
\label{fig:location_agnostic_delay_histogram}
\end{subfigure}
\vskip 5mm
\caption{Histograms of mean detection times. The histograms are smoothed by means of a kernel density estimation with a Gaussian kernel of bandwidth 70 (Blue: trouser pocket, Red: all wearing positions).
}
\label{fig:delay_histogram}
\vskip -2mm
\end{figure*}
In this section, we evaluate the two-stage feature selection performed for the distinct XGBoost classifiers trained
for different wearing locations.
Moreover, an interpretation of the selected features with respect
to the wearing location is given. Since we performed a ten-fold cross-validation, we considered the
number of times a particular feature is selected as evaluation criteria.
For the ease of understandability, we grouped different window sizes of the same features in our evaluation.
For the smart devices located in the front trouser pocket only a single feature was selected, i.e.,
the gyroscope's energy in the ground plane (with varying window sizes in different folds).
It captures the pedalling and push off movement, which happens in the early starting phase,
very well. The most relevant feature for the smart device located at the bicycle's rack
is the linear accelerometer's energy in the ground plane. This feature captures the energy of the movement in driving direction.
For the device located in the jacket pocket at the chest the most relevant features are the gyroscope's energy in the ground plane and the linear accelerometer's energy along the z-axis. The first feature captures the pedalling movement while the second feature additionally measures the downward movement of the upper body just before starting~\cite{mbHZD+17} as well as the pushing away motion to start.
The classifiers based on the smart device in the backpack use the largest number of features.
The most important features of those are the linear accelerometer's minimum, energy, and the second polynomial degree in the ground plane as well as the gyroscope's energy, minimum, and maximum in the ground plane.
The classifiers that are based on data originating from all wearing locations are heavily based on the energy of the linear accelerometer along the z-axis, the gyroscope's energy along the z-axis, and the residual of the linear accelerometer's DFT polynomial approximation along the z-axis.
It is noticeable that this classifier mainly uses features which are based on the z-axis.
Many typical starting movements (e.g., pedalling, pushing away motion) are well capture by
features computed for the accelerometers z-axis. The features based on the gyroscopes z-axis are best
explained by unsteady, swaying bicycle movements at starting.
\subsection{Starting Movement Detection}
We focus on the detailed evaluation of two selected starting movement detectors.
Moreover, we show the influence of the auxiliary class used to train the classifiers.
For evaluation, we compare a \textit{location-agnostic} with a \textit{location-specific} classifier.
We restrict ourselves to XGBoost based detectors since they outperform linear SVM based detectors.
The detectors are chosen from the Pareto frontiers (cf. Fig.~\ref{fig:res_wearing_location}). As selection criteria, we require a mean detection time of at most \SI{0.4}{\second}.
This is motivated by the scenario mentioned in the introduction. But here, we additionally consider communication
delay of modern Car2X-communication system (e.g., IEEE 802.11p) of approximately \SI{0.1}{\second}.
For the \textit{location-agnostic} classifier, we select the XGBoost classifier with an $F_1$-score of 67\% and a mean detection time of \SI{0.33}{\second}. The number of trees is 100, the maximal tree depth is 8 and the learning rate is 0.18.
We compare this detector with the \textit{location-specific} XGBoost classifier for devices worn in the trouser pocket ($F_1$-score of $94\%$ and mean detection time of \SI{0.34}{\second}). The number of trees is 300, the maximal tree depth is 3 and the learning rate is 0.11.
The $F_1$-score, precision and mean detection times for different probability thresholds on the \textit{moving} class are depicted in
Fig.\ref{fig:koehler_plot}.
We see that the optimal probability threshold for the \textit{location-specific} classifier is around 0.5.
For the \textit{location-agnostic} classifier, the $F_1$-score can be improved by applying a higher probability threshold.
We observe that the $F_1$-score scales approximately linear with the
mean detection time. We also see that an $F_1$-score of one is not reached, not even when the probability threshold is close to
one. Far from it, when approaching a threshold of one the $F_1$-score deteriorates for both classifiers, i.e.,
starting is no longer detected.
Since the classifiers are calibrated, it can be concluded that the \textit{waiting} and \textit{moving} classes
cannot be entirely separated based on the selected features, i.e., the classes posses a particular level of impurity.
Because the \textit{location-specific} XGBoost classifier for the trouser pocket uses only a single feature, this
effect gets even more pronounced. For the \textit{location-agnostic} classifier it is more moderate.
In the following, we investigate the detection times. We examined the distribution of detection times of the shortened starting
scenes. The histograms of detection times for the \textit{location-specific} and \textit{location-agnostic} XGBoost classifier are depicted in Figs.~\ref{fig:location_specific_delay_histogram}~and~\ref{fig:location_agnostic_delay_histogram}. To obtain a probability distribution, the histograms are smoothed with a kernel density estimation using a Gaussian kernel.
We see that for both, the \textit{location-specific} and \textit{location-agnostic} XGBoost classifiers, some starting movements
are detected more than \SI{1}{\second} before the first movement of the bicycle wheel. The remaining detection times can be approximately described by a Gaussian centered at the mean detection time of the respective detectors. Whereas the distribution of the \textit{location-specific} classifier is narrower and the one of the \textit{location-agnostic} classifier is heavy-tailed.
Moreover, we also investigate the effectiveness of additional output smoothing.
We found that the effects are negligible, i.e., the $F_1$-score but also the detection time is partly increased.
We examine it as a way to fine-tune the detectors.
Figure~\ref{fig:selected_sample_scenes} shows the detection results of
four selected sample starting scenes. It depicts the results
of the \textit{location-specific} XGBoost classifier evaluated for the device worn in the trouser pocket. In Fig.~\ref{fig:pred_dv1_131}
a scene with a negative starting detection time is shown.
The starting movement is detected \SI{1.3}{\second} before
the first movement of the bicycle wheel.
A starting movement with a delayed detection is depicted in
Fig.~\ref{fig:pred_dv1_121}. We notice that the \textit{starting}
phase is detected even before the labeled \textit{starting} movement.
\begin{figure}[H]
\vskip -4mm
\begin{subfigure}{\linewidth}
\centering
\includegraphics[width=0.9\columnwidth]{"images/pred_dv1_131"}
\vskip -3mm
\caption{Starting movement with negative mean detection time.}
\label{fig:pred_dv1_131}
\end{subfigure}
\begin{subfigure}{\linewidth}
\centering
\includegraphics[width=0.9\columnwidth]{"images/pred_dv1_121"}
\vskip -3mm
\caption{Starting movement detection with delayed detection.}
\label{fig:pred_dv1_121}
\end{subfigure}
\begin{subfigure}{\linewidth}
\centering
\includegraphics[width=0.9\linewidth]{"images/pred_dv1_140"}
\vskip -3mm
\caption{False positive starting movement detection.}
\label{fig:pred_dv1_140}
\end{subfigure}
\begin{subfigure}{\linewidth}
\centering
\includegraphics[width=0.9\columnwidth]{"images/pred_dv1_105"}
\vskip -3mm
\caption{Usage of auxiliary class to avoid false positive detections.}
\label{fig:pred_dv1_105}
\end{subfigure}
\vskip 4mm
\caption{\textit{Location-specific} XGBoost detection results of selected starting scenes evaluated for smart device worn in the trouser pocket. Blue, green, and red denote the \textit{waiting}, \textit{starting}, and \textit{moving} probability. The dashed black line
indicates the selected class. The vertical lines denote the beginning of labeled \textit{starting} (orange) and \textit{moving} (purple) class.}
\label{fig:selected_sample_scenes}
\end{figure}
Regarding the \textit{moving} probability, we observe a sharp transition from the \textit{starting} to the \textit{moving} phase.
Ego-movement, e.g., adjustment of pedals or swaying from one side to the other, can lead to false positive starting detections.
Figure.~\ref{fig:pred_dv1_140} shows such a scene containing a false positive detection. Figure.~\ref{fig:pred_dv1_105} shows the usage of the auxiliary \textit{starting} class to avoid false positive detections. It acts as a buffer zone between the \textit{waiting} and \textit{moving} classes.
|
\section{Introduction}
\label{intro}
A key prerequisite to develop applications of molecular nanotechnology is the ability to selectively and reversibly change measurable properties of single molecules. Materials and techniques which enable this are important for molecular-level optoelectronic device integration,\cite{Cuevas2010} molecular circuits,~\cite{Mativetsky2008} ultrasensitive molecular sensors,~\cite{Joshi2014} and plasmonic reaction control at nano-structured interfaces.~\cite{Xie2018}
Single molecules that can be reversibly switched between two or more distinguishable states, so-called molecular switches, provide an interesting playground for molecular device design.~\cite{Feringa} The changes between these states can be in the form of discrete modifications to the molecular conformation, the chemical state, the magnetic state, or other measurable molecular properties. Embedded in a matrix~\cite{Ikeda1995}, in porous metal-organic frameworks~\cite{Wang2015} or adsorbed at well-defined surfaces \cite{Katsonis2006}, such molecules are envisioned to be switchable individually or in domains, and the corresponding changes have to be detectable in terms of simple electric or spectroscopic signatures. In practice, this is an intricate task, because the function of molecular switches depends on a delicate balance of different intra- and intermolecular interactions, but also on the interactions with the environment. Conducting electrodes, such as metal substrates are of special interest as support due to the direct contact of the molecule to a conductor and the prospects of controlling function via an applied voltage or via introducing novel functionality due to electronic coupling. However, the strong molecule-metal coupling can readily lead to a loss of function -- if the intermediate electronic states are quenched, the reaction pathway is hindered, or more fundamentally, the basic stability of the equilibrium geometries changes.~\cite{Maurer2012,Titov2015}
Although clear-cut design strategies are still lacking, a large variety of molecular switches is currently investigated for their functionality. Explorative substrate-variation and molecule functionalisation have led to many switchable metal-organic interfaces including molecules that undergo changes due to chemical reactions~\cite{Schulze2012}, conformational changes~\cite{Comstock2005}, switching between chemisorbed and physisorbed states~\cite{Liu2013}, or magnetic spin transitions \cite{Venkataramani2011, Wackerlin2010}. The means to induce such switching range from light irradiation to thermal activation~\cite{Hagen2007} and inelastic electron tunnelling~\cite{Choi2006}. These fundamental studies have led to fascinating prototype applications such as molecular motors or machines~\cite{Browne2006,Chiang2012, Coskun2012} and optically contractable polymers~\cite{Hugel2002a}.
Azobenzene (Ab) and its derivatives represent one of the best studied classes of molecular switches. Its ability to change between an E and a Z conformation around the central N-N double bond (cf. Fig. \ref{fig-params}) upon light irradiation depends not only on an efficient excited-state reaction channel, but more fundamentally on the sheer fact that two meta-stable states coexist in the ground state and are fairly stable at the experimental conditions. Owing to time-resolved photoexcitation experiments and quantum dynamics simulations, the electronic and atomistic details of the gas phase photoswitching mechanism are, by now, quite well understood~\cite{Pederzoli2011,Gamez2012, weingart2011}. For a review of the literature see Ref. \cite{Bandara2012} or \cite{Maurer_Thesis2014} page 37.
\begin{figure}
\centering\includegraphics[width=3.3in]{Fig1.pdf}
\caption{\label{fig-params} (a) Schematic view of the relevant degrees of freedom for the on-surface isomerisation between E and Z-Ab on a Ag(111) surface via a transition state (TS) along the dihedral rotation of the central di-nitrogen bond. N atoms are shown in blue, C atoms in grey, and H atoms in white. The Ag atoms are shown as van-der-Waals spheres. Relevant geometry parameters are defined: $z$ refers to the vertical adsorption height of the N atoms, $\omega$ refers to the dihedral angle around the di-nitrogen bond (CNNC), $\alpha$ refers to the CNN bending angles. The bottom panels show equilibrium gas-phase structures of isolated (b) E-Azobenzene (E-Ab), (c) Z-Ab, (d) E-3,3'5,5'-tetra-\emph{tert}-butyl-Azobenzene (E-TBA), and (e) Z-TBA.}
\end{figure}
The persistence or loss of switching function of Ab and its derivatives directly adsorbed to noble metal surfaces, such as close-packed facets of coinage metals, has been intensively studied \cite{Comstock2005,Kirakosian2005,Henzl2006,alemani2008, Wolf2009, Henzl2012,Morgenstern2011}; no light-induced switching of pure Ab on Ag(111) or Au(111) has been observed to date.~\cite{comstock2007} The lack of photo-induced function has been attributed to three main factors: overly strong coupling to the surface, quenching of the excited states involved in the isomerisation, and steric hindrance due to strong dispersion interactions between the phenyl rings of the molecule and the underlying surface.~\cite{mcnellis2010a} However, successful switching of Ab on Au(111) via resonant electron tunnelling through a Scanning Tunnelling Microscopy (STM) tip~\cite{Choi2006}, suggests that there are additional factors at play when comparing different substrates. In all experimental studies, successful switching has been verified either by significant changes to spectral features, such as Two-Photon-Photoemission (2PPE) resonances, or by visual inspection of STM topograph changes.
Molecular functionalisation strategies to decouple the photochromic azo-moiety from the surface follow the basic ideas of either geometrically decoupling the central di-nitrogen bridge from the surface with spacer groups~\cite{Alemani2006,comstock2007,Hagen2007}, introducing a modified switching process via strongly chemically active functional groups~\cite{Henzl2006, Henzl2012,Morgenstern2011}, or via combination of Ab units into sterically constrained tripods.~\cite{Scheil2016} Comstock \emph{et al.} tested the former by analysing the switching behaviour of Ab and Ab substituted with \emph{tert}.-butyl groups at the two phenyl \emph{para} positions and the four \emph{meta} positions (DBA and TBA, respectively) when adsorbed to Au(111).~\cite{comstock2007} The authors were able to detect a switched state of TBA after 1 hour of irradiation with UV light, whereas no isomerisation was observed for Ab and DBA. The reason for a successful photo-induced isomerisation of TBA in contrast to DBA and Ab has been related to the electronic decoupling of the central nitrogen atoms, therewith reduced substrate effects on the excited-state lifetime, and the optical absorption properties. We note that all three molecules represent functioning molecular switches in gas phase or solvent and could be switched by an STM tip when adsorbed at a Au(111) surface~\cite{Choi2006,Alemani2006,comstock2007}.
The isomerisation mechanism of metal-mounted TBA is believed to be strongly modified compared to the gas phase. The change in electronic structure of the adsorbate can be verified by STM topographs~\cite{comstock2008}, 2PPE signatures of the molecular resonances~\cite{Hagen2008}, and changes in the work function.~\cite{Hagen2007} The photo-switched state can be identified as the molecular Z-TBA state by structural predictions from Angular Resolved Near-Edge X-Ray Absorption Fine-Structure (NEXAFS) experiments~\cite{Schmidt2010}. In contrast to the gas phase or solvent situation, a photo-induced back-reaction from this Z-TBA state to E-TBA is much less efficient and does not happen in the visible regime, but rather with photon energies in the same energy regime as the E-to-Z transition.~\cite{comstock2008} A corresponding drop in switching efficiency can thus already be understood in terms of simultaneously induced switching in both directions, as is supported by the reduced E/Z ratio at the photostationary state. Wolf and Tegeder have proposed a mechanism of cationic resonance formation in E-to-Z photo-isomerisation of TBA that is in agreement with experimental observations.~\cite{Wolf2009} Light irradiation generates electron-hole pair excitations in the metal. The generated holes rapidly diffuse to the upper edge of the metal d-bands. Energetic overlap between the d-bands and occupied molecular frontier orbitals, such as the highest occupied molecular orbital (HOMO) will enable a transient cation formation in the molecule, which subsequently drives the isomerisation.~\cite{Hagen2008, Tegeder2007} The prevalent assumption for the ensuing molecular motion is that, although the coupling to the metallic degrees of freedom will limit the lifetime of this state to tens to hundreds of femtoseconds,~\cite{campillo2000} the kinetic energy gained by the motion on the excited-state PES might suffice to enable the transition to the Z-Ab state. This is the so-called Menzel-Gomer-Redhead model~\cite{Menzel1964,Redhead1964}, which has been proposed in the context of photo-induced desorption processes~\cite{Zhou1991, Hertel1995, Guo1999}. Not much is known about the atomic degrees of freedom involved in surface-mounted TBA switching though. On the basis of STM topographs and chiral mapping, it has been argued that the isomerisation must follow a predominantly planar inversion of the central CNN bending angle $\alpha$ (\emph{cf.} definition in Fig.~\ref{fig-params}) rather than the rotational mechanism that is believed to be dominant in gas phase.~\cite{Comstock2010}
First-principles modelling of such light-induced molecular adsorbate reactions and involved excited states can in principle shed more light into the dynamics and the mechanism. Presently, this is extremely challenging. This relates to the large system size, the concomitant high computational expense, and the difficulty of describing electronic charge-transfer excitations at the molecule-metal interface. An important prerequisite for an \emph{ab initio} simulation of the excited-state dynamics involved in photo-isomerisation is a detailed understanding of the adsorption geometries of the ground-state equilibrium structures and the ability to access electronic excited states of molecular adsorbates at metal surfaces. To this end, the recent emergence of predictive-quality dispersion-corrected Density Functional approximations for metal-organic interfaces~\cite{Ruiz2012, Maurer2016Review} and efficient, approximate methods to calculate electronic excited states, such as linear expansion Delta-Self-Consistent-Field DFT ($\Delta$SCF-DFT),~\cite{Maurer2013,gavnholt2008} is providing the necessary tools to study the fundamental structural and electronic parameters that control light- or electron-driven Ab isomerisation at metal surfaces in more detail.
In this work, we use these tools to study the structural and electronic properties at the metal-organic interface that govern the switching ability of metal-adsorbed functional molecules. We do this for the case of coinage-metal adsorbed Ab, for which we previously were able to rationalise the loss of function on Ag(111) as a loss of ground state bi-stability~\cite{Maurer2012}. By analysing the effects of substrate variation and molecule functionalisation for coinage-metal-adsorbed Ab and TBA on the ground- and excited-state potential energy landscapes of the adsorbed species, we connect how chemical and electrostatic molecule-surface interactions shape these energy landscapes and provide the basis for molecular switching or the lack thereof.
\section{Computational Methods}
\label{methods}
\subsection{Density Functional Theory calculations}
All DFT calculations have been performed with the pseudopotential plane wave code CASTEP version 6.0.1~\cite{Clark2005,Payne1992} using the legacy MaterialsStudio ultrasoft pseudopotentials (USPPs)~\cite{Vanderbilt1990}. The exchange-correlation functional due to Perdew, Burke, and Ernzerhof (PBE)~\cite{Perdew1996} has been used throughout. We account for dispersion interactions using the screened Tkatchenko-Scheffler pairwise dispersion correction method (vdW$^{\mathrm{surf}}$).~\cite{Ruiz2012} This method describes the screening of van-der-Waals interactions in the metal substrate by renormalisation of the pairwise C$^6$ coefficients of the metal. It has, in the past, been shown to provide accurate adsorption geometries at a moderate overestimation of the adsorption energies.~\cite{Mercurio2013, Maurer2016, Maurer2016Review} Higher-order dispersion correction techniques, such as the Many Body Dispersion (MBD) method exist to correct for this overbinding,~\cite{Tkatchenko2012,Maurer2015}, but have not been available for geometry optimisations at the time of this study. For all structures, the dispersion interactions between periodic images of the adsorbed molecules have been switched off to simulate a low-coverage situation.
The employed computational set-up was previously described with details on careful convergence tests~\cite{Maurer2012,McNellis2009}. In short, all calculations were performed on (111)-oriented bulk-truncated 4-layer surface slabs of Cu, Ag, and Au with 350~eV or 450~eV plane wave cutoff for Azobenzene and 3,3',5,5'-tetra-\emph{tert.}-butyl-Azobenzene (TBA), respectively. The top-most two layers of the substrate slabs were allowed to relax. The PBE optimised lattice constants for Cu, Ag, and Au, are 3.62~\AA{}, 4.14~\AA{}, and 4.19~\AA{}, respectively. The vacuum region above the surface slab was chosen to exceed 20~\AA{} in all cases. We obtained optimized geometries using the delocalized internal coordinate optimization algorithm in CASTEP~\cite{Andzelm2001} with a maximum atomic force threshold of 25~meV/\AA{}. Minimum energy paths were calculated by constrained optimization where the value of the central dihedral angle $\omega$ was fixed while all other degrees of freedom of the molecule were allowed to relax. The transition states reported in Table 1 have been calculated with the quadratic synchronous transit method.~\cite{Govind2003}
The results in Fig.~\ref{fig-results-ab-rotation} have been obtained for (6$\times$4), (6$\times$3), (6$\times$3) surface unit cells containing a single Ab molecule on Cu, Ag, and Au, respectively.
TBA molecules have been studied in (6$\times$4) and (6$\times$5) surface unit cells on Ag and Au, respectively. Calculations of the gas phase molecules have been performed in 40$\times$40$\times$40~\AA{} and 50$\times$50$\times$50~\AA{} rectangular supercells for Ab and TBA, respectively, with the electronic structure calculated only at the $\Gamma$-point of the first Brillouin zone.
\subsection{Modelling molecule-surface charge-transfer excitations with linear expansion $\Delta$-Self-Consistent-Field-DFT (le$\Delta$SCF-DFT) }
The excited state results described in section \ref{results-excited-state} have been obtained with the linear expansion $\Delta$-Self-Consistent-Field DFT method (le$\Delta$SCF-DFT), which we have implemented in CASTEP 6.0.1 and which is available in the latest CASTEP 18.1 release.~\cite{Maurer2013} Conventional $\Delta$SCF-DFT, where the occupation of a single Kohn-Sham (KS) state is constrained to model an excited state, would not be able to correctly capture the charge-transfer excitation between an adsorbed molecule and the metal surface in the presence of electronic hybridisation between the two. The le$\Delta$SCF-DFT method addresses this issue by construction of a resonance orbital. This is achieved via projection of the wave function of a molecular orbital of the free molecule $\ket{\phi_c}$ onto the KS states of the surface slab containing the molecule:
\begin{center}\begin{eqnarray}
\ket{\tilde{\psi_c^{\mathbf{k}}}} = \sum_i^{\mathrm{states}} \ket{\psi_i^{\mathbf{k}}} \braket{\psi_i^{\mathbf{k}}|\phi_c}
\end{eqnarray} \end{center}
Upon construction of this resonance state $\ket{\tilde{\psi_c^{\mathbf{k}}}}$, all remaining KS states are orthonormalised with respect to it. The orbital occupations are then distributed and constrained to model the envisioned excitation. This procedure is repeated during each consecutive Self-Consistent-Field (SCF) step until convergence is reached.
In Ref.~\cite{Maurer2013}, we have shown that this method can be used to model intramolecular excitations of electrons from occupied to unoccupied adsorbate states, such as the Ab $\pi\rightarrow\pi^*$ (HOMO$\rightarrow$LUMO or S2) and the n$\rightarrow\pi^*$ (HOMO-1$\rightarrow$LUMO or S2) excitations. However, also charge-transfer excitations between the molecule and the surface can be modelled. In the case of a hole transfer from the substrate to the molecule, a so-called cationic resonance (HOMO$\rightarrow\epsilon_F$ or $h^{+}_{\mathrm{HOMO}}$), we constrain the occupation of the HOMO to be one electron less than in the ground state. The missing electron is balanced by adjusting the Fermi energy accordingly, which has a minimal effect on the electronic structure in the case of a large metal surface slab. In the case of an electron transfer from the substrate to the molecule, a so-called anionic resonance ($\epsilon_F\rightarrow$LUMO or $e^{-}_{\mathrm{LUMO}}$), one electron is taken from the Fermi level and added to the LUMO of the molecule.
\begin{figure}
\centering\includegraphics[width=3.3in]{Fig2.pdf}
\caption[Binding energy curve and excited state energies of Azobenzene on Ag(111)]{\label{fig-ledscf-bindingcurve} Ab adsorbed to Ag(111): Relative energy with respect to the ground-state equilibrium adsorption geometry as a function of the vertical adsorption height. Points at 10.5~\AA{} distance represent a molecular free-standing overlayer. Shown are the ground state energy (black curve), the first (n$\rightarrow\pi^*$, red curve), and second ($\pi\rightarrow\pi^*$, blue curve) intramolecular excited states as well as charge-transfer excitations, where an electron is added to the LUMO of the molecule (e$^-_{\mathrm{LUMO}}$, purple curve) or removed from the HOMO of the molecule (h$^+_{\mathrm{HOMO}}$, green curve). The small model in the inset on the bottom right depicts the dissociation direction orthogonal to the surface.}
\end{figure}
Figure~\ref{fig-ledscf-bindingcurve} presents the ground- and excited-state energies of E-Ab on Ag(111) as a function of the vertical distance from the surface. The ground-state energetics and geometry are calculated with DFT+vdw$^{\mathrm{surf}}$. At every geometry four different le$\Delta$SCF-DFT calculations have been calculated to yield the pictured excited-state energies; two of them are charge-neutral intramolecular excitations, the other two are molecule-surface charge-transfer excitations. Far away from the surface, at distances above 10~\AA{}, the intramolecular excitations approach the limit of the isolated molecule (gas phase S1(n$\rightarrow\pi^*$): 2.27~eV , S2($\pi\rightarrow\pi^*$): 2.75~eV ).~\cite{Maurer2011} More precisely, the values approach the limit of an infinite free-standing molecular layer in the periodic arrangement of the employed supercell geometry (S1: 2.21~eV, S2: 2.24~eV). The corresponding substrate-mediated excitations, h$^+_{\mathrm{HOMO}}$ and e$^-_{\mathrm{LUMO}}$, should, in the limit of infinite molecule-surface separation, approach the corresponding Ionization Potential (IP) and Electron Affinity (EA) of the molecule as predicted by the PBE functional (IP:~7.82~eV, EA:~1.06~eV \cite{McNellis2009a}). The corresponding value for the IP above 6~\AA\ distance from the surface is the difference between the h$^+_{\mathrm{HOMO}}$ state and the ground state (4.37~eV). The value of the EA can be calculated as the difference of the work function (including the potential drop due to the free standing overlayer this amounts to 4.21~eV \cite{McNellis2009a}) and the e$^-_{\mathrm{LUMO}}$ state and thereby amounts to 2.36~eV. The substrate-mediated excitation energies are very sensitive to the distance from the substrate and to the interactions with the neighbouring Ab images, as well as the periodic image of the substrate above~\cite{comment1}. This sensitivity stems from the excess charge on the molecule interacting via long-range Coulomb forces with the electron density of the substrate and the periodic images of the molecule.
When approaching the surface, the intramolecular excitation energies do not change dramatically, whereas e$^-_{\mathrm{LUMO}}$ and h$^+_{\mathrm{HOMO}}$ strongly shift. Close to the equilibrium geometry the anionic resonance state is only 0.38~eV above the ground state and the cationic resonance state is at 2.52~eV above the ground state. This arises from the Coulomb interaction between the localized charge on the molecule and the substrate, which steadily builds up an image charge as the molecule approaches. This difference in the renormalization behaviour of intramolecular excitations and charge-transfer excitations is well known. It can be described at the level of many-body perturbation theory~\cite{Neaton2006, Thygesen2009, Garcia-Lastra2011}, whereas linear response time-dependent DFT based on semi-local functionals fails to capture this effect. It is remarkable that by employing such an approximate scheme as le$\Delta$SCF, the main physical effects of substrate polarization and excited-state renormalization can be qualitatively captured. We therefore feel encouraged to apply this method to study the excited state landscape of charge-transfer states during the Ab and TBA on-surface isomerisation in section~\ref{results-excited-state}.
\section{Results}
\label{results}
\subsection{The role of substrate electronegativity}
\label{results-azobenzene}
The most basic prerequisite to retain the isomerisation ability of Ab-based molecular switches upon metal-surface-adsorption is that the interaction with the substrate does not destroy the existence of the two meta-stable conformers in the ground state, namely E and Z-Ab (see Fig.~\ref{fig-params}). Strong reduction or removal of the ground-state barrier or changes in the relative stability of these two states would affect the thermal stability of a Z-Ab state. As such, any viable metal substrate to support functioning metal-mounted molecular switches needs to have a weak interaction with the molecule that, ideally, affects all molecular conformations similarly as not to modify the energetic landscape along the reaction coordinate of isomerisation. Unreactive noble metals such as Ag(111) or Au(111) therefore come to mind as natural substrate choices.
In previous work, we have shown that Ab adsorption on these two surfaces does, in fact, still modify the energetic landscape along one of the three traditionally discussed pathways for the isomerisation reaction.~\cite{Maurer2012} These relevant reaction coordinates include the rotation around the central dihedral angle $\omega$, an inversion around one of the two central bending angles $\alpha$, and a symmetric inversion around the two central bending angles leading to a linearisation of the CNNC group. Various different pieces of evidence for the prevalence of one of these mechanisms over the others have been given for different enviroments.~\cite{weingart2011, Pederzoli2011, Titov2016} Calculated ground-state minimum energy paths for the three pathways are shown in Fig. S1 of the supplemental material. An important finding of our calculations is that metal-surface adsorption increases the relative energy difference between E and Z isomers due to a preferential adsorption of the E-Ab state. This additional stabilisation arises from a strong contribution due to dispersion interactions between the phenyl groups of the flat-lying E-Ab molecule and the substrate. As shown in Fig.~\ref{fig-results-ab-rotation} and further summarised as $\Delta$E(E-Z) in Table 1, for the three coinage metals (Au, Ag, Cu), this effect follows the trend of substrate reactivity with Cu showing the strongest effect and Au the weakest.
\begin{figure}[bh]
\centering\includegraphics[width=3.3in]{Fig3.pdf}
\caption{\label{fig-results-ab-rotation} Ground state minimum energy path from E-Ab to Z-Ab along the central dihedral rotation angle $\omega$ for Ab in gasphase (blue line, dashed) and adsorbed at Cu(111) (brown line, circles), Ag(111) (grey line, squares), and Au(111) (yellow line, triangles) surfaces. }
\end{figure}
The most striking finding is that, along the three reaction pathways, only the transition state along the dihedral rotation (TS) is significantly modified, which is why, in the following, we only focus on this pathway. As shown in Fig.~\ref{fig-results-ab-rotation}, the sizeable gas phase barrier of 1.09~eV when measured from the Z-Ab state, is strongly reduced upon adsorption. This effect is stronger for more reactive substrates and, in the case of Cu(111), reaches a point where the barrier vanishes completely and the Z-Ab state seizes to be a (meta)stable geometry. This finding is corroborated by the fact that no experimental evidence of a Z-Ab state on Cu(111) exists~\cite{alemani2008}. Even for a Ag(111) substrate, the remaining barrier of 0.04~eV will likely provide sufficient stabilisation of the Z-Ab state for experiments at finite temperatures. Already from this data it becomes evident that Ab can not act as a functioning molecular switch on any metal surface that is more reactive than Au(111), with experiments failing to find any evidence of switching even on more reactive Au facets such as Au(100)~\cite{alemani2008}.
As pointed out in ref.~\cite{Maurer2012}, the mechanism behind this rotational barrier reduction lies in charge-transfer between the metal and the adsorbate in the rotational transition state geometry. Upon breaking the N-N double bond, the metal stabilises the transition state by directly donating electrons to it. The electronegativity of the metal or its ability to accept electrons is inversely proportional to its reactivity and to the position of the Fermi level with respect to the molecular states. From Au to Cu, the barrier consistently decreases with decreasing electronegativity or increasing reactivity of the underlying metal. Thus, despite the potential existence of an efficient electron- or light-induced E-to-Z isomerisation, a thermal Z-to-E back-reaction becomes more and more favourable.
The electronegativity of the substrate, therefore, represents a central design parameter for metal-supported molecular switches. A potential strategy to enable the light-induced isomerisation of Ab is consequently to increase the electronegativity of the underlying substrate or, equivalently to reduce the availability of surface electrons to stabilise the transition state. However, it should be stressed that the energy landscape is highly sensitive to substrate electronegativity as already the step from a Au to a Ag surface destroys the molecular switching ability and, for adsorption on Cu(111) at higher coverages, even dissociation of the N-N bond has been observed.~\cite{Willenbockel2015} A subtle increase in electronegativity can, for example, be achieved by depleting electrons from the surface with coadsorbates that are strong electron acceptors. In the following, we will study a different avenue of tuning the switching ability, namely by molecular functionalisation with bulky spacer groups.
\begin{table*}
\begin{center}\caption{\label{tab-results-Ab-TBA} Central geometric parameters and relative energetics of Ab and TBA conformers in gas phase and adsorbed at coinage metal surfaces. In the case of TBA, $z$ refers to the vertical adsorption height of the lowest-lying N atom. In the case of Ab, $z$ refers to an average of the two N atoms. $\Delta$E(E-Z) refers to the relative energy difference between E and Z conformer and $\Delta$E(TS-Z) between transition state and Z conformer.}
\begin{tabular}{ccccccccc} \hline
& \multicolumn{2}{c}{E} & \multicolumn{2}{c}{TS} &\multicolumn{2}{c}{Z} & & \\ \hline
system & z & $\omega$ & z & $\omega$& z & $\omega$ & $\Delta$E(E-Z) & $\Delta$E(TS-Z) \\
& \AA & deg & \AA & deg & \AA & deg & \multicolumn{2}{c}{eV} \\ \hline
Ab & - & 180 & - & 90 & - & 12 & 0.58 & 1.09 \\
Ab@Cu(111) & 2.04 & 169 & 1.85 & 90 & 1.78 & 20 & 1.22$^*$ & - \\
Ab@Ag(111) & 2.52 & 173 & 2.08 & 90 & 2.10 & 50 & 0.77 & 0.06 \\
Ab@Au(111) & 3.03 & 179 & & 90 & 2.48 & 19 & 0.73 & 0.37 \\
TBA & - & 179 & - & 90 & - & 8.3 & 0.33 & 1.27 \\
TBA@Ag(111) & 2.36 & 155 & 2.06 & 89 & 2.07 & 29.7 & 0.48 & 0.04 \\
TBA@Au(111) & 2.67 & 166 & 2.00 & 90 & 2.31 & 8.7 & 0.46 & 0.48 \\ \hline
\end{tabular}\\
$^*$ The Z-Ab state on Cu(111) is not stable. For the purpose of this comparison it has been defined as the Ab molecule with an $\omega$ angle constrained at $20^{\circ}$.
\end{center}
\end{table*}
\subsection{The role of molecule functionalisation}
\label{results-tba}
A key strategy of molecular design is the derivatisation of molecules with purpose-designed functional groups to control the strength of molecular interactions. The four tetra-\emph{tert.}-butyl legs in TBA have been designed with the intent to decouple the molecule from the metal surface and to enable the photo-induced molecular switching. Contrary to Ab on Au(111), which can only be switched via electron-tunnelling through an STM tip, TBA on Au(111) can indeed be switched with light.~\cite{cho2010,comstock2008} However, the original idea that this is purely due to spatial decoupling has been debunked by quantitative x-ray standing wave measurements (XSW) and equilibrium structures of E-TBA on Au(111) calculated with DFT, which show that the vertical adsorption height of the central di-nitrogen bond is not significantly altered.~\cite{McNellis2010,mcnellis2010a} In fact, our PBE+vdW$^{\mathrm{surf}}$ calculations (see Table 1) find vertical adsorption heights of the nitrogen atoms for E-TBA and Z-TBA that are even lower than for E-Ab and Z-Ab. The strongest reduction can be found for E-TBA on Au(111). In this case, this is mostly due to an asymmetric adsorption of the di-nitrogen bridge with a difference in height between the two nitrogen atoms of 0.3~\AA{}.
\begin{figure}
\centering\includegraphics{Fig4.pdf}
\caption{\label{fig-results-TBA-substrates} Minimum energy paths from E-TBA to Z-TBA along the dihedral rotation coordinate. The paths are shown for TBA in gas phase (blue line, dashed) and adsorbed to an Ag(111) (grey line, squares) and Au(111) (yellow line, triangles) surface.}
\end{figure}
As Fig.~\ref{fig-results-TBA-substrates} shows, the spacer groups in TBA also do not have any drastic effects on the ground state energy landscape. For TBA on Ag(111) and Au(111), the rotational barriers are strongly reduced compared to the free molecule and the remaining barriers compared to the Z-TBA state are comparable to those of Ab adsorbed on the respective surfaces with an almost vanishing thermal barrier for Z-TBA on Ag(111) and a barrier of 0.48~eV for Z-TBA on Au(111). This means that there are no thermodynamic arguments against the existence of a Z-TBA state on Au(111), whereas long-time thermal stability of a Z-TBA state on Ag(111) is highly improbable. Both of these assertions, as in the case of Ab, are in agreement with experimental reports on these systems.~\cite{comstock2007,alemani2008}
An interesting structural difference that can be observed is the change in $\omega$ angles for the equilibrium E and Z-TBA structures compared to adsorbed Ab structures. Whereas E-Ab adsorbs in an almost flat-lying geometry on the Ag(111) and Au(111) surfaces, E-TBA shows deviations of 25$^{\circ}$ and 14$^{\circ}$, respectively, from a perfectly flat arrangement with the phenyl rings and side groups distorted away from the surface. On the other hand, the Z-TBA structures show much more acute angles than their Ab counterparts when adsorbed on Ag(111) and Au(111) (see also Fig. S2 and Fig. S3 for molecular structures of the metal-surface adsorbed Ab and TBA equilibrium geometries). Despite these differences, the difference in angle $\Delta\omega$ that defines the reaction path from E to Z is almost the same for Ab and TBA on Ag(111) and Au(111), namely 123/125$^\circ$ and 160/157$^\circ$, respectively. However, the effective angle distance from the E-TBA state to the rotational ground-state barrier is reduced for E-TBA compared to E-Ab. On Au(111), this difference is 89$^{\circ}$ for E-Ab and 76$^{\circ}$ for E-TBA. In addition, the energy difference between the E and Z states is significantly smaller when compared to Ab.
This means that, in the context of the proposed light- or electron-driven mechanism that involves short-lived transient excited states, which provide a funnel for energy from hot electrons into vibrational degrees of freedom, less total energy has to be provided to make it from the E-TBA to the Z-TBA state compared to Ab and an effectively shorter path has to be passed to cross the barrier. The first is accomplished by a relative destabilisation of the E-TBA state compared to the rotational transition state and the Z-TBA state due to the bulky spacer groups. The latter arises from the distorted di-nitrogen center in the ground state structure of E-TBA. We therefore note that, while bulky spacer groups do not geometrically decouple the di-nitrogen bridge from the surface, their destabilising effect on the E-TBA geometry will facilitate switching.
Whereas a clear picture arises for the switching ability of Ab and TBA on Ag(111) and Ab on Cu(111) and, by deduction, any other more reactive metal surface, the subtle differences between the ground-state potential energy landscapes and equilibrium structures of Ab and TBA on Au(111) do not provide an explanation for the difference in switching ability between the two. In order to understand the thermodynamic and kinetic differences that give rise to the photoswitching ability of TBA on Au(111), we therefore proceed to study the energetic landscapes of molecule-surface charge-transfer excitations proposed in the context of the on-surface isomerisation reaction.
\subsection{Excited state energy landscapes of transient molecular ions}
\label{results-excited-state}
The requirements of potential energy landscapes to be able to support a successful transient ion-based photoswitching mechanism, as the one described by Wolf and Tegeder,~\cite{Wolf2009} are somewhat different from the ones that apply to gas phase and solvent photochemistry. First and foremost, the transient ion formation has to be efficient, \emph{ie.} the transient ionic state has to have high overlap with high-density areas of the metallic density-of-state, such as the d-bands, so that electrons or holes can be efficiently transferred. The probability of switching due to this transient ion will be high if, within its short lifetime of few tens of femtoseconds,~\cite{campillo2000} it can efficiently transfer electronic excess energy from the vertical excitation that led to the ion formation to vibrational energy within the reactive degrees of freedom of the molecule that will drive the isomerisation. This will be the case if the equilbrium geometry exhibits strong forces upon ion formation. Whereas, in the context of an intramolecular excitation, changes to the excited state landscape upon surface-adsorption are to be avoided to retain the switching function, in the context of a transient ion formation, changes to the energy landscape of the cationic state induced by substrate variation or molecule functionalisation could be advantageous. To enable an efficient transient-ion-driven isomerisation, the energy landscapes of transient ion excitations need to show the following features:
\begin{enumerate}[1.]
\item a significant gradient towards the transition state within the Franck-Condon (FC) region of the excited state PES,
\item a barrierless path towards the ground-state transition state,
\item and a kinetic excess energy (\emph{vide infra}) that exceeds the ground state barrier, but is not too large to increase the probability of immediate thermal Z-to-E backreaction.
\end{enumerate}
The natural starting point for an investigation of the adsorbed PES topology of ionic resonances are the cation and anion PESs of isolated Ab and TBA. Corresponding gas phase calculations have been reported by F\"uchsel \emph{et al.}~\cite{Fuchsel2006} and Leyssner \emph{et al.}~\cite{Leyssner2010} in the context of electron-tunneling- and light-induced Ab switching on metal surfaces. STM-induced switching can be triggered with negative and positive bias voltage. Therefore, electron removal ('cationic resonance') and electron attachment ('anionic resonance') to the molecule are both viable mechanisms, whereas in the case of light-induced switching of Ab and TBA the possibility of an anionic resonance as dominating excitation mechanism has been excluded by experimental considerations.~\cite{Hagen2007, Wolf2009}
The above mentioned studies report DFT-based anionic and cationic potential energy surfaces of Ab and TBA along rotation and inversion. Just as in the ground-state, these PESs show two meta-stable states at very similar positions, although slightly shifted towards smaller $\alpha$ bending angles and $\omega$ dihedral angles, closer to the mid-rotation point. Correspondingly, both ionic states show significant gradients at the FC structures of E and Z-Ab and TBA. Additionally upon electron removal or attachment, the rotation and inversion barriers are significantly reduced, although still present. In general, the ionic surfaces are less corrugated and specifically the cationic PES is much more shallow. It can therefore be argued that in both ionic molecular states, the prerequisites to a substrate-mediated excitation mechanism are in principle given and if only minimal changes to these PESs occur upon metal-surface adsorption, light-induced switching along the Wolf-Tegeder mechanism is feasible.
\begin{figure*}
\centering\includegraphics[width=\textwidth]{Fig5.pdf}
\caption{\label{fig-Ab-TBA-CT} (a) and (b) Upper panel: Relaxed minimum-energy paths of Ab (a) and TBA (b) adsorbed on Au(111) along the dihedral rotation coordinate. Shown are the ground-state PES (black), an anionic resonance where an electron has been added to the LUMO of the molecule ($e^{-}$, red), and a cationic resonance where an electron has been withdrawn from the HOMO of the molecule ($h^{+}$, blue). Points marked with dashes for the anionic resonances are calculations where the excitation constraint in le$\Delta$SCF-DFT could not be fully satisfied due to the already existing occupation of the LUMO in the ground-state (see the main text for more details). Lower panel: Ground-state occupations of the molecular HOMO (dashed line) and the molecular LUMO (dotted line) along the dihedral rotation pathway. These occupations have been calculated using the orbital projections that underlie the le$\Delta$SCF-DFT method.~\cite{Maurer2013}. (c) Side view of Ab equilibrium and transition state structures along the isomerisation pathway. The order follows: E-Ab, TS-Ab, Z-Ab (d) Side view of TBA equilibrium and transition state structures along the isomerisation pathway. The order follows: E-TBA, TS-TBA, Z-TBA}
\end{figure*}
Figure \ref{fig-Ab-TBA-CT} presents the calculated cationic (in blue) and anionic (in red) resonance states for Ab and TBA adsorbed on Au(111) along the minimum energy path of dihedral rotation. As described in the previous sections, the ground state PES along this pathway (shown in black) exhibits a very similar topology for both molecules, which derives from the similar adsorption-induced changes that occur along this pathway. Both E and Z states of Ab and TBA are dominantly physisorbed, which is reflected in a net occupation of the molecular LUMO that is close to zero and a net population of the molecular HOMO that is close to full occupation (see occupation panels in Fig.~\ref{fig-Ab-TBA-CT}). These occupations are calculated by projecting the molecular orbitals of the isolated molecule from the band structure of the slab with the adsorbed molecule and integrating the corresponding occupations up to the Fermi level.~\cite{Maurer2012,Maurer2013} In a previous study, we have compared this charge analysis measure with other more established methods and have found it to provide a description of net charge flow between that of the Hirshfeld and the Bader methods.~\cite{Mueller2016} The loss of occupation in the HOMO and the gain of occupation in the LUMO around the rotational transition state is a clear sign for a covalent bond formation via a surface-to-molecule donation into the LUMO and back-donation of charge from the HOMO into the surface.
The here presented excitation energies for the cationic and anionic resonance represent the lowest possible vertical transitions of electrons between molecular states and metal states at the Fermi level based on the equilibrated ground-state geometries. Corresponding excitations can also happen at higher photon energies albeit with more excess energy. For both molecules, we find the cationic resonance at a higher excitation energy than the anionic resonance, as expected considering that the IP of the isolated molecule is much larger than the EA. At the E-Ab geometry, the corresponding energies (h$^+$: 1.65~eV, e$^-$: 1.39~eV) are close to the molecular resonance energies measured with 2PPE (HOMO: 1.77~eV, LUMO: 1.68~eV, absolute energy difference from the Fermi level~\cite{Bronner2012}) with the same qualitative trend of HOMO$>$LUMO. The anionic resonance shows a larger discrepancy, which could be a reflection of the poor description of unoccupied states with semi-local DFT and possibly also due to the difference between the high-coverage situation in experiment and the intermediate-coverage as described by our calculations.
For both molecules, the anionic resonance (shown in red in Fig.~\ref{fig-Ab-TBA-CT}) clearly satisfies the above mentioned criteria. The path from E-Ab and E-TBA towards the transition state is barrierless and shows a moderate gradient. The nominal excitation energy, not assuming any dissipation during the transition, is sufficient to overcome the barrier. We should note that the description of this excitation around the rotational transition state is limited in the le$\Delta$SCF-DFT method as the LUMO of the molecule is already more than half filled between 125 and 75$^{\circ}$ (dashed line). This means that we cannot excite a full electron into the LUMO and therefore cannot fully satisfy the occupation constraint. This is, however, also an indication that the lifetime of such a state will be extremely short. Considering also the experimental assessment that suggests that the isomerisation is not triggered by an anionic resonance, we will therefore focus on the cationic resonance state in the following.
For the cationic resonance states of TBA and Ab on Au(111), more differences are visible. The $h^{+}$ surface for both molecules shows a minimum close to the rotational transition state. However, the E-TBA FC region exhibits a significant gradient towards the transition state, whereas the E-Ab FC region shows an extended plateau across a wide range of $\omega$ values. The electronic excess energy, \emph{ie.} the energy difference between the vertical excitation energy and the ground state barrier, is very similar for both systems (0.55 and 0.52~eV for Ab and TBA on Au(111)). Therefore the cationic resonance along rotation satisfies all above defined criteria for transient-ion-mediated switching for TBA on Au(111). The cationic resonance for Ab on Au(111) instead satisfies criteria 2 and 3, but violates criterion 1. This means that, while in principle possible, the insufficient coupling of molecular vibrations with the transient cation state, could prevent sufficient energy to be transferred to the molecular vibrations within the short lifetime of the excitation. This then would rationalize why, contrary to TBA on Au(111), no light-induced switching has been observed for Ab on Au(111), whereas E-to-Z switching can be induced with sufficiently high STM electron tunnelling currents.
The excited-state topologies calculated for TBA do thus support the Wolf-Tegeder mechanism in several aspects. The transition from E to Z-TBA involves almost exclusively rotational motion around the central dihedral angle $\omega$, with the exception of a small, almost isoenergetic twisting of the phenyl rings when lifting off the phenyl ring at dihedral angles between the transition state and Z-TBA. The PES topology of a cationic resonance in the HOMO that is connected with this motion represents a fairly steep, barrierless descent towards the mid-rotation point. At this point, if not earlier, increased coupling with the substrate is likely to quench the already short-lived excitation and a corresponding motion will have a higher probability for crossing the ground-state barrier than for a momentum inversion to occur. The smaller gradient around the Z-TBA FC region and the correspondingly reduced vibrational activation of the rotational motion rationalise the lack of experimentally found Z-to-E TBA photoswitching. Furthermore, the strong vibrational dependence of the switching rate and the termal Z-to-E isomerisation that have been observed experimentally,~\cite{hagen-phd} are also consistent with the potential energy path that has been found here. A thermal activation of the dihedral angle, corresponding to a wagging motion of the molecule perpendicular to the surface, will strongly facilitate a corresponding switching upon photo-excitation.
Up to this point, we have only considered isomerisation along the dihedral rotation around the central di-nitrogen bridge. The literature around light-induced isomerisation in the gas phase discusses in detail the (co)existence of different isomerisation pathways that also include an inversion around one of the two central bending angles. For the sake of completeness, we note that cationic and anionic resonance excitations along such an inversion do not provide energy landscapes that satisfy the above criteria (see Fig. S4 in the supplemental material). Both, cationic and anionic resonance, show a topology similar to the free molecular cation and anion states with a barrier between the E-Ab and Z-Ab states. It should be noted that the steric hindrance due to bulky spacer groups of TBA will significantly impede the isomerisation along the bending angle inversion, leaving dihedral rotation as the only viable pathway for E-to-Z isomerisation. Counterintuitively, the introduction of large functional side groups, as in TBA, strongly decreases the number of degrees of freedom that can participate in isomerisation. As such, the here presented findings would suggest that, at least for the light-induced on-surface isomerisation, dihedral rotation is the dominant isomerisation pathway. This is contrary to a suggestion by Comstock \emph{et al.} that inversion must be the dominant isomerisation route.~\cite{Comstock2010}
Despite the apparent failure to stabilise a Z-Ab or Z-TBA state on Ag(111), it can be insightful to compare the effect of different substrates on the ionic PESs. Comparing the excited states along the rotation for Ab on Au(111) and Ag(111) (see supplemental material, Fig. S5), strong changes to the cationic and anionic resonance PESs are apparent. The corresponding paths of cationic and anionic resonance states along rotation are strongly modified, as compared to the gas phase or to the molecule adsorbed on a Au(111) surface. Owing to the strong adsorption, almost everywhere along the path the LUMO is more than half-filled in the ground state and therefore the anionic resonance excitation constraint in le$\Delta$SCF can not be fully satisfied. Correspondingly, the anionic resonance state is systematically downshifted and coincides with the ground state over large portions of the rotational pathway. More importantly, the cationic resonance state is systematically shifted upwards on Ag(111) as compared to Au(111) and appears more corrugated with a pronounced minimum at the transition state. At both FC regions of E and Z-Ab on the cationic resonance state, there is a significant gradient towards the mid-rotation point. Quite ironically, whereas Ab adsorbed on Ag(111) nominally fulfils the necessary criteria to efficiently activate nuclear motion upon excitation, it fails to sufficiently stabilise a Z-Ab conformer. In contrast, in the case of Ab on Au(111) sufficient stability is accompanied by highly inefficient vibronic energy transfer. The inspection of the static PESs therefore suggests that both isomerisation processes can, at best, be highly inefficient.
The excited state PES differences between Ag and Au can be understood in terms of the differences in electronegativity of the underlying surfaces. The electronegativity of the Au surface is higher than that of the Ag surface. Therefore, electrons are harder to detach from the Au surface than from the Ag surface, or in other words, adding electrons to a gold surface is connected to a smaller energetic penalty than for the case of silver. When inducing a cationic resonance state, an electron from the molecule is transferred to the substrate. For the above reasons, the energy that is necessary to do this is higher for Ag than for Au. In the case of an anionic resonance this situation is reversed. The more electronegative a surface, the higher the energetic penalty to withdraw electrons from it. Correspondingly, the energy associated with the anionic resonance on the molecule is higher for E-Ab on Au(111) than on Ag(111). This, however, does not hold for the resonances at the transition state. In both cases, the chemical bond to the surface efficiently transfers electrons from the substrate to the LUMO of the molecule. Correspondingly, the ground state is already, to some extent, an 'anionic resonance' state compared to the isolated molecule; the ground state and the $e^{-}$ state are energetically equivalent. Detaching an electron at the mid-rotation point from the HOMO, effectively yields an $n\rightarrow\pi^*$ excited Ab molecule. The corresponding excitation energy at the transition state, due to the orbital degeneracy and the conical intersection with the ground state, is zero in the gas phase. The minimum observed in the cationic state at the transition state for the adsorbed molecules stands in contrast to the barrier exhibited at this point for the gas phase molecules.~\cite{Fuchsel2006} This minimum appears because of the significant mixture of the cationic state with the n$\rightarrow\pi^*$ (S1) intramolecular excitation that stems from the molecule-surface charge-transfer in the ground-state. Correspondingly, at this point, the hybridization with the surface completely inverts the topology of the cationic resonance state and enables an otherwise infeasible substrate-mediated light-induced isomerisation process.
\section{Discussion and Outlook}
\label{conclusions}
The combined picture of ground- and excited state energy landscapes for Ab and TBA on Ag(111) and Au(111), in combination with previously presented intramolecular excited state landscapes,~\cite{Maurer2013} provides a clear rationalisation of the proposed switching mechanism in terms of the choice of substrate and the molecule functionalisation, as well as their effect on the underlying molecule-surface chemistry.
Reactive metal surfaces will not be able to sustain light- or electron-induced E-Z switching around covalent double bonds. This apparently includes noble surfaces such as silver and anything more reactive. The reason for this is an unbalanced stabilisation of different geometries along the reaction pathway, where the transition state and the E-Ab state are strongly stabilised with respect to the Z-Ab state by surface-molecule charge transfer and dispersion interactions, respectively. The effect is a loss of stability of the Z-Ab state in the ground state.
In addition to a lack of ground state stability, such reactive substrates that easily donate electrons to adsorbates will also suppress anionic resonances to the LUMO and shift cationic resonances to high energies, which equivalently means reduced overlap with the d-bands. For Ab on a more inert substrate such as Au(111), due to the high substrate electronegativity, the ionic excitations are sufficiently separated from the ground state and the ground-state stability is such that, in principle, a bistable molecule with an E and a Z state can be supported. Nevertheless, surface adsorption (hybridisation and image charge interactions) strongly modifies the topology of the ionic resonance states compared to the isolated molecule with the effect that the coupling to molecular vibrations for the E-Ab state is minimal. Whereas simple UV light exposure will not be able to transfer sufficient momentum to molecular motion, inelastic electron tunnelling can still transfer sufficient kinetic energy to achieve an isomerisation to the Z-Ab state.
The bulky spacer groups in TBA have a much more faceted effect on the potential energy landscape and switching ability than previously considered. The additional side groups do not achieve a chemical decoupling of the central di-nitrogen bridge from the surface, however they distort the E-TBA conformation and effectively destabilise it compared to all other structures. As such they contribute to a more balanced adsorption along the reaction pathway. Yet, the charge-transfer at the rotational transition state and the dipole-image charge interaction in the Z-TBA state are still sufficiently large to modify the cationic resonance landscape compared to adsorbed Ab. As a result, the excited-state PESs are shifted downwards for the rotational transition state and the Z-TBA isomer: what was a plateau for the cationic resonance state in the case of E-Ab, for E-TBA, shows a gradient that evidences the ability to transfer electronic excess energy into vibrational motion towards the barrier along a path that is significantly shorter than it is for E-Ab. These effects compound to a much higher reaction probability, while not significantly affecting the level alignment and the rate of ion formation.
The above presented results may also help to settle the controversy over the prevalence of rotation or inversion motion during photo-isomerisation. Comstock {\em et al.}, based on their STM experiments, have proposed that a dominance of an inversion-based mechanism is likely. The authors base this interpretation on the observation of a selection rule to photo-switching that applies differently to different chiral island domains of TBA on Au(111). They rationalize this effect by an initially planar linearisation of one CNN bond angle $\alpha$ and two subsequent channels towards the Z-TBA structure. In that way, E-TBA of one racemic type can convert into a left-handed and a right-handed Z-TBA molecule by changing the handedness of the di-nitrogen bridge. As a second argument, such a pathway would additionally maximise the dispersion interaction of both phenyl rings with the surface as much as possible during the isomerisation process.
The here presented ground- and excited-state potential energy landscapes paint a different picture. Whereas surface adsorption strongly affects the rotational pathway, minimal effects are apparent for the isomerisation along bending angle inversion. Both intramolecular excitation or transient ion formation would lead to barriers in the excited state and an insufficient driving force towards isomerisation. The corresponding ground state barrier that has to be overcome is considerably higher for inversion than in the case of rotation. Previously found low-lying vibrational modes contribute to vertical wagging modes that can initiate a rotational motion, whereas no such low-lying modes where found that contribute to asymmetric bending angle inversion.~ \cite{McNellis2010} On the basis of our results, dihedral rotation provides the most likely pathway to isomerisation. This does, however, not mean that some bending angle motion might not contribute to the dynamical process and the intramolecular vibrational energy distribution. The chiral selectivity that forms the argument for an inversion-based mechanism can be understood by taking a look at the dynamics of the mechanism for the molecules in solvent. Weingart \emph{et al.}~\cite{weingart2011} have found that also dihedral rotation can generate two different pro-chiral Z-Ab variants by either clockwise or anti-clockwise rotation. This finding is based on the fact that the dominant motion during isomerisation is carried out by the light-weight di-nitrogen bridge rather than the side groups. In the case of surface-adsorbed molecules, that could mean that during a wagging motion of the molecule, an asymmetric vertical rotation of the two nitrogen atoms (much like the motion of pedals on a bicycle when viewed from ahead) may change the handedness of the molecule. This assertion is supported by the recent finding of such motion for an Ab tripod structure on Ag(111).~\cite{Scheil2016}
Our findings show that the le$\Delta$SCF-DFT approach, at least qualitatively, accounts for a range of important effects that shape on-surface excited states, including electrostatic stabilisation via interaction between excited-state dipoles and the substrate image charge, as well as hybridisation and charge-transfer in chemisorbed geometries. Although no fully quantitative description can be expected from an approximate scheme like le$\Delta$SCF-DFT, the presented excited state topologies offer a convincing rationalisation of the experimental findings for the molecular switching of Ab and its derivatives on the close-packed coinage metal surfaces. One of the current limitations of our implementation of le$\Delta$SCF-DFT is the inability to calculate analytical forces in the excited state. This will be key to ensure its widespread usage and its successful application for complex and high-dimensional light-driven dynamics at surfaces. With the advent of the machine-learning-based construction of high-dimensional energy landscapes and more from {\em ab initio} data,~\cite{Behler2007b,Kolb2017,Rupp2018} this limitation could be alleviated. The authors thus see a clear remit for this efficient and approximate excited-state methodology to tackle challenging problems in photocatalysis and nonadiabatic dynamics at surfaces and nanoparticles.
\section{Conclusion}
\label{conclusion}
We have performed state-of-the-art dispersion corrected DFT and linear expansion $\Delta$SCF-DFT calculations to study the potential energy landscapes of the ground state and molecule-surface charge-transfer excited states of Azobenzene and its derivative TBA adsorbed at coinage metal surfaces. In the context of light- and electron-tunnelling-induced molecular switching of Ab, we investigated the traditional experimental parameters of functional interface design, namely the choice of substrate and its reactivity, and the molecular functionalisation. Our findings provide evidence of the drastic and somewhat unexpected effects that these parameters have on the energy landscape and the expected reactivity of the two molecules.
Our findings fully support a previously proposed mechanism based on the transient formation of a cationic molecular resonance due to light exposure or electron-tunnelling. We can clearly discern the effects that the change of substrate and the introduction of bulky spacer groups have on the energy landscapes, the molecule-surface interaction and the reactivity. The differences in the ground and excited-state potential energy landscapes that our calculations expose, clearly rationalise why TBA on Au(111) can be switched via light, whereas Ab on Au(111) can only be switched via inelastic electron tunnelling through an STM tip.
In addition to providing mechanistic insight into the photochemical reactivity of a prototypical functional interface and a differentiated picture of the effect of traditional experimental design strategies, our calculations provide evidence of the potential utility of the le$\Delta$SCF-DFT method for a diverse range of chemical challenges associated with charge-transfer in condensed matter.
\ack
Funding through the Deutsche Forschungsgemeinschaft (DFG) is gratefully acknowledged. Computer resources for this project have been provided by the Gauss Centre for Supercomputing/Leibniz Supercomputing Centre under grant id: pr63ya and the Garching Supercomputing Center of the Max-Planck Society.
\section*{References}
\providecommand{\newblock}{}
|
\section{1. General Results Regarding Commuting Projector Hamiltonians}
\subsection{1a. Partial Transposition Preserves the Set of Eigenvectors}\label{appendix_partial_transpose}
Consider a commuting projector Hamiltonian $H=H_A+H_B+H_{AB}$ , where $H_A$ and $H_B$ denote the part of $H$ with support only in real space region $A$ and $B$, and $H_{AB}$ denotes the interaction between $A$ and $B$. Define $\{O_m \}$ as the set of local commuting operators, a commuting projector Hamiltonian can be written as $H=\sum_m c_mO_m$. The thermal density matrix, $\rho=e^{-\beta H}/Z$ with $Z=\Tr e^{-\beta H}$, can be expanded as: $\rho= \sum_{\alpha} d_{\alpha}Q_{\alpha}$, where each $\{Q_{\alpha}\}$ is a tensor product of operators from the set $\{O_m\}$. Since all operators in $H$ commute, $H$, $\rho$, and $\{ O_m\}$ share the same eigenvectors. Under the partial transpose over the Hilbert space in $B$, one obtains $\rho ^{T_B}=\sum_{\alpha} d_{\alpha} Q_{\alpha}^{T_B}$. If $Q_{\alpha}$
only acts on A or B, then $Q_{\alpha}^{T_B}=Q_{\alpha}$. Only when the support of $Q_{\alpha}$ involves $A$ and $B$ simultaneously is it possible for $Q_{\alpha }$ to receive a minus sign under partial transpose. This implies that the operators basis for $\rho^{T_B}$ is still $\{Q_{\alpha} \}$, and thus the eigenvectors of $\rho^{T_B}$ are exactly the same as those of $\rho$, and the eigenvalues of $\rho^{T_B}$ can be obtained by replacing $\{O_{m} \}$ by their eigenvalues. In the argument above we implicitly assumed that all matrix elements of $\{Q_{\alpha}\}$ are real in the basis where we perform a partial transpose. If there exists complex matrix elements instead, $\{Q_{\alpha}\}$ might get a minus sign even when $\{Q_{\alpha}\}$ acts only on $A$ or $B$. Nevertheless, one can check that $\rho^{T_B}$ is still generated by tensor products of $\{O_m\}$, and therefore the conclusion remains the same.
\subsection{1b. Partial Trace Preserves the Set of Eigenvectors}\label{appendix:partial_trace}
Here we show that for commuting projector Hamiltonians, the thermal density matrix $\rho$ and the reduced density matrix $\rho_A$ obtained by tracing out all the degrees of freedom in $B$ share the same set of eigenvectors. As discussed above, $\rho= \sum_{\alpha} d_{\alpha}Q_{\alpha}$, where $\{Q_{\alpha}\}$ collects all possible operators from the product of commuting operators $\{O_m\}$. By tracing out all the degrees of freedom in $B$ for $\rho$, basis operators in $\{ Q_{\alpha } \}$ which act non-trivially on $B$ vanish. This implies that the operator basis of reduced density matrix $\rho_A$ is generated by the those operators in $\{Q_{\alpha} \}$ which act on $B$ trivially, and thus $\rho_A$ commutes with all local commuting operators.
\subsection{1c. Bipartite Negativity from Reduced Density Matrix on Boundary}\label{append:boundary_nega}
Here we show that the negativity between two spatial regions of a thermal density matrix of a commuting projector Hamiltonian equals the negativity of the reduced density matrix localized on the boundary of the bipartition. Following the notation in the main text, we define $\partial A(\partial B)$ as collection of spins on the boundary of $A(B)$ that interacts with $B(A)$, and define $A'(B')$ as the collection of spins in the bulk of $A(B)$ that only couples to spins in $A(B)$. We decompose a commuting projector Hamiltonian as $H = H_A + H_B + H_{AB}$, so that $H_A(H_B)$ denotes the interaction between the bulk spins in $A(B)$, and $H_{AB}$ denotes the interaction between the boundary spins in $\partial AB=\partial A\bigcup \partial B$. For simplicity, we also assume that the system is time reversal invariant, so that for $\rho =e^{-\beta H}/Z$, the partial transpose over the Hilbert space in $B$ acts non-trivially only on $H_{AB}$:
\begin{equation}
\left(\rho \right)^{T_B}=\frac{1}{Z} \left( e^{-\beta H_{AB}} \right)^{T_{\partial B}} e^{-\beta\left( H_{A}+H_B \right) } .
\end{equation}
As discussed above, a partial transposed density matrix is still generated by local commuting operators which are present in $H$. This implies that one can find a common eigenbasis for $H_A, H_B$ and $H_{AB}^{T_{\partial B}}$, and the eigenvalues of $\rho^{T_B}$ can be obtained by replacing all local operators by their eigenvalues. Consequently, the eigenvalues of $\rho ^{T_B}$ take the form:
\begin{equation}
\lambda=\frac{1}{Z} \left( e^{-\beta H_{AB}} \right)^{T_{\partial B}}\left(\sigma_A,\sigma_B\right) e^{-\beta\left( H_{A}(s_A,\sigma_A)+H_B(s_B,\sigma_B) \right) }.
\end{equation}
where $s_{A(B)}$ denotes the spin configuration in the bulk of $A(B)$, and $\sigma_{A(B)}$ denotes the spin configuration on the boundary that interact with $B(A)$. The one-norm of $\rho^{T_B}$ can be obtained by summing all the absolute values of eigenvalues:
\begin{equation}\label{eq:prove_one_norm}
\norm{\rho^{T_B} }_1 =\frac{ \sum_{\sigma_A,\sigma_B} \abs{ \left( e^{-\beta H_{AB}} \right)^{T_B}\left(\sigma_A,\sigma_B\right) } \sum_{s_A,s_B} e^{-\beta\left( H_{A}(s_A,\sigma_A)+H_B(s_B,\sigma_B) \right) } }{ \sum_{\sigma_A,\sigma_B} \left( e^{-\beta H_{AB}\left(\sigma_A,\sigma_B\right) } \right) \sum_{s_A,s_B} e^{-\beta\left( H_{A}(s_A,\sigma_A)+H_B(s_B,\sigma_B) \right) } }.
\end{equation}
The observation that partial transpose only affects the operators on the boundary motivates us to consider the reduced density matrix on the boundary:
\begin{equation}
\rho_{\partial AB} = \frac{1}{Z} \Tr_{A',B'} e^{-\beta H} = \frac{1}{Z} e^{-\beta H_{AB}} \Tr_{A',B'} e^{-\beta \left(H_{A}+H_B \right) } .
\end{equation}
We take the partial transpose over $\partial B$
\begin{equation}
\left( \rho_{\partial AB} \right)^{T_{\partial B}}
= \frac{1}{Z} \left\{ e^{-\beta H_{AB}} \right\}^{T_{\partial B}} \Tr_{B'} e^{-\beta H_B } \Tr_{A'} e^{-\beta H_A },
\end{equation}
where the commutative property of each local operator is used. As a result, the eigenvalue of $\left( \rho_{\partial AB} \right)^{T_{\partial B}}$ is:
\begin{equation}\label{eq:reduced_eig}
\lambda_{\sigma_A,\sigma_B} = \frac{1}{Z} \left( e^{-\beta H_{AB}} \right)^{T_{\partial B}}\left(\sigma_A,\sigma_B\right) \sum_{s_A,s_B} e^{-\beta\left( H_{A}(s_A,\sigma_A)+H_B(s_B,\sigma_B) \right) } .
\end{equation}
By summing all absolute values of $\lambda_{\sigma_A, \sigma_B}$ for $\norm{\rho_{\partial AB}^{T_{\partial B}} }_1 $ and comparing it with Eq.\ref{eq:prove_one_norm}, one finds that
\begin{equation}
\norm{\rho^{T_B} }_1 =\norm{\rho_{\partial AB}^{T_{\partial B} }}_1,
\end{equation}
which implies that the negativity of two spatial regions is given by the boundary of those two spatial regions. In fact with a similar calculation, one can show that the above equality also holds true for any commuting project Hamiltonian without time reversal symmetry.
\section{2. Calculational details of negativity for various models discussed in the main text}
\subsection{2a. Infinite-Range Commuting Projector Hamiltonian}
Consider a one-dimensional lattice of size $L$ where each lattice site has four qubits, the model Hamiltonian is
\begin{equation}
\begin{split}
H=&-\frac{1}{4L} \left( \sum_{i=1}^L \left(Z_{i1} Z_{i2} + Z_{i3} Z_{i4} \right) \right)^2 -g_z \sum_{i=1}^L Z_{i1} Z_{i2} Z_{i3} Z_{i4} -g_x \sum_{i=1}^L \left( X_{i1} X_{i2}+X_{i3} X_{i4} \right).
\end{split}
\end{equation}
The density matrix at inverse temperature $\beta$ is $\rho=\frac{1}{Z}e^{-\beta H}$ with $Z=\Tr e^{-\beta H}$. Since every local term commutes, we can perform Hubbard-Stratonovich transformation for $e^{-\beta H}$:
\begin{equation}\label{eq:ir_thermal_state}
e^{-\beta H} =\sqrt{\frac{ \beta L}{ \pi }} \int dm e^{ -\beta L m^2 - \beta \sum_{i=1 }^L H_i(m)},
\end{equation}
where a local Hamiltonian $H_i(m)$ for $i$-site of four spins is defined as :
\begin{equation}
H_i(m)= - m ( Z_{i1} Z_{i2} + Z_{i3} Z_{i4} ) - g_z Z_{i1} Z_{i2} Z_{i3} Z_{i4} -g_x \left(X_{i1} X_{i2}+X_{i3} X_{i4} \right).
\end{equation}
Eq.\ref{eq:ir_thermal_state} implies that all sites are separable since $\rho$ manifestly takes the form $\rho=\sum_k p_k \rho^1_k \otimes \cdots\otimes \rho_k^L$ where $p_k\geq 0 $, $\rho_k^i$ is a local density matrix on $i$-th site. As a result, to have non-zero negativity, an entanglement cut should be made across one of the sites (say $s$-th site) such that four spins on $s$-th site are not in the same subsystem. In the following calculation, $A$ comprises all the lattice sites with site index $i<s$ and two spins labelled by $1,3$ on $s$-th site while $B$ comprises all the lattice sites with site index $i>s$ and two spins labelled by $2,4$ on $s$-th site.
The negativity $E_N$ can be calculated via a replica trick:
\begin{equation}\label{eq:inf_nega}
E_N=\log \norm{\rho^{T_B}}_1 = \lim_{n_{e}\to 1} \frac{\Tr\left[ \left( ( e^{-\beta H} )^{T_B} \right)^{n_e} \right]}{\Tr\left[e^{-\beta H} \right]}.
\end{equation}
Notice that $n_e$ is an even number as performing trace, but analytic continuation $n_e\to 1$ is taken in the end.
First we calculate the thermal partition function:
\begin{equation}\label{eq:inf_stabilizer_Z}
Z=\Tr{ e^{-\beta H} } =\left(\frac{\beta L}{\pi} \right)^{ \frac{1}{2} } \int dm e^{-\beta L m^ 2} \Tr{e^{-\beta \sum_{i=1}^L H_i(m) } } = \left(\frac{\beta L}{\pi} \right)^{ \frac{1}{2} } \int dm e^{-\beta L f(m) }
\end{equation}
where
\begin{equation}
\beta f(m)= m^2 - \log \left[e^{\beta g_z} \cosh(2 \beta m ) +e^{-\beta g_z} \right] -\log \left[ 8 \cosh[2](\beta g_x) \right].
\end{equation}
The integral over $m$ is dominated by the saddle point $m^*$, which satisfies $\eval{\frac{\partial f(m)}{\partial m}}_{m^*} =0$:
\begin{equation}\label{eq:inf_stabilizer_saddle}
\frac{\sinh(2\beta m^*)}{ \cosh(2\beta m^* ) +e^{-2\beta g_z} } =m^*.
\end{equation}
The critical behavior of $m^*$ can be determined by expanding Eq.\ref{eq:inf_stabilizer_saddle} to $O(m^{*3})$:
\begin{equation}
\frac{2\beta m^*}{1+w} +\frac{4(w-2)}{3\left( 1+w \right)^2} \beta^2m^{*3}=m^*,
\end{equation}
where $w(\beta)\equiv e^{-2\beta g_z}$.
Define $\beta_c\equiv \frac{1+w(\beta_c)}{2}$, for $\beta>\beta_c$, we can have non-zero solution for $m^*=\pm m_0$:
\begin{equation}
m_0=\sqrt{ \frac{3\beta_c\left( \beta-\beta_c \right)}{\beta^3\left( 3-2\beta_c \right)} }
\sim \sqrt{T_c-T}
\end{equation}
while for $\beta<\beta_c$, $m^*=0$ is the only allowed solution.
Notice that the critical inverse temperature $\beta_c $ is determined by solving the transcendental equation:
\begin{equation}
2\beta_c=1+e^{-2\beta_cg_z}.
\end{equation}
On the other hand, for the calculation of $\Tr\left[ \left( ( e^{-\beta H} )^{T_B} \right)^{n_e} \right]$, since each site are separable, taking partial transpose over $B$ amounts to only taking the partial transpose on the two spins labelled by $2,4$ on the $s$-th site:
\begin{equation}\label{eq:if_transposed_trace}
\left[ e^{-\beta H} \right]^{T_B} =\sqrt{\frac{ \beta L}{ \pi }} \int dm e^{ -\beta L m^2 - \beta \sum_{i \neq s } H_i(m)} \left[ e^{ -\beta H_s(m) }\right]^{T_B} .
\end{equation}
By introducing $n_e$ replicas, we have
\begin{equation}\label{eq:nth_moment}
\begin{split}
\Tr{ \left[ \left( e^{-\beta H} \right) ^{T_B} \right]^{n_e}} &=\left(\frac{\beta L}{\pi}\right)^{\frac{n_e}{2}}\int \prod_{a=1}^{n_e} dm_{a} e^{-\beta L\sum_{a=1}^{n_e} m_a^ 2} \Tr_{i\neq s}\left\{ e^{-\beta\sum_{a=1}^{n_e} \sum_{i\neq s} H_i(m_a) } \right\} \Tr_{s}\left\{ \prod_{a=1}^{n_e} \left[ e^{-\beta H_s (m_a)} \right]^{T_B} \right\}\\
&=\left(\frac{\beta L}{\pi}\right)^{\frac{n_e}{2}}\int \prod_{a=1}^{n_e} dm_{a} e^{-\beta LF_{n_e}(\{ m_a\} )} \frac{ \Tr_{s} \left\{\prod_{a=1}^{n_e} \left[ e^{-\beta H_s (m_a)} \right]^{T_B} \right\}}{ \Tr_{s}\left\{ \prod_{a=1}^{n_e} e^{-\beta H_s (m_a)} \right\}}
\end{split}
\end{equation}
where
\begin{equation}
\beta F_{n_e}(\{m_a \} )= \sum_{a=1}^n m_a^2 - \log \left[e^{\beta n_eg_z} \cosh(2 \beta \sum_{a=1}^{n_e} m_a ) + e^{-\beta n_e g_z} \right] -\log \left[ 8 \cosh[2](\beta n_eg_x) \right].
\end{equation}
This multi-dimensional integral is again dominated by saddle points $\{m_a^*\vert a=1,2,\cdots, n_e \}$, which can be obtained from $\eval{\frac{\partial F_{n_e}(\{m_a \} )}{\partial m_a}}_{m_a^*} =0$:
\begin{equation}
\frac{\sinh(2\beta \sum_{a=1}^{n_e}m^*_a)}{ \cosh(2\beta\sum_{a=1}^{n_e} m^*_a ) +e^{-2\beta n_eg_z} } =m^*_a \quad \forall a.
\end{equation}
Assuming replica symmetry is preserved, we have $m^*_{n_e}=m^*_a~ \forall a$ with
\begin{equation}
\frac{\sinh(2n_e\beta m^*_{n_e})}{ \cosh(2n_e\beta m^*_{n_e} ) +e^{-2\beta n_eg_z} } =m^*_{n_e}.
\end{equation}
As $n_e\to 1$, the above equation is exactly the saddle point equation for the thermal partition function (Eq.\ref{eq:inf_stabilizer_saddle}). This implies $\lim_{n_e\to 1} m^*_{n_e} = m^*$.
By plugging Eq.\ref{eq:inf_stabilizer_Z} and Eq.\ref{eq:nth_moment} into Eq.\ref{eq:inf_nega}, one finds
\begin{equation}
\norm{\rho^{T_B}}_1= \frac{ \int dm e^{- \beta Lf(m,g_z,g_x)} \norm{\rho^{T_B}_s(m)}_1 }{ \int dm e^{- \beta Lf(m,g_z,g_x)} },
\end{equation}
where
\begin{equation}
\rho_s(m)\equiv \frac{ e^{-\beta H_s(m)}}{ \Tr_s\left\{ e^{-\beta H_s(m)}\right\}}.
\end{equation}
For $T>T_c$, there is an unique saddle point $m^*$, and
\begin{equation}
\norm{\rho^{T_B}}_1= \norm{\rho^{T_B}_s(m^*)}_1 \frac{ \int dm e^{- \beta Lf(m,g_z,g_x)} }{ \int dm e^{- \beta Lf(m,g_z,g_x)} }= \norm{\rho^{T_B}_s(m^*)}_1.
\end{equation}
For $T<T_c$, there are two saddle points $m^*=\pm m_0$, and thus we arrive at
\begin{equation}
\norm{\rho^{T_B}}_1= \frac{ \norm{\rho^{T_B}_s(m_0)}_1 \int_{\text{around}~ m_0} dm e^{- \beta Lf(m,g_z,g_x)} +\norm{\rho^{T_B}_s(-m_0)}_1 \int_{\text{around}~ -m_0} dm e^{- \beta Lf(m,g_z,g_x)} }{ \int_{\text{around} ~m_0} dm e^{- \beta Lf(m,g_z,g_x)} + \int_{\text{around} ~-m_0} dm e^{- \beta Lf(m,g_z,g_x)} }.
\end{equation}
Since $\norm{\rho^{T_B}_s(m_0)}_1 $ = $\norm{\rho^{T_B}_s(-m_0)}_1 $, we have
\begin{equation}
\norm{\rho^{T_B}}_1= \norm{\rho^{T_B}_s(m^*)}_1.
\end{equation}
This result implies that to calculate the bi-partite negativity between $A$ and $B$, it is sufficient to calculate the reduced density matrix for $s$-th site ($\rho_s$) where we made an entanglement cut. Incidentally, the above calculation explicitly demonstrates the claim $E_N(h = 0) = E_N(h=0^{+})$ mentioned in the main text where $E_N(h = 0)$ is the negativity in the absence of an infinitesimal symmetry breaking field (so that it receives contribution from both $m_0$ and $-m_0$) while $E_N(h=0^{+})$ is the negativity in the presence of such a field so that it receives contribution only from one saddle point (say, $m_0$).
From now on, we suppress lattice site index $s$ in the calculation since only four qubits on a single site is relevant. Meanwhile, $m$ will replace $m^*$ as the mean-field order parameter for brevity. The local density matrix is
\begin{equation}
\rho_{s}=\frac{1}{Z_{s}}e^{-\beta H_{s}}=\frac{1}{Z_{s}} e^{ \beta m ( Z_{1} Z_{2} + Z_{3} Z_{4} ) + \beta g_z Z_{1} Z_{2} Z_{3} Z_{4} +\beta g_x \left(X_{1} X_{2} +X_{3} X_{4} \right) },
\end{equation}
where the partition function $Z_s$ is
\begin{equation}
Z_s= \Tr e^{-\beta H_s}= 8\left( \cosh(\beta g_x)\right)^2 \left( e^{\beta g_z}\cosh(2\beta m ) +e^{-\beta g_z } \right)
\end{equation}
By taking partial transpose over $\{ 2,4\} \in B$, we have
\begin{equation}
\begin{split}
\left( e^{-\beta H_{s}} \right)^{T_{24}}& = e^{\beta g_z Z_1Z_2Z_3Z_4} \left[ (\cosh(\beta g_x))^2 e^{\beta m(Z_1Z_2+Z_3Z_4)} + (\sinh(\beta g_x))^2 e^{- \beta m(Z_1Z_2+Z_3Z_4)} X_1X_2X_3X_4 \right] \\
& +\frac{1}{2} \sinh(2\beta g_x) e^{-\beta g_z Z_1Z_2Z_3Z_4} \left[ e^{\beta m (-Z_1Z_2+Z_3Z_4) }X_1X_2 + e^{\beta m (Z_1Z_2-Z_3Z_4) }X_3X_4 \right].
\end{split}
\end{equation}
Due to the simple form of $\left( e^{-\beta H_{s}} \right)^{T_{24}}$, we are able to obtain all the eigenvalues of $\rho_s^{T_{24}}$, and exploit the following formula to calculate the negativity:
\begin{equation}\label{eq:nega}
E_N=\log \left[\sum_{i} \abs{\nu_i} \right] = \log \left[1-2\sum_{\nu_i<0} \nu_i \right],
\end{equation}
where $\{\nu_i \} $ denotes eigenvalues of $\rho_s^{T_{24}}$.
Since $Z_1Z_2,~Z_3Z_4,~X_1X_2,~X_3X_4 $ commute with each other, the corresponding eigenvalues of these operators $z_{12},~z_{34},~x_{12},~x_{34}=\pm 1 $ completely specify an eigenvector of $\left( e^{-\beta H_{s}} \right)^{T_{24}}$, which takes the following form
\begin{equation}
\ket{\psi }=\frac{1}{2} \left( \ket{s_1,s_2 } \pm \ket{-s_1,-s_2 } \right) \otimes \left( \ket{s_3,s_4 } \pm \ket{-s_3,-s_4 } \right).
\end{equation}
with $s_i=\pm1 $ for $i=1,2,3,4$. With this observation, the eigenvalues of $\left( e^{-\beta H_{s}} \right)^{T_{24}}$ can be obtained by replacing operators by their eigenvalues:
\begin{equation}
\begin{split}
\lambda(z_{12},z_{34},x_{12},x_{34})=&e^{\beta g_z z_{12}z_{34} } \left[ (\cosh(\beta g_x))^2 e^{\beta m(z_{12}+z_{34} )} + (\sinh(\beta g_x))^2 e^{- \beta m(z_{12}+z_{34})} x_{12}x_{34} \right] \\
& +\frac{1}{2} \sinh(2\beta g_x) e^{-\beta g_z z_{12} z_{34}} \left[ e^{\beta m (-z_{12} +z_{34}) }x_{12} + e^{\beta m (z_{12}- z_{34}) }x_{34} \right].
\end{split}
\end{equation}
For $T>T_c$, $ m=0$, one finds
\begin{equation}
\begin{split}
\lambda(z_{12},z_{34},x_{12},x_{34})=& e^{\beta g_z z_{12} z_{34}} \left[ (\cosh(\beta g_x))^2 + (\sinh(\beta g_x))^2 z_{12} z_{34} \right] +\frac{1}{2} \sinh(2\beta g_x) e^{-\beta g_z z_{12} z_{34}} \left[ x_{12} + x_{34} \right].
\end{split}
\end{equation}
When
\begin{equation}\label{eq:combination}
\begin{cases}
z_{12}=1,~z_{34}=-1,~x_{12}=-1,~x_{34} =-1\\
z_{12}=-1,~z_{34}=1,~x_{12}=-1,~x_{34} =-1,
\end{cases}
\end{equation}
we can have negative $\lambda$:
\begin{equation}
\lambda=e^{-\beta g_z}\cosh\left(2\beta g_x \right) -e^{\beta g_z} \sinh(2\beta g_x) .
\end{equation}
Thus, for $T>T_c$, the two-fold degenerate negative eigenvalue of $\rho^{T_{24}}_s$ is
\begin{equation}
\nu=\frac{e^{-\beta g_z}\cosh\left(2\beta g_x \right) -e^{\beta g_z} \sinh(2\beta g_x) }{ 16\left( \cosh(\beta g_x)\right)^2 \cosh\left(\beta g_z \right)},
\end{equation}
and the negativity can be obtained by using Eq.\ref{eq:nega} :
\begin{equation}
E_N =\log \left[ 1+\max\left\{0,\frac{e^{\beta g_z} \sinh(2\beta g_x)-e^{-\beta g_z}\cosh\left(2\beta g_x \right) }{ 4\left( \cosh(\beta g_x)\right)^2 \cosh\left(\beta g_z \right)} \right\} \right].
\end{equation}
Note that at $T_c$, one requires
\begin{equation}
e^{-2\beta_c g_z} < \tanh(2\beta_cg_x)
\end{equation}
to have non-zero negativity. This is always achievable by tuning $g_x$ since $\beta_c$ is only determined by $g_z$.
For $T<T_c$, depending on the values of $m$, there could be more choices of $(z_{12},z_{34},x_{12},x_{34})$ that can give negative eigenvalues of $\rho_s^{T_{24}}$. For simplicity, we consider $T\to T_c^{-}$, where $ m \sim \sqrt{T_c-T} \to 0^+$, and only the configurations in Eq.\ref{eq:combination} can possibly give negative eigenvalues. This is sufficient for our purpose since we only concern the possibly non-analytic behavior of the negativity. Therefore, as $T\to T_c^{-}$, the two-fold degenerate negative eigenvalue of $\rho_s^{T_{24}}$ is
\begin{equation}
\nu=\frac{e^{-\beta g_z}\cosh\left(2\beta g_x \right) -e^{\beta g_z} \sinh(2\beta g_x) \cosh(2\beta m) }{ 8\left( \cosh(\beta g_x)\right)^2 \left( e^{\beta g_z}\cosh(2\beta m ) +e^{-\beta g_z } \right)}.
\end{equation}
Finally, the negativity valid for $T>T_c^-$ is given by
\begin{equation}
\boxed{
E_N=\log \left[ 1+\max\left\{0,\frac{e^{\beta g_z} \sinh(2\beta g_x)\cosh(2\beta m)-e^{-\beta g_z}\cosh\left(2\beta g_x \right) }{ 2\left( \cosh(\beta g_x)\right)^2 \left(e^{\beta g_z}\cosh(2\beta m ) +e^{-\beta g_z} \right) } \right\} \right] }
\end{equation}
Due to the singular behavior of $m(T)$:
\begin{equation}
m=\begin{cases}
a\sqrt{T_c-T} \quad \text{for }\quad T\to T_c^-\\
0 \quad \text{for } \quad T>T_c,
\end{cases}
\end{equation}
the negativity $E_N$ is also a singular function across $T_c$.
\subsection{2b. Two dimensional Commuting Projector Hamiltonian}
Consider a two dimensional lattice, where each sites has two spins labelled by `a' and `b' respectively, the model Hamiltonian is
\begin{equation}
H=-\sum_{\expval{ij}} \widetilde{z}_i \widetilde{z}_j\ -g\sum_{i} \widetilde{x}_i,
\end{equation}
where $\widetilde{z}_i\equiv Z_{ia}Z_{ib}, \widetilde{x}_i\equiv X_{ia}X_{ib}$. Consider a thermal density matrix $\rho_T\sim \exp{-\beta H}$, here we present the calcualtion of the negativity between one spin on a single site, say, `a' spin in site 0 (subsystem $A$), and its complement (subsystem $B$). As discussed above, to calculate the negativity, we only need the reduced density matrix for spins at the boundary which in this case are the spins at site 0 and its neighboring sites (labelled as 1,2,3,4 clockwise). The corresponding reduced density matrix on these five sites is
\begin{equation}
\begin{split}
\rho=&A' e^{-\beta g \left( \widetilde{x}_1 +\widetilde{x}_2+\widetilde{x}_3+\widetilde{x}_4 \right) } \bigg[ \cosh(\beta g ) e^{\beta \widetilde{z}_0 \left( \widetilde{z}_1+\widetilde{z}_2+\widetilde{z}_3+\widetilde{z}_4 \right)} + \sinh(\beta g) e^{\beta \widetilde{z}_0 \left( \widetilde{z}_1+\widetilde{z}_2+\widetilde{z}_3+\widetilde{z}_4 \right)} \widetilde{x}_0 \bigg] \\
& \left[ 1+c_1\left( \widetilde{z}_1 \widetilde{z}_2 + \widetilde{z}_2 \widetilde{z}_3 + \widetilde{z}_3 \widetilde{z}_4 + \widetilde{z}_4 \widetilde{z}_1 \right) + c_2 \left( \widetilde{z}_1 \widetilde{z}_3 + \widetilde{z}_2 \widetilde{z}_4 \right) + c_3 \widetilde{z}_1 \widetilde{z}_2 \widetilde{z}_3 \widetilde{z}_4 \right] .
\end{split}
\end{equation}
Here $A'$ is determined by demanding $\Tr\rho=1$ and $c_1=\expval{\widetilde{z}_j\widetilde{z}_{j+1}}$; $c_2=\expval{\widetilde{z}_j\widetilde{z}_{j+2}}$; $c_3=\expval{\widetilde{z}_1\widetilde{z}_2 \widetilde{z}_3\widetilde{z}_4 } $, where the expectation values are taken with respect to the bulk thermal density matrix
$\rho_{\text{bulk}}\sim \exp{-\beta (H_A+H_B)}$. In fact, due to the property of commuting local terms, $c_i$ can be obtain by considering the thermal state of a bulk classical Hamiltonian,i.e. $g=0$, with one spin per site, and one just need to replace the composite operator $\widetilde{z}_i$ by a Pauli Z operator at site $i$ (i.e. $Z_i$). For instance,
\begin{equation}\label{eq:append_equiv}
c_1=\expval{\widetilde{z}_j\widetilde{z}_{j+1}} = \frac{ \Tr{ \widetilde{z}_j \widetilde{z}_{j+1} e^{\beta \sum_{\expval{ij}} \widetilde{z}_i \widetilde{z}_j\ +\beta g\sum_{i} \widetilde{x}_i } } }{\Tr e^{\beta \sum_{\expval{ij}} \widetilde{z}_i \widetilde{z}_j\ +\beta g\sum_{i} \widetilde{x}_i } } = \frac{ \Tr{ Z_jZ_{j+1} e^{\beta \sum_{\expval{ij}} Z_i Z_j } } }{ \Tr e^{\beta \sum_{\expval{ij}} Z_i Z_j } }
\end{equation}
Under the partial transposition over $B$, the density matrix is
\begin{equation}
\begin{split}
\rho^{T_B}=&A' e^{-\beta g \left( \widetilde{x}_1 +\widetilde{x}_2+\widetilde{x}_3+\widetilde{x}_4 \right) } \bigg[ \cosh(\beta g) e^{\beta \widetilde{z}_0 \left( \widetilde{z}_1+\widetilde{z}_2+\widetilde{z}_3+\widetilde{z}_4 \right)} + \sinh(\beta g) e^{ - \beta \widetilde{z}_0 \left( \widetilde{z}_1+\widetilde{z}_2+\widetilde{z}_3+\widetilde{z}_4 \right) } \widetilde{x}_0 \bigg] \\
& \left[ 1+c_1\left( \widetilde{z}_1 \widetilde{z}_2 + \widetilde{z}_2 \widetilde{z}_3 + \widetilde{z}_3 \widetilde{z}_4 + \widetilde{z}_4 \widetilde{z}_1 \right) + c_2 \left( \widetilde{z}_1 \widetilde{z}_3 + \widetilde{z}_2 \widetilde{z}_4 \right) + c_3 \widetilde{z}_1 \widetilde{z}_2 \widetilde{z}_3 \widetilde{z}_4 \right] ,
\end{split}
\end{equation}
The eigenvalues of $\rho^{T_B}$ can be obtained by just replacing $\widetilde{x}_i, \widetilde{z}_i$ by $\pm 1$. In fact, $e^{-\beta g \left( \widetilde{x}_1 +\widetilde{x}_2+\widetilde{x}_3+\widetilde{x}_4 \right) }$ is irrelevant since it just provides a multiplicative factor when summing negative eigenvalues, which got cancelled out by the normalization factor. Effectively, it is sufficient to consider the eigenvalues
\begin{equation}
\begin{split}
\lambda =&A \bigg[ \cosh(\beta g) e^{\beta \widetilde{z}_0 \left( \widetilde{z}_1+\widetilde{z}_2+\widetilde{z}_3+\widetilde{z}_4 \right)} + \sinh(\beta g) e^{ - \beta \widetilde{z}_0 \left( \widetilde{z}_1+\widetilde{z}_2+\widetilde{z}_3+\widetilde{z}_4 \right) } \widetilde{x}_0 \bigg] \\
& \left[ 1+c_1\left( \widetilde{z}_1 \widetilde{z}_2 + \widetilde{z}_2 \widetilde{z}_3 + \widetilde{z}_3 \widetilde{z}_4 + \widetilde{z}_4 \widetilde{z}_1 \right) + c_2 \left( \widetilde{z}_1 \widetilde{z}_3 + \widetilde{z}_2 \widetilde{z}_4 \right) + c_3 \widetilde{z}_1 \widetilde{z}_2 \widetilde{z}_3 \widetilde{z}_4 \right] ,
\end{split}
\end{equation}
where $\widetilde{x}_0$ and each $\widetilde{z}_i$ takes $\pm 1$, which gives $2^6=64$ eigenvalues, and $A$ is chosen such that the sum of these 64 eigenvalues remains unity.
$\left[ 1+c_1\cdots \right]$ part is always non-negative since it is obtained by performing partial trace for a density matrix ( positive semidefinite ). As a result, $\lambda$ can be negative only when $ \widetilde{x}_0=-1$ and $e^{2 \beta \widetilde{z}_0 \left( \widetilde{z}_1+\widetilde{z}_2+\widetilde{z}_3+\widetilde{z}_4 \right) } < \tanh(\beta g)$.
For a given $g$, there are many choices of $\widetilde{z}_i$ that can result in negative eigenvalues. As our purpose is to check whether the negativity picks up an singularity at a thermal critical point, it is sufficient to restrict $g$ in a range such that only a few eigenvalues are negative. We set $g$ in the range $e^{-8\beta} < \tanh(\beta g) < e^{-4\beta }$, and there are only two negative eigenvalues given by
\begin{equation}
\begin{cases}
&\widetilde{z}_0=1, ~~~ \widetilde{z}_1= \widetilde{z}_2= \widetilde{z}_3= \widetilde{z}_4=-1 \\
&\widetilde{z}_0=-1, ~~~ \widetilde{z}_1= \widetilde{z}_2= \widetilde{z}_3= \widetilde{z}_4=1 .
\end{cases}
\end{equation}
Finally, as
\begin{equation}
e^{-8\beta} < \tanh(\beta g) < e^{-4\beta },
\end{equation}
we obtain the expression of the negativity:
\begin{equation}\label{eq:negativity_single}
\boxed{
E_N=\log\left\{ 1-4A \left[ \cosh(\beta g) e^{-4\beta } - \sinh(\beta g) e^{4\beta} \right] \left( 1+4c_1+2c_2+c_3 \right) \right\}
}.
\end{equation}
\begin{equation}
A^{-1}=2^5\cosh(\beta g )\left[ \cosh[4](\beta )+\left( c_1+\frac{1}{2}c_2 \right) \sinh[2](2\beta) +c_3\sinh[4]( \beta ) \right]
\end{equation}
Due to the singularity of $c_i$ at the thermal critical point, the negativity $E_N$ is expected to be singular. To confirm this intuition, we now adopt a mean-field approach to calculate the coefficient $c_1,c_2,c_3$. The exact nature of singularities associated with $c_i$ for our model would of course be determined by the critical exponents of the 2D Ising model. As shown in Eq.\ref{eq:append_equiv}, $c_i$ is exactly given by the corresponding classical Hamiltonian with one spin per site. As a result, we consider the mean-field Hamiltonian
\begin{equation}
H=-(3m+Z_0) \left( Z_1+Z_2+ Z_3+ Z_4 \right),
\end{equation}
we determine $m$ from $m=\expval{Z_i}=\Tr{\rho Z_i}$ for $i=1$ to $4$, where $\rho$ is a density matrix associated with $H$. It is straightforward to obtain the mean-field equation for $m$:
\begin{equation}
m=\frac{ \cosh[4](\beta (3m+1)) \tanh(\beta (3m+1)) + \cosh[4](\beta (3m-1)) \tanh(\beta (3m-1)) }{\cosh[4](\beta (3m+1)) + \cosh[4](\beta (3m-1)) }.
\end{equation}
$T_c$ can be determined from this equation, and it is straightforward to show that $m=0$ as $T\to T_c^+$, and $m \sim \sqrt{T_c-T}$ as $T\to T_c^-$.
Finally, $c_1,c_2,c_3$ can be obtained:
\begin{equation}
\begin{split}
&c_1=c_2=\expval{Z_1Z_2} =\frac{ \cosh[2](\beta (3m+1)) \sinh[2](\beta (3m+1)) + \cosh[2](\beta (3m-1)) \sinh[2](\beta (3m-1)) }{\cosh[4](\beta (3m+1)) + \cosh[4](\beta (3m-1)) }\\
&c_3=\expval{Z_1Z_2Z_3Z_4} = \frac{ \sinh[4](\beta (3m+1)) + \sinh[4](\beta (3m-1)) }{\cosh[4](\beta (3m+1)) + \cosh[4](\beta (3m-1)) }
\end{split}
\end{equation}
Plug the coefficients into Eq.\ref{eq:negativity_single}, and expand it for small $m$,
\begin{equation}
E_N=\log\left\{ 1-4 \left[ \cosh(\beta g) e^{-4\beta } - \sinh(\beta g) e^{4\beta} \right] \left\{ \frac{16\cosh(4\beta )}{1+4\cosh(4\beta +\cosh(8\beta ))} +\frac{1728\beta^2 \left[ 1+6\cosh(4\beta ) +\cosh(8\beta )\right]m^2 }{\left[ 1+4\cosh(4\beta +\cosh(8\beta )) \right]^2} \right\} \right\}.
\end{equation}
There the negativity $E_N$ is manifestly singular at $T_c$ due to the singularity from $m$.
\subsection{2c. Quantum Spherical Model }
Consider the Hamiltonian for a quantum spherical model: $H=\frac{1}{2} g \sum_{i=1}^N p_i^2 -\frac{1}{2N} \sum_{i,j=1}^N x_ix_j + \mu \left[ \sum_{i=1}^N x_i^2 -\frac{N}{4} \right] $ where $[x_i,p_j] = i \delta_{ij}$. $\mu$ is chosen so that $\left< \sum_{i=1}^N x_i^2 \right> =\frac{N}{4}$ where the expectation value is taken with respect to the thermal density matrix. Define $x_k=\frac{1}{\sqrt{N}} \sum_{j}e^{ikj} x_j$, $p_k=\frac{1}{\sqrt{N}} \sum_{j}e^{ikj} p_j$ and introduce $a_k,a_k^{\dagger}$: $p_k=-i\sqrt{ \frac{\omega_{k}}{2g}} \left( a_k - a_{-k} ^{\dagger} \right) $,$x_k=\sqrt{ \frac{g}{2\omega_{k}}} \left( a_k + a_{-k} ^{\dagger} \right)$
the Hamiltonian can be diagonalized:
\begin{equation}\label{eq:spherical_diag_H}
H=\sum_{k} \omega_k \left(a_k^{\dagger} a_k +\frac{1}{2}\right)-\frac{\mu}{4}N,
\end{equation}
where the single particle energy $\omega_k$ is
\begin{equation}
\omega_k=\begin{cases}
\omega_0=\sqrt{2g (\mu-\frac{1}{2})} \quad \text{for} \quad k=0\\
\omega_1=\sqrt{2g \mu} \quad ~~~~~~~~ \text{for} \quad k\neq 0,
\end{cases}
\end{equation}
Note that in order to have a stable theory , $\mu\geq \frac{1}{2}$. From Eq.\ref{eq:spherical_diag_H}, the free energy density $f$ can be calculated:
\begin{equation}
f=\frac{1}{N\beta} \log\left[ 2\sinh (\frac{1}{2}\beta \omega_0) \right] +\frac{N-1}{N\beta} \log\left[ 2\sinh (\frac{1}{2}\beta \omega_1) \right]
-\mu/4.
\end{equation}
$\mu$ is determined from $\left< \sum_{i=1}^N x_i^2 \right> =\frac{N}{4}$, which is equivalent to $\frac{\partial f}{\partial \mu}=0$:
\begin{equation}
\frac{1}{2N} \sqrt{\frac{g}{2(\mu-\frac{1}{2})}} \coth\left( \frac{1}{2} \beta \sqrt {2g(\mu-\frac{1}{2})}\right) +\frac{N-1}{2N} \sqrt{\frac{g}{2\mu}} \coth\left( \frac{1}{2} \beta \sqrt {2g\mu}\right) = \frac{1}{4}.
\end{equation}
In the thermodynamic limit $N\to \infty$, $\mu$ is a singular function of $\beta,g$. For $2\sqrt{g} \coth(\frac{1}{2}\beta \sqrt{g})>1$, the system is in a disordered phase, with $\mu$ determined from
\begin{equation}\label{eq:spherical_mu}
\sqrt{\frac{g}{2\mu}} \coth(\frac{1}{2}\beta \sqrt{2g\mu})=\frac{1}{2},
\end{equation}
while the condition $2\sqrt{g} \coth(\frac{1}{2}\beta \sqrt{g})<1$ gives the ordered phase, and $\mu$ is pinned to $\frac{1}{2}$.
Here we brief describe the covariance matrix formalism for calculating the negativity of a Gaussian state $\rho$ for $N$ degrees of freedom. First we calculate the covariance matrix in displacements $(\gamma_x)_{ij}=\expval{\{ x_i-\overline{x}_i ,x_j-\overline{x}_j \}} $ and the covariance matrix in momenta $(\gamma_p)_{ij}=\expval{\{ p_i-\overline{p}_i ,p_j-\overline{p}_j\}} $, where $\overline{x}_i=\textrm{tr}{\rho x_i}$, $\overline{p}_i=\textrm{tr}{\rho p_i}$, and $\{ A,B\}=AB+BA$ is the anticommutator. Define the subsystem $A$ composed by degrees of freedom for site $i=1,2,\cdots, N_A$ and the complement $B$ composed by the rest of sites, we calculate $\tilde{\gamma}=\gamma_xR\gamma_p R$, where $R$ is diagonal matrix with $1$ for the first $N_A$ diagonal entries and $-1$ for the rest of the diagonal entries. By diagonalizing $\tilde{\gamma}$, we obtain its eigenvalues $\{\nu_i \vert i=1,2,\cdots, N \}$, from which the negativity $E_N$ can be calculated
\begin{equation}
E_N(\rho)=\sum_{i=1}^{N}\text{max}\{0,-\log \nu_i\}.
\end{equation}
For the thermal state of the spherical model, a straightforward calculation shows that
\begin{equation}
\begin{aligned}\label{eq:correlation}
&\left(\gamma_x \right)_{ij}=2\expval{x_i x_j} =m_x +\delta_{ij} d_x\\
&\left(\gamma_x \right)_{ij}=2\expval{p_i p_j} =m_p +\delta_{ij} d_p,
\end{aligned}
\end{equation}
with
\begin{equation}
\begin{aligned}
&m_x\equiv \frac{1}{N} \left[ \sqrt{\frac{g}{2\mu-1}} \coth\left(\frac{1}{2} \beta \sqrt{(2\mu-1)g} \right) -\sqrt{\frac{g}{2\mu}} \coth\left( \frac{1}{2}\beta \sqrt{2\mu g} \right) \right] \\
&d_x\equiv \sqrt{\frac{g}{2\mu}} \coth\left( \frac{1}{2}\beta \sqrt{2\mu g} \right)\\
&m_p\equiv\frac{1}{N} \left[ \sqrt{\frac{2\mu-1}{g}} \coth\left(\frac{1}{2} \beta \sqrt{(2\mu-1)g} \right) -\sqrt{\frac{2\mu}{g}} \coth\left( \frac{1}{2}\beta \sqrt{2\mu g} \right) \right] \\
&d_p\equiv \sqrt{\frac{2\mu}{g}} \coth\left( \frac{1}{2}\beta \sqrt{2\mu g} \right).
\end{aligned}
\end{equation}
Thus we have
\begin{equation}
\tilde{\gamma}=\gamma_xR\gamma_p R=d_xd_p \mathds{1}_N + m_xd_pJ_N + m_pd_x
\begin{pmatrix}
J_{N/2} & -J_{N/2}\\
-J_{N/2} & J_{N/2},
\end{pmatrix}
\end{equation}
where we define $J_N$ as an $N\cross N$ all-ones matrix. All three matrices on the R.H.S. commute with each other so they can be diagonalized with the same set of eigenvectors. Since both the second and the third matrix are rank-1 matrix, it is easy to calculate the eigenvalues. Finally, the eigenvalues of $\tilde{\gamma}$ are
\begin{equation}
\nu_k= \begin{cases}
d_xd_p = \left[ \coth\left( \frac{1}{2}\beta \sqrt{2\mu g } \right) \right]^2 \quad \text{for} \quad k=1,2, \cdots N-2\\
d_xd_p+ Nm_xd_p =\sqrt{\frac{2\mu}{2\mu-1}} \coth\left(\frac{1}{2} \beta \sqrt{ (2\mu-1)g} \right) \coth\left( \frac{1}{2}\beta \sqrt{2\mu g } \right) \quad \text{for} \quad k=N-1 \\
d_xd_p+ Nm_pd_x=\sqrt{\frac{2\mu-1}{2\mu}} \coth\left(\frac{1}{2} \beta \sqrt{ (2\mu-1)g} \right) \coth\left( \frac{1}{2}\beta \sqrt{2\mu g } \right) \quad \text{for} \quad k=N
\end{cases}
\end{equation}
One can check that $\nu_k>1 $ for $k=1,2,\cdots, N-1$ for all values of parameters in the model, and only $\nu_{N}$ can be less than $1$ to contribute to the entanglement negativity:
\begin{equation}
\boxed{
E_N=\text{Max}\{ 0, -\log \nu \} }
\end{equation}
where
\begin{equation}
\nu\equiv \nu_N=\sqrt{\frac{2\mu -1}{2\mu}} \coth\left[ \frac{1}{2} \beta \sqrt{\left(g(2\mu-1)\right)} \right] \coth\left[ \frac{1}{2} \beta\sqrt{2g\mu} \right].
\end{equation}
By using Eq.\ref{eq:spherical_mu} in the disordered phase, and $\mu=\frac{1}{2}$ in the ordered phase, $\nu$ can be further simplified:
\begin{equation}
\nu=\begin{cases}
\frac{2}{\beta\sqrt{g}} \coth(\frac{1}{2}\beta \sqrt{g}) \quad \quad \quad \quad \quad \quad \quad \text{for ordered phase} \\
\frac{1}{2} \sqrt{\frac{2\mu-1}{g}} \coth(\frac{1}{2}\beta \sqrt{(2\mu-1)g}) \quad \text{for disordered phase}.
\end{cases}
\end{equation}
To study the singularity of $E_N$ at the critical point, we calculate the first derivative of $E_N$ with respect to $g$ to observe its discontinuity at a critical point:
\begin{equation}
\begin{split}
&\eval{\frac{ \partial E_N}{\partial g}}_{g_c^{+}}= \frac{1}{g_c} +\frac{\beta_c^2}{12} \left(1 - \frac{8}{4 + \beta_c - 4 \beta g_c} \right)\\
&\eval{\frac{ \partial E_N}{\partial g}}_{g_c^{-}}=\frac{4 + \beta_c - 4 \beta_c g_c}{8 g_c} ,
\end{split}
\end{equation}
\section{3. Entanglement of Formation in a Infinite-Range Commuting Projector Hamiltonian}
To begin with, we recall the definition of the entanglement of of formation: a density matrix $\rho$ acting on a bipartite Hilbert space $\mathcal{H}=\mathcal{H}_A\otimes \mathcal{H}_B$ can be decomposed as a convex sum of pure states
\begin{equation}
\rho=\sum_k P_k\ket{k}\bra{k},
\end{equation}
and for each $\ket{k}$, we can calculate the reduced density matrix on $A$: $\rho_k^A =\Tr_{B} \ket{k} \bra{k}$, from which the entanglement entropy $S_A(\ket{k})$ is obtained: $S_A(\ket{k})= -\Tr_A \rho_k^A \log \rho_k^A$. The entanglement of formation $E_F(A,B)$ is defined as
\begin{equation}
E_F(A,B)=\text{min} \sum_k P_k S_A(\ket{k}),
\end{equation}
where minimization over all possible pure state decomposition is taken. Here we provide a model, where the entanglement of formation can be calculated analytically by showing its upper and lower bound coincide in the thermodynamic limit. Consider a one-dimensional lattice of size $L$ where each lattice site has two qubits, the model Hamiltonian is
\begin{equation}
H=-\frac{1}{2L} \left( \sum_{i=1}^L Z_{i1} Z_{i2} \right)^2 -g \sum_{i=1}^L X_{i1} X_{i2} .
\end{equation}
The density matrix at inverse temperature $\beta$ is $\rho=\frac{1}{Z}e^{-\beta H}$ with $Z=\Tr e^{-\beta H}$. We make an entanglement cut across one of the sites (say $s$-th site) such that the two spins on $s$-th site are not in the same subsystem. In the following calculation, $A$ comprises all the lattice sites with site index $i<s$ and the spin labelled by $1$ on $s$-th site while $B$ comprises all the lattice sites with site index $i>s$ and the spin labelled by $2$ on $s$-th site. For such a bipartition scheme, we prove that the entanglement of formation $E_F$ between $A$ and $B$ is exactly that from a mean-field density matrix for just two spins, where a closed form expression for $E_F$ is available. Our strategy is to find an upper bound and a lower bound on $E_F$ that happen to match each other. \\
\noindent\textbf{\textit{Upper Bound}}\\
Entanglement of formation $E_F$ requires a minimization scheme over all possible pure state decompositions. By considering a particular way of decomposition, we thus give an upper bound for $E_F$. First we perform the Hubbard-Stratonovich transformation for $\rho$:
\begin{equation}\label{eq:hs_entanglement}
\rho =\frac{1}{Z}e^{-\beta H} =\frac{1}{Z}\sqrt{\frac{ \beta L}{2 \pi }} \int dm e^{ -\frac{1}{2}\beta L m^2 - \beta \sum_{i=1 }^L H_i(m)},
\end{equation}
where a local Hamiltonian $H_i(m)$ for $i$-site of two spins is defined as :
\begin{equation}
H_i(m)= - m Z_{i1} Z_{i2} -gX_{i1} X_{i2}.
\end{equation}
Each $e^{-\beta H_i(m)}$ can be decomposed: $e^{-\beta H_i(m)}=\sum_{k_i} w^i_{k_i} (m)\ket{k_i(m)} \bra{k_i(m)} $. As a result,
\begin{equation}
\rho= \sum_{\{k_i\}} \int dm \frac{1}{Z} \sqrt{ \frac{\beta L}{2\pi}} e^{-\frac{1}{2}\beta L m^2} \left(\prod_{i} w^i_{k_i}(m) \right) \ket{k_1,\cdots,k_L}\bra{k_1,\cdots,k_L}
\end{equation}
The entanglement entropy between $A$ and $B$ in $\ket{k_1,\cdots,k_L}\bra{k_i,\cdots,k_L} $ is given by the entanglement entropy between just two spins at site $s$ due to the product state structure for different sites. Therefore,
\begin{equation}
E_F(A,B) \leq \min_{\{k_i\}} \sum_{\{k_i\}} \int dm \frac{1}{Z} \sqrt{ \frac{\beta L}{2\pi}} e^{-\frac{1}{2}\beta L m^2} \left(\prod_{i} w^i_{k_i}(m) \right) S_{s1}(\ket{k_s(m)}),
\end{equation}
where $ S_{s1}(\ket{k_s(m)})$ is the entanglement entropy between spins at $s_1$ and $s_2$ in the state $\ket{k_s(m)}$, and the minimum is taken among all possible pure state decomposition of $e^{-\beta H_i(m)}$. Since $S_{s1}(\ket{k_s})$ is independent of how we decompose $e^{-\beta H_i}$ for $i\neq s$. The summation over $k_i ~\forall i\neq s$ can be performed on $w^i_{k_i}$:
\begin{equation}
\sum_{\{k_i \vert i\neq s \}} \prod_{i\neq s} w^i_{k_i} = \left(\Tr_i e^{-\beta H_i(m)} \right)^{L-1}=e^{-\beta (L-1)f(m)},
\end{equation}
where $f(m)$ is a mean-field free energy density.
Consequently,
\begin{equation}
E_F (A,B)\leq \min_{k_s} \frac{ \int dm e^{-\beta Lf(m)} \sum_{k_s} \frac{1}{Z_s } w^s_{k_s}(m) S_{s1}(\ket{k_s(m)}) } {\int dm e^{-\beta Lf(m)}},
\end{equation}
with $Z_s\equiv \Tr_s e^{-\beta H_s(m)}$. In $L\to \infty $ limit, the argument inside the summation over $k_s$ is dominated only by saddle points, and thus
\begin{equation}\label{eq:eof_result}
E_F(A,B)\leq \min \sum_{k_s}\frac{1}{Z_s} w^s_{k_s}(m^*) S_{s1}(\ket{k_s(m^*)}),
\end{equation}
where $m^*$ is a saddle point obtained by minimizing $f(m)$. Define the mean field density matrix on a single site of two spins:
\begin{equation}
\rho_s(m^*)= \frac{1}{Z_s} e^{-\beta H_s(m^*)},
\end{equation}
we show
\begin{equation}\label{eq:upper}
E_F(A,B)\leq E_F(s1,s2),
\end{equation}
i.e., the entanglement of formation between $A$ and $B$ is upper bounded by the entanglement of formation between two spins in the mean field density matrix.\\\\
\noindent\textbf{\textit{Lower Bound}}\\
As a bona fide entanglement measure, entanglement of formation is non-increasing under a partial trace. This implies that $E_F(a,b)\leq E_F(A,B)$, where $a$ and $b$ denote a subsystem in $A$ and $B$ respectively. Here we choose two spins at the sites $s$ as $a$ and $b$. A calculation shows that the reduced density matrix at site $s$ is
\begin{equation}
\rho_s=\frac{1}{Z} \Tr_{i\neq s} e^{-\beta H} = \frac{\int dm e^{-\beta L f(m)} \frac{1}{Z_s} e^{-\beta H_s(m)} }{ \int dm e^{-\beta f(m)} }= \frac{\int dme^{-\beta Lf(m)} \rho_s(m) }{ \int dm e^{-\beta f(m)} }
\end{equation}
where $f(m)=-\frac{1}{\beta } \log Z_s=-\frac{1}{\beta } \log \Tr_s e^{-\beta H_s(m)}$ being the free energy density. In $L\to \infty $ limit, $\rho_s$ is exactly given by $\rho_s(m^*)$ where the saddle point $m^*$ is the location of the global minimum of $f(m)$. One way to see this is to expand $\rho_s$ in a complete operator basis on site $s$, and show that expectation value of any operator on site $s$ is precisely given by $\rho_s(m^*)$. This calculation shows that
\begin{equation}\label{eq:lower}
E_F(s1,s2)\leq E_F(A,B).
\end{equation}
By combining Eq.\ref{eq:upper} and Eq.\ref{eq:lower}, one finds that the bi-partite entanglement of formation between $A$ and $B$ is exactly that between two spins in the mean field density matrix which can be calculated analytically using the result of Ref.\cite{wooters1997}.
\end{document}
|
\section{Introduction}
\label{sec:intro}
In this paper, we introduce a new class of multivariate and heterogeneous point process models. In doing so, we address two challenging problems in point process analysis: we propose valid and nontrivial models for multi-type point processes, an open problem in the literature, and we produce multivariate spatial models that can flexibly accommodate anisotropy in both the marginal and joint dependence structures.
We choose to build our models using log-Gaussian Cox processes \citep{Diggle83a,Moller98} as a foundation. Thus, the observed point pattern is modelled in terms of a random intensity, generated by a random field. Recent interest in random field modelling has greatly enhanced our ability to specify flexible models for multivariate patterns. We shall build on recent progress made in this area by, for example, \citet{Gneiting10}, \citet{Apanasovich12} and \citet{Genton15}, by allowing for anisotropy in the second-order dependence structure of the latent random field model.
Datasets that require anisotropic models have been common in the point process literature over the last 20 years, for example the locations of chapels in Welsh valleys \citep{Mugglestone96a,Rajala16,Rajala18a}, the epicentral locations of earthquakes in California over a 20 year period \citep{Veen06} and clustered locations of shrubs in dryland ecosystems \citep{Haase01}. The Welsh chapels and Californian earthquakes both form elliptical clusters, indicating an anisotropic second-order interaction between points in the same pattern. Meanwhile, the dryland shrub data display a directional preference in the interaction of points of different type: \citet{Haase01} found one species to grow more often than would be expected to the east of a second species. In the point process literature, it is common to accommodate heterogeneities in the observed point pattern by using a spatially homogeneous random field model to specify an intensity process conditional upon some known covariates \citep[see, e.g.~][]{Waagepetersen08,Waagepetersen09,Diggle13b}. This approach is limited in its applicability, however, when faced with heterogeneous point pattern data with no covariate measurements, or indeed when the source of heterogeneity is unknown.
Our chosen approach to accommodating anisotropy is based upon a particular form of anisotropy known as geometric anisotropy \citep{Goff88}: whereas the spatial covariance functions that drive isotropic processes have circular contours of equivariance, those that drive geometric anisotropic processes have elliptical contours of equivariance. This approach was also considered in the univariate case by \citet{Moller14}. A great advantage of this approach is that it can be used in conjunction with well-known isotropic covariance functions; our models will use Mat\'ern-based covariance structures, which will allow the user to directly specify both the range of dependence in, and the smoothness of, the resulting random field. We will also discuss and address identifiability concerns for this class of parametric models.
Once the random field has been specified, the point process is conditionally generated as a Poisson process with intensity determined by the random field. This has the advantage of automatically generating a valid set of anisotropic point processes, where the marginal and cross-pair correlation functions have an analytic form, which we provide.
We explore the restrictions that are naturally placed on all cross-pair correlation functions, where we utilise recent results for isotropic multivariate random fields due to \citet{Apanasovich12} and \citet{Gneiting10}.
By representing our multivariate process in both the spatial and spectral domains, we will also demonstrate that allowing for distinct geometric anisotropies in each marginal process places further restrictions on valid forms of the cross-dependence structures. This is an important result that yields unique insights into the possible variation of joint co-dependence in multivariate geometric anisotropic random fields, and by extension Cox processes.
Once we have understood the constraints on possible model forms, we develop new inference methods. We detail a two-stage estimation procedure in which we first estimate the anisotropy parameters, and then use these estimates to transform the data to be isotropic; this `isotropised' point pattern is then used to estimate the covariance parameters for the underlying random field model. For this second stage, \citet{Moller14} advocated the use of minimum contrast, a method of moments approach to estimation for point process models. We appeal to the likelihood principle, and develop a maximum likelihood-based approach to inference that builds on the work of \citet{Tanaka08}. Straightforward maximum likelihood estimation of the model parameters is infeasible, due to the intractability of the LGCP likelihood, however \citet{Tanaka08} showed that the intractability of the point process likelihood can be circumvented by considering the so-called Fry process \citep{Fry79}. This is a secondary point pattern formed by the difference vectors of all point pairs in the original point pattern, and it can be treated as an inhomogeneous Poisson point process, with an associated tractable likelihood. \citet{Tanaka08} showed that the Fry process likelihood can be used to perform inference for univariate, isotropic point process models. In a novel extension of this work, we use the Fry process likelihood to perform inference for anisotropic, multivariate point processes
Finally, we apply our newly-developed methodology to real data from a tropical rainforest stand on Barro Colorado Island, Panama \citep{Condit98,Hubbell99,Hubbell10}. Recent work by \citet{Waagepetersen16} and \citet{Rajala18b} has highlighted the importance of developing realistic multivariate point process models to aid the understanding of complex species interactions within this rainforest. The need to develop anisotropic methodology in particular is characterised in Figures \ref{subfig:BCIdata_int1} and \ref{subfig:BCIdata_int2}, which show the estimated intensity of \textit{Guatteria dumetorum} and \textit{Miconia hondurensis}. Their strongly anisotropic features are clear, and we also show two simulated fields from the presented multivariate model class, exhibiting similar features.
\begin{figure}[t!]
\hspace{0.01\textwidth}
\begin{subfigure}{0.4\textwidth}
\includegraphics[width=\textwidth]{./fig1v3_BCIdata.png}
\vspace{-10pt}
\vspace{-10pt}
\caption{}
\label{subfig:BCIdata_PPs}
\end{subfigure}
\hspace{0.1\textwidth}
\begin{subfigure}{0.4\textwidth}
\includegraphics[width=\textwidth]{./fig1v3_synthdata.png}
\vspace{-10pt}
\vspace{-10pt}
\caption{}
\label{subfig:synthdata_PPs}
\end{subfigure}
\begin{center}
\vspace{-10pt}
\begin{subfigure}{0.475\textwidth}
\includegraphics[width=\textwidth]{./fig1v3_BCI_lambda1.png}
\vspace{-10pt}
\vspace{-10pt}
\caption{\hspace{15pt} }
\label{subfig:BCIdata_int1}
\end{subfigure}
\hspace{0.03\textwidth}
\begin{subfigure}{0.475\textwidth}
\includegraphics[width=\textwidth]{./fig1v3_synth_lambda1.png}
\vspace{-10pt}
\vspace{-10pt}
\caption{\hspace{15pt} }
\label{subfig:synthdata_int1}
\end{subfigure}
\vspace{-20pt}
\begin{subfigure}{0.475\textwidth}
\includegraphics[width=\textwidth]{./fig1v3_BCI_lambda2.png}
\vspace{-10pt}
\vspace{-10pt}
\caption{\hspace{15pt} }
\label{subfig:BCIdata_int2}`
\end{subfigure}
\hspace{0.03\textwidth}
\begin{subfigure}{0.475\textwidth}
\includegraphics[width=\textwidth]{./fig1v3_synth_lambda2.png}
\vspace{-10pt}
\vspace{-10pt}
\caption{\hspace{15pt} }
\label{subfig:synthdata_int2}
\end{subfigure}
\end{center}
\caption{Point pattern data showing two species of tree from the BCI tropical rainforest (a; {\em Guatteria dumetorum}, blue; {\em Miconia hondurensis}, red), along with their estimated intensity fields (c,e), and simulated point pattern data (b) from two independent univariate geometric anisotropic log-Gaussian Cox processes, with their corresponding simulated intensity fields (d,f).}
\label{fig:BCIvssynthdata}
\end{figure}
Thus, to summarize, this paper provides a number of new and important insights for multivariate spatial processes, describing the complex relationships possible when allowing for distinct geometric anisotropies in each univariate component. Our understanding gives sufficient, but not necessary, conditions to yield valid multivariate random field models and, by extension, valid multivariate Cox processes.
\section{Background}
\subsection{Log-Gaussian Cox processes}
Consider the multivariate point process $X=\{X_p\in\mathbb{R}^d, p=1,\ldots,P\}$, where the index $p$ is used to denote a univariate component of the multivariate process, and suppose that we wish to use such a process to model a multi-type point pattern. We will denote the observed point pattern $X\cap W = \{x_{p,i} \in W, i=1,\ldots,n_p ; p=1,\ldots,P\}$, where $n_p\in\mathbb{N}$ is the total number of points of type $p$ observed in the observation window $W\subset\mathbb{R}^d$. Henceforth, we will also use $x_p$ to denote an arbitrary observed point of type $p$. For many applications of interest, $d=2$, however much of the multivariate framework established here can be applied to point processes defined on a space of any dimension.
We define $X$ to be a multivariate log-Gaussian Cox process \citep[LGCP;][]{Moller98}: each univariate sub-process $X_p$ is an inhomogeneous Poisson process with intensity specified by
\begin{equation}
\label{eq_LGCPintensitydef}
\Lambda_p(x) = \exp\{S_p(x)\}, \qquad x\in\mathbb{R}^d,
\end{equation}
where $S(x) = \{S_p(x), p=1,\ldots,P\}$ is a multivariate Gaussian random field (GRF). We will assume $S_p$, and therefore $X_p$, to be stationary for all $p=1,\ldots,P$, and we denote the constant mean of $S_p(x)$ by $\mu_p$. The intensity process $\Lambda_p(x)$ will therefore also have a constant mean, which we denote $\lambda_p$, and which will take the following form:
\begin{equation}
\label{eq:lamp_mup}
\lambda_p = \mathbb{E}\left\{\Lambda_p\right\} = \exp\left\{\mu_p + \sigma_{pp}/2\right\},
\end{equation}
where $\sigma_{pp}$ denotes the variance of $S_p(x)$.
Key to the definition of a multivariate LGCP is the conditional independence of its components: given its intensity process $\Lambda_p(x)$, the univariate LGCP $X_p$ is independent of $\{X_q,q=1,\ldots,P, q\ne p\}$. As a result of this property, the second-order behaviour of the point process $X$ may be entirely, and conveniently, described through the covariance structure of the multivariate GRF $S(x)$. We do so by specifying the matrix of covariance functions $\{C_{pq}(h)\}_{p,q=1}^P$, with
\begin{equation*}
C_{pq}(h) = \textrm{cov}\left\{S_{p}(x),S_{q}(x-h)\right\},\qquad x,h\in\mathbb{R}^d.
\end{equation*}
The second-order behaviour of the multivariate point process $X$ can be directly measured through the level of clustering or separation present in the resulting point pattern. The cross-pair correlation function $g_{pq}(r)$ is defined as the expected number of points from process $q$ that lie at a distance $r$ from the typical point in process $p$, and is the standard tool for measuring aggregation and segregation, both within and between processes. For a log-Gaussian Cox process, $g_{pq}(r)$ can be straightforwardly expressed in terms of the covariance structure for the underlying multivariate GRF:
\begin{equation*}
g_{pq}(h) = \exp\{C_{pq}(h)\}, \qquad h\in\mathbb{R}^d.
\end{equation*}
From this relationship, it is clear to see that $g_{pq}(h)=1$ is equivalent to $C_{pq}(h)=0$, which indicates independence between processes $p$ and $q$ at the scale $h\in\mathbb{R}^d$. Thus, for a bivariate Poisson process $\{X_p,X_q\}$, i.e.~under an assumption of complete spatial randomness, we would expect $g_{pq}(h)=1$, whereas significant departures from this would indicate aggregation ($g_{pq}(h)>1$) or segregation ($g_{pq}(h)<1$) of points from processes $p$ and $q$ at separation $h\in\mathbb{R}^d$.
The dependence structure for the multivariate GRF $S(x)$ can equivalently be described in the frequency (spectral) domain. The (cross-)spectral density function $f_{pq}(\omega)$ forms a Fourier transform pair with the (cross-)covariance function:
\begin{equation*}
f_{pq}(\omega) = \frac{1}{(2\pi)^d}\int_{\mathbb{R}^d} \exp(-i\omega^T x) C_{pq}(x)dx, \qquad \omega\in\mathbb{R}^d.\\
\end{equation*}
By considering the spectral-domain behaviour of our multivariate GRF $S(x)$, we will demonstrate the difficulties inherent in multivariate modelling of geometric anisotropic spatial dependence, and we will consider the complex coherence at frequency $\omega\in\mathbb{R}^d$, $\gamma_{pq}(\omega)$:
\begin{equation}
\label{eq:compcoherence}
\gamma_{pq}(\omega) = \frac{f_{pq}(\omega)}{\left\{f_{pp}(\omega)f_{qq}(\omega)\right\}^{\frac12}}.
\end{equation}
\subsection{Geometric Anisotropic LGCPs}
\label{subsec:GALGCPs}
We describe here the approach to modelling geometric anisotropy in univariate LGCPs, as introduced by \citet{Moller14}. In brief, the required dependence structure is specified through the application of an isotropic covariance structure to a geometrically manipulated version of the space on which the process lives. Since a LGCP is fully defined by the first and second order characteristics of the underlying Gaussian random field, \citet{Moller14} showed that one can therefore construct a geometric anisotropic LGCP through using standard geometric manipulations to modify the space on which the latent univariate GRF is defined. In Section \ref{sec:mvga}, we will show how this can be flexibly extended to the multivariate case by specifying individual components of a population of $P$ GRFs through $P$ potentially distinct geometric manipulations of $\mathbb{R}^2$.
Given an isotropic covariance function $C_0(\|h\|)$, $h\in\mathbb{R}^d$, we can define a geometric anisotropic version as \citep{Christakos92}
\begin{equation}
\label{eq:geoanisotcov}
C(h) = C_0\left(\sqrt{h^T\Sigma^{-1} h}\right),
\hspace{25pt}
h\in\mathbb{R}^d,
\end{equation}
where
\begin{equation}
\label{eq:geoanisotSigmadef}
\Sigma = R_{\theta}
\left(
\begin{array}{cc}
1 & 0 \\
0 & \zeta^2
\end{array}
\right)
R^T_{\theta},
\end{equation}
for $\theta\in[0,\pi)$ and $\zeta\in(0,1]$, and where $R_{\theta}$ is the rotation matrix. Under this parameterisation, $\Sigma$ is defined such that the ellipse $E=\{h\in\mathbb{R}^2:h^T\Sigma^{-1}h=1\}$ has a semi-major axis of unit length at angle $\theta$, relative to the abscissa axis of the original coordinate system, and a semi-minor axis of length $\zeta$ at angle $\theta+\pi/2$. Accordingly, we can describe the covariance function defined in \eqref{eq:geoanisotcov} as `elliptic', and we have that the LGCP driven by a GRF with elliptic covariance structure will also display elliptic, or geometric anisotropic, second-order behaviour, described by the pair correlation function and spectral density function as:
\begin{eqnarray*}
g(h) & = & g_0\left(\sqrt{h^T\Sigma^{-1} h}\right) = \exp\left\{C_0\left(\sqrt{h^T\Sigma^{-1} h}\right)\right\} \\
f(\omega) & = & \left|\Sigma\right|^{1/2} f_0\left(\sqrt{\omega^T\Sigma\omega}\right),
\end{eqnarray*}
for $h,\omega\in\mathbb{R}^d$, where $f_0(\|\omega\|)$ is the isotropic spectral density that forms a Fourier transform pair with $C_0(\|h\|)$, and $g_0(\|h\|)$ is the corresponding isotropic pair correlation function.
Our specification of geometric anisotropy differs slightly from that of \citet{Moller14}, who include an additional scale parameter in their definition of the deformation matrix $\Sigma$; this is used to scale the axes in the resulting elliptical covariance structure. In practice, however, the majority of parametric covariance functions of interest incorporate a scale parameter that directly controls the correlation length, and so including a separate scale parameter in \eqref{eq:geoanisotSigmadef} creates nonidentifiability issues when performing parameter inference. We avoid this issue by assuming all scale information to be controlled by the parametric form of $C_0(\|h\|)$.
Since we are considering processes that display anisotropy, it will be useful for their analysis to be able to express their second-order properties in polar coordinates. We therefore define the anisotropic pair correlation function, replacing the vector $h\in\mathbb{R}^d$ with its length $r$ and angle $\phi$:
\begin{equation}
\label{eq:anisotpcf}
g^a(r,\phi) = g([r\cos\phi,r\sin\phi]) = g_0\left(\frac{r}{\zeta}\sqrt{1-(1-\zeta^2)\cos^2(\phi-\theta)}\right).
\end{equation}
\section{Defining the Model}
\label{sec:model}
\subsection{Accommodating multivariate geometric anisotropy}
\label{sec:mvga}
For a population of $P$ LGCPs, we specify the multivariate dependence through the covariance structure of the $P$-dimensional GRF that drives the $P$ conditionally independent intensity processes. We extend the definition of geometric anisotropy in \eqref{eq:geoanisotcov} and we define the following family of geometric anisotropic auto- and cross-covariance functions:
\begin{equation*}
C_{pq}(h) = C_{0,pq}\left(\sqrt{h^T\Sigma_{pq}^{-1} h}\right),
\hspace{25pt}
p,q = 1,\ldots,P,
\;\;
h\in\mathbb{R}^d,
\end{equation*}
for some corresponding family of isotropic covariance functions $\{C_{0,pq}(\|h\|);p,q=1,\ldots,P\}$, and for a collection of deformation matrices $\{\Sigma_{pq}; p,q=1,\ldots,P\}$, where $\Sigma_{pq}$ is defined in terms of the parameter pair $(\theta_{pq},\zeta_{pq})$ according to \eqref{eq:geoanisotSigmadef}.
This framework will allow for the possibility of distinct geometric anisotropies in each of the marginal processes. Such processes can be used to model, for example, bivariate point patterns in which each component displays elliptical clustering at different orientations, or with differing degrees of ellipticity. Care must be taken in specifying the parameters for the cross-covariance functions $C_{pq}$, however, in order to ensure a valid multivariate model. In the spatial domain, we require the matrix of covariance functions $(C_{pq}(h))_{p,q=1}^P$ to be nonnegative definite for all $h\in\mathbb{R}^d$; the equivalent requirement in the spectral domain is that the matrix of spectral densities $(f_{pq}(\omega))_{p,q=1}^P$ is nonnegative definite for all $\omega\in\mathbb{R}^d$. If we consider the bivariate dependence structure for two processes $X_p$ and $X_q$, $p,q=1,\ldots,P$, then we can see that the restriction in the spectral domain is equivalent to requiring the magnitude squared coherence $|\gamma(\omega)|^2$ to be bounded above by 1, where the complex coherence is defined as in \eqref{eq:compcoherence}. This restriction can alternatively, and unsurprisingly, be written at every frequency as
\begin{equation}
\label{eq:crossspectrumUB}
0\le |f_{pq}(\omega)| \le \left\{f_{pp}(\omega)f_{qq}(\omega)\right\}^{\frac12},\qquad\omega\in\mathbb{R}^d,
\end{equation}
and this gives an upper bound on the magnitude of the cross-spectrum. This upper bound is displayed in Figure \ref{fig2:MVGAspectra} for a bivariate process with distinct marginal geometric anisotropies. By considering the behaviour of \eqref{eq:crossspectrumUB} over the two-dimensional Fourier domain, we can now make some general comments about the level of dependence between components in a bivariate geometric anisotropic LGCP; this discussion also applies to pairwise dependences in multivariate LGCPs of higher dimension. For the remainder of this subsection, the only assumption that we make is that each autospectrum and cross-spectrum in the bivariate process is decreasing for increasing frequencies $\omega$. In particular, the following discussion is valid for any family of spectral densities that satisfies this assumption.
The inequality in \eqref{eq:crossspectrumUB} implies that, for any two processes, between-process dependence can only be non-negligible at those frequencies that contribute significantly to the marginal dependence in {\em both} processes.
For two processes with distinct marginal geometric anisotropies, this restriction impacts the high-frequency behaviour of the bivariate process more than the low-frequency behaviour This can be seen by considering the spectra displayed in Figure \ref{fig2:MVGAspectra}: when constructing the upper bound for the cross-spectrum according to \eqref{eq:crossspectrumUB}, the high-frequency contributions of each of the autospectra are killed by the negligible power at the same frequency in the other autospectrum; the contrasting behaviour of the marginal processes at high frequencies kills any high-frequency dependence \textit{between} the processes.
As a result, for any two processes that display contrasting anisotropic behaviour, significant between-process dependence will be more evident at low frequencies, or large spatial scales.
Due to our modelling assumption of geometric anisotropy in the cross-dependence structure, the cross-spectrum will have elliptical contours of equal power density. From Figure \ref{fig2:MVGAspectra} we can also see that the elliptical geometries of the autospectra can dictate a nontrivial geometric structure for the upper bound of the cross spectrum.
For any given pair of marginal spectra, and thus a given upper bound to the corresponding cross-spectrum, the ellipticity of the true cross-spectrum will therefore impact its permissible coverage of the frequency space, as its elliptical structure must fit within the upper bound's nontrivial geometry. Indeed, we can see from Figure \ref{fig2:MVGAspectra} that, in order for our elliptical cross-spectrum to extend further into the higher-frequency regions of the Fourier space, the ellipticity of the cross-spectrum should be more pronounced; if we were to assume a more isotropic cross-dependence structure, then the non-negligible cross-spectrum would be more restricted to the low-frequency region around the origin. Since the overall power in the cross-process dependence is obtained by integrating the cross-spectrum over the entire Fourier domain, this gives us a link between the power and the degree of anisotropy in the cross-process dependence. We will formalise this relationship towards the end of the next section, in the context of a Mat\'ern specification for our multivariate dependence structure.
\begin{figure}
\centering
\begin{subfigure}{0.275\textwidth}
\includegraphics[width=\textwidth]{./fig2_autospectrum1_v3.png}
\label{subfig:MVGAautospec1}
\end{subfigure}
\begin{subfigure}{0.275\textwidth}
\includegraphics[width=\textwidth]{./fig2_autospectrum2_v3.png}
\label{subfig:MVGAautospec2}
\end{subfigure}
\begin{subfigure}{0.3\textwidth}
\includegraphics[width=\textwidth]{./fig2_crossspectrumUB_v3.png}
\label{subfig:MVGAcrossspecUB}
\end{subfigure}
\vspace{-20pt}
\caption{Geometric anisotropic autospectra (left and centre) for a bivariate GRF, along with the upper bound on the corresponding cross-spectrum (right), as given in equation \eqref{eq:crossspectrumUB}.}
\label{fig2:MVGAspectra}
\end{figure}
\subsection{A multivariate Mat\'ern correlation structure}
\label{subsec:MVMaterncorr}
The Mat\'ern family of correlation functions \citep{Stein99,Guttorp06} provides a flexible route to modelling multi-scale dependence in stationary spatial processes. For univariate random fields, one can use a single three-parameter covariance function to replicate dependence structures that act over any positive scale, whilst additionally controlling the smoothness of any realisations. The flexibility of this model has made it the tool of choice for modelling univariate processes in the spatial statistics literature, and there has naturally been a great deal of interest in extending its use to the multivariate setting. \citet{Gneiting10} and \citet{Apanasovich12} have recently addressed this interest, proposing the use of a Mat\'ern function to describe all auto- and cross-covariances for a multivariate isotropic stationary random field. This work has been further extended by \citet{Kleiber12}, who accommodate nonstationarity by allowing the Mat\'ern parameters to vary with respect to location; this allows for the possibility of local anisotropic behaviour, as well as local variances and smoothnesses. We will incorporate elements from these approaches in our modelling framework, though we retain an assumption of stationarity. Our aim is to ascertain whether observed second-order point process characteristics can be modelled independently of first-order covariates; this goal would be best served under assumptions of stationarity in the underlying GRF.
We develop the stationary Mat\'ern covariance structure for the multivariate GRF $S(x)$, ensuring that all auto- and cross-covariances have a valid Mat\'ern form. The multivariate Mat\'ern model was introduced by \citet{Gneiting10}, who established necessary and sufficient conditions for the validity of the bivariate model, and sufficient conditions for the validity of a restricted subclass of the multivariate ($P\ge3$) model. This work was extended by \citet{Apanasovich12}, who relax the restrictions on the multivariate model and provide sufficient conditions for its validity for any dimension $P\ge1$. We present a class of Mat\'ern auto- and cross-covariance functions that can accommodate multivariate geometric anisotropy, and we adapt the work of \citet{Apanasovich12} in order to provide sufficient conditions for its validity, for $P\ge1$.
Following \citet{Gneiting10}, we define the isotropic multivariate Mat\'ern covariance function to be
\begin{eqnarray}
\label{eq:mvMatern}
\nonumber
C_{0,pq}(\|h\|;\alpha_{pq},\nu_{pq},\sigma_{pq}) & = & \frac{\sigma_{pq}}{2^{\nu_{pq}-1}\Gamma\left(\nu_{pq}\right)} \left(\frac{2\displaystyle\sqrt{\nu_{pq}}}{\alpha_{pq}} \|h\|\right)^{\nu_{pq}} \mathcal{K}_{\nu_{pq}}\left(\frac{2\displaystyle\sqrt{\nu_{pq}}}{\alpha_{pq}} \|h\|\right),
\hspace{10pt}
h\in\mathbb{R}^d,
\\
\end{eqnarray}
where $\mathcal{K}_{\nu}(\cdot)$ is the modified Bessel function of the second kind \citep[][pp.374--379]{Abramowitz65}. Here, $\sigma_{pq}\in\mathbb{R}$ ($\sigma_{pp}>0$) is the zero-lag covariance between field components $S_p$ and $S_q$, and $\alpha_{pq}>0$ and $\nu_{pq}>0$ are scale and smoothness parameters, respectively. The latter two parameters control the rate of decay of covariance between the same two processes with respect to distance. As a scale parameter, $\alpha_{pq}$ determines the `practical range' of the covariance function, i.e.~the separation distance at which $S_p$ and $S_q$ may be considered approximately independent. The smoothness parameter $\nu_{pq}$ determines the shape of the covariance function, and in particular the speed with which it decays close to the origin. For the marginal processes, $\nu_{pp}$ clearly then controls the smoothness of the realisations; indeed, the marginal process $X_p$ will be $m$ times mean-square differentiable if and only if $\nu_{pp}>m$.
Throughout the literature, the Mat\'ern covariance function has been defined using a variety of parametric forms, with the three parameters interacting in a different manner in each specification. The predominant material difference between the parameterisations is the formulation of the term used to scale the absolute distance $\|h\|$. The inverse of this distance-scaling factor is also known as the correlation length and this is proportional to the practical range; as can be seen from \eqref{eq:mvMatern}, in the current parameterisation, the correlation length is equal to $\alpha_{pq}/2\sqrt{\nu_{pq}}$. For alternative parameterisations of the model where the correlation length is independent of $\nu_{pq}$, it is often found that the effects of $\alpha_{pq}$ and $\nu_{pq}$ on the practical range and shape of $C_{0,pq}(\|h\|;\alpha_{pq},\nu_{pq},\sigma_{pq})$ cannot be well separated. The parameterisation of the Mat\'ern function given in \eqref{eq:mvMatern}, attributable to \cite{Handcock94} in the univariate scenario, is chosen to allow maximal separation of the roles of $\alpha_{pq}$ and $\nu_{pq}$ in determining the second-order behaviour of $S(x)$ and, ultimately, the resulting point process $X$.
As $\alpha_{pq}$ increases for fixed $\nu_{pq}$, so too will the practical range of $C_{0,pq}(\|h\|;\alpha_{pq},\nu_{pq},\sigma_{pq})$. This will increase the maximum distance at which one can expect to find cross-process aggregation and segregation of points in $X_p$ and $X_q$. Note that the corresponding effect for the marginal scale parameters is that an increase (decrease) in $\alpha_{pp}$ will result in an increase (resp.~decrease) in the width of the observed clusters in $X_p$. Recall that the smoothness parameter controls the shape of the covariance function; as $\nu_{pq}$ increases, $C_{0,pq}(\|h\|;\alpha_{pq},\nu_{pq},\sigma_{pq})$ becomes smoother around $\|h\|=0$. As a result of our parameterisation, as $\nu_{pq}$ increases for fixed $\alpha_{pq}$, \eqref{eq:mvMatern} will increase for small values of $\|h\|$ and decrease for large values of $\|h\|$; the distribution of variance shifts from high scales to low scales. Thus, where the scale parameter $\alpha_{pq}$ determines the width of areas in which processes $X_p$ and $X_q$ will have similar intensities, $\nu_{pq}$ will determine {\it how} similar these intensity processes are within these regions.
Having established the Mat\'ern form of the auto- and cross-covariances for a multivariate isotropic GRF, we now generalise to allow for anisotropic multivariate covariance structures. Recall from Section \ref{sec:mvga} that we obtain our geometric anisotropic (cross-)covariance function by applying the deformation matrix $\Sigma_{pq}$:
\begin{eqnarray}
C_{pq}(h;\alpha_{pq},\nu_{pq},\sigma_{pq},\Sigma_{pq}) & = & \frac{\sigma_{pq}}{2^{\nu_{pq}-1}\Gamma\left(\nu_{pq}\right)} \left(\frac{2\sqrt\nu_{pq}}{\alpha_{pq}}\left\|\Sigma_{pq}^{-1/2}h\right\|\right)^{\nu_{pq}} \mathcal{K}_{\nu_{pq}}\left(\frac{2\sqrt\nu_{pq}}{\alpha_{pq}}\left\|\Sigma_{pq}^{-1/2}h\right\|\right),\nonumber\\
\label{eq:mvgaMatern}
\end{eqnarray}
which is defined for any $h\in\mathbb{R}^d$.
Recall that we require the matrix $\left(C_{pq}(h;\alpha_{pq},\nu_{pq},\sigma_{pq},\Sigma_{pq})\right)_{p,q=1}^P$ to be nonnegative definite for all $h\in\mathbb{R}^d$, in order for \eqref{eq:mvgaMatern} to define a valid multivariate covariance model. Satisfaction of this requirement can be guaranteed by placing the following conditions on the cross-covariance parameters $\{\alpha_{pq},\nu_{pq},\sigma_{pq},\theta_{pq},\zeta_{pq}, p\ne q\}$.
\begin{condition}
\label{cond:nu}
There exists a nonnegative constant $\Delta_\nu$ such that $\nu_{pq} - (\nu_{pp}+\nu_{qq})/2 = \Delta_{\nu}(1-A_{\nu,pq})$, $p,q=1,\ldots,P$, where $A_\nu$ is a valid $P\times P$ correlation matrix, with entries $0\le A_{\nu,pq}\le1$.
\end{condition}
\begin{condition}
\label{cond:alpha}
The matrix with elements $-4\nu_{pq}/\alpha_{pq}^2$, $p,q=1,\ldots,P$, is {\em conditionally} nonnegative definite. This is a weaker assumption than that of nonnegative definiteness, and it may be satisfied by a matrix containing only negative elements.
\end{condition}
\begin{condition}
\label{cond:sigma}
The matrix with elements
$$
\frac{|\Sigma_{pq}|^{1/2}\sigma_{pq}\Gamma(\nu_{pq}+d/2)}{\pi^{d/2}\Gamma(\frac{\nu_{pp}+\nu_{qq}}{2}+\frac{d}{2})\Gamma(\nu_{pq})}\left(\frac{4\nu_{pq}}{\alpha_{pq}^2}\right)^{\Delta_{\nu}+\frac{\nu_{pp}+\nu_{qq}}{2}},
\qquad
p,q=1,\ldots,P,
$$
is nonnegative definite.
\end{condition}
\begin{condition}
\label{cond:Sigma}
The matrix with elements $-\|\Sigma_{pq}^{1/2}\omega\|^2$, $p,q=1,\ldots,P$, is conditionally nonnegative definite for any $\omega\in\mathbb{R}^d$.
\end{condition}
\begin{proposition}
\label{prop:mvgaMatern}
For $p,q=1,\ldots,P$, let $\alpha_{pq}>0$, $\nu_{pq}>0$, $\sigma_{pq}\in\mathbb{R}$, $\theta\in[0,2\pi)$ and $\zeta\in(0,1]$. Then the multivariate geometric anisotropic Mat\'ern function \eqref{eq:mvgaMatern} specifies a valid multivariate covariance model if Conditions \ref{cond:nu}-\ref{cond:Sigma} are met.
\end{proposition}
The proof of Proposition \ref{prop:mvgaMatern} is given in the Appendix,
and follows a similar argument to the proof of Theorem 1 of \citet{Apanasovich12}.
\begin{remark}
\label{rmk:zeta}
If Condition \ref{cond:Sigma} holds, then the $P\times P$ matrix with $(p,q)$-element $\left|\Sigma_{pq}\right|^{-1/2} = \zeta_{pq}^{-1},$ will be nonnegative definite; in particular, we can deduce $\zeta_{pq}^2\ge\zeta_{pp}\zeta_{qq}$, for all $p,q=1,\ldots,P.$
\end{remark}
Conditions \ref{cond:nu}-\ref{cond:Sigma} are similar in spirit to those placed by \citet{Apanasovich12} on the Mat\'ern parameters in order to guarantee a valid multivariate dependence structure in an isotropic setting. In the simpler isotropic framework, the three conditions specified by \citet{Apanasovich12} are sufficient to guarantee nonnegative definiteness of the resulting spectral density, and also to guarantee that all absolute zero-lag cross-correlations are bounded above by one. In the more general geometric anisotropic setting, we require a more extensive specification. Conditions \ref{cond:nu}-\ref{cond:Sigma}, above, are sufficient to guarantee nonnegative definiteness of the geometric anisotropic spectral density, and are also sufficient for the absolute colocated cross-correlations to be bounded above by 1.
These conditions constitute a set of implicit relationships that, between them, specify a valid multivariate geometric anisotropic LGCP. We will now provide explicit restrictions on the cross-dependence parameters in terms of the marginal dependence parameters. This will allow users to sequentially construct a valid multivariate model by first specifying the marginal covariances, and then conditionally specifying the cross-covariance structures. This sequential approach to model construction will also be reflected in our model-fitting procedures in Section \ref{sec:modelfitting}.
Trivial rearrangement of Condition \ref{cond:nu} yields an explicit expression for $\nu_{pq}$ in terms of the corresponding marginal values. In Remarks \ref{rmk:alpha}-\ref{rmk:equicorr}, below, we provide similar constructions for the cross-covariance parameters $\alpha_{pq},$ $\sigma_{pq}$, $\theta_{pq}$ and $\zeta_{pq}$, such that Conditions \ref{cond:nu}-\ref{cond:Sigma} may be satisfied. The proofs for Remarks \ref{rmk:alpha}-\ref{rmk:equicorr} are given in the Appendix.
\begin{remark}
\label{rmk:alpha}
Condition \ref{cond:alpha} is satisfied by the parameters $\{\nu_{pq},\alpha_{pq}\;;\;p,q=1,\ldots,P\}$ if
\begin{equation*}
\frac{4\nu_{pq}}{\alpha_{pq}^2} = \frac12\left(\frac{4\nu_{pp}}{\alpha_{pp}^2}+\frac{4\nu_{qq}}{\alpha_{qq}^2}\right) + \Delta_{\alpha}\left(1-A_{\alpha,pq}\right),
\end{equation*}
for some constant $\Delta_{\alpha}\ge0$ and for some $0\le A_{\alpha,pq}\le1$ that form a valid correlation matrix. This remark is also made by \citet{Apanasovich12} in their chosen Mat\'ern parameterisation.
\end{remark}
\begin{remark}
\label{rmk:sigma}
Condition \ref{cond:sigma} is satisfied by the parameters $\{\nu_{pq},\alpha_{pq},\zeta_{pq},\sigma_{pq}\;;\;p,q=1,\ldots,P\}$ if
\begin{equation*}
\sigma_{pq} = \frac{\pi^{d/2} V_{p}V_{q}A_{\sigma,pq} }{\zeta_{pq}}
\left(\frac{4\nu_{pq}}{\alpha_{pq}^2}\right)^{-\Delta_{\nu}-\frac{\nu_{pp}+\nu_{qq}}{2}}
\frac{\Gamma(\frac{\nu_{pp}+\nu_{qq}}{2}+\frac{d}{2})\Gamma(\nu_{pq})}{\Gamma(\nu_{pq}+\frac{d}{2})}
\qquad p,q=1,\ldots,P,
\end{equation*}
for constants $V_p,V_q\ge0$ and for some $A_{\sigma,pq}\in[-1,1]$ that form a valid correlation matrix.
\end{remark}
\begin{remark}
\label{rmk:Sigma}
Condition \ref{cond:Sigma} is satisfied by the deformation matrices $\{\Sigma_{pq}\;;\;p,q=1,\ldots,P\}$ if their diagonal elements $\left[\Sigma_{pq}\right]_{ii}$, can be written
$$
\left[\Sigma_{pq}\right]_{ii} = \frac12\left[\Sigma_{pp}+\Sigma_{qq}\right]_{ii} + \Delta^{(i)}_{\Sigma}\left(1-A^{(i)}_{\Sigma,pq}\right),\qquad i=1,2.
$$
\end{remark}
\begin{remark}
\label{rmk:equicorr}
For small $P$, we can follow the lead of \citet{Apanasovich12} and use equicorrelated matrices $A^{(i)}_{\Sigma}$, $i=1,2$, setting $A^{(i)}_{\Sigma,pq}=\rho^{(i)}_{\Sigma}$, $p\ne q$; in this scenario, for the sake of identifiability, we redefine $\Delta_{\Sigma}^{(i)}:=\Delta_{\Sigma}^{(i)}(1-\rho_{\Sigma}^{(i)})$, $i=1,2$.
\end{remark}
Conditions \ref{cond:nu}-\ref{cond:Sigma}, along with Remarks \ref{rmk:alpha}-\ref{rmk:Sigma}, indicate a sequential approach to specifying a valid multivariate geometric anisotropic Mat\'ern covariance structure in practice. As mentioned previously, Condition \ref{cond:nu} and Remarks \ref{rmk:alpha}-\ref{rmk:Sigma} suggest that one must specify the parameters for the marginal covariance function before conditionally specifying the parameters for each cross-covariance function. These statements also indicate that, within each individual component of the joint model, i.e.~for fixed $p,q$, there is a particular order in which the five parameters $(\theta_{pq},\zeta_{pq},\alpha_{pq},\nu_{pq},\sigma_{pq})$ should necessarily be specified. From Remark \ref{rmk:sigma}, we can see that, for each $(p,q)$ pairing, the specification of the Mat\'ern power parameter $\sigma_{pq}$ is dependent upon the corresponding ratio of anisotropy $\zeta_{pq}$, as well as the other Mat\'ern parameters, $\alpha_{pq}$ and $\nu_{pq}$, and Remark \ref{rmk:Sigma} indicates that the anisotropy parameters $(\theta_{pq},\zeta_{pq})$ should be jointly specified. In addition, Condition \ref{cond:nu} and Remark \ref{rmk:alpha} indicate that the smoothness parameter $\nu_{pq}$ should be specified before the scale parameter $\alpha_{pq}$. We conclude that, for each marginal or bivariate component of the joint covariance model, the anisotropy parameters should be specified before the Mat\'ern parameters, with the Mat\'ern smoothness, scale and power parameters being specified third, fourth and fifth, respectively.
We conclude this section by considering the limitations placed on the zero-lag cross-correlation coefficients $\rho_{pq}:=\sigma_{pq}/\sqrt{\sigma_{pp}\sigma_{qq}}$. By rearranging Condition \ref{cond:sigma}, we can write:
\begin{equation}
\label{eq:colocxcorrs}
\rho_{pq}^{2} = \frac{\sigma_{pq}^2}{\sigma_{pp}\sigma_{qq}} \le \prod_{i=1}^4\tau_{pq}^{(i)}\le1,
\end{equation}
with
\begin{equation*}
\tau_{pq}^{(1)} = \frac{\mathcal{B}^2(\nu_{pq},\frac{d}{2})}{\mathcal{B}^2(\frac{\nu_{pp}+\nu_{qq}}{2},\frac{d}{2})},
\hspace{15pt}
\tau_{pq}^{(2)} = \left[\frac{\frac{4\nu_{pp}}{\alpha_{pp}^2}\frac{4\nu_{qq}}{\alpha_{qq}^2}}{\left(\frac{4\nu_{pq}}{\alpha_{pq}^2}\right)^2}\right]^{\Delta_{\nu}},
\end{equation*}
\begin{equation*}
\tau_{pq}^{(3)} = \frac{\Gamma^2(\frac{\nu_{pp}+\nu_{qq}}{2})\left(\frac{\alpha_{pq}^2}{4\nu_{pq}}\right)^{\nu_{pp}+\nu_{qq}}}
{\Gamma(\nu_{pp})\left(\frac{\alpha_{pp}^2}{4\nu_{pp}}\right)^{\nu_{pp}}\Gamma(\nu_{qq})\left(\frac{\alpha_{qq}^2}{4\nu_{qq}}\right)^{\nu_{qq}}},
\hspace{10pt}
\tau_{pq}^{(4)} = \frac{|\Sigma_{pp}|^{1/2}|\Sigma_{qq}|^{1/2}}{|\Sigma_{pq}|} = \frac{\zeta_{pp}\zeta_{qq}}{\zeta_{pq}^2},
\end{equation*}
where $\mathcal{B}(\cdot,\cdot)$ is the Beta function \citep{Abramowitz65}. The first inequality in \eqref{eq:colocxcorrs} is directly implied by Condition \ref{cond:sigma}. The second inequality in \eqref{eq:colocxcorrs} can be shown componentwise: By Remark \ref{rmk:zeta}, Condition \ref{cond:Sigma} ensures that $\tau_{pq}^{(4)}\le1$, and as noted by \citet{Apanasovich12} in the isotropic framework, Conditions \ref{cond:nu} and \ref{cond:alpha} are sufficient to guarantee that $\tau_{pq}^{(i)}\le 1$, $i=1,2,3$
In the isotropic framework, $\tau_{pq}^{(4)}=1$, and we are left with the limitations noted by \citet{Apanasovich12}: the zero-lag cross-correlation will be bounded above by 1 when the corresponding univariate isotropic processes share identical Mat\'ern parameters. When the marginal parameter specifications differ, this upper bound will decrease as the smoothness and inverse correlation length of the cross-covariance structure depart from the arithmetic mean of the corresponding marginal quantities.
In our more general anisotropic framework, we can see from $\tau_{pq}^{(4)}$ in \eqref{eq:colocxcorrs} that the upper bound on the colocated cross-correlations will also be affected by the relationship between the cross-covariance ratio of anisotropy $\zeta_{pq}$ and the ratios of anisotropy in the corresponding marginal covariance structures.
If we assume Condition \ref{cond:Sigma} to hold, then by Remark \ref{rmk:zeta}, $\zeta_{pq}$ will be restricted to the closed interval $[\zeta_{pp}^{1/2}\zeta_{qq}^{1/2},1]$. If $\zeta_{pq}=\zeta_{pp}^{1/2}\zeta_{qq}^{1/2}$, then $\tau_{pq}^{(4)}$ will reduce to 1, and the upper bound of the colocated cross-correlation $\rho_{pq}$ will behave as in the isotropic framework, i.e.~as described above. Increasing $\zeta_{pq}$ away from this geometric mean, however, will decrease $\tau_{pq}^{(4)}$, which will in turn shrink the upper bound on $\rho^2_{pq}$, given in \eqref{eq:colocxcorrs}.
In other words, as the ellipticity of the cross-covariance function becomes less pronounced, the maximum possible degree of zero-lag correlation between the two components of the field will decrease. This formalizes the relationship between the power and the anisotropy of the cross-process dependence, discussed at the end of Section \ref{sec:mvga}.
\section{Fitting the Model}
\label{sec:modelfitting}
\subsection{Parameter Estimation Procedure}
\label{subsec:paramest}
In order to fit our parametric model to an observed multitype point pattern, we must estimate both the marginal and joint anisotropy parameters $\{\theta_{pq},\zeta_{pq};p,q=1,\ldots,P\}$, as well as the parameters that specify the mean and Mat\'ern covariance structure of the underlying Gaussian random field, $\{\mu_p,\alpha_{pq},\nu_{pq},\sigma_{pq};p,q=1,\ldots,P\}$. At a high level, we follow the approach of \citet{Moller14}, who fit a univariate version of our model by first estimating the angle and ratio of anisotropy in the observed data, before using these estimates to back-transform the data into an isotropic framework. The resulting `isotropised' point pattern is then used to estimate the mean parameters and the Mat\'ern parameters. Our approach to each component of this two-stage model-fitting procedure will differ from the methods of \citet{Moller14}, however. We use an approach to estimating anisotropy that is less sensitive to user-specified tuning parameters, which we adapt from the work of \citet{Rajala16}, and we use a more automatable approach to estimating the mean and Mat\'ern parameters, which we develop from the work of \citet{Tanaka08}.
In developing our parameter estimation methodology, we are faced with the question of whether to put measures into place to guarantee that the fitted model satisfies Conditions \ref{cond:nu}-\ref{cond:Sigma}, therefore ensuring validity of the multivariate dependence structure. This is the approach taken by \citet{Apanasovich12} for fitting multivariate isotropic Mat\'ern GRFs; they fit the marginal dependence structures then use the estimated marginal parameters to restrict the parameter subspace for the Mat\'ern cross-covariances. Since Conditions \ref{cond:nu}-\ref{cond:Sigma} are sufficient, and not necessary, the resulting restriction on the joint dependence structure could be overstated, potentially resulting in inconsistent estimators for the Mat\'ern cross-covariance parameters. Under the assumption that the smoothness is known, however, the power and scale parameters for a univariate Mat\'ern covariance function cannot be consistently estimated under infill asymptotics \citep{Zhang04}; consistency can only be achieved by increasing the observation window $W$. As noted by \citet{Apanasovich12}, constraining $\sigma_{pq}^2$ and $\alpha_{pq}^2$ ($p\ne q$) conditional on their corresponding marginal values therefore provides no additional penalty in terms of estimator consistency when assuming a fixed observation window. Furthermore, the numerical tests of \citet{Apanasovich12} show that reasonable accuracy can indeed be obtained when using this constrained approach to parameter estimation; this approach therefore warrants examination in the current framework. In order to avoid compromising the consistency of the anisotropy estimators, we do not use our conditions from Section \ref{sec:model} to restrict the parameter pair $(\theta_{pq},\zeta_{pq})$, $p\ne q$.
\subsection{Estimating the Anisotropy Parameters}
We focus first on quantifying the anisotropy present in both the marginal and joint dependence structures in a multi-type point pattern. \citet{Moller14} estimate the angle of anisotropy in a univariate geometric anisotropic point pattern by finding the angle $\phi$ at which the $r$-integrated difference between the anisotropic pair correlation function $g^a(r,\phi)$ and its phase-shifted self $g^a(r,\phi+\pi/2)$, is maximised. This is achieved by estimating $g^a(r,\phi)$ over a discrete lattice of polar coordinates $(r,\phi)$, and numerically approximating the required integral in $r$. Accuracy of the resulting estimator is therefore sensitive to the resolution of the polar lattice, as well as the choice of two bandwidth parameters used in estimating the anisotropic pair correlation function; for details of these bandwidth parameters, see \citet{Moller14}. Finally, use of this estimation method is also dependent on the assumption that the isotropic pair correlation function is strictly decreasing. Whilst this assumption holds true for our assumed Mat\'ern model, it can be violated by real data. The approach we detail below is more widely applicable, as it does not depend on such an assumption, and it is also less sensitive to subjective choices of bandwidth parameters.
We adopt and adapt the method introduced by \citet{Rajala16} for estimating the angle of anisotropy: we adopt this method for characterising anisotropy in the marginal covariance structures, and we adapt it for estimating the angle of anisotropy in the cross-covariance structures. For the sake of generality, we describe the procedure for estimating $\theta_{pq}$, $p\ne q$. We start by constructing the point pattern formed by the difference vectors $\
x_{p,i}-x_{q,j}\;;\; i=1,\ldots,n_p, j=1,\ldots,n_q\}$; this is the (bivariate) Fry process \citep{Fry79}, and when $p=q$, this will be rotationally symmetric of order 2, about the origin.
The Fry process is useful here as its first-order properties will reflect the second-order properties of the original point pattern. We can therefore estimate any second-order anisotropy in the original bivariate point pattern by estimating the anisotropy in the intensity of the bivariate Fry process.
Dividing the polar plane into a selected number, $n_F$, of distinct sectors, and for $l\in L\subset\mathbb{N}$, we collect the $l$th nearest Fry point in each sector into a set, $G_l$, of $n_F$ points, such that each $G_l$ sketches out a noisy contour around the origin, and such that the intensity of the Fry process is reflected in the proximity of the $G_l$s to one another. For point patterns that display segregation, the anisotropy in the joint second-order dependence structure will be shared by the contours of the intensity field for the Fry process; for aggregated point patterns, the angle of anisotropy will be phase-shifted by $\pi/2$ in the Fry process. In order to quantify the anisotropy in the original point pattern then, we can treat the $G_l$s as sampled versions of the Fry intensity's contours, and assuming Gaussian measurement error we can infer the corresponding true contours using adjusted ordinary least squares, and subsequently derive the angle of anisotropy in the original point pattern.
For full technical details of this method, we direct the reader to \citet{Rajala16}.
For each marginal process, as described by \citet{Moller14}, we can transform the observed point pattern $X_p\cap W$ and the corresponding observation window $W$ by assuming fixed values for $\theta\in[0,2\pi)$ and $\zeta\in[0,1]$:
\begin{eqnarray*}
X_{p,\theta,\zeta} = X_p R_{\theta}^T
\left(
\begin{array}{cc}
1 & 0 \\
0 & \zeta^{-1}
\end{array}
\right)
,\qquad&&\qquad
W_{\theta,\zeta} = W R_{\theta}^T
\left(
\begin{array}{cc}
1 & 0 \\
0 & \zeta^{-1}
\end{array}
\right)
.
\end{eqnarray*}
If the chosen values of $\theta$ and $\zeta$ are equal to the values that describe the anisotropy of $X_p$, then the transformed point process $X_{p,\theta,\zeta}$ will be isotropic and the corresponding anisotropic pair correlation function $g^a_{pp,\theta,\zeta}(r,\phi)$ will be constant with respect to its second argument. This motivates our chosen method for estimating the \textit{marginal} anisotropy ratios $\zeta_{pp}$, which we also adopt from the work of \citet{Rajala16}.
Following \citet{Rajala16}, we define the following directional discrepancy statistic:
\begin{equation}
\label{eq:intdiffK}
V_{pp,\theta}(\zeta) = \int_{b_1}^{b_2} \left[K^a_{pp,\theta,\zeta}(r,0) - K^a_{pp,\theta,\zeta}(r,\pi/2)\right] dr,
\end{equation}
where
\begin{equation}
\label{eq:sectorKfun}
K^a_{pp,\theta,\zeta}(r,\phi) = \int_0^r g^a_{pp,\theta,\zeta}(s,\phi)ds
\end{equation}
is the sector-$K$-function, an anisotropic variant of Ripley's $K$-function, evaluated on the isotropised point pattern $X_{p,\theta,\zeta}$. To estimate the marginal ratio of anisotropy $\zeta_{pp}$, we back-transform our observed point pattern using the estimated angle of anisotropy $\hat{\theta}_{pp}$ and a sequence of candidate ratios $\{\zeta_{pp,k}:=k\zeta_{max}/(1+n_{\zeta}),\; k=1,\ldots,n_{\zeta}\}$, for some user-defined upper bound $\zeta_{max}$. We then choose $\hat{\zeta}_{pp} = \zeta_{pp,k}\in(0,\zeta_{max})$ to be the candidate value that minimises the estimate $\hat{V}_{pp,\hat{\theta}_{pp}}(\zeta_{pp,k})$. Note that, although we defined $\zeta_{pp}\in(0,1)$ in Section \ref{subsec:GALGCPs}, the sampling variance of the estimated sector-$K$-function can result in an estimated ratio $\hat{\zeta}_{pp}>1$.
\citet{Moller14} use a similar approach, in effect minimising the directional discrepancy statistic \eqref{eq:intdiffK}, but using the anisotropic pair correlation function in place of the sector-$K$-function. Indeed, it is possible to use any directional second-order statistic in the integrand of \eqref{eq:intdiffK}. We choose to use $K^a_{pp,\theta,\zeta}$ for two reasons. Firstly, the analysis of \citet{Redenback09} suggests that the sector-$K$-function is better-suited to characterising anisotropy than nearest-neighbour statistics; the authors conclude that, for detecting anisotropy in point patterns, statistical tests based on the sector-$K$-function have greater power, in general, than those based on nearest-neighbour orientation statistics. Secondly, estimation of the sector-$K$-function requires the choice of only one tuning parameter, an angular bandwidth, whereas the use of the anisotropic pair correlation function would require the specification of both angular and radial bandwidths
Our chosen approach to estimating $\zeta_{pp}$ can be extended to the multivariate scenario, where we are interested in the geometric anisotropic cross-dependence exhibited by a given pair of Cox processes $X_p$ and $X_q$. By manipulating the space $\mathbb{R}^d$ on which both processes live, we also manipulate the cross-covariance function $C_{pq}(h)$ that specifies the dependence between the random fields that drive $X_p$ and $X_q$. For each pair of processes, we once again define a discrete set of candidate multivariate anisotropy ratios $\{\zeta_{pq,k}\in(0,\zeta_{max}), k=1,\ldots,n_{\zeta}\}$, and we choose $\hat{\zeta}_{pq} = \zeta_{pq,k}$ for which the estimated value of $V_{pq,\hat{\theta}_{pq}(\zeta_{pq,k})}$ is minimised, where $V_{pq,\hat{\theta}_{pq}(\zeta_{pq,k})}$ is defined through transforming both $X_p\cap W$ and $X_q\cap W$, along with their common observation window $W$.
The above approach to estimating the anisotropy parameters requires the selection of a number of control parameters: the number of sectors $n_F$, into which we partition the Fry process; the number $n_\zeta$ of candidate ratios of anisotropy, as well as their upper bound $\zeta_{max}$; and the limits of integration, $b_1$ and $b_2$ in \eqref{eq:intdiffK}, which we use to calculate $\hat{V}_{pq,\hat{\theta}_{pq}}(\zeta_{pq,k})$ when estimating $\zeta$. As a rule of thumb, and for reasons outlined below, \citet{Rajala16} suggest choosing $n_F\approx\lambda|W|/6$, where
$\lambda|W|$ is the expected number of points in the original point process.
We adopt this guideline for choosing $n_F$ when estimating the anisotropy in the marginal processes, and we derive a similar rule of thumb for $n_F$ when estimating the between-process anisotropy, by following the same arguments as \citet{Rajala16}. For the bivariate Poisson process with intensity vector $(\lambda_p,\lambda_q)$ in a circular spatial window $W$, the expected number of bivariate Fry points per sector is approximately $\lambda_p\lambda_q|W|^2/3n_F$.
Each point in the bivariate process can be expected to contribute if there are at least $(\lambda_p+\lambda_q)|W|$ points per sector, and so we have a bivariate direction count rule of $n_F\approx \lambda_p\lambda_q|W|/3(\lambda_p+\lambda_q)$.
Selection of both $n_{\zeta}$ and $\zeta_{max}$ is straightforward: $\zeta_{max}$ should be chosen such that $(0,\zeta_{max})$ covers the majority of the sampling distribution of $\zeta_{pq}$, and selection of $n_{\zeta}$ involves a trade-off between accuracy in the resulting estimates and computational expense of the estimation procedure. In Section \ref{sec:implementation}, where we implement our model fitting procedure for both simulated data and tropical rainforest data, we use $\zeta_{max}=2$ and $n_{\zeta}=199$ for estimating all marginal and joint ratios of anisotropy. Choice of the limits of integration, $b_1$ and $b_2$ in \eqref{eq:intdiffK}, is a more subjective task, and should be determined by the range of scales over which dependence (either within, or between processes) is sought to be characterised; these need not be the same for all marginal and cross-dependence relationships being estimated. In Section \ref{sec:implementation}, we detail our choices of these limits of integration
\subsection{Estimating the Mat\'ern Parameters}
\label{subsec:Maternestimation}
Once we have estimated our anisotropy parameters, we can isotropise the point pattern and its observation window, and use this transformed data to estimate the remaining parameters. In order to ensure that the Mat\'ern parameters satisfy Conditions \ref{cond:nu}-\ref{cond:sigma}, we define $\nu_{pq}$, $\alpha_{pq}$ and $\sigma_{pq}$ according to the specifications in Condition \ref{cond:nu}, Remark \ref{rmk:alpha} and Remark \ref{rmk:sigma}, respectively.
Techniques for modelling the correlation matrices $A_{\nu}$, $A_{\alpha}$ and $A_{\sigma}$ are discussed by \citet{Apanasovich10}, and the reader is directed there for further details.
When $P$ is small, however, we can simplify our task by assuming the off-diagonal elements of $A_{\alpha},A_{\nu},A_{\sigma}$ to be constant \citep{Apanasovich12}
In order to estimate both the mean and Mat\'ern parameters, we maximise the Palm log-likelihood, first proposed by \citet{Tanaka08}.
For estimating the marginal parameters, we use the version of the Palm log-likelihood given by \citet{Dvorak12}, where the inner region correction is proposed to deal with edge effects:
\begin{multline}
\label{eq:palmloglik}
\ell(\lambda_p,\alpha_{pp},\nu_{pp},\sigma_{pp}) \approx \sum_{\substack{x_{p,i}\in X_{p,\theta,\zeta}\cap W_{\theta,\zeta}\setminus R \\x_{p,j}\in X_{p,\theta,\zeta}\cap W_{\theta,\zeta}\\r_{ij}<R}}^{\ne}
\log\left\{\lambda_p g_{pp}(r_{ij};\alpha_{pp},\nu_{pp},\sigma_{pp})\right\}
\\
- \lambda_p|X_p\cap W\setminus R| K_{p}(R;\alpha_{pp},\nu_{pp},\sigma_{pp}),
\end{multline}
where
$r_{ij}=\|x_{p,i}-x_{p,j}\|$, $K_{p}(r;\alpha_{pp},\nu_{pp},\sigma_{pp})$ is Ripley's univariate $K$-function, which we approximate by numerically integrating the corresponding pair correlation function, and $|X_{p,\theta,\zeta}\cap W\setminus R|$ denotes the number of points in the isotropised pattern $X_{p,\theta,\zeta}$ that lie further than a distance $R$ from the boundary of $W_{\theta,\zeta}$. $R$ is a user-defined tuning parameter that can be objectively set based on the data; this is discussed further in Section \ref{sec:implementation}. As is common in the point pattern literature, we use $\neq$ in the summation notation to indicate summation over pairs of distinct points.
The Palm log-likelihood \eqref{eq:palmloglik} can be analytically maximised with respect to $\lambda_p$, yielding the maximum Palm-likelihood estimate (MPLE) $\hat{\lambda}_p$, and we obtain MPLEs for the remaining marginal Mat\'ern parameters by numerically maximising $\ell(\hat{\lambda}_p,\alpha_{pp},\nu_{pp},\sigma_{pp})$. The MPLE for $\mu_p$ can be subsequently calculated according to \eqref{eq:lamp_mup}.
We further develop the Palm log-likelihood approach, in order to estimate the parameters for the cross-covariance structure; our bivariate Palm log-likelihood follows a similar construction to the marginal version. First, we obtain the symmetric bivariate Fry process for components $X_p$ and $X_q$, using the inner region correction to deal with edge effects. We then treat this Fry process as an inhomogeneous Poisson process, with intensity equal to a bivariate version of the Palm intensity \citep{Daley08,Prokesova13}, which we define heuristically as follows: for $x$ at distance $r$ from the origin $o$, the occurrence rate of process $q$ at $x\in\left\{\mathbb{R}^2:\|x\|=r\right\}$, assuming there to be a point of process $p$ at the origin, is
$$
\lambda_{0,pq}(x)dx = \mathbb{P}\left(|X_q \cap dx|=1\big||X_p\cap \{o\}|=1\right),
$$
where $dx$ is the Lebesgue measure for the infinitesimal set at $x$. Following this definition, we can relate the bivariate Palm intensity to the (isotropic) cross-pair correlation function for the original process:
$$
\lambda_{0,pq}(r) = \lambda_q g_{0,pq}(r),
$$
and this allows us to obtain the following bivariate Palm log-likelihood, which can be maximised to obtain estimates for $\alpha_{pq}$, $\nu_{pq}$, $\sigma_{pq}$, $p\ne q$:
\begin{multline}
\label{eq:bivpalmloglik}
\ell(\lambda_p,\lambda_q,\alpha_{pq},\nu_{pq},\sigma_{pq}) \approx \sum_{\substack{x_{p,i}\in X_p\cap W_{\theta,\zeta}\\x_{q,j}\in X_q\cap W_{\theta,\zeta}\\r_{ij}<R}}^{\ne}
\log\left\{(\lambda_p+\lambda_q) g_{pq}(r_{ij};\alpha_{pq},\nu_{pq},\sigma_{pq})\right\}
\\
- \Big(|X_q\cap W\setminus R|\lambda_p+|X_p\cap W\setminus R|\lambda_q\Big) K_{pq}(R;\alpha_{pq},\nu_{pq},\sigma_{pq}),
\end{multline}
where $r_{ij}=\|x_{p,i}-x_{q,j}\|$, $K_{pq}(r;\alpha_{pq},\nu_{pq},\sigma_{pq})$ is Ripley's bivariate $K$-function, and $|X_p\cap W\setminus R|$ denotes the number of observed points in process $p$ that lie further than a distance $R$ from the boundary of the window $R$. By substituting our previous estimates of $\lambda_p$ and $\lambda_q$ into \eqref{eq:bivpalmloglik}, we obtain an expression in terms of the Mat\'ern cross-covariance parameters only. We numerically maximise this expression in $(\alpha_{pq},\nu_{pq},\sigma_{pq})$ over the constrained parameter space described by Condition \ref{cond:nu}, Remark \ref{rmk:alpha} and Remark \ref{rmk:sigma}, and dependent on the corresponding estimated marginal Mat\'ern parameters. As described in Section \ref{subsec:paramest}, the use of constrained optimisation should not affect the consistency of the cross-covariance parameter estimators, however they may display some bias due to the truncation of their supports.
\section{Implementation}
\label{sec:implementation}
\subsection{Proof of concept simulations}
\label{subsec:proofofconcept}
We demonstrate the validity of the model fitting procedure described in Section \ref{sec:modelfitting}, through a series of Monte Carlo simulation studies. Using the restrictions in Section \ref{sec:model}, we define four distinct bivariate geometric anisotropic LGCPs with valid Mat\'ern covariance structures; the parameter values for each model are given in Table \ref{table:MCsims_fullMatern}. For all four models, the parameters are chosen such that the expected log-intensity for each process component, $\log(\lambda_p)= 6.75$ ($p=1,2$), specifying point patterns with a similar intensity to the ecological data to be considered in Section \ref{subsec:BCIimplementation}. For each of the four fully-specified models, we simulate 500 distinct point patterns on the unit square, $W=[0,1]^2$.
For each model, and for $p,q=1,2$, we executed our parameter estimation procedure as described in Section \ref{subsec:paramest}.
For both the marginal and cross-dependence relationships, we estimate $\theta_{pq}$ using Fry processes consisting of only those point pairs separated by $r\in(0,0.25)$. Similarly, when estimating $\zeta_{pq}$, we numerically approximate the integral $V_{pq,\hat{\theta}_{pq}}(\zeta)$ as defined in \eqref{eq:intdiffK}, using the limits of integration $b_1=0$, $b_2=0.25$. Approximation of $V_{pq,\hat{\theta}_{pq}}(\zeta)$ involves estimating the sector-$K$-function over a discrete, high-resolution set of distances $r$, using an angular bandwidth parameter which we choose to be $h_{\phi} = \pi/8$ following \citet[][\S 4.3.1]{Rajala18a}, and details of the chosen sector-$K$-function estimator are given in the Appendix.
In estimating the anisotropy parameters, our choice of interval for $r$ is deliberately large relative to the true scale of dependence in all of our models, as we intend to show that reasonable results can be obtained without prior knowledge of the true scale of dependence in the data.
When estimating the Mat\'ern parameters, despite using a favourable form of the Mat\'ern parameterisation as discussed in Section \ref{subsec:MVMaterncorr},
there proved to be insufficient separation of the effects of $\nu_{pq}$ and $\alpha_{pq}$ in practice for both parameters to be allowed to vary freely during estimation. In order to avoid this issue, a common strategy \citep[e.g.][]{Diggle13b} is to restrict $\hat{\nu}_{pq}$ to three candidate values, representing three sufficiently distinct levels of smoothness in the resulting random fields: we seek $\hat{\nu}_{pq}\in\{0.05,0.5,5.0\}$, $p,q=1,2$. For the case $p\ne q$, this candidate vector was further restricted, to ensure that $\hat{\nu}_{12}$ satisfied Condition \ref{cond:nu}. The remaining Mat\'ern parameters were allowed to vary on continuous bounded intervals: $\hat{\alpha}_{pq}\in(0,\alpha^{UB}_{pq})$ and $\hat{\sigma}_{pq}\in(0,\sigma^{UB}_{pq})$. In the marginal cases, $\alpha^{UB}_{pp}=10$ and $\sigma^{UB}_{pp}=50$, $p=1,2$, were chosen such that these constituted generous intervals around the corresponding true values. For estimating the cross-covariance parameters, $\alpha^{UB}_{12}$ and $\sigma^{UB}_{12}$ were chosen to ensure compliance with Conditions \ref{cond:alpha} and \ref{cond:sigma}.
Our implementation was carried out in Matlab, where we used the default interior-point algorithm to carry out constrained maximisation of the Palm-log likelihood with respect to $(\alpha_{pq},\sigma_{pq})$, for each candidate value of $\nu_{pq}$. Since this algorithm requires the user to initialise the parameters being sought, we did so using a computationally inexpensive version of the widely-used minimum contrast method, minimising the difference between the estimated (isotropic) pair correlation function and its closed-form expression across a coarse grid of parameter pairs $(\alpha_{pq},\sigma_{pq})$.
We also detail our choice of the MPLE tuning parameter $R$. Following the guidance of \citet{Prokesova13}, we chose $R$ to be approximately equal to the range of interaction in the relevant dataset. The practical range of dependence is defined in the geostatistics literature to be the distance at which the spatial auto- or cross-correlation decays to 0.05. We calculated the practical range for each of our models, motivating our choice of $R=0.1$ for models 1 and 2, and $R=0.25$ for models 3 and 4.
In a small proportion of runs, the MPLE procedure returned seemingly degenerate estimates of either $\hat{\alpha}_{pq}$ or $\hat{\sigma}_{pq}$, $p=1,2$, with one or the other being returned equal to their upper bound. This was found to occur when the majority of the points in the corresponding dataset lay in the boundary region created using the above values of $R$. In this scenario, the number of points contributing to the Palm log-likelihoods \eqref{eq:bivpalmloglik}-\eqref{eq:palmloglik} is reduced, leading to a loss of accuracy in the MPLE procedure. We therefore counter this phenomenon by decreasing $R$ when necessary. When the initial attempt returns estimates of any of the Mat\'ern scale or power parameters greater than 95\% of their corresponding upper bound, we iteratively repeat the MPLE procedure, reducing $R$ by 0.01 each time, until all scale and power estimates are below this 95\% threshold. We found this to be an adequate, if somewhat ad-hoc remedy to the problem. After applying our iterative fix, for each of the four models considered, fewer than 8 of the 500 Monte Carlo runs returned any Mat\'ern scale or power estimates greater than $50\%$ of their corresponding upper bound.
In Table \ref{table:MCsims_fullMatern}, we provide summary statistics for the Monte Carlo sampling distributions of the parameters in Models 1-4. For the smoothness parameters, we report the modal estimate from our Monte Carlo simulations, as we consider only three potential values for these parameters. For the estimated scales of anisotropy, we provide the median of the Monte Carlo samples, along with the sample standard deviation, since their sampling distributions display evidence of skewness. For the estimated angles of anisotropy, as well as the Mat\'ern scale and power parameters, we provide the MC sample mean and the MC sample standard deviation.
The sampling distributions of the parameter estimates for Model 1 are depicted in Figure \ref{fig:MCsims_Model1}, and the corresponding figures for Models 2-4 are provided in the Appendix
\begin{table}
\begin{center}
\begin{tabular}{|l|ccc|ccc||cc|}
\hline
& $\theta_{11}$ & $\theta_{22}$ & $\theta_{12}$
& $\zeta_{11}$ & $\zeta_{22}$ & $\zeta_{12}$
& $\mu_{1}$ & $\mu_2$ \\
\hline
Dataset 1
& $36^{\circ}$ & $72^{\circ}$ & $54^{\circ}$
& 0.20 & 0.20 & 0.35
& 4.75 & 4.5 \\
MC estimate
& $40.36^{\circ}$ & $73.87^{\circ}$ & $67.07^{\circ}$
& 0.25 & 0.24 & 0.41
& 3.14 & 3.06 \\
MC std.~dev.
& $21.72^{\circ}$ & $16.19^{\circ}$ & $37.49^{\circ}$
& 0.33 & 0.27 & 0.48
& 1.62 & 2.31 \\
\hline
Dataset 2
& $36^{\circ}$ & $72^{\circ}$ & $54^{\circ}$
& 0.40 & 0.40 & 0.60
& 4.75 & 4.5 \\
MC estimate
& $42.51^{\circ}$ & $75.38^{\circ}$ & $70.90^{\circ}$
& 0.41 & 0.41 & 0.58
& 3.65 & 3.47 \\
MC std.~dev.
& $26.54^{\circ}$ & $22.22^{\circ}$ & $43.54^{\circ}$
& 0.35 & 0.28 & 0.44
& 1.91 & 2.82 \\
\hline
Dataset 3
& $36^{\circ}$ & $72^{\circ}$ & $54^{\circ}$
& 0.20 & 0.20 & 0.35
& 5.75 & 5.625 \\
MC estimate
& $32.49^{\circ}$ & $69.87^{\circ}$ & $52.07^{\circ}$
& 0.22 & 0.23 & 0.34
& 2.95 & 2.81 \\
MC std.~dev.
& $14.49^{\circ}$ & $8.53^{\circ}$ & $27.22^{\circ}$
& 0.08 & 0.12 & 0.29
& 1.40 & 2.27 \\
\hline
Dataset 4
& $36^{\circ}$ & $72^{\circ}$ & $54^{\circ}$
& 0.40 & 0.40 & 0.60
& 5.75 & 5.625 \\
MC estimate
&$39.73^{\circ}$ & $74.24^{\circ}$ & $64.94^{\circ}$
& 0.39 & 0.40 & 0.52
& 4.16 & 4.28 \\
MC std.~dev.
& $21.98^{\circ}$ & $18.31^{\circ}$ & $37.58^{\circ}$
& 0.19 & 0.23 & 0.25
& 1.72 & 1.36 \\
\hline
\end{tabular}
\begin{tabular}{|l|ccc|ccc|ccc|}
\hline
& $\alpha_{11}$ & $\alpha_{22}$ & $\alpha_{12}$ & $\nu_{11}$ & $\nu_{22}$ & $\nu_{12}$ & $\sigma_{11}$ & $\sigma_{22}$ & $\sigma_{12}$ \\
\hline
Dataset 1
& 0.045 & 0.065 & 0.050
& 0.5 & 0.5 & 0.5
& 4.00 & 4.50 & 1.97 \\
MC estimate
& 0.042 & 0.116 & 0.047
& 0.5 & 0.5 & 0.5
& 4.52 & 4.13 & 1.28 \\
MC std.~dev.
& 0.033 & 0.066 & 0.021
& - & - & -
& 2.93 & 2.11 & 1.00 \\
\hline
Dataset 2
& 0.045 & 0.065 & 0.050
& 0.5 & 0.5 & 0.5
& 4.00 & 4.50 & 2.30 \\
MC estimate
& 0.084 & 0.138 & 0.049
& 0.5 & 0.5 & 0.5
& 4.19 & 4.14 & 1.37 \\
MC std.~dev.
& 0.586 & 0.689 & 0.029
& - & - & -
& 2.47 & 2.50 & 0.98 \\
\hline
Dataset 3
& 0.090 & 0.120 & 0.100
& 0.5 & 0.5 & 0.5
& 2.00 & 2.25 & 0.98 \\
MC estimate
& 0.147 & 0.174 & 0.110
& 0.5 & 0.5 & 5.0
& 2.91 & 2.98 & 0.58 \\
MC std.~dev.
& 0.593 & 0.497 & 0.067
& - & - & -
& 2.36 & 2.40 & 0.93 \\
\hline
Dataset 4
& 0.090 & 0.120 & 0.100
& 0.5 & 0.5 & 0.5
& 2.00 & 2.25 & 1.15 \\
MC estimate
& 0.191 & 0.181 & 0.111
& 0.5 & 0.5 & 5.0
& 2.85 & 2.52 & 0.72 \\
MC std.~dev.
& 0.763 & 0.487 & 0.075
& - & - & -
& 2.19 & 1.77 & 1.02 \\
\hline
\end{tabular}
\end{center}
\caption{Monte Carlo estimates and standard errors for the anisotropy (top) and Mat\'ern (bottom) parameters in four distinct models. All estimates and errors are given to 2dp, apart from those for the scale parameters; these are presented to 3dp, due to the magnitude of the errors
}
\label{table:MCsims_fullMatern}
\end{table}
\begin{figure}
\includegraphics[width=\textwidth]{./Results1.png}
\vspace{-40pt}
\caption{Histograms of the parameter distributions for the synthetic bivariate geometric anisotropic LGCP with Mat\'ern covariance structure specified by Model 1. The parameter values used to generate each dataset are marked by vertical dashed lines.}
\label{fig:MCsims_Model1}
\end{figure}
From these results, we can identify some general conclusions regarding the performance of our model fitting procedure. Firstly, for all datasets, the estimated anisotropy parameters are in reasonable agreement with their corresponding true values. There is room for improvement in accuracy, especially in the estimated values of $\hat{\theta}_{pq}$, $p,q=1,2$; as noted above, this can be achieved through reducing the range of distances, $r$, over which we seek to characterise anisotropy. The broad accuracy of these estimates, however, suggests that our bivariate generalisation of \citeauthor{Rajala16}'s method of estimating anisotropy has been successful.
Similarly, we can see that the Mat\'ern scale parameters, $\alpha_{pq}$, have been satisfactorily estimated for all four models. For models 1 and 2, we have also recovered the correct values of the Mat\'ern smoothness parameters $\nu_{pq}$. For models 3 and 4, however, there are some notable inaccuracies in estimating $\nu_{12}$. We attribute this to the lower power in the between-process dependence structures for models 3 and 4. For all four models, the power parameter estimates $\sigma_{pq}$, $p,q=1,2$ show reasonable accuracy, though there is consistent underestimation of the joint dependence power parameter. Further examination of the empirical distributions of $\hat{\sigma}_{12}$ suggests that this can be attributed to our use of constrained optimisation of the bivariate Palm log-likelihood.
In the final panel of Figures \ref{fig:MCsims_Model1} and \ref{fig:MCsims_Model2}-\ref{fig:MCsims_Model4}, we have overlain the empirical parameter distribution for $\hat{\sigma}_{12}$, restricted to those MC simulations where $\hat{\sigma}_{12}$ was not equal to the upper bound dictated by $\hat{\sigma}_{11}$ and $\hat{\sigma}_{22}$. This suggests that our use of constrained optimisation limits the accuracy of the estimated power parameter; this is the cost of ensuring that each fitted parameter vector specifies a valid multivariate dependence structure.
Finally we note that, across all models, the estimation of $\mu_1$ and $\mu_2$ is poor. We found that this can be improved by reducing the MPLE tuning parameter $R$, but with a loss of accuracy in the resulting covariance parameters. In practice, we can of course avoid this trade-off by instead using the classical estimator for the intensity, $\hat{\lambda}_p=n_p/|W|$, and combining this with $\hat{\sigma}_{pp}$ to obtain a more accurate estimate for $\mu_{p}$.
Overall, these results indicate reasonable success for our model fitting procedure, and warrant its use in exploring the model's effectiveness in characterising real data.
\subsection{Application to ecological data}
\label{subsec:BCIimplementation}
In order to demonstrate the utility of our multivariate geometric anisotropic framework, we fit our multivariate Mat\'ern geometric anisotropic LGCP to a bivariate point pattern from a 50ha plot in the BCI forest stand in Panama. Our point pattern of interest comprises two tree species, \textit{Cecropia obtusifolia} and \textit{Spondias radlkoferi}. To ease comparison with the studies in the previous section, we rescale the coordinates to the half-unit window $[0,1]\times[0,0.5]$; this rescaled bivariate point pattern is displayed in Figure \ref{fig:BCIdata}. \textit{C. obtusifolia} and \textit{S. radlkoferi} were chosen as
a preliminary study of the data revealed empirical evidence of between-process anisotropy at a range of $r=50m$. This is demonstrated in Figure \ref{fig:MCsecKfun}, which we describe below. This preliminary evidence also motivates the scales over which we seek to characterise anisotropy in the data: for both the marginal and cross-dependence relationships, we estimate $\theta_{pq}$ using Fry processes consisting of only those point pairs separated by $r\in(0,0.05)$, and we estimate $\zeta_{pq}$ using $b_1=0$ and $b_2=0.05$ as the limits of integration in $\hat{V}_{pq,\hat{\theta}_{pq}}(\zeta)$.
As in Section \ref{subsec:proofofconcept}, we must also choose a value of the MPLE tuning parameter $R$. Once again, we do so by consulting the marginal and cross-pair correlation functions for the isotropised data, once the marginal and between-process anisotropy parameters have been estimated. We found that the corresponding implied practical range, both within-species and between-species, was in the interval $[0.1,0.2]$. We therefore executed the MPLE portion of our model fitting procedure for each of $R=0.1,0.15,0.2$.
\begin{figure}
\includegraphics[width=\textwidth]{cecropia+spondias.png}
\caption{Rescaled point pattern data from the 50ha tropical rainforest census plot on Barro Colorado Island. Two species are shown: \textit{Cecropia obtusifolia} (blue circles) and \textit{Spondias radlkoferi} (red crosses).}
\label{fig:BCIdata}
\end{figure}
For the proof-of-concept studies in Section \ref{subsec:proofofconcept}, we were able to avoid constraining the anisotropy parameters during the estimation procedure, as we knew that their true values satisfied the relevant model validity conditions of Section \ref{subsec:MVMaterncorr}. When fitting the model to observed data, however, we have no such assurance. Instead of introducing any new constraints on the anisotropy parameters here, we acknowledge this uncertainty by checking each fitted model against Conditions 1-4; all of the fitted models we present here were found to satisfy these validity conditions.
In Section \ref{subsec:proofofconcept}, we also found that estimating the Mat\'ern parameters via constrained optimisation can result in underestimation of the overall power in the between-process covariance. This occurs when the estimated value of $\sigma_{12}$ is equal to the upper bound specified by the marginal dependence structures. By calculating this upper bound explicitly, and comparing with $\hat{\sigma}_{12}$, we can therefore ascertain whether each fitted model accurately represents the between-species dependence structure; this is important, as it describes the interspecific interaction between individual trees in our dataset.
To begin with, we applied our model-fitting procedure as in Section \ref{subsec:proofofconcept}. This resulted in the model specified by the first three rows of parameter estimates in Table \ref{table:BCIfittedmodels}. Since $\hat{\sigma}_{12}=\sigma^{UB}_{12}$ in each of these specifications, we conclude that none of these fitted models accurately represent the interspecific interaction in our dataset.
Motivated by the observation that distinct values of the marginal smoothness parameters lead to prohibitively small values of $\sigma^{UB}_{12}$, we next proceeded to fix the smoothness parameters, $\nu_{11}=\nu_{22}=\nu_{12}=0.5$, such that we sought to fit a geometric anisotropic bivariate Exponential covariance structure to our data. Crucially, all of the discussion from Sections \ref{sec:model} and \ref{sec:modelfitting} is valid for fixed values of $\nu_{pq}$, $p,q=1,2$. The resulting parameter estimates for this model are given in rows 4-6 of Table \ref{table:BCIfittedmodels}.
In order to demonstrate the utility of our multivariate anisotropic framework, we also fit an isotropic version of the multivariate Exponential LGCP to the same data, for the purpose of comparison. In practice, we achieve this by fixing $\zeta_{pq}=1$ and $\theta_{pq}=0$ for $p,q=1,2$, and implementing the MPLE portion of the model fitting procedure as described above, using $R=0.1,0.15,0.2$. The resulting three sets of estimated scale and power parameters for this model are given in the bottom three rows of Table \ref{table:BCIfittedmodels}. As is shown in this table, the interspecific interaction is well-represented in only one of each of the anisotropic and isotropic fitted Exponential models. We henceforth restrict our attention to these two fitted models.
\begin{table}[t!]
\begin{center}
\begin{tabular}{|ccc|ccc||cc|}
\hline
$\hat{\theta}_{11}$ & $\hat{\theta}_{22}$ & $\hat{\theta}_{12}$
& $\hat{\zeta}_{11}$ & $\hat{\zeta}_{22}$ & $\hat{\zeta}_{12}$
& $\hat{\lambda}_{1}$ & $\hat{\lambda}_{2}$ \\
\hline
$158.89^{\circ}$ & $87.65^{\circ}$ & $127.39^{\circ}$
& 0.51 & 0.39 & 0.53
& 824 & 878 \\
\hline
\end{tabular}
\begin{tabular}{|l|c|ccc|ccc|ccc|}
\hline
Covariance model
& $R$
& $\hat{\alpha}_{11}$ & $\hat{\alpha}_{22}$ & $\hat{\alpha}_{12}$
& $\hat{\nu}_{11}$ & $\hat{\nu}_{22}$ & $\hat{\nu}_{12}$
& $\hat{\sigma}_{11}$ & $\hat{\sigma}_{22}$ & $\hat{\sigma}_{12}$ \\
\hline
Anisotropic
& 0.1
& 0.03 & 0.71 & 0.15
& 0.05 & 0.5 & 0.5
& 11.55 & 13.66 & 1.16$^*$ \\
Mat\'ern
& 0.15
& 10.00 & 0.08 & 0.11
& 0.05 & 5.0 & 5.0
& 19.05 & 3.09 & 1.01e-07$^*$ \\
& 0.2
& 0.07 & 0.10 & 0.14
& 0.05 & 5.0 & 5.0
& 12.50 & 2.00 & 0.03$^*$ \\
\hline
Anisotropic
& 0.1
& 0.03 & 0.71 & 0.04
& - & - & -
& 3.09 & 13.66 & 1.61$^*$ \\
Exponential
&\bf{0.15}
&\bf{0.10} &\bf{0.18} &\bf{0.12}
& - & - & -
&\bf{3.47} &\bf{5.22} &\bf{2.45}
\\
& 0.2
& 0.07 & 0.17 & 0.08
& - & - & -
& 3.30 & 3.03 & 1.96$^*$ \\
\hline
Isotropic
& 0.1
& 0.03 & 0.75 & 0.04
& - & - & -
& 3.32 & 8.29 & 1.44$^*$ \\
Exponential
& 0.15
& 0.08 & 0.13 & 0.08
& - & - & -
& 3.34 & 2.79 & 2.33$^*$ \\
&\bf{0.2}
&\bf{0.12} &\bf{0.09} &\bf{0.10}
& - & - & -
&\bf{4.28} &\bf{3.40} &\bf{3.50}
\\
\hline
\end{tabular}
\end{center}
\caption{Parameter estimates for three bivariate LGCPs, fitted to the tropical rainforest data described in the text. The anisotropy parameters in the top table apply to both anisotropic models described in the bottom table. Those values of $\hat{\sigma}_{12}$ marked with an asterisk ($^*$) are equal to the corresponding upper bound $\sigma^{UB}_{12}$. The two models that we choose to assess using global envelope tests, are highlighted in bold.}
\label{table:BCIfittedmodels}
\end{table}
In order to assess different aspects of each model's performance, we use global envelope tests (GETs), in which we compare second-order statistics for the observed data with those of $M$ bivariate point patterns, independently simulated from the fitted model. Such tests were developed by \citet{Myllymaki17} to address multiple testing concerns with regards to the popular use of Monte Carlo envelope tests \citep{Loosemore06,Baddeley14}. The envelopes provided by the GETs describe a proper statistical test: if the observed test statistic lies outside the simulated envelope \textit{at any instance}, then the null hypothesis that the observed data belong to the fitted model may be rejected. In order to construct the envelopes, we use one of the following two approaches, both of which are described in detail by \citet{Myllymaki17}. For symmetric second-order statistics, we use the scaled studentized maximum absolute difference (MAD) to construct the critical bounds. For asymmetric second-order statistics, it is more appropriate to construct the envelopes using the scaled directional quantile MAD. In both cases, we configure our tests such that they have a global type I error probability of 0.1, using $M=499$.
To assess each model's description of bivariate anisotropy in the data, we estimate the the sector-$K$-function \eqref{eq:sectorKfun} at a distance $r=0.05$, and at the angles $\phi=k\pi/60$, $k=0,\ldots,60$. This assessment is summarised in Figure \ref{fig:MCsecKfun}, which gives the estimated marginal and between-process sector-$K$-functions for the observed BCI data, along with their corresponding directional quantile MAD envelopes.
To assess the fitted model's ability to replicate aggregation in the observed multi-type point pattern, we use the `$p$-to-$q$' nearest-neighbour distance distribution function $G_{pq}(r)$ \citep{VanLieshout99}. This describes the empirical distribution of the absolute distance between the typical point of type $p$ and its nearest point of type $q$ and is, in general, asymmetric in $p,q$.
We calculate $G_{pq}(r)$ at a range of distances, using a bivariate version of the border-corrected estimators detailed by \citet[][\S 8.11.3]{Baddeley15}. The resulting estimates for the observed BCI data are provided in Figure \ref{fig:MCnnddfun
, along with their corresponding studentized MAD envelopes.
From Figure \ref{fig:MCsecKfun}, we can see that the two chosen species in the BCI forest stand exhibit anisotropic interspecific interaction at a range of $50m$: for $\phi\in[7\pi/60,9\pi/60]\cup[12\pi/60,18\pi/60]$, the estimated sector-$K$-function $K_{12}^a(0.05,\phi)$ lies outside of the envelope generated by a multivariate isotropic LGCP. The $p$-value for the global envelope test that was carried out for this statistic was $0.058$, indicating departure from the isotropic model when using a global type I error probability of 0.1. From the bottom row of panels, we can see that our multivariate geometric anisotropic LGCP can comfortably replicate this observed heterogeneity.
Finally, the bottom row of panels in Figure \ref{fig:MCnnddfun} demonstrates that our fitted multivariate anisotropic LGCP also accurately captures the clustering behaviour exhibited by the observed bivariate data. The top panels in this Figure suggest that, despite not being able to account for the anisotropy in the data, the multivariate isotropic LGCP has also captured the clustering behaviour evident in this particular example.
\begin{figure}[t!]
\includegraphics[width=\textwidth]{./BCI_GETs_secKfun_isot.png}
\vspace{-30pt}
\\
\includegraphics[width=\textwidth]{./BCI_GETs_secKfun_anisot.png}
\caption{Estimates of the sector-$K$-function \eqref{eq:sectorKfun}, at a fixed range of $50m$, for the observed bivariate BCI point pattern (black line), along with the corresponding 90\% directional-quantile MAD envelopes obtained from a fitted multivariate geometric anisotropic LGCP (bottom row) and from a fitted multivariate isotropic LGCP (top row). Departure of the data from the adopted model is highlighted with red circles. The vertical axes are presented on a log-scale to highlight the departure of the data from the isotropic model in the third panel.
}
\label{fig:MCsecKfun}
\end{figure}
\begin{figure}[t!]
\includegraphics[width=\textwidth]{BCI_GETs_Gfun_isot.png}
\vspace{-30pt}
\\
\includegraphics[width=\textwidth]{BCI_GETs_Gfun_anisot.png}
\caption{Estimated nearest-neighbour distance distribution functions for the observed bivariate BCI point pattern (black line), along with the corresponding 90\% studentized MAD envelopes obtained from a fitted multivariate geometric anisotropic LGCP (bottom row) and from a fitted multivariate isotropic LGCP (top row).}
\label{fig:MCnnddfun}
\end{figure}
\section{Discussion}
Using the model-fitting methodology described in Section \ref{sec:modelfitting}, we have shown that by incorporating geometric anisotropy into the between-process dependence, as well as the marginal dependence, we can construct a LGCP that more accurately replicates any rotationally heterogeneous interaction between points in a multi-type point pattern.
We have focussed here on a covariate-free approach, motivated in part by the desire to allow the description of anisotropic between-process dependence in data for which there are no explanatory spatial variables. Nevertheless, the models presented here are flexible enough to use (potentially incomplete) covariate information where it is available.
Indeed, an interesting first extension of this work would be to incorporate covariates into the first-order description of the GRF underlying our LGCPs; for instance, the expected value of the GRF could be specified through a linear regression model, and inference with respect to the regression parameters may be achievable through the use of estimating functions \citep[e.g.][]{Waagepetersen08,Waagepetersen09}. Such an approach would allow the user to exploit any knowledge of spatial covariates whilst being confident that any residual heterogeneity in the data would be accounted for by the increased flexibility of the multivariate geometric anisotropic second-order dependence structure.
\section{Acknowledgements}
The work of J. S. Martin, D. J. Murrell and S. C. Olhede was supported by the UK Engineering and Physical Sciences Research Council via EP/N007336/1, and EP/L001519/1. S. C. Olhede also acknowledges support from the 7th European Community Framework Programme via a Grant CoG 2015- 682172NETS (Olhede).
The BCI forest dynamics research project was founded by S. P. Hubbell and R. B. Foster and is now managed by R. Condit, S. Lao, and R. Perez under the Center for Tropical Forest Science and the Smithsonian Tropical Research in Panama. Numerous organizations have provided funding, principally the U.S. National Science Foundation, and hundreds of field workers have contributed.
\newpage
|
\section{Introduction}
The journal \textit{Monthly Notices of the Royal Astronomical Society} (MNRAS) encourages authors to prepare their papers using \LaTeX.
The style file \verb'mnras.cls' can be used to approximate the final appearance of the journal, and provides numerous features to simplify the preparation of papers.
This document, \verb'mnras_guide.tex', provides guidance on using that style file and the features it enables.
This is not a general guide on how to use \LaTeX, of which many excellent examples already exist.
We particularly recommend \textit{Wikibooks \LaTeX}\footnote{\url{https://en.wikibooks.org/wiki/LaTeX}}, a collaborative online textbook which is of use to both beginners and experts.
Alternatively there are several other online resources, and most academic libraries also hold suitable beginner's guides.
For guidance on the contents of papers, journal style, and how to submit a paper, see the MNRAS Instructions to Authors\footnote{\label{foot:itas}\url{http://www.oxfordjournals.org/our_journals/mnras/for_authors/}}.
Only technical issues with the \LaTeX\ class are considered here.
\section{Obtaining and installing the MNRAS package}
Some \LaTeX\ distributions come with the MNRAS package by default.
If yours does not, you can either install it using your distribution's package manager, or download it from the Comprehensive \TeX\ Archive Network\footnote{\url{http://www.ctan.org/tex-archive/macros/latex/contrib/mnras}} (CTAN).
The files can either be installed permanently by placing them in the appropriate directory (consult the documentation for your \LaTeX\ distribution), or used temporarily by placing them in the working directory for your paper.
To use the MNRAS package, simply specify \verb'mnras' as the document class at the start of a \verb'.tex' file:
\begin{verbatim}
\documentclass{mnras}
\end{verbatim}
Then compile \LaTeX\ (and if necessary \bibtex) in the usual way.
\section{Preparing and submitting a paper}
We recommend that you start with a copy of the \texttt{mnras\_template.tex} file.
Rename the file, update the information on the title page, and then work on the text of your paper.
Guidelines for content, style etc. are given in the instructions to authors on the journal's website$^{\ref{foot:itas}}$.
Note that this document does not follow all the aspects of MNRAS journal style (e.g. it has a table of contents).
If a paper is accepted, it is professionally typeset and copyedited by the publishers.
It is therefore likely that minor changes to presentation will occur.
For this reason, we ask authors to ignore minor details such as slightly long lines, extra blank spaces, or misplaced figures, because these details will be dealt with during the production process.
Papers must be submitted electronically via the online submission system; paper submissions are not permitted.
For full guidance on how to submit a paper, see the instructions to authors.
\section{Class options}
\label{sec:options}
There are several options which can be added to the document class line like this:
\begin{verbatim}
\documentclass[option1,option2]{mnras}
\end{verbatim}
The available options are:
\begin{itemize}
\item \verb'letters' -- used for papers in the journal's Letters section.
\item \verb'onecolumn' -- single column, instead of the default two columns. This should be used {\it only} if necessary for the display of numerous very long equations.
\item \verb'doublespacing' -- text has double line spacing. Please don't submit papers in this format.
\item \verb'referee' -- \textit{(deprecated)} single column, double spaced, larger text, bigger margins. Please don't submit papers in this format.
\item \verb'galley' -- \textit{(deprecated)} no running headers, no attempt to align the bottom of columns.
\item \verb'landscape' -- \textit{(deprecated)} sets the whole document on landscape paper.
\item \verb"usenatbib" -- \textit{(all papers should use this)} this uses Patrick Daly's \verb"natbib.sty" package for citations.
\item \verb"usegraphicx" -- \textit{(most papers will need this)} includes the \verb'graphicx' package, for inclusion of figures and images.
\item \verb'useAMS' -- adds support for upright Greek characters \verb'\upi', \verb'\umu' and \verb'\upartial' ($\upi$, $\umu$ and $\upartial$). Only these three are included, if you require other symbols you will need to include the \verb'amsmath' or \verb'amsymb' packages (see section~\ref{sec:packages}).
\item \verb"usedcolumn" -- includes the package \verb"dcolumn", which includes two new types of column alignment for use in tables.
\end{itemize}
Some of these options are deprecated and retained for backwards compatibility only.
Others are used in almost all papers, but again are retained as options to ensure that papers written decades ago will continue to compile without problems.
If you want to include any other packages, see section~\ref{sec:packages}.
\section{Title page}
If you are using \texttt{mnras\_template.tex} the necessary code for generating the title page, headers and footers is already present.
Simply edit the title, author list, institutions, abstract and keywords as described below.
\subsection{Title}
There are two forms of the title: the full version used on the first page, and a short version which is used in the header of other odd-numbered pages (the `running head').
Enter them with \verb'\title[]{}' like this:
\begin{verbatim}
\title[Running head]{Full title of the paper}
\end{verbatim}
The full title can be multiple lines (use \verb'\\' to start a new line) and may be as long as necessary, although we encourage authors to use concise titles. The running head must be $\le~45$ characters on a single line.
See appendix~\ref{sec:advanced} for more complicated examples.
\subsection{Authors and institutions}
Like the title, there are two forms of author list: the full version which appears on the title page, and a short form which appears in the header of the even-numbered pages. Enter them using the \verb'\author[]{}' command.
If the author list is more than one line long, start a new line using \verb'\newauthor'. Use \verb'\\' to start the institution list. Affiliations for each author should be indicated with a superscript number, and correspond to the list of institutions below the author list.
For example, if I were to write a paper with two coauthors at another institution, one of whom also works at a third location:
\begin{verbatim}
\author[K. T. Smith et al.]{
Keith T. Smith,$^{1}$
A. N. Other,$^{2}$
and Third Author$^{2,3}$
\\
$^{1}$Affiliation 1\\
$^{2}$Affiliation 2\\
$^{3}$Affiliation 3}
\end{verbatim}
Affiliations should be in the format `Department, Institution, Street Address, City and Postal Code, Country'.
Email addresses can be inserted with the \verb'\thanks{}' command which adds a title page footnote.
If you want to list more than one email, put them all in the same \verb'\thanks' and use \verb'\footnotemark[]' to refer to the same footnote multiple times.
Present addresses (if different to those where the work was performed) can also be added with a \verb'\thanks' command.
\subsection{Abstract and keywords}
The abstract is entered in an \verb'abstract' environment:
\begin{verbatim}
\begin{abstract}
The abstract of the paper.
\end{abstract}
\end{verbatim}
\noindent Note that there is a word limit on the length of abstracts.
For the current word limit, see the journal instructions to authors$^{\ref{foot:itas}}$.
Immediately following the abstract, a set of keywords is entered in a \verb'keywords' environment:
\begin{verbatim}
\begin{keywords}
keyword 1 -- keyword 2 -- keyword 3
\end{keywords}
\end{verbatim}
\noindent There is a list of permitted keywords, which is agreed between all the major astronomy journals and revised every few years.
Do \emph{not} make up new keywords!
For the current list of allowed keywords, see the journal's instructions to authors$^{\ref{foot:itas}}$.
\section{Sections and lists}
Sections and lists are generally the same as in the standard \LaTeX\ classes.
\subsection{Sections}
\label{sec:sections}
Sections are entered in the usual way, using \verb'\section{}' and its variants. It is possible to nest up to four section levels:
\begin{verbatim}
\section{Main section}
\subsection{Subsection}
\subsubsection{Subsubsection}
\paragraph{Lowest level section}
\end{verbatim}
\noindent The other \LaTeX\ sectioning commands \verb'\part', \verb'\chapter' and \verb'\subparagraph{}' are deprecated and should not be used.
Some sections are not numbered as part of journal style (e.g. the Acknowledgements).
To insert an unnumbered section use the `starred' version of the command: \verb'\section*{}'.
See appendix~\ref{sec:advanced} for more complicated examples.
\subsection{Lists}
Two forms of lists can be used in MNRAS -- numbered and unnumbered.
For a numbered list, use the \verb'enumerate' environment:
\begin{verbatim}
\begin{enumerate}
\item First item
\item Second item
\item etc.
\end{enumerate}
\end{verbatim}
\noindent which produces
\begin{enumerate}
\item First item
\item Second item
\item etc.
\end{enumerate}
Note that the list uses lowercase Roman numerals, rather than the \LaTeX\ default Arabic numerals.
For an unnumbered list, use the \verb'description' environment without the optional argument:
\begin{verbatim}
\begin{description}
\item First item
\item Second item
\item etc.
\end{description}
\end{verbatim}
\noindent which produces
\begin{description}
\item First item
\item Second item
\item etc.
\end{description}
Bulleted lists using the \verb'itemize' environment should not be used in MNRAS; it is retained for backwards compatibility only.
\section{Mathematics and symbols}
The MNRAS class mostly adopts standard \LaTeX\ handling of mathematics, which is briefly summarised here.
See also section~\ref{sec:packages} for packages that support more advanced mathematics.
Mathematics can be inserted into the running text using the syntax \verb'$1+1=2$', which produces $1+1=2$.
Use this only for short expressions or when referring to mathematical quantities; equations should be entered as described below.
\subsection{Equations}
Equations should be entered using the \verb'equation' environment, which automatically numbers them:
\begin{verbatim}
\begin{equation}
a^2=b^2+c^2
\end{equation}
\end{verbatim}
\noindent which produces
\begin{equation}
a^2=b^2+c^2
\end{equation}
By default, the equations are numbered sequentially throughout the whole paper. If a paper has a large number of equations, it may be better to number them by section (2.1, 2.2 etc.). To do this, add the command \verb'\numberwithin{equation}{section}' to the preamble.
It is also possible to produce un-numbered equations by using the \LaTeX\ built-in \verb'\['\textellipsis\verb'\]' and \verb'$$'\textellipsis\verb'$$' commands; however MNRAS requires that all equations are numbered, so these commands should be avoided.
\subsection{Special symbols}
\begin{table}
\caption{Additional commands for special symbols commonly used in astronomy. These can be used anywhere.}
\label{tab:anysymbols}
\begin{tabular}{lll}
\hline
Command & Output & Meaning\\
\hline
\verb'\sun' & \sun & Sun, solar\\[2pt]
\verb'\earth' & \earth & Earth, terrestrial\\[2pt]
\verb'\micron' & \micron & microns\\[2pt]
\verb'\degr' & \degr & degrees\\[2pt]
\verb'\arcmin' & \arcmin & arcminutes\\[2pt]
\verb'\arcsec' & \arcsec & arcseconds\\[2pt]
\verb'\fdg' & \fdg & fraction of a degree\\[2pt]
\verb'\farcm' & \farcm & fraction of an arcminute\\[2pt]
\verb'\farcs' & \farcs & fraction of an arcsecond\\[2pt]
\verb'\fd' & \fd & fraction of a day\\[2pt]
\verb'\fh' & \fh & fraction of an hour\\[2pt]
\verb'\fm' & \fm & fraction of a minute\\[2pt]
\verb'\fs' & \fs & fraction of a second\\[2pt]
\verb'\fp' & \fp & fraction of a period\\[2pt]
\verb'\diameter' & \diameter & diameter\\[2pt]
\verb'\sq' & \sq & square, Q.E.D.\\[2pt]
\hline
\end{tabular}
\end{table}
\begin{table}
\caption{Additional commands for mathematical symbols. These can only be used in maths mode.}
\label{tab:mathssymbols}
\begin{tabular}{lll}
\hline
Command & Output & Meaning\\
\hline
\verb'\upi' & $\upi$ & upright pi\\[2pt]
\verb'\umu' & $\umu$ & upright mu\\[2pt]
\verb'\upartial' & $\upartial$ & upright partial derivative\\[2pt]
\verb'\lid' & $\lid$ & less than or equal to\\[2pt]
\verb'\gid' & $\gid$ & greater than or equal to\\[2pt]
\verb'\la' & $\la$ & less than of order\\[2pt]
\verb'\ga' & $\ga$ & greater than of order\\[2pt]
\verb'\loa' & $\loa$ & less than approximately\\[2pt]
\verb'\goa' & $\goa$ & greater than approximately\\[2pt]
\verb'\cor' & $\cor$ & corresponds to\\[2pt]
\verb'\sol' & $\sol$ & similar to or less than\\[2pt]
\verb'\sog' & $\sog$ & similar to or greater than\\[2pt]
\verb'\lse' & $\lse$ & less than or homotopic to \\[2pt]
\verb'\gse' & $\gse$ & greater than or homotopic to\\[2pt]
\verb'\getsto' & $\getsto$ & from over to\\[2pt]
\verb'\grole' & $\grole$ & greater over less\\[2pt]
\verb'\leogr' & $\leogr$ & less over greater\\
\hline
\end{tabular}
\end{table}
Some additional symbols of common use in astronomy have been added in the MNRAS class. These are shown in tables~\ref{tab:anysymbols}--\ref{tab:mathssymbols}. The command names are -- as far as possible -- the same as those used in other major astronomy journals.
Many other mathematical symbols are also available, either built into \LaTeX\ or via additional packages. If you want to insert a specific symbol but don't know the \LaTeX\ command, we recommend using the Detexify website\footnote{\url{http://detexify.kirelabs.org}}.
Sometimes font or coding limitations mean a symbol may not get smaller when used in sub- or superscripts, and will therefore be displayed at the wrong size. There is no need to worry about this as it will be corrected by the typesetter during production.
To produce bold symbols in mathematics, use \verb'\bmath' for simple variables, and the \verb'bm' package for more complex symbols (see section~\ref{sec:packages}). Vectors are set in bold italic, using \verb'\mathbfit{}'.
For matrices, use \verb'\mathbfss{}' to produce a bold sans-serif font e.g. \mathbfss{H}; this works even outside maths mode, but not all symbols are available (e.g. Greek). For $\nabla$ (del, used in gradients, divergence etc.) use \verb'$\nabla$'.
\subsection{Ions}
A new \verb'\ion{}{}' command has been added to the class file, for the correct typesetting of ionisation states.
For example, to typeset singly ionised calcium use \verb'\ion{Ca}{ii}', which produces \ion{Ca}{ii}.
\section{Figures and tables}
\label{sec:fig_table}
Figures and tables (collectively called `floats') are mostly the same as built into \LaTeX.
\subsection{Basic examples}
\begin{figure}
\includegraphics[width=\columnwidth]{example}
\caption{An example figure.}
\label{fig:example}
\end{figure}
Figures are inserted in the usual way using a \verb'figure' environment and \verb'\includegraphics'. The example Figure~\ref{fig:example} was generated using the code:
\begin{verbatim}
\begin{figure}
\includegraphics[width=\columnwidth]{example}
\caption{An example figure.}
\label{fig:example}
\end{figure}
\end{verbatim}
\begin{table}
\caption{An example table.}
\label{tab:example}
\begin{tabular}{lcc}
\hline
Star & Mass & Luminosity\\
& $M_{\sun}$ & $L_{\sun}$\\
\hline
Sun & 1.00 & 1.00\\
$\alpha$~Cen~A & 1.10 & 1.52\\
$\epsilon$~Eri & 0.82 & 0.34\\
\hline
\end{tabular}
\end{table}
The example Table~\ref{tab:example} was generated using the code:
\begin{verbatim}
\begin{table}
\caption{An example table.}
\label{tab:example}
\begin{tabular}{lcc}
\hline
Star & Mass & Luminosity\\
& $M_{\sun}$ & $L_{\sun}$\\
\hline
Sun & 1.00 & 1.00\\
$\alpha$~Cen~A & 1.10 & 1.52\\
$\epsilon$~Eri & 0.82 & 0.34\\
\hline
\end{tabular}
\end{table}
\end{verbatim}
\subsection{Captions and placement}
Captions go \emph{above} tables but \emph{below} figures, as in the examples above.
The \LaTeX\ float placement commands \verb'[htbp]' are intentionally disabled.
Layout of figures and tables will be adjusted by the publisher during the production process, so authors should not concern themselves with placement to avoid disappointment and wasted effort.
Simply place the \LaTeX\ code close to where the figure or table is first mentioned in the text and leave exact placement to the publishers.
By default a figure or table will occupy one column of the page.
To produce a wider version which covers both columns, use the \verb'figure*' or \verb'table*' environment.
If a figure or table is too long to fit on a single page it can be split it into several parts.
Create an additional figure or table which uses \verb'\contcaption{}' instead of \verb'\caption{}'.
This will automatically correct the numbering and add `\emph{continued}' at the start of the caption.
\begin{table}
\contcaption{A table continued from the previous one.}
\label{tab:continued}
\begin{tabular}{lcc}
\hline
Star & Mass & Luminosity\\
& $M_{\sun}$ & $L_{\sun}$\\
\hline
$\tau$~Cet & 0.78 & 0.52\\
$\delta$~Pav & 0.99 & 1.22\\
$\sigma$~Dra & 0.87 & 0.43\\
\hline
\end{tabular}
\end{table}
Table~\ref{tab:continued} was generated using the code:
\begin{verbatim}
\begin{table}
\contcaption{A table continued from the previous one.}
\label{tab:continued}
\begin{tabular}{lcc}
\hline
Star & Mass & Luminosity\\
& $M_{\sun}$ & $L_{\sun}$\\
\hline
$\tau$~Cet & 0.78 & 0.52\\
$\delta$~Pav & 0.99 & 1.22\\
$\sigma$~Dra & 0.87 & 0.43\\
\hline
\end{tabular}
\end{table}
\end{verbatim}
To produce a landscape figure or table, use the \verb'pdflscape' package and the \verb'landscape' environment.
The landscape Table~\ref{tab:landscape} was produced using the code:
\begin{verbatim}
\begin{landscape}
\begin{table}
\caption{An example landscape table.}
\label{tab:landscape}
\begin{tabular}{cccccccccc}
\hline
Header & Header & ...\\
Unit & Unit & ...\\
\hline
Data & Data & ...\\
Data & Data & ...\\
...\\
\hline
\end{tabular}
\end{table}
\end{landscape}
\end{verbatim}
Unfortunately this method will force a page break before the table appears.
More complicated solutions are possible, but authors shouldn't worry about this.
\begin{landscape}
\begin{table}
\caption{An example landscape table.}
\label{tab:landscape}
\begin{tabular}{cccccccccc}
\hline
Header & Header & Header & Header & Header & Header & Header & Header & Header & Header\\
Unit & Unit & Unit & Unit & Unit & Unit & Unit & Unit & Unit & Unit \\
\hline
Data & Data & Data & Data & Data & Data & Data & Data & Data & Data\\
Data & Data & Data & Data & Data & Data & Data & Data & Data & Data\\
Data & Data & Data & Data & Data & Data & Data & Data & Data & Data\\
Data & Data & Data & Data & Data & Data & Data & Data & Data & Data\\
Data & Data & Data & Data & Data & Data & Data & Data & Data & Data\\
Data & Data & Data & Data & Data & Data & Data & Data & Data & Data\\
Data & Data & Data & Data & Data & Data & Data & Data & Data & Data\\
Data & Data & Data & Data & Data & Data & Data & Data & Data & Data\\
\hline
\end{tabular}
\end{table}
\end{landscape}
\section{References and citations}
\subsection{Cross-referencing}
The usual \LaTeX\ commands \verb'\label{}' and \verb'\ref{}' can be used for cross-referencing within the same paper.
We recommend that you use these whenever relevant, rather than writing out the section or figure numbers explicitly.
This ensures that cross-references are updated whenever the numbering changes (e.g. during revision) and provides clickable links (if available in your compiler).
It is best to give each section, figure and table a logical label.
For example, Table~\ref{tab:mathssymbols} has the label \verb'tab:mathssymbols', whilst section~\ref{sec:packages} has the label \verb'sec:packages'.
Add the label \emph{after} the section or caption command, as in the examples in sections~\ref{sec:sections} and \ref{sec:fig_table}.
Enter the cross-reference with a non-breaking space between the type of object and the number, like this: \verb'see Figure~\ref{fig:example}'.
The \verb'\autoref{}' command can be used to automatically fill out the type of object, saving on typing.
It also causes the link to cover the whole phrase rather than just the number, but for that reason is only suitable for single cross-references rather than ranges.
For example, \verb'\autoref{tab:journal_abbr}' produces \autoref{tab:journal_abbr}.
\subsection{Citations}
\label{sec:cite}
MNRAS uses the Harvard -- author (year) -- citation style, e.g. \citet{author2013}.
This is implemented in \LaTeX\ via the \verb'natbib' package, which in turn is included via the \verb'usenatbib' package option (see section~\ref{sec:options}), which should be used in all papers.
Each entry in the reference list has a `key' (see section~\ref{sec:ref_list}) which is used to generate citations.
There are two basic \verb'natbib' commands:
\begin{description}
\item \verb'\citet{key}' produces an in-text citation: \citet{author2013}
\item \verb'\citep{key}' produces a bracketed (parenthetical) citation: \citep{author2013}
\end{description}
Citations will include clickable links to the relevant entry in the reference list, if supported by your \LaTeX\ compiler.
\defcitealias{smith2014}{Paper~I}
\begin{table*}
\caption{Common citation commands, provided by the \texttt{natbib} package.}
\label{tab:natbib}
\begin{tabular}{lll}
\hline
Command & Ouput & Note\\
\hline
\verb'\citet{key}' & \citet{smith2014} & \\
\verb'\citep{key}' & \citep{smith2014} & \\
\verb'\citep{key,key2}' & \citep{smith2014,jones2015} & Multiple papers\\
\verb'\citet[table 4]{key}' & \citet[table 4]{smith2014} & \\
\verb'\citep[see][figure 7]{key}' & \citep[see][figure 7]{smith2014} & \\
\verb'\citealt{key}' & \citealt{smith2014} & For use with manual brackets\\
\verb'\citeauthor{key}' & \citeauthor{smith2014} & If already cited in close proximity\\
\verb'\defcitealias{key}{Paper~I}' & & Define an alias (doesn't work in floats)\\
\verb'\citetalias{key}' & \citetalias{smith2014} & \\
\verb'\citepalias{key}' & \citepalias{smith2014} & \\
\hline
\end{tabular}
\end{table*}
There are a number of other \verb'natbib' commands which can be used for more complicated citations.
The most commonly used ones are listed in Table~\ref{tab:natbib}.
For full guidance on their use, consult the \verb'natbib' documentation\footnote{\url{http://www.ctan.org/pkg/natbib}}.
If a reference has several authors, \verb'natbib' will automatically use `et al.' if there are more than two authors. However, if a paper has exactly three authors, MNRAS style is to list all three on the first citation and use `et al.' thereafter. If you are using \bibtex\ (see section~\ref{sec:ref_list}) then this is handled automatically. If not, the \verb'\citet*{}' and \verb'\citep*{}' commands can be used at the first citation to include all of the authors.
\subsection{The list of references}
\label{sec:ref_list}
It is possible to enter references manually using the usual \LaTeX\ commands, but we strongly encourage authors to use \bibtex\ instead.
\bibtex\ ensures that the reference list is updated automatically as references are added or removed from the paper, puts them in the correct format, saves on typing, and the same reference file can be used for many different papers -- saving time hunting down reference details.
An MNRAS \bibtex\ style file, \verb'mnras.bst', is distributed as part of this package.
The rest of this section will assume you are using \bibtex.
References are entered into a separate \verb'.bib' file in standard \bibtex\ formatting.
This can be done manually, or there are several software packages which make editing the \verb'.bib' file much easier.
We particularly recommend \textsc{JabRef}\footnote{\url{http://jabref.sourceforge.net/}}, which works on all major operating systems.
\bibtex\ entries can be obtained from the NASA Astrophysics Data System\footnote{\label{foot:ads}\url{http://adsabs.harvard.edu}} (ADS) by clicking on `Bibtex entry for this abstract' on any entry.
Simply copy this into your \verb'.bib' file or into the `BibTeX source' tab in \textsc{JabRef}.
Each entry in the \verb'.bib' file must specify a unique `key' to identify the paper, the format of which is up to the author.
Simply cite it in the usual way, as described in section~\ref{sec:cite}, using the specified key.
Compile the paper as usual, but add an extra step to run the \texttt{bibtex} command.
Consult the documentation for your compiler or latex distribution.
Correct formatting of the reference list will be handled by \bibtex\ in almost all cases, provided that the correct information was entered into the \verb'.bib' file.
Note that ADS entries are not always correct, particularly for older papers and conference proceedings, so may need to be edited.
If in doubt, or if you are producing the reference list manually, see the MNRAS instructions to authors$^{\ref{foot:itas}}$ for the current guidelines on how to format the list of references.
\section{Appendices and online material}
To start an appendix, simply place the \verb'
\section{Introduction}
With astronomy becoming increasingly characterized by large surveys and big data sets, machine-learning techniques have become staples of astronomical data analysis that are used for low-level data processing, classification, interpolation, pattern recognition, and parameter inference. Among modern machine-learning techniques, deep learning using artificial neural networks (ANNs) is getting increasing attention from astronomers, because of its great potential for data-driven astronomy and its recent successes in many fields such as computer vision, voice recognition, machine translation, etc. While ANNs have been around for decades, they have only become one of the dominant machine-learning techniques in the last few years. The reasons behind this recent progress are the combination of big data, cheap availability of fast computational hardware, advances in the methodology of ANNs, and the availability and accessibility of software platforms implementing this technology \citep{2017arXiv171205855S}. In astronomy, the big data era is now fully upon us: large photometric \citep[e.g., SDSS;][]{2011ApJS..193...29A}, spectroscopic \citep[e.g., APOGEE;][]{2017AJ....154...94M}, astrometric \citep[Gaia;][]{2016A&A...595A...1G}, and time-domain \citep[e.g., PTF;][]{2009PASP..121.1395L} data sets already exist and will grow exponentially in terms of quality and quantity with upcoming projects like the Large Synoptic Survey Telescope \citep[LSST;][]{2009arXiv0912.0201L}, Euclid \citep{2011arXiv1110.3193L}, and projects like the Maunakea Spectroscopic Explorer \citep[MSE;][]{2016arXiv160600043M}. ANNs are poised to play an outsized role in this data-rich future.
The hardware and software landscape for machine learning has vastly changed in the last decade. In particular, the availability of cheap graphics processing units (GPUs) is driven by the development and demand of high performance low cost personal gaming, but modern machine learning methods are ideally suited to be run on GPUs. For example, the code and analysis that we describe in this paper is entirely run on a personal desktop computer with a $\approx\$ 500$ consumer GPU accelerated by the NVIDIA CUDA Deep Neural Network library \citep{2014arXiv1410.0759C}. The accessibility of software technology for machine learning is supported by open source communities. For example, the open source python deep learning libraries used in this work---\texttt{Tensorflow} \citep{2016arXiv160304467A}, \texttt{keras} \citep{keras2015}, and the package developed by us and described in this paper \texttt{astroNN} \footnote{\url{https://github.com/henrysky/astroNN}} (see Appendix \ref{appendix:graph:astroNN}). The combination of these factors allows astronomers to easily exploit deep learning in astronomical big data analysis without the high cost of development in human resources, time, hardware, and software.
In this work we investigate the application of deep learning to the analysis of high-resolution spectroscopic data. Such data contain a wealth of information about the overall physical state of stars and about the abundances of different elements in their photospheres \citep{Gray05a}. This information is traditionally extracted using tools such as the curve of growth, equivalent widths, or forward modeling with synthetic spectra in what is often a laborious and tedious effort. Here we demonstrate that, as long as a small---thousands of stars---training set of data analyzed with more traditional means is available, ANNs can process high-resolution spectra faster and more reliably than other methods.
ANNs have been used for spectrosopic analysis before. Because of the paucity of high-resolution spectra until recently, early use of ANNs was mostly limited to training the network on libraries of synthetic spectra \citep[e.g.,][]{1997MNRAS.292..157B,2000A&A...357..197B,2015MNRAS.452..158Y}. The large spectroscopic databases provided by the SEGUE \citep{2009AJ....137.4377Y} and APOGEE surveys allowed ANNs to be trained directly on observed stellar spectra and map them onto stellar parameters (SEGUE: \citealt{2007A&A...467.1373R}, \citealt{2008AJ....136.2022L}; APOGEE: \citealt{2018MNRAS.475.2978F}). Uncertainty estimation has been explored using generative ANNs (GANs) for the \emph{Gaia} RVS data by \citet{2016A&A...594A..68D}.
The availability of the large and rich APOGEE spectroscopic data set has spurred many applications of machine-learning methods to these data. Examples of these are the \texttt{Cannon 2} \citep{2016arXiv160303040C} method for data-driven abundance analysis, dimensionality reduction of the spectral space to determine the dimensionality of abundance space in the Milky Way \citep{2018MNRAS.475.1410P}, and machine-learned outlier detection and similarity directly using the spectra \citep{2018MNRAS.476.2117R}.
The work most directly related to that described in this paper is the recent \texttt{StarNet} ANN, which is trained on spectroscopic data from APOGEE \citep{2018MNRAS.475.2978F}. \texttt{StarNet} uses a convolutional neural network to infer three labels $[\ensuremath{T_\mathrm{eff}}, \ensuremath{\log g}, \xh{Fe}]$ from high-resolution spectra and demonstrated that deep learning is an effective way both in terms of performance and of accuracy to do spectroscopic analysis when the number of training data is large.
In this work, we go beyond the \texttt{StarNet} method in various way: (a) we present a robust objective function for the neural network to learn from incomplete data while taking uncertainty in the training labels into account, (b) we use a Bayesian neural network with dropout variational inference with this objective function to estimate the uncertainties of labels determined by the neural network \citep{2015arXiv150602142G}, (c) we simultaneously infer 22 stellar and elemental abundance labels accurately and precisely for both high and low signal-to-noise ratio (SNR) spectra while constraining the model to reflect our physical understanding of stellar spectra, (d) we implement the method on a GPU using standard tools allowing for more than an order of magnitude speed-up and make these easily accessible (see Appendix \ref{subsec:fastMC}), and (e) we demonstrate that a large neural network can work well with a limited amount of training data (thousands of high SNR stellar spectra). We also present these 22 stellar parameters and elemental abundance predictions with uncertainties for the entire APOGEE DR14 data set.
The outline of this paper is as follows. Section \ref{sec:nn-general} describes the basics of ANNs, of dropout variational inference, and of our robust objective function. Section \ref{subsec:apogee-data} discusses the data selection and processing from APOGEE DR14 to construct training and test sets. Section \ref{sec:main_nn} describes the performance of our trained neural network on unseen individual and combined spectra in the test sets, on stars in open and globular clusters, and we present the results from a sensitivity analysis to understand the neural network. Section 5 describes variations in NN training such as: training on the full, uncensored spectrum, training with small data sets with only thousands of spectra, and training with a different continuum normalization process. Section 6 discusses the fast performance of the neural network and what types of future work this allows, and comparisons to other, similar spectroscopic analysis approaches. Section 7 gives our conclusions. Appendix A describes the \texttt{astroNN} python package developed for this work and gives instructions on how to perform variational inference on arbitrary APOGEE spectra.
Code to reproduce all of the plots in this paper as well as the \texttt{FITS}\footnote{\url{https://github.com/henrysky/astroNN_spectra_paper_figures/raw/master/astroNN_apogee_dr14_catalog.fits}} data file containing our neural network's predictions for 22 stellar parameters and abundances for the whole APOGEE DR14 is available at \url{https://github.com/henrysky/astroNN_spectra_paper_figures}.
\section{Bayesian Neural Networks with Drop-out}\label{sec:nn-general}
Deep learning refers to the usage of multi-layer (``deep'') ANNs to achieve both supervised and unsupervised machine learning. As opposed to task-specific algorithms, ANNs provide a general learning method that can be used on a variety of learning tasks. Bayesian neural network refers to the application of Bayesian inference to neural networks to obtain a posterior distribution function (PDF) on the weights that characterize the ANN given some input data. This PDF can then be used to propagate training uncertainty into predictions made with the ANN using new input data. Here, we use dropout variational inference as an approximation to Bayesian neural networks. Dropout variational inference is a new method proposed by \citet{2015arXiv150602142G} that can be applied to a wide variety of neural network architectures, is easy to implement, and is computationally cheap. This technique is previously used in \citet{2017ApJ...850L...7P} for strong gravitational lensing parameters estimation with uncertainty
In this section, we give a brief introduction to ANNs, describe the idea of dropout variational inference, and then discuss the loss function that we use to train our ANN using incomplete and noisy training data.
\subsection{Artificial Neural Networks}
ANNs were originally inspired by biological systems such as human brains, which consist of numerous neurons interconnected by synapses. The information or stimuli from the external world travel in the form of electrical impulses called action potentials. An ANN mimics this configuration and dynamics by representing a general learning task as a set of layers consisting of neurons that communicate through connections (the ``synapses''). The ``strength'' of each connection is given by a simple linear functional form $y = w\,x+b$, which is transformed by each neuron using a non-linear function. The final decision or output of the ANN depends on the input and on the strength of the connections (the weights $\vec{W}$). There is no need to pre-program any knowledge in an ANN, i.e., neural networks consist of random weights at the initial training stage. In order to achieve learning, error signals representing the agreement between the true output and the ANN output for (a subset of) the training set is back-propagated through the neural network and the connection strength of the synapses is adjusted to obtain better agreement between truth and prediction. Figure \ref{figure:ann} shows an example of a typical ANN.
Mathematically, an ANN is a real-valued, smooth function approximation to a general function. Consider data $\{\vec{x}, \vec{y}\}$ and a neural network $f$ parameterized by a set of parameters $\vec{W}$ that takes input $\vec{x}$ and maps it to $\hat{\vec{y}}=f^{\vec{W}}(\vec{x})$. Each neuron $i$ in an ANN takes an input vector $\vec{v}$ (either the actual input $\vec{x}$ or the output of a previous layer) and maps it to an output number $o$ as $o = \vec{w}_i \,\vec{v}+b_i$; this output number is then optionally transformed using a non-linear function before going to the next layer or the output. The parameters $\vec{W}$ in our notation represent the total set of $\{\vec{w}_i,b_i\}$ of all the neurons.
The quality of the ANN is represented by an objective function $J(\vec{y}, \hat{\vec{y}})$ that we want to minimize. At each step in the training process, a new set of parameters $\vec{W}_{\mathrm{new}}$ in ANN optimization can be obtained by descending along the gradient computed using back-propagation \citep{1986Natur.323..533R}
\begin{equation} \label{eq:grad_descent}
\vec{W}_\mathrm{new} = \vec{W} - \eta\frac{\partial{J}}{\partial{\vec{W}}}\,,
\end{equation}
where $\eta$ is the learning rate. For small $\eta$, this update step should lead to $J(\vec{y}, f^{\vec{W}_{\mathrm{new}}}(\vec{x}))$ that is smaller than $J(\vec{y}, f^\vec{W}(\vec{x}))$. To deal with large data sets, the gradients in each step are typically computed using only a small, random subset of the training data set that is different in each update step; this corresponds to a \emph{stochastic gradient descent} algorithm. In our work, we use a more sophisticated version of this type of optimizer, the ADAM optimizer \citep{2014arXiv1412.6980K}.
``Convolutional neural networks'' (CNNs) refers to the method of learning a set of convolution filters as part of the learning process. These convolution filters act as feature extractors, by convolving the input data (or the input data to a given layer in the ANN) with the filter. Useful features are extracted after these convolutional layers and, because they usually are not densely packed in the input data, we can apply a technique called max-pooling to reduce the size of our neural network, thus prevent overfitting. Max-pooling of size $n$ takes the maximum of $n$ pixels in non-overlapping subregions of incoming data, thus reducing the size of the input. Another advantage of max-pooling for spectroscopic analysis is that it induces a degree of translational invariance, making the analysis independent of small errors in the radial velocity correction.
\begin{figure}
\centering
\begin{tikzpicture}
\node[circle, draw, thick, label=left:{Spectrum Pixel}] (i1) {};
\node[circle, draw, thick, above=2em of i1, label=left:{Spectrum Pixel}] (i2) {};
\node[circle, draw, thick, above=2em of i2, label=left:{Spectrum Pixel}] (i3) {};
\node[circle, draw, thick, below=2em of i1, label=left:{Spectrum Pixel}] (i4) {};
\node[circle, draw, thick, below=2em of i4, label=left:{Spectrum Pixel}] (i5) {};
\node[circle, draw, thick, right=4em of i1] (h1) {};
\node[circle, draw, thick, right=4em of i2] (h2) {};
\node[circle, draw, thick, right=4em of i3] (h3) {};
\node[circle, draw, thick, right=4em of i4] (h4) {};
\node[circle, draw, thick, right=4em of i5, label=below:{Hidden}] (h5) {};
\node[circle, draw, thick, right=4em of h1] (hh1) {};
\node[circle, draw, thick, right=4em of h2] (hh2) {};
\node[circle, draw, thick, right=4em of h3] (hh3) {};
\node[circle, draw, thick, right=4em of h4] (hh4) {};
\node[circle, draw, thick, right=4em of h5, label=below:{Hidden}] (hh5) {};
\node[circle, draw, thick, right=4em of hh2, label=right:{Output}] (o1) {};
\node[circle, draw, thick, right=4em of hh4, label=right:{Output}] (o2) {};
\draw[-stealth, thick] (i1) -- (h1);
\draw[-stealth, thick] (i1) -- (h2);
\draw[-stealth, thick] (i1) -- (h3);
\draw[-stealth, thick] (i1) -- (h4);
\draw[-stealth, thick] (i1) -- (h5);
\draw[-stealth, thick] (i2) -- (h1);
\draw[-stealth, thick] (i2) -- (h2);
\draw[-stealth, thick] (i2) -- (h3);
\draw[-stealth, thick] (i2) -- (h4);
\draw[-stealth, thick] (i2) -- (h5);
\draw[-stealth, thick] (i3) -- (h1);
\draw[-stealth, thick] (i3) -- (h2);
\draw[-stealth, thick] (i3) -- (h3);
\draw[-stealth, thick] (i3) -- (h4);
\draw[-stealth, thick] (i3) -- (h5);
\draw[-stealth, thick] (i4) -- (h1);
\draw[-stealth, thick] (i4) -- (h2);
\draw[-stealth, thick] (i4) -- (h3);
\draw[-stealth, thick] (i4) -- (h4);
\draw[-stealth, thick] (i4) -- (h5);
\draw[-stealth, thick] (i5) -- (h1);
\draw[-stealth, thick] (i5) -- (h2);
\draw[-stealth, thick] (i5) -- (h3);
\draw[-stealth, thick] (i5) -- (h4);
\draw[-stealth, thick] (i5) -- (h5);
\draw[-stealth, thick] (h1) -- (hh1);
\draw[-stealth, thick] (h1) -- (hh2);
\draw[-stealth, thick] (h1) -- (hh3);
\draw[-stealth, thick] (h1) -- (hh4);
\draw[-stealth, thick] (h1) -- (hh5);
\draw[-stealth, thick] (h2) -- (hh1);
\draw[-stealth, thick] (h2) -- (hh2);
\draw[-stealth, thick] (h2) -- (hh3);
\draw[-stealth, thick] (h2) -- (hh4);
\draw[-stealth, thick] (h2) -- (hh5);
\draw[-stealth, thick] (h3) -- (hh1);
\draw[-stealth, thick] (h3) -- (hh2);
\draw[-stealth, thick] (h3) -- (hh3);
\draw[-stealth, thick] (h3) -- (hh4);
\draw[-stealth, thick] (h3) -- (hh5);
\draw[-stealth, thick] (h4) -- (hh1);
\draw[-stealth, thick] (h4) -- (hh2);
\draw[-stealth, thick] (h4) -- (hh3);
\draw[-stealth, thick] (h4) -- (hh4);
\draw[-stealth, thick] (h4) -- (hh5);
\draw[-stealth, thick] (h5) -- (hh1);
\draw[-stealth, thick] (h5) -- (hh2);
\draw[-stealth, thick] (h5) -- (hh3);
\draw[-stealth, thick] (h5) -- (hh4);
\draw[-stealth, thick] (h5) -- (hh5);
\draw[-stealth, thick] (hh1) -- (o1);
\draw[-stealth, thick] (hh1) -- (o2);
\draw[-stealth, thick] (hh2) -- (o1);
\draw[-stealth, thick] (hh2) -- (o2);
\draw[-stealth, thick] (hh3) -- (o1);
\draw[-stealth, thick] (hh3) -- (o2);
\draw[-stealth, thick] (hh4) -- (o1);
\draw[-stealth, thick] (hh4) -- (o2);
\draw[-stealth, thick] (hh5) -- (o1);
\draw[-stealth, thick] (hh5) -- (o2);
\end{tikzpicture}
\caption{A simple multilayer neural network.}
\label{figure:ann}
\centering
\end{figure}
\subsection{Dropout Variational Inference}
Variational inference is a general Bayesian inference method where the PDF is obtained not by sampling---as is the case when one uses Markov Chain Monte Carlo (MCMC) methods---but rather by fitting an approximation to the PDF using an objective function obtained by variational calculus. This method of Bayesian inference has the advantage that it can be applied to problems with large numbers of parameters, because optimization is in general faster and easier than sampling. This advantage comes at the expense of accuracy: the obtained PDF is only an approximation to the true PDF.
A Bayesian neural network with dropout variational inference works by approximating the true PDF for the weights ---which can number in the millions in our application below ---as a product of Bernoulli distributions. It can then be shown that a neural network trained with dropout applied to every layer except the last one and which has a Gaussian prior on the weights is an approximation to the full Bayesian neural network \citep{2015arXiv150602142G}. The Gaussian prior is achieved by imposing L2 regularization, parameterized by a regularization constant $\lambda$, on the loss function as
\begin{equation} \label{eq:l2_reg}
J_\mathrm{regularized}(\vec{y}, \hat{\vec{y}}) = J(\vec{y}, \hat{\vec{y}}) + \lambda\vec{W}^2\,.
\end{equation}
The hyper-parameter $\lambda$ is determined using the validation set: We set $\lambda$ to the value that optimizes the neural network precision on a validation set. L2 regularization is equivalent to a Gaussian prior in the Bayesian interpretation.
Dropout \citep{2012arXiv1207.0580H} is a technique that is primarily used to avoid over-fitting in neural networks, because deep neural networks usually have more parameters than data points. Dropout multiplies the hidden neurons---that is, those not in the input or output layer---by a Bernoulli distributed random variable that take the value 1 with a certain probability and 0 otherwise. When neurons are multiplied by zero they are effectively dropped; doing this during training prevents neurons from co-adapting to the training data, which would otherwise lead to overfitting. An example of a given instance of dropout for the example ANN in Figure \ref{figure:ann} is shown in Figure \ref{figure:ann_dropout}.
To use dropout for uncertainty estimation, we run $N$ times Monte Carlo dropout in forward passes through the network; in other words, we keep dropout turned on to make predictions using the ANN. Since dropout drops weights randomly, the neural network is probabilistic and has different predictions in every forward pass through the network. The mean value of predictions will be the final prediction and the standard deviation of predictions will be the model uncertainty. In addition to this ``model uncertainty'', the neural network that we use also gives a ``predictive uncertainty'' (see below for how we obtain the predictive uncertainty). The total uncertainty is the sum of model and predictive uncertainty in quadrature \citep{2017arXiv170304977K}.
In more mathematical terms, the Bayesian neural network predicts $\{\hat{\vec{y}}, \hat{\vec{\sigma}}^2\}=f^{\widehat{\vec{W}}}(x)$ where $\hat{\vec{y}}$ is the prediction of the labels, $\hat{\vec{\sigma}}^2$ is the predictive variance, $f^{\widehat{\vec{W}}}$ is the neural network with randomly masked weights due to dropout and $\vec{x}$ is the input data. We run the forward pass in the neural network $N$ times and obtain a set of $\{\hat{\vec{y}}_i, \hat{\vec{\sigma}}^2_i\}^N_{i=1}$. The final prediction $\hat{\vec{y}}$ and uncertainty intervals $\hat{\vec{\sigma}}$ is
\begin{equation} \label{eq:1sigma}
\hat{\vec{y}}\pm\hat{\vec{\sigma}} = \frac{1}{N}\sum^{N}_{i=1}\hat{\vec{y}}_i\pm\sqrt{\frac{1}{N}\sum^{N}_{i=1}\hat{\vec{y}}^2_i - \left(\frac{1}{N}\sum^{N}_{i=1}\hat{\vec{y}}_i \right)^2 + \frac{1}{N}\sum^{N}_{i=1}\hat{\vec{\sigma}}^2_i}\,.
\end{equation}
\begin{figure}
\centering
\begin{tikzpicture}
\node[circle, draw, thick, label=left:{Spectrum Pixel}] (i1) {};
\node[circle, draw, thick, above=2em of i1, label=left:{Spectrum Pixel}] (i2) {};
\node[circle, draw, thick, above=2em of i2, label=left:{Spectrum Pixel}] (i3) {};
\node[circle, draw, thick, below=2em of i1, label=left:{Spectrum Pixel}] (i4) {};
\node[circle, draw, thick, below=2em of i4, label=left:{Spectrum Pixel}] (i5) {};
\node[circle, draw, thick, red, fill=red!10, right=4em of i1] (h1) {};
\node[circle, draw, thick, right=4em of i2] (h2) {};
\node[circle, draw, thick, red, fill=red!10, right=4em of i3] (h3) {};
\node[circle, draw, thick, red, fill=red!10, right=4em of i4] (h4) {};
\node[circle, draw, thick, right=4em of i5, label=below:{Hidden}] (h5) {};
\node[red] (icr) at (h1) {$\mathlarger{\mathlarger{\mathlarger{\mathlarger{\mathlarger{\bm{\times}}}}}}$};
\node[red] (icr) at (h3) {$\mathlarger{\mathlarger{\mathlarger{\mathlarger{\mathlarger{\bm{\times}}}}}}$};
\node[red] (icr) at (h4) {$\mathlarger{\mathlarger{\mathlarger{\mathlarger{\mathlarger{\bm{\times}}}}}}$};
\node[circle, draw, thick, right=4em of h1] (hh1) {};
\node[circle, draw, thick, red, fill=red!10, right=4em of h2] (hh2) {};
\node[circle, draw, thick, right=4em of h3] (hh3) {};
\node[circle, draw, thick, red, fill=red!10, right=4em of h4] (hh4) {};
\node[circle, draw, thick, right=4em of h5, label=below:{Hidden}] (hh5) {};
\node[red] (icr) at (hh2) {$\mathlarger{\mathlarger{\mathlarger{\mathlarger{\mathlarger{\bm{\times}}}}}}$};
\node[red] (icr) at (hh4) {$\mathlarger{\mathlarger{\mathlarger{\mathlarger{\mathlarger{\bm{\times}}}}}}$};
\node[circle, draw, thick, right=4em of hh2, label=right:{Output}] (o1) {};
\node[circle, draw, thick, right=4em of hh4, label=right:{Output}] (o2) {};
\draw[-stealth, thick] (i1) -- (h2);
\draw[-stealth, thick] (i1) -- (h5);
\draw[-stealth, thick] (i2) -- (h2);
\draw[-stealth, thick] (i2) -- (h5);
\draw[-stealth, thick] (i3) -- (h2);
\draw[-stealth, thick] (i3) -- (h5);
\draw[-stealth, thick] (i4) -- (h2);
\draw[-stealth, thick] (i4) -- (h5);
\draw[-stealth, thick] (i5) -- (h2);
\draw[-stealth, thick] (i5) -- (h5);
\draw[-stealth, thick] (h2) -- (hh1);
\draw[-stealth, thick] (h2) -- (hh3);
\draw[-stealth, thick] (h2) -- (hh5);
\draw[-stealth, thick] (h5) -- (hh1);
\draw[-stealth, thick] (h5) -- (hh3);
\draw[-stealth, thick] (h5) -- (hh5);
\draw[-stealth, thick] (hh1) -- (o1);
\draw[-stealth, thick] (hh1) -- (o2);
\draw[-stealth, thick] (hh3) -- (o1);
\draw[-stealth, thick] (hh3) -- (o2);
\draw[-stealth, thick] (hh5) -- (o1);
\draw[-stealth, thick] (hh5) -- (o2);
\end{tikzpicture}
\caption{An example of dropout: a certain fraction of neurons are randomly dropped when evaluating the ANN. This is primarily used to prevent overfitting, but can also provide uncertainty estimation when it is used as an approximation to a Bayesian neural network.}
\label{figure:ann_dropout}
\centering
\end{figure}
\subsection{Objective function for incomplete and noisy training data} \label{subsec:objective}
The objective function $J(\vec{y}, \hat{\vec{y}})$ is the function that the neural network aims to minimize to train the network. Generally, neural networks for regression use the Mean Squared Error (MSE), defined as
\begin{equation} \label{eq:mse}
\text{Mean Squared Error} = J_{MSE}(\vec{y}, \hat{\vec{y}}) = \frac{1}{N} \sum^N_{i=1}(\hat{\vec{y}_i}-\vec{y}_i)^2\,.
\end{equation}
The MSE is prone to overfitting to outliers due to the squared term and it does not take uncertainty in the training labels into account.
However, astronomical data and observations are often incomplete and noisy. For example, in the case of spectroscopic data, the abundance of an element $\xh{X_1}$ may have only a single absorption line in the wavelength range by the detector. But due to reasons such as cosmic rays or if the line's Earth-frame wavelength happens to overlap with a strong sky emission line, the only absorption line $\xh{X_1}$ may not be measurable. But other abundances $\xh{X_2}$ can still be accurately measured. Previous data-driven approaches on spectroscopic data like the \texttt{Cannon 2} \citep{2016arXiv160303040C} or \texttt{StarNet} \citep{2018MNRAS.475.2978F} needed to filter such spectra from the training set because of the data incompleteness.
For the Bayesian neural network used in this work, we employ the following robust objective to get the loss for label $i$ and assume that unavailable data are labeled using \texttt{MAGIC NUM}
\begin{equation} \label{eq:mmse}
J(y_i, \hat{y}_i) =
\begin{cases}
\frac{1}{2} (\hat{y_i}-y_i)^2 e^{-s_i} + \frac{1}{2}(s_i) & \text{ for } y_i \neq \texttt{MAGIC NUM}\\
0 & \text{ for } y_i = \texttt{MAGIC NUM}
\end{cases}
\end{equation}
In this expression, $s_i = \ln \left[\sigma^2_{\mathrm{known}, i} + \sigma^2_{\mathrm{predictive}, i}\right]$, which corresponds to the natural logarithm of the sum of the known uncertainty variance in the labels and an additional predictive variance. This predictive variance is also learned by the neural network and forms another output from the neural network for each input. The known uncertainty variance is that returned by the reduction pipeline that produces the labels for the training subset. In general, the neural network can be trained to give the predictive variance without any known variance in the labels, which is a form of unsupervised training. The predictive variance from the loss function learned by the neural network represents any variance in the training set that cannot be explained by the known variance. This predictive variance contributes to the error budget for predictions on new data, see Equation \eqref{eq:1sigma}.
The final loss for the stochastic gradient descent is calculated from a mini-batch partition of the data consisting of $N$ data point and $D$ labels
\begin{equation} \label{eq:fcorrect}
J(\vec{y}, \hat{\vec{y}}) = \frac{1}{N} \sum_{i=1}^{N} \left( \frac{1}{D} \sum_{i=1}^{D} J(y_i, \hat{y}_i) \right)\mathcal{F}_{\mathrm{correction},i}\,,
\end{equation}
where $\mathcal{F}_{\mathrm{correction},i}$ is a correction term to correct for the fact that in Equation \eqref{eq:mmse} we effectively assume that the neural network made no error for missing data. If $D$ is the overall number of labels and $D_i$ is the number of labels not equal to \texttt{MAGIC NUM} for data point $i$, then
\begin{equation} \label{eq:fcorrect_itself}
\mathcal{F}_{\mathrm{correction},i} = \frac{D}{D_i}\,.
\end{equation}
The objective function in Equation \eqref{eq:mmse} acts as a robust version of the conventional MSE objective in Equation \eqref{eq:mse}, allowing the neural network to take the effect of uncertain and missing labels into account. This makes the model more robust because high uncertainty in a training label will have a smaller effect on the loss, preventing the neural network to learn from such labels. Equation \eqref{eq:mmse} also assumes that, without any other information, the prediction from the neural network is accurate and thus back-propagates zero loss for incomplete labels. The correction term $\mathcal{F}_{\mathrm{correction},i}$ in Equation \eqref{eq:fcorrect} is factored into the final loss in order to prevent the effective learning rate from decreasing due to the presence of missing labels. $\mathcal{F}_{\mathrm{correction},i}$ equals one in which there are no missing labels, in which case it resembles a conventional loss function.
\section{High-resolution spectroscopic data from APOGEE}\label{subsec:apogee-data}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{SNR.pdf}
\caption{Signal-to-noise (SNR) ratio distribution in the training set and in the two test sets used in this work. $\widetilde{SNR}$ represents median SNR. All spectra with ASPCAP reported SNR$>400$ are set to SNR$=400$, leading to the peak at high SNR. High SNR combined spectra test set refers to combined spectra with $100<\mathrm{SNR}<200$. Individual visits test set refers to the set of individual visits of stars in the high SNR combined-spectra test set.}
\label{figure:snr_traintest}
\centering
\end{figure}
The main source of data that we use are spectra and derived labels from the APO Galactic Evolution Experiment (APOGEE) Data Release 14 \citep{2018arXiv180709773H,2018arXiv180709784J,2018ApJS..235...42A}. APOGEE spectra are obtained with a 300-fibre spectrograph \citep{Wilson10a} attached to the Sloan Foundation 2.5m telescope at Apache Point Observatory \citep{2006AJ....131.2332G}. APOGEE is an infrared ($1.5 \mu m - 1.7 \mu m$), high resolution ($R\sim 22,500$), high signal-to-noise ratio (typical $\mathrm{SNR}>100$) spectroscopic survey. The APOGEE data set contains stellar parameter and chemical abundances obtained using the APOGEE Stellar Parameter and Chemical Abundances Pipeline (ASPCAP; \citealt{2016AJ....151..144G}). ASPCAP is an automated pipeline for determining the stellar labels from observed spectra by comparing observed spectra to a precomputed library of theoretical spectra using $\chi^2$ minimization. We describe how we select data from the overall APOGEE DR14 catalog and how we define our training and tests sets in Section \ref{subsec:selection}. In Section \ref{subsec:reduction} we discuss the method to process the data in the training and test sets.
\subsection{Training, test, and validation data selection from APOGEE DR14} \label{subsec:selection}
We have created one training and two test sets from the set of APOGEE DR14 spectra. Each data set consists of continuum normalized spectra, 22 ASPCAP labels (\ensuremath{T_\mathrm{eff}}, \ensuremath{\log g}, \xh{C}, \xh{CI}, \xh{N}, \xh{O}, \xh{Na}, \xh{Mg}, \xh{Al}, \xh{Si}, \xh{P}, \xh{S}, \xh{K}, \xh{Ca}, \xh{Ti}, \xh{TiII}, \xh{V}, \xh{Cr}, \xh{Mn}, \xh{Fe}, \xh{Co}, \xh{Ni})\footnote{\xh{CI} and \xh{TiII} are measurements of the carbon and titanium abundance using spectral regions that only have neutral atomic carbon and singly-ionized atomic titanium features, respectively. The overall \xh{C} and \xh{Ti} are mostly determined by molecular (for carbon) and neutral atomic (for titanium) features. Because ASPCAP returns these measurements separately, we include them also as part of our label set. The masking windows used can be accessed with the python function described in \url{https://astronn.readthedocs.io/en/v1.0.0/tools_apogee.html\#retrieve-aspcap-elements-window-mask}.} and their associated ASPCAP uncertainty. ASPCAP determines the abundances of all of these elements using synthetic spectra computed using the line list from \citet{2015ApJS..221...24S}. The SNR distributions of these subsets are shown in Figure \ref{figure:snr_traintest}. In Figure \ref{figure:train_tefflogg} we display the training sample in the space of \ensuremath{T_\mathrm{eff}}\ and \ensuremath{\log g}\ colored by \xh{Fe}. The training set consists of 33,407 spectra that have $\mathrm{SNR}>200$. At the start of training, $90\%$ of the training set is randomly selected to train the neural network---that is, used to compute the gradients of the objective function in the training steps---and the remaining $10\%$ constitute a separated validation set used to validate the performance of the neural network during the training process. $4.6\%$ of the combination of all training ASPCAP labels are $-9999$, the value used by ASPCAP to represent highly uncertain or unavailable labels; as discussed in Section \ref{subsec:objective}, these spectra are still used in our robust objective function (thus, \texttt{MAGIC NUM} $= -9999$ for APOGEE in Equation [\ref{eq:mmse}]). Two test sets are used, one consists of spectra with SNR between 100 to 200, which are called the high-SNR test set. These spectra are picked from the set of ``combined'' APOGEE spectra, which are the combinations of the individual exposures and it is these combinations that are used by ASPCAP for their analysis (most APOGEE spectra are obtained as a set of at least three individual hour-long exposures to obtain $\mathrm{SNR}>100$; \citealt{2015AJ....150..148H}). This set of spectra is entirely separate from the training set. The other test set consists of 81,483 spectra picked from the set of individual visit spectra that go into the combined spectra in the high-SNR test set. These spectra in the individual visit test set have much lower SNR than the spectra in the set of combined training spectra (which all have $\mathrm{SNR}>200$), see Figure \ref{figure:snr_traintest}. The advantage of this test set is that we are interested in the performance of our neural network for spectra with low SNR, but ASPCAP does not provide labels for individual-visit spectra with low SNR. However, we do have neural network or ASPCAP labels for the high-SNR combinations of the individual visit spectra, which we can use to test our method. Because the low SNR test set consists of the same \emph{stars} as the high SNR test set, the labels in the high SNR test set are representative of those in the low SNR test set, even though the \emph{noise} in them is not. Thus, the low SNR test set provides a stringent test of how this method performs at low SNR.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{logg_teff_fe.pdf}
\caption{$\ensuremath{\log g}$ vs $\ensuremath{T_\mathrm{eff}}$ colored by $\xh{Fe}$ abundances in the training set; all the labels in the training set are determined by ASPCAP. All main-sequence stars in APOGEE DR14 have $\ensuremath{\log g}$ set to \texttt{MAGIC NUM} $= -9999$ and are not displayed in the plot.}
\label{figure:train_tefflogg}
\centering
\end{figure}
On top of the SNR cut, we perform cuts on the values of the stellar parameters. This is necessary, because all of the knowledge learned by the neural network is solely driven by the training data. Therefore, we need to make sure that the training labels are as accurate as possible, because any systematic inaccuracy such as bias at lower SNR will be captured by the neural network and propagated to new test data. For this reason, we exclude spectra with surface temperature $\ensuremath{T_\mathrm{eff}}$ smaller than $4000\,\mathrm{K}$ or higher than $5500\,\mathrm{K}$, because at these temperatures ASPCAP may not be accurate \citep{2016AJ....151..144G}. Additionally, we remove spectra flagged with the \texttt{APOGEE\_ASPCAPFLAG} or \texttt{APOGEE\_STARFLAG} flags and spectra with a radial velocity scatter larger than $1\,\mathrm{km\,s}^{-1}$, because these represent potential issues with specific labels, issues with spectra, and potential binary stars, respectively. These cuts ensure the quality of the training and test set.
ASPCAP determines abundances by performing a $\chi^2$ fit on an interpolated, large grid of synthetic spectra computed for the APOGEE wavelength region \citep{2016AJ....151..144G}. After performing the fit, ASPCAP made several calibrations of the stellar parameters and abundances based on the consistency of abundances within open and globular clusters and by comparing to external data. The synthetic spectra are computed under various simplifying assumptions and using line lists that are not 100\,\% complete and this limits the quality of the synthetic grid used by ASPCAP and thus ultimately limits the Neural Network accuracy. Any systematic bias will propagated from ASPCAP to Neural Network during training. This is a disadvantage of the specific supervised machine-learning method that we are using here (we discuss advantages and disadvantages of our method compared to other methods in more detail in Sections \ref{subsec:assumption} and \ref{subsec:comparesynth}).
The resulting training set contains both main-sequence dwarfs and red giants. However, ASPCAP does not report calibrated values of \ensuremath{\log g}\ for main-sequence stars and these are all set to \texttt{MAGIC NUM} $=-9999$. These \ensuremath{\log g}\ labels are therefore ignored in our training procedure, although other labels such as \ensuremath{T_\mathrm{eff}}\ for main-sequence are used. This means that we cannot hope to determine good \ensuremath{\log g}\ for main-sequence stars with our neural network. We discuss this further below.
\subsection{Data reduction for training} \label{subsec:reduction}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{normalization_aspcap.pdf}
\caption{Example of continuum normalization: this Figure shows the continuum-normalized combined APOGEE spectrum of 2M19060637+4717296, showing the difference between the continuum-normalization performed by ASPCAP spectra and and that used in our method. The top three panels display the spectrum in three parts (corresponding to the three detectors used in the APOGEE instrument), while the bottom panels shows the difference between the two. The difference between the two normalization methods is a relatively-smooth function of wavelength and there is an overall offset due to the fact that ASPCAP does not attempt to trace the actual continuum, while our method does.}
\label{figure:norm_aspcap}
\centering
\end{figure}
All methods for spectroscopic analysis work with continuum-normalized spectra, because the information about stellar labels is mainly contained in narrow spectral features and the overall continuum is typically not well calibrated. The method used for continuum normalization is important, because we require a method that is invariant with respect to SNR due to the fact that we train the neural network on high SNR spectra only but also test it on low SNR spectra. To accomplish this, we use the same method as employed by the \texttt{Cannon 2} method \citep{2016arXiv160303040C}, where continuum normalization is performed by using a set of ``continuum pixels'' ---identified as pixels that depend little on the stellar labels using their data-driven model for APOGEE spectra---and fitting a continuum to the flux in only these pixels. The spectrum in each APOGEE detector is normalized separately in this method, with a 2-degree polynomial which has been demonstrated to be effective by \citet{2016arXiv160303040C}. While we use DR14 spectra, we use a set of continuum pixels obtained using DR12 spectra that is included in the \texttt{apogee} software package \citep{2016ApJ...817...49B}. Spectra in DR14 extend over a slightly wider wavelength range for each detector than those in DR12, which means that our continuum normalization does not perfectly capture the behavior of the continuum at the edges of the detector. But as we show below, the exact method of continuum normalization does not have a large effect on the performance of our method.
Figure \ref{figure:norm_aspcap} shows the difference between the continuum normalization used by ASPCAP and that described above for the combined spectrum of 2M19060637+4717296 as an example. An ideal continuum normalization would place continuum pixels to have a flux of 1. It is clear that the ASPACP normalization fails to do so (note that this is by design: the DR14 ASPCAP analysis explicitly does not attempt to do this; see \citealt{2018arXiv180709773H}). Aside from an overall offset, the continuum-normalized spectra are similar for both methods.
After continuum normalization, we check the APOGEE pixel-level mask bits in the \texttt{APOGEE\_PIXMASK} bitmask and set the flux value of pixels that contain the following bits to $1$ (the expected continuum): \textbf{0}: bad pixel, \textbf{1}: cosmic ray, \textbf{2}: saturated, \textbf{3}: unfixable, \textbf{4}: bad from dark, \textbf{5}: bad from flat, \textbf{6}: high error, \textbf{7}: no sky info, \textbf{12}: overlaps a significant sky line.
Besides the continuum normalization of the spectra, we need to standardize the labels and continuum normalized spectra to be able to easily use them with standard neural-network methods. For the labels, we subtract the mean and divide by the standard deviation such that the training labels all approximately have a mean of 0 and a standard deviation of 1. In our case we have 22 labels, and therefore we calculate 22 means and 22 standard deviations to standardize the labels. Labels which are \texttt{MAGIC NUM} $=-9999$ (that is, missing) are not involved in the normalization process, i.e. $-9999$ values remain constant such that objective function in Equation \eqref{eq:mmse} can recognize these missing labels during optimization. For continuum-normalized spectra, we calculate the mean of the flux pixel-by-pixel for the spectra in training set, and subtract the means from all the spectra in the training set. In other words ideally an average spectrum should be a flat straight line at $\mathrm{flux}=0$ after this final normalization step. The reason behind this normalization is that this maps the average spectrum to the average label for a neural network with all weights set to zero and we expect that the average spectrum should have stellar labels close to the average of those in the training set. Thus, it will be easier and faster for the neural network to converge to a minimum.
While performing variational inference on test sets, it is important to use the same normalization procedure and the same set of means and standard deviations as used for the training set to normalize and denormalize all the data. This is also the reason why we do not scale the training spectra to have standard deviation of $1$, because the training and testing spectra have completely different SNR. In other words, training set spectra have high SNR, thus low overall standard deviation and test set spectra have low SNR, thus higher overall standard deviation. Using the parameters to standardize training spectra will not standardize testing spectra so we chose not to scale any spectra.
\section{Performance on APOGEE data} \label{sec:main_nn}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth, clip]{NN_diagram.pdf}
\caption{The neural-network architecture mainly used in this work, defined as \texttt{ApogeeBCNNCensored()} in \texttt{astroNN}.}
\label{figure:nn_flow}
\centering
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{logg_teff_fe_panel_2only.pdf}
\caption{Neural Network $\ensuremath{T_\mathrm{eff}}$ and $\ensuremath{\log g}$ prediction color-coded by $\xh{Fe}$ (left panel) and $\ensuremath{\log g}$ uncertainties (right panel). The group at high $\ensuremath{\log g}$ is due to 2,611 spectra in the high-SNR test set that are dwarfs with \texttt{MAGIC NUM} $=-9999$ ASPCAP value. \ensuremath{\log g}\ values for these stars are absent in the training data, because ASPCAP in DR14 does not provide the $\ensuremath{\log g}$ for dwarfs. The neural network clearly predicts the wrong values for \ensuremath{\log g}\, but this is also reflected in the large uncertainties for these stars in the right panel. For test set objects along the giant branch, the neural network returns reasonable parameters (compare to Figure~\ref{figure:train_tefflogg}). Similarly, the uncertainty in \ensuremath{\log g}\ is high for low-metallicity, low \ensuremath{\log g}\ giants, for which training data are sparse.}
\label{figure:logg_teff_isochrones}
\centering
\end{figure*}
The main neural network that we train and test with APOGEE spectra in this work is the network \verb|ApogeeBCNNCensored()| in \texttt{astroNN} (see Appendix~\ref{appendix:graph:astroNN}). The neural network architecture is shown in Figure \ref{figure:nn_flow}. Users only have to provide a continuum-normalized spectrum and are returned a prediction and associated uncertainty. Rather than using a single, simple neural network to predict stellar parameters and elemental abundances from an input spectrum, we use a combination of (a) a large neural network trained on the full spectral wavelength range to predict [$\ensuremath{T_\mathrm{eff}}$, $\ensuremath{\log g}$, $\xh{Fe}$] (the big gray-colored network on the right side in Figure \ref{figure:nn_flow}) and (b) 19 mini neural networks used to predict the 19 \xh{X} abundances based on fixed regions of the spectrum that contain known spectral features for each element and the overall [$\ensuremath{T_\mathrm{eff}}$, $\ensuremath{\log g}$, $\xh{Fe}$]. This architecture mimics that of traditional spectroscopic analysis and of ASPCAP, where the overall stellar parameters ($\ensuremath{T_\mathrm{eff}}, \ensuremath{\log g}, \xh{Fe}$, etc.) are determined first and individual abundances are determined afterwards from specific spectral features. However, our method differs from this in a crucial aspect: the network trained to predict [$\ensuremath{T_\mathrm{eff}}$, $\ensuremath{\log g}$, $\xh{Fe}$] from the full spectrum also has a two-neuron connection characterized by two latent variables to the 19 mini-networks used to predict the individual abundances (in addition to feeding [$\ensuremath{T_\mathrm{eff}}$, $\ensuremath{\log g}$, $\xh{Fe}$] to the mini-networks as well). We choose to use two neurons as the connection to mimic ASPCAP, in which \xfe{C} and $\xfe{\alpha}$ are fitted to the full spectrum, because these elements strongly affect the stellar photosphere and thus the formation of all spectral lines. By using two neurons we can learn a latent space similar to these two elements, but we do not require the latent variables to exactly correspond to \xfe{C} and $\xfe{\alpha}$ to give the network the opportunity to learn a better low-dimensional set of latent variables. This allows the mini-networks to use a limited amount of information from the full spectrum that is not captured by [$\ensuremath{T_\mathrm{eff}}$, $\ensuremath{\log g}$, $\xh{Fe}$] in making their predictions. To produce the 19 masked spectra for the 19 mini-networks (one mask per element) we use the windows employed by ASPCAP DR14 to determine individual abundances \citep{2016AJ....151..144G}. These windows were derived by the ASPCAP team using synthetic spectral syntheses to isolate regions of the spectra most sensitive to individual elements. Unlike ASPCAP, we only use the windows as a binary mask---pixels are either in or out---we do not use the weights assigned to pixels within the windows.
The entire combination of the large ANN to predict [$\ensuremath{T_\mathrm{eff}}$, $\ensuremath{\log g}$, $\xh{Fe}$] and the 19 mini-networks, including their two-neuron connection, is trained simultaneously. To achieve this, two unconventional layers are included in the network, \texttt{StopGrad} and \texttt{MaxNorm}. \texttt{StopGrad} is an identity transformation layer with the property that its gradient is always set to 0 during training, but otherwise is simply 1 (for example, when computing the sensitivity of the network output to input; see Section \ref{subsec:jac}). This layer prevents the error from an individual abundance $\xh{X}$ to be back-propagated during training to the network predicting [$\ensuremath{T_\mathrm{eff}}$, $\ensuremath{\log g}$, $\xh{Fe}$]. That is, we do not allow prediction errors in $\xh{X}$ to affect the training of the network that predicts the stellar parameters. \texttt{MaxNorm} is a weight-constraint layer that requires $\sqrt{\sum{w^2}} \leq \delta$, where $w$ are the weights and $\delta$ is a constraint constant. $\delta$ is determined using the validation data set, such that $\delta$ optimizes the performance of neural network on the validation set. This layer prevents the mini neural networks that predict $\xh{X}$ from paying too much attention to the full spectrum. The reason why we construct this rather complex network rather than the more straightforward single network that predicts all the labels from the full spectrum is discussed in detail in Section \ref{subsec:full_spec}. Briefly, the reason is that when allowing the full spectrum to inform individual abundances, the ANN predictions in regions of label-space with few training data become highly correlated due to correlations in the training data.
We discuss the performance of the \verb|ApogeeBCNNCensored()| network in detail in the following subsections. We will see that we find that the neural network trained on high SNR training data and tested on high SNR testing data displays a fairly high bias, which may result from errors in ASPCAP, but a small amount of scattering. When testing on low SNR individual-visit spectra, the network shows almost no bias and small amounts of scattering (but larger than that at higher SNR). The results are considerably better than any previous work on applying machine-learning techniques to high-resolution spectral analysis. The uncertainty estimation from dropout variational inference that we find makes sense, because it correlates with $1 / \sqrt{SNR}$ and is similar to the scatter in the residuals between our method and ASPCAP. For both test sets, we find that the NN performs the best for elements that ASPCAP reports as being their most accurate elements (for example \xh{Mg} and \xh{Ni} in an independent validation of ASPCAP DR13/14 by \citet{2018arXiv180709784J}). A sensitivity analysis of how the ANN outputs depend on the input spectra shows that the model depends in a reasonable manner on wavelength. All our results can be reproduced using our online code (see Appendix~\ref{appendix:graph:astroNN}), but due to the stochastic nature of dropout and the neural-network training process, it is impossible to reproduce the exact same results. However, statistically, the results should be very close to those described in this work.
In the following, we define the residual as
\begin{equation} \label{eq:residual}
\mathrm{Residual} = \textit{NN } \mathrm{Prediction}-\mathrm{ASPCAP}\,,
\end{equation}
where $NN$ refers to the neural network. As a robust measurement of the scatter, we use a measure based on the Median Absolute Deviation (MAD): $\ensuremath{\sigma^{\mathrm{MAD}}} = 1.4826\,\mathrm{MAD}$, where the factor is such that for a Gaussian distribution $\ensuremath{\sigma^{\mathrm{MAD}}}$ equals the Gaussian standard deviation. Thus, for a set of residual $R:[R_1, R_2,...,R_n]$, \ensuremath{\sigma^{\mathrm{MAD}}}\ is
\begin{equation} \label{eq:scattering}
\ensuremath{\sigma^{\mathrm{MAD}}} = 1.4826\,\mathrm{median}\left(|R_i - \mathrm{median}(R)|\right)\,.
\end{equation}
In all of these calculations, ASPCAP labels which are equal to $-9999$ (or, more generally, labels equal to \texttt{MAGIC NUM}) are excluded from the calculation.
\subsection{Comparison to ASPCAP at high signal-to-noise ratio}\label{subsec:highsnr}
\begin{table}
\centering
\caption{Neural-network prediction result on the high SNR test set from comparing NN predictions to ASPCAP}
\label{table:highsnr_result}
\begin{tabular}{lrr}
\hline
Label & Median of residual & \ensuremath{\sigma^{\mathrm{MAD}}}\ of residual\\
\hline
$\ensuremath{T_\mathrm{eff}}$ & $-20 \text{ K}$ & $30 \text{ K}$ \\
$\ensuremath{\log g}$ & $0.012 \text{ dex}$ & $0.051 \text{ dex}$ \\
$\xh{C}$ & $0.003 \text{ dex}$ & $0.040 \text{ dex}$ \\
$\xh{CI}$ & $0.013 \text{ dex}$ & $0.058 \text{ dex}$ \\
$\xh{N}$ & $-0.004 \text{ dex}$ & $0.041 \text{ dex}$ \\
$\xh{O}$ & $-0.021 \text{ dex}$ & $0.046 \text{ dex}$ \\
$\xh{Na}$ & $-0.01 \text{ dex}$ & $0.16 \text{ dex}$ \\
$\xh{Mg}$ & $0.000 \text{ dex}$ & $0.027 \text{ dex}$ \\
$\xh{Al}$ & $-0.038 \text{ dex}$ & $0.071 \text{ dex}$ \\
$\xh{Si}$ & $0.000 \text{ dex}$ & $0.029 \text{ dex}$ \\
$\xh{P}$ & $-0.02 \text{ dex}$ & $0.10 \text{ dex}$ \\
$\xh{S}$ & $0.006 \text{ dex}$ & $0.051 \text{ dex}$ \\
$\xh{K}$ & $-0.013 \text{ dex}$ & $0.049 \text{ dex}$ \\
$\xh{Ca}$ & $-0.015 \text{ dex}$ & $0.033 \text{ dex}$ \\
$\xh{Ti}$ & $-0.029 \text{ dex}$ & $0.052 \text{ dex}$ \\
$\xh{TiII}$ & $0.06 \text{ dex}$ & $0.17 \text{ dex}$ \\
$\xh{V}$ & $-0.009 \text{ dex}$ & $0.097 \text{ dex}$ \\
$\xh{Cr}$ & $-0.002 \text{ dex}$ & $0.048 \text{ dex}$ \\
$\xh{Mn}$ & $-0.018 \text{ dex}$ & $0.038 \text{ dex}$ \\
$\xh{Fe}$ & $-0.004 \text{ dex}$ & $0.020 \text{ dex}$ \\
$\xh{Co}$ & $-0.02 \text{ dex}$ & $0.14 \text{ dex}$ \\
$\xh{Ni}$ & $0.003 \text{ dex}$ & $0.029 \text{ dex}$ \\
\hline
\end{tabular}
\end{table}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{lowSNR_comparison.pdf}
\caption{Median offset between our NN method, ASPCAP, and the \texttt{Cannon 2}. Each panel shows the Median Absolute Error (MAE) of ASPCAP or the \texttt{Cannon 2} assuming that the NN is the ground truth in bins of SNR for the combined APOGEE spectra. The panels show the MAE of \ensuremath{T_\mathrm{eff}}, \ensuremath{\log g}, and \xh{Fe}, as well as \xh{Al} as a representative element. The MAE bias between the NN and ASPCAP has a strong SNR dependence for \ensuremath{T_\mathrm{eff}}\ and somewhat less strong for \ensuremath{\log g}. This trend is much smaller between the NN and the \texttt{Cannon 2}, especially for \xh{Al}. Typically, the bias between ASPCAP and the \texttt{Cannon 2} (the difference between the blue and orange curves) is higher than that between either of them and the NN. The strong trend with SNR when comparing the NN to the \texttt{Cannon 2} or to ASPCAP may be due to a SNR dependent bias in ASPCAP.}
\label{figure:lowSNR_issue}
\centering
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{teff_logg_xh_allin1_highSNR.pdf}
\caption{Comparison between neural-network predictions for $\ensuremath{T_\mathrm{eff}}$, $\ensuremath{\log g}$, and $\xh{X}$ and those from ASPCAP at high SNR (SNR between 100 and 200). The blue curve shows the MAE between the NN and ASPCAP in bins of the ASPCAP label, while the green and orange error bars give the NN model and total uncertainty. Overall the MAE is small and the uncertainties are similar to the MAE, but there are bigger residuals at low \xh{X}, because the training set contains few low-metallicity stars and spectral features are weaker for such stars, leading to worse performance of the NN.}
\label{figure:all_you_can_residue_highSNR}
\centering
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{delta_xh_teff_highSNR.pdf}
\caption{Comparison between neural-network predictions for $\xh{X}$ and the ASPCAP labels for the high SNR test set as a function of \ensuremath{T_\mathrm{eff}}. Curves and errorbars are as in Figure \ref{figure:all_you_can_residue_highSNR}. Spectral features are weaker at higher \ensuremath{T_\mathrm{eff}}\ leading to an increase in the typical residuals that is matched by an increase in the NN uncertainty.}
\label{figure:delta_xh_teff_highSNR}
\centering
\end{figure*}
We first test the performance of the neural network for high SNR spectra, which we define here as spectra with SNR between $100$ to $200$ (the standard cuts described in Section \ref{subsec:selection} apply as well). In this section, we evaluate the performance of the NN by comparing it to the predictions from ASPCAP for the same stars. Thus, any comparison is affected by biases in the ASPCAP results themselves and we will see that there are reasons to believe that ASPCAP suffers from SNR-dependent biases, even at these high SNRs. These biases inflate the size of the residuals between the neural-network predictions and ASPCAP. We also do not account for the random uncertainties in the ASPCAP predictions, which also contribute to the size of the residuals. In this sense, the biases and errors derived from comparing to ASPCAP from this section are an upper limit on the size of the biases and errors of the neural network's predictions. In the next section, we will compare the neural-network's prediction against itself and find much smaller residuals.
A summary of the results is given in Table \ref{table:highsnr_result}. Figure \ref{figure:logg_teff_isochrones} displays the predicted $\ensuremath{\log g}$ versus $\ensuremath{T_\mathrm{eff}}$. In this figure, the left panel is color-coded by $\xh{Fe}$ and we overlay PARSEC isochrones \citep{2012MNRAS.427..127B} at 4 different metallicities to indicate the expected location of stars in this space. It is clear that the predicted parameters for stars at different metallicities conform well to these expectations. That the predicted \ensuremath{\log g}\ is highly precise for giants is clear from the narrowness of the red clump. The right panel of Figure \ref{figure:logg_teff_isochrones} shows the same predicted $\ensuremath{\log g}$ versus $\ensuremath{T_\mathrm{eff}}$, but now color-coded by the neural-network uncertainty in \ensuremath{\log g}. The neural network is highly confident in its \ensuremath{\log g}\ prediction along the well-populated parts of the giant branch. However, the uncertainty in \ensuremath{\log g}\ is large for the group of stars at high \ensuremath{\log g}. This is reasonable, because these are main-sequence dwarfs for which we have no \ensuremath{\log g}\ training data (see discussion in Section \ref{subsec:selection} above). The predicted \ensuremath{\log g}\ are clearly wrong, but this is reflected in the high uncertainties for these stars. Similarly, the uncertainty in \ensuremath{\log g}\ is high for low-metallicity, low-\ensuremath{\log g}\ giants, for which training data are sparse (see Figure~\ref{figure:train_tefflogg}).
The results in Table \ref{table:highsnr_result} show that the neural network prediction displays a relatively high bias in all labels when compared to ASPCAP, especially in $\ensuremath{T_\mathrm{eff}}$. It is possible that the majority of this bias comes from the ASPCAP prediction on lower SNR spectra instead of the neural network. The neural network's predictions are consistent across a wide range of SNR, while the ASPCAP $\ensuremath{T_\mathrm{eff}}$ is probably biased at SNR$<200$. Figure \ref{figure:lowSNR_issue} shows the Median Absolute Error (MAE) of ASPCAP$-$NN and \texttt{Cannon}$-$NN between SNR 50 to 300 for the entire APOGEE DR14 catalog with good \texttt{APOGEE\_ASPCAPFLAG} and \texttt{APOGEE\_STARFLAG} flags, velocity scatter smaller than $1\,\mathrm{km\,s}^{-1}$, and ASPCAP $4000< \ensuremath{T_\mathrm{eff}} <5500$. Most labels show strong SNR-dependent trends that are above the empirical precision found for the neural network (see discussion below and Figure \ref{figure:snr_acc}). The bias with respect to ASPCAP is generally larger than that compared to the \texttt{Cannon 2}. These SNR dependent trends demonstrate that both ASPCAP and the \texttt{Cannon 2} are probably biased to a larger degree and at higher SNR than previously thought.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{teff_logg_xh_SNR_allin1.pdf}
\caption{Comparison between neural-network predictions for $\ensuremath{T_\mathrm{eff}}$, $\ensuremath{\log g}$, and $\xh{X}$ from low-SNR, individual-exposure spectra and those from the same neural network applied to the combined spectra in high-SNR test set. Curves and errorbars are as in Figure \ref{figure:all_you_can_residue_highSNR}. The number beside or below the parameter/abundance label in each plot represents the MAE at SNR$=50$. When compared to its own predictions at high SNR, the performance of the neural network at low SNR is excellent, with residuals of size $0.01 \text{ to } 0.02\,\mathrm{dex}$ at SNR$\gtrsim100$ and residuals that are only slightly bigger even at SNR$\approx50$.}
\label{figure:snr_acc}
\centering
\end{figure*}
Figure \ref{figure:all_you_can_residue_highSNR} shows more detailed results on the performance of the neural network. In this figure, the blue line represents the median absolute error between the neural network and ASPCAP in bins of ASPCAP labels, the orange error bars represent the median total NN uncertainty in these bins, and the green line gives the contributions of the NN model uncertainty to the total uncertainty. The difference between the orange and green error bar therefore gives a sense of the contribution of the predictive uncertainty (the uncertainties are added in quadrature, so the predictive uncertainty is not simply the difference between the orange and green uncertainties). Both the accuracy and the precision are good for $\xh{X}\gtrsim-0.7$, especially around solar metallicity where there is the most training data. The neural network has lower accuracy at low $\xh{X}$ mainly due to two reasons: first, there are not much training data at low metallicity because $\xh{Fe}\approx -0.7$ is the lower bound of typical abundances in the Galactic disk and second, spectral features at low metallicity are less defined and therefore less informative about the abundances. The \ensuremath{\log g}\ prediction is accurate and precise in the well-populated lower and mid regions of the giant branch, but becomes more uncertain for very luminous, low-\ensuremath{\log g}\ giants. By and large, the uncertainties returned by the NN are similar to the typical MAE difference between the NN and ASPCAP. For some individual elements, the uncertainties are smaller than the typical MAE difference at low metallicity, but it is always the case that these uncertainties are quite large ($\gtrsim 0.25\,\mathrm{dex}$), thus signaling that the NN prediction is noisy.
Figure \ref{figure:delta_xh_teff_highSNR} similarly shows how the accuracy and uncertainties in the individual abundances depend on the surface temperature. Spectral features in stars with higher surface temperatures are weaker, so we expect the accuracy to decrease and the uncertainties to increase. Figure \ref{figure:delta_xh_teff_highSNR} demonstrates that this is indeed the case, although in general the trend with temperature is quite weak.
\subsection{Results at low signal-to-noise ratio} \label{subsec:lowsnr}
\begin{table}
\centering
\caption{Neural-network prediction result on the individual visits of low SNR test set spectra from comparing results on the combined spectra of the same test set}
\label{table:indi_result}
\begin{tabular}{lrr}
\hline
Label & Median of residual & \ensuremath{\sigma^{\mathrm{MAD}}}\ of residual\\
\hline
$\ensuremath{T_\mathrm{eff}}$ & $-3 \text{ K}$ & $29 \text{ K}$ \\
$\ensuremath{\log g}$ & $0.000 \text{ dex}$ & $0.047 \text{ dex}$ \\
$\xh{C}$ & $-0.004 \text{ dex}$ & $0.045 \text{ dex}$ \\
$\xh{CI}$ & $-0.003 \text{ dex}$ & $0.054 \text{ dex}$ \\
$\xh{N}$ & $-0.002 \text{ dex}$ & $0.046 \text{ dex}$ \\
$\xh{O}$ & $0.000 \text{ dex}$ & $0.029 \text{ dex}$ \\
$\xh{Na}$ & $-0.013 \text{ dex}$ & $0.060 \text{ dex}$ \\
$\xh{Mg}$ & $-0.001 \text{ dex}$ & $0.023 \text{ dex}$ \\
$\xh{Al}$ & $-0.005 \text{ dex}$ & $0.045 \text{ dex}$ \\
$\xh{Si}$ & $-0.001 \text{ dex}$ & $0.024 \text{ dex}$ \\
$\xh{P}$ & $0.00 \text{ dex}$ & $0.10 \text{ dex}$ \\
$\xh{S}$ & $0.002 \text{ dex}$ & $0.054 \text{ dex}$ \\
$\xh{K}$ & $-0.003 \text{ dex}$ & $0.030 \text{ dex}$ \\
$\xh{Ca}$ & $-0.002 \text{ dex}$ & $0.021 \text{ dex}$ \\
$\xh{Ti}$ & $-0.003 \text{ dex}$ & $0.026 \text{ dex}$ \\
$\xh{TiII}$ & $-0.001 \text{ dex}$ & $0.043 \text{ dex}$ \\
$\xh{V}$ & $-0.003 \text{ dex}$ & $0.052 \text{ dex}$ \\
$\xh{Cr}$ & $-0.006 \text{ dex}$ & $0.029 \text{ dex}$ \\
$\xh{Mn}$ & $-0.005 \text{ dex}$ & $0.030 \text{ dex}$ \\
$\xh{Fe}$ & $-0.004 \text{ dex}$ & $0.020 \text{ dex}$ \\
$\xh{Co}$ & $-0.010 \text{ dex}$ & $0.086 \text{ dex}$ \\
$\xh{Ni}$ & $-0.005 \text{ dex}$ & $0.024 \text{ dex}$ \\
\hline
\end{tabular}
\end{table}
To test the neural network at lower SNR, we make use of the set of individual-exposures for the stars in the high-SNR test set, as explained in detail in Section \ref{subsec:selection}. For the combined spectra in the high-SNR test set, we have results from the neural network and we can therefore test the performance of the neural network on the low SNR individual exposures by comparing the predictions from the low SNR spectra to the results from the NN on the high SNR combined spectra.
The results from comparing the NN predictions on low SNR spectra to the NN measurements from their counterpart combined spectra are shown in Table \ref{table:indi_result}. The first thing to note is that the bias (median of the residual) with respect to the NN is much smaller than it was in the high SNR comparison in Table \ref{table:highsnr_result}. The main reason for this is that ASPCAP parameters and abundances are accurate at SNR$>200$ and the predictions from the neural network using the individual exposures are essentially the same parameters as the prediction from their high SNR counterpart. This demonstrates that the neural network's predictions are robust at low SNR.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{open_clusters.pdf}
\caption{Spread in the predicted NN and ASPCAP labels for stars in the open clusters M67 (14 stars) and NGC6819 (20 stars). The spread in the NN predictions is small for all elements and well matched by the NN uncertainties (green errorbars). The NN has a smaller spread in each element than ASPCAP. The overall level of chemical homogeneity is $0.030\pm0.029\,\mathrm{dex}$.}
\label{figure:open_clusters}
\centering
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{m13.pdf}
\caption{Spread in the predicted NN and ASPCAP labels for stars in the globular cluster M13 (23 stars). As expected, the spread in the abundances of heavier elements is small, while that in light elements is larger. This is especially the case for Al, which has a $\approx0.5\,\mathrm{dex}$ spread in M13 \citep{2015AJ....149..153M} that is well recovered by the NN predictions.}
\label{figure:m13}
\centering
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{jac_Mg_h.pdf}
\caption{The neural network's $\xh{Mg}$ sensitivity, that is mean $\frac{\partial\xh{Mg}}{\partial\lambda_j}$, for metal-poor ($\xh{Fe}<-1.5$; blue curve) and metal-rich stars ($\xh{Fe}>0.4$; orange curve). The green regions show the ASPCAP \xh{Mg} windows. We also label the hydrogen lines and a strong FeI feature. Because of the way the network is structured, information about \xh{Mg} in the spectrum is mainly extracted in the ASPCAP windows, but especially for metal-poor stars the network also depends on features outside of the windows.}
\label{figure:jacobian}
\centering
\end{figure*}
Figure \ref{figure:snr_acc} shows the self-consistency of the neural network at low SNR in more detail. Here, the median absolute error is calculated between the predicted NN label for the individual exposure and the predicted NN label from the high-SNR combined counterpart. All abundances show significant errors at SNR$<30$. This is expected, because the noisy low-SNR spectra limit the ability of the neural network to get information from spectral features. However, the neural network still performs well at SNR$\approx 50$, with most abundances measured to a few hundredths of a dex. The performance of the neural network at SNR$\approx 50$ is better than that of the \texttt{Cannon 2} \citep{2016arXiv160303040C}, even though we only use regions of the spectra with known spectral features for the element of interest, while the \texttt{Cannon 2} uses the full spectrum for each element. The exact precision at SNR$\approx 50$ is shown in each panel, with the best performance of $0.014\,\mathrm{dex}$ for $\xh{Fe}$ and $\xh{Ca}$ to the worst performance of $0.068\,\mathrm{dex}$ for $\xh{P}$. Note that we measure \emph{each} element's abundance to better than $0.10\,\mathrm{dex}$, often considered the target uncertainty in large spectroscopic surveys, at SNR$\approx 50$. Of course, at lower metallicities the performance is worse, similar to what is seen in Figure \ref{figure:all_you_can_residue_highSNR}.
At high SNR, the residual in the predictions for most labels flattens out to $0.01 \text{ to } 0.02\,\mathrm{dex}$. This demonstrates that for most of the abundances, the neural network can reach a precision of $\approx 0.01\,\mathrm{dex}$ for high SNR spectra.
\subsection{Results on open and globular clusters}\label{subsec:clusters}
To further test the performance of the neural network's predictions for the individual elemental abundances, we apply it to open and globular clusters. Open clusters provide a good testbed for data-driven abundance analyses, because they are a chemically homogeneous population of stars, at least at the level of current abundance precision \citep[e.g.,][]{2006AJ....131..455D,2007AJ....133.1161D,2016ApJ...817...49B,2018ApJ...853..198N}. Ideally, the neural-network predictions for the abundances of stars in open clusters should exhibit a very small spread. Similarly, globular clusters should be homogeneous in all but their lightest elements. For the light elements we expect to see spreads and anti-correlations between pairs of elements (e.g., Mg and Al; \citealt{2015AJ....149..153M}).
To perform these tests, we select stars from two open clusters that are well populated in the APOGEE database: M67 and NGC6819. We select members from the catalog provided by \citet{2013AJ....146..133M} and we exclude spectra with the \texttt{APOGEE\_STARFLAG} bitmask set. The resulting spreads in the abundances of each element in each cluster are shown in Figure \ref{figure:open_clusters}. The overall level of chemical homogeneity in these two clusters found by using the NN predictions is $0.030 \pm 0.029\,\mathrm{dex}$, obtained by calculating the mean abundance scatter and the mean uncertainty in the scatter. This is the same level of homogeneity as that reported by using the \texttt{Cannon 2} \citep{2018ApJ...853..198N} and \texttt{The Payne} \citep{2018arXiv180401530T} by benchmarking on open clusters.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{jac_logg_fe.pdf}
\caption{The neural network's $\ensuremath{\log g}$ sensitivity, that is mean $\frac{\partial\ensuremath{\log g}}{\partial\lambda_j}$, for metal-poor ($\xh{Fe}<-1.5$; blue curve) and metal-rich stars ($\xh{Fe}>0.4$; orange curve). We also label the hydrogen lines and a strong FeI feature. Because of the way the network is structured, information about $\ensuremath{\log g}$ in the spectrum is mainly extracted in the hydrogen lines.}
\label{figure:jacobian_logg}
\centering
\end{figure*}
To test the neural network's performance at low metallicity, we use the globular cluster M13, which has many members in the APOGEE catalog. The spread in the abundances from the NN and from ASPCAP for M13 is displayed in Figure \ref{figure:m13}. This figure shows the expected behavior: the spread in the abundances of heavier elements is small, while that in the abundance of lighter elements is larger, especially for Al. A boutique analysis of the APOGEE spectra for stars in M13 by \citet{2015AJ....149..153M} showed that the spread in Al in M13 is particularly large: $\approx0.5\,\mathrm{dex}$. This is similar to the spread in the NN Al abundances in M13, while that in ASPCAP is significantly larger. In Figure \ref{figure:m13_al_mg} displayed in Section \ref{subsec:full_spec} below, we show the abundances of both Mg and Al for stars in M13 and these fall roughly along the expected sequence (which in M13 is almost vertical, that is, there is very little Mg spread in M13; \citealt{2015AJ....149..153M}).
\subsection{Sensitivity analysis}\label{subsec:jac}
In order to better understand what the neural network is doing, we can compute the sensitivity of each label to the input spectrum as this provides a glimpse into how the neural network makes its predictions and which regions in wavelength space are crucial for the neural network to predict each label. In mathematical terms, every neural network maps inputs $x$ to outputs $y$ in a differentiable manner (indeed, this differentiability is crucial in allowing the neural network to be optimized by gradient descent). In practice, neural-network frameworks allow this gradient to be computed analytically by making use of automatic differentiation. This procedure can be applied to compute the derivatives $\frac{\partial\text{Label}}{\partial\lambda}$ of each label with respect to every wavelength pixel $\lambda$ of the input spectrum. This derivative represents the sensitivity of the neural network to each pixel for every label. A negative $\frac{\partial\text{Label}_i}{\partial\lambda_j}$ indicates that if the flux at the $j^{th}$ wavelength bin $\lambda_j$ goes up, the value of the $i^{th}$ label decreases and vice versa for a positive value of $\frac{\partial\text{Label}_i}{\partial\lambda_j}$.
An example of this type of sensitivity analysis is shown in Figure \ref{figure:jacobian}. This figure displays the derivative $\frac{\partial\xh{Mg}}{\partial\lambda_j}$, that is, the sensitivity of the neural network for the \xh{Mg} abundance. The derivative is averaged over two sets of stars in the high-SNR test set: all metal-poor stars with $\xh{Fe}<-1.5$ and all metal-rich stars with $\xh{Fe}>0.4$. The green regions in this figure show the ASPCAP windows used to derive the \xh{Mg} abundance and the same windows that we use when making the \xh{Mg} prediction with the neural network. It is clear that for the metal-rich stars which have strong Mg features, the neural network mainly pays attention to the regions of the spectrum within the ASPCAP windows and only limited attention to the rest of the spectrum (recall that our neural-network architecture is such that a limited amount of information about the full spectrum can be used in the \xh{Mg} prediction, through the connection between the large neural network that predicts the $[\ensuremath{T_\mathrm{eff}},\ensuremath{\log g},\xh{Fe}]$ parameters and the mini-network that predicts \xh{Mg}). For metal-poor stars, the network still pays much attention to the region of the spectrum within the ASPCAP windows, but it also pays stronger attention to regions outside of the windows (and example is the strong FeI feature in the red part of the spectrum). This is because for metal-poor stars, the spectral Mg features are weaker, and so the neural network can improve its predictions by making use of a limited amount of information from the full spectrum. This shows that letting the neural network see the whole spectra is essential for it to make sensible predictions in extreme cases.
The behavior for other individual elements is similar to that for \xh{Mg} shown in Figure \ref{figure:jacobian}. For the \ensuremath{T_\mathrm{eff}}\ and \ensuremath{\log g}\ predictions, the neural network uses the entire spectral range. For \ensuremath{T_\mathrm{eff}}\ the network mainly gets information from a large number of spectral features. For \ensuremath{\log g}, we display the derivative $\frac{\partial\ensuremath{\log g}}{\partial\lambda_j}$ in Figure \ref{figure:jacobian_logg}. While the derivative is non-zero over the full wavelength range, it is especially large near the hydrogen lines (the brackett series). This behaviour makes sense as the strong hydrogen lines are known to be strongly sensitive to \ensuremath{\log g}. Therefore, we see that even when the neural network is allowed to use the full spectrum, it uses a physically-plausible set of features in the spectrum to predict \ensuremath{\log g}.
\section{Variations}
\subsection{Training on the full, uncensored spectrum}\label{subsec:full_spec}
Before we settled on the \verb|ApogeeBCNNCensored()| NN architecture shown in Figure \ref{figure:nn_flow} that uses censored versions of the spectrum over the full wavelength range when determining the abundances of individual elements, we attempted using a simple multi-layered Bayesian convolutional NN with dropout. This network is available as \verb|ApogeeBCNN()| in the \texttt{astroNN} python package. Rather than splitting the label determination into a large NN to infer the main stellar parameters $[\ensuremath{T_\mathrm{eff}},\ensuremath{\log g},\xh{Fe}]$ and mini-networks to determine individual element abundances, this simple neural network is trained on the full spectra to infer all 22 parameters and abundances without any censorship. The performance of this \verb|ApogeeBCNN()| network on both the high-SNR and the individual-visits
test sets is similar to that of \verb|ApogeeBCNNCensored()| described in Sections \ref{subsec:highsnr} and Section \ref{subsec:lowsnr} for all labels. Thus, there is almost no loss in information in using only the censored spectra to determine individual abundances.
However, the \verb|ApogeeBCNN()| network fails to perform well in regions of abundance space that are not well covered by the training set and where the intrinsic abundance trends are different from those for the majority of the sample. This is clearly seen when we apply the \verb|ApogeeBCNN()| network to the M13 globular cluster. As discussed in Section \ref{subsec:clusters} above, M13, like many globular clusters, has a wide spread in Al abundances and the Al abundances are anti-correlated with the Mg abundances (although for M13 the actual spread in Mg is very small; \citealt{2015AJ....149..153M}). In Figure \ref{figure:m13_al_mg}, we show the \xh{Al} vs. \xh{Mg} abundances for stars in M13 for the \verb|ApogeeBCNN()| and \verb|ApogeeBCNNCensored()| networks, as well as those for all stars in the training set. It is clear that for \verb|ApogeeBCNN()|, \xh{Al} is very strongly correlated with \xh{Mg} in M13. This likely results from the fact that the \xh{Al} abundance in M13 is difficult to measure (in large part because for the epoch at which most M13 stars were observed by APOGEE, one of the prominent Al lines in the $H$ band overlapped with a strong sky-emission line, rendering it unusable). In the absence of information on \xh{Al} from Al lines, \verb|ApogeeBCNN()| falls back on the correlation between \xh{Al} and \xh{Mg} seen in the training set and therefore, the \verb|ApogeeBCNN()| \xh{Al} and \xh{Mg} are almost entirely correlated. This happens because \verb|ApogeeBCNN()| does not know which features in the full spectrum belong to \xh{Al} and which to \xh{Mg}.
Because the \verb|ApogeeBCNNCensored()| network uses only regions of the spectrum with Al features to determine \xh{Al}, it provides \xh{Al} measurements that are more in line with the results from \citet{2015AJ....149..153M}. As discussed above, the spread in \xh{Al} as determined by \verb|ApogeeBCNNCensored()| is about the same as that determined by \citet{2015AJ....149..153M}.
Despite the fact that the neural network trained on the full spectrum fails in certain regions of parameter space, we do allow a limited amount of information from the full spectrum to be used when determining the individual-element abundances. On the one hand these are the determinations of the overall stellar parameters $[\ensuremath{T_\mathrm{eff}},\ensuremath{\log g},\xh{Fe}]$, but we also include an additional, trainable two-neuron connection in the censored network as shown in Figure \ref{figure:nn_flow}. This allows the abundance prediction to depend on the full spectrum in a way that is not predetermined by us, but is learned from the training set. Such a connection makes physical sense, because the abundance of certain elements has a strong effect on the structure of the stellar photosphere, which in turn affects all parts of the spectrum. This is the case, for example, for carbon and oxygen in the cool stars observed by APOGEE \citep[e.g.,][]{2012AJ....144..120M}, but is also the case for other elements that are strong electron donors (e.g., Mg, Si), which therefore may also cause small effects through the spectral region. The two-neuron connection in \verb|ApogeeBCNNCensored()| allows such effects to be determined directly from the training data in a limited manner, without letting abundances of individual elements be determined entirely through correlations with other elements in the training data.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{m13_al_mg.pdf}
\caption{\xh{Al} vs. \xh{Mg} as determined by the \texttt{ApogeeBCNNCensored()} NN, which only uses regions of the spectrum containing spectral features for each element to be determined, and by the \texttt{ApogeeBCNN()} NN, which uses the full spectrum for each element, for stars in the globular cluster M13. The black points show the distribution of \xh{Al} vs. \xh{Mg} in the training set while the blue and red point with errorbars show the median error with \texttt{ApogeeBCNNCensored()} and \texttt{ApogeeBCNN()} respectively. Because little information about \xh{Al} is available from regions containing Al features for M13 stars, the \texttt{ApogeeBCNN()} predictions of \xh{Al} follow the correlation with \xh{Mg} that is present in the data set. The \texttt{ApogeeBCNNCensored()} NN avoids this and displays the correct Al spread in M13 and show a better agreement with the analysis of \citet{2013AJ....146..133M} than ASPCAP does.}
\label{figure:m13_al_mg}
\centering
\end{figure}
\subsection{Training with small data sets}
We have trained our neural network using high-SNR, high-resolution APOGEE spectra. Such high-quality data are expensive to obtain, because they require a large amount of telescope time. If we want to use a data-driven approach such as the neural network trained here to ``transfer'' labels from a high-resolution, high-SNR survey (e.g., APOGEE) to a low-resolution, low-SNR survey (e.g., LAMOST, as done using the \texttt{Cannon 2} by \citet{2017ApJ...836....5H}, we need to obtain a number of spectra for stars in common between the surveys, which takes away from the ability to observe new targets. Therefore, we investigate in this section whether we can train the neural network with smaller data sets and retain the good performance of the approach.
In traditional machine-learning techniques, the number of parameters is typically kept smaller than the size of the training data, because otherwise the machine-learning method will end up overfitting to the training data and therefore fail to generalize to the test data. However, a modern, medium-sized ANN often has billions of parameters that are optimized with less than a million training data. In our application, the network we use in Section \ref{sec:main_nn} has $\approx 3.5$ million trainable parameters, while the training set has $\approx 30,000$ objects. Therefore, the number of trainable parameters is more than a hundred times the number of training data. However, each spectrum consists of 7,514 flux values at different wavelengths, so the total number of training data points is $\approx 200$ million, more than the number of parameters. Therefore, we expect that we may be able to train the network with even fewer training objects.
The key to being able to train a large ANN with smaller training data sets lies in regularization \citep{2016arXiv161103530Z}. In our case, this regularization is provided by the dropout variational inference method used as a Bayesian approximation. Without this regularization, the network would simply memorize the results for the training set. The dropout procedure reduces the effective capacity of the network, by not allowing this type of memorization to work, and therefore helps the network to generalize to the test data.
We may ask whether we can reduce the number of training data. To do this, we train neural networks with exactly the same architecture and parameters as \texttt{ApogeeBCNNCensored()} but with training sets that are factors of $2^k$ smaller, from $2$ to $32$. The resulting bias and scatter of the residuals for the individual visit test set are displayed in Table \ref{table:smalldata_result} for the three main stellar parameters \ensuremath{T_\mathrm{eff}}, \ensuremath{\log g}, and \xh{Fe} (we do not show the results for the individual abundances, but these display similar trends). The bias in $\ensuremath{T_\mathrm{eff}}$ stays around $0\,\mathrm{K}$ for any size training set, similar to the $\ensuremath{T_\mathrm{eff}}$ bias shown in Table \ref{table:indi_result}. Compared to training on the whole training set, the network trained on 4,175 spectra (which is $12.5\%$ of the original training set) has larger scatter by $5$K, $0.012\,$dex and $0.003\,$dex in $\ensuremath{T_\mathrm{eff}}$, $\ensuremath{\log g}$, and $\xh{Fe}$ without introducing too much bias. Since the test set is mostly dominated by solar abundance stars, metrics in Table \ref{table:smalldata_result} are also mostly dominated by those stars. In regions of parameter space that are sparsely covered by the original training set, the performance becomes much worse with small training set. With even smaller training sets, the scatter across the whole parameters space as well as NN uncertainty increases significantly. Therefore, what is more important than the overall size of the training set is that it covers a wide range of possible parameter space.
Thus, we can obtain almost the same performance with a network trained on only a few thousand stars. Usage of the neural-network approach described in this paper for transferring APOGEE labels to surveys like LAMOST therefore look promising.
\begin{table}
\centering
\caption{Training on small data sets: median and \ensuremath{\sigma^{\mathrm{MAD}}}\ of the \ensuremath{T_\mathrm{eff}}, \ensuremath{\log g}, \xh{Fe} residuals between the NN prediction and ASPCAP results for the individual test set when the NN is trained with a limited amount of data. Each label column shows median / \ensuremath{\sigma^{\mathrm{MAD}}}.}
\label{table:smalldata_result}
\begin{tabular}{lccc}
\hline
\# of objects \\ (7,514 pixels each) & \ensuremath{T_\mathrm{eff}}\ (k) & \ensuremath{\log g}\ (dex) & \xh{Fe} (dex) \\
\hline
$33407$ & $-1\ / \ 24$ & $0.000\ / \ 0.046$ & $-0.004\ / \ 0.018$ \\
$16703$ & $3\ / \ 24$ & $0.010\ / \ 0.055$ & $-0.005\ / \ 0.019$ \\
$8351$ & $1\ / \ 27$ & $0.011\ / \ 0.059$ & $-0.006\ / \ 0.022$ \\
$4175$ & $1\ / \ 28$ & $0.007\ / \ 0.058$ & $-0.004\ / \ 0.021$ \\
$2087$ & $-4\ / \ 35$ & $0.013\ / \ 0.083$ & $-0.014\ / \ 0.029$ \\
$1043$ & $9\ / \ 53$ & $0.03\ / \ 0.16$ & $-0.012\ / \ 0.039$ \\
\hline
\end{tabular}
\end{table}
\subsection{Importance of continuum normalization}
\begin{table}
\centering
\caption{Neural-network prediction results on the high-SNR test set from comparing to ASPCAP, when using ASPCAP's procedure for continuum-normalization on the training and testing spectra.}
\label{table:aspcapnorm_result}
\begin{tabular}{lrr}
\hline
Label & Median of residual & \ensuremath{\sigma^{\mathrm{MAD}}}\ of residual\\
\hline
$\ensuremath{T_\mathrm{eff}}$ & $-22 \text{ K}$ & $31 \text{ K}$ \\
$\ensuremath{\log g}$ & $0.010 \text{ dex}$ & $0.050 \text{ dex}$ \\
$\xh{C}$ & $-0.003 \text{ dex}$ & $0.051 \text{ dex}$ \\
$\xh{CI}$ & $0.010 \text{ dex}$ & $0.056 \text{ dex}$ \\
$\xh{N}$ & $0.008 \text{ dex}$ & $0.064 \text{ dex}$ \\
$\xh{O}$ & $-0.017 \text{ dex}$ & $0.047 \text{ dex}$ \\
$\xh{Na}$ & $-0.01 \text{ dex}$ & $0.13 \text{ dex}$ \\
$\xh{Mg}$ & $-0.000 \text{ dex}$ & $0.025 \text{ dex}$ \\
$\xh{Al}$ & $-0.040 \text{ dex}$ & $0.070 \text{ dex}$ \\
$\xh{Si}$ & $-0.002 \text{ dex}$ & $0.029 \text{ dex}$ \\
$\xh{P}$ & $-0.02 \text{ dex}$ & $0.11 \text{ dex}$ \\
$\xh{S}$ & $0.011 \text{ dex}$ & $0.060 \text{ dex}$ \\
$\xh{K}$ & $-0.010 \text{ dex}$ & $0.046 \text{ dex}$ \\
$\xh{Ca}$ & $-0.014 \text{ dex}$ & $0.030 \text{ dex}$ \\
$\xh{Ti}$ & $-0.024 \text{ dex}$ & $0.050 \text{ dex}$ \\
$\xh{TiII}$ & $-0.04 \text{ dex}$ & $0.11 \text{ dex}$ \\
$\xh{V}$ & $-0.007 \text{ dex}$ & $0.099 \text{ dex}$ \\
$\xh{Cr}$ & $0.000 \text{ dex}$ & $0.048 \text{ dex}$ \\
$\xh{Mn}$ & $-0.016 \text{ dex}$ & $0.038 \text{ dex}$ \\
$\xh{Fe}$ & $-0.003\text{ dex}$ & $0.018 \text{ dex}$ \\
$\xh{Co}$ & $-0.02 \text{ dex}$ & $0.14 \text{ dex}$ \\
$\xh{Ni}$ & $0.003 \text{ dex}$ & $0.030 \text{ dex}$ \\
\hline
\end{tabular}
\end{table}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{xfe_alpha.pdf}
\caption{Neural network and ASPCAP predictions on 157,598 APOGEE DR14 stars after a cut on neural network \ensuremath{\log g}\ uncertainty $<0.2\,\mathrm{dex}$ as a means to filter out (a) main-sequence stars (as discussed in the right panel of Figure~\ref{figure:logg_teff_isochrones} ) and (b) problematic spectra that result in a high uncertainty in labels including \ensuremath{\log g}. This figure shows all $\alpha$ and odd-Z light elements Na, Al, and K among our 20 labels prediction. For most abundances, the NN abundances display less scatter and a clearer high/low alpha sequence than ASPCAP.}
\label{figure:xfe_alpha}
\centering
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{xfe_not_alpha.pdf}
\caption{Like Figure \ref{figure:xfe_alpha}, but for the remaining elements. For abundances like \xfe{Cr} and \xfe{Ni}, the NN abundances show much tighter scatter at \xfe{X}=0 as expected.}
\label{figure:xfe_not_alpha}
\centering
\end{figure*}
As described in Section \ref{subsec:reduction}, we have chosen to employ a custom continuum normalization procedure that uses a set of pixels assumed to represent the (pseudo-)continuum on both individual exposures and on the combined spectra, instead of using ASPCAP's pseudo-continuum normalized spectra, which simply fit a polynomial to each detector's spectrum. Both of these procedures are expected to be independent of SNR, but the procedure we use produces continuum-normalized spectra that are more similar to those used in traditional spectroscopic analyses. In this section, we test what happens when we use ASPCAP's procedure instead.
Table \ref{table:aspcapnorm_result} shows the result of a \verb|ApogeeBCNNCensored()| neural network trained on the same data but with spectra that are continuum normalized using ASPCAP's procedure (we simply use the continuum-normalized spectra provided in the APOGEE data release for this, rather than attempting to reproduce ASPCAP's normalization). We then similarly test on the same testing data as in the high-SNR test set, but with spectra normalized using ASPCAP's procedure. Comparing the results in Table \ref{table:aspcapnorm_result} with those in Table \ref{table:highsnr_result}, we see that the resulting biases are similar, but the scatter in some of the abundance labels (e.g., \xh{C} and \xh{N}) is considerably larger. This test therefore demonstrates that the continuum normalization that we use does improve the neural network's performance, so there is use in attempting to obtain a (pseudo-)continuum that is close to the true continuum.
\section{Abundance distributions in the Milky Way}
To further illustrate the power of the NN approach, we show the \xfe{X} vs. \xh{Fe} distributions of stars for all elements measured by ASPCAP and compare it to the ASPCAP results for the same stars. We only use a single cut to the full APOGEE DR14 catalog of spectra: we remove all stars with \ensuremath{\log g}\ uncertainty larger than 0.2 dex. This cut removes any problematic spectra which result in large label uncertainties (traced by the \ensuremath{\log g}\ uncertainty) and also cuts out essentially all dwarfs; because the NN \ensuremath{\log g}\ results for dwarfs are based on extrapolation---the training set contains no \ensuremath{\log g}\ for dwarfs---the actual NN \ensuremath{\log g}\ values are very wrong, so it is essential to cut on their uncertainty instead (see Figure \ref{figure:logg_teff_isochrones}). After this cut, we are left with a sample of 157,598 stars.
Figure \ref{figure:xfe_alpha} displays the abundance distribution for $\alpha$ and odd-Z light elements Na, Al, and K for these stars, both for the NN and for ASPCAP. Figure \ref{figure:xfe_not_alpha} shows the distribution of the remaining elements. As above, we show results for CI and TiII separately, as these are measured separately by ASPCAP. It is clear that the NN abundance patterns are tighter than those obtained from ASPCAP. This is especially clear for the $\alpha$ elements (O, Mg, Si, S, Ca, Ti), which has a distinct bimodal structure at intermediate metallicities ($\xh{Fe} \approx -0.2$ to $-0.5$) in \emph{all} $\alpha$ elements. The low-$\alpha$ sequence in all of these distributions is also significantly tighter than in the ASPCAP results. We also see that Al and K essentially behave as $\alpha$ elements, both displaying a clear bimodal structure with the same pattern as the $\alpha$ elements.
\section{Discussion}
\subsection{Performance}
Deep learning is a promising tool for the big-data era in astronomy. Besides the better precision and accuracy of the neural network compared to other approaches to spectroscopic analysis, the development of modern hardware-accelerated deep-learning technology means that our analysis is also much faster than traditional and other data-driven approaches. The neural network described in this work determines 22 labels from $\approx 100,000$ APOGEE spectra consisting of $\approx 7,500$ wavelength pixels each, while doing $100$ Monte Carlo dropout runs---equivalent to determining 22 parameters from 10,000,000 APOGEE spectra without dropout---in $\approx 300$ seconds or $\approx3\,\mathrm{ms}$ per star ($\approx30\,\mu\mathrm{s}$ per star without dropout). This performance is obtained on a Nvidia consumer graphics processor with $4,375$ GFLOPS at single precision, while the training steps with 40 epochs and a batch size of 64 takes $\approx 700$ seconds to complete in total\footnote{\textit{Batch size} refers to the number of training examples in one forward/backward pass and \textit{epoch} refers one forward pass and one backward pass of all the training examples. For example, with 6400 training data and a batch size of 64 with 5 epochs, each epoch will perform 100 gradient updates, each gradient is an average of 64 training data without replacement, and do it 5 times.}. The performance of neural networks is expected to get much better with upcoming graphics processors specializing in deep learning, such as Tensor Processing Units (TPUs) in upcoming consumer Nvidia GPUs \citep{2018arXiv180406826J}.
This fast performance offers great advantages in the era of large spectroscopic surveys. For example, it means that computational needs are far lower when only a small, high-SNR subset of the data (the training set) is analyzed with traditional, slow tools, while the majority of the data set can be processed much faster with the neural-network approach. Currently, the APOGEE ASPCAP pipeline requires a large cluster both to produce the library of synthetic spectra used in the ASPCAP fitting and for performing the fits themselves. Our fast framework allows for a much faster development cycle. This is the case in the narrow sense of providing the opportunity for fast prototyping and exploration of new neural-network model architectures. But when coupled with development of the tools to produce the small training set, this is also the case in the broader sense of seeing how changes to the input physics used in producing the training set affect the larger test sample.
Our extremely fast analysis coupled with the fact that the dropout procedure produces realistic uncertainties also opens up the possibility of real-time analysis of whether high enough SNR is obtained for a given level of abundance uncertainties. Stellar spectra could be analyzed on-the-fly as they are being collected and integration stopped when an acceptable uncertainty in the abundances is reached. With future, large fibre-positioner systems this could be used to run efficient, large spectroscopic surveys. Because our approach automatically assigns large uncertainties for objects far outside the training-set boundaries, such objects, which are likely to be of interest, would automatically obtain high SNR spectra in this approach.
\subsection{Comparison to other data-driven approaches to spectral analysis} \label{subsec:assumption}
Another data-driven approach to high-resolution analysis is provided by the \texttt{Cannon} \citep{2016arXiv160303040C}, which we have already discussed in the text above. The \texttt{Cannon}'s approach is fundamentally different from ours, in that the \texttt{Cannon} builds a data-driven model of the spectra as a function of stellar labels that is then used to fit labels to observed spectra, while our approach directly determines the mapping from spectra to labels. We therefore make slightly different assumptions about the relation between spectra and labels. To be clear, we list our methods and assumptions (or lack thereof) and how they compare to those made by the \texttt{Cannon}:
\begin{itemize}
\item NNs in this work map spectra directly to labels in a single step (and a sequence of these single steps to determine the uncertainty using dropout). The \texttt{Cannon 2} is a generative model that generates realistic spectra from labels and then matches spectra by $\chi^2$ minimization to find abundances.
\item We assume that the value of the labels is a continuous, smooth function of the flux. Thus, we assume that similar spectra have similar labels. This assumption is shared by the \texttt{Cannon 2} and essentially by all approaches to spectroscopic fitting.
\item Despite that, we do not require that spectra with similar labels are similar. This limits our ability to generate a new spectrum for certain set of labels and we cannot generate a set of $\Delta$flux for a given spectrum that changes only a single label, while keeping the others constant. The \texttt{Cannon 2} assumes that spectra with similar labels are similar.
\item The flexibility of the NNs mean that it only performs well on what it is trained and the ability to extrapolate is limited (but note that when extrapolation occurs, the returned uncertainties are very large). This is similar to the \texttt{Cannon 2}, although their simpler model may approximate physical models better and thus have better performance when extrapolating.
\end{itemize}
Our NN approach has some disadvantages with respect to forward-modeling approaches like the \texttt{Cannon}:\\
\emph{Interpretability:} Because the \texttt{Cannon} builds a forward model of the spectra it allows one to generate spectra from the model for a given set of labels. This can be used to inspect the internal functioning of the model and one can ask questions such as: does the behaviour when changing \xh{Al} while keeping all other parameters fixed make sense? Because of the reasons given above, our NN cannot answer such questions and therefore does not allow inspection of the model in this sense. It is possible to compute derivatives of each label value with respect to the input spectra (e.g., see Figures \ref{figure:jacobian} and \ref{figure:jacobian_logg}), but as these derivatives do not keep other labels fixed, they are less meaningful.\\
\emph{Handling of uncertainties in the input:} Our approach ignores uncertainties in the spectra. An approach that forward models the spectra can fit in both the training and test step while taking the flux uncertainties into account. Nevertheless, we have shown that even though we train on high SNR spectra, the NN performs well even at very low SNR. Part of the reason for this is that the flux uncertainties are relatively constant with wavelength. This is not the case for the label uncertainties in the training set (see above) and because it is far easier to only take one of input or output uncertainties into account in any machine-learning technique, it appears more important for the analysis of stellar spectra to take the label uncertainty into account.
On the other hand, the advantages of the NN approach over an approach like the \texttt{Cannon} are manifold:\\
\emph{Speed:} Because our method learns a mapping from spectra to labels and does not require any fitting when making predictions, it is much faster. As discussed above, we can determine labels for $\approx 100,000$ APOGEE stars in about three seconds on a single $\approx\$500$ GPU (and in five minutes if we also want the uncertainty), while running the \texttt{Cannon 2} on the same number of spectra takes about 20 minutes on a small cluster \citep{2016arXiv160303040C}.\\
\emph{Realistic uncertainties:} The \texttt{Cannon 2} can obtain uncertainty estimates from its $\chi^2$ fitting to the spectra and their uncertainties, but these uncertainties are typically much smaller than the scatter in the results obtained from cross-validation, demonstrating that the uncertainties are underpredicted. Our approach returns uncertainties that, at least within the well-populated regions of the training set, are approximately equal to the scatter from cross-validation or equal to the scatter in open clusters. Thus, our uncertainties can be used in practice to determine whether observed scatter is real or due to noise in the data.\\
\emph{Extrapolation warnings:} Data-driven approaches only perform well within the training set and do not perform well when extrapolating, especially approaches as flexible as our NN approach. Data-driven approaches therefore typically need a way to determine first whether or not a test object is within the bounds of the training set. For high-dimensional input and label spaces this boundary can be a complicated, high-dimensional surface that is difficult to determine. In tests, this issue is typically ignored, as most test sets are chosen to represent the training set (e.g., in the standard random partition of the full data set into a training and test set). As we demonstrated above, the uncertainties returned by our NN are very large whenever a spectrum or label outside of the training set is analyzed (e.g., the main-sequence stars' \ensuremath{\log g}\ in Figure \ref{figure:logg_teff_isochrones}). These large uncertainties can therefore be used as a warning flag for spectra or labels outside of the training set and our method thus provides an automatic check for the training-set boundaries.\\
\emph{Training on noisy and incomplete data:} Because we take the uncertainties on the training labels into account and can also use training stars for which not all labels are measured, we can train on an imperfect training set. This is important, because for various reasons many stars in the training set, even at high SNR, do not have complete and high-precision set of labels. The \texttt{Cannon 2} in its current implementation does not take uncertainties in the training labels into account and can only train on training objects with a complete set of labels.
The advantage of all data-driven approaches is that they allow properties such as stellar mass, age, or luminosity to be determined directly from stellar spectra. This is difficult to do with theoretical stellar models, because how these properties affect the spectra is not well known. For example, a common way to infer luminosity is by using isochrones based on stellar evolutionary models and stellar parameters and abundances measured from stellar spectra \citep{2016AA...585A..42S}, but stellar isochrones are not well calibrated for all types of stars and stages of stellar evolution. Spectra may contain direct indicators of mass \citep[e.g.,][]{2016ApJ...823..114N}, age, or luminosity that can be extracted by training on data sets for which these are known quantities (e.g., masses from asteroseismology, e.g., APOKASC: \citealt{2014ApJS..215...19P}; or luminosities from \emph{Gaia}).
\subsection{Comparison to neural-network approaches to spectral analysis based on synthetic spectra}\label{subsec:comparesynth}
The training set potentially does not represent all stars in the test set and, as discussed above, NNs trained on an unrepresentative training set will perform poorly for stars outside of the training set. Moreover, abundances determined by ASPCAP contain systematic errors that propagate to the NN during training and we cannot easily quantify these errors or correct them. These systematic errors result from the quality of ASPCAP synthetic grid and errors introduced by the data reduction and calibration steps in ASPCAP. One solution is using the techniques used by \texttt{StarNet} or \texttt{The Payne} \citep{2018arXiv180401530T} by training on theoretical synthetic spectra, because theoretical spectra with different combinations of abundances can be generated and we can choose these combinations to represent the likely set of abundances in the test set. In terms of performance, \texttt{StarNet} uses neural networks similar to those used here and, when implemented on a GPU, could be as fast to train and evaluate as our method. \texttt{The Payne}, however, exhibits far slower performance because it is a forward model of the spectra given the labels similar to the \texttt{Cannon}, but with higher complexity. Training \texttt{The Payne} takes about 5 CPU minutes per wavelength pixel or 26 days for all pixels, compared to our 10 minutes. Evaluating for a test spectrum takes about 1 CPU s per spectrum or about 28 hours for $100,000$ APOGEE stars, versus our 3 seconds for $100,000$ APOGEE stars (300 s including uncertainties).
Training on synthetic spectra can lead to bad performance when the theoretical spectra do not match the observed spectra well. This can be due to calibration, uncertainties, continuum-normalization issues, or unmodeled physics in the spectral synthesis. For example, without correcting systematic differences from the observed spectra before training, \texttt{StarNet} reports that their neural network trained on synthetic spectra does not perform well on observed spectra due to differences in the feature distributions between theoretical and real spectra. Of course, because data-driven approaches rely on a training set that is typically analyzed using synthetic spectra, they suffer from limitations in the input physics as well. However, through clever use of star clusters, binaries, and similar systems it may be possible to create training sets that are less sensitive to the theoretical modeling of stellar spectra. This is because if we can assume that these systems are chemically homogeneous, we can transfer the chemical labels from well-understood stellar types onto those of poorly-understood stellar types, thus creating an empirical training set for the poorly-understood types.
Another goal described in \texttt{Cannon 2} and the version of \texttt{StarNet} which trained on real world spectra is to identify potential unknown abundance lines. The use of a censored neural network in this work will render this practically impossible, because we limited the attention of the network to known spectral lines. However, even for a network trained on the full spectrum or on synthetic spectra, it will be questionable whether a previously unknown spectral feature that is identified really belongs to a certain element.
\section{Conclusions}
Large spectroscopic surveys are now routinely obtaining high-resolution spectra for hundreds of thousands of stars and upcoming surveys such as WEAVE \citep{2014SPIE.9147E..0LD}, SDSS-V \citep{2017arXiv171103234K}, 4MOST \citep{2012SPIE.8446E..0TD}, and MSE \citep{2016arXiv160600043M} will soon provide such spectra for millions of stars. Traditionally, such spectra are analyzed one-by-one with procedures that are largely manual, which is clearly impractical for large data sets. Fitting spectra with synthetic libraries (e.g., ASPCAP) is the currently favored approach, but this method is slow and current implementations do not deal well with low SNR spectra. As an alternative approach, we have presented neural network models for spectroscopic analysis to infer stellar parameters and chemical abundances with associated uncertainty using Bayesian ANNs with dropout variational inference. We implemented this method in a general, open-source python framework for ANNs called \texttt{astroNN} (see Appendix \ref{appendix:graph:astroNN}). We also release a catalog of abundances for the APOGEE DR14 data set determined by our \verb|ApogeeBCNNCensored()| network (see link in the Introduction).
Our neural-network method has various special ingredients beyond what would be used in a standard deep learning application. We (a) present a robust objective function for the neural network to learn from incomplete data while taking uncertainty in the training labels into account, (b) use a Bayesian neural network with dropout variational inference with this objective function to estimate the uncertainties on the labels determined by the neural network, and (c) combine a large NN to obtain the overall stellar parameters $[\ensuremath{T_\mathrm{eff}},\ensuremath{\log g},\xh{Fe}]$ with mini-networks to determine individual elemental abundances that use versions of the full spectrum censored to only include regions with spectral lines for a given element. With this approach, we simultaneously determine 22 stellar and elemental abundance labels accurately and precisely for both high and low SNR spectra. We implemented the method on a GPU using standard tools, which allows speed-ups of more than an order of magnitude and we make these tools easily accessible (see Appendix \ref{subsec:fastMC}). Our method is extremely fast, allowing stellar parameters and abundances for the entire APOGEE database to be determined in about ten minutes on a single GPU.
We performed detailed tests of our method by comparing to results from the standard APOGEE ASPCAP pipeline and by comparing results on high-SNR, combined spectra to those from low-SNR, individual exposures. At high SNR, we obtain abundances precise to $\approx 0.01$ to $0.02\,\mathrm{dex}$ and even at low-SNR (SNR$\approx 50$) we get precisions of $\approx 0.02$ to $0.03\,\mathrm{dex}$ for most elements. These precisions are confirmed by looking at the scatter in the abundances within open clusters, which we find to be $0.03\pm0.03\,\mathrm{dex}$. We also recovered the expected abundance trends in globular clusters, but found that the censoring in the network is crucial to obtain this, because training on the full spectrum for all elements causes the NN to depend too strongly on correlations between elements within the training set and therefore to fail when these correlations are absent, like in globular clusters. We also demonstrated that a large neural network can work well with a limited amount of training data, finding barely degraded performance for training sets that only consists of thousands of spectra.
The speed and flexibility of neural networks mean that they are a highly useful tool for spectroscopic data analysis. They allow the results obtained from a detailed analysis of a small calibration set to be transfered to an entire large data set of millions of spectra in a matter of minutes and can thus be a great aid in the development of the next generation of spectroscopy tools. They could also be of use in earlier stages of the data processing, for example, for homogenizing spectra taken with different instruments (e.g., the northern and southern APOGEE spectrographs) or for removing instrumental systematics such as persistence \citep[e.g.,][]{2017MNRAS.470.4782J}. Such applications would be easy to pursue with the \texttt{astroNN} software package described in the appendix.
\section*{Acknowledgements}
It is our pleasure to thank Kim Venn and other members of the \texttt{StarNet} group for valuable feedback and for releasing their code publicly. We also thank Natalie Price-Jones for help with the APOGEE data. HL and JB received support from the Natural Sciences and
Engineering Research Council of Canada (NSERC; funding reference number RGPIN-2015-05235) and from an Ontario Early Researcher Award (ER16-12-061). JB also received partial
support from an Alfred P. Sloan Fellowship.
Funding for the Sloan Digital Sky Survey IV has been
provided by the Alfred P. Sloan Foundation, the U.S. Department
of Energy Office of Science, and the Participating
Institutions. SDSS-IV acknowledges support and resources
from the Center for High-Performance Computing at the
University of Utah. The SDSS web site is www.sdss.org.
|
\section{Introduction}
In \cite{J} and \cite{M}, algebraic structures called \textit{quandles} (or
\textit{distributive groupoids}) were introduced
as an abstraction of the Wirtinger presentation of the fundamental group
of the complement of a knot in $\mathbb{R}^3$. Colorings of knot diagrams
by elements of finite quandles define an integer-valued invariant known as
the \textit{quandle counting invariant}. Invariants of quandle-colored
knots, e.g. Boltzmann weights defined from quandle 2-cocycles, can be used
to strengthen this invariant, defining new invariants known
as \textit{enhancements}. See \cite{EN} for more.
Initially defines in \cite{AG},
algebraic structures called \textit{quandle modules} were
used in \cite{CEGS} to enhance the quandle counting invariant, inspiring later
generalizations by one of the authors
to the cases of rack modules in \cite{HHNYZ}, biquandle modules in \cite{BN}
and birack shadow modules in \cite{NP}, among others. In each of these
cases, a counting invariant is enhanced with secondary colorings by elements
of a commutative ring with identity obeying an Alexander-style relation
which depends on the quandle colors at the crossing.
In \cite{MN}, the notion of using sets with ternary operations to define knot
invariants was considered, with colorings of regions in the planar complement
of a knot or link diagram by elements of structures known as \textit{ternary
quasigroups}. These structures can be seen as an abstraction of the Dehn
presentation of the knot group analogous to the way quandles abstract the
Wirtinger presentation. In \cite{MN2}, ternary quasigroup invariants
were enhanced with a homology theory. A related structure called
\textit{biquasiles} was introduced in \cite{NN} by two of the authors
with applications to surface-links in \cite{KN} by one of the authors.
Recently ternary quasigroup operations known as \textit{Niebrzydowski
tribrackets} have been studied with
additional generalizations to the cases of virtual knots in \cite{NP2} and
trivalent spatial graphs in \cite{GNT}.
In this paper we apply the idea behind quandle modules to the case of
Niebrzydowski
tribrackets, obtaining an infinite family of ehancements of the tribracket
counting invariant. The paper is organized as follows. In Section \ref{T}
we recall the basics of Niebrzydowski tribrackets and see some examples,
and introduce an enhancement for Alexander tribrackets. In Section
\ref{TM} we define tribracket modules and introduce the tribracket module
enhancement of the counting invariant. We compute some examples to show that
the enhancement is nontrivial. We conclude in Section \ref{Q} with some
questions for future work.
\section{Tribrackets}\label{T}
We begin with a defintion.
\begin{definition} (see e.g. \cite{NOO})
Let $X$ be a set. A \textit{horizontal tribracket} on $X$ is a
map $[\ ,\ ,\ ]:X\times X\times X\to X$ satisfying
\begin{itemize}
\item[(i)] For any subset $\{a,b,c,d\}\subset X$, any three elements uniquely
determine the fourth such that $[a,b,c]=d$, and
\item[(ii)]
\[[b,[a,b,c],[a,b,d]]=[c,[a,b,c],[a,c,d]]=[d,[a,b,d],[a,c,d]].\]
\end{itemize}
\end{definition}
\begin{example}
Let $X$ be any module over a commutative ring $R$ wth identity. Then any
pair of units $x,y\in R^{\times}$ defines a tribracket structure on $X$
by setting
\[[a,b,c]=-xya+xb+yc.\]
We call this an \textit{Alexander tribracket} and denote it by
$X=(R,x,y)$. Let us verify axiom (ii):
\begin{eqnarray*}
{}[b,[a,b,c],[a,b,d]] & = & -xyb+x(-xya+xb+yc)+y(-xya+xb+yd) \\
& = & (-x^2y-xy^2)a+x^2b+xyc+y^2d \\
{}[c,[a,b,c],[a,c,d]] & = & -xyc+x(-xya+xb+yc)+y(-xya+xc+yd) \\
& = & (-x^2y-xy^2)a+x^2b+xyc+y^2d \\
{}[d,[a,b,d],[a,c,d]] & = & -xyd+x(-xya+xb+yd)+y(-xya+xc+yd) \\
& = & (-x^2y-xy^2)a+x^2b+xyc+y^2d.
\end{eqnarray*}
\end{example}
\begin{example}
Let $G$ be a group. Then $G$ has the structure of a tribracket by setting
\[[a,b,c]=ba^{-1}c.\]
We call this a \textit{Dehn tribracket}. As with the Alexander case, let us
verify axiom (ii):
\begin{eqnarray*}
{}[b,[a,b,c],[a,b,d]]
& = & [a,b,c]b^{-1}[a,b,d] \\
& = & ba^{-1}cb^{-1}ba^{-1}d \\
& = & ba^{-1}ca^{-1}d \\
{}[c,[a,b,c],[a,c,d]]
& = & [a,b,c]c^{-1}[a,c,d] \\
& = & ba^{-1}cc^{-1}ca^{-1}d \\
& = & ba^{-1}ca^{-1}d \\
{}[d,[a,b,d],[a,c,d]]
& = & [a,b,d]d^{-1}[a,c,d] \\
& = & ba^{-1}dd^{-1}ca^{-1}d \\
& = & ba^{-1}ca^{-1}d
\end{eqnarray*}
as required.
\end{example}
\begin{example}
We can specify a tribracket structure on a finite set $X=\{1,2,\dots, n\}$
with an operation 3-tensor, i.e. an ordered list of $n$ $n\times n$ matrices
with elements in $X$ such that the element in matrix $a$, row $b$, column $c$
is $[a,b,c]$. This notation enables us to compute with tribrackets for
which we lack algebraic formulas. For example, the set $X=\{1,2,3\}$
has tribracket structures including
\[\left[
\left[\begin{array}{rrr}
1 & 3 & 2 \\
2 & 1 & 3 \\
3 & 2 & 1
\end{array}\right],
\left[\begin{array}{rrr}
2 & 1 & 3 \\
3 & 2 & 1 \\
1 & 3 & 2
\end{array}\right],
\left[\begin{array}{rrr}
3 & 2 & 1 \\
1 & 3 & 2\\
2 & 1 & 3
\end{array}\right]
\right].\]
In this case for example, we verify axiom (ii) for the case $a=1,$ $b=2,$ $c=3,$ $d=1$ by the computation
\begin{eqnarray*}
{}[2, [1,2,3], [1,2,1]] & = & [2,3,2] = 3, \\
{}[3, [1,2,3], [1,3,1]] & = & [3,3,3] = 3 \quad \mathrm{and} \\
{}[1, [1,2,1], [1,3,1]] & = & [1,2,3] = 3.
\end{eqnarray*}
\end{example}
The tribracket axioms are motivated by the Reidemeister moves using the
following region coloring rule:
\[\includegraphics{dn-sn-ys-1.pdf}\]
We call the invertibility conditions in axiom (i)
\textit{left, center} and \textit{right invertibility} for the ability to
uniquely recover $a,b,$ and $c$ respectively in $[a,b,c]=d$ given the other
three. These are the conditions required to guarantee that for every coloring
on one side of an oriented Reidemeister I or II move, there is a \textit{unique}
coloring of the diagram on the other side of the move which agrees with the
original coloring outside the neighborhood of the move. Axiom (ii) is
the condition required by the Reidemeister III move needed to complete a
generating set of oriented Reidemesiter moves:
\[\includegraphics{dn-sn-ys-2.pdf}\]
It follows that for any tribracket $X$, the number of $X$-colorings of an
oriented knot or link $L$ diagram is an integer-valued link invariant, which
we call the \textit{tribracket counting invariant}, denoted
$\Phi_X^{\mathbb{Z}}(L)$. We will denote the set of $X$-colorings of $L$ as
$\mathcal{C}_X(L)$, and we have $\Phi_X^{\mathbb{Z}}(L)=|\mathcal{C}_X(L)|$.
\begin{example}
If $X$ is an Alexander tribracket, we can compute $\Phi_X^{\mathbb{Z}}(L)$
using linear algebra. Let $X$ be the Alexander tribracket on $\mathbb{Z}_3$
with $x=1$ and $y=2$, so we have $[a,b,c]=a+b+2c$. Then the trefoil knot $3_1$
below has system of coloring equations
\[\raisebox{-0.5in}{\includegraphics{dn-sn-ys-3.pdf}}
\quad\begin{array}{rcl}
{}[a,b,c] & = & e \\
{}[a,c,d] & = & e \\
{}[a,d,b] & = & e
\end{array}
\]
and after row-reduction mod $3$,
\[
\left[\begin{array}{rrrrr}
1 & 1 & 2 & 0 & 2 \\
1 & 0 & 1 & 2 & 2 \\
1 & 2 & 0 & 1 & 2
\end{array}\right]
\rightarrow
\left[\begin{array}{rrrrr}
1 & 0 & 0 & 2 & 2 \\
0 & 1 & 1 & 1 & 0 \\
0 & 0 & 0 & 0 & 0
\end{array}\right]
\]
we see that $\Phi_X^{\mathbb{Z}}(3_1)=3^3=27$. This distinguishes the trefoil
from the unknot $0_1$, which has $\Phi_X^{\mathbb{Z}}(0_1)=3^2=9$ $X$-colorings.
\end{example}
\begin{remark}
Writing the region Alexander tribracket coloring equations from an oriented
link diagram as a system of linear equations yields a matrix from which
an Alexander matrix and the Alexander polynomial of the knot or link can be
derived.
\end{remark}
An \textit{enhancement} of a counting invariant is a generally stronger
invariant from which we can recover the counting invariant. Any invariant
$\phi$ of $X$-colored knots and links defines an enhancement by taking
the multiset of $\phi$-values over the set of colorings $L_f$
of the knot or link $L$,
\[\Phi_X^{\phi,M}(L)=\{\phi(L_f)\ |\ L_f\in \mathcal{C}_X(L)\}.\]
Many such examples have been previously studied in the cases of quandle and
biquandle-colored knots and links, starting with the 2-cocycle enhancements
in \cite{CJKLS}; see \cite{EN} for a whole chapter of examples.
\begin{remark}
For Alexander tribrackets $X=(R,x,y)$, we can enhance the tribracket counting
invariant by setting $\phi(L_f)$ equal to the rank of the image submodule
of the coloring, analogous to the $(t,s)$-rack enhancements in \cite{CN}.
The multiset of the ranks of these image submodules over the complete
set of colorings is the \textit{Alexander image enhancement} of the tribracket
counting invariant,
\[\Phi_X^{A}(L)=\{\mathrm{rank}(\mathrm{Span}(\mathrm{Im}(f)))\ |\ f\in \mathcal{C}_X(L)\}.\]
\end{remark}
\section{Tribracket Modules}\label{TM}
We would like to ehance the tribracket counting invariant by finding an
invariant of $X$-colored oriented knot and link diagrams. To this end, we
make the following definition, analogous to the cases of quandles, racks,
and biracks in papers such as \cite{BN,CEGS,HHNYZ,NP}.
\begin{definition}\label{def:tbmod}
Let $X$ be a tribracket and $R$ a commutative ring with identity. A
\textit{tribracket module structure} on $R$, also called an \textit{$X$-module},
is a choice of units $x_{a,b,c}$ and $y_{a,b,c}$ for each triple of elements
of $X$ satisfying the conditions
\begin{eqnarray*}
x_{c,[a,b,c],[a,c,d]}x_{a,b,c} & = & x_{d,[a,b,d],[a,c,d]}x_{a,b,d}\\
& = & x_{b,[a,b,c],[a,b,d]}x_{a,b,c} +y_{b,[a,b,c],[a,b,d]}x_{a,b,d} \\
& & -x_{b,[a,b,c],[a,b,d]}y_{b,[a,b,c],[a,b,d]} \\
y_{c,[a,b,c],[a,c,d]}y_{a,c,d} & = & y_{b,[a,b,c],[a,b,d]}y_{a,b,d}\\
& = & x_{d,[a,b,d],[a,c,d]}y_{a,b,d} +y_{d,[a,b,d],[a,c,d]}y_{a,c,d} \\
& & -x_{d,[a,b,d],[a,c,d]}y_{d,[a,b,d],[a,c,d]} \\
x_{b,[a,b,c],[a,b,d]}y_{a,b,c} & = & y_{d,[a,b,d],[a,c,d]}x_{a,c,d}\\
& = & x_{c,[a,b,c],[a,c,d]}y_{a,b,c} +y_{c,[a,b,c],[a,c,d]}x_{a,c,d} \\
& & -x_{c,[a,b,c],[a,c,d]}y_{c,[a,b,c],[a,c,d]}\\
x_{c,[a,b,c],[a,c,d]}x_{a,b,c}y_{a,b,c} + y_{c,[a,b,c],[a,c,d]}x_{a,c,d}y_{a,c,d}
& = & x_{b,[a,b,c],[a,b,d]}x_{a,b,c}y_{a,b,c} + y_{b,[a,b,c],[a,b,d]}x_{a,b,d}y_{a,b,d} \\
& = & x_{d,[a,b,d],[a,c,d]}x_{a,b,d}y_{a,b,d} + y_{d,[a,b,d],[a,c,d]}x_{a,c,d}y_{a,c,d}
\end{eqnarray*}
for all $a,b,c,d\in X$.
\end{definition}
A tribracket module over a tribracket $X=\{1,2,\dots, n\}$
is specified with a pair $V=(x,y)$ of $3$-tensors such that
the entries in matrix $a$, row $b$, column $c$ are $x_{a,b,c}$, $y_{a,b,c}$
respectively.
\begin{remark}
The term ``module'' here follows the use of the term in \cite{AG} and other
previous work; it is justified by the fact that the invariant we define below
associates an $R$-module, generated by the regions in $L$ with relations
determined by the coloring, to each
$X$-coloring $L_f$ of a link $L$.
\end{remark}
\begin{example}
A \textit{constant tribracket module} is one in which the $x_{a,b,c}$ and
$y_{a,b,c}$-values do not depend on $a,b,c\in X$. In this case, $V$ is
an Alexander tribracket on $R$ and the sticker colorings are independent
of the $X$ colorings. For instance the tribracket
\[X=\left[\left[\begin{array}{rr}
1 & 2 \\
2 & 1
\end{array}\right],\left[\begin{array}{rr}
2 & 1 \\
1 & 2
\end{array}\right]\right]\]
has constant tribracket modules with $\mathbb{Z}_3$ coefficients including
\begin{eqnarray*}
V_1 & = & \left[\left[\begin{array}{rr}
1 & 1 \\
1 & 1
\end{array}\right],\left[\begin{array}{rr}
1 & 1 \\
1 & 1
\end{array}\right]\right], \quad
\left[\left[\begin{array}{rr}
1 & 1 \\
1 & 1
\end{array}\right],\left[\begin{array}{rr}
1 & 1 \\
1 & 1
\end{array}\right]\right],\\
V_2 & = & \left[\left[\begin{array}{rr}
1 & 1 \\
1 & 1
\end{array}\right],\left[\begin{array}{rr}
1 & 1 \\
1 & 1
\end{array}\right]\right], \quad
\left[\left[\begin{array}{rr}
2 & 2 \\
2 & 2
\end{array}\right],\left[\begin{array}{rr}
2 & 2 \\
2 & 2
\end{array}\right]\right],\\
V_3 & = & \left[\left[\begin{array}{rr}
2 & 2 \\
2 & 2
\end{array}\right],\left[\begin{array}{rr}
2 & 2 \\
2 & 2
\end{array}\right]\right], \quad
\left[\left[\begin{array}{rr}
1 & 1 \\
1 & 1
\end{array}\right],\left[\begin{array}{rr}
1 & 1 \\
1 & 1
\end{array}\right]\right] \quad \mathrm{and}\\
V_4 & = & \left[\left[\begin{array}{rr}
2 & 2 \\
2 & 2
\end{array}\right],\left[\begin{array}{rr}
2 & 2 \\
2 & 2
\end{array}\right]\right], \quad
\left[\left[\begin{array}{rr}
2 & 2 \\
2 & 2
\end{array}\right],\left[\begin{array}{rr}
2 & 2 \\
2 & 2
\end{array}\right]\right].\end{eqnarray*}
\end{example}
The tribracket module axioms are motivated by the Reidemeister moves with
the coloring scheme below. Given an oriented knot or link diagram with
a tribracket coloring, we put a secondary labeling on the regions with a
\textit{sticker} (an element of $R$, represented in our diagrams as
an element of $R$ surrounded by a box) in each region.
\[\includegraphics{dn-sn-ys-4.pdf}\]
The sticker colorings must then satisfy the rule
\[z=-x_{a,b,c}y_{a,b,c}v+x_{a,b,c}u+y_{a,b,c}w,\]
a customized Alexander tribracket-style coloring with coefficients
depending on the $X$-colors at the crossing.
\begin{proposition}
Let $X$ be a tribracket, $R$ a commutative ring with identity and $V$ an
$X$-module over $R$. Then sticker colorings of the regions in an
oriented knot diagram's planar complement are in one-to-one correspondence
before and after $X$-colored Reidemeister moves.
\end{proposition}
\begin{proof} Recall (see \cite{P} for example) that the set of all four
oriented Reidemeister I moves, all four oriented Reidemeister II moves, and
the all-positive Reidemeister III move forms a generating set of oriented
Reidemeister moves.
Invertibility of $x_{a,b,c}$ and $y_{a,b,c}$ satisfies the claim for Reidemeister
I and II moves; let us illustrate with one of the four oriented
Reidemeister II moves.
\[\includegraphics{dn-sn-ys-8.pdf}\]
The condition we need for uniqueness of the sticker $u$ given $v,w$ and $z$
is that the equation
\[z=-x_{a,b,c}y_{a,b,c}v+x_{a,b,c}u+y_{a,b,c}w\]
should be solvable for $u$ in terms of $v,w$ and $z$; this is possible
provided $x_{a,b,c}$ is invertible in $R$:
\[u=y_{a,b,c}v+x_{a,b,c}^{-1}z-x_{a,b,c}^{-1}y_{a,b,c}w\]
The other Reidemeister I and II moves are similar. It remains only to
verify for the all-positive Reidemeister III move.
\[\includegraphics{dn-sn-ys-5.pdf}\]
The region marked $\framebox{$\ast$}$ gets three sticker colorings
-- one on the left side of the move and two on the right --
which must all agree. Each of
these is an expression in the independent variables $u,v,z,w$, so we can
compare these coefficients to obtain the necessary equations, i.e., the
conditions in Defintion \ref{def:tbmod}.
\end{proof}
\begin{definition}
Let $L$ be an oriented link diagram, $X$ a tribracket, $R$ a commutative ring
with identity and $V=(x,y)$ an $X$-module structure on $R$. For each
$X$-coloring $f\in \mathcal{C}_X(L)$ of $L$, let $A_f$ be the coefficient
matrix of the homogeneous system of linear equations
in the $R$-module generated by the regions of $L$ determined by the
sticker coloring equations. Then the
\textit{tribracket module multiset enhancement} of the tribracket counting
invariant is the multiset
\[\Phi_X^{V,M}(L)=\{|\mathrm{Ker} \ A_f|\ :\ f\in\mathcal{C}_X(L)\}\]
or if $R$ is infinite,
\[\Phi_X^{V,M}(L)=\{\mathrm{rank}(\mathrm{Ker} \ A_f)\ :\ f\in\mathcal{C}_X(L)\}.\]
We can optionally convert these multisets to ``polynomial''
form by replacing multiplicities with coefficients and elements with exponents
of a formal variable $u$
\[\Phi_X^{V}(L)=\sum_{f\in\mathcal{C}_X(L)}u^{|\mathrm{Ker} \ A_f|}\]
or if $R$ is infinite,
\[\Phi_X^{V}(L)=\sum_{f\in\mathcal{C}_X(L)}u^{\mathrm{rank}(\mathrm{Ker} \ A_f)}.\]
This notation has the advantage that evaluation at $u=1$ yields the original
counting invariant and provides easier visual comparison of invariant values.
\end{definition}
By construction, we have the following proposition:
\begin{proposition}
For any $X$-module $V$ over a tribracket $X$ and commutative ring with
identity $R$, $\Phi_X^{V,M}(L)$ and $\Phi_X^{V}(L)$ are invariants of oriented
knots and links.
\end{proposition}
\begin{example}
If $V$ is a constant $X$-module, then $|\mathrm{Ker}\ A_f|$ is just
the number of colorings of $L$ by the Alexander tribracket $A$ on $R$ with
parameters $(x,y)$ and we have
\[\Phi_X^{V}(L)=|\Phi_X^{\mathbb{Z}}(L)| u^{|\Phi_A^{\mathbb{Z}}(L)|}.\]
\end{example}
\begin{example}\label{ex:nt}
Let $V$ be an $X$-module with coefficients in a finite ring $R$. The unlink
of $n$ components has $n+1$ regions with no crossings and hence no
restrictions on $X$-colorings, so there are $|X|^{n+1}$ region colorings.
Each of these has similarly no restrictions on sticker colorings, so there are
$|R|^{n+1}$ sticker colorings for each region coloring. Hence,
the value of $\Phi_X^{V}(L)$ on the unlink of $n$ components is
$\Phi_X^{V}(L)=|X|^{n+1}u^{|R|^{n+1}}$.
\end{example}
\begin{example}\label{ex:nt8}
Let $X$ be the set $\{1,2\}$ with tribracket operation given by
\[\left[\left[\begin{array}{rr}
1 & 2 \\
2 & 1
\end{array}\right],\left[\begin{array}{rr}
2 & 1 \\
1 & 2
\end{array}\right]\right].\]
The trefoil knot $3_1$ has four $X$-colorings:
\[\includegraphics{dn-sn-ys-7.pdf}\]
Then we compute via \texttt{python} that $X$ has modules with $\mathbb{Z}_3$
coefficients including
\[V=
\left[\left[\begin{array}{rr}
2 & 2 \\
2 & 1
\end{array}\right],\left[\begin{array}{rr}
1 & 2 \\
2 & 2
\end{array}\right]\right], \quad
\left[\left[\begin{array}{rr}
1 & 2 \\
2 & 2
\end{array}\right],\left[\begin{array}{rr}
2 & 2 \\
2 & 1
\end{array}\right]\right].
\]
Consider the $X$-coloring of the trefoil with all regions colored $1\in X$:
\[\includegraphics{dn-sn-ys-6.pdf}\]
Let us compute the set of sticker colorings. We obtain linear system of
coloring equations over $R=\mathbb{Z}_3$
\[\begin{array}{rcl}
-x_{11}y_{11}u_2+x_{11}u_1+y_{11}u_4 & = & u_3 \\
-x_{11}y_{11}u_2+x_{11}u_1+y_{11}u_4 & = & u_3 \\
-x_{11}y_{11}u_2+x_{11}u_1+y_{11}u_4 & = & u_3 \\
\end{array}\Rightarrow
\begin{array}{rcl}
-(1)(2)u_2+1u_1+2u_4 & = & u_3 \\
-(1)(2)u_2+1u_4+2u_5 & = & u_3 \\
-(1)(2)u_2+1u_5+2u_1 & = & u_3 \\
\end{array}\]
which via row-reduction over $\mathbb{Z}_3$
\[
\Rightarrow
\left[\begin{array}{rrrrr}
1 & 1 & 2 & 2 & 0 \\
0 & 1 & 2 & 1 & 2 \\
2 & 1 & 2 & 0 & 1 \\
\end{array}\right]
\Rightarrow
\left[\begin{array}{rrrrr}
1 & 0 & 0 & 1 & 1 \\
0 & 1 & 2 & 1 & 2 \\
0 & 0 & 0 & 0 & 0 \\
\end{array}\right]
\]
has kernel of dimension $3$. Similarly, the other $X$-colorings have
$|R|^3=3^3=27$ colorings.
In particular the trefoil has $\Phi_X^{V}(3_1)=4u^{27}$ which is different from
the unknot's value of $4u^9$, and the enhancement detects the difference between
the unknot and the trefoil.
\end{example}
\begin{example}
Let $X$ be the tribracket in example \ref{ex:nt8}. Via \texttt{python}
computation we selected tribracket modules $V_1$ and $V_2$
with coffecients in
$\mathbb{Z}_3$ and $V_3$ with coefficients in $\mathbb{Z}_8$,
\begin{eqnarray*}V_1&=&
\left[\left[\begin{array}{rr}
2 & 1 \\
2 & 2
\end{array}\right],\left[\begin{array}{rr}
2 & 2 \\
1 & 2
\end{array}\right]\right], \quad
\left[\left[\begin{array}{rr}
2 & 2 \\
1 & 1
\end{array}\right],\left[\begin{array}{rr}
1 & 1 \\
2 & 2
\end{array}\right]\right], \\
V_2 & = &
\left[\left[\begin{array}{rr}
1 & 1 \\
1 & 1
\end{array}\right],\left[\begin{array}{rr}
1 & 1 \\
1 & 1
\end{array}\right]\right], \quad
\left[\left[\begin{array}{rr}
1 & 1 \\
2 & 2
\end{array}\right],\left[\begin{array}{rr}
2 & 2 \\
1 & 1
\end{array}\right]\right]
\quad \mathrm{and}\\
V_3 & = &
\left[\left[\begin{array}{rr}
1 & 3 \\
1 & 7
\end{array}\right],\left[\begin{array}{rr}
7 & 1 \\
3 & 1
\end{array}\right]\right], \quad
\left[\left[\begin{array}{rr}
1 & 5 \\
1 & 1
\end{array}\right],\left[\begin{array}{rr}
1 & 1 \\
5 & 1
\end{array}\right]\right],
\end{eqnarray*}
and computed the $\Phi_X^V(L)$ values for the prime links of up to seven
crossings at the knot atlas \cite{KA}. The results are collected in the tables.
\[
\begin{array}{r|l}
\Phi_X^{V_1}(L) & L \\ \hline\hline
2u^9+6u^{27} & L2a1, L4a1, L5a1, L6a2, L7a4, L7a6 \\
2u^9+4u^{27}+2u^{81} & L7a2, L7a3, L7n1, L7n2 \\
8u^{27} & L6a1, L6a3, L7a1, L7a5 \\ \hline
8u^{27}+8u^{81} & L6a5 \\
2u^9+6u^{27}+8u^{81} & L6n1, L7a7 \\
2u^9+14u^{81} & L6a4 \\
\end{array}
\]
\[
\begin{array}{r|l}
\Phi_X^{V_2}(L) & L \\ \hline\hline
6u^9+2u^{27} & L2a1, L62, L7a6 \\
2u^9+6u^{27} & L4a1, L5a1, L7a2, L7a3, L7a4, L7n1, L7n2 \\
4u^9+4u^{27} & L6a3, L7a5 \\
8u^{27} & L6a1, L7a1 \\ \hline
2u^9+6u^{27}+8u^{81} & L6a4 \\
6u^9+8u^{27}+2u^{81} & L6a5 \\
8u^9+6u^{27}+2u^{81} & L6n1, L7a7 \\
\end{array}
\]
\[
\begin{array}{r|l}
\Phi_X^{V_3}(L) & L \\ \hline\hline
2u^{128}+4u^{256}+2u^{512} & L2a1, L6a2, L6a3, L7a5, L7a6 \\
2u^{256}+6u^{512} & L4a1, L6a1, L7a2,L7n1 \\
8u^{512} & L5a1, L7a1, L7a3, L7a4, L7n2 \\\hline
2u^{256}+6u^{1024}+6u^{2048}+2u^{4096} & L6a5, L6n1,L7a7 \\
2u^{1024}+6u^{2048}+8u^{4096} & L6a4
\end{array}
\]
\end{example}
\begin{example}
For our final example let $X$ be the 4-element tribracket
\[
\left[
\left[\begin{array}{rrrr}
4 & 3 & 2 & 1 \\
2 & 4 & 1 & 3 \\
3 & 1 & 4 & 2 \\
1 & 2 & 3 & 4
\end{array}\right],
\left[\begin{array}{rrrr}
3 & 1 & 4 & 2 \\
4 & 3 & 2 & 1 \\
1 & 2 & 3 & 4 \\
2 & 4 & 1 & 3 \\
\end{array}\right],
\left[\begin{array}{rrrr}
2 & 4 & 1 & 3 \\
1 & 2 & 3 & 4 \\
4 & 3 & 2 & 1 \\
3 & 1 & 4 & 2 \\
\end{array}\right],
\left[\begin{array}{rrrr}
1 & 2 & 3 & 4 \\
3 & 1 & 4 & 2 \\
2 & 4 & 1 & 3 \\
4 & 3 & 2 & 1
\end{array}\right]\right]
\]
with the module $V$ with $\mathbb{Z}_3$ coefficients specified by
\[
\left[
\left[\begin{array}{rrrr}
1 & 1 & 1 & 1 \\
2 & 1 & 1 & 2 \\
2 & 1 & 1 & 2 \\
1 & 1 & 1 & 1
\end{array}\right],
\left[\begin{array}{rrrr}
1 & 2 & 2 & 1 \\
1 & 1 & 1 & 1 \\
1 & 1 & 1 & 1 \\
1 & 2 & 2 & 1
\end{array}\right],
\left[\begin{array}{rrrr}
1 & 2 & 2 & 1 \\
1 & 1 & 1 & 1 \\
1 & 1 & 1 & 1 \\
1 & 2 & 2 & 1
\end{array}\right],
\left[\begin{array}{rrrr}
1 & 1 & 1 & 1 \\
2 & 1 & 1 & 2 \\
2 & 1 & 1 & 2 \\
1 & 1 & 1 & 1
\end{array}\right]
\right],
\]\[
\left[
\left[\begin{array}{rrrr}
2 & 2 & 2 & 2 \\
1 & 1 & 1 & 1 \\
1 & 1 & 1 & 1 \\
2 & 2 & 2 & 2 \\
\end{array}\right],
\left[\begin{array}{rrrr}
1 & 1 & 1 & 1 \\
2 & 2 & 2 & 2 \\
2 & 2 & 2 & 2 \\
1 & 1 & 1 & 1
\end{array}\right],
\left[\begin{array}{rrrr}
1 & 1 & 1 & 1 \\
2 & 2 & 2 & 2 \\
2 & 2 & 2 & 2 \\
1 & 1 & 1 & 1
\end{array}\right],
\left[\begin{array}{rrrr}
2 & 2 & 2 & 2 \\
1 & 1 & 1 & 1 \\
1 & 1 & 1 & 1 \\
2 & 2 & 2 & 2
\end{array}\right]
\right].
\]
We computed the invariant on prime knots with up to eight crossings and links
with up to seven crossings. The results are collected in the table. In
particular the knots in the table all have counting invariant value 16 but
are sorted into three classes by the enhancement, while the invariant is quite
effective at distinguishng the links in the table.
\[\begin{array}{r|l}
\Phi_X^V(L) & L \\ \hline
16u^9 & 4_1, 5_1, 5_2, 6_2, 6_3, 7_1, 7_2, 7_3, 7_5, 7_6, 8_1, 8_2, 8_3, 8_4, \\
& 8_6, 8_7, 8_8, 8_9, 8_{12}, 8_{13}, 8_{14}, 8_{16}, 8_{17} \\
8u^9+8u^{27} & 3_1, 6_1, 7_4,7_7, 8_5, 8_{10}, 8_{11}, 8_{15}, 8_{19}, 8_{20}, 8_{21} \\
8u^9+8u^{81} & 8_{18} \\\hline
16u^9+16u^{27} & L2a1, L6a2, L7a6 \\
32u^{27} & L6a3, L7a5\\\hline
16u^9+32u^{27}+16u^{81} & L7a3, L7n1, L7n2 \\
16u^9+48u^{27} & L4a1, L5a1, L7a4 \\
32u^9+32u^{81} & L6n1, L7a7 \\
32u^{27}+32u^{81} & L6a5 \\\hline
64u^{27} & L6a1, L7a1 \\\hline
32u^9+ 224u^{81} & L6a4 \\
\end{array}
\]
\end{example}
\section{Questions}\label{Q}
We conclude with some questions and directions for future work.
First and foremost, more efficient methods than our (relatively) brute-force
axiom testing for finding tribracket modules would be highly desirable. While
even the relatively small examples we have found are fairly good at
distinguishing classical knots and links, we expect that modules over larger
finite rings or infinite rings should yield even stronger invariants.
What is the relationship between tribracket modules and tribracket cocycles?
In the case of racks, rack modules are closely related to structures known
as \textit{dynamical cocycles} \cite{AG}; what is the appropriate definition
for tribracket dynamical cocycles?
As in the case of \cite{CN2}, we can consider tribracket modules with
coefficients in a polynomial algebra as defining a kind of tribracket-colored
Alexander polynomial for each coloring. Do these invariants satisfy skein
relations? Do their coefficients define Vassiliev invariants?
|
\section{Introduction}
In the \emph{nucleated instability\/} (also called core
instability) hypothesis of giant planet
formation, a critical mass for static core envelope
protoplanets has been found. Mizuno (\cite{mizuno}) determined
the critical mass of the core to be about $12 \,M_\oplus$
($M_\oplus=5.975 \times 10^{27}\,\mathrm{g}$ is the Earth mass), which
is independent of the outer boundary
conditions and therefore independent of the location in the
solar nebula. This critical value for the core mass corresponds
closely to the cores of today's giant planets.
Although no hydrodynamical study has been available many workers
conjectured that a collapse or rapid contraction will ensue
after accumulating the critical mass. The main motivation for
this article
is to investigate the stability of the static envelope at the
critical mass. With this aim the local, linear stability of static
radiative gas spheres is investigated on the basis of Baker's
(\cite{baker}) standard one-zone model.
Phenomena similar to the ones described above for giant planet
formation have been found in hydrodynamical models concerning
star formation where protostellar cores explode
(Tscharnuter \cite{tscharnuter}, Balluch \cite{balluch}),
whereas earlier studies found quasi-steady collapse flows. The
similarities in the (micro)physics, i.e., constitutive relations of
protostellar cores and protogiant planets serve as a further
motivation for this study.
\section{Baker's standard one-zone model}
\begin{figure*}
\centering
\caption{Adiabatic exponent $\Gamma_1$.
$\Gamma_1$ is plotted as a function of
$\lg$ internal energy $\mathrm{[erg\,g^{-1}]}$ and $\lg$
density $\mathrm{[g\,cm^{-3}]}$.}
\label{FigGam}%
\end{figure*}
In this section the one-zone model of Baker (\cite{baker}),
originally used to study the Cephe{\"{\i}}d pulsation mechanism, will
be briefly reviewed. The resulting stability criteria will be
rewritten in terms of local state variables, local timescales and
constitutive relations.
Baker (\cite{baker}) investigates the stability of thin layers in
self-gravitating,
spherical gas clouds with the following properties:
\begin{itemize}
\item hydrostatic equilibrium,
\item thermal equilibrium,
\item energy transport by grey radiation diffusion.
\end{itemize}
For the one-zone-model Baker obtains necessary conditions
for dynamical, secular and vibrational (or pulsational)
stability (Eqs.\ (34a,\,b,\,c) in Baker \cite{baker}). Using Baker's
notation:
\[
\begin{array}{lp{0.8\linewidth}}
M_{r} & mass internal to the radius $r$ \\
m & mass of the zone \\
r_0 & unperturbed zone radius \\
\rho_0 & unperturbed density in the zone \\
T_0 & unperturbed temperature in the zone \\
L_{r0} & unperturbed luminosity \\
E_{\mathrm{th}} & thermal energy of the zone
\end{array}
\]
\noindent
and with the definitions of the \emph{local cooling time\/}
(see Fig.~\ref{FigGam})
\begin{equation}
\tau_{\mathrm{co}} = \frac{E_{\mathrm{th}}}{L_{r0}} \,,
\end{equation}
and the \emph{local free-fall time}
\begin{equation}
\tau_{\mathrm{ff}} =
\sqrt{ \frac{3 \pi}{32 G} \frac{4\pi r_0^3}{3 M_{\mathrm{r}}}
}\,,
\end{equation}
Baker's $K$ and $\sigma_0$ have the following form:
\begin{eqnarray}
\sigma_0 & = & \frac{\pi}{\sqrt{8}}
\frac{1}{ \tau_{\mathrm{ff}}} \\
K & = & \frac{\sqrt{32}}{\pi} \frac{1}{\delta}
\frac{ \tau_{\mathrm{ff}} }
{ \tau_{\mathrm{co}} }\,;
\end{eqnarray}
where $ E_{\mathrm{th}} \approx m (P_0/{\rho_0})$ has been used and
\begin{equation}
\begin{array}{l}
\delta = - \left(
\frac{ \partial \ln \rho }{ \partial \ln T }
\right)_P \\
e=mc^2
\end{array}
\end{equation}
is a thermodynamical quantity which is of order $1$ and equal to $1$
for nonreacting mixtures of classical perfect gases. The physical
meaning of $ \sigma_0 $ and $K$ is clearly visible in the equations
above. $\sigma_0$ represents a frequency of the order one per
free-fall time. $K$ is proportional to the ratio of the free-fall
time and the cooling time. Substituting into Baker's criteria, using
thermodynamic identities and definitions of thermodynamic quantities,
\begin{displaymath}
\Gamma_1 = \left( \frac{ \partial \ln P}{ \partial\ln \rho}
\right)_{S} \, , \;
\chi^{}_\rho = \left( \frac{ \partial \ln P}{ \partial\ln \rho}
\right)_{T} \, , \;
\kappa^{}_{P} = \left( \frac{ \partial \ln \kappa}{ \partial\ln P}
\right)_{T}
\end{displaymath}
\begin{displaymath}
\nabla_{\mathrm{ad}} = \left( \frac{ \partial \ln T}
{ \partial\ln P} \right)_{S} \, , \;
\chi^{}_T = \left( \frac{ \partial \ln P}
{ \partial\ln T} \right)_{\rho} \, , \;
\kappa^{}_{T} = \left( \frac{ \partial \ln \kappa}
{ \partial\ln T} \right)_{T}
\end{displaymath}
one obtains, after some pages of algebra, the conditions for
\emph{stability\/} given
below:
\begin{eqnarray}
\frac{\pi^2}{8} \frac{1}{\tau_{\mathrm{ff}}^2}
( 3 \Gamma_1 - 4 )
& > & 0 \label{ZSDynSta} \\
\frac{\pi^2}{\tau_{\mathrm{co}}
\tau_{\mathrm{ff}}^2}
\Gamma_1 \nabla_{\mathrm{ad}}
\left[ \frac{ 1- 3/4 \chi^{}_\rho }{ \chi^{}_T }
( \kappa^{}_T - 4 )
+ \kappa^{}_P + 1
\right]
& > & 0 \label{ZSSecSta} \\
\frac{\pi^2}{4} \frac{3}{\tau_{ \mathrm{co} }
\tau_{ \mathrm{ff} }^2
}
\Gamma_1^2 \, \nabla_{\mathrm{ad}} \left[
4 \nabla_{\mathrm{ad}}
- ( \nabla_{\mathrm{ad}} \kappa^{}_T
+ \kappa^{}_P
)
- \frac{4}{3 \Gamma_1}
\right]
& > & 0 \label{ZSVibSta}
\end{eqnarray}
For a physical discussion of the stability criteria see Baker
(\cite{baker}) or Cox (\cite{cox}).
We observe that these criteria for dynamical, secular and
vibrational stability, respectively, can be factorized into
\begin{enumerate}
\item a factor containing local timescales only,
\item a factor containing only constitutive relations and
their derivatives.
\end{enumerate}
The first factors, depending on only timescales, are positive
by definition. The signs of the left hand sides of the
inequalities~(\ref{ZSDynSta}), (\ref{ZSSecSta}) and (\ref{ZSVibSta})
therefore depend exclusively on the second factors containing
the constitutive relations. Since they depend only
on state variables, the stability criteria themselves are \emph{
functions of the thermodynamic state in the local zone}. The
one-zone stability can therefore be determined
from a simple equation of state, given for example, as a function
of density and
temperature. Once the microphysics, i.e.\ the thermodynamics
and opacities (see Table~\ref{KapSou}), are specified (in practice
by specifying a chemical composition) the one-zone stability can
be inferred if the thermodynamic state is specified.
The zone -- or in
other words the layer -- will be stable or unstable in
whatever object it is imbedded as long as it satisfies the
one-zone-model assumptions. Only the specific growth rates
(depending upon the time scales) will be different for layers
in different objects.
\begin{table}
\caption[]{Opacity sources.}
\label{KapSou}
$$
\begin{array}{p{0.5\linewidth}l}
\hline
\noalign{\smallskip}
Source & T / {[\mathrm{K}]} \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
Yorke 1979, Yorke 1980a & \leq 1700^{\mathrm{a}} \\
Kr\"ugel 1971 & 1700 \leq T \leq 5000 \\
Cox \& Stewart 1969 & 5000 \leq \\
\noalign{\smallskip}
\hline
\end{array}
$$
\end{table}
We will now write down the sign (and therefore stability)
determining parts of the left-hand sides of the inequalities
(\ref{ZSDynSta}), (\ref{ZSSecSta}) and (\ref{ZSVibSta}) and thereby
obtain \emph{stability equations of state}.
The sign determining part of inequality~(\ref{ZSDynSta}) is
$3\Gamma_1 - 4$ and it reduces to the
criterion for dynamical stability
\begin{equation}
\Gamma_1 > \frac{4}{3}\,\cdot
\end{equation}
Stability of the thermodynamical equilibrium demands
\begin{equation}
\chi^{}_\rho > 0, \;\; c_v > 0\, ,
\end{equation}
and
\begin{equation}
\chi^{}_T > 0
\end{equation}
holds for a wide range of physical situations.
With
\begin{eqnarray}
\Gamma_3 - 1 = \frac{P}{\rho T} \frac{\chi^{}_T}{c_v}&>&0\\
\Gamma_1 = \chi_\rho^{} + \chi_T^{} (\Gamma_3 -1)&>&0\\
\nabla_{\mathrm{ad}} = \frac{\Gamma_3 - 1}{\Gamma_1} &>&0
\end{eqnarray}
we find the sign determining terms in inequalities~(\ref{ZSSecSta})
and (\ref{ZSVibSta}) respectively and obtain the following form
of the criteria for dynamical, secular and vibrational
\emph{stability}, respectively:
\begin{eqnarray}
3 \Gamma_1 - 4 =: S_{\mathrm{dyn}} > & 0 & \label{DynSta} \\
\frac{ 1- 3/4 \chi^{}_\rho }{ \chi^{}_T } ( \kappa^{}_T - 4 )
+ \kappa^{}_P + 1 =: S_{\mathrm{sec}} > & 0 & \label{SecSta} \\
4 \nabla_{\mathrm{ad}} - (\nabla_{\mathrm{ad}} \kappa^{}_T
+ \kappa^{}_P)
- \frac{4}{3 \Gamma_1} =: S_{\mathrm{vib}}
> & 0\,.& \label{VibSta}
\end{eqnarray}
The constitutive relations are to be evaluated for the
unperturbed thermodynamic state (say $(\rho_0, T_0)$) of the zone.
We see that the one-zone stability of the layer depends only on
the constitutive relations $\Gamma_1$,
$\nabla_{\mathrm{ad}}$, $\chi_T^{},\,\chi_\rho^{}$,
$\kappa_P^{},\,\kappa_T^{}$.
These depend only on the unperturbed
thermodynamical state of the layer. Therefore the above relations
define the one-zone-stability equations of state
$S_{\mathrm{dyn}},\,S_{\mathrm{sec}}$
and $S_{\mathrm{vib}}$. See Fig.~\ref{FigVibStab} for a picture of
$S_{\mathrm{vib}}$. Regions of secular instability are
listed in Table~1.
\begin{figure}
\centering
\caption{Vibrational stability equation of state
$S_{\mathrm{vib}}(\lg e, \lg \rho)$.
$>0$ means vibrational stability.
}
\label{FigVibStab}
\end{figure}
\section{Conclusions}
\begin{enumerate}
\item The conditions for the stability of static, radiative
layers in gas spheres, as described by Baker's (\cite{baker})
standard one-zone model, can be expressed as stability
equations of state. These stability equations of state depend
only on the local thermodynamic state of the layer.
\item If the constitutive relations -- equations of state and
Rosseland mean opacities -- are specified, the stability
equations of state can be evaluated without specifying
properties of the layer.
\item For solar composition gas the $\kappa$-mechanism is
working in the regions of the ice and dust features
in the opacities, the $\mathrm{H}_2$ dissociation and the
combined H, first He ionization zone, as
indicated by vibrational instability. These regions
of instability are much larger in extent and degree of
instability than the second He ionization zone
that drives the Cephe{\"\i}d pulsations.
\end{enumerate}
\begin{acknowledgements}
Part of this work was supported by the German
\emph{Deut\-sche For\-schungs\-ge\-mein\-schaft, DFG\/} project
number Ts~17/2--1.
\end{acknowledgements}
\section*{Appendix A}
\label{sec:appA}
FACT and MAGIC are located within 100~m of each other and therefore
observe under exactly the same atmospheric conditions. The
instrumentation used in these two telescopes is different (e.g. FACT
uses SiPMs as light detectors, instead of PMTs), and the observation
mode (stereo vs mono) and analysis chains are completely separate.
It is interesting and useful therefore to compare the two experiments' light curves.
Figure~\ref{fig:corFACT} shows the MAGIC light curve above 1 TeV superimposed with
the FACT light curve. The scale is chosen so that for the night of the highest flux the points
overlap. As FACT monitors Mrk~501 every possible night, the FACT light curve is
more densely sampled than that of MAGIC.
The right panel of Figure~\ref{fig:corFACT} shows that there is an
excellent agreement between the MAGIC VHE fluxes above 1 TeV and the
excess rates measured with FACT.
The flux-flux plot is fit to a first order polynomial resulting in a $\chi^{2}$/DoF of 10.4/10,
with a slope of (8.51$\pm$0.61)$\times$10$^{-13}$ cm$^{-2}$ s$^{-1}$ hr$^{-1}$ and an offset of
($-$0.0021$\pm$0.0021)$\times$10$^{-10}$ cm$^{-2}$ s$^{-1}$ hr$^{-1}$ , which is consistent with zero, as expected.
This function was then used to normalize the FACT excess rate to the
fluxes (above 1 TeV) reported in the bottom panel of Figure~\ref{fig:lightcurvecombinedTeV}.
\begin{figure}[h]
\centering
\includegraphics[width=25pc]{figures3/magic_fact_lc_high.eps}
\includegraphics[width=25pc]{figures3/magic_fact_corr_high.eps}
\caption{Top panel:
superposition of the FACT excess rates and MAGIC light curve above 1 TeV.
Bottom panel: correlation between the FACT excess rate and the MAGIC flux
above 1~TeV on the 12 days when data were taken by both instruments.
}
\label{fig:corFACT}
\end{figure}
\section*{Appendix B}
\label{sec:appB}
This section reports the spectral parameters resulting from the fits to the X-ray and gamma-ray spectra.
\begin{table}[h]
\caption{
Parameters resulting from the fit with a power law F(E) =
N$_{0}$(E/0.5~TeV)$^{-{\Gamma}}$ to the measured MAGIC spectra shown
in Figures~\ref{fig:index_xrt_pl}, \ref{fig:sedall}, \ref{fig:sedall2}, and \ref{fig:sed}.
The fitted spectra were EBL corrected using \citet{2008A&A...487..837F}. }
\label{tab:magicIndex}
\centering
\begin{tabular}{l c c c }
\hline\hline
MJD & N$_\text{{o}}$ [10$^{-11}$ s$^{-1}$ cm$^{-2}$ TeV$^{-1}$] & $\Gamma$ & $\chi^{2}$/dof \\
\hline
56007 & ~1.90$\pm$0.25 & 1.85$\pm$0.25 &~0.70/3 \\
56032 & ~2.39$\pm$0.23 & 1.89$\pm$0.15 &~1.49/3 \\
56036 & ~4.36$\pm$0.30 & 1.88$\pm$0.08 &~0.33/4 \\
56040 & ~2.46$\pm$0.23 & 2.31$\pm$0.18 &~1.36/3 \\
56070 & ~5.15$\pm$0.25 & 1.94$\pm$0.08 &~1.99/3 \\
56071 & ~2.85$\pm$0.18 & 2.22$\pm$0.09 &~6.76/5 \\
56072 & ~5.60$\pm$0.20 & 2.02$\pm$0.04 &~8.94/6 \\
56073 & ~4.73$\pm$0.23 & 2.20$\pm$0.07 &18.37/6 \\
56074 & ~2.41$\pm$0.18 & 2.01$\pm$0.09 &~5.63/5 \\
56075 & ~1.83$\pm$0.15 & 2.11$\pm$0.11 &~5.49/5 \\
56076 & ~2.24$\pm$0.23 & 2.20$\pm$0.15 &~6.79/5 \\
56087 & 15.43$\pm$0.31 & 2.02$\pm$0.03 &96.26/6 \\
56093 & ~2.50$\pm$0.21 & 2.15$\pm$0.12 &~4.26/5 \\
56094 & ~2.60$\pm$0.17 & 2.24$\pm$0.09 &~3.09/5 \\
56095 & ~2.75$\pm$0.21 & 2.05$\pm$0.10 &~2.70/5 \\
56096 & ~7.15$\pm$0.30 & 2.09$\pm$0.05 &12.94/6 \\
56097 & ~6.17$\pm$0.29 & 2.20$\pm$0.07 &~5.82/5 \\
\hline
\end{tabular}
\end{table}
\begin{table}[h]
\caption{
Parameters resulting from the fit with a power law F(E) =
N$_{0}$(E/0.5~TeV)$^{-{\Gamma}}$ to the measured VERITAS spectra shown
in Figures~\ref{fig:index_xrt_pl}, \ref{fig:sedall}, and \ref{fig:sedall2}.
The fitted spectra were EBL corrected using \citet{2008A&A...487..837F}. }
\label{tab:veritasIndex}
\centering
\begin{tabular}{l c c c }
\hline\hline
MJD & N$_\text{{o}}$ [10$^{-11}$ s$^{-1}$ cm$^{-2}$ TeV$^{-1}$] & $\Gamma$ & $\chi^{2}$/dof \\
\hline
56007 & 1.08$\pm$0.09 & 2.05$\pm$0.10 & 10.73/6 \\
56009 & 2.26$\pm$0.19 & 1.99$\pm$0.09 & 20.46/7 \\
56015 & 1.82$\pm$0.17 & 1.75$\pm$0.09 & ~6.80/6 \\
56034 & 0.89$\pm$0.11 & 1.59$\pm$0.21 & 10.82/6 \\
56038 & 0.58$\pm$0.09 & 1.91$\pm$0.24 & ~9.01/5 \\
56046 & 0.65$\pm$0.08 & 1.80$\pm$0.18 & ~4.85/6 \\
56050 & 0.70$\pm$0.08 & 2.09$\pm$0.23 & ~5.39/6 \\
56061 & 0.71$\pm$0.08 & 2.19$\pm$0.19 & ~9.15/4 \\
56066 & 1.12$\pm$0.09 & 2.10$\pm$0.09 & ~1.74/6 \\
56069 & 1.05$\pm$0.08 & 2.02$\pm$0.13 & ~4.05/4 \\
56073 & 2.27$\pm$0.12 & 2.08$\pm$0.05 & 10.22/5 \\
56075 & 2.52$\pm$0.45 & 2.12$\pm$0.22 & ~3.97/2 \\
56076 & 1.35$\pm$0.14 & 1.94$\pm$0.11 & ~1.73/3 \\
56077 & 1.62$\pm$0.28 & 1.23$\pm$0.26 & ~0.42/3 \\
56090 & 2.08$\pm$0.40 & 1.76$\pm$0.29 & 12.39/4 \\
56092 & 3.57$\pm$0.30 & 2.48$\pm$0.19 & 11.57/6 \\
56093 & 1.46$\pm$0.35 & 1.43$\pm$0.23 & ~3.39/2 \\
56094 & 2.14$\pm$0.31 & 1.81$\pm$0.29 & ~2.42/3 \\
56095 & 2.06$\pm$0.30 & 1.95$\pm$0.24 & ~2.22/4 \\
56096 & 2.55$\pm$0.31 & 1.54$\pm$0.26 & ~5.07/3 \\
56097 & 1.76$\pm$0.23 & 1.91$\pm$0.17 & ~8.97/3 \\
56099 & 0.85$\pm$0.15 & 1.66$\pm$0.20 & ~1.16/2 \\
\hline
\end{tabular}
\end{table}
\begin{table*}[h]
\caption{\textit{Fermi}--LAT flux and power-law index above 0.2~GeV
from the gamma-ray spectra shown
in Figures~\ref{fig:sedall}, \ref{fig:sedall2}, and \ref{fig:sed}.}
\label{tab:fermiIndex}
\centering
\begin{tabular}{c c c c c}
\hline\hline
Observation [MJD] & Flux [10$^{-8}$ cm$^{-2}$ s$^{-1}$] & $\Gamma$ & TS \\
\hline
56007.5--56010.5 & ~6.3$\pm$2.1 & 1.7$\pm$0.2 & 43 \\
56013.5--56016.5 & ~7.3$\pm$2.4 & 1.8$\pm$0.2 & 61 \\
56030.5--56033.5 & ~4.9$\pm$2.0 & 1.9$\pm$0.3 & 34 \\
56032.5--56035.5 & ~3.6$\pm$1.9 & 1.9$\pm$0.4 & 19 \\
56034.5--56037.5 & ~5.0$\pm$2.7 & 2.4$\pm$0.6 & 16 \\
56036.5--56039.5 & ~2.2$\pm$1.2 & 1.4$\pm$0.3 & 33 \\
56038.5--56041.5 & ~2.9$\pm$1.9 & 1.6$\pm$0.3 & 28 \\
56044.5--56047.5 & ~8.4$\pm$2.7 & 2.0$\pm$0.2 & 52 \\
56059.5--56062.5 & ~1.1$\pm$1.1 & 1.8$\pm$0.7 & ~4 \\
56064.5--56067.5 & ~7.6$\pm$2.7 & 2.1$\pm$0.3 & 40 \\
56071.5--56074.5 & ~5.7$\pm$1.9 & 1.7$\pm$0.2 & 56 \\
56074.5--56077.5 & ~2.3$\pm$1.3 & 1.4$\pm$0.3 & 24 \\
56075.5--56078.5 & ~4.0$\pm$1.6 & 1.5$\pm$0.2 & 41 \\
56086.5--56087.5 & 12.0$\pm$5.0 & 1.5$\pm$0.2 & 59 \\
56088.5--56091.5 & ~8.4$\pm$2.9 & 1.7$\pm$0.2 & 71 \\
56092.5--56095.5 & ~8.1$\pm$2.5 & 1.8$\pm$0.2 & 65 \\
56093.5--56096.5 & ~7.5$\pm$2.3 & 1.7$\pm$0.2 & 72 \\
\hline
\end{tabular}
\end{table*}
\begin{table*}[h]
\caption{Spectral models (log-parabola and power-law functions) fitted to the \textit{Swift}--XRT data used
in Figures~\ref{fig:index_xrt_pl}, \ref{fig:sedall}, \ref{fig:sedall2}, and \ref{fig:sed}.}
\label{tab:xrtIndex}
\centering
\begin{tabular}{l c c c c c c c}
\hline\hline
Start Time & $\alpha$ & $\beta$ & $\chi^{2}$/dof & 0.3--2~keV & 2--10~keV & $\alpha$ & $\chi^{2}$/dof \\
MJD & & & & [10$^{-10}$ erg cm$^{-2}$ s$^{-1}$] & [10$^{-10}$ erg cm$^{-2}$ s$^{-1}$] & & \\
\hline
55972.079 & 1.532$\pm$0.052 & 0.356$\pm$0.094 & 189/221 & 1.545$\pm$0.024 & 2.062$\pm$0.069 & 1.687$\pm$0.028 & 233/222 \\
55977.296 & 1.624$\pm$0.053 & 0.411$\pm$0.103 & 226/205 & 1.144$\pm$0.018 & 1.255$\pm$0.046 & 1.778$\pm$0.028 & 277/206 \\
55981.242 & 1.694$\pm$0.055 & 0.282$\pm$0.107 & 176/198 & 1.041$\pm$0.019 & 1.147$\pm$0.047 & 1.809$\pm$0.033 & 196/199 \\
55985.247 & 1.757$\pm$0.052 & 0.261$\pm$0.102 & 191/191 & 1.427$\pm$0.026 & 1.443$\pm$0.054 & 1.862$\pm$0.033 & 210/192 \\
55989.256 & 1.669$\pm$0.047 & 0.237$\pm$0.093 & 211/219 & 1.346$\pm$0.021 & 1.609$\pm$0.051 & 1.761$\pm$0.028 & 230/220 \\
55994.206 & 1.712$\pm$0.046 & 0.347$\pm$0.092 & 190/222 & 1.306$\pm$0.021 & 1.316$\pm$0.044 & 1.849$\pm$0.028 & 233/223 \\
56000.142 & 1.634$\pm$0.047 & 0.402$\pm$0.091 & 270/229 & 1.650$\pm$0.023 & 1.797$\pm$0.057 & 1.804$\pm$0.026 & 329/230 \\
56005.166 & 1.638$\pm$0.055 & 0.363$\pm$0.108 & 208/195 & 1.628$\pm$0.029 & 1.822$\pm$0.074 & 1.782$\pm$0.033 & 243/196 \\
56009.429 & 1.662$\pm$0.047 & 0.267$\pm$0.072 & 221/229 & 1.374$\pm$0.021 & 1.617$\pm$0.054 & 1.776$\pm$0.027 & 247/230 \\
56011.450 & 1.519$\pm$0.049 & 0.378$\pm$0.088 & 264/245 & 1.447$\pm$0.020 & 1.930$\pm$0.054 & 1.691$\pm$0.026 & 320/246 \\
56015.451 & 1.606$\pm$0.055 & 0.321$\pm$0.095 & 203/217 & 1.408$\pm$0.023 & 1.725$\pm$0.055 & 1.758$\pm$0.029 & 237/218 \\
56017.316 & 1.685$\pm$0.051 & 0.214$\pm$0.095 & 204/220 & 1.370$\pm$0.024 & 1.629$\pm$0.057 & 1.780$\pm$0.029 & 219/221 \\
56019.188 & 1.714$\pm$0.044 & 0.250$\pm$0.087 & 232/230 & 1.257$\pm$0.020 & 1.380$\pm$0.043 & 1.812$\pm$0.027 & 256/231 \\
56019.457 & 1.623$\pm$0.051 & 0.342$\pm$0.099 & 177/202 & 1.186$\pm$0.018 & 1.387$\pm$0.053 & 1.755$\pm$0.029 & 214/203 \\
56023.080 & 1.760$\pm$0.059 & 0.260$\pm$0.116 & 185/181 & 1.095$\pm$0.024 & 1.104$\pm$0.044 & 1.866$\pm$0.034 & 200/182 \\
56027.548 & 1.767$\pm$0.057 & 0.302$\pm$0.113 & 198/187 & 1.173$\pm$0.024 & 1.125$\pm$0.044 & 1.890$\pm$0.033 & 220/188 \\
56032.174 & 1.664$\pm$0.054 & 0.297$\pm$0.100 & 183/215 & 1.330$\pm$0.025 & 1.518$\pm$0.054 & 1.796$\pm$0.030 & 209/216 \\
56034.249 & 1.607$\pm$0.054 & 0.298$\pm$0.103 & 219/190 & 1.080$\pm$0.019 & 1.350$\pm$0.053 & 1.728$\pm$0.031 & 244/191 \\
56036.043 & 1.590$\pm$0.057 & 0.305$\pm$0.104 & 176/208 & 1.229$\pm$0.021 & 1.568$\pm$0.055 & 1.727$\pm$0.031 & 202/209 \\
56038.376 & 1.677$\pm$0.058 & 0.397$\pm$0.112 & 186/189 & 1.003$\pm$0.019 & 1.024$\pm$0.040 & 1.846$\pm$0.033 & 223/190 \\
56040.123 & 1.710$\pm$0.054 & 0.313$\pm$0.109 & 213/196 & 1.169$\pm$0.022 & 1.218$\pm$0.047 & 1.836$\pm$0.033 & 237/197 \\
56042.395 & 1.795$\pm$0.061 & 0.258$\pm$0.125 & 162/167 & 1.199$\pm$0.028 & 1.144$\pm$0.051 & 1.894$\pm$0.038 & 174/168 \\
56044.005 & 1.706$\pm$0.075 & 0.326$\pm$0.135 & 119/165 & 1.100$\pm$0.025 & 1.142$\pm$0.049 & 1.857$\pm$0.039 & 136/166 \\
56046.412 & 1.691$\pm$0.057 & 0.298$\pm$0.111 & 197/193 & 1.108$\pm$0.021 & 1.208$\pm$0.048 & 1.815$\pm$0.034 & 218/194 \\
56047.413 & 1.673$\pm$0.057 & 0.265$\pm$0.099 & 200/205 & 1.128$\pm$0.021 & 1.306$\pm$0.045 & 1.793$\pm$0.031 & 219/206 \\
56048.148 & 1.574$\pm$0.057 & 0.323$\pm$0.111 & 195/196 & 1.020$\pm$0.019 & 1.312$\pm$0.053 & 1.711$\pm$0.033 & 220/197 \\
56053.089 & 1.678$\pm$0.054 & 0.296$\pm$0.107 & 186/192 & 1.053$\pm$0.019 & 1.175$\pm$0.044 & 1.796$\pm$0.033 & 208/193 \\
56054.947 & 1.732$\pm$0.059 & 0.341$\pm$0.112 & 167/185 & 1.111$\pm$0.022 & 1.090$\pm$0.044 & 1.877$\pm$0.034 & 193/186 \\
56056.029 & 1.624$\pm$0.056 & 0.321$\pm$0.085 & 181/197 & 1.067$\pm$0.019 & 1.270$\pm$0.047 & 1.764$\pm$0.032 & 208/198 \\
56059.358 & 1.712$\pm$0.054 & 0.226$\pm$0.102 & 171/198 & 1.069$\pm$0.019 & 1.203$\pm$0.043 & 1.807$\pm$0.032 & 185/199 \\
56061.296 & 1.700$\pm$0.057 & 0.289$\pm$0.107 & 178/187 & 1.115$\pm$0.022 & 1.207$\pm$0.047 & 1.825$\pm$0.034 & 199/188 \\
56066.310 & 1.578$\pm$0.070 & 0.509$\pm$0.144 & 164/130 & 1.645$\pm$0.035 & 1.779$\pm$0.093 & 1.770$\pm$0.039 & 202/131 \\
56073.333 & 1.576$\pm$0.047 & 0.225$\pm$0.088 & 224/236 & 1.355$\pm$0.020 & 1.908$\pm$0.060 & 1.671$\pm$0.027 & 243/237 \\
56074.059 & 1.628$\pm$0.062 & 0.168$\pm$0.113 & 183/184 & 1.330$\pm$0.028 & 1.817$\pm$0.074 & 1.704$\pm$0.035 & 190/185 \\
56076.021 & 1.565$\pm$0.050 & 0.336$\pm$0.095 & 189/224 & 1.371$\pm$0.022 & 1.771$\pm$0.057 & 1.709$\pm$0.028 & 226/225 \\
56077.273 & 1.608$\pm$0.048 & 0.274$\pm$0.088 & 212/232 & 1.451$\pm$0.024 & 1.850$\pm$0.061 & 1.729$\pm$0.027 & 241/233 \\
56078.208 & 1.565$\pm$0.065 & 0.383$\pm$0.122 & 172/166 & 1.311$\pm$0.026 & 1.619$\pm$0.068 & 1.723$\pm$0.035 & 202/167 \\
56078.423 & 1.636$\pm$0.056 & 0.236$\pm$0.100 & 200/215 & 1.316$\pm$0.024 & 1.661$\pm$0.058 & 1.745$\pm$0.031 & 216/216 \\
56086.973 & 1.538$\pm$0.051 & 0.190$\pm$0.090 & 249/241 & 2.055$\pm$0.031 & 3.192$\pm$0.103 & 1.627$\pm$0.027 & 262/242 \\
56089.047 & 1.637$\pm$0.053 & 0.339$\pm$0.101 & 208/216 & 2.187$\pm$0.037 & 2.506$\pm$0.084 & 1.786$\pm$0.030 & 241/217 \\
56090.049 & 1.542$\pm$0.045 & 0.291$\pm$0.083 & 229/255 & 1.591$\pm$0.022 & 2.225$\pm$0.063 & 1.669$\pm$0.025 & 265/256 \\
56090.923 & 1.660$\pm$0.046 & 0.327$\pm$0.088 & 207/226 & 1.404$\pm$0.022 & 1.568$\pm$0.051 & 1.790$\pm$0.027 & 248/227 \\
56093.988 & 1.704$\pm$0.086 & 0.198$\pm$0.155 & 105/105 & 0.954$\pm$0.028 & 1.117$\pm$0.061 & 1.796$\pm$0.049 & 110/106 \\
56095.050 & 1.543$\pm$0.050 & 0.321$\pm$0.095 & 236/235 & 1.532$\pm$0.022 & 2.078$\pm$0.069 & 1.679$\pm$0.028 & 271/236 \\
56101.280 & 1.528$\pm$0.056 & 0.384$\pm$0.106 & 246/214 & 1.620$\pm$0.025 & 2.120$\pm$0.075 & 1.697$\pm$0.030 & 285/215 \\
56103.281 & 1.563$\pm$0.045 & 0.273$\pm$0.082 & 261/261 & 1.555$\pm$0.023 & 2.137$\pm$0.061 & 1.689$\pm$0.025 & 294/262 \\
56110.359 & 1.555$\pm$0.077 & 0.340$\pm$0.147 & 108/122 & 1.081$\pm$0.025 & 1.411$\pm$0.069 & 1.700$\pm$0.042 & 124/123 \\
56116.297 & 1.803$\pm$0.062 & 0.110$\pm$0.119 & 165/163 & 1.040$\pm$0.025 & 1.120$\pm$0.054 & 1.849$\pm$0.038 & 168/164 \\
56131.468 & 1.538$\pm$0.057 & 0.293$\pm$0.105 & 171/198 & 1.296$\pm$0.023 & 1.820$\pm$0.068 & 1.665$\pm$0.032 & 194/199 \\
56138.152 & 1.438$\pm$0.042 & 0.179$\pm$0.072 & 281/288 & 1.807$\pm$0.023 & 3.348$\pm$0.082 & 1.524$\pm$0.023 & 299/289 \\
55959.895 & 1.562$\pm$0.047 & 0.321$\pm$0.084 & 240/247 & 1.494$\pm$0.021 & 1.964$\pm$0.057 & 1.709$\pm$0.025 & 283/248 \\
55966.582 & 1.652$\pm$0.054 & 0.280$\pm$0.103 & 161/190 & 1.501$\pm$0.028 & 1.774$\pm$0.067 & 1.764$\pm$0.032 & 183/191 \\
\hline
\end{tabular}
\end{table*}
\section{Summary and conclusion}
\label{sec:conclussion}
We have presented the results from the 2012 Mrk~501 multiwavelength campaign.
An excellent set of data was taken using more than 25 instruments over the period
covering March to June 2012. The source flux was observed
to vary between 0.5 and 4.9~CU above 1~TeV, with an average flux of $\sim$1~CU.
The highest VHE gamma-ray flux was observed on a single night, June 9.
This outburst was also observed in the X-ray band by the \textit{Swift}/XRT instrument.
The fractional variability was seen to increase as a function of energy
and peak in the VHE regime.
This is similar to what has been seen in other multiwavelength campaigns targeting
Mrk~501, but different to the behavior of Mrk~421 which peaks at X-ray energies.
This remains the case even when the flare information is removed or when we consider
only data taken simultaneously in the X-ray and VHE bands, thereby
underlining the difference already observed between these two sources.
Investigating possible correlations between X-ray and VHE bands,
in two energy ranges each, a maximum Pearson correlation coefficient
of 4.9$\sigma$ was observed for energies above 2~keV and 1~TeV.
The significance of this correlation drops to 2.5$\sigma$ when the flare day is excluded.
A further search for correlation
using a discrete correlation function found no evidence for a time lag between X-rays and VHE
gamma rays.
Interestingly, the X-ray and VHE power-law index corresponded to an extremely hard spectrum during
the entire three-month period.
The VHE spectral index above 0.2 TeV, as observed by both MAGIC
and VERITAS, was around $\sim$2,
compared to the typical power-law index of
2.5 \citep{2011ApJ...729....2A,2011ApJ...727..129A,2015A&A...573A..50A}.
The source did not show the previously observed harder-when-brighter
behavior at VHE energies.
In the X-ray domain, Mrk~501 showed a hardening of the spectral shape
with increasing X-ray flux, but the X-ray power-law spectral
index was always less than 2 (for both low and high activity). Therefore,
the synchrotron peak was located above 5 keV and the inverse-Compton peak above 0.5 TeV,
making Mrk 501 an extreme HBL during the entire observing period.
This suggests that being an EHBL is a temporary state of the source, instead of an instrinsic characteristic.
We were able to form 17 SEDs, where the time difference between X-ray and VHE
data taking was less than four hours.
The X-ray and VHE data were modeled using a one-zone SSC model; however,
the model underestimated the amount of optical and UV radiation required to fit the observed light curves.
During 2012 the optical light curve was observed to be at a 10-year low and we assume that this component
comes from a different region.
The emission at GeV is systematically (within 1--2$\sigma$) above the 17 SSC model curves, two
showing a data-model difference of $\sim$3$\sigma$, which may be interpreted as a hint for the existence of an
additional contribution at GeV energies, and is not considered in the current theoretical scenario.
A two-zone model was also used in order to model the flare day of \mbox{June~9 (MJD~56087).}
The two-zone SSC scenario improved the data-model agreement with respect to the one-zone SSC model and theoretical assumptions,
although it still does not describe the broadband data well.
Despite the caveats mentioned above, the one-zone SSC framework provides a reasonable description of the segments
of the SED where most energy is emitted, and where most of the variability occurs.
The direct relation between the electron energy density
and the gamma-ray activity shows that most of
the variability can be explained by injection of high-energy electrons.
The very hard EED (\textit{p}$_\text{{1}}$<2) obtained in our fits, together with the
trend of lower magnetic field strength $B$ for higher electron energy
densities $U_e$, suggests that magnetic reconnection plays a dominant
role in the acceleration of the particles.
However, we cannot exclude scenarios that incorporate other mechanisms
for acceleration of the radiating (high-energy) particle population,
such as efficient shock drift acceleration, or second-order Fermi
acceleration if the electrons can be effectively trapped in regions of
strong turbulence.
\section{Correlation between the X-ray and VHE gamma-ray emission}
\label{sec:corr}
\begin{figure*}
\centering
\includegraphics[width=20pc]{figures3/flux_xrt_1_vhe_1.eps}
\includegraphics[width=20pc]{figures3/flux_xrt_1_vhe_2.eps}
\includegraphics[width=20pc]{figures3/flux_xrt_2_vhe_1.eps}
\includegraphics[width=20pc]{figures3/flux_xrt_2_vhe_2.eps}
\caption{VHE flux as a function of the \textit{Swift}/XRT flux for the energy ranges shown.
Open circles represent data that were taken within 12 hours of each other, red circles within
6 hours and blue circles within 3 hours.
In each case, linear fits to the closed
circle points (6 hours or less) are depicted with a red
line (when considering the June 9 flare) and with a
dotted-dashed grey line (when excluding the June 9
flare).
}
\label{fig:corrtev}
\end{figure*}
\begin{table*}
\caption{Correlation Results: VHE vs X-ray flux. See Section \ref{sec:corr} and
Figure \ref{fig:corrtev}.
The normalised slope is the gradient of the fit in Figure \ref{fig:corrtev}, divided by the
ratio of the average of each distribution, in order to create a dimensionless scaling factor.
Pearson correlation function 1~$\sigma$ errors and the
significance of the correlation are calculated following \citet{2002nrca.book.....P}.
Discrete correlation function (DCF) and errors are calculated as prescribed in \cite{1988ApJ...333..646E}.
}
\label{tab:Corr1}
\centering
\begin{tabular}{l | l | l }
\hline
& VHE (0.2 -- 1~TeV) & VHE ($>$ 1~TeV) \\
\hline & \begin{tabular}{l l l} Normalized & Pearson correlation& \\
Slope of fit & coefficient ($\sigma$) & DCF \\
\end{tabular}
& \begin{tabular}{l l l} Normalized & Pearson correlation& \\
Slope of fit & coefficient ($\sigma$) & DCF \\
\end{tabular} \\
\hline \textit{Swift}/XRT (0.3 -- 2~keV) & \begin{tabular}{l l l}
4.34$\pm$0.19 & \hspace{.05cm} 0.76$^{+0.10}_{-0.15}$ (3.7) \hspace{0.1cm} & 0.72$\pm$0.59 \\
\end{tabular} & \begin{tabular}{l l l}
4.14$\pm$0.27 & \hspace{.05cm} 0.78$^{+0.10}_{-0.15}$ (3.9) \hspace{0.1cm} & 0.74$\pm$0.59 \\
\end{tabular} \\
Excluding Flare & \begin{tabular}{l l l}
1.01$\pm$0.21 & \hspace{.05cm} 0.38$^{+0.24}_{-0.30}$ (1.4) \hspace{0.1cm} & 0.37$\pm$0.14 \\
\end{tabular} & \begin{tabular}{l l l}
1.28$\pm$0.25 & \hspace{.05cm} 0.39$^{+0.23}_{-0.29}$ (1.6) \hspace{0.1cm} & 0.42$\pm$0.17 \\
\end{tabular} \\
\textit{Swift}/XRT (2 -- 10~keV) & \begin{tabular}{l l l}
2.57$\pm$0.13 & \hspace{.05cm} 0.87$^{+0.06}_{-0.10}$ (4.7) \hspace{0.1cm} & 0.81$\pm$0.64 \\
\end{tabular} & \begin{tabular}{l l l}
2.72$\pm$0.16 & \hspace{.05cm} 0.88$^{+0.06}_{-0.10}$ (4.9) \hspace{0.1cm} & 0.83$\pm$0.64 \\
\end{tabular} \\
Excluding Flare & \begin{tabular}{l l l}
1.64$\pm$0.20 & \hspace{.05cm} 0.56$^{+0.18}_{-0.25}$ (2.2) \hspace{0.1cm} & 0.54$\pm$0.21 \\
\end{tabular} & \begin{tabular}{l l l}
1.66$\pm$0.21 & \hspace{.05cm} 0.59$^{+0.16}_{-0.24}$ (2.5) \hspace{0.1cm} & 0.60$\pm$0.20 \\
\end{tabular} \\
\hline
\end{tabular}
\end{table*}
This section focuses on the cross-correlation between the X-ray and
VHE emission, which are the energy bands with the largest variability
in the emission of Mrk~501 (as shown in Figure~\ref{fig:fvar_all}).
Figure \ref{fig:corrtev} shows the integral flux for the two VHE ranges,
0.2--1~TeV and $>$1~TeV,
plotted against that for the two \textit{Swift}/XRT flux bands, 0.3--2~keV and 2--10~keV.
The symbols are color-coded depending on the time difference
between the observations: 3, 6 or 12 hours. The correlation studies
are performed with data taken within 6 hours (the red and blue
symbols), which is approximately the largest temporal coverage provided by a Cherenkov
telescope for one source during one night.
\begin{figure*}
\centering
\includegraphics[width=20pc]{figures3/xrt_flux210_index_v2.eps}
\includegraphics[width=20pc]{figures3/Mrk501_2012_index_flux_500.eps}
\caption{Left Panel: \textit{Swift}/XRT X-ray power-law spectral index vs flux in the 2-10 keV band.
Right Panel: Measured VHE power-law spectral index vs VHE
Flux above 0.2 TeV. Red points represent VERITAS data and
black points MAGIC data. The data are EBL corrected using \citet{2008A&A...487..837F}.
The \textit{Swift}/XRT spectrum of Mrk\,501 is often curved, and can
be described at keV energies with a spectral index that is typically between
1.8 and 2.1, while the VHE spectral index measured with MAGIC and
VERITAS during typical non-flaring activity is about 2.5
\citep{2011ApJ...727..129A,2015A&A...573A..50A}. The typical spectral
indices at X-ray and VHE are marked with a dashed line. For
comparison purposes, the panels depict with solid lines the result of a fit with a
constant to the X-ray and VHE spectral indices.
}
\label{fig:index_xrt_pl}
\end{figure*}
Three methods were used to test for correlation in each of the four panels shown
in Figure \ref{fig:corrtev}, and the results are shown in Table~\ref{tab:Corr1}. A Pearson's correlation test
was applied to the data and a maximum correlation of 4.9$\sigma$
is found
between the higher-energy component of the X-ray band (2--10~keV) and the higher-energy
component of the VHE band ( $\textgreater$ 1~TeV).
However, this falls to 2.5~$\sigma$ when the day of the flare is removed.
We also quantified the correlations using the discrete correlation
function \citep[DCF,][]{1988ApJ...333..646E} which has the advantage
over the Pearson correlation that
the errors in the individual flux measurements (which contribute to the dispersion in the flux values)
are naturally taken into account.
Using the data shown in Figure \ref{fig:corrtev} the correlation for
the two higher-energy bands of the X-ray and VHE light curves yields
0.83$\pm$0.64 when using all data, and 0.60$\pm$0.20
after removing the June 9 flare. The three-times-larger error in the DCF
when the big VHE flare is included is due to the fact that the error in the DCF is
given by the dispersion in the individual (for a given pair of
X-ray/VHE data points) unbinned discrete correlation function,
and this single (flaring) data point
deviates substantially from the behavior of the others.
The DCF value for the data without the flaring activity corresponds to a
marginal correlation at the level of 3$\sigma$, which
is consistent with the Pearson
correlation analysis.
The DCF method is often used to look for a time delay between the
emission at different wavelengths.
Such a search was carried out for the two X-ray and VHE gamma-ray bands, and no significant delay was found.
Neither a linear (shown in
Figure~\ref{fig:corrtev}) nor a quadratic fit function describes the data
well; the linear fit of the highest-energy component in each band,
gives a $\chi^{2}$/DoF of 148.7/15.
In summary, this correlation study yields only a marginal correlation,
which is greatest when comparing the high-energy X-ray and higher-energy VHE gamma-ray components.
In Figure \ref{fig:index_xrt_pl} we present the correlation between the
spectral index (derived from a power-law fit) and the integral flux for both
the \textit{Swift}/XRT and VHE data. The spectral fit results
with power-law functions are reported in Tables~\ref{tab:magicIndex},
\ref{tab:veritasIndex}, \ref{tab:xrtIndex} (in Appendix~B),
respectively for MAGIC, VERITAS and \textit{Swift}/XRT.
At X-rays, the source shows the harder-when-brighter
behavior reported several times for Mrk~501
\citep[e.g.][]{1998ApJ...492L..17P,2007ApJ...669..862A}, but such behaviour is not
observed in the VHE domain during the observing campaign in 2012.
The Mrk~501 spectra measured in the X-ray and VHE ranges were harder than previously observed, during both
high and low activity. The very hard X-ray and VHE gamma-ray spectra
observed during the full campaign will be further discussed in Section~\ref{Mrk501vsEHBL}.
\section{Discussion}
\label{sec:discussion}
\input{notes/discussion1.tex}
\input{notes/discussion2.tex}
\input{notes/discussion3.tex}
\subsection{Mrk~501 as an Extreme BL Lac object in 2012}
\label{Mrk501vsEHBL}
The BL Lac objects known to emit VHE gamma rays have the maximum of their
high-energy component typically peaking in the 1--100~GeV band,
which implies that IACTs measure soft VHE spectra (power-law
indices $\Gamma > 2$, where dN/dE $\propto$ E$^{-\Gamma}$). However, there is also a small number of VHE BL Lacs
where the maximum of the gamma-ray peak is located well within the VHE band
\citep{2011MNRAS.414.3566T}, which implies that IACTs would measure
hard VHE spectra (power-law indices $\Gamma < 2$),
once the spectra are corrected for the absorption in the EBL. These objects have the peak of their
synchrotron peak also at higher energies ($>$1--10 keV), a property which was
initially used to flag them as special sources, and categorise them as
``extreme HBLs'' (EHBLs, \citet{2001A&A...371..512C}).
Archetypal objects belonging to this class, and extensively studied
in the last few years, are 1ES~0229+200
\citep[][]{2007A&A...475L...9A,2012ApJ...747L..14V,2014ApJ...782...13A,2013arXiv1307.8091C}
and 1ES~0347-121 \citep[][]{2007A&A...473L..25A,2014ApJ...787..155T}.
Some of the sources classified as EHBLs according to the position of
their synchrotron peak have been shown to have a very soft VHE
gamma-ray spectrum \citep[e.g. RBS~0723,][]{RBS0723}, which indicates that there is not a
uniform class of EHBLs, and hence some diversity within this
classification of sources (entirely based on observations). In this section we
focus on those EHBLs that also have a hard VHE gamma-ray
component (e.g. 1ES~0229+200), which are actually the most relevant
objects for EBL and intergalactic magnetic field (IGMF) studies \citep{2015ApJ...813L..34D,2015ApJ...814...20F}
In order to model the broadband SEDs of these EHBLs with hard VHE
gamma-ray spectral components, one
requires special physical conditions \citep[see e.g.][]{2009MNRAS.399L..59T,2011ApJ...740...64L,2014ApJ...787..155T}, such as large minimum electron
energies ($\gamma_{min} > 10^{2-3}$) and low magnetic fields
(\textit{B}$\lesssim$10--20 mG).
Moreover, leptonic models are also challenged by the limited variability
of the VHE emission of EHBLs, which differs very much from the
typically high variability observed in the VHE emission of
HBLs. For that reason, several authors have proposed that the VHE
gamma-ray emission is the result of electromagnetic cascades occurring in the intergalactic space,
possibly triggered by a beam of high-energy hadrons produced in the
jet of the EHBLs \citep[e.g.][]{2010APh....33...81E}, or alternatively
produced within leptohadronic models \citep[e.g.][]{2015MNRAS.448..910C}.
Aside from its importance in blazar emission models, the
extremely hard gamma-ray emission allows constraints to be placed on
the IGMF
\citep[e.g.][]{2010Sci...328...73N}, and provides a powerful tool study the absorption
of gamma rays in the EBL \citep[e.g.][]{2013IJMPD..2230025C}.
Potential deviations from this absorption are also of interest,
as they could be related to the mixing of photons with new spin-zero bosons such as axion-like particles
\citep[e.g.][]{2007PhRvD..76l1301D,2009PhRvD..79l3511S,2011PhRvD..84j5030D}.
Therefore, it is evident that EHBLs with hard VHE gamma-ray spectral components are fascinating objects that can be
used to study blazar jet phenomenology, high-energy cosmic rays, EBL
and IGMF. The main problem is that there are only a few sources
identified as EHBLs and detected using IACTs and that they are typically rather faint.
implying the need for very long observations, which complicates
the studies mentioned above.
For instance, 1ES~0229+200,
which is probably the most studied EHBL, has a VHE flux above 580~GeV of only
$\sim$0.02 CU, and for many years it was thought to be a
steady gamma-ray source. A 130-hour observation performed by H.E.S.S. recently
showed that the source is variable \citep{2015arXiv150904470C}, which has
strong implications for example on the lower limits derived on the IGMF.
Mrk~501 has been observed for a number of years by MAGIC and VERITAS, and it has
typically shown a soft VHE gamma-ray spectrum, with a power-law
index $\Gamma \sim$2.5
\citep[e.g.][]{2011ApJ...727..129A,2011ApJ...729....2A,2015A&A...573A..50A}.
It is known that during strong gamma-ray activity, such as the
activity in 1997 and 2005, the VHE spectra
became harder, with $\Gamma \sim$2.1--2.2
\citep{1999A&A...350...17D,1998ApJ...501L..17S,2007ApJ...669..862A} and recently
\citet{2016A&A...594A..76A}
have also reported similar spectral hardening during
the outstanding activity in May 2009. It is worth noticing that during
the big flare in April 1997, the synchrotron peak of Mrk~501 shifted to
energies beyond 100~keV, and that Mrk~501 was identified as an EHBL by
\citet{2001A&A...371..512C}. However, this happened only during
extreme flaring events.
On the contrary, as displayed in Figure~\ref{fig:index_xrt_pl}, during this campaign, Mrk~501 shows very hard X-ray and VHE gamma-ray
spectra during both very high and the quiescent or low activity.
The measured VHE spectra show power-law indices harder than 2.0, which has never been measured
before, and the hardness of the VHE
spectrum is independent of the measured activity.
A fit to the spectral indices with a constant yields $\Gamma$=2.041$\pm$0.015
($\chi^{2}$/NDF = 86/38). The left panel of Figure~\ref{fig:index_xrt_pl}
shows an average X-ray spectral index value of 1.752$\pm$0.004
($\chi^{2}$/NDF = 330/51). In both cases we have clear spectral variability,
hence spectra which statistically differ from the mean value.
In contrast to the VHE spectra, in the X-ray spectra one can
observe a dependence on the source activity, with the spectrum
getting harder with increasing flux; but Mrk~501 shows
spectra with photon index $\textless$ 2.0 even for the lowest-activity days.
We did not find any relation between the X-ray and VHE spectral indices.
In summary, during the 2012 campaign, both the X-ray and VHE spectra were persistently harder than
$\Gamma$=2 (during low and high source activity), which implies that the maximum of the synchrotron peak is above 5 keV,
and the maximum of the inverse-Compton peak is above 0.5~TeV. In other words,
Mrk~501 behaved effectively like an EHBL during 2012.
This suggests that being an EHBL may not be a permanent characteristic
of a blazar, but rather a state which may change over time.
\subsection{Modeling temporal evolution of the broadband SED}
\label{Mrk501SSCModel}
The accurate description of the broadband SED of Mrk~501 and its
temporal evolution can be provided by a complex theoretical scenario involving
the superposition of several emitting regions, as reported in
Section \ref{sec:sed}. We have shown that the optical-UV emission and the soft X-ray
emission cannot be parameterized with a single synchrotron component,
something that had already been observed during the campaign
from 2009 \citep{Mrk501MW2009_Variability}. Additionally, the \mbox{3-day-integrated} GeV emission, as measured by
\textit{Fermi}-LAT, is systematically above (at 1--2 $\sigma$ for
single SEDs) the model curves and, in two SEDs, we found indications of an additional component at MeV--GeV energies,
something which had been also reported by \citet{2015ApJ...798....2S} using observations
from 2010 and 2011.
However, the X-ray and VHE gamma-ray bands, which
are the segments of the SED with the highest energy flux, and the most variable ones,
can be described in a satisfactory way with a simple one-zone
SSC model. This fact allows one to draw straightforward physical
conclusions with a reduced number of model parameters.
The electron spectral
indices vary slightly across the models while the break energy
changes by a factor of three,
reaching the highest value during the night of the flare. The particle spectra
were found to be hard, with $p_\text{{1}}$ $\leq$ 2 for most cases, which is needed
to explain the very hard X-ray and VHE spectra. We also find a strong
positive correlation between the electron energy density,
\textit{U}$_\text{{e}}$ (derived from the one-zone SSC model) and the VHE
gamma-ray emission measured by MAGIC and VERITAS. The Pearson's
correlation coefficient between \textit{U}$_\text{{e}}$ and both the 0.2--1~TeV and above 1~TeV flux is
0.97$^{+0.01}_{-0.02}$, with the significance of the correlation being larger than 7$\sigma$.
The value of \textit{U}$_\text{{e}}$ depends on the value used for $\gamma_\text{{min}}$, which is not well constrained by the data.
But we noted that \textit{U}$_\text{{e}}$ only changes by 10--20\% when changing $\gamma_\text{{min}}$
by one order of magnitude. Given that the SSC modeling requires changes in \textit{U}$_\text{{e}}$ by factors of a few
to explain the 17 broadband SEDs of Mrk~501, we consider that the dependency on the chosen value for $\gamma_\text{{min}}$ does not have any
relevant impact in the significant correlation between the measured VHE flux and the SSC model \textit{U}$_\text{{e}}$ values.
This relation indicates that, within the one-zone SSC used here, the main cause of the broadband SED variability is the injection or
acceleration of electrons.
The average broadband SED of Mrk~501 during the observing campaign in 2009 was successfully modelled with a
one-zone SSC scenario, where the energisation of the electrons was attributed to diffusive first-order
Fermi acceleration \citep{2011ApJ...727..129A}. Yet during the multi-instrument observations in 2012 we
measured substantially harder X-ray and VHE spectra that required EEDs with harder spectra in the models.
Such hard-spectrum EEDs may be produced through second-order Fermi acceleration
\citep{2015MNRAS.447..530C,2011ApJ...740...64L,2009MNRAS.393.1063T,2016ApJ...832..177S}.
Additionally, the radiative cooling of a monoenergetic pileup particle energy distribution can result in a
power-law particle distribution with index of 2
\citep{2004ApJ...616..136S}.
These narrow distributions of particles may arise through stochastic
acceleration by energy exchanges with resonant Alfv\'en waves in a
turbulent medium as described by \cite{1985A&A...143..431S},
\cite{2008ApJ...681.1725S}, and \cite{2014ApJ...780...64A}.
In this case quasi-Maxwellian
distributions are obtained: these have been suggested by multiwavelength modeling
of Mrk~421 \citep{2015A&A...578A..22A}.
Magnetic reconnection in blazar jets
\citep{2010MNRAS.402.1649G}, which has been invoked by \citet{2015ApJ...811..143P}
to explain the variability of Mrk~421, is another process that can
effectively produce hard EEDs \citep{2012ApJ...746..148C,2012ApJ...754L..33C}.
Additionally, as reported in \citet{2014ApJ...789...66Z,2015ApJ...804...58Z}, through magnetic
reconnection, the dissipated magnetic energy is converted into
non-thermal particle energy, hence leading to a decrease in the magnetic field strength $B$ for increasing
gamma-ray activity and $U_e$. This trend is
also observed in the parameter values retrieved from our SSC model
parameterisation (see Table~\ref{tab:sed1}), thus supporting the hypothesis of
magnetic reconnection occurring in the jets of Mrk~501.
In principle, obtaining hard EEDs from a diffusive shock acceleration
process is difficult, as first-order Fermi acceleration produces a power-law index with value of 2, and
the spectrum then evolves in time due to radiative cooling and steepens further. However, Baring and
collaborators \citep{2017MNRAS.464.4875B} have recently shown that
shock acceleration can also produce hard EEDs with indices as hard as one, primarily because of
efficient drift acceleration in low levels of MHD turbulence near
relativistic shocks: see \citet{2012ApJ...745...63S} for a complete
discussion.
As reported in Section \ref{sec:sed}, the X-ray and gamma-ray segments from the SED related to the large
VHE flare on MJD 56087
(2012 June 9) were modelled with a two-zone SSC model in order to better describe
the high-energy peak, with a maximum at $\sim$2~TeV. In
this scenario, the X-ray and VHE spectra are completely dominated by the emission of a region that is
smaller (by one order of magnitude), and with a narrower EED characterized by a very high minimum Lorentz
factor $\gamma_\text{{min}}$. This multizone SSC scenario was successfully used to model the broadband SEDs of
Mrk~421 that also showed peaked or multi-peaked structures during a 13-day period of flaring activity in March 2010
\citep{2015A&A...578A..22A}. The relatively steady optical and GeV emission could be produced in a
shock-in-jet component while the variable X-ray and VHE gamma-ray emission could arise from a component
originating in the base of the jet and producing this relatively narrow EED. The more compact zone is probably
intimately connected to the injector site, perhaps a jet shock, thereby
more directly sampling the acceleration characteristics since there
has been less time for electrons to cool in the ambient magnetic field \citep{2017MNRAS.464.4875B}.
It is also worth noting that the large VHE flare from June 9 2012 occurred when the degree of
polarization was at its lowest value ($\sim$2\%) during the 2012
campaign (see bottom panels of Figure~\ref{fig:lightcurvecombinedTeV}). On
the other hand, the large VHE flare from May 1 2009 occurred when the degree of polarization was at its
highest value ($\sim$5\%) during the 2009 campaign \citep{2016A&A...594A..76A,Mrk501MW2009_Variability}. Since enhanced
polarization is naturally anticipated in short duration flares where
smaller length scales are sampled, this observational dichotomy
complicates the picture. This observation suggests that
there is a diversity in gamma-ray flares in Mrk~501, and at least some of them
seem not to involve any change in the degree of polarization, which may occur naturally if the optical and
the VHE emission are produced in different regions of the jet.
\subsection{Multiband variability and correlations}
\label{discussion:variability}
Section \ref{sec:variability} reports a general increase in the flux variability with
increasing energy. At radio, optical and UV bands we observe relatively
low variability ($F_{\text{var}}\leq 0.1$), but for the variability at the 37 GHz
radio fluxes from Metsahovi, which
is 0.13$\pm$0.02. This variability is not produced by a flare, or by a
slow temporal evolution (weeks or months long) of the light curve, which is
often observed at radio, but by a consistent flickering in the radio fluxes.
Such flickering is rare in blazars, but it has been
already reported in previous observing campaigns of Mrk~501
\citep[][]{2015A&A...573A..50A,2015ApJ...812...65F}. In the X-rays and GeV
gamma-ray bands we observe high variability ($F_{\text{var}}\sim
0.2-0.4$). However, we note that we do not have sensitivity to
determine the fractional variability in the
band 0.2-2~GeV (where the excess variance is negative), and hence the
fractional variability in this band could be lower than that measured at
X-rays.
In the VHE gamma-ray band we observe very high variability
($F_{\text{var}}\sim 0.5-0.9$), i.e. about three times larger than that at X-rays.
Such a multiband (from radio to VHE) variability pattern
was first reported with observations from 2008 in \citet{2015A&A...573A..50A} and then confirmed with more
precise measurements from the 2009 campaign \citep{Mrk501MW2009_Variability}. The repeated occurrence of
this variability pattern in 2012 demonstrates that this is a typical characteristic in the broadband
emission of Mrk~501.
On the other hand, \citet{2015ApJ...812...65F} show that, during the observing
campaign in 2013, the multiband variability pattern was somewhat different, with the variability at X-rays
being similar to that at VHE, thereby showing that somewhat different dynamical processes occurred in
Mrk~501 during that year.
It is worth comparing this multi-year variability pattern of Mrk~501 with that from the other archetypical
TeV blazar, Mrk~421. During the multi-instrument campaigns from 2009, 2010 and 2013, as reported in
\cite{2015A&A...576A.126A}, \cite{2015A&A...578A..22A} and \cite{2016ApJ...819..156B}, Mrk~421 showed a
double-peak structure in the plot of $F_\text{{var}}$ against energy,
where the largest variability occurs in X-rays and VHE (instead of a
broad increase with energy, with the variability at VHE being much larger than that at X-rays).
These observations show a fundamentally different behavior when compared to that of Mrk~501.
\begin{figure}
\centering
\includegraphics[width=10.0cm]{figures3/05kev1tev.eps}
\includegraphics[width=10.0cm]{figures3/5kev1tev.eps}
\includegraphics[width=10.0cm]{figures3/50kev1tev.eps}
\caption{Flux-flux plots derived from the one-zone SSC model used
to fit 17 broadband SEDs (see Section~\ref{sec:sed}). The SSC model flux at 1 TeV is compared to the
SSC model flux at 0.5~keV (top), 5~keV (middle) and 50~keV
(bottom).
The data point with the highest X-ray and VHE activity
corresponds to the one-zone SSC model for the June 9 flare (MJD 56087).
}
\label{fig:sedcorrtev}
\end{figure}
During large VHE gamma-ray flaring activity, the X-ray and VHE gamma-ray emission of Mrk~501 have
been found to be correlated. This occurred during the long and historical flare of 1997
\citep[][]{1998ApJ...492L..17P,Gliozzi:2006it}, and the large few-days-long flare observed in 2013
\citep{2015ApJ...812...65F}. During non-flaring activity, a positive X-ray/VHE correlation was reported
at the 99\% confidence level \citep{2015A&A...573A..50A}. On the other hand, using the measurements from
the 4.5-month-long 2009 campaign, where significant variability was observed in both X-ray and VHE bands,
the emission from these two bands was found to be uncorrelated \citep{Mrk501MW2009_Variability}. Using the
data collected during the 2012 campaign, with many more observations than in 2009, one also finds only marginal
correlation between the X-ray emission and the VHE gamma-ray emission.
This is an interesting result
because, under the most simplistic and widely
accepted theoretical scenarios, the X-ray emission and the VHE gamma-ray
emission are produced by the same population of high-energy particles (electrons and positrons). We note
also that the situation for Mrk~421 is radically different from that of Mrk~501. The various multi-instrument
campaigns performed on Mrk~421 always show a clear and positive correlation between the X-ray emission and the VHE
gamma-ray emission, during both high and low source activity
\citep{2008ApJ...677..906F,2011ApJ...738...25A,2015A&A...578A..22A,2015A&A...576A.126A,2016A&A...593A..91A,2016ApJ...819..156B}.
\begin{table}
\caption{Correlations derived for several combinations of X-ray
bands and VHE flux above 1 TeV using the data (upper part) and the
one-zone SSC theoretical model used to describe 17 broadband
SEDs. See text for further details.}
\label{tab:datamodelcorr}
\begin{center}
\begin{tabular}{ l | r }
\hline
& Pearson correlation \\
& coefficient ($\sigma$) \\
\hline
Data: 0.3--2~keV vs > 1~TeV & 0.78$^{+0.10}_{-0.15}$ (3.9) \\
Excluding June 9 flare: & 0.39$^{+0.23}_{-0.29}$ (1.6) \\
\vspace{-0.3cm}
~~~~ & ~~~~~~ \\
Data: ~2--10~keV vs > 1~TeV & 0.88$^{+0.06}_{-0.10}$ (4.9) \\
Excluding June 9 flare: & 0.59$^{+0.16}_{-0.24}$ (2.5) \\
\vspace{-0.2cm}
~~~ & ~~~~ \\
\hline
\vspace{-0.2cm}
~~~~ & ~~~~~~ \\
Model: 0.5~~keV vs 1~TeV & 0.71$^{+0.11}_{-0.16}$ (3.3) \\
Excluding June 9 flare: & 0.44$^{+0.19}_{-0.25}$ (1.7) \\
\vspace{-0.3cm}
~~~~ & ~~~~~~ \\
Model: 5.0~~keV vs 1~TeV & 0.87$^{+0.06}_{-0.09}$ (4.8) \\
Excluding June 9 flare: & 0.63$^{+0.14}_{-0.20}$ (2.7) \\
\vspace{-0.3cm}
~~~~ & ~~~~~~ \\
Model: 50.0~keV vs 1~TeV & 0.77$^{+0.09}_{-0.13}$ (3.8) \\
Excluding June 9 flare: & 0.73$^{+0.11}_{-0.16}$ (3.4) \\
\hline
\end{tabular}
\end{center}
\end{table}
\citet{Mrk501MW2009_Variability} put forward two scenarios to explain the measured multiband variability
and correlations seen in the emission of Mrk~501: a) the high-energy electrons that are responsible for a large
part of the TeV emission do not dominate the keV emission; b) there is an additional (and very
variable) component contributing to the TeV emission, such as external inverse-Compton.
In this manuscript, we
use the results from our one-zone SSC modelling of the 17 broadband SEDs from 2012 to probe the first
scenario. We compared the one-zone SSC model fluxes at 0.5~keV, 5~keV and 50~keV with
the one-zone SSC model fluxes at 1~TeV, which are, by construction of the theoretical model, produced by
the same population of electrons. The results are depicted in Figure~\ref{fig:sedcorrtev}, and the
correlations obtained are reported in Table~\ref{tab:datamodelcorr}. The X-ray
vs VHE gamma-ray correlation
as a function of the X-ray energy is: 3.3$\sigma$ for 0.5~keV, 4.8
$\sigma$ for 5~keV, and 3.8$\sigma$ for 50~keV. As reported in section \ref{sec:sed}
(see Table \ref{tab:sed1}), the June 9 flare (MJD 56087) is not
properly described by the one-zone SSC scenario used here. If we remove the
results derived with the SSC model for this large flare, hence providing a
more reliable description of the typical behaviour of Mrk~501 during
the campaign in 2012, the significance of the X-ray/VHE correlation is 1.7$\sigma$ for 0.5~keV, 2.7$\sigma$ for 5~keV, and 3.4$\sigma$
for 50~keV. Therefore, the one-zone SSC model provides X-ray/VHE correlations at a level that are
consistent with correlations obtained with the measured X-ray and VHE gamma-ray fluxes reported in
Table~\ref{tab:Corr1} \citep{2017A&A...603A..31A}. This shows that an
additional high-energy component (e.g. external inverse-Compton) is not necessary to explain the
variability and correlation patterns observed in Mrk~501.
This exercise also shows the importance of sampling with accuracy a large portion of the electromagnetic
spectrum.
In particular, sensitivity in the 50--100~keV range comparable to that currently provided by
MAGIC and VERITAS in the 0.1--1~TeV range would greatly increase the potential for studying flux variability
and interband correlations.
The main differences in the multiband
variability and correlation patterns with Mrk~421 may be related to the fact that, for Mrk~421,
the electrons dominating the emission of the $\sim$1~TeV photons also dominate the emission at
$\sim$1~keV (the peak of the synchrotron spectrum).
This is the energy region sampled with \textit{Swift}/XRT with exquisite accuracy and extensive temporal coverage.
Recent and future publications devoted to multiwavelength campaigns on blazars will continue to benefit from a new
generation of X-ray telescopes, including \textit{NuSTAR}\footnote{\url{http://www.nustar.caltech.edu}} and
\textit{Astrosat}\footnote{\url{http://astrosat.iucaa.in}}, which
operate at 3--79~keV and 2--80~keV respectively. \textit{NuSTAR}
represents a significant improvement (by a factor of 100~in sensitivity)
over coded-mask instruments like \textit{Swift}/BAT, and hence
provides a much better view into the hard X-ray emission of blazars.
\section{Participating Instruments}
\label{sec:experiments}
\input{notes/magic.tex}
\input{notes/veritas.tex}
\input{notes/fact.tex}
\input{notes/fermi.tex}
\input{notes/swift.tex}
\input{notes/optical.tex}
\input{notes/radio.tex}
\begin{figure*}
\centering
\includegraphics[height=48.5pc,width=46pc]{figures3/lightcurvePol.eps}
\caption{Multiwavelength light curve for Mrk~501 during
the 2012 campaign.
The bottom two panels report the electric vector polarization angle (P.A.) and polarization degree (P.D.).
The correspondence between the instruments and the measured quantities is given in the legends.
The horizontal dashed line in the VHE light curves represents 1 CU as reported in
\citet{2016APh....72...76A}, and the blue vertical dotted lines in the panels with the
polarization light curves depict the day of the large VHE flare (MJD~56087).
}
\label{fig:lightcurvecombinedTeV}
\end{figure*}
\subsection{FACT}
The First G-APD Cherenkov Telescope (FACT) is
the first Cherenkov telescope to use silicon photomultipliers (SiPM/G-APD) as photodetectors.
As such, the camera consists of 1440
G-APD sensors, each
with a field of view of 0.11$^{\circ}$ providing a total field of view of 4.5$^{\circ}$.
FACT is located next to the MAGIC telescopes at the Observatorio del Roque de Los Muchachos.
The telescope makes use of the old HEGRA CT3
\citep{1998NewAR..42..547M} mount, and has a focal length of 5~m and an
effective dish diameter of $\sim$3~m. The telescope operates in the
energy range from $\sim$0.8~TeV to $\sim$50~TeV.
For more details about the design and experimental setup see \cite{2013JInst...8P6008A}.
Since 2012, FACT has been continuously monitoring known TeV blazars,
including Mrk~501 and Mrk~421.
FACT provides a dense sampling rate by focusing on a subset of sources
and the ability of the instrument to operate safely during nights of bright ambient light.
The data are analyzed and processed immediately, and results are available publicly
online \footnote{\texttt{http://fact-project.org/monitoring/}} within minutes of the observation
\citep{2013arXiv1311.0478D,2014arXiv1407.1988B,2015arXiv150202582D}.
\subsection{\textit{Fermi} LAT}
The Large Area Telescope (LAT, \citet{2009ApJ...697.1071A}) is a pair-conversion
telescope ({\it Fermi Gamma-ray Space Telescope}) operating in the
energy range from $\sim$30~MeV to $>$TeV. \textit{Fermi} scans the
sky continuously, completing one scan every 3~hours. The \textit{Fermi}-LAT data presented
in this paper cover the period from 2011 December 29 (MJD 55924)
to 2012 August 13 (MJD 56152). The data were analyzed using the standard \textit{Fermi} analysis software tools
(version \textit{v10r1p1}), using the \textit{P8R2\_SOURCE\_V6} response function.
Events with energy above 0.2 GeV and coming from a 10$^{\circ}$ region of interest (ROI) around Mrk 501 were
selected, with a 100$^{\circ}$ zenith-angle cut to avoid contamination from the Earth's limb. Two
background templates were used to model the diffuse Galactic and
isotropic extragalactic background, \textit{gll\_iem\_v06.fits} and
\textit{iso\_P8R2\_SOURCE\_V6\_v06.txt},
respectively\footnote{http://fermi.gsfc.nasa.gov/ssc/data/access/lat/BackgroundModels.html}.
All point sources in the third {\it Fermi}-LAT source catalog
\citep[3FGL,][]{2015ApJS..218...23A} located in the 10$^{\circ}$ ROI
and an additional surrounding 5$^{\circ}$-wide annulus were included
in the model. In the unbinned likelihood fit, the spectral parameters were set to
the values from the 3FGL, while the normalization parameters of the
nine sources within the ROI identified as variable were
left free. The normalisation of the diffuse components, as well as the
the model parameters related to Mrk~501 were also left free.
Because of the moderate sensitivity of \textit{Fermi}-LAT to detect
Mrk~501 (especially when the source is not flaring), we performed the
unbinned likelihood analysis on one-week time intervals for determining the light
curves in the two energy bands 0.2--2~GeV and $>$2~GeV reported in
Section~\ref{sec:lightcurves}. In both cases we fixed the PL index to 1.75, as was
done in \cite{Mrk501MW2009_Variability}. On the other hand, in order to increase the
simultaneity with the VHE data, we used 3-day time intervals (centered at the night of the VHE
observations) for most of the unbinned likelihood spectral analyses reported in
Section~\ref{sec:sed}. For those spectral analyses, we performed first
the PL fit in the range from 0.2~GeV to 300~GeV (see spectral
results in Table~\ref{tab:fermiIndex}). Then, we performed the
unbinned likelihood analysis
in three energy bins (split equally in log space from 0.2~GeV to
300~GeV) where the PL index was fixed to the value retrieved from
the spectral fit to the full energy range. Flux upper limits at 95\% confidence level were calculated whenever
the test statistic (TS) value\footnote{The TS value quantifies the probability of having a point gamma-ray source at
the location specified. It is roughly the square of the significance
value \citep{1996ApJ...461..396M}.} for the source
was below 4 \footnote{A TS value of 4 corresponds to a $\sim$2$\sigma$ flux measurement, which is a commonly used threshold for flux measurements of known sources.}.
\section{Fractional Variability}
\label{sec:variability}
\begin{figure*}
\centering
\includegraphics[width=20pc]{figures3/fvar7.eps}
\includegraphics[width=20pc]{figures3/fvar7_noflare.eps}
\caption{Fractional variability $F_\text{{var}}$ for each instrument as a function of energy.
Left panel includes all data, while the right panel includes
all data except for the day of the VHE flare (MJD 56087).
$F_\text{{var}}$ values computed with X-ray and VHE data
taken within the same night are shown with gray open markers.
}
\label{fig:fvar_all}
\end{figure*}
In order to characterize the variability at each wavelength we follow the prescription of
\cite{2003MNRAS.345.1271V} where the fractional variability ($F_\text{{var}}$) is defined as
\begin{equation}
F_{var} = \sqrt{\frac{S^{2} - \langle \sigma_{err}^{2} \rangle }{ {\langle x \rangle}^2}}
\end{equation}
Here \textit{S} is the standard deviation of the flux measurement, $\langle\sigma_{err}^{2}\rangle$
the mean squared error and ${\langle x \rangle}^{2}$ the square of the
average photon flux.
The error on $F_\text{{var}}$ is estimated following the prescription of
\cite{2008MNRAS.389.1427P}, as described by \cite{2015A&A...573A..50A}
\begin{equation}
\Delta F_\text{{var}} = \sqrt{ F_\text{{var}}^{2} + err( \sigma_\text{{NXS}}^{2} ) } - F_\text{{var}}
\end{equation}
and $err( \sigma_\text{{NXS}}^{2} )$ is taken from equation 11 in \cite{2003MNRAS.345.1271V}
\begin{equation}
err( \sigma_\text{{NXS}}^{2} ) = \sqrt{ \
\left( \sqrt{\frac{2}{N}} \frac{\langle \sigma^{2}_\text{{err}} \rangle}{\langle x \rangle ^{2}} \right)^{2} \
+ \
\left( \sqrt{\frac{\langle \sigma^{2}_\text{{err}} \rangle}{N}} \frac{2F_\text{{var}}}{\langle x \rangle} \right)^{2} \
}
\end{equation}
where \textit{N} is the number of flux measurements. This method, commonly used to quantify the variability, has the
caveat that the resulting $F_\text{{var}}$ and its related
uncertainty depend on the instrument sensitivity and the observing strategy performed. For instance, densely
sampled light curves with small uncertainties in the flux measurements may allow us to see flux variations that are
hidden otherwise, and hence may yield a larger $F_\text{{var}}$ and/or smaller uncertainties in the calculated values
of $F_\text{{var}}$. This introduces differences in the ability to detect variability in the different
energy bands. Issues regarding the application of this method, in the context of multiwavelength campaigns,
are discussed by \citet{2014A&A...572A.121A,2015A&A...573A..50A,2015A&A...576A.126A}. In the multi-instrument dataset
presented in this case, the sensitivity of the instruments {\it Swift}/BAT and {\it Fermi}-LAT precludes the detection of
Mrk~501 on hour timescales, and hence integration over several days is required (and still yields flux measurements
with relatively large uncertainties). This means that the {\it Swift}/BAT and {\it Fermi}-LAT $F_\text{{var}}$ values are not
directly comparable to those of the other
instruments, for which $F_\text{{var}}$ values computed with nightly observations (and typically smaller error bars) are reported.
Figure \ref{fig:fvar_all} shows the $F_\text{{var}}$ as a function of energy.
The left panel uses all the data presented in Figure \ref{fig:lightcurvecombinedTeV}, with the exception of nights where
there were simultaneous FACT and MAGIC data. In these cases the FACT data are removed.
The figure displays $F_\text{{var}}$ values for those bands with positive excess variance
($S^2$ larger than \mbox{$<\sigma_{err}>^2$}); a negative excess
variance is interpreted as
absence of variability either because there was no variability or because the instruments
were not sensitive enough to detect it. We obtained negative excess variances
for the 15 GHz radio fluxes measured with OVRO and the 0.2-2~GeV fluxes
measured with {\it Fermi}-LAT.
The right panel shows the same data except for the flare day (MJD 56087), which has been removed from the multi-instrument dataset, and hence
shows a more typical behavior of the source
during the 2012 multi-instrument campaign.
Figure \ref{fig:fvar_all} also reports the values of $F_\text{{var}}$ obtained by using the
X-ray/VHE observations taken simultaneously\footnote{The $F_\text{{var}}$ in the radio and optical bands does not change much when selecting sub-samples of the full dataset because the
variability in these energy bands is small and the flux variations have longer timescales, in comparison with those
from the X-ray and VHE bands.}. Additionally, the right panel in Figure \ref{fig:fvar_all} also shows that,
when the large VHE flare from MJD 56087 is removed, the
$F_\text{{var}}$ changes substantially in the VHE gamma-ray band
(e.g. from 0.93$\pm$0.04 down to 0.53$\pm$0.05 above 1 TeV) but the
variability changes mildly in the X-ray band (e.g. from
0.301$\pm$0.003 to 0.241$\pm$0.003 at 2-10~keV). In both panels there is a
general increase of the fractional variability with increasing
energy of the emission. These results will be further discussed in
Section~\ref{discussion:variability}.
\section{Introduction}
\label{sec:introduction}
The galaxy Markarian 501 (Mrk~501; $z$=0.034) was first cataloged,
along with Markarian 421,
in an ultra-violet survey \citep{1972Ap......8...89M}.
At very high energies (VHE; $E$ > 100 GeV) it was first detected
by the pioneering Whipple imaging atmospheric-Cherenkov telescope \citep[IACT,][]{1996ApJ...456L..83Q}.
Mrk~501 is a BL Lacertae (BL Lac) object, a
member of the blazar subclass of active galactic nuclei (AGN),
the most common source class in the extragalactic VHE catalog\footnote{http://tevcat.uchicago.edu}.
Since the discovery of Mrk~501's VHE emission, it has been extensively studied across all
wavelengths. The spectral energy distribution (SED) shows the two characteristic broad peaks,
the low-frequency peak from radio to
X-ray and the high-frequency peak from X-ray to very high energies.
The first peak is thought to originate from synchrotron emission. The second either from inverse-Compton scattering of electrons from the lower-energy component
\citep{1985ApJ...298..114M,1992ApJ...397L...5M,1992A&A...256L..27D,1994ApJ...421..153S} or
from the acceleration of hadrons which produce synchrotron emission or interact to produce pions and, in turn, gamma rays
\citep{1993A&A...269...67M,2000NewA....5..377A,2000A&A...354..395P}.
Whilst the typical flux of Mrk~501, above 1~TeV, in a non-flaring
state, is about one-third that of the Crab Nebula (Crab units; CU)\footnote{
In this study we use the Crab Nebula VHE emission reported in
\citet{2016APh....72...76A}. The photon flux of the Crab
Nebula above 1 TeV is \mbox{2 $\times$ 10$^{-11}$ cm$^{-2}$ s$^{-1}$.}}, it has shown extraordinary flaring activity,
the first notable examples occuring in 1997 \citep{1997ApJ...487L.143C,1999A&A...350...17D,1999ApJ...518..693Q}.
Another such flaring episode in the same year \citep{1999A&A...349...11A} showed the flux above 2~TeV ranged from a fraction of 1~CU
to 10~CU, with an average of 3~CU, and the doubling timescale was found to be as short as 15
hours. In the same period the \textit{BeppoSAX} X-ray satellite reported a hundredfold increase in
the energy of the synchrotron peak in coincidence with a hardening of the spectrum.
Mrk~501 is an excellent object with which to study blazar phenomena because it
is bright and nearby, which permits significant detections in
relatively short observing times in essentially all energy bands. Therefore,
the absorption of gamma rays in the extragalactic background light (EBL, \citet{2013APh....43..112D,2007A&A...475L...9A,2015MNRAS.451..611B}),
although not negligible, plays a relatively small role below a few
TeV. The flux attenuation factor, $\exp(-\tau)$, at a photon energy of 5~TeV is smaller than 0.5 (for z=0.034) for most EBL models
\citep{2008A&A...487..837F,2011MNRAS.410.2556D,2012MNRAS.422.3189G}.
In 2008, an extensive multi-instrument program was organised in order to
perform an objective (unbiased by flaring states) and detailed study of the temporal evolution, over many years, of
the broadband emission of Mrk~501 \citep[see e.g.][]{2011ApJ...727..129A,2015A&A...573A..50A,2015ApJ...812...65F,2017A&A...603A..31A}.
Here, we report
on one of those campaigns, that took place in 2012 and serendipitously observed the largest
flare since 1997. This paper is organized as follows: In Section \ref{sec:experiments} the experiments that took part
in the campaign are described along with their data analysis. Section \ref{sec:lightcurves}
describes the multiwavelength light curves from these instruments and is followed by Sections
\ref{sec:variability} and \ref{sec:corr}, in which the multiband
variability and related correlations are characterized.
Section \ref{sec:sed} characterizes the broadband SED within a
standard leptonic scenario, and in Section \ref{sec:discussion}, we discuss the
implications of the osbservational results reported in this
paper. Finally, in Section \ref{sec:conclussion}, we make some concluding remarks.
\section{Multiwavelength Light Curve}
\label{sec:lightcurves}
Figure \ref{fig:lightcurvecombinedTeV} shows a complete set of light curves for all
participating instruments, from radio to VHE.
The first panel from the top shows the radio data from the Mets\"{a}hovi and Owens Valley radio
observatories.
Each data point represents the average over one night of observations.
Optical data in the R-band, after host galaxy subtraction as prescribed in section 2.6,
are shown in the second panel. The light curve also
shows very little variability, with just a slow change in flux of
about $\sim$10-20\% on timescales of many tens of days.
When compared to the 13 years of optical observations from the
Tuorla group\footnote{See \url{http://users.utu.fi/kani/1m/Mkn_501_jy.html}}, one can note
that in 2012 the flux was at a historic minimum.
The ultraviolet data from the \textit{Swift}/UVOT
are presented in the third panel and follows the same pattern as the
R band fluxes. Overall, the low-frequency observations (radio to ultraviolet) show little variation during this period.
In the X-ray band the \textit{Swift}/XRT and BAT light curves show a
large amount of variation, occurring on timescales of days (i.e. much
faster than those in the optical band).
The \textit{Swift}/XRT points represent nightly fluxes derived from
$\sim$1~ks observations (where the error bars are smaller than the markers),
while the \textit{Swift}/BAT points are
the weighted average of all measurements performed within
5-day intervals.
On the day of June 9 2012 (MJD 56087)
a flare is observed where the \textit{Swift}/XRT flux reached 3.2 $\times$ 10$^{-10}$
erg cm$^{-2}$ s$^{-1}$
in the 2--10~keV band. Interestingly, the largest flux point in the 0.3--2~keV band
occurs two days later, indicating
that the X-ray activity can have a different variability pattern
below and above 2~keV.
The \textit{Fermi}-LAT light curves, which are binned in 7~day time intervals, show some
mild variability. The ability to detect small amplitude variability at these energies is
strongly limited by the relatively large statistical uncertainty in the flux measurements.
The seventh and eighth panels of Figure~\ref{fig:lightcurvecombinedTeV} show the VHE light curves
from MAGIC, VERITAS and FACT.
Here we split the VHE information from MAGIC and VERITAS into two bands,
from 200~GeV to 1~TeV and above 1~TeV.
Each point represents a nightly average, with the 18 MAGIC
observations, obtained from an average observation of
1.25~hours, and the 28 VERITAS data points, obtained
from an average observation of 0.5~hours.
The VHE emission is highly variable, with
the average in both bands being approximately 0.7~CU above 1~TeV.
The largest VHE flux is observed on 2012 June 9, where the light
curves show a very clear flare (which is also visible in the X-ray
light curve) with a 0.2--1~TeV flux of 5.6 $\times$ 10$^{-10}$
cm$^{-2}$ s$^{-1}$ (2.8 CU), and the $>$1~TeV flux reaching 1.0 $\times$
10$^{-10}$ cm$^{-2}$ s$^{-1}$ (4.9 CU).
Unfortunately, VERITAS was not scheduled to observe Mrk~501 on June 9.
FACT observed Mrk~501 for an average of 3.3~hours per night over 73 nights during the
campaign. As with the other TeV instruments, the data shown are binned nightly.
The FACT fluxes reported in Figure \ref{fig:lightcurvecombinedTeV} were obtained with a
first-order polynomial that relates the MAGIC flux (ph cm$^{-2}$ s$^{-1}$ above 1 TeV) and the FACT
excess rates (events/hour), as explained in Appendix~A.
Measurements of the degree of optical linear polarization and its position
angle are displayed in the bottom panels of Figure~\ref{fig:lightcurvecombinedTeV}.
As with the optical photometry, the polarization shows only mild variations on time scales of
weeks to months during the campaign. Variations of the degree of
polarization are muted by the strong contribution of unpolarized starlight
from the host galaxy falling within the observation apertures.
At these optical flux levels and with the apertures used for the ground-based
polarimetry, the optical flux from the host galaxy is about 2/3 of the
flux measured, and hence the intrinsic polarization of the blazar is about a factor of
three higher than observed.
Different instruments used somewhat different apertures and optical
bands, which implies that the
contribution of the host galaxy to the optical flux and polarization
degree will be somewhat different for the different instruments.
Since the host galaxy is not subtracted, this leads to small offsets
(at the level of $\sim$1\%) in the measurements of the degree of
polarization. The position angle of the polarization (which is not
affected by the host galaxy) remains at 120-140$^{\circ}$ for more than a month before and after the VHE
flare. For comparison, the position angle of the 15~GHz VLBI jet is at
$\sim$150$^{\circ}$ \citep{2009AJ....137.3718L}. Overall, there is no
apparent optical signature, either in flux or linear polarization, that
can be associated with the gamma-ray activity observed in Mrk~501 during
2012.
\subsection{MAGIC}
The Major Atmospheric Gamma-ray Imaging Cherenkov Telescopes (MAGIC) comprise two telescopes located at the Observatorio del Roque de Los
Muchachos, La Palma, Canary Islands, Spain (2.2~km a.s.l., 28$^{\circ}$ 45$^{\prime}$ N 17$^{\circ}$ 54$^{\prime}$ W). Both
telescopes are 17~m in diameter and have a parabolic dish. The system is able to detect air showers initiated by gamma rays in the
energy range from $\sim$50~GeV to $\sim$50~TeV.
During 2011 and 2012 the readout systems of both telescopes were upgraded,
and the camera of MAGIC-I (operational since 2003) was replaced, increasing the density of pixels.
This
resulted in a telescope performance enabling a detection of a $\sim$0.7\% Crab Nebula-like source
within 50 hours, or a 5\% Crab-like flux in 1 hour of observation. The systematic uncertainties in the
spectral measurements for a Crab-like point source were estimated to be 11\% in the normalization factor (at $\sim$200 GeV)
and 0.15 in the power-law slope. The systematic uncertainty in the absolute energy determination is estimated to be 15\%.
Further details about the performance of the MAGIC telescopes after the hardware upgrade in 2011--2012 can be found in
\cite{2016APh....72...76A}. The data were analyzed using MARS, the
standard analysis package of MAGIC \citep{Zanin2013-MARS,2016APh....72...76A}.
The data from March and April 2012 were taken in stereo mode,
whilst the data taken in May and June 2012 were taken with MAGIC-II
operating as a single telescope due to a technical issue which
precluded the operation of MAGIC-I.
\subsection{Optical Instruments}
\label{optical:description}
Optical data in the R band were provided by various telescopes around the world, including the ones from
the GASP-WEBT program \citep{2008A&A...481L..79V,2009A&A...504L...9V}. In this paper we report observations
performed in the R band from the following observatories:
Crimean Astrophysical Observatory, St. Petersburg,
Sierra de San Pedro M\`{a}rtir,
Roque de los Muchachos (KVA),
Teide (IAC80),
Lulin (SLT),
Rozhen (60cm),
Abastumani (70 cm),
Skinakas,
the robotic telescope network AAVSOnet,
ROVOR and
iTelescopes.
The calibration was performed using the stars
reported by \citet{1998A&AS..130..305V}, the Galactic extinction was corrected using the coefficients given
in \citet{2011ApJ...737..103S}, and the flux from the host galaxy in the R band was estimated using
\citet{2007A&A...475..199N} for the apertures of 5 arcsecond and 7.5 arcsecond used by the various
instruments. The reported fluxes include instrument-specific offsets of a few mJy, owing to the different
filter spectral responses and analysis procedures of the various optical data sets (e.g. for signal and
background extraction) in combination with the strong host-galaxy contribution, which is about 2/3 of the
total flux measured in the R band. The offsets applied are the following ones: Abastumani = 3.0 mJy; San
Pedro Martir = -1.8 mJy; Teide = 2.1 mJy; Rozhen = 3.9 mJy; Skinakas = 0.8 mJy; AAVSOnet = -3.8 mJy;
iTelescopes = -2.5 mJy; ROVOR = -2.7 mJy. These offsets were determined using several of the GASP-WEBT instruments as
reference, and scaling the other instruments (using simultaneous observations) to match them.
Additionally, a point-wise fluctuation of 0.2 mJy ($\sim$0.01 mag) was added in quadrature to the
statistical errors in order to account for
potential day-to-day differences for observations with the same instrument.
We also report on polarization measurements from five facilities: Lowell Observatory (Perkins telescope),
St. Petersburg (LX-200), Crimean (AZT-8+ST7), Steward Observatory (2.3m Bok and 1.54m Kuiper telescopes) and Roque de los
Muchachos (Liverpool telescope). All polarization measurements
were obtained from R band imaging polarimetry, except for the measurements from Steward Observatory, which
are derived from spectropolarimetry between 4000~$\r{A}$ and 7550~$\r{A}$ with a resolution of
$\sim$15~$\r{A}$. The
reported values are constructed from the median \textit{Q}/\textit{I} and \textit{U}/\textit{I}
in the 5000--7000~$\r{A}$ band. The
effective wavelength of this bandpass is similar to the Kron-Cousins R band. The wavelength dependence
in the polarization of Mrk~501 seen in the spectro-polarimetry is small and does not significantly affect
the variability analysis of the various instruments presented here, as can be deduced from the good
agreement between all the instruments shown in the bottom panels of Figure~\ref{fig:lightcurvecombinedTeV}.
The details related to the observations and analysis of the polarization data are reported by
\cite{2008A&A...492..389L,2009arXiv0912.3621S,2010ApJ...715..362J,2016MNRAS.462.4267J}.
\subsection{Polarization}
\begin{figure}
\centering
\includegraphics[width=23pc]{figures3/Mrk501_polarization_all.eps}
\caption{Electric vector polarization angle and polarization
degree during the multi-instrument campaign of Mrk~501 in
2012. The correspondence between the instruments and the measured
quantities is given in the legend. Further details about
these measurements are given in Section
\ref{optical:description}. The blue vertical line is represents
the day of the VHE flare (MJD=56087).
}
\label{fig:polar}
\end{figure}
\subsection{Radio Observations}
We report here radio observations from telescopes at the Mets\"{a}hovi Radio Observatory
and the Owens Valley Radio Observatory (OVRO).
The 14~m Mets\"{a}hovi Radio telescope operates at 37 GHz and the OVRO at 15 GHz.
Details of the observation strategies can be found in \citet{1998A&AS..132..305T} and
\citet{2011ApJS..194...29R}. For both instruments Mrk~501 is a point source, and therefore the
measurements represent an integration of the full source extension,
which is much larger than the region that is expected to produce the
blazar emission at optical/X-ray and gamma-ray energies that we wish
to study.
However, as reported by \citet{2011ApJ...741...30A}, there is a correlation between radio and GeV
emission of blazars.
In the case of Mrk~501,
\citet{Mrk501MW2009_Variability} showed that the radio core emission
increased during a period of high gamma-ray activity,
therefore
part of the radio emission seems to be related to the gamma-ray component, and should be considered when studying the blazar emission.
\section{Temporal evolution of the broadband spectral energy distribution}
\label{sec:sed}
In order to model the data, several time-resolved spectral energy distributions were formed.
Spectral measurements were selected in cases where a \textit{Swift}/XRT spectrum and a MAGIC/VERITAS
spectrum were obtained within 6 hours of each other (i.e. from observations performed during the same night).
This allowed 17 distinct SEDs to be constructed, spanning three months.
The mean absolute time difference between the X-ray and VHE data are
1.2~hours, with the maximum time difference being 4.0~hours.
Because of the substantially lower variability at radio and
optical (see Figure~\ref{fig:fvar_all}) in comparison to that at
X-rays and VHE gamma-rays, strict simultaneity in these
bands is not relevant. Nevertheless, the \textit{Swift}/UVOT data are naturally simultaneous to that of
\textit{Swift}/XRT, and the high sampling performed by optical
instruments provides a flux measurement well within half day of
the X-ray and VHE observations.
\subsection{Theoretical model and fitting methodology}
The broadband SED of Mrk~501 has previously been modeled well using one-zone synchrotron self-Compton (SSC)
scenarios during high and low activity
\citep{2001ApJ...554..725T,2011ApJ...727..129A,2015A&A...573A..50A,2015ApJ...812...65F}.
The emission is assumed to come from a spherical region, containing a population of relativistic electrons, traveling along
the jet. The region has a radius $R$, is permeated by a magnetic field
of strength $B$ and is moving relativistically with a Doppler factor $\delta$.
The electron energy distribution (EED) is assumed
to have an energy density $U_e$, and be parameterized by a broken power law with index
$p_{1}$ from $\gamma_{1}$ to $\gamma_{b}$ and $p_{2}$ from
$\gamma_{b}$ to $\gamma_{2}$, where $\gamma_{i}$ is Lorentz factor of the electrons.
A $\chi^{2}$-minimization fit was performed to find the best-fit SED model to the observed spectra.
An SSC code developed by \cite{2004ApJ...601..151K} was incorporated into the XSPEC spectral fitting
software \citep{1996ASPC..101...17A} as an external model to perform the minimization
using the Levenburg-Marquadt algorithm\footnote{{\scriptsize \url{https://heasarc.gsfc.nasa.gov/xanadu/xspec/manual/XSappendixLocal.html}}}.
In order to decrease the degeneracy among the model parameters, and after inspecting the 17 broadband SEDs,
we decided to fix the values of the
parameters $\gamma_\text{{min}}$, $\gamma_\text{{max}}$, \textit{R} and $\delta$,
and to set the location of $\gamma_\text{{brk}}$ to be the cooling
break, along with a canonical index change of 1 at $\gamma_\text{{brk}}$ (i.e. $p_2-p_1=1$).
The parameters $\gamma_{min}$,$\gamma_{max}$ are very difficult to constrain with the available broadband SED,
as described in \citet{Mrk501MW2009_Variability}, and it was decided to fix them to 3$\times$10$^2$ and 8$\times$10$^6$ (log$\gamma$=2.5 and 6.9),
which are reasonable values used in the literature \citep[see][]{2011ApJ...727..129A,2015A&A...573A..50A}.
Additionally, the values of $\delta$ and \textit{R} were fixed to reasonable
values that could successfully describe the data and ensure a minimum
variability timescale of 1~day, as no intra-night variability was observed, making this
the fastest variability observed during the three-month period considered in this paper.
A $\delta$ $\sim$10 (which results in \textit{R}$\sim$2.65$\times$10$^{16}$cm by variability arguments)
is a suitable value used to model the emission of
high-peaked BL Lacs such as Mrk 501 \citep[e.g.][]{Mrk501MW2009_Variability}, though it is larger than the modest bulk
Lorentz factors suggested by Very Long Baseline Array measurements \citep{2010ApJ...723.1150P,2004ApJ...600..115P,2002ApJ...579L..67E}.
First, we fit the synchrotron peak to adjust the characteristics of the EED and \textit{B} field. The synchrotron
peak is more accurately determined than
the inverse-Compton, and has a more direct relation to the EED. Then, we fit the inverse-Compton peak, using all parameters from the fit to
the synchrotron peak, and leaving the electron energy density \textit{U}$_\text{{e}}$ as the only free parameter. After that, we fit the broadband SED using the parameter values from
the previous step as starting values. Lastly, we perform a broadband SED fit, using the parameter values from the previous step as starting
values, and loosen slightly the condition that the cooling break occurs at $\gamma_\text{{brk}}$, and that the indices in the EED change by
exactly 1.0. In this last step, we allow \textit{B} and $\gamma_\text{{brk}}$ to vary within $\pm$2\%, and \textit{p}$_\text{{1}}$ and
\textit{p}$_\text{{2}}$ to
vary within $\pm$1\% of the values obtained from the previous step. This last step in the fitting procedure provides a non-negligible
improvement in the data-model agreement, with minimal (a few \%) departures from the canonical values of $\gamma_\text{{brk}}$ and spectral--index
change within the one-zone SSC scenario.
\subsection{Model results}
The results for the 17 broadband SEDs mentioned above can be seen in
Figures~\ref{fig:sedall}, \ref{fig:sedall2} and \ref{fig:sed}. The
corresponding SSC model parameters are listed in Table~\ref{tab:sed1}.
\begin{figure*}
\centering
\includegraphics[height=11.5pc,width=21pc]{figures3/2012_03_23_veritas_F.eps}
\includegraphics[height=11.5pc,width=21pc]{figures3/2012_03_29_veritas_F.eps}
\includegraphics[height=11.5pc,width=21pc]{figures3/2012_04_15_magic_F.eps}
\includegraphics[height=11.5pc,width=21pc]{figures3/2012_04_17_veritas_F.eps}
\includegraphics[height=11.5pc,width=21pc]{figures3/2012_04_19_magic_F.eps}
\includegraphics[height=11.5pc,width=21pc]{figures3/2012_04_21_veritas_F.eps}
\includegraphics[height=11.5pc,width=21pc]{figures3/2012_04_23_magic_F.eps}
\includegraphics[height=11.5pc,width=21pc]{figures3/2012_04_29_veritas_F.eps}
\caption{ Spectral energy distributions (SED) for 8 observations
between MJD 56009 and MJD 57046.
The markers match the following experiments; the
green open triangle OVRO (radio 15 GHz), blue open square Mets$\"{a}$hovi (radio 37 GHz),
red open circle (R--band optical, corrected for host galaxy),
blue open triangles \textit{Swift}/UVOT (UV),
black filled circles \textit{Swift}/XRT (X--ray),
pink open triangles \textit{Swift}/BAT (X--ray),
blue open circles \textit{Fermi}--LAT (gamma rays) and
red/green filled squares/triangles MAGIC/VERITAS (VHE gamma rays).
VHE data are EBL--corrected using \citet{2008A&A...487..837F}.
The BAT energy flux relates to a one-day
average, while the \textit{Fermi}--LAT energy flux relates to
three-day average centered at the VHE observation.
Filled markers are those fit by the theoretical model, while open markers are not.
The black line represents the best fit with a one--zone SSC model,
with the results of the fit reported in Table \ref{tab:sed1}.
}
\label{fig:sedall}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[height=11.5pc,width=21pc]{figures3/2012_05_14_veritas_F.eps}
\includegraphics[height=11.5pc,width=21pc]{figures3/2012_05_19_veritas_F.eps}
\includegraphics[height=11.5pc,width=21pc]{figures3/2012_05_26_veritas_F.eps}
\includegraphics[height=11.5pc,width=21pc]{figures3/2012_05_29_magic_F.eps}
\includegraphics[height=11.5pc,width=21pc]{figures3/2012_05_30_veritas_F.eps}
\includegraphics[height=11.5pc,width=21pc]{figures3/2012_06_12_veritas_F.eps}
\includegraphics[height=11.5pc,width=21pc]{figures3/2012_06_16_magic_F.eps}
\includegraphics[height=11.5pc,width=21pc]{figures3/2012_06_17_magic_F.eps}
\caption{ Spectral energy distributions (SED) for 8 observations
between MJD~57061 and MJD 57095. See
Figure~\ref{fig:sedall} for explanation of markers and
other details.
The black line represents the best fit with a one--zone SSC model,
with the results of the fit reported in Table \ref{tab:sed1}.
}
\label{fig:sedall2}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[height=27.0pc,width=41pc]{figures3/2012_06_09_magic_F.eps}
\includegraphics[height=27.0pc,width=41pc]{figures3/2012_06_09_2zoneB_magic_F.eps}
\caption{
Broadband SED for MJD 56087 (VHE flare from 2012 June 9), fitted with a one-zone SSC model
(top) and a two-zone SSC model (bottom). See Figure~\ref{fig:sedall} for explanation of markers and
other details. In the bottom panel, the green line
depicts the emission of the first (large) zone
responsible for the baseline emission, and the red
line the emission from the second (smaller) zone,
that is responsible for the flaring state. The fit
results from the one-zone SSC model fit are reported
in Table~\ref{tab:sed1}, while those from the two-zone
SSC model fit are
reported in Table~\ref{tab:sed2}.
}
\label{fig:sed}
\end{figure*}
\begin{table*}
\caption{One--zone SSC model results.
The following parameters were fixed: region size (\textit{R}) 2.65$\times$10$^{16}$cm, the Doppler
factor ($\delta$) 10, $\gamma_\text{{min}}$ 3.17$\times$10$^{2}$ and $\gamma_\text{{max}}$ 7.96$\times$10$^{6}$.
V refers to VERITAS and M to MAGIC observations.}
\label{tab:sed1}
\centering
\begin{tabular}{l c c c c c c}
\hline\hline
MJD ($\chi^{2}$/DoF) & \textit{B} & $\gamma_\text{{brk}}$ & \textit{p}$_\text{{1}}$ & \textit{p}$_\text{{2}}$ & \textit{U}$_\text{{e}}$ & $\eta$ \\
& [10$^{-2}$ G] & [10$^6$] & & & [10$^{-3}$ erg/cm$^{3}$] & [\textit{U}$_\text{{e}}$/\textit{U}$_\text{{B}}$] \\
\hline
56009 V (34.0/13) & 2.26 & 0.85 & 1.90 & 2.87 & 11.96 & ~589 \\
56015 V (29.9/11) & 2.34 & 0.81 & 1.90 & 2.87 & ~9.27 & ~425 \\
56032 M (19.9/10) & 2.99 & 0.49 & 1.88 & 2.77 & ~5.20 & ~146 \\
56034 V (24.3/12) & 2.22 & 0.90 & 1.86 & 2.90 & ~6.88 & ~350 \\
56036 M (21.0/11) & 2.00 & 1.07 & 1.93 & 2.96 & 10.50 & ~659 \\
56038 V (19.8/10) & 2.55 & 0.63 & 1.78 & 2.82 & ~4.50 & ~173 \\
56040 M (18.8/11) & 3.00 & 0.51 & 1.91 & 2.93 & ~5.98 & ~166 \\
56046 V (23.5/12) & 3.26 & 0.41 & 1.81 & 2.82 & ~4.30 & ~102 \\
56061 V (24.0/10) & 2.65 & 0.65 & 1.78 & 2.82 & ~4.66 & ~166 \\
56066 V (36.0/12) & 3.39 & 0.42 & 1.70 & 2.73 & ~5.11 & ~112 \\
56073 V (13.3/11) & 2.00 & 1.28 & 1.93 & 2.96 & 11.70 & ~736 \\
56076 M (19.7/10) & 2.13 & 0.81 & 1.69 & 2.70 & ~6.57 & ~361 \\
56077 V (17.7/9) & 1.96 & 1.07 & 1.80 & 2.82 & ~9.29 & ~607 \\
56087 M (62.5/12) & 1.64 & 1.70 & 1.89 & 2.91 & 21.30 & 1398 \\
56090 V (32.7/10) & 2.21 & 0.91 & 1.86 & 2.83 & 10.10 & ~520 \\
56094 M (18.0/10) & 2.98 & 0.50 & 2.00 & 2.97 & ~7.04 & ~199 \\
56095 M (16.8/10) & 2.25 & 0.84 & 1.68 & 2.73 & ~6.78 & ~336 \\
\hline
\end{tabular}
\end{table*}
\begin{table*}
\caption{Two--zone SED model results.
The fixed parameters are the same as in Table \ref{tab:sed1} except
for the size and the energy span of the EED for the flaring zone,
which are $R=$3.3$\times$10$^{15}$~cm, $\gamma_\text{{min}}=$ 2$\times$10$^{3}$ and $\gamma_\text{{max}}=$ 2$\times$10$^{6}$.
}
\label{tab:sed2}
\centering
\begin{tabular}{l c c c c c c}
\hline\hline
MJD ($\chi^{2}$/DoF) & \textit{B} & $\gamma_\text{{brk}}$ & \textit{p}$_\text{{1}}$ & \textit{p}$_\text{{2}}$ & \textit{U}$_\text{{e}}$ & $\eta$ \\
& [10$^{-2}$ G] & [10$^6$] & & & [10$^{-3}$ erg/cm$^{3}$] & [\textit{U}$_\text{{e}}$/\textit{U}$_\text{{B}}$] \\
\hline
Quiescent state & 2.1 & 1.0 & 2.17 & 3.18 & 14.1 & 775 \\
56087 M (31.2/7) & 6.8 & 0.74 & 1.50 & 2.52 & 420 & 2280 \\
\hline
\end{tabular}
\end{table*}
We found that the one-zone SSC model approximately describes the X-ray
and VHE gamma--ray data.
However, the model is not able to produce sufficient emission at eV energies to describe the
optical-UV emission and the soft X--ray emission with a single component.
A similar problem in modelling the broadband SED of Mrk~501 within a
one-zone SSC framework was reported in \cite{Mrk501MW2009_Variability}.
During 2012, the variability in the optical-UV band was less
than 10\%, as reported in Section \ref{sec:variability}, and the
R-band flux was at a historical minimum (over 13 years of observations
performed by the Tuorla group), as mentioned in Section \ref{sec:lightcurves}.
It is therefore
reasonable to assume that this part of the spectrum is dominated by
the emission from a distinct region of the jet, where the emission is
slowly changing on timescales of many weeks.
This new region, if populated by high electron density, could also contribute to the GeV emission.
But this contribution should be characterised by low flux variability (lower than the one measured),
as occurs in the optical emission.
For simplicity, we will not consider the description of the
optical-UV emission in the theoretical scenario presented here, which
focuses on the X-ray and VHE gamma--ray bands, i.e. the
most variable portions of the electromagnetic spectrum and where most
of the energy is emitted.
While only the X--ray and VHE data are strictly simultaneous (within
4 hours) and therefore used in the one-zone SSC model fits,
we note that the three-day average GeV emission (centered on the VHE
observation) measured with
\textit{Fermi}--LAT matches well most model curves on a case-by-case basis.
The notable exceptions are on MJD 56046 and MJD 56095, where the LAT
spectral points (especially the one at the lower energy) deviate
from the theoretical curve, worsening the $\chi^{2}$/DoF of the fit from
23.5/12 to 33.0/14
for the first day and from
16.8/10 to 27.2/12
for the second one.
The combination of the LAT and MAGIC/VERITAS spectral points for these
two days shows a flat gamma--ray bump over four orders of magnitude (from 0.2~GeV to
2~TeV).
The $p$ values of those fits, when considering also the agreement with
the two LAT data points, are 0.3\% and 0.7\%, which is comparable to the data-model agreement from other
broadband SEDs where the LAT spectral
points match well with the model curves (e.g. MJD 56015, 56034).
These two broadband SEDs may hint at the
existence of an additional component emitting at GeV energies, as has already
been proposed by \citet{2015ApJ...798....2S}.
However, using
the data presented in this paper,
the statistical significance is not large enough to make that claim, and we will not
consider additional (and variable) GeV components in our theoretical model.
On the other hand, it is also worth noticing that most of the
\textit{Fermi}--LAT data points are systematically located above (within
1--2 $\sigma$) the SSC model curves, which may be taken as another
hint for the existence of an additional contribution at GeV
energies that is constantly present at some level.
From the fit parameters, we can derive a value for $\eta$, the ratio
of the electron energy density to the magnetic field energy density,
which gives an indication of the departure from equipartition \citep{2016MNRAS.456.2374T}.
Here the values differ from unity by more than two orders of
magnitude, indicating
that the particle population has an excess of energy compared to the magnetic field.
This is a common situation when modeling the broadband SEDs of Mrk~501
(and TeV blazars in general) with a one-zone SSC scenario
\citep[see][]{2001ApJ...554..725T,2011ApJ...727..129A,2015A&A...573A..50A,2015ApJ...812...65F},
which implies more energy in the particles than in the magnetic
field, at least locally where the broadband blazar emission is
produced. It is interesting to note that \citet{2017MNRAS.464.4875B}
employ complete thermal plus
non-thermal distributions in their shock acceleration modeling
of Mrk~501 (2009 campaign) and other blazar multiwavelength spectra,
determining \textit{U}$_\text{{e}}$ consistently, and, using a \textit{B} field of
$\sim$10$^{-2}$~G, arrive at a value of $\eta
\sim 300$ for Mrk~501,
which is very similar (within a factor of $\sim$2) to the
values reported in Table~\ref{tab:sed1} of this manuscript.
The worst SSC model fit by far is the one for MJD 56087 (2012 June 9),
where $\chi^{2}$/DoF=62.5/12 ($p$=$8 \times 10^{-9}$).
This day corresponds to the large VHE gamma--ray flare reported in Section
\ref{sec:lightcurves}, for which the SED shows a peak-like
structure centered at $\sim$2 TeV.
For the sake of completeness, we attempted a fit leaving all the model
parameters free, apart from the relation between \textit{R} and Doppler factor to
ensure a minimum variability of 1 day. This fit yielded a $\chi^{2}$/DoF=30.3/9 (\textit{p}=4$\times$10$^{-4}$). While this fit provides a better data-model agreement,
the obtained model is less physically meaningful because the model parameters are not related as expected in the canonical one-zone SSC
framework (e.g. $\gamma_\text{{br}}$ and $B$, or $p_{\text{1}}$ and $p_{\text{2}}$). Moreover,
this fit requires a $\gamma_\text{{min}}$=6$\times$10$^{4}$, which is an unusually high value for
HBLs such as Mrk~501. Because of that, we attempted a fit with a two-zone SSC scenario with model parameters physically related as we did for
the one-zone SSC scenario described in Section 6.1. In this framework, one relatively large zone
dominates the emission at optical and MeV energies (and
is presumed steady or slowly changing with time).
The other, smaller zone, which is spatially separated from the first, is
characterized by a very narrow electron energy distribution and dominates
the variable emission occurring at X-rays and VHE gamma rays, and
eventually also produces narrow inverse-Compton bumps. This
scenario was successfully used to model
a 13-day-long period of flaring activity
in Mrk~421, as reported in \citet{2015A&A...578A..22A}.
To describe the broadband SED of Mrk~501 measured for MJD 56087, the
EED of the second region was chosen to span over three orders of
magnitude, from $\gamma_\text{{min}}=$2$\times$10$^{3}$ to $\gamma_\text{{max}}=$
2$\times$10$^{6}$, and to have a radius $R$ of 3.3 $\times$ 10$^{15}$
cm which, for a Doppler factor of 10, corresponds to a light--crossing time of three hours,
and hence suitable to describe variability with timescales much shorter than one day.
The broadband SED fitting using the two-zone SSC model is done in the same way as the one-zone SSC model fit described above, but now with twice as many parameters.
The resulting model fit is displayed in Figure~\ref{fig:sed}, and the model parameters reported in
Table~\ref{tab:sed2}. The data-model agreement achieved with this
two-zone SSC scenario yielded $\chi^{2}$/DoF=31.2/7 ($p$=$6 \times
10^{-5}$) which, although this scenario still does not describe the broadband data
satisfactorily, is still several orders of magnitude better than the p value obtained with a single-zone SSC scenario.
\subsection{\textit{Swift}}
The study reported in this paper makes use of the three instruments on
board the \textit{Neil Gehrels Swift Gamma-ray Burst Observatory}
\citep{2004ApJ...611.1005G}; namely the
Burst Alert Telescope \citep[BAT,][]{2005ApJ...633L..77M}, the X-ray
Telescope \citep[XRT,][]{2005SSRv..120..165B} and the Ultraviolet/Optical
Telescope \citep[UVOT,][]{2005SSRv..120...95R}.
The 15-50 keV hard X-ray fluxes from BAT were retrieved from the transient monitor results
provided by the \textit{Swift}/BAT team \citep{0067-0049-209-1-14}\footnote{See
\url{http://swift.gsfc.nasa.gov/results/transients/}}, where we made a
weighted average of all the observations performed within temporal
bins of five days. The BAT count rates are converted to energy flux
using that 0.00022 counts cm$^{-2}$ s$^{-1}$ corresponds to
1.26$\times$10$^{-11}$erg~cm$^{-2}$ s$^{-1}$ \citep{Krimm:2013lwa}.
This conversion is strictly correct only
for sources with the Crab Nebula spectral index in the BAT energy
domain ($\Gamma$=2.1), but the systematic error for sources with
different indices is small and often negligible in comparison with the
statistical uncertainties, as reported in \citet{0067-0049-209-1-14}.
The XRT and UVOT data come from
dedicated observations organized and performed within the framework of
the planned extensive multi-instrument campaign. In this study we consider the 52 Mrk~501 observations performed
between 2012 February 2 (MJD 55959) and 2012 July 30 (MJD 56138). All observations were carried out in
the Windowed Timing (WT) readout mode, with an average exposure of 0.9~ks. The data were
processed using the XRTDAS software package (v.2.9.3), which was developed by the ASI Science Data
Center and released by HEASARC in the HEASoft package (v.6.15.1). The data are
calibrated and cleaned with standard filtering criteria using the \textit{xrtpipeline} task and
calibration files available from the \textit{Swift}/XRT CALDB (version 20140120). For the spectral
analysis, events are selected within a 20-pixel ($\sim$46~arcsecond) radius, which contains 90\% of the
point-spread function (PSF).
The background was estimated from a nearby circular region with a radius of 40~pixels. Corrections for
the PSF and CCD defects are applied from response files generated using the \textit{xrtmkarf} task
and the cumulative exposure map. Before the spectra are fitted the 0.3-10 keV data are binned to
ensure that there are at least 20~counts in each energy bin. The spectra are then corrected for
absorption with a neutral-hydrogen column density fixed to the Galactic 21-cm value in the
direction of Mrk 501, namely 1.55 $\times$ 10$^{20}$ cm$^{-2}$ \citep{2005A&A...440..775K}.
\textit{Swift}/UVOT made between 31 and 52 measurements, depending on the filter used.
The data telemetry volume was reduced using
the \textit{image mode}, where the photon timing information is discarded and the image is directly
accumulated on-board.
In this paper we considered UVOT image data taken within the same observations
acquired by XRT. Here we use the UV lenticular filters, W1, M2 and W2,
which are the ones that are not affected by the strong flux of the host galaxy.
We evaluated the photometry of the source according to the
recipe in \citet{2008MNRAS.383..627P}, extracting source counts with an aperture
of 5~arcsecond radius and an annular background aperture with inner and outer radii
of 20~arcsecond and 30~arcsecond. The count rates were converted to fluxes using the
updated calibrations \citep{2011AIPC.1358..373B}. Flux values were then corrected for mean Galactic extinction using an $E (B - V )$
value of 0.017 \citep{2011ApJ...737..103S} for the UVOT filter effective
wavelength and the mean Galactic interstellar extinction curve in \citet{1999PASP..111...63F}.
\subsection{VERITAS}
The VERITAS experiment (Very Energetic Radiation Imaging Telescope Array System) is an array of
IACTs located at the Fred Lawrence Whipple Observatory in
southern Arizona (1.3~km a.s.l., N 31$^\circ$40$^{\prime}$, W 110$^\circ$57$^{\prime}$). It
consists of four Davies-Cotton-type telescopes.
Full array operations
began in September 2007. Each telescope has a focal length and dish diameter of 12~m.
The total effective mirror area is 106~m$^{2}$ and the camera of each telescope is made up of 499 photomultiplier tubes (PMTs). A single pixel
has a field of view of 0.15$^{\circ}$. The system operates in the
energy range from $\sim$100~GeV to $\sim$50~TeV.
VERITAS has also undergone several upgrades.
In 2009 one of the telescopes was moved in order to make the array more symmetric.
During the summer of 2012 the VERITAS cameras were upgraded by replacing all of the
photo-multiplier tubes \citep{2013arXiv1308.4849D}.
For more details on the VERITAS instrument see
\citet{2008AIPC.1085..657H}.
The performance of VERITAS is characterized by a sensitivity of
$\sim$1\% of the Crab Nebula flux to detect (at 5$\sigma$) a
point-like source in 25 hours of observation, which is equivalent to detecting
(at 5$\sigma$) a $\sim$5\% Crab flux in 1 hour. The uncertainty on the
VERITAS energy calibration is approximately 20\%. The systematic uncertainty on
reconstructed spectral indices is estimated at $\pm$0.2, independent of the source
spectral index, according to studies of \citet{Madhavan2013}. Further details
about the performance of VERITAS can be found on the VERITAS
website\footnote{\url{http://veritas.sao.arizona.edu/specifications}}.
|
\section{Introduction\label{S1}}
Recently, Advanced LIGO has made the first detection of gravitational waves (GWs) and opened a new window to explore very energetic events \citep{2016PhRvL.116f1102A}. The event responsible for the GWs revealed by this first detection, GW150914, was the merger of black holes (BHs) in a binary system and it has been followed by four more detections of merging binary BHs \citep[BBHs,][]{2016PhRvL.116x1103A,2016PhRvX...6d1015A,2017PhRvL.118v1101A,2017PhRvL.119n1101A} and one merging binary neutron stars \citep{2017PhRvL.119p1101A}.
A number of different scenarios for the formation of these merging compact binaries have been proposed so far; the different formation mechanisms proposed have invoked isolated binary evolution
\citep[e.g.][]{2002ApJ...572..407B,2007ApJ...662..504B,2012ApJ...759...52D},
three-body interactions in dense stellar systems \citep[e.g.][]{2000ApJ...528L..17P,2006ApJ...637..937O,2016ApJ...824L...8R,2018ApJ...855..124S},
the orbital evolution of hierarchical systems \citep[e.g.][]{2012MNRAS.422..841A,2014ApJ...781...45A,2016ApJ...816...65A,2018ApJ...856..140H,2018arXiv180508212R},
relativistic captures \citep[e.g.][]{2009MNRAS.395.2127O,2015MNRAS.448..754H,2017PhRvD..96h4009B,2018ApJ...860....5G}.
As for the environment in which these compact binaries might form, the scenarios proposed include globular clusters
\citep[GCs,][]{2010MNRAS.402..371B,2011MNRAS.416..133D,2013MNRAS.435.1358T,2014MNRAS.440.2714B,2015PhRvL.115e1101R,2016PhRvD..93h4029R,2017ApJ...834...68C,2017PASJ...69...94F,2017MNRAS.464L..36A,2017MNRAS.469.4665P,2017arXiv170607053B,2018A&A...615A..91B,2018PhRvD..97j3014S}, young/open clusters \citep[e.g.][]{2014MNRAS.441.3703Z,2017MNRAS.467..524B,2018MNRAS.473..909B} and galactic nuclei \citep[e.g.][]{2009MNRAS.395.2127O,2012ApJ...757...27A,2018MNRAS.474.5672L}.
An important aspect concerning the formation of BHs is the mass fall-back after the supernova explosions. As discussed by \citet{2002ApJ...572..407B}, this fall-back (i.e., failed supernovae) mechanism can increase the remnant BH masses and reduce the natal kicks, which, in turn, can lead to a larger fraction of BHs retained inside the host stellar system \citep{2015ApJ...800....9M,2016MNRAS.458.1450W, 2016MNRAS.463.2109R,2018MNRAS.478.1844A,2018MNRAS.479.4652A}.
The retention of a large number of BHs can significantly influence not only the internal dynamics \citep[e.g.][]{2013MNRAS.432.2779B} but also the observational properties of star clusters \citep[e.g.][]{2008MNRAS.386...65M,2017ApJ...834...68C,2017arXiv171203979W,2018ApJ...855L..15K,2018MNRAS.476.5274L,2018MNRAS.479.4652A,2018MNRAS.478.1844A}. The retained BHs in a dense stellar system rapidly sink to the centre of the system due to the effects of dynamical friction and form a compact subsystem predominantly composed of BHs, on a timescale of few hundreds Myr \citep{2013ApJ...763L..15M}. Due to its short relaxation timescale, a BH subsystem quickly undergoes core collapse and generate energy through the formation and dynamical interactions of BBHs \citep{2013MNRAS.432.2779B}. Recoil velocities acquired during binary-single and binary-binary interactions can result in BHs ejection from GCs, and some numerical studies \citep[e.g.][]{2015ApJ...800....9M,2017MNRAS.469.4665P} suggested that $\sim$30\% of dynamically escaping BHs are in binary systems, some of which are expected to merge within the age of the Universe. Moreover, \citet{2018MNRAS.478.1844A} suggested that some massive Galactic GCs (GGCs) are still harbouring a large number of BHs and that the formation and ejection of BBHs can still be ongoing in those GGCs. The BBHs' properties as well as the merger and detection rates of these BBHs are significantly affected not only by the global properties of host GCs such as the initial mass, size and the metallicity \citep[e.g.][]{2016PhRvD..93h4029R,2017MNRAS.464L..36A,2017ApJ...836L..26C} but also by the GC's stellar initial mass function and the prescriptions for the mass fall-back and the stellar wind \citep[see e.g.][and the references therein]{2017ApJ...834...68C}.
In this paper, we present an analysis of the survey of Monte-Carlo simulations of GCs evolution from \citet{2017MNRAS.464.2511H} and of another set of simulations performed specifically for this paper aimed at a detailed characterization of the link between the properties of BBHs formed in GCs and the structure of the host GCs. Understanding the connection between the properties of the BBHs and those of their host GCs is an important step for more realistic estimates of the merger rate of BBHs. We extracted the information of all escaping BBHs from our GC simulations and found some empirical relations between the properties of merging BBHs and those of the host GCs. These relations provide an essential ingredient to estimate the merger and detection rates of BBHs for any assumed GC system properties and GC formation rate. We also provided examples of estimates of the local merger rate density for various assumptions concerning the properties of GC systems.
The structure of this paper is as follows. We briefly describe the numerical method, the initial conditions and assumptions of our GC simulations in Section \ref{S2}. The relations between the properties of merging BBHs and those of the host GC are presented in Section \ref{S3}. In Section \ref{S4}, we then estimate the local merger rate density based on these relations. We conclude with a summary of our results in Section \ref{S5}.
\section{Methods and Initial Conditions\label{S2}}
The models used for this study are those of the survey of Monte-Carlo simulations presented in \citet{2017MNRAS.464.2511H}.
The simulations followed the evolution of 81 cluster models with a variety of initial number of stars ($N=2\times10^5, 5\times10^5, 10^6$), half-mass radii ($r_{\rm h}=1, 2, 4$ pc), binary fractions (10, 20, 50 per cent) and galactocentric distances ($r_{\rm G}=4, 8, 16$ kpc) \citep[see Table 1 in ][]{2017MNRAS.464.2511H} and were run with the {\sc mocca} code \citep{2013MNRAS.431.2184G,2013MNRAS.429.1221H}. The initial density structure of clusters follows the \citet{1966AJ.....71...64K} density profile with the dimensionless central potential, $W_{0} = 7$. We have adopted \citet{2001MNRAS.322..231K} initial mass function with the mass range of stars from 0.1 to 100 M$_{\odot}$. The metallicity is fixed to $Z = 0.001$ for all our simulation models. All single and binary stars in the simulations evolve according to the stellar evolution recipes \citep[SSE \& BSE,][]{2000MNRAS.315..543H,2002MNRAS.329..897H} implemented in the {\sc mocca} code. We used the stellar wind prescription of SSE and BSE. In all of our simulations, we adopt the mass fall-back mechanism \citep{2002ApJ...572..407B} modifying the natal kicks for BHs.
All our GC simulation models are limited by the tidal field from the host galaxies with a realistic treatment of escaping stars based on \citet{2000MNRAS.318..753F} \citep[see also][]{2013MNRAS.431.2184G}. For the parameter spaces of the initial conditions considered in \citet{2017MNRAS.464.2511H}, the ratio of the half-mass radius to the tidal radius for the GC simulation models ranges from 0.005 to 0.09.
In \citet{2017MNRAS.464.2511H}, we used the initial binary distribution (e.g. eccentricity, semi-major axis and the mass ratio) based on the {\it initial binary population} (hereafter IBP) in which the orbital parameters of short-period proto-binaries are redistributed by mutual interactions between binary components (e.g. mass transfer and tidal circularization) due to the large stellar radii during pre-main-sequence stage as suggested by \citet{1995MNRAS.277.1507K} \citep[see also ][]{2013pss5.book..115K}.\footnote{This {\it pre-main-sequence eigenevolution} \citep{1995MNRAS.277.1507K} was originally postulated to explain the observed properties of Galactic field binary populations originating in embedded clusters. Most recently, \citet{2017MNRAS.471.2812B,2018MNRAS.474.3740B} have provided a modified prescription of the Kroupa IBP for the binary distributions in GCs, which is, however, not applied in this study.} As the results of this proto-binary evolution, short period binaries with large eccentricity are preferentially depleted and tend to have similar masses.
For this study, we have also run another set of 81 simulations with the {\it birth binary population} (hereafter BBP) from \citet{1995MNRAS.277.1507K} that follows the \citet{1991A&A...248..485D} period distribution and the thermal eccentricity distribution to investigate the effects of the initial orbital properties of primordial binaries on the formation, dynamical evolution of BBHs and the rate of merger events among these BBHs from GCs.
\section{bbh mergers from gc simulation models\label{S3}}
\subsection{Correlation between the number of BBH mergers and GC properties\label{S3.1}}
\begin{figure}
\centering
\includegraphics[trim=15 10 5 5,width=1.0\columnwidth]{figure_nmer.eps}
\caption{Number of BBH mergers versus the parameter $\gamma$ defined in Eq. (\ref{E1}). Different symbol types, sizes and colors represent models with different initial number of stars, half-mass radii, galactocentric distances and binary fractions \citep[see also][]{2017MNRAS.464.2511H}. Dashed lines indicate the two times of Poisson errors of the locus line.}\label{F1}
\end{figure}
\begin{figure}
\centering
\includegraphics[trim=15 10 5 5,width=1.0\columnwidth]{figure_npr.eps}
\caption{Same as Fig. \ref{F1} but for only primordial BBH mergers that escape from GCs due to the natal kicks during supernova explosions.}\label{F2}
\end{figure}
\begin{figure}
\centering
\includegraphics[trim=15 10 5 5,width=1.0\columnwidth]{figure_ndy.eps}
\caption{Same as Fig. \ref{F1} but for only dynamical BBH mergers that form dynamically inside GCs by three-body or exchange encounters and subsequently escape.}\label{F3}
\end{figure}
We first focus on the presentation of our results for the simulations with the IBP distribution.
After 12 Gyr of evolution, the 81 GC models explored for this study produced 9519 escaping BBHs and 3402 of them emit GWs and merge within 12 Gyr.
To illustrate the dependence of the GW events on the properties of the host GCs in which the BBH formed, we show the correlation between the number of merging BBHs, $N_{\rm merg}$ and the initial properties of GCs in Fig. \ref{F1}.
We determine the number of BBH mergers that escape from GCs and subsequently merge within 12 Gyr, i.e., $t_{\rm merg}\equiv t_{\rm esc}+t_{\rm Peters} <$ 12 Gyr, where $t_{\rm esc}$ and $t_{\rm Peters}$ are, respectively, the BBH escaping time and the \citet{1964PhRv..136.1224P} timescale for GW coalescence of BBHs calculated using their semi-major axis and eccentricity at the moment of escape. We found that the number of merging BBHs is closely correlated with a parameter, $\gamma$, defined as
\begin{equation}\label{E1}
\gamma\equiv A\cdot \frac{M_{0}}{10^{5}{\rm M}_{\odot}}\times \Big(\frac{\rho_{\rm h}}{10^{5}{\rm M}_{\odot}{\rm pc}^{-3}}\Big)^{\alpha} + B\cdot \frac{M_{0}}{10^{5}{\rm M}_{\odot}}\times f_{\rm b,0}
\end{equation}
where $M_{0}$, $\rho_{\rm h}$ and $f_{\rm b,0}$ are, respectively, the initial total mass, initial half-mass density (i.e., mean density within the half-mass radius) and the initial primordial binary fraction; $A$, $B$ and $\alpha$ are the fitting parameters. Our best fitting result that minimizes the $\chi^{2}$ value is ($A$, $B$, $\alpha$) $=$ (12.53$\pm$0.22, 6.89$\pm$0.84, 0.33$\pm$0.02). The uncertainty on the best-fit parameters is determined by calculating the values for which $\chi^{2}=\chi_{\rm min}^{2}+1$ \citep{1976ApJ...210..642A} by assuming that the $\chi^{2}$ distribution in 1D parameter space is simply a quadratic function.
Eq. (\ref{E1}) implies that there are two main formation channels for BBH mergers from GCs; one, the primordial channel, is related only to the primordial binary fraction and binary stellar evolution and is described by the second term of Eq. (\ref{E1}). The other channel is affected by the cluster's internal dynamics and its contribution to the total number of merging BBHs is described by the first term in Eq. (\ref{E1}) (hereafter we will refer to this as the dynamical channel). The number of BBH mergers from the primordial channel, $N_{\rm merg,p}$, depends, as was to be expected, only on the initial binary fraction and the total mass of GCs. The number of merging BBHs from the dynamical channel, $N_{\rm merg,d}$, on the other hand is the result of the combined effects of a number of processes affected by a variety of structural parameters (e.g. encounter rates, hardening rate per encounters, central velocity dispersion, ejection rate, etc.); our results show that the number of merging BBHs resulting from the complex interplay of all these processes is well described by a parameter with a simple dependence on the cluster's mass and half-mass density.
We point out the number of merging BBHs does not show any significant dependence on the galactocentric distance. For the compact clusters explored in our survey, this is to be expected as the population of merging BBHs escape from clusters as a result of either natal kicks following supernova explosion or ejection from close encounters in the cluster's inner regions. The galactocentric distance and the strength of the tidal field, on the other hand, are relevant for the more gradual evaporation process which is not important for the BBH population studied here. Note, however, that there is a broad trend of metallicity with galactocentric distance \citep[e.g.][]{1994AJ....108.1292D} that can affect the properties of BH populations and the formation of BBHs accordingly.
In order to better illustrate the relative importance of the two channels, we have divided the merging BBHs escaping from our GC simulation models into two groups according to their origin. We have classified BBHs that escape through the natal kick after supernova explosions as merger candidates with the primordial origin (789 of all merging BBHs); all the other merging BBHs are classified as dynamical BBH mergers (2613 of all merging BBHs).
Figs. \ref{F2} and \ref{F3} show the dependence of the number of primordial and dynamical BBH mergers on, respectively, $\gamma_{\rm pri}$ and $\gamma_{\rm dyn}$, the two terms already introduced in the definition of the parameter $\gamma$ in Eq. (\ref{F1}),
\begin{equation}\label{E2}
\gamma_{\rm dyn}\equiv A\cdot \frac{M_{0}}{10^{5}{\rm M}_{\odot}}\times \Big(\frac{\rho_{\rm h}}{10^{5}{\rm M}_{\odot}{\rm pc}^{-3}}\Big)^{\alpha},
\end{equation}
\begin{equation}\label{E3}
\gamma_{\rm pri}\equiv B\cdot \frac{M_{0}}{10^{5}{\rm M}_{\odot}}\times f_{\rm b,0}.
\end{equation}
The best fitting parameters are ($A$, $\alpha$) $=$ (12.30$\pm$0.44, 0.33$\pm$0.01) for dynamical BBH mergers and ($B$) $=$ (6.64$\pm$0.25) for primordial BBH mergers, separately. It is apparent that the best-fit parameters calculated for the two merging BBH population separately are very similar to those obtained from fitting all the BBH mergers together.
\begin{figure}
\centering
\includegraphics[trim=15 10 5 5,width=1.0\columnwidth]{figure_nmer_bbp.eps}
\caption{Same as Fig. \ref{F1} but for GC simulation models with the BBP distribution.}\label{F4}
\end{figure}
\begin{figure}
\centering
\includegraphics[trim=15 10 5 5,width=1.0\columnwidth]{figure_ndy12.eps}
\caption{Same as Fig. \ref{F3} but $\gamma_{\rm dyn, 12}$ is defined by using current GC properties.}\label{F5}
\end{figure}
\begin{figure}
\centering
\includegraphics[trim=15 10 5 5,width=1.0\columnwidth]{figure_ndy12_bbp.eps}
\caption{Same as Fig. \ref{F5} but for the BBP distribution.}\label{F6}
\end{figure}
For models with the BBP distribution, we find 6755 BBHs escaping from GCs in total and 2382 of them merge before 12 Gyr. In Fig. \ref{F4}, we show the relation between $N_{\rm merg}$ and $\gamma$ for the simulations with the BBP distribution. The best fitting parameters for this set of simulations are ($A$, $B$, $\alpha$) $=$ (14.55$\pm$0.31, -0.03$\pm$0.03, 0.38$\pm$0.01). For the BBP distribution, we obtained only 1 primordial BBH merger while all the other BBH mergers have a dynamical origin and this explains the lack of dependence of the number of merging BBHs on the initial binary fraction.
This is due to the pairing of primordial binary components of massive stars. For the BBP distribution, we used a uniform pairing so that the masses of secondary stars are uniformly chosen in between [$m_{\rm min}$, $m_{\rm pri}$]. For the IBP distribution with mass feeding algorithm \citep{2013pss5.book..115K} used in the simulations, the pairing rule for the massive binaries is different.
Some theoretical studies \citep[e.g.][]{2000MNRAS.314...33B,2002MNRAS.336..705B,2007ApJ...661.1034K} for the formation of massive binaries in star forming regions show that massive binaries tend to evolve to the mass ratio $q\sim1$ during the proto-binary stage \citep[note that, however, some mechanisms such as the magnetic breaking during proto-binary evolution can prevent the evolution of mass ratio of proto-binaries toward $q \sim 1$; see e.g.][]{2013ApJ...763....7Z}.
GC simulation models with IBP distribution in this study adopted this condition so that the massive binaries (especially for $m>5$M$_{\odot}$) that can be the progenitors of BHs are more likely to have $q\sim1$ and evolve to BBHs.\footnote{However, \citet{2017MNRAS.471.2812B,2018MNRAS.474.3740B} provided a modified prescription of the Kroupa IBP distribution for GC environments suggesting that the pre-main-sequence eigenevolution and mass feeding algorithm are not applied to massive binaries and that the pairing rule for massive binaries is a uniform pairing based on \citet{2012Sci...337..444S}. If this is the case, the number of primordial BBH mergers from GCs will be negligible and the relation between the number of BBH mergers and GC properties will be similar to that for the BBP distribution in our study.} On the other hand, with the BBP distribution, massive stars are initially coupled with less massive stars and therefore require exchange encounters to become BBHs.
Figs. \ref{F5} and \ref{F6} show the correlation between $N_{\rm merg}$ and the current (at $T = 12$ Gyr) GC properties for models with different initial binary distributions.
We only present the correlation for dynamical BBH mergers in this figure.
The best fitting parameters of Eq. (\ref{E2}) for the correlation between the number of dynamical BBH mergers and the current GC properties are ($A$, $\alpha$) $=$ (303.1$\pm$6.1, 0.43$\pm$0.02) for the IBP distribution and ($A$, $\alpha$) $=$ (527.7$\pm$11.1, 0.52$\pm$0.02) for the BBP distribution, respectively. This relation can be used to estimate current merger rate for the MW GCs or Local Group GCs (see Section \ref{S4.3}). The values of $A$ and $B$ in the relation with the current GC properties are larger than those obtained when the initial GC properties are used because the current masses of GCs are smaller than the initial ones.
\subsection{Time evolution of merger rates\label{S3.2}}
\begin{figure}
\centering
\includegraphics[trim=15 10 5 5,width=1.0\columnwidth]{figure_tmer.eps}
\caption{Time evolution of BBH merger rate normalized to the average merger rate over 12 Gyr. Different colors for histograms represent the time evolution of merger rates from GC simulation models with different initial number of stars. Solid lines show the best fitting results (see Section \ref{S3.2} for further details). }\label{F7}
\end{figure}
In the previous section, we provided a relation between the (initial or current) GC properties and the expected number of BBHs that escape from GCs and merge by emitting GWs within 12 Gyr. However, many numerical studies aimed at the estimation of the merger rate of BBHs from GCs have shown that the merger rate is time-dependent \citep[e.g.][]{2017MNRAS.464L..36A,2017PASJ...69...94F}. Since one of our main goals is to provide an empirical relation that allows to calculate the merger rates and the detection rates of BBH merger events from GCs, we have also calculated a model for the time dependence of the rate of BBH mergers from GCs. In Fig. \ref{F7}, we present the histograms of the merging time, $t_{\rm merg}$, of BBHs, with the numbers normalized to the merger rate averaged over the 12 Gyr of evolution. This figure clearly shows that the merger rate decreases with time very rapidly. Initially the merger rate is as high as 10 times the average merger rate while the merger rate at 12 Gyr is $\sim$5 times lower than the average merger rate. We found that the time evolution of the merger rate is well described by the following expression,
\begin{equation}\label{E4}
\mathcal{R} \equiv \left<\mathcal{R}\right> ae^{-b(t/t_{12})^{c}},
\end{equation}
where $\left<\mathcal{R}\right>$ is the average merger rate over 12 Gyr, and $a$, $b$ and $c$ are the fitting parameters. $\mathcal{R}$ is defined as the number of mergers per unit time bin.
Our best-fit parameters for the time evolution of the merger rate are ($a$, $b$, $c$) $=$ (13.01$\pm$3.00, 4.14$\pm$0.19, 0.35$\pm$0.04) for the IBP distribution.
In order to test the dependence of the best-fit parameters on the cluster's initial number of stars, we repeated the fit for subsets of the simulation data with different initial number of stars ($N=2\times 10^5$, $5\times 10^5$ and $10^6$) and found that the time evolution of the normalized merger rates does not significantly depend on the initial number of stars in GCs. We also tested other subsets with different half-mass radii, galactocentric distances, binary fractions and we did not find any significant discrepancy from the best fitting parameters obtained for the entire survey of simulations.
\begin{figure}
\centering
\includegraphics[trim=15 10 5 5,width=1.0\columnwidth]{figure_tmer_org.eps}
\caption{Same as Fig. \ref{F7} but the BBH mergers used for the fitting are separated into the primordial and dynamical BBH mergers. Dotted line indicates the time evolution of the merger rate for GC simulation models with the BBP distribution. }\label{F8}
\end{figure}
\begin{table}
\begin{center}
\caption{Best fitting results for empirical relations. ``pri'' and ``dyn'' denote the best-fit results using the primordial and dynamical BBH mergers separately. $\gamma_{\rm dyn,12}$ present the expected number of dynamical BBH mergers using current GC properties. Note that for the BBP distribution, the BBH mergers originating from GCs are mostly dynamical BBHs.}
\begin{tabular}{c c c c c}
\hline
\hline
binary & \multicolumn{4}{c}{$N_{\rm merg}$ vs. GC properties (Eq. \ref{E1}, \ref{E2}, \ref{E3})}\\ \cline{2-5}
distribution & & $A$ & $B$ & $\alpha$ \\
\hline
IBP & $\gamma_{\rm tot}$ & 12.53$\pm$0.22 & 6.89$\pm$0.84 & 0.33$\pm$0.02\\
IBP & $\gamma_{\rm pri}$ & - & 6.64$\pm$0.25 & -\\
IBP & $\gamma_{\rm dyn}$ & 12.30$\pm$0.44 & - & 0.33$\pm$0.01\\
IBP & $\gamma_{\rm dyn,12}$ & 303.1$\pm$6.1 & - & 0.43$\pm$0.02 \\
BBP & $\gamma_{\rm tot}$ & 14.55$\pm$0.31 & -0.03$\pm$0.03 & 0.38$\pm$0.01\\
BBP & $\gamma_{\rm dyn,12}$ & 527.7$\pm$11.1 &-& 0.52$\pm$0.02\\\hline\hline
binary & \multicolumn{4}{c}{Time evolution of merger rate (Eq. \ref{E4})}\\ \cline{2-5}
distribution & & $a$ & $b$ & $c$ \\\hline
IBP & $\mathcal{R_{\rm tot}}$ & 13.01$\pm$3.00 & 4.14$\pm$0.19 & 0.35$\pm$0.04\\
IBP & $\mathcal{R_{\rm pri}}$ & 21.80$\pm$8.82 & 5.33$\pm$0.29 & 0.29$\pm$0.05 \\
IBP & $\mathcal{R_{\rm dyn}}$ & 6.15$\pm$1.23 & 3.27$\pm$0.17 & 0.51$\pm$0.06\\
BBP & $\mathcal{R_{\rm tot}}$ & 7.96$\pm$2.03 & 3.54$\pm$0.21 & 0.42$\pm$0.06\\\hline
\label{T1}
\end{tabular}
\end{center}
\end{table}
In Fig. \ref{F8}, we present the time evolution of the merger rates for BBHs with different formation origins and find some differences between the primordial and dynamical BBH merger rates. Since the progenitor BBHs for primordial mergers form in a very short time interval during a GC's early evolution ($T<30$ Myr), the merger rate decreases more rapidly than that for dynamical BBH mergers. The best-fit results are ($a$, $b$, $c$) $=$ (21.80$\pm$8.82, 5.33$\pm$0.29, 0.29$\pm$0.05) for primordial mergers and ($a$, $b$, $c$) $=$ (6.15$\pm$1.23, 3.27$\pm$0.17, 0.51$\pm$0.06) for dynamical mergers. For the BBP distribution, the best-fit parameters are ($a$, $b$, $c$) $=$ (7.96$\pm$2.03, 3.54$\pm$0.21, 0.42$\pm$0.06), similar to the parameters found for the dynamical BBH mergers in models with the IBP distribution.
We summarize our results for the correlation between $N_{\rm merg}$ and GC properties and the time evolution of the merger rates in Table \ref{T1}.
\subsection{Chirp mass and mass ratio of BBH mergers\label{S3.3}}
In this section we investigate some fundamental properties of the merging BBHs. Fig. \ref{F9} shows the distribution of chirp masses, $\mathcal{M}_{\rm chirp}\equiv (m_1m_2)^{3/5}(m_1+m_2)^{-1/5}$ of merging BBHs. It is interesting to note that the chirp mass distribution for BBH mergers throughout all look-back time, $t_{\rm lb}$ and that for BBHs merging in the local Universe with red-shift, $z\leq0.2$ (i.e., $t_{\rm lb}\lesssim2.4$ Gyr with standard cosmological parameters assumed) are slightly different because the formation and merging timescales depend on the mass of BBHs \citep{2017ApJ...836L..26C,2018PhRvL.120o1101R}.
In Fig. \ref{F9} we also plot the values of the chirp masses of BBH mergers detected by LIGO so far and show that they approximately fall within the range of values corresponding to the broad peak in the distribution of $\mathcal{M}_{\rm chirp}$ for merging BBHs.
As illustrated by this figure, for the set of simulations considered in this paper the high-$\mathcal{M}_{\rm chirp}$ events are likely to belong to the dynamical BBH groups while the low-$\mathcal{M}_{\rm chirp}$ could be either primordial or dynamical BBHs. It is also possible that these low-$\mathcal{M}_{\rm chirp}$ sources come from metal-richer environments \citep[see e.g.][]{2017ApJ...836L..26C,2018arXiv180601285A}.
\begin{figure}
\centering
\includegraphics[trim=15 10 5 5,width=1.0\columnwidth]{figure_chmass.eps}
\caption{Distribution of chirp masses, $\mathcal{M}_{\rm chirp}$ of merging BBHs escaping from GC simulation models. Black line shows the distribution of all BBH mergers from GC simulation models with the IBP distribution. Blue and red lines are for primordial and dynamical BBH mergers, respectively. Dotted line shows $\mathcal{M}_{\rm chirp}$ distribution of BBH mergers from models with the BBP distribution. Thin and thick lines represent $\mathcal{M}_{\rm chirp}$ distribution for the BBH mergers throughout all look-back time and those merging in the local Universe ($z\leq0.2$), respectively. Note that the model lines correspond to merger rate densities and are not corrected for observational selection effects. The $\mathcal{M}_{\rm chirp}$ of 5 GW events that have been detected by LIGO so far are marked with the range of 90 per cent confidence intervals (data from \url{https://losc.ligo.org/events/}). }\label{F9}
\end{figure}
We point out that the distribution of $\mathcal{M}_{\rm chirp}$ of dynamical BBH mergers for the IBP distribution is almost identical to that for the BBP distribution (which has only dynamical BBH mergers). This implies that the host GC dynamics is the key factor determining the BBH mergers' properties and differences between the IBP and the BBP initial properties of primordial binaries do not play an important role in the chirp mass of the BBH mergers.
It is interesting to note that recent numerical studies \citep{2018PhRvL.120o1101R,2018ApJ...855..124S} for GCs with post-Newtonian calculations for BBHs suggested that mergers of BBHs can occur inside GCs and the merger product can form a binary with other BHs and merge again in/outside of clusters \citep[see also][for the retention of in-cluster BBH mergers]{2018arXiv180201192M}.
The mass distribution of merging BBHs, especially for high masses, will be affected by this process.
\begin{figure}
\centering
\includegraphics[trim=15 10 5 5,width=1.0\columnwidth]{figure_q_mch.eps}
\caption{Distribution of BBH mergers for the IBP distribution in $q$-$\mathcal{M}_{\rm chirp}$ plane. Grey dots show all mergers and red and blue dots represent, respectively, dynamical and primordial BBHs that merge in the local Universe ($z\leq0.2$).
Black stars and boxes indicate the ranges of $\mathcal{M}_{\rm chirp}$ and $q$ based on 90 per cent confidence intervals of LIGO detections. Upper panel shows the distribution of mass ratio, $q$, of primordial (blue) and dynamical (red) BBHs merging in the local Universe.}\label{F10}
\end{figure}
Fig. \ref{F10} shows the distribution of the $\mathcal{M}_{\rm chirp}$ versus the mass ratio $q$ ($\equiv m_{2}/m_{1}$, where $m_1>m_2$) of BBH mergers from models with the IBP distribution.
The sequence of blue points corresponds to the primordial BBH mergers obtained with the binary stellar evolution and the mass fall-back mechanism used in the simulations (see Section \ref{S2}). Three overdense regions in this plane can be easily identified at $\mathcal{M}_{\rm chirp}\sim20$ and $q\sim1$, $\mathcal{M}_{\rm chirp}\sim13$ and $q\sim0.55$, $\mathcal{M}_{\rm chirp}\sim10$ and $q\sim1$, respectively. This features are related to the shape of the mass function of BHs produced in our simulations, which has a bi-modal distribution with peaks at $m_{\rm BH}\sim12$ and 24 M$_{\odot}$.
In the upper panel in Fig. \ref{F10}, the distribution of $q$ shows that dynamical BBH mergers tend to have similar masses \citep[see also][]{2016MNRAS.458.3075A,2016PhRvD..93h4029R,2017MNRAS.469.4665P}.
It is interesting to point out the presence of a sequence of BBH mergers with high-$\mathcal{M}_{\rm chirp}$ and low-$q$. This group comprises BHs that may have increased their mass due to mergers with other stars or black holes. In the latter case, it may be possible that some of the merger remnants may already have been ejected from the host stellar system due to gravitational wave recoil kicks \citep{2018PhRvL.120o1101R,2018arXiv180201192M}.
An important general point to emphasize is that the $\mathcal{M}_{\rm chirp}$ distribution of merging BBHs strongly depends on the metallicity \citep{2016PhRvD..93h4029R,2017ApJ...836L..26C,2017MNRAS.464L..36A,2018MNRAS.474.2959G}.
The metallicity affects not only the number of BHs produced, the number of BBH mergers but also the mass range of BBHs.
However, \citet{2017ApJ...836L..26C} pointed out that the $\mathcal{M}_{\rm chirp}$ distribution of BBHs formed dynamically and merging in the local Universe ($z\leq0.2$) does not depend on metallicity for $Z\leq0.001$.
We note that the masses of BHs depend also on the single and binary stellar evolution recipes,
\footnote{Note that the common-envelope phase (CEP) is also important for the binary stellar evolution and the formation of compact binaries \citep[e.g.][]{2017MNRAS.468.2429B,2018arXiv180600001G}. In this study we used the CEP parameters, $\alpha_{\rm CE}=3$ and $\lambda=0.5$. However, recent studies \citep[e.g.][]{2017MNRAS.468.2429B} suggested $\alpha_{\rm CE}\sim 0.5$, and lower $\alpha_{\rm CE}$ and $\lambda$ value may lead to more binary mergers during the CEP and the subsequent production of single BHs. The uncertainty in the value of these parameters may affect the number of primordial BBHs and the mass distribution of merging BBHs from GCs.}
and, in particular, on the fall-back prescription \citep[e.g.][]{2012ApJ...749...91F,2015MNRAS.451.4086S}. Additional observations and numerical simulations are therefore needed to constrain the values of the BH masses after supernova explosions.
\section{Merger rate density\label{S4}}
In this section we estimate the merger rate density of BBHs escaping from the GCs using the empirical relations obtained in the previous sections. Since the relation between the expected number, the mass distribution and the time evolution of the rates are different for primordial and dynamical BBH mergers, we calculate the merger rates separately for BBH mergers with different origins.
\subsection{Rate density for primordial BBH mergers\label{S4.1}}
To estimate the merger rate density, we follow the calculation of \citet{2017MNRAS.464L..36A} \citep[see also][]{2004A&A...415..407B}.
For this calculation, we use the merging time and the chirp mass for each primordial BBH that will merge within 12 Gyr from all the simulation models. Having this data and the total and average initial mass of all simulated GCs, we can estimate the merger rate density per unit chirp mass using a GC star formation rate as a function of redshift and the contribution of the merger rate from individual GCs to the rate density according to the age distribution of GCs based on the GC star formation history. For this purpose, the GC star formation rate estimated by \citet{2013MNRAS.432.3250K} has been adopted in this calculation.
We already pointed out that the number of primordial BBH mergers depends on the initial mass and binary fraction of GCs. If we simply assume that the initial binary fraction is universal for all GCs, the merger rate density for primordial BBHs only depends on the GC formation rate.
The number of primordial BBH mergers over 12 Gyr based on the IBP distribution is $\sim$6.64$f_{\rm b,0}$ per $10^5$M$_{\odot}$ from Eq. (\ref{E3}) and its best-fit parameters. On the other hand the number of primordial BBH mergers from the simulations with the BBP distribution is very small and we estimate the contribution of primordial BBH mergers to be negligible in this case.
Our estimate of the number of mergers per unit mass is consistent with that in \citet{2018MNRAS.474.2959G} with similar metallicity. From this number it follows that the local merger rate density of primordial BBHs ranges from 0.18 to 1.8 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ for an initial binary fraction ranging from 10 to 100 per cent.
\subsection{Rate density for dynamical BBH mergers from initial GC properties\label{S4.2}}
In order to estimate the merger rate density for dynamical BBH mergers from our empirical relations, we first need to calculate the number of BBH mergers per unit GC mass. For the calculation of the number of dynamical BBH mergers we need to make an assumption on the initial GC mass and size distribution and then combine these with our estimate of the number of dynamical BBH mergers from Eq. (\ref{E2}). For the initial GC mass function (ICMF), we adopt a \citet{1976ApJ...203..297S} function
\begin{equation}
dN_{\rm GC}\propto M^{-\beta}{\rm exp}(-M/M_{*})dM
\end{equation}
where $\beta=2$ \citep{1999ApJ...527L..81Z,2003A&A...397..473B,2003AJ....126.1836H} and $M_{*}$ is the exponential cut-off mass of the ICMF.
We consider different combinations of the minimum mass of GCs, $M_{\rm min}=10^3, 10^4$M$_{\odot}$ and exponential cut-off mass $M_{*}=10^6, 10^{6.5}$M$_{\odot}$ \citep[for the selection of $M_{*}$, see e.g.][]{2017ApJ...839...78J}.
No firm prediction on the distribution of the initial sizes of GCs is currently available. Instead, we tried to find a realistic distribution of the initial size of GCs from the observations of young massive clusters (YMCs) in extra-galactic systems although it is possible that old GCs forming in the early Universe formed with a different size distribution. There are a number of observational studies \citep[e.g.][]{2010ApJ...709..411H,2012MNRAS.419.2606B,2015MNRAS.452..525R} showing that the effective radii of YMCs tend to increase with the YMC's age. This might be due to the combined effects of the primordial gas expulsion, initial mass loss by the stellar evolution and/or the presence of a significant number of retained BHs \citep[e.g.][]{2008MNRAS.386...65M}. By correcting the age dependence of the effective radii of YMCs in M83 \citep{2015MNRAS.452..525R}, we obtain a log-normal distribution of the initial half-mass radius with $\sigma=0.4$ and $\left<r_{\rm h}\right>=2.8$ pc which is comparable with the initial half-mass radii used in the numerical simulations by \citet{2010ApJ...719..915C,2013MNRAS.429.2881C} reproducing the distribution of GGCs.
Many studies of YMCs found that there is a weak correlation between the mass and the effective radius of YMCs \citep[e.g.][]{1999AJ....118..752Z,2004A&A...416..537L,2010ApJ...709..411H,2012A&A...543A...8M,2017ApJ...841...92R}. We take the relation for the average value of the initial half-mass radius, $\left<r_{\rm h}\right>/{\rm pc}=2.8\times(M/10^4{\rm M}_{\odot})^{0.1}$ from \citet{2004A&A...416..537L}. In order to investigate the effects of the initial size distribution on the merger rate density, we consider another distribution of the initial half-mass radius, $\left<r_{\rm h}\right>/{\rm pc}=0.33\times(M/10^4{\rm M}_{\odot})^{0.13}$ from \citet{2012A&A...543A...8M}, which is much smaller that the previous one \citep[note that these ``small'' and ``large'' size distributions are roughly consistent with the half-mass radii for massive clusters and open clusters/associations from the simulations for the cluster formation done by][]{2016ApJ...817....4F}. We, however, ignore the effects of the host galaxy tidal field on the initial distribution of half-mass radii since there is no correlation between the effective radii and galactocentric distances of YMCs found in nearby galaxies \citep{2017ApJ...841...92R}. \citet{2012ApJ...756..167M} also showed that the galactocentric distance does not significantly affect the early (less than a few hundreds Myr) evolution of half-mass radii of star clusters.
Using the initial mass and size distributions discussed above,
we can estimate the expected number of dynamical BBH mergers per GC masses through Eq. (\ref{E2}) and its fitting parameters as
\begin{equation}
\frac{N_{\rm merg}}{M_{\rm GCSF}}=\frac{\iint \gamma_{\rm dyn} N(M)N(r_{\rm h})dMdr_{\rm h}}{\int N(M)MdM},
\end{equation}
where $M_{\rm GCSF}$ is the total mass of all GCs, and $N(M)$ and $N(r_{\rm h})$ are, respectively, the mass and half-mass radius distribution of initial GC systems (and in which, as explained above, the mean of the half-mass radius distribution depends on the cluster mass).
For the different ICMF we have considered, we find that the total number of dynamical BBHs mergers per unit mass over 12 Gyr based on the ``large'' size distribution \citep{2004A&A...416..537L} is
$\sim$2.45 (2.34) per $10^5$M$_{\odot}$ for [$M_{\rm min}, M_{*}$] = [$10^4$M$_{\odot}, 10^{6.5}$M$_{\odot}$], $\sim$2.11 (1.96) for [$10^4$M$_{\odot}, 10^{6}$M$_{\odot}$], $\sim$1.96 (1.82) for [$10^3$M$_{\odot}, 10^{6.5}$M$_{\odot}$], and $\sim$1.69 (1.53) for [$10^3$M$_{\odot}, 10^{6}$M$_{\odot}$], respectively for the simulation models with the IBP (BBP) distribution.
From these estimates, the corresponding local merger rate densities for the different ICMF are 1.91 (1.73), 1.64 (1.45), 1.52 (1.34) and 1.31 (1.13) ${\rm Gpc}^{-3}{\rm yr}^{-1}$, respectively.
We point out that the ratio of the local merger rate density to the number of mergers per GC masses for dynamical BBHs is larger than that for primordial BBHs due to the chirp mass distribution and time evolution of the merger rates (see Figs. \ref{F8} and \ref{F9}).
We also emphasize that the local merger rate density for dynamically-formed BBHs does not show any significant dependence on the binary distributions and weakly dependent on the ICMF with a variation for the different ICMFs considered of a factor of $\lesssim2$. Along with the local merger rate density for primordial BBHs, our calculation of local merger rate density of $\sim$4 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ is consistent with that of $\sim$5 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ from other literature \citep[e.g.][]{2016PhRvD..93h4029R,2017MNRAS.464L..36A}. Some discrepancies may be due to the different distribution of GC models.
For more compact initial size distribution of GCs suggested by \citet{2012A&A...543A...8M}, we obtain the merger rate density of 14.3 (17.5), 12.6 (15.1), 11.7 (13.9) and 10.3 (12.0) ${\rm Gpc}^{-3}{\rm yr}^{-1}$, respectively for GC models with the IBP (BBP) distribution with the different ICMF. The merger rate density is larger for the BBP distribution because the expected number of dynamical BBH mergers has stronger correlation with the initial density of GCs. The many more additional detections of BBH mergers will be needed to shed light on the initial binary distribution in GCs as well as the distribution of the initial properties of GCs.
We point out that
the mass distribution and the merger rate are nearly independent of the metallicity for $Z\leq0.001$ for either primordial \citep{2018MNRAS.474.2959G} or dynamical \citep{2017ApJ...836L..26C} BBH mergers. By combining the GC star formation history from \citet{2013MNRAS.432.3250K} and the red-shift metallicity relation from \citet{2016Natur.534..512B}, we expect that approximately $\lesssim$10\% of GCs especially forming at lower red-shift ($z\sim2$--3) will be affected by the effects of metallicity. Although we fixed the metallicity to $Z=0.001$, there will not be significant effects of cosmological metallicity variation on the estimation of the local merger rate density.
\subsection{Rate density for dynamical BBH mergers from current GC properties\label{S4.3}}
In Section \ref{S3.1}, we discussed the correlation between the expected number of dynamical BBH mergers and the current GCs' mass and half-mass density. Combining the correlation and the distribution of observed GCs, we also can estimate the merger rate density for BBH mergers originating from surviving GCs. To reproduce the mass distribution of GGCs, we generated the GCMF following an evolved Schechter function \citep{2007ApJS..171..101J},
\begin{equation}
\frac{dN}{dM}\propto \frac{1}{(M+\Delta)^2}{\rm exp} \left( -\frac{M+\Delta}{M_{\rm c}}\right),
\end{equation}
where $\Delta$ is a factor for the mass loss of GCs, and $M_{\rm c}$ is the exponential cut-off mass for the GCMF. We adopted the values of $\Delta=10^{5.4}$M$_{\odot}$ and $M_{\rm c}=10^{5.9}$M$_{\odot}$ from \citet{2007ApJS..171..101J}.
\begin{figure}
\centering
\includegraphics[trim=15 10 5 5,width=1.0\columnwidth]{figure_gcdist.eps}
\caption{Present-day distribution of GCs in the half-mass density and the mass of GCs. Red dots are the Milky Way GCs from the \citet{1996AJ....112.1487H} GGC catalog assuming that the mass-to-light ratio for all GCs is equal to 2 and the half-mass radius is equal to $\sim$1.7 of the projected half-light radius in the catalog. Grey dots are modeled GCs following an evolved Schechter function GCMF \citep{2007ApJS..171..101J} and a log-normal distribution with $\left<r_{\rm h}\right>=6.1$ pc and $\sigma=0.63$ for the distribution of the half-mass radius. Dashed lines indicate the expected number of dynamical BBH mergers produced in the individual GCs over 12 Gyr lifetime. }\label{F11}
\end{figure}
In Fig. \ref{F11}, we show the distribution of GGCs from \citet{1996AJ....112.1487H} catalog in the $M_{\rm GC}$-$\rho_{\rm h}$ plane. We simply assumed that the mass-to-light ratio $\Upsilon=2$ and the half-mass radius $r_{\rm h}= \sim 1.7 R_{\rm hl}$ \citep[where $R_{\rm hl}$ is the projected half-light radius, see ][]{2013MNRAS.429.2881C} for all GGCs.
We then distributed the half-mass radius of GCs at present-day by using a log-normal distribution with parameters $\left<r_{\rm h}\right>=6.1$ pc and $\sigma=0.63$ which give a best-fit with the distribution of \citet{1996AJ....112.1487H} GGCs in the $M_{\rm GC}$-$\rho_{\rm h}$ plane.
Dashed lines in this figure indicate the expected number of dynamical BBH mergers generated from the individual GCs over 12 Gyr cluster lifetime. Many of GGCs are expected to produce between $\sim$10 and $\sim$ 1000 BBH mergers within 12 Gyr.
\begin{table}
\begin{center}
\caption{Estimate of the local merger rate density, $R_{\rm local}$. We used the GC star formation rate from \citet{2013MNRAS.432.3250K} for the calculation of $R_{\rm local}$ of primordial and dynamical BBHs based on the initial GC properties. We used a \citet{1976ApJ...203..297S} function with different parameters for the ICMF. ``large'' and ``small'' denote the initial distribution of the half-mass radii of GCs based on \citet{2004A&A...416..537L} and \citet{2012A&A...543A...8M}, respectively. For the GCMF for the calculation of $R_{\rm local}$ from current GC properties, we used an evolved Schechter \citep{2007ApJS..171..101J} function and a log-normal distribution used in \citet{2016PhRvD..93h4029R}. A log-normal distribution for the initial size ($r_{\rm h}$) distribution of GCs was used. For $\rho_{\rm GC}$, we took conservative, standard and optimistic cases from \citet{2016PhRvD..93h4029R}. We also considered a time-dependent $\rho_{\rm GC}$ from \citet{2017PASJ...69...94F}, as denoted by ``F2017''. See the text for more details.}
\begin{tabular}{c c c c c c}
\hline
\hline
\multicolumn{6}{c}{Primordial BBH mergers (Section \ref{S4.1})}\\
\hline
$f_{\rm b,0}$ & \multicolumn{5}{c}{$R_{\rm local}$ (Gpc$^{-3}$yr$^{-1}$)}\\
\hline
10\% & \multicolumn{5}{c}{0.18} \\
100\% & \multicolumn{5}{c}{1.8}\\
\hline
\hline
\multicolumn{6}{c}{Dynamical BBH mergers w/ initial GC properties (Section \ref{S4.2})}\\
\hline
binary & ICMF & \multicolumn{4}{c}{initial GC size distribution} \\\cline{3-6}
distribution & [$M_{\rm min}, M_{*}$] & \multicolumn{2}{c}{large} & \multicolumn{2}{c}{small}\\\hline
IBP & [$10^4, 10^{6.5}$] M$_{\odot}$ & \multicolumn{2}{c}{1.91} & \multicolumn{2}{c}{14.3}\\
IBP & [$10^4, 10^{6}$] M$_{\odot}$ & \multicolumn{2}{c}{1.64} & \multicolumn{2}{c}{12.6}\\
IBP & [$10^3, 10^{6.5}$] M$_{\odot}$ & \multicolumn{2}{c}{1.52} & \multicolumn{2}{c}{11.7}\\
IBP & [$10^3, 10^{6}$] M$_{\odot}$ & \multicolumn{2}{c}{1.31} & \multicolumn{2}{c}{10.3}\\
BBP & [$10^4, 10^{6.5}$] M$_{\odot}$ & \multicolumn{2}{c}{1.73} & \multicolumn{2}{c}{17.5}\\
BBP & [$10^4, 10^{6}$] M$_{\odot}$ & \multicolumn{2}{c}{1.45} & \multicolumn{2}{c}{15.1}\\
BBP & [$10^3, 10^{6.5}$] M$_{\odot}$ & \multicolumn{2}{c}{1.34} & \multicolumn{2}{c}{13.9}\\
BBP & [$10^3, 10^{6}$] M$_{\odot}$ & \multicolumn{2}{c}{1.31} & \multicolumn{2}{c}{12.0}\\
\hline
\hline
\multicolumn{6}{c}{Dynamical BBH mergers w/ current GC properties (Section \ref{S4.3})}\\
\hline
binary & GCMF & \multicolumn{4}{c}{$\rho_{\rm GC}$ (Mpc$^{-3}$)}\\\cline{3-6}
distribution & & 0.33 & 0.77 & 2.31 & F2017\\\hline
IBP & eSchechter & 0.63 & 1.46 & 4.39 & 5.16 \\
IBP & log-normal & 0.89 & 2.02 & 6.20 & 7.28 \\
BBP & eSchechter & 0.84 & 1.96 & 5.88 & 6.77 \\
BBP & log-normal & 1.15 & 2.68 & 8.04 & 9.26 \\\hline
\label{T2}
\end{tabular}
\end{center}
\end{table}
To compute the local merger rate density from the expected number of BBH mergers from current GCs properties, $\gamma_{\rm dyn,12}$ from Eq. (\ref{E2}) and the distributions of GCs' present-day properties introduced above, we need the number density of GCs, $\rho_{\rm GC}$ in the local Universe. We simply adopt 0.33, 0.77 and 2.31 Mpc$^{-3}$ \citep{2016PhRvD..93h4029R} for conservative, standard and optimistic assumptions of $\rho_{\rm GC}$, respectively. We obtain the local merger rate densities for dynamical BBH mergers are 0.63, 1.46 and 4.39 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ for conservative, standard and optimistic cases, respectively, assuming the age of all GCs is 12 Gyr. The merger rate at the present-day has been corrected by using the Eq. (\ref{E4}) for the case of dynamical BBH mergers (i.e. $\mathcal{R} \sim 0.24 \left<\mathcal{R}\right>$).
We have also considered the time-dependent $\rho_{\rm GC}$ from \citet{2017PASJ...69...94F} for old (T $\ge$ 10 Gyr) GCs. The total $\rho_{\rm GC}$ is slightly smaller (2.2 Mpc$^{-3}$) than the optimistic case of \citet{2016PhRvD..93h4029R}. However, we obtain the merger rate density of 5.16 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ which is slightly larger than our optimistic case based on $\rho_{\rm GC}$ from \citet{2016PhRvD..93h4029R} because there are younger GCs with higher merger rates compared to GCs with ages of 12 Gyr (see Figs. \ref{F7} and \ref{F8}). Our estimates are similar in order of magnitude but systematically smaller than those from other studies such as $\sim$5 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ from \citet{2016PhRvD..93h4029R} (for the standard case) and 13 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ from \citet{2017PASJ...69...94F} \citep[see also][]{2017MNRAS.464L..36A,2017MNRAS.469.4665P}
since our GCMF includes a larger number of GCs with lower masses which contribute less to the merger rate density compared to the more massive GCs.
Using the same GCMF adopted in \citet{2016PhRvD..93h4029R}, we obtain the merger rate density of 0.89, 2.02, 6.20 and 7.28 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ for conservative, standard, optimistic and time-dependent $\rho_{\rm GC}$.
On the other hand, as discussed in previous sections, most of BBH mergers based on the BBP distribution are dynamical mergers. We estimate the merger rate density of 0.84, 1.96, 5.88 and 6.77 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ for conservative, standard, optimistic and time-dependent $\rho_{\rm GC}$, respectively, using the evolved Schechter function GCMF. The estimate of the merger rate density becomes 1.15, 2.68, 8.04 and 9.26 assuming that the GCMF follows a log-normal distribution as used in \citet{2016PhRvD..93h4029R}. We summarize our estimates of the local merger rate density in Table \ref{T2}.
We point out that the calculation of the local merger rate density based on the cluster current properties includes only the contribution of surviving clusters.
GCs dissolving before 12 Gyr of course can contribute to the population of BBH mergers and to take their contribution into account, a calculation like that presented in the previous section must be carried out.
Alternatively we can use a simple toy model for the evolution of a globular cluster system and assume the GC disruption proceeding from the low-mass GCs; using this simple model we can calculate the cumulative fraction of BBH mergers from surviving GCs as a function of the fraction of surviving GCs from our GC models introduced in Section \ref{S4.2}.
We show the result of this calculation in Fig. \ref{F12}. This figure provides an approximate estimate of the fraction of BBH mergers from GCs that still survive at the present-day. If we assume that only $\sim$3 per cent of GCs survive up to now as suggested by \citet{2014ApJ...785...71G} for the Milky Way, the fraction of BBH mergers from surviving GCs varies from 0.3 to 0.7 \citep[0.5 with $M_{\rm min}=10^4 {\rm M}_{\odot}$ used in][]{2014ApJ...785...71G} depending on the ICMF.
The local merger rates based on the current GC properties become comparable with those based on the initial GC properties with ``small'' size distribution when the contribution of merging BBHs from dissolving GCs is taken into account.
\begin{figure}
\centering
\includegraphics[trim=15 10 5 5,width=1.0\columnwidth]{figure_gcdisolv.eps}
\caption{Cumulative fraction of dynamical BBH mergers from surviving GCs as a function of the fraction of surviving GCs. We simply assume that GCs are disrupted from lower-mass GCs. Different lines present the ICMF with different parameters.}\label{F12}
\end{figure}
We conclude this section by pointing out that it is possible that YMCs can contribute to the local merger rate density.
\citet{2017MNRAS.467..524B,2018MNRAS.473..909B} have performed direct $N$-body simulations for YMC-like systems with post-Newtonian approximation implemented and found that YMCs can contribute the detection rate to a similar extent as more massive GC counterpart. \citet{2014MNRAS.441.3703Z} have estimated the local merger rate density of 3.6 Gpc$^{-3}$yr$^{-1}$ for BBHs originating from YMCs in the local Universe.
\citet{2017PASJ...69...94F} also have suggested from their direct $N$-body simulations that the local merger rate density can be up to a factor of $\sim$3 times larger when younger clusters with ages between 2 and 10 Gyr are included in the estimation of the merger rate density. Using the empirical relation for the time evolution of the merger rates in Eq. (\ref{E4}) and the GC formation rate adopted by \citet{2017PASJ...69...94F}, we obtain a local merger rate density about $\sim$6 times higher when younger GCs are included.
When younger clusters are included, an important aspect to consider is the well-known age-metallicity relation for GCs \citep[e.g.][]{2013MNRAS.436..122L}. \citet{2017PASJ...69...94F} considered the effects of the metallicity by limiting the mass of BHs and found that there is no significant effect on the local merger rates. This is, however, in contrast with the findings of \citet{2018MNRAS.474.2959G} who suggested that the number of BBH mergers per unit mass strongly depends on the metallicity. The study of \citet{2018MNRAS.474.2959G} is focused on primordial BBHs, but in the dense environment like GCs, the internal dynamics can in part compensate the effects of the metallicity \citep[see e.g.][]{2017MNRAS.464L..36A,2017ApJ...836L..26C}.
Finally in this study, we did not consider the contribution to the merger rate density by the BBHs that merge inside GCs through the dynamical interactions and binary evolution. However, these in-cluster mergers become more important for very young clusters \citep{2017MNRAS.464L..36A,2017MNRAS.467..524B,2018MNRAS.473..909B}. According to \citet{2017MNRAS.464L..36A}, the contribution of these in-cluster mergers is about 20 per cent of the total merger rate through the entire evolution however becomes less than 1 per cent if the host GCs are old (T $>$ 10 Gyr). It is also important to note that the rate of in-cluster mergers can increase substantially if the dissipative effects connected with GW radiation (i.e. three-body GW capture) are taken into account \citep{2018ApJ...855..124S}.
\section{summary and conclusions\label{S5}}
In this paper we have studied the formation of binary black holes (BBHs) in globular clusters (GCs) and explored the relation between the number and properties of merging BBHs and the structural properties of their host GCs. Our study is based on a large survey of Monte Carlo simulations following the dynamical evolution of GCs with a broad range of different initial masses, sizes and primordial binary properties.
Our results have revealed a close correlation between the number of BBH mergers escaping from GCs and the properties of host GCs such as the initial mass, half-mass radius and the fraction of primordial binaries (Figs. \ref{F1} and \ref{F4}).
We identified two groups of BBH mergers; one group is composed of primordial BBH mergers forming simply as a result of binary stellar evolution and escaping from GCs due to the natal kicks by supernova explosions. The second group is composed of dynamical BBH mergers forming as a result of binary-binary and binary-single interactions in the GC dense environments and ejected from GCs through the dynamical interactions.
The number of primordial BBH mergers is correlated with the GC's initial mass and binary fraction (see Eq. \ref{E3}), while we found that the number of dynamical BBH mergers produced in 12 Gyr is correlated with a parameter $\gamma_{\rm dyn}$ (see Eq. \ref{E2}) depending on the GC's initial mass and half-mass density (Figs. \ref{F2} and \ref{F3}). Interestingly we have shown that the number of dynamical BBH mergers correlates also with the same $\gamma_{\rm dyn}$ parameter but defined in terms of the GC's current properties (Figs. \ref{F5} and \ref{F6}). We provide analytic expressions describing the correlations between the number of BBH mergers and the host GC's properties and apply them to estimate the BBH merger rate for a few different models of GC populations but the expression provided in our study can be used more in general for GC populations with initial conditions different from those adopted in our calculations.
The specific properties of primordial and dynamical BBH mergers such as the merging time and the chirp mass distribution are very important for the estimate of the local merger rate and the detection rate. In general, we find that the merger rate decreases with time due to the continuous ejection of single and binary BHs from GCs (Fig. \ref{F7}). We showed that the time evolution of the merger rate for primordial BBH mergers decreases more rapidly than that for dynamical BBH mergers; this difference is due to differences between the formation and ejection timescales of the two groups of BBH mergers (Fig. \ref{F8}).
The two groups of BBH mergers are characterized also by differences in the chirp masses.
The dynamical BBH mergers contribute more massive BBH mergers compared to the primordial BBH mergers (Figs. \ref{F9} and \ref{F10}).
Based on the analytic expressions obtained from study, we estimated the local merger rates of BBHs escaping from GCs. The local merger rate for primordial BBHs depends only on the cosmological GC formation rate and we obtained a rate of 0.18--1.8 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ (Section \ref{S4.1}) depending on the primordial binary fraction.
To estimate the local merger rate for dynamical BBHs, on the other hand, it is necessary make an assumption on the initial distribution of GC masses and size. As pointed out above, the analytic expressions obtained in this paper allow to calculate the local merger rate for any assumption concerning these initial distributions. We estimated a local merger rate for dynamical BBHs of 1.3--18 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ depending on a variety of combinations of the initial GC mass function and size distribution (Section \ref{S4.2}).
We also estimated a local rate for dynamical BBH mergers from the current properties of surviving GCs equal to 0.6--9.3 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ (Section \ref{S4.3}; see also Table \ref{T2}), assuming all GCs have the same age and metallicity.
The production of BBH mergers from GCs also can be influenced by the formation and the presence of intermediate mass black holes (IMBHs) in GCs. \citet{2015MNRAS.454.3150G} suggested that a seed BH for an IMBH can be formed by the runaway collisions of massive main-sequence (MS) stars \citep[see also][]{2002ApJ...576..899P,2017MNRAS.472.1677S}. This process will preferentially deplete the massive MS progenitors for stellar-mass BHs. Moreover, \citet{2007MNRAS.374..857T} have found that hard binaries can be disrupted by the interactions with the IMBH. These interactions between the IMBH and BBHs might result in the capture of one BH to the IMBH and the ejection of the companion BH \citep[this IMBH-BH binary can deplete the stellar-mass BH population by ejection; see][]{2014MNRAS.444...29L}, which is the possible source of intermediate mass ratio inspirals (IMRIs) for space-based GW detectors \citep[e.g.][]{2002ApJ...581..438M,2009ApJ...698L.129S}. Detailed investigations for the effects of the formation of IMBHs in GCs on the merger rate of stellar-mass BBHs will be studied in our forthcoming papers.
\section*{acknowledgement}
We thank an anonymous referee whose suggestions helped to improve this manuscript.
JH acknowledges support from the China Postdoctoral Science Foundation, Grant No. 2017M610694.
AA was partially supported by the National Science Center (NCN), Poland, through the grant UMO-2015/17/N/ST9/02573.
MG and AA were partially supported by NCN, Poland, through the grant UMO-2016/23/B/ST9/02732 and is currently supported by the Carl Tryggers Foundation through the grant CTS 17:113.
TB was supported by the grant TEAM/2016-3/19 from the Foundation for Polish Science (FNP).
This research was supported in part by Lilly Endowment, Inc., through its support for the Indiana University Pervasive Technology Institute, and in part by the Indiana METACyt Initiative. The Indiana METACyt Initiative at IU is also supported in part by Lilly Endowment, Inc. This work benefited from support by the International
Space Science Institute (ISSI), Bern, Switzerland, through its International
Team programme ref. no. 393 \textit{The Evolution of Rich Stellar Populations
\& BH Binaries} (2017-18).
|
\section{Introduction}
In recent years deep neural networks have revolutionized many domains,
e.g., image recognition, speech recognition, speech synthesis, and knowledge discovery~\citep{krizhevsky2012imagenet, lecun2012efficient, schmidhuber2015deep, lecun2015deep, van2016wavenet}.
Due to their ability to naturally learn from structured data and exhibit superior performance,
they are increasingly used in practical applications and critical decision processes,
such as novel knowledge discovery techniques, autonomous driving or medical image analysis.
To fully leverage their potential it is essential that users can \emph{comprehend and analyze} these processes.
E.g., in neural architecture~\citep{zoph2017learning} or chemical compound searches~\citep{montavon2013machine, schutt2017quantum} it would be extremely useful to know which properties help a neural network to choose appropriate candidates.
Furthermore for some applications understanding the decision process might be a legal requirement.
Despite these arguments neural networks are often treated as black boxes,
because their complex internal workings and the basis for their predictions are not fully understood.
In the attempt to alleviate this shortcoming several methods were proposed, e.g., Saliency Map~\citep{baehrens2010explain, simonyan2013deep}, SmoothGrad~\citep{smilkov2017smoothgrad}, IntegratedGradients~\citep{sundararajan2017axiomatic}, Deconvnet~\citep{zeiler2014visualizing}, GuidedBackprop~\citep{springenberg2015striving}, PatternNet and PatternAttribution~\citep{kindermans2018learning}, LRP~\citep{BachPLOS15, LapCVPR16, LapJMLR16, montavon2018methods}, and DeepTaylor~\citep{MonPR17}.
Theoretically it is not clear which method solves the stated problems best, therefore an empirical comparison is required~\citep{SamTNNLS17, kindermansreliability}.
In order to evaluate these methods, we present \textbf{\emph{iNNvestigate}} which provides a common interface to a variety of analysis methods.
\iffalse
\fi
In particular, \emph{iNNvestigate}\ contributes:
\begin{itemize}
\item A common interface for a growing number of analysis methods that is applicable to a broad class of neural networks.
With this instantiating a method is as uncomplicated as passing a trained neural network to it and allows for easy qualitative comparisons of methods.
For quantitative evaluations of (image) classification task we further provide an implementation of the method ``perturbation analysis''~\citep{SamTNNLS17}.
\item Support of all methods listed above---this includes the first reference implementation for PatternNet and PatternAttribution and an extended implementation for LRP---and an open source repository for further contributions.
\item A clean and modular implementation, casting each analysis in terms of layer-wise forward and backward computations. This limits code redundancy, takes advantage of automatic differentiation, and eases future integration of new methods.
\end{itemize}
\emph{iNNvestigate}\ is available at repository: \url{https://github.com/albermax/innvestigate}.
It can be simply installed as Python package and contains documentation for code and applications.
To demonstrate the versatility of \emph{iNNvestigate}\ we provide examples for the analysis of image classifications for a variety of state-of-the-art neural networks.
\paragraph{Terminology}
The different methods pose different assumption to tasks and are designed for different objectives, yet they are related to ``explaining'' or ``interpreting'' neural networks (see~\citet{montavon2018methods}).
We actively refrain from using this terminology in order to prevent misunderstandings between the design choices of the algorithms and the implicit assumption these terms bring along.
Therefore we will solely use the neutral term \textit{analyzing} and leave any interpretation to the user.
\section{Library}
\paragraph{Interface} The main feature is a common interface to several analysis methods.
The workflow is as simple as passing a Keras neural network model to instantiate an analyzer object for a desired algorithm.
Then, if needed, the analyzer will be fitted to the data and eventually be used to analyze the model's predictions.
The corresponding Python code is:
\begin{lstlisting}
import innvestigate
model = create_a_keras_model()
analyzer = innvestigate.create_analyzer("analyzer_name", model)
analyzer.fit(X_train) # if needed
analysis = analyzer.analyze(X_test)
\end{lstlisting}
\paragraph{Implemented methods} At publication time the following algorithms are supported: Gradient Saliency Map, SmoothGrad, IntegratedGradients, Deconvnet, GuidedBackprop, PatternNet and PatternAttribution, DeepTaylor, and LRP including LRP-Z, -Epsilon, -AlphaBeta.
In contrast, current related work \citep{raghakot2017kerasvis, ancona2018towards} is limited to gradient-based methods.
We intend to further extend this selection and invite the community to contribute implementations as new methods emerge.
\paragraph{Documentation} The library's documentation contains several introductory scripts and example applications. We demonstrate how the analyses can be applied to the following state-of-the-art models: VGG16 and VGG19~\citep{simonyan2014very}, InceptionV3~\citep{szegedy2016rethinking}, ResNet50~\citep{he2016deep}, InceptionResNetV2~\citep{szegedy2017inception}, DenseNet~\citep{huang2017densely}, NASNet mobile, and NASNet large~\citep{zoph2017learning}. Figure~\ref{figdiffmethods} shows the result of each analysis on a subset of these networks.
\begin{figure}[h]
\centering
\includegraphics[height=140px]{different_methods_updated.pdf}
\caption{Result of methods applied to various neural networks (blank, if a method does not support a network's architecture yet).}
\label{figdiffmethods}
\end{figure}
\subsection{Details}
\paragraph{Modular implementation} All of the methods have in common that they perform a back-propagation from the model outputs to the inputs.
The core of \emph{iNNvestigate}\ is a set of base classes and functions that is designed to allow for rapid and easy development of such algorithms.
The developer only needs to implement specific changes to the base algorithm and the library will take care of the complex and error-prone handling of the propagation along the graph structure.
Further details can be found in the repositories documentation.
Another advantage of the modular design is that one can extend any analyzer with a given set of wrappers.
One application of this is the smoothing of the analysis results by adding Gaussian noise to the copies of the input and averaging the outcome.
E.g., SmoothGrad is realized in this way by combining a smoothing wrapper with a gradient analyzer.
\paragraph{Training} PatternNet and PatternAttribution~\citep{kindermans2018learning} are two novel approaches that condition their analysis on the data distribution.
This is done by identifying the signal and noise direction for each neuron of a neural network.
Our software scales favorably, e.g., one can train required patterns for the methods on large datasets like Imagenet~\citep{deng2009imagenet} in less than an hour using one GPU.
We present the first reference implementation of these methods.
\paragraph{Quantitative evaluation} Often analysis methods for neural networks are compared by qualitative (visual) inspection of the result.
This is can lead to subjective evaluations and one approach to create a more objective and quantitative comparison of analysis algorithms is the method ``perturbation analysis''~\citep[also known as ``PixelFlipping'']{SamTNNLS17}.
The intuition behind this method is that perturbing regions which are recognized as important for the classification task by the analyzing method, will impact the classification most.
This allows to assess which analysis method best identifies regions that matter for a specific task and neural network.
\emph{iNNvestigate}\ contains an implementation of this method.
\paragraph{Installation \& license} \emph{iNNvestigate}\ is published as open-source software under the MIT-license and can be downloaded from: \url{https://github.com/albermax/innvestigate}.
It is build as a Python 2 or 3 application on top of the popular and established Keras~\citep{chollet2015keras} framework.
This allows to use the library on various platforms and devices like CPUs and GPUs.
At the time of publication only the TensorFlow~\citep{abadi2016tensorflow} Keras-backend is supported.
The library can be simply installed as Python package.
\iffalse
\subsection{Graph reversal}
\fi
\section{Conclusion}
We have presented \emph{iNNvestigate}, a library that makes it easier to analyze neural networks' predictions and to compare different analysis methods.
This is done by providing a common interface and implementations for many analysis methods as well as making tools for training and comparing methods available.
In particular it contains reference implementations for many methods (PatternNet, PatternAttribution, LRP) and
example application for a large number of state-of-the-art applications.
We expect that this library will support the field of analyzing machine learning and
facilitate research using neural networks in domains such as drug design or medical image analysis.
\acks{Correspondence to MA, SL, KRM, WS and PJK.
This work was supported by the Federal Ministry of Education and Research (BMBF) for the Berlin Big Data Center BBDC (01IS14013A).
Additional support was provided by the BK21 program funded by Korean National Research Foundation grant (No.\ 2012-005741) and the Institute for Information \& Communications Technology Promotion (IITP) grant funded by the Korea government (no.\ 2017-0-00451, No.\ 2017-0-01779).
}
\newpage
\vskip 0.2in
\bibliographystyle{plainnat}
|
\section{The era of precision spectroscopy}
Astronomical spectroscopy is rapidly evolving into a precision science. The possibility to acquire visible spectra of distant objects -- such as medium-to-high-redshift QSOs
at high resolution and with a wavelength accuracy below 1$\,\mathrm{m\,s^{-1}}$
is opening exciting opportunities in
the field fundamental physics, e.g.~the possibility to determine a variation of the fundamental constants \citep{2017RPPh...80l6902M}, or to measure the accelerated expansion of the universe from a redshift drift of distant sources \citep{1962ApJ...136..319S,2008MNRAS.386.1192L}.
Several instruments are being conceived and realized to meet unprecedented requirements in terms of stability and repeatability of the observations;
one is ESPRESSO \citep{2013Msngr.153....6P}, a ultra-stable spectrograph for the ESO Very Large Telescope (VLT), which was intended since its inception as a precursor of the future high-resolution spectrograph \citep{2014SPIE.9147E..23Z} for the ESO Extremely Large Telescope (ELT). ESPRESSO is the first ESO instrument to be equipped with a dedicated Data Analysis Software or DAS \citep{2012SPIE.8448E..1OD,2015ASPC..495..289C,2016SPIE.9913E..3RD}, which is included in the instrument package together with the Data Reduction Software or DRS. This article describes the lessons learned in developing the ESPRESSO DAS, and presents first implementation of the new ``Astrocook'' Python package, which is meant to pave the way towards the next generation of data processing systems \citep{2016SPIE.9910E..2FC}.
\articlefigure{O4-5_f1}{cont}{A portion of the Lyman-$\alpha$ forest of QSO J0515-4410 (black line: flux density; red line: error on flux density) as fitted by the Astrocook package. Information obtained from automated Voigt-profile fitting (green line) of the detected lines (red crosses) is used to locally adjust the guess continuum (dotted blue line); the final continuum is determined by smoothing the result after iteration (blue line).}
\section{The ESPRESSO Data Analysis Software}
ESPRESSO has been purposely designed as an end-to-end ``science machine'', which is fed photons from the telescope and outputs not just calibrated spectra, but actual astrophysical information about the observed targets \citep{2013Msngr.153....6P}. In the case of QSO spectra, such information includes (i) the determination of the emission continuum of the QSO; (ii) the detection of individual absorption features (lines) that can be modeled as the superposition of different Voigt-profile components; and (iii) the interpretation of such lines in terms of absorption systems, including different neutral hydrogen and metal transition at the same redshift, with a given column density and thermal broadening. Emission and absorption features are entangled in a way that requires a simultaneous treatment; most significantly, neutral hydrogen lines giving rise to the so-called Lyman forest require to be fitted with respect to a reference continuum emission, but in turn they provide information about the opacity $\tau_\mathrm{HI}$ as a function of redshift, which can be used to refine the estimation of the continuum itself. Only an iterative approach can cope with such situation and obtain a proper fit of both continuum and lines \citep[see also Sect.~\ref{astrocook}]{2016SPIE.9913E..3RD}.
In developing the ESPRESSO DAS, much effort was put in having the software modules mimic what human observers would do ``by hand'' or ``by eye''. The human brain is extremely powerful at detecting spectral features and mentally subtract them to let the underlying pattern emerge (e.g.~removing a line system to guess the shape of the original emission), but it lacks the capability to handle large quantities of data and is generally prone to subjective bias. The DAS procedures reproduce some manual operations (e.g.~using a detected C\textsc{iv} doublet as a starting point to identify different metal transitions at the same redshift, or adding components to a line system where the fit residuals are high, to improve the model) and perform then automatically along the whole spectrum, or sequentially across a catalogue of spectra.
The DAS code is written mostly in ANSI C, taking advantage of the Common Pipeline Library \citep[CPL,][]{2004SPIE.5493..444M} implemented for the whole ESO Data Flow System. Code modules (``recipes'') can be launched as stand-alone plugins, or invoked through the ESO Reflex interface \citep{2013A&A...559A..96F}, which can be used to implement different pipelines for cascade execution (``workflow'') through a graphical user interface. A fixed workflow has been set up for the analysis of QSO spectra, allowing the users to inspect the recipe products and interactively set up the recipe parameters through a set of Python scripts. The code is distributed under the GNU General Public License; its first public release will be issued after the commissioning of ESPRESSO (early 2018).
The instrument-specific strategy adopted by the DAS has of course some limitations: (i) being tailored to the output of the DRS, it is not easily applied to spectra from other instruments; (ii) its workflow execution is limited by the capability of the ESO Reflex environment, which may not be suited to all situations; (iii) it doesn't allow for finer configuration below the recipe level: recipes are provided as black boxes not meant to be modified by the users; only recipes parameter can be tuned at will. While such features do not impact in the DAS ability to properly handle ESPRESSO spectra, they may hinder the prospective generalization of the DAS model to the future generation of high-resolution spectrographs. To overcome the issue, the Astrocook package was developed.
\section{The Astrocook package}\label{astrocook}
Astrocook is a new Python package to analyze quasar spectra. The name was originated by the tagline ``a thousand recipes to cook a spectrum'', which is meant to emphasize the versatility of the tool. The project is still in its infancy, but a working copy of the package can be downloaded for testing from its GitHub webpage (\url{https://github.com/DAS-OATs/astrocook}). The idea is to create a general representation of a spectrum (and its associated metadata) in the form of a Python object, and to develop an instrument-agnostic set of procedures that the users may invoke through simple scripts, designing their own workflows with the desired level of control in the sequence of operations and in the iteration schemes.
The algorithms developed for the ESPRESSO DAS were divided into atomic constituents, thus breaking the recipe-as-a-black-box constraint of the original code; in addition, several new solutions were added, to improve functionality and address specific analysis requirement. Python 3 was chosen as a development language due to its current prevalence in the astrophysical data analysis community; we took advantage, in particular, of the NumPy \citep{van2011numpy} and Astropy \citep{2013A&A...558A..33A} packages for general data handling, and of the LMFIT package \citep{newville_2014_11813} for non-linear least-square minimization of Voigt profiles.
As an example of the difference between the DAS and the Astrocook we discuss the algorithm to fit the emission continuum by removing the absorption lines. In the DAS, this task is performed by the recipe \texttt{espda\_fit\_qsocont}, which proceeds as follows: (i) the spectrum is split into a blue and a red part, using the Lyman-$\alpha$ emission as a demarcation; (ii) in the red part, previously-detected absorption lines are fitted all at once with respect to an interpolated continuum; (iii) in the blue part, lines are fitted iteratively in spectral chunks, from the strongest to the weakest, with respect to a guess continuum which is determined by taking into account the residual $\tau_\mathrm{HI}$ of the lines yet to be fitted; (iv) the final continuum is determined with a cubic spline interpolation, after the fitted lines are removed from the spectrum. Such procedure is effective in simultaneously interpreting the emission and absorption features, but does not accommodate for further iterations and is subject to local failure which require an ad hoc treatment. Conversely, Astrocook provides a set of methods to perform the operations separately: (i) estimate the guess continuum by detecting and masking the absorption lines and applying a suitable smoothing technique to the masked spectrum (e.g.~a Savitzky-Golay filter); (ii) identify the absorption features as Lyman-$\alpha$ lines or metal doublets and organize them into groups to be fitted together; (iii) fit the absorption features with respect to the guess continuum, automatically adding Voigt component to improve the goodness of fit; (iv) refine the guess continuum based on the information extracted from the line, possibly including the continuum normalization as a free parameter in the line fit; etc. An example of the results is shown in Fig.~\ref{cont}. Each module may run independently or taking prior information from the products of other modules, providing an extreme liberty in designing the procedure. This is precisely the kind of flexibility which is required to tune the code to specific analysis cases while maintaining control over the reproducibility and repeatability of the execution.
The scientific exploitation of Astrocook is ongoing. The package is meant to collect contributions from the QSO data analysis community as large; a porting of the QSFit package for emission line fitting \citep{2017MNRAS.472.4051C} is currently being implemented, and several additions (ranging from the template-based flux calibration to the creation of mock spectra for evaluating the statistical significance of the results) are under consideration. The experience of the DAS ``on the field'', when ESPRESSO comes into operation, will motivate the further development of the package in the years leading to the ELT first light.
|
\section{Introduction and main results}
Quantum entanglement is a profound property of many-body quantum systems~\cite{Horodecki2009}.
Entanglement recently surfaced in many areas of quantum physics, both in condensed matter~\cite{Rev2008Amico} and in high energies~\cite{calabrese2009entanglement,nishioka2009t,Srednicki1993Entropy}. Moreover, entanglement is crucial for the investigation of strongly correlated many-body systems~\cite{Rev2008Amico} and thus a cornerstone in building effective numerical techniques for strongly correlated many-body systems~\cite{White92,Vidal2003Efficient,Schollwock2005,Verstraete2008,schollwock2011density}. It is therefore important to develop experimental protocols for the detection and characterization of entanglement.
Entanglement was recently experimentally measured in bosonic cold atoms~\cite{Islam2015Measuring,Kaufman794}, photonic chips~\cite{pitsios2017photonic}, and trapped ions~\cite{linke2017measuring,brydges2018probing}. Moreover, there have also been many theoretical proposals~\cite{Horodecki2002Method,Abanin2012Measuring,hauke2016measuring} to measure entanglement in a multitude of physical systems, such as quantum dots~\cite{Banchi2016Entanglement}, optical lattices~\cite{alves2004multipartite,daley2012measuring,pichler2013thermal,Pichler2016Measurement,gray2017measuring,Elben2018Renyi,Vermersch2018Unitary},
and Gaussian states~\cite{Weedbrook2012Gaussian}. The method of Refs.~\onlinecite{Islam2015Measuring,Kaufman794} based on many-particle interference is specifically appealing for condensed matter systems since it gives entanglement between macroscopic subsystems containing many particles.
Nearly all of experimental and theoretical advancements in entanglement measurement protocols apply either only to bosonic systems, or require the application of interaction between fermions~\cite{gray2017measuring,Elben2018Renyi,Vermersch2018Unitary}. Measurement protocols for fermionic systems had so far remained elusive due to the inherent difference of their statistics.
Nevertheless, an experimental protocol of the 2\textsuperscript{nd} R\'{e}nyi entropy entanglement measure for fermionic systems, was suggested by Pichler \textit{et al}.~\cite{pichler2013thermal}. Their result, however, was derived in methods that do not directly allow generalizations to other entanglement measures.
In this paper we present measurement protocols for fermionic systems that are directly applicable using current experimental settings~\cite{Islam2015Measuring,Kaufman794,petta2005coherent,trotzky2008time,lloyd2014quantum,folling2007direct,simon2011quantum,Theis2004Tuning,Pichler2016Measurement}. Results are presented that generalize the known bosonic results for arbitrary R\'{e}nyi entropies and negativities.
We also show how they may be used to quantum simulate fermions on manifolds with spin structures.
The generic types of systems we will consider consist of identical particles, with a special focus on fermions, hopping on lattices which are themselves partitioned into two or more subregions. We remark that entanglement emerging due to quantum statistics of identical particles, either bosonic or fermionic, is receiving attention and raising a number of fundamental issues~\cite{ghirardi2002entanglement,PhysRevLett.91.097902,Tichy_2012,PhysRevLett.121.150501}. In view of the diverse literature on this fundamental topic, our focus in this paper is treatment of the specific entanglement-measuring protocols~\cite{alves2004multipartite,daley2012measuring,pichler2013thermal,Pichler2016Measurement,gray2017measuring,GoldsteinSela2018,cornfeld2018imbalance} implemented in Refs.~\onlinecite{Islam2015Measuring,Kaufman794}.
\subsection{Entanglement Entropy}
Entanglement is naturally quantified by the entanglement entropy~\cite{Horodecki2009}, which measures the information in a subsystem with no knowledge of the remaining system. The entanglement entropy was shown to be useful in probing numerous properties of many-body quantum systems~\cite{Rev2008Amico,calabrese2009entanglement,LAFLORENCIE2016,Nielsen1999Conditions,Calabrese2008ES,GoldsteinSela2018,Nielsen2001Separable}, such as quantum critical behaviour~\cite{Vidal2003Entanglement} and non-equilibrium dynamics~\cite{Bardarson2012,daley2012measuring}. For example, the entanglement entropy of scales differently for bosonic and fermionic systems~\cite{PhysRevLett.96.010404,Eisert2010Colloquium}; furthermore, many novel states of matter that cannot be defined by their symmetries, such as topological phases~\cite{Kitaev2006,Levin2006,jiang2012identifying} and spin liquids~\cite{Zhang2011,isakov2011topological}, are discernible by their entanglement entropy scaling properties~\cite{Eisert2010Colloquium}.
For a system in a pure state \(\tket{\psi}\), its density matrix (DM) is given by \(\hat{\rho}=\tket{\psi}\tbra{\psi}\). If one bi-partitions the system into \(A\cup B\), the quantum information available to an observer in region \(A\) is encoded by the reduced DM, \(\hat{\rho}_A=\mathrm{Tr}_B\hat{\rho}\). One may thus use the subsystem von-Neumann entropy \(S(A)=-\mathrm{Tr}\hat{\rho}_A\log\hat{\rho}_A\) to quantify the entanglement between the subsystems. If \(A\) and \(B\) are unentangled, \(\tket{\psi^{AB}}=\tket{\psi^A}\tket{\psi^B}\), then \(S(A)=0\) and \(\mathrm{Tr}\{\hat{\rho}_A^n\}=1\) for any integer \(n\ge1\). One may therefore use other entanglement measures such as the R\'{e}nyi entanglement entropies \cite{Horodecki2009,Mintert2007Observable},
\begin{equation}\label{rendef}
S_n(A)=\frac{1}{1-n}\log\mathrm{Tr}\hat{\rho}_A^n,
\end{equation}
which give various entanglement bounds~\cite{Mintert2005Concurrence,Aolita2006Measuring,mintert2007entanglement} and are directly related to the von-Neumann entropy \(S(A)=\lim_{n\to1}S_n(A)\). Moreover, even if the system \({A\cup B}\) is in a mixed state, one may use these entropies to evaluate the mutual information~\cite{Srednicki1993Entropy,Rev2008Amico,Eisert2010Colloquium} \({I_n(A:B)=S_n(A)+S_n(B)-S_n(A\cup B)}\).
\subsection{Measurement Protocols}\label{sec:meas}
Following theoretical proposals in Refs.~\onlinecite{alves2004multipartite,daley2012measuring}, a measurement of the 2\textsuperscript{nd} R\'{e}nyi entropy was realized by Islam \textit{et al}.~\cite{Islam2015Measuring}. This experimental advancement was accomplished using many-copy protocols where one creates \(n\) identical copies~\cite{Ekert2002Direct,Horodecki2002Method} of a system with a DM \(\hat{\rho}^{\otimes n}=\hat{\rho}\otimes\hat{\rho}\otimes\cdots\otimes\hat{\rho}\). The key theoretical idea put forwards by Daley \textit{et al}.~\cite{daley2012measuring}, is that, for any \emph{bosonic} system, the entropies may be directly extracted from occupancy measurements of the particle number \(N_k^A\) in copy \(k=1\dots n\) of region \(A\).
\begin{figure}[t]
\centering
\includegraphics[width=1.\columnwidth]{FerEntFig1v8.png}
\caption{Schematic description of the measurement protocols, Eqs.~(\ref{mainres}), (\ref{mainres2}), for either entanglement entropy between $A=A_1\cup A_2$ and $B$, or entanglement negativity between $A_1$ and $A_2$. For discussion, see Sec.~\ref{sec:meas}.\label{fig:1}}
\end{figure}
The basic steps of the protocol are depicted in Fig~\ref{fig:1} and are as follows:
(i) Prepare \(n\) copies of a quantum system; Fig.~\ref{fig:1}(a).
(ii) Apply a unitary evolution realizing a Fourier transform (FT) in the \(n\)-copy spase; Fig.~\ref{fig:1}(b).
(iii) Measure the occupancies, \(N_k^A\), in every copy, and evaluate a function of the occupancies, \(f(\{N\})={\textstyle\prod\nolimits_{k=1}^n} e^{\frac{2\pi ik}{n}N^A_k}\); Fig.~\ref{fig:1}(c). It is this function which is affected by fermionic minus signs; see Eq.~(\ref{mainres}).
(iv) Repeat these steps and calculate the average \(\langle f(\{N\})\rangle\).
Such an experimental protocol was carried out for \(n=2\) by Islam \textit{et al}.~\cite{Islam2015Measuring}; see Ref.~\onlinecite{daley2012measuring}. The protocol is encapsulated by the following operator relation,
\begin{equation}\label{renbos}
\mathrm{Tr}\hat{\rho}_A^n=\mathrm{Tr}\big\{f(\{\hat{N}\})\widetilde{\hat{\rho}^{\otimes n}}\big\},
\end{equation}
where, \(\widetilde{\hat{\rho}^{\otimes n}}\) is a Fourier transformed DM; see Eqs.~(\ref{FT}), (\ref{renres}) for a technical definition.
The \emph{first main result} of this paper is the novel adaptation of this protocol to \emph{fermionic} systems, where we find
\begin{equation}\label{mainres}
f(\{N\})=
\begin{cases} \delta_{N_\mathrm{avg}^A\in\mathbb{N}}~(-1)^{{N}_\mathrm{avg}^A}~\prod\limits_{k=1}^n e^{\frac{2\pi ik}{n} {N}_{k}^A} & n~\mathrm{even},\\
\delta_{N_\mathrm{avg}^A\in\mathbb{N}}~\prod\limits_{k=1}^n e^{\frac{2\pi ik}{n} {N}_{k}^A} & n~\mathrm{odd}.
\end{cases}
\end{equation}
Here, \({N}_\mathrm{avg}^A=\frac{1}{n}{N}_\mathrm{tot}^A=\frac{1}{n}\sum_{k=1}^n{N}_k^A\).
In fact, all protocols in this paper are to be interpreted using the same (i-iv) steps; the various entanglement measures differ only by the choice of FTs and of \( f(\{N\})\), they are also brought in a similar form to that of Eq.~(\ref{renbos}).
\subsection{Entanglement Negativity}
\begin{figure*}[t]
\centering
\includegraphics[width=1\linewidth]{Figure2v6.pdf}
\caption{Simulations of our measurement protocols for the Klich-Levitov quench model; see Sec.~\ref{sec:fur}. (top) Potential barrier between two spinless leads is eliminated at ${t=0}$. (bottom) Simulated entanglement measurements at times ${t>0}$, in units of lattice spacing over Fermi velocity. Dots with error bars display protocols' outcomes, solid lines depict exact entanglements. Additional parameters are given Sec.~\ref{sec:num}.
\label{fig:quench}}
\end{figure*}
The entropy ceases being a good measure of entanglement between two subsystems when the system is either open, mixed, or multi-partitioned. Therefore, in these generic cases, other entanglement measures must be deployed~\cite{Horodecki2009,plenio2007introduction}. Entanglement negativity \cite{Peres1996Separability} is both computationally tractable and experimentally viable~\cite{Vidal2002Computable,PhysRevA.58.883,lee2000partial,huang2014computing,gray2017measuring}. It emerged as a principal entanglement measure~\cite{Horodecki2009,plenio2007introduction,eisler2014entanglement,lanyon2017efficient,calabrese2012negativity,Rev2008Amico,LAFLORENCIE2016,Ruggiero2016Negativity,PhysRevA.60.3496,lee2000partial,eisert1999comparison,eisert2006entanglement} and was explored in various contexts ranging from numerical techniques~\cite{Ruggiero2016Negativity,Chung2014negativity,eisler2015partial,Eisler2016Entanglement} to field-theory calculations~\cite{calabrese2012negativity,calabrese2013entanglement,calabrese2014finite,coser2016towards,hoogeveen2015entanglement}.
In such generic cases, it is desirable to quantify the entanglement between two subsystems \(A=A_1\cup A_2\) coupled to an environment \(B\). Using the Peres-Horodecki entanglement criterion~\cite{Peres1996Separability}, one sees that entanglement is entailed by nonvanishing logarithmic negativity \(\mathcal{E}({A_1:A_2})=\log|\hat{\rho}_A^{\mathrm{T}_2}|\).
Here, the partial transpose \(\hat{\rho}_A^{\mathrm{T}_2}\) of a DM \(\hat{\rho}_A\) is given by transposing the indices at region $A_2$, \textit{i.e.} \(\tbra{I^{\sA{1}};J^{\sA{2}}}\hat{\rho}_A^{\mathrm{T}_2}\tket{K^{\sA{1}};L^{\sA{2}}}\!=\!\tbra{I^{\sA{1}};L^{\sA{2}}}\hat{\rho}_A\tket{K^{\sA{1}};J^{\sA{2}}}\).
In an analogous fashion to the entropic case, one may similarly study the R\'{e}nyi entanglement negativities
\begin{equation}\label{negdef}
\mathcal{E}_n({A_1:A_2})=\log\mathrm{Tr}\{(\hat{\rho}_A^{\mathrm{T}_2})^n\},
\end{equation}
such that \(\mathcal{E}({A_1:A_2})=\lim_{n\to1/2}\mathcal{E}_{2n}({A_1:A_2})\).
Recently, a practical proposal, for accurately estimating the negativity in a bosonic setting, using an efficient number of measurements,
was suggested by Gray \textit{et al}.~\cite{gray2017measuring}, utilizing a similar type of many-copy scheme~\cite{cornfeld2018imbalance}.
Our \emph{second main result} is a protocol for measuring the R\'{e}nyi negativities in \emph{fermionic} systems, using the protocol steps (i)-(iv) described above, with suitably modified FT and $f(\{N\})$; see Eqs.~(\ref{FT2}), (\ref{mainres2}).
\subsection{Further Applications and Examples}\label{sec:fur}
As the main results of this paper we obtain measurement protocols that generalize the bosonic protocols for all R\'{e}nyi entanglement entropies \(S_n(A)\thicksim\mathrm{Tr}\hat{\rho}_A^n\) and all R\'{e}nyi entanglement negativities \(\mathcal{E}_n({A_1:A_2})\thicksim\mathrm{Tr}\{(\hat{\rho}_A^{\mathrm{T}_2})^n\}\). This is done in a manner suiting \emph{fermionic} systems, and even, as our \emph{third main result}, allows the simulation of fermions on manifolds with nontrivial spin structure; see Sec.~\ref{sec:spin}.
As an example simulating an experimental implementation of our protocols we consider the Klich-Levitov quench model~\cite{Klich2009Quantum,klich2009many,klich2008scaling,PhysRevB.85.035409,PhysRevB.83.161408,PhysRevB.80.235412,PhysRevB.91.125406,calabrese2009entanglement}, connecting two decoupled noninteracting fermionic tight-binding chains at a certain time, and tracking the resulting growth of entanglement between two subsystems thereof.
While in this model we can easily compute the R\'{e}nyi entropies or negativities exactly, showing entanglement growth and finite size oscillations, extracting quantum averages in a real experiment with a finite sampling of the quantum measurements may become demanding. In order to account for this hurdle and demonstrate that reasonable results can be obtained with a finite sampling, we explicitly simulated a probabilistic measurements of $N^{\sA{1,2}}_k$ within the exact state.
This is done using the techniques of Refs.~\onlinecite{Klich2002FCS,Klich2014FCS,eisler2015partial,Eisler2016Entanglement,GoldsteinSela2018,cornfeld2018imbalance} and the results are presented in Fig.~\ref{fig:quench}.
A more detailed discussion of the model and analyses is given in Sec.~\ref{sec:num}
To simplify the discussion below we concentrate on states where the total particle number in the \emph{entire} system \(A\cup B\) is conserved. However, a closer look shows that only the conservation of total fermion parity is needed, making our results applicable to, \textit{e.g.}, mean-field superconductors~\cite{cornfeld2018entanglement}.
Furthermore, following recent progress in resolving both entropy and negativity into symmetry sectors~\cite{GoldsteinSela2018,cornfeld2018imbalance}, our protocols can be straightforwardly generalized as to directly measure both the symmetry resolved entropy, \(S(N)\thicksim\mathrm{Tr}\{\delta_{\hat{N}^A,N}\hat{\rho}_A^n\}\), and negativity, \(\mathcal{E}(\Delta N)\thicksim\mathrm{Tr}\{\delta_{(\hat{N}^{A_1}-\hat{N}^{A_2}),\Delta N}(\hat{\rho}_A^{\mathrm{T}_2})^n\}\).
Moreover, knowledge of the R\'{e}nyi {} entropies and negativities yields useful information about the entanglement spectrum~\cite{Horodecki2009,Nielsen1999Conditions,Calabrese2008ES,Pichler2016Measurement,GoldsteinSela2018,Nielsen2001Separable} of \(\hat{\rho}_A\) and the negativity spectrum~\cite{eisler2014entanglement,Ruggiero2016Negativity,gray2017measuring,coser2016towards} of \(\hat{\rho}_A^{\mathrm{T}_2}\). These spectra fully characterize all entanglement attributes, and even partial knowledge of the R\'{e}nyi {} entropies and negativities may be utilized to extract valuable entanglement spectra properties.
The paper is organized as follows. In Sec.~\ref{sec:entropies} we introduce the shift operators in $n$-copy spaces as our main tool to compute the R\'{e}nyi entropy. We show how fermionic signs enter this quantity, and how one can use charge conservation to correctly account for these signs in the entanglement measurement protocol. In Sec.~\ref{sec:neg} we move to the case of entanglement negativity and show that it is intrinsically more complicated. We resolve this case by utilizing the local Majorana operator formalism, and show how the fermionic signs can be accounted for by the Fourier transform. In Sec.~\ref{sec:num} we provide an example where we estimate the utility of our protocol in an experiment; we finally conclude in Sec.~\ref{sec:conc}.
\section{Entanglement Entropies}\label{sec:entropies}
\subsection{Shift Operators in $n$-copy Space and Fermionic Signs}
We begin our discussion by introducing the key player in this paper, which is the shift operator \(\hat{V}_A\). It acts on the states \(\tket{ \psi^A_{1},\psi^A_{2},\dots,\psi^A_{n}}\) in the $n$-copy Hilbert space of region $A$, by shifting them among the copies~\cite{Ekert2002Direct},
\begin{eqnarray}\label{Vact}
\hat{V}_A\tket{\psi_{1}^A,\psi_{2}^A,\dots,\psi_{n}^A} &=& \tket{\psi_{n}^A,\psi_{1}^A,\dots,\psi_{n-1}^A},\\
\mathrm{Tr}\hat{\rho}_{A}^n &=& \mathrm{Tr}\{\hat{V}_A\hat{\rho}^{\otimes n}\}.
\end{eqnarray}
One should thus come up with a protocol that measures \(\hat{V}_A\) on the \(n\)-copy system \(\hat{\rho}^{\otimes n}\) to obtain \(\mathrm{Tr}\hat{\rho}_A^n\) and thus the entropies, Eq.~(\ref{rendef}).
Note, that all the following derivations hold more generally for the trace of a product of different DMs, \(\mathrm{Tr}\{\hat{\rho}_{1A}\cdots \hat{\rho}_{nA}\} = \mathrm{Tr}\{\hat{V}_A(\hat{\rho}_1\otimes\cdots\otimes \hat{\rho}_n)\}\), and that we restrict our attention to identical \(\hat{\rho}_k\) for simplicity.
Consider a fermionic state in the occupation basis,
\begin{equation}\label{Mma}
\tket{\mathbf{M}}=\tket{\mathbf{m}_{1},\mathbf{m}_{2},\dots,\mathbf{m}_{n}} = (\hat{\mathbf{a}}^{\dagger}_1)^{\mathbf{m}_{1}}(\hat{\mathbf{a}}^{\dagger}_2)^{\mathbf{m}_{2}}\cdots(\hat{\mathbf{a}}^{\dagger}_n)^{\mathbf{m}_{n}}\tket{0}.
\end{equation}
Here, \([m_{k}]_j\in\{0,1\}\) is the occupation at site \(j\in A\) of copy \(k=1\dots n\), and \((\hat{\mathbf{a}}^{\dagger}_k)^{\mathbf{m}_{k}}\mathrel{\overset{\makebox[0pt]{\mbox{\tiny\rm{def}}}}{=}} \prod_{j\in A}[\hat{a}^{\dagger}_k]_j^{[m_{k}]_j}\), where the product is taken at some fixed order. Let us investigate what becomes of the state under the action of a unitary evolution manifesting a FT in the \(k=1\dots n\) copy space,
\begin{equation}
\label{FT}
\hat{F}\hat{\mathbf{a}}^{\dagger}_{k'}\hat{F}^\dagger=\frac{1}{\sqrt{n}}\sum_{k=1}^n \omega^{k'k}\hat{\mathbf{a}}^{\dagger}_k,
\end{equation}
where \(\omega=e^{\frac{2\pi i}{n}}\) is a primitive root of unity. This transformation may be performed by evolving the system under noninteracting Hamiltonians, and its implementation may be carried out in numerous ways~\cite{Reck1994Experimental,Bovino2005Direct,folling2007direct,daley2012measuring,Pichler2016Measurement} such as a series of beam-splitters.
The evolution acts on the general state as
\begin{equation}\label{FTact}
\hat{F}\tket{\mathbf{m}_{1}, \dots, \mathbf{m}_{n}}= \Big( {\textstyle\sum\limits_{k}}\tfrac{\omega^{1k}}{\sqrt{n}}\hat{\mathbf{a}}^{\dagger}_k \Big)^{\mathbf{m}_{1}}\!\!\!\!\!\!\cdots\Big( {\textstyle\sum\limits_{k}}\tfrac{\omega^{nk}}{\sqrt{n}}\hat{\mathbf{a}}^{\dagger}_k \Big)^{\mathbf{m}_{n}
\tket{0}.
\end{equation}
This can be related to the shift operator \(\hat{V}_A\) by looking at the phase operator \(\hat{U}_A\), used in the bosonic variant with \(f(\{N\})=\prod\nolimits_{k=1}^n \omega^{kN^A_k}\),
\begin{equation}
\label{Udef}
\hat{U}_A\mathrel{\overset{\makebox[0pt]{\mbox{\tiny\rm{def}}}}{=}}{\prod_{k=1}^n}\omega^{k \hat{N}_{k}^A},\qquad \hat{U}_A\hat{\mathbf{a}}^{\dagger}_k=\omega^{k}\hat{\mathbf{a}}^{\dagger}_k\hat{U}_A.
\end{equation}
When acting on the FTed state it yields
\begin{align}
\label{steps}
& \hat{U}_A\hat{F}\tket{\mathbf{M}}
= \Big({\textstyle\sum\limits_{k}}\tfrac{\omega^{(1+1)k}}{\sqrt{n}}\hat{\mathbf{a}}^{\dagger}_{k} \Big)^{\mathbf{m}_{1}}\!\!\!\!\!\!\cdots\Big( {\textstyle\sum\limits_{k}}\tfrac{\omega^{(n+1)k}}{\sqrt{n}}\hat{\mathbf{a}}^{\dagger}_{k} \Big)^{\mathbf{m}_{n}
\tket{0}\nonumber\\
& =\Big({\textstyle\sum\limits_{k}}\tfrac{\omega^{2k}}{\sqrt{n}}\hat{\mathbf{a}}^{\dagger}_{k} \Big)^{\mathbf{m}_{1}}\!\!\!\!\!\!\cdots\Big( {\textstyle\sum\limits_{k}}\tfrac{\omega^{nk}}{\sqrt{n}}\hat{\mathbf{a}}^{\dagger}_{k} \Big)^{\mathbf{m}_{n-1}}\Big({\textstyle\sum\limits_{k}}\tfrac{\omega^{1k}}{\sqrt{n}}\hat{\mathbf{a}}^{\dagger}_{k} \Big)^{\mathbf{m}_{n}
\tket{0} \nonumber\\
& =(-1)^{|\mathbf{m}_n|(|\mathbf{m}_1|+\ldots+|\mathbf{m}_{n-1}|)}\hat{F}\tket{\mathbf{m}_{n},\mathbf{m}_{1}, \dots, \mathbf{m}_{n-1}} \nonumber\\
& =\hat{F}\hat{V}_A(-1)^{\hat{N}_n^A(\hat{N}_{\mathrm{tot}}^A-\hat{N}_n^A)}\tket{\mathbf{M}},
\end{align}
where ${|\mathbf{m}_k|=\sum_{j\in A} [m_k]_j}$.
In exchanging the order of creation operators, we used the fermionic commutation relations and accumulated a phase in order to get the form of Eqs.~(\ref{Vact}), (\ref{FTact}).
Since this equation holds for all states $\tket{\mathbf{M}}$, one has
\begin{equation}
\label{keyrel}
\hat{V}_A=\hat{F}^\dagger\hat{U}_A\hat{F}(-1)^{\hat{N}_n^A(\hat{N}_{\mathrm{tot}}^A-\hat{N}_n^A)}.
\end{equation}
This formal result is crucial for our further analyses. In the bosonic case, where the minus sign is absent, it provides an alternative proof of the measurement protocol by Pichler \textit{et al.}~\cite{Pichler2016Measurement} . Namely, measurements of the function $f(\{N\})$ following the FT, which is nothing but $\hat{F}^\dagger \hat{U}_A \hat{F}$, yield the shift operator in the many-copy space. The latter gives the $n$-th R\'{e}nyi entropy by definition.
Below, we will use Eq.~(\ref{keyrel}) to change the function $f(\{N\})$ in order to account for the fermionic minus sign. We remark that, a priory, this is not obvious since the number operators, $\hat{N}_k^A$, determining this sign are number operators of individual copies, before the FT. The experimental protocol, however, measures the number operators only after the FT. However, as we hereafter demonstrate, conservation laws relate the number operators in different copies and allow one to measure the fermionic signs just from the total number operator.
\subsection{Conservation Laws}
Using this operator relation, we may now successfully turn our attention to entanglement measures, whereby particle conservation becomes useful~\cite{GoldsteinSela2018,cornfeld2018imbalance}.
Let us examine the expectation value of the shift operator over the many-copy reduced density matrix,
\begin{equation}
\hat{\rho}_{A}^{\otimes n}
={\sum\limits_{\mathbf{M}',\mathbf{M}}}\tket{\mathbf{M}'}{\prod\limits_{k}}[\rho_{A}]_{\mathbf{m}'_k,\mathbf{m}^{\phantom{|}}_k}\tbra{\mathbf{M}}.
\end{equation}
A result of particle number conservation is the commutation~\cite{GoldsteinSela2018} of the density matrices with the particle number \({[\hat{\rho}_{A},\hat{N}^A]=0}\), and hence, using index notation,
\(
[\rho_{A}]_{\mathbf{m}'_k,\mathbf{m}^{\phantom{|}}_k}=[\rho_{A}]_{\mathbf{m}'_k,\mathbf{m}^{\phantom{|}}_k}\delta_{|\mathbf{m}'_k|,|\mathbf{m}^{\phantom{|}}_{k}|}
\).
By also utilizing Eq.~(\ref{Vact}) in index form,
\(
\tbra{\mathbf{M}'}\hat{V}_A\tket{\mathbf{M}}={\prod_k}\delta_{\mathbf{m}'_{k+1},\mathbf{m}^{\phantom{|}}_{k}}
\),
one thus gets
\begin{align}\label{delta1}
&\mathrm{Tr}\{\hat{V}_A\hat{\rho}^{\otimes n}\}
=\sum\limits_{\mathbf{M},\mathbf{M}'}\tbra{\mathbf{M'}}\hat{V}_A\tket{\mathbf{M}}\prod\limits_{k}[\rho_{A}]_{\mathbf{m}^{\phantom{|}}_k,\mathbf{m}'_k} \\
&=\sum\limits_{\mathbf{M},\mathbf{M}'}\tbra{\mathbf{M'}}\hat{V}_A\tket{\mathbf{M}}\prod\limits_{k}\big(\delta_{\mathbf{m}'_{k+1},\mathbf{m}^{\phantom{|}}_{k}}\delta_{|\mathbf{m}^{\phantom{|}}_k|,|\mathbf{m}'_{k}|}\big)[\rho_{A}]_{\mathbf{m}^{\phantom{|}}_k,\mathbf{m}'_k}.\nonumber
\end{align}
Upon examining these combined relations one finds
\begin{equation}
|\mathbf{m}_{1}^{\phantom{1}}|=\ldots=|\mathbf{m}_{n}^{\phantom{1}}|=|\mathbf{m}'_{1}|=\ldots=|\mathbf{m}'_{n}|,
\end{equation}
In the \(n\)-copy space, one thus find \(\hat{N}_k^A\) as good quantum numbers. Therefore, when substituting Eq.~(\ref{keyrel}) within the trace, one may impose
\begin{equation}\label{NinN}
N_k^A\cong\frac{1}{n}N_\mathrm{tot}^A=\hat{N}_\mathrm{avg}^A\in\mathbb{N},
\end{equation}
and hence, \((-1)^{\hat{N}_n^A(\hat{N}_{\mathrm{tot}}^A-\hat{N}_1^A)}\cong(-1)^{(n-1)\hat{N}_\mathrm{avg}^A}\), \textit{i.e.},
\begin{equation}
\mathrm{Tr}\{\hat{V}_A\hat{\rho}^{\otimes n}\}=\mathrm{Tr}\{\hat{F}^\dagger\hat{U}_A\hat{F}(-1)^{(n-1)\hat{N}_\mathrm{avg}^A}\hat{\rho}^{\otimes n}\}.
\end{equation}
Moreover, since the total particle number in region $A$ is a FT invariant, \(\hat{N}_\mathrm{tot}^A=\sum_k\hat{\mathbf{a}}^{\dagger}_ {k}\cdot\hat{\mathbf{a}}_{k}^{\phantom{|}}=\hat{F}\hat{N}_\mathrm{tot}^A\hat{F}^\dagger\), we deduce
\begin{equation}\label{Ufer}
\begin{gathered}
\mathrm{Tr}\hat{\rho}_A^n=\mathrm{Tr}\{\hat{V}_A\hat{\rho}^{\otimes n}\} = \mathrm{Tr}\{\hat{U}_A^\mathrm{fer}(\hat{F}\hat{\rho}^{\otimes n}\hat{F}^\dagger)\},\\
\hat{U}_A^\mathrm{fer}\mathrel{\overset{\makebox[0pt]{\mbox{\tiny\rm{def}}}}{=}}(-1)^{(n-1)\hat{N}_\mathrm{avg}^A}\prod_{k=1}^n e^{\frac{2\pi ik}{n} \hat{N}_{k}^A}.
\end{gathered}
\end{equation}
As compared to Eq.~(\ref{keyrel}), this formula which uses particle number conservation, allows the determination of fermionic signs from measurements of the total particle number. The latter is compatible with the experimental protocol for the bosonic case, namely, it does not require one to measure an additional non-commuting observable.
\subsection{Entanglement Entropy Measurement Protocols}
Using the particle number conservation constraints,
Eq.~(\ref{Ufer}) may be recast into the following form
to attain our \emph{first main result}, Eq.~(\ref{mainres}),
\begin{align}\label{renres}
&n~\mathrm{even:} &&\mathrm{Tr}\hat{\rho}_A^n=\mathrm{Tr}\Big\{\delta_{N_\mathrm{avg}^A\in\mathbb{N}}(-1)^{\hat{N}_\mathrm{avg}^A}\!{\textstyle\prod\limits_{k=1}^n}\! e^{\frac{2\pi ik}{n} \hat{N}_{k}^A}\widetilde{\hat{\rho}^{\otimes n}}\Big\}, \nonumber\\
&n~\mathrm{odd:} &&\mathrm{Tr}\hat{\rho}_A^n=\mathrm{Tr}\Big\{\delta_{N_\mathrm{avg}^A\in\mathbb{N}}{\textstyle\prod\limits_{k=1}^n}\! e^{\frac{2\pi ik}{n} \hat{N}_{k}^A}\widetilde{\hat{\rho}^{\otimes n}}\Big\},
\end{align}
where \(\widetilde{\hat{\rho}^{\otimes n}}=\hat{F}\hat{\rho}^{\otimes n}\hat{F}^\dagger\) is the FTed DM.
This operator identity implies that measurements of \(\hat{U}_A^\mathrm{fer}\) on the FTed system, according to the protocol of Eq.~(\ref{renbos}), yield the expectation value of $\hat{V}_A$ and hence the R\'{e}nyi entanglement entropy.
Note that (i) by setting \(n=2\), this result reduces to the known results of Pichler \textit{et al}.~\cite{pichler2013thermal}; and (ii) for odd \(n\), the fermionic minus signs cancel out, and one recovers the known results~\cite{daley2012measuring,GoldsteinSela2018} for the bosonic case.
Let us note a relation with field theory. One may unify the even/odd expressions by absorbing the fermionic minus sign in a relabelling of the FT index $k$ as to run over half integers for even $n$, such that \(\hat{U}_A^\mathrm{fer}=\prod_{k=-(n-1)/2}^{(n-1)/2} e^{\frac{2\pi ik}{n} \hat{N}_{k}^A}\).
This was noticed by the conformal field theory community and used to solve for the entropies in critical fermionic systems~\cite{casini2005entanglement,Cardy2008Form,casini2008analytic,casini2009entanglement,Cornfeld2017Ising}; we herein showed this to hold in fact for any (non-critical) charge conserving system.
\section{Entanglement Negativities and Spin Structures}\label{sec:neg}
Using the valuable properties of the shift operator we may now address the evaluation of the negativities, Eq.~(\ref{negdef}), of subregions \(A=A_1\cup A_2\) for fermionic systems.
In the bosonic case, the simplest way to evaluate the negativities relies on the properties of partial transposition,
\begin{eqnarray}\label{negrel}
\mathrm{Tr}\{(\hat{\rho}_A^{\mathrm{T}_2})^n\}&=&\textstyle\mathrm{Tr}\{\hat{V}_A(\hat{\rho}_A^{\mathrm{T}_2})^{\otimes n}\}=\mathrm{Tr}\{\hat{V}_A(\hat{\rho}_A^{\otimes n})^{\mathrm{T}_2}\} \nonumber\\
&=&\mathrm{Tr}\{\hat{V}_A^{\mathrm{T}_2}\hat{\rho}_A^{\otimes n}\}=\mathrm{Tr}\{\hat{V}_{A_1}^{\phantom{+}}\hat{V}_{A_2}^{-1}\hat{\rho}^{\otimes n}\}.
\end{eqnarray}
Similar to the analysis of R\'{e}nyi entropies in Sec.~\ref{sec:entropies}, when exploring particle number conservation constraints applying to the negativity~\cite{cornfeld2018imbalance}, one gets
\begin{align}\label{delta2}
\tbra{\mathbf{M}'^{\sA{1}}\mathbf{M}'^{\sA{2}}}\hat{V}_A^{\mathrm{T}_2}\tket{\mathbf{M}^{\sA{1}}\mathbf{M}^{\sA{2}}}&=\prod\limits_k\delta_{\mathbf{m}'^{1}_{k+1},\mathbf{m}^{1}_{k}}\delta_{\mathbf{m}'^{2}_{k-1},\mathbf{m}^{2}_{k}},\nonumber\\
\tbra{\mathbf{M}^{\sA{1}}\mathbf{M}^{\sA{2}}}\hat{\rho}_A^{\otimes n}\tket{\mathbf{M}'^{\sA{1}}\mathbf{M}'^{\sA{2}}}&\propto\prod\limits_k\delta_{|\mathbf{m}^{1}_{k}|+|\mathbf{m}^{2}_{k}|,|\mathbf{m}'^{1}_{k}|+|\mathbf{m}'^{2}_{k}|},
\end{align}
\textit{cf.}~Eq.~(\ref{delta1}). Here \({\tket{\mathbf{M}}^{\sA{\alpha}}=\tket{\mathbf{m}_{1}^\alpha,\dots,\mathbf{m}_{n}^\alpha}}\) are the occupations in the \(n\) copies of subregion \(A_{\alpha=1,2}\); see Eq.~(\ref{Mma}).
Upon examining these combined relations we find
\begin{equation}
|\mathbf{m}_n^{\sA{1}}|-|\mathbf{m}_1^{\sA{2}}|=|\mathbf{m}_1^{\sA{1}}|-|\mathbf{m}_2^{\sA{2}}|=\ldots=|\mathbf{m}_{n-1}^{\sA{1}}|-|\mathbf{m}_n^{\sA{2}}|,
\end{equation}
and so within the trace one may impose
\begin{equation}
{N_\mathrm{avg}^{A_1}-N_\mathrm{avg}^{A_2}=\frac{1}{n}(N_\mathrm{tot}^{A_1}-N_\mathrm{tot}^{A_2})}\in\mathbb{Z},
\end{equation}
\textit{cf.}~Eq.~(\ref{NinN}).
Unfortunately, this is a dead end,
as the fermionic minus signs of Eq.~(\ref{keyrel}) are not resolved by these constraints. To proceed, we must utilize a different venue.
\subsection{Local Operators}
A useful approach towards negativities (from both the analytical and numerical perspectives) is the local operator formalism. Any operator, and hence any reduced DM, may be expanded~\cite{Vidal2003Entanglement} by a basis of local Majorana operators, \(\hat{\mathbf{a}}^{\sA{1,2}}=\tfrac{1}{2}(\hat{\bm{\gamma}}^{\sA{1,2}}+i\hat{\bm{\gamma}}'^{\sA{1,2}})\), in the form of
\begin{equation}
\hat{\rho}_A=\sum_{\bm{\mu}^{1}\bm{\mu}^{2}}w_{\bm{\mu}^{1}\bm{\mu}^{2}}(\hat{\bm{\gamma}}^{\sA{1}})^{\bm{\mu}^{1}}(\hat{\bm{\gamma}}^{\sA{2}})^{\bm{\mu}^{2}},
\end{equation}
where \([\mu^{1,2}_{(k)}]_j\in\{0,1\}\) for all Majorana operators in region \(A_{1,2}\) (of copy \(k\)). When performing such an expansion, a natural choice of transposition is~\cite{eisler2015partial}
\begin{equation}\label{T2def}
\begin{gathered}
\hat{\rho}_A^{\mathrm{T}_2}=\tfrac{1-i}{2}\hat{\rho}_A^{+}+\tfrac{1+i}{2}\hat{\rho}_A^{-},\\
\hat{\rho}_A^{\pm}\mathrel{\overset{\makebox[0pt]{\mbox{\tiny\rm{def}}}}{=}}\sum_{\bm{\mu}^{1}\bm{\mu}^{2}}w_{\bm{\mu}^{1}\bm{\mu}^{2}}(\hat{\bm{\gamma}}^{\sA{1}})^{\bm{\mu}^{1}}(\hat{\bm{\gamma}}^{\sA{2}})^{\bm{\mu}^{2}}(\pm i)^{|\bm{\mu}^2|},
\end{gathered}
\end{equation}
where \(|\bm{\mu}^{1,2}|=\sum_{j\in A_{1,2}} [\mu^{1,2}]_j\).
This decomposition is highly beneficial, as originally noted in the context of Gaussian DMs~\cite{eisler2015partial,coser2015partial,coser2016towards,coser2016spin,Eisler2016Entanglement,cornfeld2018imbalance}. It was proven~\cite{eisler2015partial} that \(\hat{\rho}_A^{\pm}\) are Gaussian as well.
To appreciate this relation, let us look at the 3\textsuperscript{rd} R\'{e}nyi negativity, \(\mathcal{E}_3({A_1:A_2})\),
\begin{equation}\label{neg3}
\mathrm{Tr}\{(\hat{\rho}_{A}^{\mathrm{T}_2})^3\}=-\frac{1}{2}\mathrm{Tr}\{(\hat{\rho}^+_A)^3\}+\frac{3}{2}\mathrm{Tr}\{(\hat{\rho}^+_A)^2\hat{\rho}^-_A\}.
\end{equation}
The decomposition hereby allows one to evaluate the negativity by tracing Gaussian matrices, which is a tractable task both numerically and analytically. We herein show that the decomposition is useful beyond the Gaussian case, and is valuable for any fermionic system.
Measurements of negativities \(\mathrm{Tr}\{(\hat{\rho}_{A}^{\mathrm{T}_2})^n\}\) are reduced to separate evaluations of monomials, \(\mathrm{Tr}\{\hat{\rho}_A^{\sigma_1}\hat{\rho}_A^{\sigma_2}\cdots\hat{\rho}_A^{\sigma_n}\}\) with \(\sigma_k=\pm\), \textit{cf.} Eq.~(\ref{neg3}). Let us begin by examining the simplest case of a pure monomial \(\mathrm{Tr}\{(\hat{\rho}_{A}^\pm)^n\}\).
Key properties of the partial transposition, Eq.~(\ref{T2def}), are that \((\hat{N}^{A_2})^{\mathrm{T}_2}=|A_2|-\hat{N}^{A_2}\), where $|A_2|$ is the number of sites in region $A_2$, and that therefore~\cite{cornfeld2018imbalance}
\begin{equation}
0=[\hat{\rho}_A,\hat{N}^A]^{\mathrm{T}_2}=[\hat{\rho}_A^{\mathrm{T}_2},\hat{N}^{A_1}-(\hat{N}^{A_2})^{\mathrm{T}_2}]=[\hat{\rho}_A^{\mathrm{T}_2},\hat{N}^A].
\end{equation}
It is straightforward to check that \([\hat{\rho}_A^\pm,\hat{N}^A]=0\) by construction. We hence recover charge conservation, which was the only prerequisite for our entanglement entropy results. We thus get \(\mathrm{Tr}\{(\hat{\rho}_{A}^\pm)^n\}=\mathrm{Tr}\{\hat{U}_A^\mathrm{fer}\hat{F}(\hat{\rho}_A^\pm)^{\otimes n}\hat{F}^\dagger\}\), by applying Eq.~(\ref{Ufer}). Following a rather technical Majorana-operator calculation (see Appendix~\ref{app:pure}) one finds that within the trace \(\hat{U}_{A_2}^\mathrm{fer}\cong(\hat{U}_{A_2}^\mathrm{fer})^{-1}(-1)^{\frac{1}{2}\sum_{k=1}^n|\bm{\mu}^{2}_k|}\). This cancels the excess phases \((\pm i)\) of the decomposed DMs Eq.~(\ref{T2def}) and restores the structure of the negativity measurements, Eq.~(\ref{negrel}),
\begin{align}\label{Uneg}
&\mathrm{Tr}\{(\hat{\rho}_{A}^\pm)^n\}= \mathrm{Tr}\{\delta_{(N_\mathrm{avg}^{A_1}-N_\mathrm{avg}^{A_2})\in\mathbb{Z}}\hat{U}_{A}^\mathrm{neg}(\hat{F}\hat{\rho}^{\otimes n}\hat{F}^\dagger)\},\nonumber\\
&\hat{U}_A^\mathrm{neg}\mathrel{\overset{\makebox[0pt]{\mbox{\tiny\rm{def}}}}{=}}\hat{U}_{A_1}^\mathrm{fer}(\hat{U}_{A_2}^\mathrm{fer})^{-1}\\
&\phantom{\hat{U}_A^\mathrm{neg}}=(-1)^{(n-1)\big(\hat{N}_{\mathrm{avg}}^{A_1}-\hat{N}_{\mathrm{avg}}^{A_2}\big)}\prod_{k=1}^n e^{\frac{2\pi ik}{n} \big(\hat{N}_{k}^{A_1}-\hat{N}_{k}^{A_2}\big)}.\nonumber
\end{align}
This result incorporates the particle number conservation constraints and directly generalizes both the known bosonic negativity measurement schemes~\cite{gray2017measuring,cornfeld2018imbalance} and our fermionic entropy results, Eqs.~(\ref{mainres}), (\ref{Ufer}), (\ref{renres}). When applying the general protocol of Eq.~(\ref{renbos}), it implies that measurements of \(\hat{U}_A^\mathrm{fer}\) on the FTed system yields the pure monomial part of the negativity; \textit{e.g.}, \(\mathrm{Tr}\{(\hat{\rho}^+_A)^3\}\) of Eq.~(\ref{neg3}). Note, that these pure monomials are in fact equivalent to the recently proposed partial time-reversal negativity entanglement measure~\cite{Shapourian2017Partial}. This partial time-reversal appears in the topological characterization of interacting fermionic systems and constitutes a signature of many-body topological phases~\cite{Shapourian2017Many,Shiozaki2018Many}.
\subsection{Generalized Fourier Transform}
All that is left in order to complete the protocol for the measurement of the full negativities is the treatment of mixed monomials \(\mathrm{Tr}\{\hat{\rho}_A^{\sigma_1}\hat{\rho}_A^{\sigma_2}\cdots\hat{\rho}_A^{\sigma_n}\}\), \textit{e.g.} \(\mathrm{Tr}\{(\hat{\rho}^+_A)^2\hat{\rho}^-_A\}\) of Eq.~(\ref{neg3}). This may be accomplished by considering the parity operator, \(\hat{P}^{A_2}=(-1)^{\hat{N}^{A_2}}\),
which relates the decomposed matrices \(\hat{\rho}_A^\mp=\hat{P}^{A_2}\hat{\rho}_A^\pm\hat{P}^{A_2}\). By deploying the parity on appropriate copies, one finds
\begin{equation}
\mathrm{Tr}\{\hat{\rho}_A^{\sigma_1}\hat{\rho}_A^{\sigma_2}\cdots\hat{\rho}_A^{\sigma_n}\} =\mathrm{Tr}\{\hat{V}_A\hat{P}^{A_2}_{\{\sigma\}}(\hat{\rho}_A^{+})^{\otimes n} \hat{P}^{A_2}_{\{\sigma\}}\},
\end{equation}
where
\(\hat{P}^{A_2}_{\{\sigma\}} \mathrel{\overset{\makebox[0pt]{\mbox{\tiny\rm{def}}}}{=}}\prod_{k=1}^n(\hat{P}^{A_2}_k)^{\delta_{\sigma_1,(-)}}\).
We thus get a generalization of the pure monomial result of Eq.~(\ref{Uneg}),
\begin{equation}\label{mixed}
\mathrm{Tr}\{\hat{\rho}_A^{\sigma_1}\cdots\hat{\rho}_A^{\sigma_n}\!\} \!=\! \mathrm{Tr}\{\delta_{(N_\mathrm{avg}^{A_1}-N_\mathrm{avg}^{A_2})\in\mathbb{Z}}\hat{U}_{A}^\mathrm{neg}(\hat{F}_{\{\sigma\}}^{\phantom{|}}\hat{\rho}^{\otimes n}\hat{F}_{\{\sigma\}}^{\dagger})\}.
\end{equation}
Here, we utilized a generalized FT, \(\hat{F}_{\{\sigma\}}\mathrel{\overset{\makebox[0pt]{\mbox{\tiny\rm{def}}}}{=}}\hat{F}^{A_{1}}_{\phantom{|}}\hat{F}^{A_{2}}_{\phantom{|}} \hat{P}^{A_{2}}_{\{\sigma\}}\), which satisfies
\begin{equation}
\label{FT2}
\hat{F}_{\{\sigma\}}^{\phantom{|}}\hat{\mathbf{a}}^{A_{2}\dagger}_{k'}\hat{F}_{\{\sigma\}}^{\dagger}=\frac{1}{\sqrt{n}}\sum_{k=1}^n \sigma_{k'}\omega^{k'k}\hat{\mathbf{a}}^{A_{2}\dagger}_k,
\end{equation}
\textit{cf.} Eqs. (\ref{FT}), (\ref{Udef}). This modified FT may be readily achieved using similar unitary evolution protocols applied for the standard FT~\cite{Reck1994Experimental,Bovino2005Direct,folling2007direct}.
\begin{figure}[t]
\centering
\includegraphics[width=1.\columnwidth]{fig3.png}
\caption{An example of a manifold with a nontrivial spin structure around a homology cycle; see Coser \textit{et al}.~\cite{coser2016spin}.\label{fig:spin}}
\end{figure}
\subsection{Entanglement Negativity Measurement Protocols}
By combining the operator identities, Eqs.~(\ref{Uneg}), (\ref{mixed}), we obtain our \emph{second main result},
\begin{multline}\label{mainres2}
f(\{N\})=\delta_{(N_\mathrm{avg}^{A_1}-N_\mathrm{avg}^{A_2})\in\mathbb{Z}}\prod_{k=1}^n e^{\frac{2\pi ik}{n} \big(N_{k}^{A_1}-N_{k}^{A_2}\big)} \\
\times\begin{cases} (-1)^{N_\mathrm{avg}^{A_1}-N_\mathrm{avg}^{A_2}}~ & n~\mathrm{even},\\
\phantom{(-}1 & n~\mathrm{odd}.
\end{cases}
\end{multline}
The above analysis implies that measurements of \(\hat{U}_A^\mathrm{neg}\) on the FTed system yield the R\'{e}nyi entanglement negativity.
As an example, in order to measure the 3\textsuperscript{rd} R\'{e}nyi negativity one would have to average over a set of measurements using the standard FT, \(\hat{F}=\hat{F}_{+++}\), with weight \((-\frac{1}{2})\), and a set of measurements using a generalized FT, \(\hat{F}_{++-}\), with weight \((+\frac{3}{2})\); see Eq.~(\ref{neg3}).
\subsection{Spin Structures}\label{sec:spin}
The above calculation of negativities is just a special case of a general fermionic system with nontrivial spin structure.
Similar quantities appear in, e.g., string theory~\cite{SEIBERG1986272} or the classification of interacting topological systems~\cite{Shapourian2017Many,Shiozaki2018Many}. We use the connection of such manifolds to entanglement measures of fermionic systems~\cite{coser2016spin} in order to show that they are not mere theoretical constructs, but could actually be measured with ultracold atoms.
As shown in Ref.~\onlinecite{coser2016spin}, spin structures of many (1+1)-dimensional manifolds are related to polynomials in \(\hat{\rho}_{A},\hat{\rho}_A^+,\hat{P}^{\scriptscriptstyle A_\alpha}_k,\mathrm{~and~}\hat{P}^{\scriptscriptstyle B_\beta}_k\), where \(B_\beta\) are subregions of \(B\) connecting subregions \(A_\alpha\) in copy \(k=1\ldots n\).
Using the results of this paper, one may experimentally simulate spin structure partition functions on arbitrary genus manifolds, as depicted in Fig.~\ref{fig:spin}, by performing additional parity measurements, \(\hat{P}_k^{\scriptscriptstyle B_\beta}=(-1)^{\hat{N}_k^{B_\beta}}\), in region \(B_\beta\) of copy \(k\).
Similar protocols apply to ($d$+1)-dimensional manifolds obtained by cutting and gluing $\mathbb{R}^{d+1}$ at given constant-time hyperplanes.
\section{Example}\label{sec:num}
As an example simulating an experimental implementation of our protocols we study the Klich-Levitov quench model~\cite{Klich2009Quantum,klich2009many,klich2008scaling} of a tight-binding fermionic chain of length $L$ at half-filling which is initially disconnected at the middle at times ${t<0}$ and later connected at times ${t>0}$, see Fig.~\ref{fig:quench}(top). The time-dependent Hamiltonian is given by
\begin{align}
&\hat{H}(t)=\begin{cases}
\hat{H}_0 & t<0,\\
\hat{H}_1 & t>0,
\end{cases}\\
&\hat{H}_0=\sum\limits_{j=1}^{L/2-1}(\hat{a}^\dagger_j \hat{a}^{\phantom{|}}_{j+1}+\mathrm{h.c.})+\sum\limits_{j=L/2+1}^{L-1}(\hat{a}^\dagger_j \hat{a}^{\phantom{|}}_{j+1}+\mathrm{h.c.}), \nonumber\\
&\hat{H}_1=\sum\limits_{j=1}^{L-1}(\hat{a}^\dagger_j \hat{a}^{\phantom{|}}_{j+1}+\mathrm{h.c.}). \nonumber
\end{align}
The time-dependent density matrix at times ${t>0}$ both before and after the (generalized) Fourier transform are thus given by
\begin{eqnarray}
\hat{\rho}(t)&=&\lim\limits_{\beta\to\infty}e^{-i\hat{H}_1t}\frac{e^{-\beta\hat{H}_0}}{\mathrm{Tr}\, e^{-\beta\hat{H}_0}}e^{i\hat{H}_1t},\\
\widetilde{\hat{\rho}^{\otimes n}}(t)&=&\hat{F}_{\{\sigma\}}^{\phantom{|}}\hat{\rho}(t)^{\otimes n}\hat{F}^\dagger_{\{\sigma\}}.
\end{eqnarray}
In order to simulate our measurement protocols, we must determine the probability $P[f(\{N\})]$ of a result for $f(\{N\})$ in our measurement protocols, Eqs.~(\ref{mainres}), (\ref{mainres2}).
\begin{align}
& P\big[f(\{N\})=e^{\frac{2\pi i m}{n}}\big]=\mathrm{Tr}\Big\{\delta_{f(\{\hat{N}\}),\exp(\frac{2\pi i m}{n})}\widetilde{\hat{\rho}^{\otimes n}}\Big\} \nonumber\\
&={\sum\limits_{q,q'=1}^n} e^{-\frac{2\pi i m q}{n}}\,\mathrm{Tr}\Big\{{\prod\limits_{k=1}^{n}} e^{\frac{2\pi i}{n} \big(\hat{N}_{k}^{A_1}\pm \hat{N}_{k}^{A_2}\big)[[k(+\frac{1}{2})]q+q']}\widetilde{\hat{\rho}^{\otimes n}}\Big\}.
\end{align}
Here, the ``$\pm$" distinguishes between the negativity and the entropy, and the ``$(+\frac{1}{2})$" applies only for even $n$.
The evaluation of this trace of exponentials is done using the methods of Refs.~\onlinecite{Klich2002FCS,Klich2014FCS,GoldsteinSela2018} for the entanglement entropy, and using the methods of Refs.~\onlinecite{eisler2015partial,Eisler2016Entanglement,cornfeld2018imbalance} for the entanglement negativity.
We simulate \(A\cup B\) of 32 sites, with ${|A|=16}$ for the entropy and ${|A_{1,2}|=2}$ for the negativity.
The data presented in Fig.~\ref{fig:quench}(bottom) is attained by randomly sampling $f(\{N\})$ from $P[f(\{N\})]$ and averaging over the instances.
We sample 600 instances for the R\'{e}nyi entanglement entropy as well as for the R\'{e}nyi entanglement negativity.
The error bars display the statistical standard error of the mean (SEM).
\section{Conclusions}\label{sec:conc}
We described, for the first time, experimental protocols for direct detection of various entanglement measures in fermionic systems beyond the 2\textsuperscript{nd} R\'{e}nyi entropy.
We presented a complete paradigm that enables the evaluation of all R\'{e}nyi entropies and negativities for both pure and mixed states.
Possible applications include: testing the scaling laws of entanglement in many-body quantum states, determining the coherence of quantum simulators, and incorporating entanglement detections in quantum computations. Our protocols also allows the use of cold atoms to quantum simulate the partition function of fermions on certain manifolds with nontrivial spin structures.
All our protocols are based on performing Fourier transforms on the quantum system, which may be realized by applying sequences of beam splitters, and are readily attainable using existing technologies.
\section*{Acknowledgements}
E.S. was supported in part by the Israel Science Foundation (Grant No. 1243/13) and by the US-Israel Binational Science Foundation (Grant No. 2016255). M.G. was supported by the Israel Science Foundation (Grant No. 227/15), the German Israeli Foundation (Grant No. I-1259-303.10), the US-Israel Binational Science Foundation (Grant No. 2016224), and the Israel Ministry of Science and Technology (Contract No. 3-12419).
|
\section{Introduction}
Supergravity solutions of the form $AdS_d\times S^n$ have long been known, and have particular importance in the context of AdS/CFT. Perhaps the most familiar cases are those associated with the near-horizon geometry of M2, D3 and M5 branes. In contrast, the case of $AdS_6/CFT_5$ has been less well explored, in part because it does not admit a single brane interpretation. Nevertheless, there has been a recent resurgence of interest in this case, driven both on the field theory and the gravity sides of the duality. In contrast with the more familiar cases, holography in this dimension is interesting as there are no maximally supersymmetric 5D SCFTs (with 16 Poincar\'e supercharges and 16 superconformal supercharges). In particular, the maximal possible amount of supersymmetry in 5D is 8+8 supercharges \cite{Nahm:1977tg}. The superconformal algebra in 5D is then based on the Lie superalgebra $F(4)$ with maximal bosonic subalgebra $SO(2,5)\times SO(3)$ \cite{Minwalla:1997ka}. The natural six-dimensional dual is then $F(4)$ gauged supergravity which is a non-chiral theory with 16 real supercharges \cite{Romans:1985tw}.
Of course, a string theory realization of $AdS_6/CFT_5$ starts not in six dimensions, but rather with 10 or 11 dimensional supergravity (in its low-energy limit) compactified to $AdS_6$ times some internal manifold. Such backgrounds can preserve at most half of the total supersymmetries \cite{Figueroa-OFarrill:2002ecq}, in accordance with the $CFT_5$ picture. Until recently, string theory realizations of such $AdS_6$ duals have been hard to come by, with the prime example being a construction of stacks of D4-branes, D8-branes, and O8-planes in type IIA string theory \cite{Seiberg:1996bd,Brandhuber:1999np,Bergman:2012kr,Passias:2012vp}. The presence of D8-branes means that the supergravity background is obtained in the massive IIA theory.
In type IIB, $(p,q)$ five-brane webs \cite{Aharony:1997ju,Aharony:1997bh} with D7-branes added \cite{DeWolfe:1999hj} can realize large classes of 5D SCFTs. The holographic dual of these five-brane webs corresponds to supersymmetric $AdS_6$ solutions of IIB supergravity. The Killing spinor equations for this system were investigated in \cite{Apruzzi:2014qva,Kim:2015hya} and reduced to a pair of coupled PDEs, and a complete family of local solutions was constructed in \cite{DHoker:2016ujz,DHoker:2017mds,DHoker:2017zwj}. The local $AdS_6$ solutions were found by studying the IIB Killing spinor equations, which led to a family of solutions of the form $AdS_6\times S^2$ warped over a Riemann surface $\Sigma$; remarkably, these solutions are completely determined by a pair of holomorphic functions $\mathcal{A}_{\pm}$ on $\Sigma$ which can be chosen freely up to global regularity conditions.
More generally, the $AdS_6$ solutions of \cite{DHoker:2016ujz} ought to be viewed as vacuum solutions of six-dimensional $F(4)$ gauged supergravity obtained by reducing ten-dimensional IIB supergravity on $S^2\times\Sigma$. Curiously, the existence of a whole family of solutions suggests that the lifting of $F(4)$ gauged supergravity to ten dimensions is far from unique. This is in contrast to the 32 supercharge cases such as gauged $\mathcal N=8$ supergravity in five-dimensions, which has a unique lift to IIB supergravity on $S^5$. Before this family of IIB vacua was discovered, a full non-linear Kaluza-Klein reduction of massive IIA supergravity to $F(4)$ gauged supergravity on (one hemisphere of) $S^4$ was obtained in \cite{Cvetic:1999un}. This was subsequently dualized to a IIB reduction in \cite{Jeong:2013jfc} using non-abelian T-duality. In particular, the IIA reduction ansatz of \cite{Cvetic:1999un} was based on $AdS_6$ times a squashed $S^4$ foliated by 3-spheres, with the $SU(2)$ $R$-symmetry corresponding to the gauging of $SU(2)_L$ inside the $SO(4)\simeq SU(2)_L\times SU(2)_R$ isometry of $S^3$. Non-abelian T-dualizing then gives rise to a IIB reduction where the $S^4$ is replaced by $S^2\times\Sigma$, with the $R$-symmetry now corresponding to the $SO(3)$ isometry of $S^2$. However, this procedure gives rise to only a single background, first obtained in \cite{Lozano:2012au}, and not the entire family of solutions constructed in \cite{DHoker:2016ujz}.
In this paper, we present a complete family of consistent truncations of IIB supergravity to $F(4)$ gauged supergravity parametrized by the same holomorphic functions $\mathcal{A}_\pm$ of \cite{DHoker:2016ujz}. This generalizes the family of $AdS_6$ vacua into complete reduction ans\"atze, in agreement with the conjecture that any supersymmetric vacuum solution can be extended to a consistent truncation incorporating the full lower-dimensional supergravity multiplet \cite{Duff:1985jd,Gauntlett:2007ma}. (General half-maximal consistent truncations were recently considered in the framework of exceptional field theory in \cite{Malek:2017njj}.) The non-linear reduction to $F(4)$ gauged supergravity was inspired by the construction of \cite{Jeong:2013jfc}, but is in fact much more general. Naturally, it incorporates the reduction ansatz of \cite{Jeong:2013jfc} as a special case, and we give the explicit functions $\mathcal{A}_{\pm}$ necessary to recover the results of \cite{Jeong:2013jfc} in section \ref{sec:discussion}.
The rest of the paper is structured as follows. In section \ref{sec:10Dsetup}, we review the basics of 10D type IIB supergravity, and briefly discuss the 1/2 BPS $AdS_6\times S^2\times \Sigma$ solutions of \cite{DHoker:2016ujz} that depend on the pair of holomorphic functions $\mathcal{A}_{\pm}$ on $\Sigma$. Then, in section \ref{sec:6Dsugra+reduction}, we discuss Romans' 6D $F(4)$ gauged supergravity and give our full non-linear reduction ansatz, which also depends on the holomorphic functions $\mathcal{A}_{\pm}$. Finally, in section \ref{sec:discussion}, we discuss how our ansatz reduces to previous results in the literature in special cases, and the connection to $AdS_6/CFT_5$ holography. The appendices discuss our conventions for indices and coordinates on the various manifolds we consider (appendix \ref{app:ourconventions}), the relation between different conventions for IIB supergravity (appendix \ref{app:10Dconventions}), and more details of checking the equations of motion on our ansatz (appendix \ref{app:EOMchecks}).
\paragraph{Main result:} Our main result is the full non-linear Kaluza-Klein reduction ansatz given in section \ref{sec:genansatz} in (\ref{general:ansatz:1})-(\ref{general:ansatz:G5}). On this ansatz, the 10D type IIB supergravity equations of motion (\ref{2B:eoms}) are equivalent to the 6D $F(4)$ gauged supergravity equations of motion (\ref{F(4):eoms}). The reduction ansatz depends on a pair of unrestricted holomorphic functions $\mathcal{A}_{\pm}$ on the Riemann surface $\Sigma$.
\paragraph{Note added:} After this work appeared on the arXiv, the work \cite{Malek:2018zcz} was released, which also constructs the consistent truncation to $F(4)$ gauged supergravity.
\section{IIB supergravity and warped \texorpdfstring{$AdS_6\times S^2\times\Sigma$}{AdS6 x S2 x Sigma} vacua} \label{sec:10Dsetup}
Before presenting the non-linear Kaluza-Klein ansatz, we quickly review the equations of motion of IIB supergravity and the family of supersymmetric $AdS_6$ solutions of \cite{DHoker:2016ujz}. This sets up our conventions and lays the groundwork for the reduction.
\subsection{Type IIB supergravity}
Type IIB supergravity is a chiral theory with 32 real supercharges. Because of the self-dual five-form, it does not admit a covariant Lagrangian formulation (although one can come close). Of course, for presenting the consistent truncation, we only need the equations of motion, which we give in the $SU(1,1)$ formulation \cite{Schwarz:1983qr,Howe:1983sra}.
The bosonic fields consist of the Einstein frame metric $g_{MN}$, a complex scalar with kinetic term $P_1$ and composite connection $Q_1$, a complex three-form field strength $G_3$ and a self-dual five-form $F_5=*F_5$. From a stringy point of view, the one-forms $P$ and $Q$ are related to the dilaton and RR axion, while $G_3$ contains both NSNS and RR (real) three-forms. The exact map is spelled out in appendix \ref{app:10Dconventions}.
The IIB supergravity fields satisfy the Bianchi identities:
\begin{subequations}
\begin{align}
dP&=2iQ\wedge P,\\
dQ&=-iP\wedge\overline{P},\\
dG_3&=iQ\wedge G_3-P\wedge\overline{G}_3.
\end{align}
\ese
These are automatically satisfied if we introduce a complex scalar $B$ and a complex two form $C_2$ as:
\begin{subequations}
\begin{align}
\label{eq:10D:PforB} P&=(1-|B|^2)^{-1}dB,\\
\label{eq:10D:QforB} Q&=(1-|B|^2)^{-1}\Im[Bd\overline{B}],\\
\label{eq:10D:G3forC2} G_3&=(1-|B|^2)^{-\frac{1}{2}}(dC_2-Bd\overline{C}_2).
\end{align}
\label{eq:10D:PG3def}\ese
The equations of motion are:
\begin{subequations}
\begin{align}
(d-2iQ)\wedge*P&=-\frac{1}{4}G_3\wedge* G_3,\label{2B:eom:P}\\
(d-iQ)\wedge*G_3&=P\wedge*\overline{G}_3-4iG_3\wedge* F_5,\label{2B:eom:G3}\\
d*F_5&=\fft{i}8G_3\wedge\overline{G}_3,\label{2B:eom:F5}
\end{align}
\label{2B:eoms}
\ese
along with the Einstein equation
\begin{align}
R_{IJ}&=P_I\overline{P}_J+\overline{P}_I P_J+\frac{1}{6}F_{IKLMN}F\ind{_J^{KLMN}}\nn\\
&\quad+\frac{1}{8}\left(G_{IKL}\overline{G}\ind{_J^{KL}}+\overline{G}_{IKL}G\ind{_J^{KL}}\right)-\frac{1}{48}g_{IJ}G_{KLM}\overline{G}^{KLM}.\label{2B:eom:Einstein}
\end{align}
Note that the five-form equation can equally well be viewed as a Bianchi identity, $dF_5=\fft{i}8G_3\wedge\overline{G}_3$, which may be solved by taking $F_5=dC_4+\fft{i}8\Im[C_2\wedge d\overline{C}_2]$. However, as in other IIB reductions, it is more convenient to make the ansatz on the field strength, and not the potential, in which case the equation of motion (\ref{2B:eom:F5}) will need to be verified.
\subsection{A family of warped \texorpdfstring{$AdS_6$}{AdS6} solutions}\label{sec:SUSYAdS6sols}
A large family of 1/2 BPS solutions to IIB supergravity of the form $AdS_6\times S^2$ warped over a 2D Riemann surface $\Sigma$ (with complex coordinates $z,\bar{z}$) were found by a detailed analysis of the IIB supersymmetry equations in \cite{DHoker:2016ujz}. These solutions are (locally) completely determined by a pair of holomorphic functions $\mathcal{A}_{\pm}(z)$ on the Riemann surface. Remarkably, there is no restriction on the pair of holomorphic functions in order to obtain a local supersymmetric solution.\footnote{A different but presumably equivalent classification of these $AdS_6$ solutions was found in \cite{Apruzzi:2018cvq}. These solutions were given in terms of solutions to the spherically symmetric cylindrical Laplace equation instead of holomorphic functions.} To obtain a well-behaved global solution, considerably more care is needed: $\mathcal{A}_{\pm}$ are allowed to be multi-valued as holomorphic sections of a holomorphic bundle over $\Sigma$ with structure group contained in $SU(1,1)\times \mathbb{C}$; the boundaries and other global properties of $\Sigma$ are important in constructing such solutions \cite{DHoker:2016ujz,DHoker:2017mds,DHoker:2017zwj,DHoker:2016ysh}. There has been much interest in these global solutions and interpreting them as near-horizon limits of $(p,q)$ 5-brane webs with additional 7-branes in the web \cite{DHoker:2017mds,DHoker:2017zwj,DHoker:2016ysh,Gutperle:2018vdd,Bergman:2018hin}. (See also section \ref{sec:discussionholo} for more discussion regarding the holographic duals of these solutions.)
Here we briefly summarize the solutions of \cite{DHoker:2016ujz}. The starting point is a pair of holomorphic functions $\mathcal{A}_\pm(z)$, from which we may define a holomorphic function $\mathcal{B}$ whose derivative is given by (using obvious notation $\partial=\partial_z)$:
\begin{equation} \label{eq:Bdef} \partial_z \mathcal{B} = \mathcal{A}_+\partial\mathcal{A}_- - \mathcal{A}_- \partial \mathcal{A}_+.\end{equation}
Of the undetermined integration constant in $\mathcal{B}$, only the real part is relevant. From $\mathcal{A}_\pm$ and $\mathcal{B}$, we define
\begin{equation}
\kappa_\pm=\partial\mathcal A_\pm,\qquad\kappa^2=-|\kappa_+|^2+|\kappa_-|^2,
\end{equation}
along with
\begin{equation}
\mathcal G=|\mathcal A_+|^2-|\mathcal A_-|^2+\mathcal B+\bar{\mathcal B}.
\end{equation}
Note that $\kappa^2 = -\partial\bar{\partial}G$. We also find it useful to introduce
\begin{equation}
\mathcal{Y}= \fft{\kappa^2\mathcal G}{|\partial\mathcal G|^2},\qquad
\hat{\mathcal C}=\fft{\kappa_+\bar\partial\mathcal G+\bar\kappa_-\partial\mathcal G}{\kappa^2}.
\label{eq:YCdef}
\end{equation}
The supersymmetric $AdS_6$ vacua of \cite{DHoker:2016ujz} are then given by the metric
\begin{equation}
ds^2 = f_6^2 ds^2_{AdS_6} +f_2^2 ds^2_{S^2} + 4\rho^2 dz d\bar{z},
\label{eq:susysolmet}
\end{equation}
with metric functions
\begin{equation}
f_6^2=\fft{c_6^2}{\rho^2}\kappa^2\sqrt{\tilde{\mathcal D}},\qquad
f_2^2=\fft{c_6^2}{9\rho^2}\fft{\kappa^2}{\sqrt{\tilde{\mathcal D}}},\qquad
\rho^4=\fft{c_6^2}6\fft{\kappa^4\sqrt{\tilde{\mathcal D}}}{\mathcal G},
\label{eq:metfuncs}
\end{equation}
and matter fields
\begin{subequations}
\begin{align}
B &= \fft{(\mathcal A_+-\bar{\mathcal A}_-)-\hat{\mathcal C}/\sqrt{\tilde{\mathcal D}}}{(\bar{\mathcal A}_+-\mathcal A_-)+\bar{\hat{\mathcal C}}/\sqrt{\tilde{\mathcal D}}}, \label{eq:Bvac}\\
C_2 &= \frac{2ic_6}{9}\left[\fft{\hat{\mathcal C}}{\tilde{\mathcal D}}-3(\bar{\mathcal A}_-+\mathcal A_+)\right]\mathrm{vol}(S^2),\label{eq:Cvac}\\
F_5 &= 0,\label{eq:Fvac}
\end{align}%
\label{eq:susysolmat}%
\end{subequations}%
Here $c_6$ is a non-vanishing constant, and we have furthermore introduced the recurring factor
\begin{equation}
\tilde{\mathcal D}=1+\fft{2}{3\mathcal Y},
\label{eq:calD}
\end{equation}
which will play an important role in the generalization to the full KK reduction below.
Finally, note that there are in fact two branches of solutions, the first with
\begin{equation} \label{eq:conditionskappaG} \kappa^2\geq 0, \qquad \mathcal{G}\geq 0,\end{equation}
and the second with both quantities non-positive. (These two branches of solutions are mapped into each other by complex conjugation \cite{DHoker:2016ujz}.) Of course, the metric functions $f_2^2$, $f_6^2$ and $\rho^2$ must all be positive. This will be the case provided we take the positive square-root ($\sqrt{\tilde{\mathcal D}}>0$) on the first branch and the negative square-root ($\sqrt{\tilde{\mathcal D}}<0$) on the second branch.
\section{6D Supergravity and the reduction ansatz}\label{sec:6Dsugra+reduction}
The existence of a family of $AdS_6$ solutions suggests the possibility of a complete non-linear Kaluza-Klein reduction of IIB supergravity to six-dimensional $F(4)$ gauged supergravity. Here we first present a brief overview of the six-dimensional theory, and then turn to the full reduction ansatz generalizing the vacuum solution discussed above in section \ref{sec:SUSYAdS6sols}.
\subsection{\texorpdfstring{$F(4)$}{F(4)} gauged supergravity}\label{sec:6Dsugra}
Romans' six-dimensional $F(4)$ gauged supergravity \cite{Romans:1985tw} is a non-chiral theory with 16 real supercharges. The bosonic fields consist of a metric $\tilde g_{\mu\nu}$, a real scalar $\tilde{\phi}$, an Abelian two-form $\tilde F_2$ and three-form $\tilde F_3$, and three $SU(2)$ gauge two-forms $\tilde F^i$. The bosonic Lagrangian may be written in a form notation \cite{Romans:1985tw,Cvetic:1999un} as
\begin{align}
\mathcal L&=R*_6\mathbbm1-4\fft{*_6dX\wedge dX}{X^2}-\tilde g^2\left(\fft29X^{-6}-\fft83X^{-2}-2X^2\right)*_61\nn\\
&\qquad-\fft12X^4*_6\tilde F_3\wedge\tilde F_3-\fft12X^{-2}\left(*_6\tilde F_2\wedge\tilde F_2+\fft1{\tilde g^2}*_6\tilde F^i\wedge\tilde F^i\right)\nn\\
&\qquad-\tilde A_2\wedge\left(\fft12d\tilde A_1\wedge d\tilde A_1+\fft13\tilde g\tilde A_2\wedge d\tilde A_1+\fft2{27}\tilde g^2\tilde A_2\wedge\tilde A_2+\fft1{2\tilde g^2}\tilde F^i\wedge\tilde F^i\right),
\label{eq:F4lag}
\end{align}
where $X=e^{-\frac{1}{2\sqrt{2}}\tilde\phi}$, and the field strengths are given in terms of the gauge potentials by
\begin{subequations}
\begin{align}
\tilde F_3&=d\tilde A_2,\\
\tilde F_2&=d\tilde A_1+\frac{2}{3}\tilde g\tilde A_2,\\
\tilde F^i&=d\tilde A^i+\frac{1}{2}\epsilon_{ijk}\tilde A^j\wedge\tilde A^k.
\end{align}
\ese
The equations of motion corresponding to this Lagrangian are given by
\begin{subequations}
\begin{align}
d(X^4*_6\tilde F_3)&=-\frac{1}{2}\tilde F_2\wedge\tilde F_2-\frac{1}{2\tilde g^2}\tilde F^i\wedge\tilde F^i-\frac{2}{3}\tilde gX^{-2}*_6\tilde F_2,\label{F(4):eom:1}\\
d(X^{-2}*_6\tilde F_2)&=-\tilde F_2\wedge\tilde F_3,\label{F(4):eom:2}\\
D(X^{-2}*_6\tilde F^i)&=-\tilde F_3\wedge\tilde F^i,\label{F(4):eom:3}\\
d(X^{-1}*_6dX)&=\frac{1}{8}X^{-2}\left(*_6\tilde F_2\wedge\tilde F_2+\frac{1}{\tilde g^2}*_6\tilde F^i\wedge\tilde F^i\right)-\frac{1}{4}X^4*_6\tilde F_3\wedge\tilde F_3\nonumber\\&\quad+\tilde g^2\left(\frac{1}{6}X^{-6}-\frac{2}{3}X^{-2}+\frac{1}{2}X^2\right)*_6\mathbbm{1},\label{F(4):eom:4}
\end{align}
\label{F(4):eoms}\ese
along with the Einstein equation
\begin{align}
\tilde R_{\mu\nu}&=4X^{-2}\partial_\mu X\partial_\nu X+\tilde g^2\left(\frac{1}{18}X^{-6}-\frac{2}{3}X^{-2}-\frac{1}{2}X^2\right)g_{\mu\nu}+\frac{1}{4}X^4\left((\tilde F_3)^2_{\mu\nu}-\frac{1}{6}g_{\mu\nu}(\tilde F_3)^2\right)\nonumber\\&\quad+\frac{1}{2}X^{-2}\left((\tilde F_2)^2_{\mu\nu}-\frac{1}{8}g_{\mu\nu}(\tilde F_2)^2\right)+\frac{1}{2\tilde g^2}X^{-2}\left((\tilde F^i)^2_{\mu\nu}-\frac{1}{8}g_{\mu\nu}(\tilde F^i)^2\right).\label{F(4):eom:5}
\end{align}
In the above, the $SU(2)$ gauge covariant derivative $D$ is defined by
\begin{equation}
D\tilde F^i=d\tilde F^i+\epsilon_{ijk}\tilde A^j\wedge\tilde F^k.
\end{equation}
This $F(4)$ gauged supergravity theory admits a supersymmetric $AdS_6$ vacuum with $\tilde F_2=\tilde F_3=\tilde F^i = 0$ and $X=1$ that preserves all 16 of the supersymmetries \cite{Romans:1985tw}. There is also a non-supersymmetric $AdS_6$ vacuum with $X = 3^{-1/4}$. Note that to get a unit radius $AdS_6$, we must choose:
\begin{equation} \tilde{g} = \frac{3}{\sqrt{2}}.
\label{eq:tildeg}
\end{equation}
We will make this choice from here on.
\subsection{Generalized reduction ansatz}\label{sec:genansatz}
The $F(4)$ gauged supergravity theory was obtained from a warped $S^4$ reduction of massive IIA supergravity in \cite{Cvetic:1999un}, and more recently from a reduction of IIB supergravity over an $S^2$ and a Riemann surface in \cite{Jeong:2013jfc}. In both cases, the supersymmetric $AdS_6$ vacua uplifted to 10D are 1/2 BPS solutions of either IIA or IIB supergravity.
We now present our main result, which is a consistent truncation of type IIB supergravity to 6D $F(4)$ gauged supergravity generalizing the reduction of \cite{Cvetic:1999un,Jeong:2013jfc}. We start with the metric
\begin{equation}
ds^2=f_6^2ds_6^2+f_2^2ds^2_{\tilde S^2} +4\rho^2dzd\bar{z},\label{general:ansatz:1}
\end{equation}
where
\begin{equation}
f_6^2=\fft{c_6^2}{\rho^2}\kappa^2\sqrt{\mathcal D},\qquad
f_2^2=\fft{c_6^2}{9\rho^2}\fft{\kappa^2X^2}{\sqrt{\mathcal D}},\qquad
\rho^4=\fft{c_6^2}{6}\fft{\kappa^4X^2\sqrt{\mathcal D}}{\mathcal G}.
\label{eq:fXs}
\end{equation}
The matter fields are given by
\begin{subequations}
\begin{align}
B&=\fft{(\mathcal A_+-\bar{\mathcal A}_-)-\hat{\mathcal C}X^2/\sqrt{\mathcal D}}{(\bar{\mathcal A}_+-\mathcal A_-)+\bar{\hat{\mathcal C}}X^2/\sqrt{\mathcal D}},\label{general:ansatz:2}
\\
C_2&=\frac{2ic_6}{9}\left[\fft{\hat{\mathcal C}}{\mathcal D}-3(\bar{\mathcal A}_-+\mathcal A_+)\right]\mathrm{vol}(\tilde S^2)-\sqrt2c_6(\bar{\mathcal A}_--\mathcal A_+)\tilde F_2+\fft{2ic_6}3(\bar{\mathcal A}_-+\mathcal A_+)\mu^i\tilde F^i,\label{general:ansatz:3}\\
F_5&=G_5+*G_5,\label{general:ansatz:4}
\end{align}
\end{subequations}
with
\begin{equation} \label{general:ansatz:G5} G_5=c_6^2\left[\fft{i\kappa^2X^4}{2}(*_6\tilde F_3)\wedge dz\wedge d\bar z+\fft{1}{2\sqrt{2}X^2}(*_6\tilde F_2)\wedge*_2d\mathcal G-\fft{1}{6X^2}(*_6\tilde F^i)\wedge D(\mu^i\mathcal G)\right].\end{equation}
It is also useful to give the expression for $*G_5$:
\begin{align}
*G_5&=c_6^2\Bigg[\bigg(\fft{\mathcal GX^4}{6\mathcal D}\tilde F_3+\fft{\sqrt{2}}{36\mathcal D}\tilde F_2\wedge d\mathcal G+\fft{1}{54\mathcal D}\mu^i\tilde F^i\wedge *_2d\mathcal G\bigg)\wedge\mathrm{vol}(\tilde S^2)\nn\\&\kern3em+\fft{i\kappa^2}{18}\tilde F^i\wedge*_2D\mu^i\wedge dz\wedge d\bar{z}\Bigg].
\end{align}
Here $\mathcal{D}$ is defined as
\begin{equation}
\mathcal D=X^4+\fft{2}{3\mathcal Y},
\label{eq:newcalD}
\end{equation}
which generalizes (\ref{eq:calD}) in the presence of a non-trivial scalar. Note that $ds^2_{\tilde S^2}$ and $\mathrm{vol}(\tilde S^2)$ have been used instead of $ds^2_{S^2}$ and $\mathrm{vol}(S^2)$ since the $SU(2)$ isometry of the $S^2$ is gauged by the $\tilde A^i$ fields. (See appendix \ref{app:ourconventions} for more information.) The remaining definitions, (\ref{eq:Bdef}) through (\ref{eq:YCdef}), are unchanged.
Our reduction ansatz thus provides a consistent truncation of IIB supergravity on an $S^2$ and warped over a Riemann surface $\Sigma$ to 6D $F(4)$ gauged supergravity. In fact, our ansatz contains a family of such truncations --- one for each pair of holomorphic functions $\mathcal{A}_{\pm}$ on the Riemann surface. This fully generalizes and contains the supersymmetric $AdS_6$ solutions of \cite{DHoker:2016ujz} discussed in section \ref{sec:SUSYAdS6sols}. In particular, (\ref{eq:fXs}) and (\ref{general:ansatz:2}) generalize the corresponding expressions (\ref{eq:metfuncs}) and (\ref{eq:Bvac}) to incorporate the scalar $X$, and (\ref{general:ansatz:3}) and (\ref{general:ansatz:4}) generalize (\ref{eq:Cvac}) and (\ref{eq:Fvac}) to incorporate the six-dimensional gauge fields. When $X=1$ (i.e.\ the scalar $\tilde{\phi}=0$) and the gauge fields are turned off, $\tilde F_2 = \tilde F_3 = \tilde F^i = 0$, this reduction ansatz reduces as it must to that of \cite{DHoker:2016ujz}.
We have explicitly checked that our ansatz (\ref{general:ansatz:1})-(\ref{general:ansatz:G5}) satisfies the 10D IIB equations of motion (\ref{2B:eoms}) if and only if the 6D fields $\tilde g_{\mu\nu}$, $\tilde{F}_2$, $\tilde{F}_3$, $\tilde{F}^i$ and $X$ satisfy the 6D equations of motion (\ref{F(4):eoms}) and (\ref{F(4):eom:5}). In particular, the IIB axi-dilaton equation of motion (\ref{2B:eom:P}) is equivalent to the 6D scalar equation of motion (\ref{F(4):eom:4}), the IIB three-form equation (\ref{2B:eom:G3}) is equivalent to (\ref{F(4):eom:1}), (\ref{F(4):eom:2}), (\ref{F(4):eom:3}), and (\ref{F(4):eom:4}), and the self-dual five-form equation (\ref{2B:eom:F5}) is equivalent to (\ref{F(4):eom:1}), (\ref{F(4):eom:2}), and (\ref{F(4):eom:3}). Finally, the IIB Einstein equation, (\ref{2B:eom:Einstein}), is satisfied if and only if (\ref{F(4):eom:3}), (\ref{F(4):eom:4}) and the 6D Einstein equation (\ref{F(4):eom:5}) are satisfied. For more details of these calculations, see appendix \ref{app:EOMchecks}.
Our ansatz thus gives a family of consistent truncations of 10D IIB supergravity to 6D $F(4)$ gauged supergravity, in the sense that any solution of the 6D theory can be uplifted to a family of 10D solutions using (\ref{general:ansatz:1})-(\ref{general:ansatz:G5}). We stress once more that this ansatz is consistent for \emph{any} pair of holomorphic functions $\mathcal{A}_{\pm}$, at least up to regularity conditions of the $AdS_6$ vacuum.
\section{Discussion}\label{sec:discussion}
The reduction ansatz (\ref{general:ansatz:1})-(\ref{general:ansatz:G5}) generalizes the IIB reduction of \cite{Jeong:2013jfc}, corresponding to the $AdS_6$ background of \cite{Lozano:2012au}, to encompass the complete family of $AdS_6$ backgrounds of \cite{DHoker:2016ujz}. In fact, our method for finding this generalized ansatz was to translate the specific reduction of \cite{Jeong:2013jfc} into the language of the holomorphic functions $\mathcal{A}_\pm$ of \cite{DHoker:2016ujz}. For example, starting with the string frame metric embedding of \cite{Jeong:2013jfc}
\begin{equation}
d\hat s^2=X^{-\frac{1}{2}}s^{-\frac{1}{3}}\Delta^{\frac{1}{2}}\left[ds_6^2+2\tilde g^{-2}X^2d\xi^2\right]+e^{-2A}dr^2+\frac{r^2e^{2A}}{r^2+e^{4A}}ds^2_{\tilde S^2},
\label{particular:ansatz:met}
\end{equation}
with
\begin{equation}
e^A=\frac{1}{\sqrt{2}\tilde g}s^{-\frac{1}{6}}c\,X^{-\fft34}\Delta^{-\fft14}, \qquad \Delta = X c^2 + X^{-3}s^2, \qquad s = \sin \xi, \quad c = \cos\xi,
\end{equation}
we deduce the coordinate transformation on the Riemann surface $\Sigma$ mapping $\{\xi,r\}\to\{z,\bar z\}$
\begin{equation}
z=z_1+iz_2, \qquad z_1 = \frac23 \tilde{g}^2 r,\qquad z_2 = \sin^{2/3}\xi.
\label{eq:particular:ansatz:coords}
\end{equation}
This allows us to rewrite (\ref{particular:ansatz:met}) as
\begin{equation}
d\hat s^2=X^{-\frac{1}{2}}s^{-\frac{1}{3}}\Delta^{\frac{1}{2}}ds_6^2+\frac{r^2e^{2A}}{r^2+e^{4A}}ds^2_{\tilde S^2}+\fft9{4\tilde g^4}e^{-2A}dz\,d\bar z.
\end{equation}
Converting to the Einstein frame and comparing with the $AdS_6$ metric (\ref{general:ansatz:1}) then yields the $X$-dependent metric functions (\ref{eq:fXs}), provided we identify
\begin{equation}
\kappa^2=\fft{z_1z_2}{36c_6^2},\qquad\mathcal{G}=\fft{(1-z_2^3)z_1}{54c_6^2},\qquad|\partial\mathcal{G}|^2=\fft{z_1^2z_2^4+\fft19(1-z_2^3)^2}{(36c_6^2)^2},
\label{eq:kapcalG}
\end{equation}
where we have additionally set $\tilde g=3/\sqrt2$ as in (\ref{eq:tildeg}).
Working backwards from (\ref{eq:kapcalG}), it is not too difficult to deduce the form of the holomorphic functions $\mathcal{A}_\pm$
\begin{align}\label{eq:particular:ansatz:calA}
\mathcal A_\pm(z)&=\fft1{c_6}\left(\frac{1}{216}z^3\mp\frac{i}{4}z-\frac{i}{108}\right),
\end{align}
along with the auxiliary function
\begin{align}
\mathcal B(z)&=\fft1{c_6^2}\left(-\frac{i}{864}z^4+\frac{1}{216}z\right).
\end{align}
Given these functions, it is then possible to work out (\ref{general:ansatz:2}) as the appropriate generalization of the IIB axi-dilaton reduction (\ref{eq:Bvac}) in the presence of a non-trivial scalar $X$. The remaining expressions are then for the complex three-form $G_3$ and self-dual five-form $F_5$. These take a bit more effort, but can be obtained by translating the particular ansatz of \cite{Jeong:2013jfc} into the $SU(1,1)$ Einstein frame and reexpressing the result in terms of the holomorphic functions $\mathcal{A}_{\pm}$ and their derived quantities such as $\mathcal{G}$ and $\kappa_{\pm}$. Finally, as it was not guaranteed that this construction would be a consistent truncation, it was essential to check the IIB equations of motion and verify that they were equivalent to the $F(4)$ gauged supergravity equations of motion --- and in particular, that they did not impose any extra conditions on the functions $\mathcal{A}_{\pm}$. Additional details of these checks are presented in appendix~\ref{app:EOMchecks}.
\subsection{Holography}\label{sec:discussionholo}
Finally, we conclude with a few remarks about $AdS_6/CFT_5$ holography. From a 10D point of view, we may use the supersymmetric $AdS_6$ solutions of \cite{DHoker:2016ujz,DHoker:2017mds,DHoker:2017zwj} discussed in section \ref{sec:SUSYAdS6sols}, see e.g. \cite{DHoker:2016ysh, Gutperle:2017tjo, Gutperle:2018wuk,Gutperle:2018vdd, Bergman:2018hin, Fluder:2018chf}. These efforts have been limited to the supersymmetric $AdS_6$ vacua as the extension of these solutions to include non-BPS excitations were not known until now.
Alternatively, one can approach 6D/5D holography in supergravity from the 6D perspective; efforts using the 6D $F(4)$ supergravity discussed in section \ref{sec:6Dsugra} include \cite{Ferrara:1998gv,DAuria:2000afl,Karndumri:2012vh,Karndumri:2014lba,Alday:2014rxa,Alday:2014bta,Alday:2014fsa,Hama:2014iea,Alday:2015jsa,Gutperle:2017nwo,Chang:2017mxc,Gutperle:2018axv}. The 6D supergravity is then the effective theory that describes (a consistent truncation of the set of) excitations around the $AdS_6$ vacuum. However, without an understanding of the uplift of these 6D solutions to 10D, a microscopic understanding of the CFT described by the $AdS_6$ vacuum and its excitations was lacking.
Our reduction ansatz (\ref{general:ansatz:1})-(\ref{general:ansatz:G5}) provides a key link between the 10D and 6D approaches above. In the 10D approach, it provides a way to include (non-BPS) excitations to the family of $AdS_6$ vacua. From the 6D perspective, it gives a way to understand the microscopic dual 5D CFT through the brane picture obtained by the 10D uplift of the 6D solutions. For example, the fact that our uplift is independent of the choice of holomorphic functions $\mathcal{A}_{\pm}$ shows that it is perhaps not particularly surprising that the massive spin-2 excitations around the supersymmetric $AdS_6$ vacua found in \cite{Gutperle:2018wuk} are universally present for any $\mathcal{A}_{\pm}$.
It would be interesting to extend the ansatz (\ref{general:ansatz:1})-(\ref{general:ansatz:G5}) further to include more fields in the truncation from 10D. For example, one could consider coupling vector multiplets to the 6D $F(4)$ theory; such an ansatz would then provide a 10D uplift of the 6D Janus solutions found in \cite{Gutperle:2017nwo}, which should correspond to mass deformations of the dual 5D SCFTs. It is not obvious, however, whether the consistent truncation obtained here can be generalized to include the addition of matter multiplets, as non-linear Kaluza-Klein consistency often requires a delicate balance between internal harmonics unless a symmetry argument can be made \cite{Duff:1984hn,Hoxha:2000jf}, which does not appear to be the case for the Janus solutions. A possibly more fruitful approach would be to consider instead gauge fields that arise from D3-branes wrapping 3-cycles \cite{Bergman:2018hin}, as they are not subject to the standard constraints of Kaluza-Klein consistency.
\acknowledgments
This project was initiated following discussions with C.F.~Uhlemann at the 2018 USU Workshop on Strings and Black Holes. We wish to thank C.F.~Uhlemann for enlightening discussions and E.~O~Colgain and O.~Varela for useful comments. This work was supported in part by the U.S.~Department of Energy under grant DE-SC0007859.
|
\section{Introduction}
\label{sec:intro}
There are several ways by which one can approach the study of galaxy
evolution. One approach is to classify the galaxies in the nearby
Universe into different morphological types and study their
properties. In such an approach, the study of S0 galaxies has received
considerable attention in the last decade. S0 galaxies are
characterised by the presence of, what appears to be, an
elliptical-like bulge and a smooth outer disk devoid of any apparent
spiral arms. Indeed, the study of these galaxies is important because
hidden within their inter-relationship with galaxies of other types
are clues to the varied processes that play important roles in the
overall picture of galaxy evolution (See \citet{Aguerri2012} and
references therein).
For example, observational studies suggest that the central components
of S0 galaxies viz. bulges resemble elliptical galaxies
\citep{Renzini99}. They have thus likely formed in a similar manner
to ellipticals, i.e. through hierarchical clustering and
mergers. However, the observation that the fraction of S0 galaxies
within a cluster of galaxies grows at the cost of spirals
\citep{Dressler1980}, suggests that S0s originate from spirals. Gas
stripping via ram pressure is often considered a viable process for
such a transformation in dense cluster environments \citep{Gunn1972,
Aragon-Salamanca2006, Laurikainen2010, Maltby2015}. However, we know
that S0's are equally common in groups \citep{Wilman2009, Just2010,
Bekki2011}. These observations suggest that the processes such as
ram pressure stripping are not responsible for the formation of the
majority of S0s in groups where minor mergers and tidal effects can
control their evolution \citep{Mazzei2014, Mapelli2015}.
Simulations too suggest that there are multiple methods for forming
S0s - they can arise in major mergers, can grow through slow/secular
processes and can also form through gas stripping of spirals
\citep{Querejeta2015, Tapia2017}. Traces of past merger events which
impact the observed kinematics have been observed in some S0s
\citep{Falcon-Barroso2004}. The observed star formation (SF) and SF
history (SFH) characteristics of S0s can be complex
(e.g. \citet{Barway2013} and references therein).
\citet{Barway2007, Barway2009} suggest that the dominant formation
process in case of S0s is a function of luminosity, with brighter S0s
forming in a manner similar to ellipticals and fainter S0s forming
through secular processes or perhaps by gas stripping of spirals
\citep{Rizzo2018},
which in turn may have formed through secular evolution.
\citet{Vaghmare2013} used archival deep mid-infrared images from the
Spitzer Infrared Array Camera to systematically study the bulges of S0
galaxies. Bulges, as defined photometrically, are excesses of light
over an inward extrapolation of an exponential profile fitted to the
outer disk. They are known to occur in two varieties - classical,
resembling ellipticals in almost every aspect including morphology,
colour, nature of stellar populations and kinematics, and pseudo,
resembling disks and believed to have formed through secular
evolution \citep{Kormendy2004}. \citet{Vaghmare2013} point out
that pseudobulges do exist in S0 galaxies and preferentially
occur in fainter S0s. But what is perhaps a more interesting result is
that disks of pseudobulge hosting S0s have a lower scale length than
those hosting classical bulges.
In \citet{Vaghmare2013}, the authors speculate that the lower scale length could imply either
(i) that the population of disks with lower scale length on average preferentially
host pseudobulges or (ii) that the lowered scale length is a result of some process
responsible for either the bulge's or the overall galaxy's growth. In a
follow-up study, \citet{Vaghmare2015} compare the disk properties of spirals and
S0s and find that pseudobulge hosting S0s have a lower scale length than the
pseudobulge hosting spirals. The authors explain this as being due to lowered disk
luminosity in case of S0s hosting pseudobulges. They further compare bulge
luminosities to demonstrate that only the early-type spirals may transform into
S0s with pseudobulges through disk fading via gas stripping, while late-type
spirals will need additional processes, especially to aid the growth of bulges to
observed luminosities.
The conclusions reached by \citet{Vaghmare2013, Vaghmare2015} and by Barway et
al. rely on statistical arguments. Correlations are derived between
various parameters, which in turn are derived from imaging data. While such
studies indeed offer insights into broad trends among galaxies in a given class,
they cannot provide information on individual galaxies and their star formation /
assembly histories. For this, one requires data from multiple wavebands.
\citet{Barway2013} use broad band photometry from multiple surveys
spanning UV, optical and infrared wavelengths but again use statistical
arguments to point out broad trends among S0 galaxies. In order to truly
constrain the formation histories of S0s, one needs to have spectroscopic data.
Using stellar population synthesis techniques, one can then derive very detailed
star formation histories for these galaxies and confront the conclusions reached
by the photometric studies. To this effect, we have been using the Southern
African Large Telescope's (SALT) Robert Stoby Spectrograph (RSS) to observe a
select sample of S0 galaxies.
The work presented in this paper
focusses on a subset of the galaxies - those shown to host pseudobulges by
\citet{Vaghmare2015}, whose photometric properties and their implications were
summarised earlier. The aim here is to check
for consistency between star formation histories and the conclusions based on
the photometric observations. This paper is organised
as follows - the data and the reduction methods are described in section
\ref{sec:data}, section \ref{sec:specmod} explains the process of modelling the
spectra using \textsc{starlight} while sections \ref{sec:result}, \ref{sec:discussion}
and \ref{sec:bar_dn4000}
present the detailed results and possible interpretations.
Throughout this paper, we use the standard concordance cosmology with $\Omega_M=
0.3$, $\Omega_\Lambda= 0.7$ and $h_{100}=0.7$. Also, magnitudes used
in this work are in the AB system.
\begin{figure*}
\centering
\includegraphics[width=0.8\textwidth]{figure1.pdf}
\caption{The 3.6 micron images obtained using the Spitzer IRAC, for the eight
galaxies studied in the present paper.}
\label{fig:spitzer_images}
\end{figure*}
\section{Sample and Observations}
\label{sec:data}
\begin{table*}
\begin{center}
\begin{minipage}{166mm}
\caption{Basic parameters of the sample galaxies.}
\label{tab:pb_observations}
\begin{tabular}{l c c c c l}
\hline
Object Name & RA (J2000) & Dec (J2000) & Total magnitude & Exp. Time & Position Angle \\
& hh:mm:ss & dd:mm:ss & B band & seconds & degree \\
\hline
ESO079-007 & 00:50:04 & -66:33:10 & 13.73 & 4500 & 2.6 \\
NGC1522 & 04:06:08 & -52:40:06 & 13.97 & 5000 & 38.6 \\
NGC1533 & 04:09:52 & -56:07:06 & 11.79 & 3600 & 170 \\
NGC1510 & 04:03:33 & -43:24:01 & 13.47 & 4500 & 102.9 \\
IC2085 & 04:31:24 & -54:25:01 & 13.95 & 5000 & 112 \\
ESO085-030 & 05:01:30 & -63:17:36 & 13.79 & 4500 & 148.2 \\
NGC5750 & 14:46:11 & -00:13:23 & 12.55 & 4500 & 70.6 \\
NGC7709 & 23:35:27 & -16:42:18 & 13.60 & 4500 & 55 \\
\hline
\end{tabular}
\textbf{Notes:} Column 1 is the galaxy name ; Column 2 \& 3 are Right
Ascension and Declination of galaxy ; Column 4 is the total B band
magnitude; Column 5 is the total exposure time for RSS spectrum and
Column 6 is the slit position angle, measured North to East.
\end{minipage}
\end{center}
\end{table*}
The sample used in the present study is described in detail in
\citet{Vaghmare2013} and \citet{Vaghmare2015}. For convenience,
we summarise its essential features here. The parent sample comprises
3657 S0 galaxies as found in the \emph{Third Revised Catalogue}
\citep{deVauc91}. A total B-band magnitude cut-off of 14.0 was imposed
to reduce this sample to 1031 galaxies which were then cross-matched
with the Spitzer Heritage Archive (SHA). We found mid-infrared images
taken by the Spitzer Infrared Array Camera (IRAC) for 247 galaxies at
3.5 $\mu m$. These images were processed to achieve the desired
quality for 2D decompositions and structural parameters were obtained
by using GALFIT \citep{Peng2010} for a subset of 185 S0s. Of these, 25
were classified as having pseudobulges using a very conservative
criterion -- bulges should lie more than 3-$\sigma$ below the Kormendy
relation \citet{Kormendy2004} and should have an S\'{e}rsic index
$n<2$. This is a combination of the criteria used by
\citet{Gadotti2009} and \citet{Fisher2008} (see
\citet{Vaghmare2013} for full details).
We used the 10m class Southern African Large Telescope (SALT) \citep{Buckley2006} and its
Robert Stobie Spectrograph (RSS)\citep{Burgh2003, Kobulnicky2003} to study the sample. The RSS is
a versatile instrument providing several observing modes and is designed to work in a wavelength range of 320 - 900 nm.
Of the 25 pseudobulges from \citet{Vaghmare2015}, 15 were in the
declination range accessible to SALT. We were given time to observe 8
of these galaxies which we present in this paper. A brief summary of the
observational parameters is presented in Table \ref{tab:pb_observations}.
Spitzer 3.6 micron imaging for the galaxies is shown in Figure \ref{fig:spitzer_images}.
The long-slit RSS spectra were obtained using a 1'' wide slit and the PG0900 grating, placing the slit
along the major axis of the galaxy in each case.
RSS uses an array of three
CCDs of $2048 \times 4096$ pixels, and the grating angle in the instrument was set in a way
so as to avoid any important spectral features falling in the resulting two 15 pixel size gaps.
We used a $2 \times 4$ CCD binning scheme to improve the
signal-to-noise ratio (SNR). The spectra obtained in the said configuration have an observed
wavelength range of 3640-6765 \AA \ and thus cover the required optical range
for unravelling the stellar formation histories of these galaxies. The spectral resolution is $\sim$ 3 \AA.
We used multiple 600-900 sec individual exposures for achieving the total
exposure times per galaxy as listed in Table~\ref{tab:pb_observations}.
For 7 of the 8 galaxies in our sample, this was done over two
separate visits to the targets. Calibration unit Arc lamps and flat-field images were observed with the same configuration as
used for the science targets and spectrophotometric standards.
\subsection{Spectroscopic Reduction}
We made use of the SALT \emph{product data} generated by the in-house pipeline called
PySALT \citep{PySALT}, which mosaics the individual CCD data to a single FITS file, corrects for cross-talk effects,
and performs bias and gain corrections. We then carried out further reduction
steps using our own custom tools,
written in Python/PyRAF, which consist of several
modules designed to handle specific steps of processing.
First, we trim the regions of the CCD not containing any usable
data and then fill the CCD gaps using a gradient fill. To do this, we
calculate the
the gradient across the gap by measuring the flux values of
the pixels to the right and the left of the gap, computing the difference
and normalising it to the number of pixels belonging to the CCD gap (the
gradient). We then use the gradient to determine the flux values for
pixels in each row of the image and replace the zero value pixels by
these flux values. This step is
important to prevent discontinuities in the brightness distribution which
affect the flat-fielding process which relies on the modelling of the large
scale illumination structure. Another advantage of this gradient fill is
seen in the subsequent step of background subtraction. A CCD gap which
appears straight in the original images becomes curved as a result of the
coordinate transformation step. This causes background subtraction to result
in discontinuous patches. By filling the CCD gap, these patches can be
avoided.
We then remove the cosmic rays using the LACosmic
algorithm \citep{LACosmic}. Calibration in wavelength is achieved using arc
lamps and has typical error of $\pm 0.35$ \AA. We use the arc lamp to
determine a coordinate transformation to `straighten' frames so that every
column corresponds to a unique wavelength making the background sky shape
fitting and subtraction possible. Using the spectrophotometric
standard data, we determine the relative flux calibration. We then check
individual frames for alignment and co-add them. During this step,
a standard deviation frame is constructed from which an error frame is obtained.
Before the co-addition, to have consistent noise characteristics, we equalize
the effective exposures of the frames -- this is needed because the pupil of SALT
changes by design during the observation.
If $O_i(x, y)$ are
the individual observed frames, the final spectrum $F(x,y)$ and error spectrum
$E(x, y)$ can be written as,
\begin{equation}
F(x,y) = \left< O(x,y) \right> = \frac{\sum_i^N O_i(x, y)}{N}
\label{eq:final_frame}
\end{equation}
\begin{equation}
E(x,y) = \sqrt{\frac{\sum_i^N (O_i(x,y) - \left<O(x,y)\right>)^2}{N-1}}
\label{eq:error_frame}
\end{equation}
An alternate way of obtaining the error frames is to start from the raw CCD
images whose noise properties are relatively well understood to try to model the effects of
every single reduction step on the errors. This is however very complicated and
so
we adopt the
above technique of determining our final error frame. Since we are combining
only 6 - 8 frames, we might overestimate our errors, but this
over-estimation does not affect the findings presented in the
paper. Our final step is to correct for foreground extinction and this is done using
the \emph{deredden} task in IRAF.
\subsection{Post Reduction Procedure}
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{aperture_lab.png}
\includegraphics[width=0.4\textwidth]{aperture_lab2.png}
\caption{A plot of the brightness profile of a galaxy along the slit. The
apertures, rows which are summed up to produce the 1-d spectrum in the
given region of the galaxy, are marked in alternating colors. The rows lying
between a pair of successive red and blue lines are summed to obtain the
spectrum corresponding to the centre of this aperture.
As can be seen, the width of these apertures increases in inverse ratio to the level of flux.
The lower panel shows a close-up of the central region.}
\label{fig:aperture}
\end{figure}
The modelling of the spectrum to study its star formation history has been
done using \textsc{starlight} \citep{Fernandes2005}. This program requires a 1-d spectrum
in the form of an ASCII table comprising at least two columns, namely wavelength
and flux. Furthermore, we wish to study the spectra at different spatial positions.
We thus extract 1-d spectra from positions along the
long slit by summing up a suitable number of rows. If this number
is kept constant as a function of the position on the long slit, the
signal-to-noise ratio degrades for spectra extracted far away from the centre due to
the decreased brightness of the galaxy in this region. So, we came up with a
scheme of increasing the number of rows summed up as an inverse function of the
total amount of flux from the galaxy in a given region of the slit. This is
illustrated in Figure \ref{fig:aperture}, which shows the flux in scaled
counts as a function of position along the slit. The routine that
determines the sizes of these apertures requires as an input the number of
apertures to be extracted and the range of the spatial intensity profile over which the
extraction is to be carried out, and then determines these aperture sizes in units
of pixels so as to keep the total flux within the aperture roughly constant.
The routine does not take into account the physical scale of the galaxy. The
number of apertures to be extracted are chosen subjectively based on the overall
quality of the spectrum which in turn depends on the brightness of the galaxy.
Due to the rotation of stars in the galaxies, the spectra
represented by individual rows are shifted with respect to each other.
We constructed a Python based module to characterise this by tracing
the centroid of individual absorption features along the long slit. A rotation curve
of the galaxy is thus obtained. By fitting polynomials of suitable order to this
curve, it is also possible to \emph{derotate} individual rows for the purpose of
stellar population fitting -- this is necessary to not introduce broadening
of spectral features due to rotation. We shift each row in the wavelength space
using the fitted rotation curve information before co-adding rows to be fit with \textsc{starlight}.
During the coaddition of the rows, the corresponding rows from the error frame are also
added in quadrature and divided by number of rows to get the error spectrum.
The spectrum and the related uncertainties can now be read and modelled
by \textsc{starlight}.
\section{Spectral Modelling}
\label{sec:specmod}
There are several methods to derive the star formation history of a galaxy
from its spectrum (see \citet{Sanchez2016} for a critical summary).
In the present study, we adopt the method implemented in \textsc{starlight}
\citep{Fernandes2005}. We briefly explain the method for the convenience
of the reader.
Any arbitrary star formation history can be approximated as a series of star
bursts in each of which a population of stars with a given initial mass function (IMF)
is formed almost at the same time. The spectrum of such a population, known as
a Simple Stellar Population (SSP), can be determined as a function of age and
metallicity, knowing the IMF and adopting a stellar spectral library. The
observed spectrum can then be modelled as a linear combination of spectra of
different SSP. A galaxy whose formation is consistent with a single
episodic burst of star formation would be described by a single SSP and thus any
attempt to model such a galaxy spectrum would lead us to finding virtually no
contribution from any SSP but one. And for any other formation history, the
relative contributions would be proportional to the strength of the star bursts
at each epoch. This way, the star formation history of the galaxies can be
obtained.
Such a method is implemented by \textsc{starlight} which accepts an input 1-d spectrum, a
set of base spectra and data flags/masks and fits a \emph{population vector} to
the observed spectrum, which tells us the relative contribution of each spectrum
in the base to the observed spectrum. Mathematically, the model spectrum can be
written as
\begin{equation}
M_{\lambda} = \sum_{j=1}^{N_*} P_j L_{\lambda,j}^0 \ast G(v_*, \sigma_*)
10^{-0.4A_{\lambda,j}}
\label{eq:starlight_model}
\end{equation}
\noindent Here, $P_j$ refers to fractional contribution of jth base spectrum to
the total spectrum. $L_{\lambda,j}^0$ refers to the jth base spectrum without any
extinction or kinematic filters applied, G refers to a Gaussian characterized by
a centroid $v_*$ and a width $\sigma_*$. The operation $\ast$ implies a
convolution. The last multiplicative factor is to account for intrinsic
extinction along the line of slight. $N_*$ refers to the total number of base
spectra.
However, an important issue concerning the
use of \textsc{starlight} is with regards to the base selection. A hand picked base may be biased
towards certain ages and metallicities and thus introduce biases in the deduced
star formation history. An attempt to conservatively choose all possible ages
and metallicities increases computation overheads (as $N_*^2$). We thus adopt the prescription of \citet{Richards2009}
who suggest using diffusion mapping with K-means clustering to obtain a base of
M spectra starting from a very large number N of base spectra covering all or most of
the parameter space of interest, with $M \ll N$. For the present study we obtain
150 base spectra from the MILES \citep{MILES} spectral library obtained using the
Salpeter IMF,
and use diffusion mapping with K-means clustering to construct
a base of 45 spectra, which uniformly covers the age-metallicity parameter
space. The initial base spectra are obtained from a web application provided by
the MILES group. The code for applying the \citet{Richards2009} algorithm is
provided by them as a MATLAB code. We wrote a converter to reformat the spectra
obtained from the MILES site into a form compatible with the MATLAB code. We
wrote another script in Python to transform the output of the MATLAB code into a
form compatible with \textsc{starlight}. The final 45 spectra used for modeling
the spectra of the galaxies is presented in Appendix A.
In order to compute the errors on the parameters returned by \textsc{starlight}, we adopt
the following approach. We use the derived error spectrum to construct several
($\sim 100$) realisations of the observed spectrum and run \textsc{starlight} for each
one of them. These realisations are constructed by approximating the
underlying error distribution to be a Gaussian and drawing random samples
from a Gaussian distribution with zero mean and a standard deviation equal
the corresponding uncertainty read off from the error spectrum. The standard deviation of any given parameter obtained for these
realisations is adopted as an estimate of its uncertainty. As noted earlier, the
method by which error frames are being constructed is likely to lead to an
over-estimation of the uncertainty. This means that the error bars obtained on
the \textsc{starlight} parameters will also be over-estimated. As can be gauged from the
discussion of the science results in the subsequent sections, this
over-estimation does not affect the conclusions drawn in this paper.
\section{Results}
\label{sec:result}
\begin{figure*}
\centering
\includegraphics[width=0.9\textwidth]{spectra_all.png}
\caption{Observed spectrum (grey line) and best-fit \textsc{starlight} model
(red line) for the central regions of galaxies in our sample.
The grey regions indicate wavelengths ignored during the \textsc{starlight} fit as
they contain emission from hot gas or are lost in the CCD gaps.}
\label{fig:spectra_all}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.9\textwidth]{all_histories.png}
\caption{Star formation histories for the galaxies in our sample.
The X-axis indicates the logarithm of the age and the Y-axis the
percentage of light contributed to the observed spectrum by a stellar
population of the given age. These percentages are computed from the
population vector returned by \textsc{starlight}.The hatch patterns are circles for the first six
galaxies which are unbarred and forward leaning diagonals for the last two
galaxies which are barred.}
\label{fig:sfh_all}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.9\textwidth]{all_metals.png}
\caption{Metallicity contributions for the galaxies in our
sample. The X-axis is the metallicity while the Y-axis represents the
fractional contribution to the flux of the observed spectrum as determined from the
population vector. The hatch patterns are circles for the first six
galaxies which are unbarred and forward leaning diagonals for the last two
galaxies which are barred.}
\label{fig:metal_all}
\end{figure*}
\begin{figure*}
\includegraphics[width=\textwidth]{figure6_1.png}
\caption{Age and metallicity gradients for the galaxies in our sample. The
X-axis in all the plots is the distance relative to the centre of the galaxy
in kpc. For plots on the left, the Y-axis indicates the fraction of the
stars with age $ > 10^9$ yr, while for the plots on the right, it indicates the light
weighted metallicity. Both quantities are computed by \textsc{starlight} and can be
obtained from the population vector output by it.}
\label{fig:grads_others1}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{figure6_2.png}
\contcaption{Age and metallicity gradients for the galaxies in our sample. The
X-axis in all the plots is the distance relative to the centre of the galaxy
in kpc. For plots on the left, the Y-axis indicates the fraction of the
stars with age $ > 10^9$ yr, while for the plots on the right, it indicates the light
weighted metallicity. Both quantities are computed by \textsc{starlight} and can be
obtained from the population vector output by it.}
\label{fig:grads_others2}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{figure6_3.png}
\contcaption{Age and metallicity gradients for the galaxies in our sample. The
X-axis in all the plots is the distance relative to the centre of the galaxy
in kpc. For plots on the left, the Y-axis indicates the fraction of the
stars with age $ > 10^9$ yr, while for the plots on the right, it indicates the light
weighted metallicity. Both quantities are computed by \textsc{starlight} and can be
obtained from the population vector output by it.}
\label{fig:grads_barred}
\end{figure*}
\begin{table*}
\begin{center}
\begin{minipage}{166mm}
\caption{Best fit Bulge and Disk parameters for the S0 galaxies in our
sample.}
\label{tab:phot_params}
\begin{tabular}{l c c c c c c c c c c c c}
\hline
Name & $T$ & $L_K$ & \multicolumn{3}{c}{Bulge parameters} &
\multicolumn{2}{c}{Disk parameters} & \multicolumn{3}{c}{Bar Parameters} & B/T
& Bar/T \\
& & & $<\mu_e>$ & $\rm{r}_e$ & $n$ & $\mu_{0d}$ & $\rm{r}_d$ &
$\rm{m}_{bar}$ & $n_{\rm{bar}}$ & $\rm{r}_{e(\rm{bar})}$ & & \\
\hline
\hline
ESO079-007 & -2.0 & -18.73 & 23.25 & 0.87 & 0.53 & 21.23 & 1.72 & -- &
--
& -- & 0.04 & -- \\
NGC1522 & -2.3 & -18.33 & 20.63 & 0.36 & 0.44 & 21.83 & 0.81 & -- & --
&
-- & 0.37 & -- \\
NGC1510 & -1.8 & -20.24 & 19.18 & 0.24 & 1.18 & 21.41 & 0.82 & -- & --
&
-- & 0.40 & -- \\
IC2085 & -1.2 & -18.48 & 20.37 & 0.53 & 0.88 & 21.79 & 1.24 & -- & -- &
-- & 0.41 & -- \\
ESO085-030 & -0.4 & -18.76 & 20.50 & 0.86 & 1.41 & 20.94 & 0.78 & -- &
-- & -- & 0.64 & -- \\
NGC7709 & -1.9 & -20.92 & 20.65 & 0.72 & 0.41 & 21.03 & 1.70 & -- & --
&
-- & 0.20 & -- \\
NGC1533 & 0.1 & -18.79 & 16.73 & 0.22 & 0.95 & 19.77 & 1.58 & 12.94 & 0.50
&
0.89 & 0.22 & 0.07 \\
NGC5750 & -2.0 & -21.62 & 17.84 & 0.32 & 0.98 & 19.94 & 2.56 & 13.64 & 0.20
& 2.23 & 0.09 & 0.09 \\
\hline
\end{tabular}
\textbf{Notes:} Column 1 - The common name of the galaxy; Column 2 - Hubble
stage parameter $T$, Column 3 - K-band absolute magnitude (AB system), Column 4 - the
average surface brightness of the bulge within its effective radius, in mag per
arcsec$^2$, Column 5 - bulge effective radius in kpc, Column 6 - the bulge
S\'{e}rsic index, Column 7 - disk central brightness in mag per arcsec$^2$,
Column 8 - disk scale length in kpc, Column 9 - integrated apparent magnitude
of the bar, Column 10 - S\'{e}rsic index of bar, Column 11 - Bar effective
radius in kpc, Column 12 - Bulge-to-Total luminosity ratio, Column 13 -
Bar-to-Total luminosity ratio.
\end{minipage}
\end{center}
\end{table*}
For the eight galaxies in the current sample, we have the photometric parameters
from \citet{Vaghmare2015}
obtained by modelling the mid-infrared surface brightness distribution. These parameters are reproduced in Table
\ref{tab:phot_params}. In Table \ref{tab:spec_params}, we have various spectral
parameters derived by fitting \textsc{starlight} models to the spectra from the central
regions of these galaxies. Also included in this table are $D_n(4000)$ index
\citep{Balogh1999}
values computed as the ratio of the
integrated fluxes from 4000 - 4100 \AA\ and 3850 - 3950 \AA\ bands. Using standard statistical techniques, we have studied
correlations among photometric and spectroscopic
parameters which we discuss in Section \ref{sec:corr_specphoto}.
Using the tools and techniques explained in the previous sections, we were able
to model the spectra from the central regions of the galaxies in our sample and
derive their various stellar population parameters. The central spectra and the best fit
\textsc{starlight} models are shown in Figure \ref{fig:spectra_all}.
The star formation history as output by \textsc{starlight} is a
population vector which contains percentage contributions of SSPs of various
ages and metallicities, present in the base spectra, to the observed spectrum. To represent the star formation
history of the galaxy on a plot, we consider uniform bins of age in the log
space and use the population vector to determine the total percentage contribution by
all SSPs in a given age bin. So the plot is made with the logarithm of the age
on the X axis and percentage of
light contributed by an SSP at that age along the Y-axis. The star formation
histories of the galaxies are shown in Figure \ref{fig:sfh_all} and the metallicity
histograms in Figure
\ref{fig:metal_all}.
In Figure \ref{fig:grads_others1},
we show the age and
metallicity gradients for the six unbarred galaxies and the two barred galaxies.
These plots are constructed as follows. For any given galaxy, we define the the
center using light weighted average row number for the brightness profile along
the slit. This is labelled as zero in all the plots. For each spectrum derived,
a light weighted centroid is assigned to it using the spatial profile in the
series of rows which were summed up to obtain the spectrum. \textsc{starlight} models
give us light weighted age and metallicity for this spectrum as well as
the errors on them. This way, one can obtain the gradients of the light weighted
age as well as the light weighted metallicity.
\subsection{Broad Trends}
In this section, we will discuss the broad trends found in this sample of eight
galaxies and reserve the next section for comments on individual objects. As
discussed earlier, the galaxies in this sample are a subset of the sample of
photometrically classified S0s hosting pseudobulges from \citet{Vaghmare2015}.
Pseudobulge hosting galaxies are in general expected to have a complex star
formation history viz. their spectrum cannot be described using that of a Simple Stellar Population.
This is because pseudobulges are believed to have been formed through slow and
secular processes with sustained star formation over a long period of time. An
SSP, on the other hand, describes a stellar population which has taken birth in
a single episodic burst.
The photometric studies on the current sample done by \citet{Vaghmare2015}
reveal that their disks have a smaller scale length than their counterparts
among spiral galaxies. The authors argue that since the total luminosity of the
disk goes as $ L \sim I_d(0) r_d^2 $ the smaller scale length $r_d$ implies
that disks have a lower luminosity, assuming similar values of $I_d(0)$ which
they observationally confirm. They argue that this lower luminosity is due to
stripping of gas, a process often invoked to explain the fading of spiral arms
in spiral galaxies resulting in a morphological transformation to S0s. The
authors further argue in favor of gas stripping of spiral disks by comparing bulge luminosities.
Pseudobulges in S0s have a lower luminosity, which they say is due to the lower
availability of gas in S0s to drive bulge growth.
Based on these two known aspects of the current sample of S0s - that
they are pseudobulge hosts and they have undergone gas stripping, one
would expect two results from a detailed spectral study of such
galaxies. Firstly, we expect the spectrum of these galaxies to not be
well fitted by a single SSP spectrum. Secondly, we expect no strong
indications of recent star formation. Both these expectations can now be
tested using the spectral data obtained in the present study.
In Figure \ref{fig:spectra_all} we show the
final reduced spectra from the central regions of the eight galaxies. It is
clear that the spectra of the non-barred galaxies (first six) indicate a population
dominated by young and recently formed stars while the spectra of barred
galaxies indicate an old and evolved stellar population likely to be well described
using an SSP.
A spectrum of a typical elliptical galaxy, a class of galaxies
known to be consistent with having an SSP, are devoid of strong emission lines
and exhibit a strong break at the 4000 \AA\ region due to metal absorption
in the outer atmospheres of old stars.
Such a break is seen in the spectrum of the two barred galaxies. But the break
is absent from the spectrum of five of six unbarred galaxies in Figure
\ref{fig:spectra_all}, indicating that the spectrum in their case is
dominated by contributions from a younger stellar population.
Further, one can see that most of the spectra show strong emission lines such as
$H\alpha$. This is counter to
the expectation that there should be no ongoing star formation in the current
sample of S0 galaxies, assuming that they are fully gas stripped and transformed
spirals.
In Figure \ref{fig:sfh_all}, we see the derived star
formation history of the central regions of these galaxies. One can again see that the
barred and the unbarred galaxies behaving differently. In case of
unbarred galaxies, we see peaks towards the lower end of the age axis indicating
that a large fraction of the light contributed to the observed spectrum comes
from the lower age population. We see the opposite result in the case of barred
galaxies which show higher peaks at the higher end of the age axis. An important
point to note regarding Figure \ref{fig:sfh_all}
concerns the lower limit on the age axis. The strong $H\alpha$ lines suggest
there must be a population of stars with $\log(\rm{age}) <7.0$ but the lowest
point on the age axis is 7.5. This is because the MILES spectral library does
not allow one to probe ages lower than this and not necessarily because there is
no population of age lower than the above limit. So, the lowest age bin should
be interpreted as containing percentage contribution from populations at that
age and lower.
Next, we explore the metallicity distributions in the unbarred and barred
galaxies, as shown in Figure
\ref{fig:metal_all} respectively. Again, it
is easy to spot the differences. In case of unbarred galaxies, there is a common
peak at 0.005 - 0.006 subsolar metallicity with little contribution from higher
metallicities. On the other hand, in case of barred galaxies, metallicities of
different ranges have comparable contributions to the stellar populations.
\begin{table*}
\caption{Best fit \textsc{starlight} parameters for the spectra from the central regions
of the eight galaxies in our sample.}
\label{tab:spec_params}
\begin{tabular}{l c c c c c c c c c c c c c}
\hline
Name & LWZ & MWZ & LWAge & MWAge & Av & Young & Inter & Old & $\textrm{Mass}_{\textrm{tot}}$ & $D_n(4000)$ \\
\hline
\hline
ESO079-007 & $ 0.007 $ & $ 0.009 $ & $ 8.275 \pm 0.039 $ & $ 9.706 \pm 0.053 $ & $ 0.488 \pm 0.056 $ & 0.768 & 0.0 & 0.232 & 8.429 & 1.076 \\
NGC1522 & $ 0.006 $ & $ 0.007 $ & $ 8.018 \pm 0.021 $ & $ 9.324 \pm 0.084 $ & $ 0.314 \pm 0.041 $ & 0.849 & 0.063 & 0.088 & 7.715 & 0.983 \\
NGC1510 & $ 0.006 $ & $ 0.008 $ & $ 8.029 \pm 0.021 $ & $ 9.356 \pm 0.094 $ & $ 0.282 \pm 0.035 $ & 0.754 & 0.194 & 0.052 & 7.917 & 1.002 \\
IC2085 & $ 0.009 $ & $ 0.013 $ & $ 8.953 \pm 0.025 $ & $ 9.961 \pm 0.03 $ & $ 0.892 \pm 0.019 $ & 0.429 & 0.03 & 0.541 & 8.125 & 1.329 \\
ESO085-030 & $ 0.008 $ & $ 0.01 $ & $ 8.627 \pm 0.019 $ & $ 9.82 \pm 0.048 $ & $ 0.582 \pm 0.022 $ & 0.597 & 0.0 & 0.403 & 8.285 & 1.201 \\
NGC7709 & $ 0.006 $ & $ 0.01 $ & $ 8.066 \pm 0.027 $ & $ 9.697 \pm 0.103 $ & $ 0.116 \pm 0.046 $ & 0.871 & 0.0 & 0.129 & 8.581 & 1.011 \\
NGC1533 & $ 0.02 $ & $ 0.02 $ & $ 10.176 \pm 0.01 $ & $ 10.224 \pm 0.003 $ & $ 0.042 \pm 0.011 $ & 0.018 & 0.0 & 0.982 & 9.065 & 1.731 \\
NGC5750 & $ 0.011 $ & $ 0.015 $ & $ 9.615 \pm 0.026 $ & $ 10.053 \pm 0.022 $ & $ 0.279 \pm 0.018 $ & 0.171 & 0.0 & 0.829 & 9.358 & 1.543 \\
\hline
\end{tabular}
\textbf{Notes:} Column 1 - The common name of the galaxy;
Column 2 - Light Weighted Metallicity;
Column 3 - Mass Weighted Metallicity;
Column 4 - Log of Light Weighted Age [yr];
Column 5 - Log of Mass Weighted Age [yr];
Column 6 - Extinction Parameter;
Column 7 - Fraction of Young Stars (Age $< 10^8$ yr);
Column 8 - Fraction of Intermediate Stars ($10^8 < $ Age $ <10^9$ yr);
Column 9 - Fraction of Old Stars (Age $>10^9$ yr);
Column 10 - Log of Galaxy's Total Mass [$M_{\sun}$];
Column 11 - The $D_n(4000)$ index. The typical uncertainties in case of both
light and mass weighted metallicities are of the order of $10^{-4}$ and thus
have not been provided in the table above.
\end{table*}
Having summarised the broad trends in spectra, the star formation histories and
the metallicity distributions of the galaxies in the current sample, we now move
towards a more detailed description of individual galaxies.
\subsection{Discussion of Individual Objects}
\subsubsection{ESO079-007}
\begin{figure}
\centering
\includegraphics[width=0.8\columnwidth]{ESO079_asy2.png}
\caption{A narrow region of the 2-d spectrum of the galaxy ESO079-003, near
the $H\alpha$, N[II] and S[II] emissions. The continuum has been marked along
with the emission lines. As can be seen, the distribution of the emission
lines about the continuum is asymmetric.}
\label{fig:eso079_asy}
\end{figure}
The NASA Extragalactic Database (NED), describes this S0 galaxy as being
peculiar. What sets this galaxy apart from conventional S0 galaxies is the
presence of asymmetric star forming regions i.e. the star forming regions are
more towards one side of the centre than the other. To illustrate this better,
we show in Figure \ref{fig:eso079_asy}, the 2-d spectrum near the wavelength
region dominated by $H\alpha$, N[II] and S[II] emission lines. The figure shows
the stellar continuum in the centre of the galaxy. It is easy to see that there
is a stronger star formation activity on one side of the galaxy centre than the
the opposite side. The overall metallicity of the galaxy is also sub-solar.
It is likely that the galaxy is accreting gas only from one direction
in the disc which builds up the gas reservoir to induce disc
instabilities and trigger asymmetric star formation \citep{Vulcani2018}.
Galaxy interactions in the form of fly-bys or
recent accretion of smaller satellites could also account for the
asymmetric distribution of star formation within the disc \citep{Mapelli2008}.
\subsubsection{NGC 1522}
NED classifies this galaxy as a peculiar S0 galaxy. The spectrum of this galaxy
suggests a very young population and strong ongoing star formation dominating
the light budget at optical wavelengths. \textsc{starlight} models suggest that 95\%
of the flux is explained by a stellar population younger than $10^8$ yr. What
makes the object even more interesting is that it is highly metal poor
suggesting that the gas responsible for the ongoing star formation is not
enriched by any Population 2 stars.
Bulge-disk decomposition for this galaxy shows that the bulge component of the
stars accounts for 40\% of the total galaxy light and the rest by a faded disk. Yet again we
have a case of a galaxy whose star formation indicates availability of a large
amount of gas but the disk shows signs of having faded, assuming a spiral
progenitor with a brighter disk.
\subsubsection{NGC 1533}
This S0 galaxy is a part of the Dorado Group with a clearly discernable bar structure. It exhibits
a smooth appearance and one look at the spectrum of the galaxy reveals that it
is dominated by a very old population. An age gradient plot reveals that indeed,
95\% of the galaxy light is dominated by old stars. From 3.6 micron imaging, we
know the bulge effective radius to be $\sim 4$''. It has been pointed out in the
literature \citep{Gadotti2009} that bulge light dominates the disk light for
atleast twice the bulge effective radius. So, it reasonable to expect that to
obtain a disk dominated spectrum sufficiently free from any bulge light
contamination, one would have to go twice as far (16'').
This means we do not have much
signal from the disk dominated region to be able to compare the populations of
bulges and disks. The metallicity is also near solar which is not too surprising
given other features of this galaxy.
This galaxy is clearly quite different from the first two examples in our
sample - devoid of star formation and dominated by older stars. It is
also different in another major way - it is a barred galaxy. Its formation
history is consistent with an old coeaval population but with a small component
of recently formed stars. These stars could have formed through gas infall
induced by bars.
\subsubsection{NGC 1510}
This galaxy is known to be a part of a pair with the other galaxy
being NGC 1512. According to \citet{Muerer2006}, this is a pair of
galaxies exhibiting starbursts and a common HI envelope. The
presence of an intense star burst is confirmed by our spectrum which
shows extremely strong $H\alpha$ emission. The age variation as a
function of distance from the centre reveals an asymmetric
feature. This means that there are more old stars on one side of the
galaxy. This population of old stars on one side is also metal rich
as can be seen in Figure \ref{fig:grads_others1}.
These features seem to be a result of the tidal effects of its
partner galaxy NGC 1512. The tidal features are clearly visible in
NGC 1510 from archival {\it GALEX} FUV and NUV images. The tidal interactions
have likely led to the asymmetric formation of star burst regions in
this galaxy. Again, the B/T ratio in this galaxy is quite high at
0.4.
\subsubsection{IC 2085}
This galaxy is part of NGC~1566 group of galaxies which is a part
of the Dorado Group. In comparison to other galaxies in this sample, the spectrum
obtained for this galaxy is of a relatively lower SNR. Even without
performing any modelling, one can see that this galaxy contains
both a very old and a very young stellar population in the centre. The
presence of strong emission lines such as $H\alpha$ support ongoing star
formation as well as recently formed stars while the 4000 \AA\ break caused by
several absorption features below 4000 \AA\
is indicative of an older population. Figure 9 shows an asymmetric
age gradient around the centre however the stellar metallicity
distribution is symmetric. This could be due to the bursts
of star formation in the presence of asymmetric dust features. The
optical image from Digital Sky Survey (DSS) confirms the presence of
disturbed dust lanes.
The best-fit model from \textsc{starlight} reveals that 50\% of the light is accounted for by a
young population with age $<10^8$ yr and the rest by very old stars.
Such a mixture of very young and very old stars hints at a composite
bulge system.
\subsubsection{ESO085-030}
This galaxy shares several features with ESO079-007 except for
the star forming regions which are not distributed asymmetrically.
About 40\% of the light in the centre of this galaxy is being
accounted for by old stars and this fraction falls to 15\% as one
moves outwards. The B/T ratio as determined from the 3.6-micron
imaging is 0.6 with the disk scale length smaller than the bulge
effective radius and the Sersic index is highest in the sample. This
means that at least in the centre, the bulge which is surrounded by the
small disk dominates significantly and thus the trend being observed
is due to the bulge. This indicates that pseudobulges grow in a
manner similar to disks i.e. inside-out.
That the pseudobulge can have a B/T ratio of 0.6 needs further explanation here.
Based on prior studies carried out on pseudobulges, one generally does not expect
such a high value of B/T. These studies have largely focussed on late-type
galaxies which are known to be dominated by disks. However, it is possible that
in the case of an S0 galaxy, the disk has lowered in luminosity due to a
morphology transformation through gas stripping. This would cause an increase in
the B/T ratio of these objects.
\subsubsection{NGC 5750}
This is an interesting object which exhibits a ring like structure. It is
classified as an S0/a galaxy which designates it as a transition class object,
i.e. an object which might be speculated to be transforming between S0 and Sa
morphology. Sandage, in \emph{The Carnegie Atlas of Galaxies} speculates that
the ring might actually be made of two very tightly bound spiral arms.
The galaxy is barred and in terms of spectral properties, is similar to NGC
1533, comprising of an old population with near solar metallicity. It is
also the second barred galaxy in our sample which has been classified as a
pseudobulge based on photometric properties, but it shows characteristics that
are atypical and not expected of pseudobulges.
\subsubsection{NGC 7709}
All conclusions about the NGC 1522 apply to this galaxy as is - there is
evidence of strong ongoing star formation, very little flux contribution from
the old stars and an overall sub-solar metallicity.
\subsection{Correlations between Photometric and Spectroscopic Parameters}
\label{sec:corr_specphoto}
\begin{figure}
\centering
\includegraphics[width=0.8\columnwidth]{bulmue_lwage.png}
\includegraphics[width=0.8\columnwidth]{bulmue_lwz.png}
\caption{A plot of bulge average surface brightness within effective radius as
a function of light weighted age and metallicity.}
\label{fig:bul_trend}
\end{figure}
In Figure \ref{fig:bul_trend}, are two plots showing the variation of bulge
average surface brightness within effective radius against light weighted age and
metallicity respectively. We see a clear trend that average surface brightness
increases with increasing age and metallicity.
All other correlations show significant scatter and thus have not been
discussed here. We would like to point out that the present study cannot conclusively
determine whether the scatter is because of low number statistics or indicates
an absence of a real correlation altogether. Further investigation with a larger sample
is warranted.
The Spearman rank correlation coefficient between average surface
brightness of the bulge within its effective radius and the logarithm of the
light weighted age obtained from the \textsc{starlight} model, is $-0.757$ with a
significance of $ > 95$ \%. A similar correlation is seen with metallicity; the rank
correlation for this is found to be $-0.737$ with similar significance.
These correlations point at a relation between age and metallicity of
the population and the growth / build up of bulges.
These correlations could also be driven by the galaxy mass. For example, if the
age or metallicity are correlated with the mass of the galaxy
\citep{Gallazzi2005}, then one should
expect to see the correlation we find in the upper panel of Figure
\ref{fig:bul_trend}. A simple way to
test the possibility is to evaluate the Spearman partial rank correlation
between the average surface brightness and age, with the effect of total mass of
the galaxy factored out. The masses of our galaxies were determined by a
prescription given in \citet{Cook2014} from the M/L ratios at 3.6
$\mu$m. We find that the partial correlation coefficient is
$-0.601$, which is not significantly different from the Spearman rank
correlation coefficient between the two variables, which is $-0.757$. Similarly,
the partial correlation coefficient between the average surface brightness and
metallicity is $-0.581$, which is comparable to the correlation coefficient if
$-0.737$. This seems to suggest that the correlations are not completely driven
by the total mass of the galaxy. Note that these correlations hold even if we
use the corresponding mass weighted age and metallicity, albeit with a very minor
change in the correlation strengths.
Again, we would like to point out that these results are not statistically
robust owing to the small number of data points. But the presence of such
correlations suggests the need for a study which aims at obtaining both
photometric and spectroscopic parameters for a larger sample of galaxies.
\section{Discussion}
\label{sec:discussion}
We have seen above that the barred galaxies in our sample have largely similar
stellar population properties and so is the case with the unbarred galaxies
taken as a group but there are important differences between these
groups.
\subsection{The unbarred group}
\label{subsec:unbarred_group}
Recall that the parent sample of this study is the sample of 25 pseudobulge
hosting S0 galaxies, as described in \citet{Vaghmare2013}. The authors
classified pseudobulges using a very conservative criterion. Typically, studies
aimed at bulge classification adopt a very simple criterion where all bulges
with S\'{e}rsic index $n<2$ are classified as pseudobulges and others as
classical. But as \citet{Vaghmare2013} argue, the parameter $n$ is not easy to
constrain and thus there is scope for misclassification. \citet{Gadotti2009}
proposed an alternate criterion based on the Kormendy diagram and argued on the
merits of such an approach. \citet{Vaghmare2013} adopted a combination of the
two criteria to ensure that a secure classification is obtained. The chances of
misclassifying a galaxy as a pseudobulge host is thus quite low.
\citet{Vaghmare2015} suggested that the pseudobulge hosting S0s have a
lower disk luminosity with respect to their spiral counterparts, which
the authors interpret as a disk that has faded while undergoing a
transformation in morphology from spiral to S0. The authors
argue that this transformation involves stripping of gas which is
needed for sustaining active star formation. The signs of active star
formation in these galaxies, as indicated by their spectra is in
contradiction with this conclusion. Further to confirm if the gas
stripping process is a viable scenario, we investigated the
environment of these unbarred galaxies. We found that excluding
IC\,2085, the galaxies in the unbarred group are in an isolated
environment.
One possible resolution is that these isolated galaxies are not
gas stripped spirals. In this case, we have to conclude that these
disks are inherently less luminous and somehow such disks have an
affinity to be a part of pseudobulge hosting S0 galaxies. Another
explanation is that these objects are gas stripped spirals but have
had their gas replenished through numerous gas rich minor mergers
which in turn led to complex star formation \citep{Penoyre2017}.
\subsection{The barred group}
\label{subsec:barred_group}
An interesting commonality in the two galaxies in this group, apart from the
stellar population, is that these galaxies are barred. With the usual caveat
of lack of robustness due to small numbers, this common observation allows for
some speculation on why, despite such a conservative classification criterion, the
bulges of these galaxies are found not to be consistent with pseudobulges.
Recall that the photometric parameters of the bulges as derived by performing
2-d bulge-disk decomposition, have been used for classifying them. One may thus
argue that the presence of a bar has a systematic effect on the derived bulge
properties \citep{Laurikainen2005, Gadotti2009}. But \citet{Vaghmare2013}
recognize this problem and take care to fit a bar component as well. Thus
the unaccounting of a bar component cannot explain this
observation.
Another explanation for the old stellar population in these galaxies is as
follows. The presence of the bar led to an amplified funnelling of gas
towards the centre of the galaxies in the past. This allowed the pseudobulge
to form much earlier than the rest of the galaxies where there was no bar
component to aid the quick formation of a pseudobulge. As a result the
pseudobulges in both these galaxies contain a very old population as can be seen
in the Figure \ref{fig:sfh_all}. In the case of NGC 1533, there does not seem to
have been any gas available to lead to recent star formation while in the
case of NGC 5750, there is some evidence of recent star formation.
A confirmation of whether the bar action can explain the older age could be
obtained from a high spatial resolution rotation curve for these galaxies. The
relative domination of a random component of stellar velocities vs systematic
rotation may be able to shed light on whether there is funnelling of gas taking
place. However, such a study is not possible using the present data.
\section{Stellar Populations in Pseudobulge Hosting S0s - Barred vs Unbarred}
\label{sec:bar_dn4000}
In section \ref{subsec:barred_group}, we saw that the only pseudobulge hosting S0 galaxies
which exhibited a significant departure from the overall nature of
star formation history observed among other galaxies were
barred. However, with such low numbers, it is not appropriate to firmly conclude
that the nature of pseudobulges in S0 galaxies is a function of the
presence or absence of a bar. But it is reasonable to expect a
difference since we know that the bar does affect the nature of
stellar orbits and gas dynamics in a galaxy. Further, a bar is known
to be a driver of growth in bulges in the secular evolution scenario.
We thus decided to perform a quick analysis to see if such a
difference between pseudobulges in barred S0s and unbarred S0s, is indeed
real. \citet{Mishra2017} have presented evidence for the bimodal
stellar age distribution of pseudo-bulges of unbarred S0 galaxies as
probed by the $D_n(4000)$ index. This index is a reliable indicator of
the mean age of galaxy stellar population which is quantified using
the strength of the 4000 \AA{} spectral break arising due to the
accumulation of absorption lines of metals in the atmosphere of old,
low-mass stars in galaxies. The light-weighted mean stellar ages of
$\sim$2 Gyr corresponds to $D_n(4000)$ index of 1.5 \citep{Kauffmann2003}
and using this \citet{Mishra2017} have divided the sample of
unbarred pseudo-bulges in S0 galaxies with a young (having $D_n(4000) <
1.5$) and old (having $D_n(4000) \ge 1.5$) stellar populations. In
their study, Mishra et al. (2017) did not use S0 galaxies that host a
bar. In this work, we use the sample and data described in \citet{Mishra2017}
but for pseudo-bulges in S0 galaxies that host a bar.
In Figure \ref{fig:barwise_dn4000}, we show the distribution of the $D_n(4000)$ index for all
the barred S0s with pseudobulge from the sample in
\citet{Mishra2017}. This figure reveals
that the barred galaxies preferentially have $D_n(4000) \ge 1.5$ ( older stellar
populations), which is consistent with our current study. The figure also shows the $D_n(4000)$ index
for the eight galaxies in our sample. Our current spectroscopic analysis seems to suggest that the bimodality
presented in \citet{Mishra2017} could be caused by the presence of a
bar. This, in turn, suggests that pseudobulges, when found in barred
S0s exhibit a dominant older stellar population
while an unbarred pseudobulge hosting S0 can exhibit significant contribution from
both young and old populations. Our current spectroscopic sample,
however, does not have pseudobulge hosting S0s which have a
very old population and do not contain a bar. Commenting on the exact nature of
such differences in S0s with pseudobulge however, is rather difficult.
\begin{figure}
\centering
\includegraphics[width=0.8\columnwidth]{FinalDn4000.png}
\caption{A plot showing the distribution of $D_n(4000)$ index for
barred pseudobulge hosting S0 galaxies from the sample used in
\citet{Mishra2017}. The blue vertical lines represent
the $D_n(4000)$ index for unbarred galaxies in our sample while the red
vertical lines represent the barred galaxies.}
\label{fig:barwise_dn4000}
\end{figure}
\section{Summary}
\label{sec:summary}
In this paper, we have studied eight S0 galaxies established by
\citet{Vaghmare2015} to host pseudobulges, using long-slit spectra obtained from
the SALT-RSS. We have modelled the spectra using \textsc{starlight} to derive detailed
star formation histories of these galaxies as well as age and metallicity
gradients. The present study focusses primarily on the spectra of the central
regions of the galaxy and, using the spectral information, confronts the conclusions drawn by \citet{Vaghmare2015}.
The most important conclusion is that these eight galaxies are not all similar
in terms of their star formation histories. This
means that objects grouped in one class using photometric criteria
(the class here being pseudobulge hosting S0s) can exhibit diverse properties.
The present study brings out the need for further studies that are based on
spectral analysis to truly constrain the formation mechanisms at work in
the formation of these objects.
We discovered that six of these eight galaxies, which are unbarred, exhibit a
complex star formation history with active star formation. We attempt
reconciling the observed ongoing star formation with the conclusion reached by
\citet{Vaghmare2015} that these are gas stripped spirals. It has been suggested
by \citet{Poggianti2017} that the process of ram pressure stripping in galaxies
can cause initial fuelling of gas inwards, which could lead to star formation in
the central regions. Once the gas is stripped, the disk can fade, but residual
gas in the central region could be responsible for the observed star formation.
This would make fading due to stripping consistent with our observations. But
since most of the galaxies are in an isolated environment the question would
remain as to what caused the stripping.
In the other two galaxies, which are barred, we find a population similar to
those expected in elliptical galaxies. We have detected a trace of recent star
formation in one of these galaxies which likely indicates that the bar is
funnelling / has funnelled gas towards the center leading to a formation of a
younger group of stars.
We have also used a
sample of SDSS galaxies to show that the preferential occurrence of older populations
in barred S0 galaxies, suggested by the spectral data, is likely a real
phenomenon. The study of this SDSS sample also suggests that there are
unbarred pseudobulge hosting S0s with very old populations \citep{Mishra2017}.
However, no such object
has been found in the present study.
Using the data obtained from the SALT-RSS, it is possible to perform a very
detailed analysis of the star formation history of the off-center regions of the
galaxy. This can lead to new insights into how galaxies formed.
We defer such a detailed study to a future paper.
\section*{ACKNOWLEDGEMENTS}
We thank the anonymous referee for insightful comments that
have improved both the content and presentation of this paper.
All of the observations reported in this paper were obtained with the
Southern African Large Telescope (SALT) under program 2014-1-IUCAA-RSA-OTH-001 (PI:
A.K. Kembhavi). KV would like to acknowledge Alexei Kniazev, Steve Crawford and R
Srianand for useful discussions and guidance. KV also acknowledges
the Council of Scientific and Industrial Research (CSIR), India for
financial assistance and SAAO \& SALT for travel and hospitality
support. Part of this work for KV was possible due to a grant from the National
Knowledge Network (NKN), India. SB acknowledges SAAO where part of this work
has been done. PV acknowledges support from the SA National Research Foundation.
YW thanks IUCAA for hosting him on his sabbatical where a part of this work was done. We acknowledge support from a South African National
Research Foundation grant (PID-93727) and from a bilateral grant under
the Indo-South Africa Science and Technology Cooperation (PID-102296)
funded by the Department of Science and Technology (DST) of the Indian
and South African governments. PyRAF is a product of the Space Telescope Science
Institute, which is operated by AURA for NASA and was extensively utilised for
the spectral data reduction pipelines.
\bibliographystyle{mnras}
|
\section{Introduction} \label{sec:introduction}
Revealing cosmic star formation history is one of the biggest challenges in astronomy.
Because a significant fraction of star formation is obscured by dust at high redshift (e.g., \cite{mada14}, for a review), infrared (IR)--submillimeter/millimeter (submm/mm) observations are required to understand the true star-forming activity.
The intensity of the extragalactic background light (EBL) in the IR--submm/mm is known to be comparable to that of the EBL in the optical, also showing the importance of IR--submm/mm observations for revealing the dust-obscured activity in the Universe.
Deep surveys at submm/mm (850~$\mu$m and 1~mm wavelengths) with ground-based telescopes uncovered a population of bright ($S_{\rm 1mm} \gtrsim 1$ mJy) submm/mm galaxies (SMGs; \cite{blai02, case14}, for reviews).
SMGs are highly obscured by dust, and the resulting thermal dust emission dominates the bolometric luminosity.
The energy source of submm/mm emission is primarily from intense star formation activity, with IR luminosities of $L_{\rm IR} \gtrsim$ a few $\times 10^{12}$~$L_{\odot}$ and star formation rates of SFRs $\gtrsim$ a few $\times 100 M_{\odot}$~yr$^{-1}$.
The redshift distribution of SMGs is characterized by a median redshift of $z \sim 2$--3 (e.g., \cite{chap05, yun12, simp14, chen16, mich17, bris17}).
The stellar masses and SFRs of SMGs show that they are located above or at the massive end of the main sequence of star-forming galaxies (e.g., \cite{dadd07, mich12, mich14, dacu15}).
It is thought that SMGs are progenitors of massive elliptical galaxies in the present-day Universe observed during their formation phase (e.g., \cite{lill96, smai04}).
The contribution of SMGs to the EBL is estimated by integrating the number counts.
Blank field surveys with single-dish telescopes resolved $\sim$20\%--40\% of the EBL at 850~$\mu$m (e.g., \cite{barg99, eale00, bory03, copp06}) and $\sim$10\%--20\% at 1 mm (e.g., \cite{grev04, pere08, scot08, scot10, hats11}).
It is expected that deeper submm/mm observations trace less dust-obscured star-forming galaxies, which may overlap galaxies detected in rest-frame ultraviolet (UV) and optical wavelengths.
\citet{whit17} found a dependence of the fraction of obscured star formation (SFR$_{\rm IR}$) on stellar mass out to $z = 2.5$:
50\% of star formation is obscured for galaxies with $\log(M/M_{\odot}) = 9.4$, and $>$90\% for galaxies with $\log(M/M_{\odot}) > 10.5$.
Deep surveys probing fainter submm objects ($S_{\rm 1mm} < 1$~mJy), which are expected to be more normal star-forming galaxies rather than ``classical'' SMGs, are essential to understand the cosmic star-formation history and the origin of EBL, however, such observations have been hampered by the confusion limit of observations with single-dish telescopes since they have large beam sizes ($\sim$$15''$--$30''$).
Interferometric observations enable us to reveal faint submm sources by substantially reducing the confusion limit.
The Atacama Large Millimeter/submillimeter Array (ALMA) is now detecting submm sources more than an order of magnitude fainter than ``classical'' SMGs.
Because of its high sensitivity and high angular resolution, ALMA can collect serendipitous sources from a variety of data sets to probe the fainter end of the number counts (\cite{hats13, ono14, carn15, fuji16, oteo16}).
These studies show that more than 50\% of the EBL at 1 mm is resolved into discrete sources at a flux limit of $\sim$0.1~mJy.
These studies are based on serendipitous sources detected in fields where faint submm sources are not the main targets, which could introduce biases due to the clustering of sources around the targets or sidelobes caused by bright targets.
It is necessary to conduct ``unbiased'' surveys in a contiguous field rather than collecting discrete fields in order to obtain a census on the population of faint submm sources.
Surveys in a contiguous field are also beneficial for clustering analysis.
During ALMA Cycle 1, the central 2 arcmin$^2$ area of the Subaru/XMM-Newton Deep Survey Field (SXDF) was observed as an ALMA deep blank field survey \citep{kohn16, tada15, hats16, wang16, yama16}.
From Cycle 1 to present, the GOODS-S/Hubble Ultra Deep Field (HUDF) has been observed with ALMA in different surveys \citep{walt16, arav16, dunl17, fran18}.
There are also deep surveys in overdense regions such as the ALMA deep field in the $z = 3.09$ protocluster SSA~22 field (ADF22; \cite{umeh15, umeh17, umeh18}) and the ALMA Frontier Fields Survey of gravitational lensing clusters \citep{gonz17}.
The GOODS-S/HUDF field has the deepest multi-wavelengths data from X-ray to radio with ground-based telescopes and satellites such as
{\sl Chandra} \citep{xue11, luo17},
{\sl XMM-Newton} \citep{coma11},
{\sl HST}/ACS/WFC3 (HUDF, CANDELS, XDF; \cite{beck06, grog11, koek11, elli13, illi13}),
VLT/HAWK-I (HUGS; \cite{font14}),
Magellan/FourStar (ZFOURGE; \cite{stra16}),
{\sl Spitzer} (S-CANDELS; \cite{ashb15}),
{\sl Herschel}/PACS (PEP; \cite{lutz11}) and SPIRE (HerMES; \cite{oliv12}),
APEX/LABOCA (LESS; \cite{weis09}),
ASTE/AzTEC \citep{scot10, yun12},
SCUBA-2/JCMT \citep{cowi17},
and VLA \citep{mill13, rujo16}.
Spectroscopic observations have also been conducted extensively (e.g., \cite{lefe04, bram12, skel14}).
The VLT/MUSE spectroscopic survey of HUDF (the $3' \times 3'$ deep region region and $1' \times 1'$ ultra-deep region) provides 3-D data cubes of this field \citep{baco15, baco17}.
JWST will conduct deep multi-band imaging and spectroscopy, offering the ability to diagnose optically-faint galaxies which are difficult to study with existing optical/near-IR telescopes.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig1.eps}
\end{center}
\caption{
ASAGAO region consisting of nine sub-regions (red) overlaid on the {\sl HST}/WFC3 F160W image.
The orange, purple, and green regions represent the ALMA survey areas of ASPECS \citep{walt16, arav16} at 1.2~mm, HUDF \citep{dunl17} at 1.3~mm, and GOODS-ALMA \citep{fran18} at 1.1~mm, respectively.
}
\label{fig:region}
\end{figure}
The ALMA surveys of the GOODS-S field have been conducted with different survey strategies:
a deep but narrow survey (4.5~arcmin$^2$, $1\sigma = 34$~$\mu$Jy~beam$^{-1}$) at 1.3 mm (HUDF; \cite{dunl17}),
a shallower and wider survey (69~arcmin$^2$, $1\sigma \sim 180$~$\mu$Jy~beam$^{-1}$) at 1.1 mm (GOODS-ALMA; \cite{fran18}),
and spectral scans in an area of 1~arcmin$^2$ (ALMA Spectroscopic Survey; ASPECS) at 3 mm and 1.2 mm \citep{walt16, arav16} (figure~\ref{fig:region}).
The spectral scans cover the full window of the bands, offering the deepest continuum maps ($1\sigma_{\rm 3mm} = 3.8$~$\mu$Jy~beam$^{-1}$ and $1\sigma_{\rm 1.2mm} = 12.7$~$\mu$Jy~beam$^{-1}$).
The faint submm sources detected in these studies are found to be on the main sequence, but located at higher stellar mass and SFR ranges (e.g., \cite{hats15, yama16, arav16, dunl17}) due to the survey detection limit.
In addition, the numbers of sources studied in these surveys are still very limited,
and the demand for deeper and wider surveys remains high.
In this paper, we present the results of ALMA twenty-six arcmin$^2$ survey of GOODS-S at one-millimeter (ASAGAO).
ASAGAO is a deep ($1\sigma \sim 61$~$\mu$Jy~beam$^{-1}$ for a 250 k$\lambda$-tapered map) and wide-area (26~arcmin$^2$) survey on a contiguous field at 1.2~mm.
The observing area matches the deepest VLA C-band 5 cm (6 GHz) observations (\cite{rujo16}; Rujopakarn et al. in prep.) and the ultra-deep VLT/HAWK-I $K_S$-band images.
The primary goal of this survey is to obtain a census of galaxies with $L_{\rm IR} \gtrsim 3 \times 10^{11}$~$L_{\odot}$ or SFR $\gtrsim 50$~$M_{\odot}$~yr$^{-1}$ for the understanding of the dust-obscured star-formation history of the Universe.
The initial results based on the ASAGAO data have been reported by \citet{ueda18} for the X-ray active galactic nucleus (AGN) properties, and by \citet{fuji18} for morphological studies.
The results of the multi-wavelength analysis are discussed in \citet{yama18}, and the clustering analysis is conducted by Yoshimura et al. (in prep.).
The arrangement of this paper is as follows.
Section~\ref{sec:data} outlines the ALMA observations, data reduction, and archival data used in this study, and shows the obtained images.
Section~\ref{sec:source} describes the detected sources, and we list the source catalog.
In Section~\ref{sec:counts}, we describe the method of creating number counts, and compare with previous studies.
We present the method of constructing luminosity functions and compare with previous studies in Section~\ref{sec:lf}.
The conclusions are presented in Section~\ref{sec:conclusions}.
Throughout the paper, we adopt a cosmology with $H_0=70$ km s$^{-1}$ Mpc$^{-1}$, $\Omega_{\rm{M}}=0.3$, and $\Omega_{\Lambda}=0.7$, and a \citet{chab03} IMF.
All magnitudes are given in the AB system.
\section{Observations and Data Reduction} \label{sec:data}
\begin{table}
\tbl{ALMA observations.}{
\begin{tabular}{ccccc}
\hline
Date & Tuning & Sub-region& $N_{\rm ant}$ & Baseline (max) \\
& & & & (m) \\
\hline
2016-09-02 & 2 & NW & 39, 45 & 1808.012, 2732.660 \\
2016-09-03 & 2 & NE & 41 & 1770.782 \\
2016-09-06 & 2 & NE & 39 & 2483.450 \\
2016-09-07 & 1 & N & 39 & 2483.450 \\
2016-09-08 & 2 & SW & 39 & 2483.450 \\
2016-09-12 & 2 & SE & 38 & 3143.756 \\
2016-09-14 & 2 & SE & 38 & 3247.644 \\
2016-09-18 & 1, 2 & NW, W & 38 & 2483.451 \\
2016-09-19 & 2 & W & 40 & 3143.756 \\
2016-09-20 & 1, 2 & E & 39 & 3143.756 \\
2016-09-21 & 1, 2 & E, SW, S & 39 & 3143.756 \\
2016-09-22 & 1, 2 & SW, S & 39 & 3143.756 \\
2016-09-24 & 2 & N, C & 39 & 3143.756 \\
2016-09-25 & 1, 2 & C, NE & 39 & 3143.756 \\
2016-09-26 & 1 & NE, C & 40 & 3247.644 \\
2016-09-27 & 1 & W, C & 43 & 3247.644 \\
2016-09-28 & 1 & W, S, SE & 40 & 3143.756 \\
2016-09-29 & 1 & SE & 39 & 3247.644 \\
\hline
\end{tabular}}\label{tab:observations}
\end{table}
\begin{table}[t]
\tbl{Center frequencies of spectral windows used in the surveys of ASAGAO, HUDF \citep{dunl17}, and GOODS-ALMA \citep{fran18}.}{
\begin{tabular}{ccccc}
\hline
spw ID & \multicolumn{2}{c}{ASAGAO} & HUDF & GOODS-ALMA \\
& tuning 1 & tuning 2 & & \\
& (GHz) & (GHz) & (GHz) & (GHz) \\
\hline
0 & 254.12 & 245.12 & 212.2 & 255.9 \\
1 & 256.00 & 247.00 & 214.2 & 257.9 \\
2 & 269.12 & 260.12 & 228.2 & 271.9 \\
3 & 271.00 & 262.00 & 230.2 & 273.9 \\
\hline
\end{tabular}}\label{tab:tunings}
\end{table}
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig2.eps}
\end{center}
\caption{
Frequency setups of ASAGAO tuning 1 (red), tuning 2 (blue), HUDF (purple), and GOODS-ALMA (green).
Solid line represents the atmospheric transmission at the ALMA site for a precipitable water vapor of 1 mm calculated using the Atmospheric Transmission at Microwaves code (ATM; \cite{pard01})\footnotemark (left axis).
The dashed line shows the modified black body spectrum with a dust emissivity index of $\beta = 1.5$, a dust temperature of 35~K, and $z = 2$, scaled to a flux density at 243~GHz of 1 mJy (right axis).
}
\label{fig:tunings}
\end{figure}
\footnotetext{https://almascience.eso.org/about-alma/atmosphere-model}
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig3.eps}
\end{center}
\caption{
Signal-to-noise ratio map with a 250~k$\lambda$ taper (left)
and the primary beam coverage map (right) based on the original ASAGAO data.
}
\label{fig:asagao_map}
\end{figure*}
\subsection{Observations} \label{subsec:observations}
ALMA band 6 observations of the GOODS-S field were conducted in September 02--29, 2016 for the Cycle 3 program (Project code: 2015.1.00098.S, PI: K. Kohno) as summarized in table~\ref{tab:observations}.
The $\sim$$5' \times 5'$ survey area centered at (R.A., Dec.) = (\timeform{03h31m38.601s}, \timeform{-27D46'59.830''}) consists of 9 tiles (figure~\ref{fig:region}) and each tile was covered by $\sim$90-pointing mosaic observations with Nyquist sampling.
Two frequency tunings were adopted to cover a wider frequency range, providing a larger survey volume for searching serendipitous line emitting galaxies.
The center frequencies of the tunings are 262.56~GHz (1.14~mm) and 253.56~GHz (1.18~mm), which were selected to avoid strong atmospheric absorption lines (table~\ref{tab:tunings} and figure~\ref{fig:tunings}).
The correlator was used in the time domain mode (TDM).
Four basebands were used for each tuning, and a spectral window (spw) was placed for each baseband with a bandwidth of 2000 MHz (15.625~MHz $\times$ 128 channels), providing a total nominal bandwidth of 16 GHz (effective bandwidth of 15 GHz) centered at 258.6~GHz (1.16~mm).
The observations were done in 37 execution blocks in the C40-6 array configuration (maximum recoverable scale of $\theta_{\rm MRS} \approx 1\farcs2$) with a minimum baseline length of 15.065~m and a maximum baseline length ranging from 1770 m to 3247 m.
The number of available antenna was 38--45.
The total observing time is 45 hours, and the on-source integration time is 29 hours.
The bandpass was calibrated with quasars J0522$-$3627, J0238+1636, and J0334$-$4008, and the phase was calibrated with J0348$-$2749.
J0334$-$4008 and J2357$-$5311 were observed as flux calibrators.
\subsection{Data Reduction} \label{subsec:reduction}
To reduce the data volume for easier handling in continuum imaging, we average the data in frequency and time directions with 32 channels ($\Delta \nu = 0.5$~GHz) and 10.08~sec, respectively.
The effect of bandwidth smearing on the peak flux density of a source caused by the channel averaging is less than 1\% even at the edge of the primary beam \citep{cond98}.
We also confirm that the effect of the time averaging on the flux density is negligible based on the imaging of the bandpass calibrator.
The data were reduced with Common Astronomy Software Applications (CASA; \cite{mcmu07}).
Data calibration was done with the ALMA Science Pipeline Software of CASA version 4.7.2.
The maps were processed by the task {\verb tclean } of CASA version 5.1.1 with natural weighting, a cell size of 0.1 arcsec, a gridding option of standard, the spectral definition mode of multi-frequency synthesis, the number of Taylor coefficients in the spectral model of 2 for a spectrum with a slope, and a primary beam limit of 0.2 (default value).
Clean boxes are placed when a component with a peak signal-to-noise ratio (SN) above 5 is identified, and {\verb CLEAN }ed down to a $2\sigma$ level.
Because the observations were done with a higher angular resolution ($\sim$$0.2''$) than requested because of the restriction of array configuration, we adopt a $uv$-taper of 250~k$\lambda$ to weight extended components, which gives a synthesized beam size of $0\farcs51 \times 0\farcs45$.
The signal-to-noise ratio map and the primary beam coverage map are shown in figure~\ref{fig:asagao_map}.
In this study, we use the region where the primary beam coverage is larger than or equal to 0.2 in the map, which is a 26-arcmin$^2$ area.
A sensitivity map was created by using the {\sc BANE} program \citep{hanc12}, which performs $3\sigma$ clipping in the signal map and calculate the standard deviation on a sparse grid of pixels and then interpolate to make a noise image.
Figure~\ref{fig:pix_hist} shows the histograms of flux density of the signal map (before primary beam correction).
The pixel-flux distribution is well explained by a Gaussian curve, and a Gaussian fit gives $1\sigma$ of 61~$\mu$Jy~beam$^{-1}$.
The excess from the fitted Gaussian at $\gtrsim$0.3~mJy indicates the contribution from real sources.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig4.eps}
\end{center}
\caption{
Distribution of flux density of the signal map based on the original ASAGAO data (uncorrected for primary beam attenuation).
The dashed curve shows the result of a Gaussian fit ($1\sigma = 61$~$\mu$Jy~beam$^{-1}$).
}
\label{fig:pix_hist}
\end{figure}
\subsection{ALMA Archival Data}
In addition to our data, we also use the ALMA archival data of 1-mm (band 6) surveys in the GOODS-S field of HUDF \citep{dunl17} and GOODS-ALMA \citep{fran18}.
We do not use the data set of ASPECS, where the synthesized beam size ($1\farcs68 \times 0\farcs92$) is largely different from those of the others ($\lesssim 0.5''$).
The ALMA survey of HUDF by \citet{dunl17} covered a 4.5 arcmin$^2$ area at 1.3~mm during Cycle 1 and 2 (Project code: 2012.1.00173.S, PI: J. Dunlop).
The correlator was configured with four spectral windows with a 2000 MHz bandwidth (15.625 MHz $\times$ 128 channels).
The synthesized beam with natural weighting is $0\farcs59 \times 0\farcs50$.
An $uv$-tapering of $\simeq$$220 \times 180$~k$\lambda$ they adopted gives a final synthesized beam of $0\farcs71 \times 0\farcs67$ and a noise level of 34~$\mu$Jy~beam$^{-1}$.
A wider area of 69 arcmin$^2$ ($\sim$$10' \times 7'$) was observed in the GOODS-ALMA survey \citep{fran18} at 1.13~mm during Cycle 3 (Project code: 2015.1.00543.S, PI: D. Elbaz).
The survey consists of six sub-mosaics, encompassing the survey fields of ASAGAO, HUDF, and ASPECS.
The correlator was set to have four spectral windows with 15.625 MHz $\times$ 128 channels.
The synthesized beam with natural weighting is $\sim$0\farcs20--0\farcs29 depending on the sub-regions.
The rms noise level is $\sim$180~$\mu$Jy~beam$^{-1}$ and $\sim$110~$\mu$Jy~beam$^{-1}$ for the tapered map with a synthesized beam of 0\farcs6 and for the untapered map, respectively.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig5.eps}
\end{center}
\caption{
$uv$-plane coverage of the combined data.
}
\label{fig:uv}
\end{figure}
\begin{figure*}
\begin{center}
\includegraphics[width=.8\linewidth]{fig6.eps}\\
\end{center}
\caption{
The combined signal map (ASAGAO + HUDF + GOODS-ALMA) with a 250 k$\lambda$ taper (corrected for primary beam attenuation).
The squares represent the detected sources.
}
\label{fig:combined_sigmap}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig7.eps}
\end{center}
\caption{
The primary beam coverage map (left) and the rms noise map (right) for the combined data.
The rms noise map is corrected for primary beam attenuation.
}
\label{fig:combined_covmap}
\end{figure*}
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig8.eps}
\end{center}
\caption{
Cumulative area of the combined map as a function of rms noise level (corrected for primary beam attenuation) for the ASAGAO only data (dashed) and the combined data (solid).
}
\label{fig:effective_area}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig9.eps}
\end{center}
\caption{
\textit{Top}: Distribution of flux density of the signal map based on the combined data (uncorrected for primary beam attenuation).
The dashed curve shows the result of a Gaussian fit ($1\sigma = 34$~$\mu$Jy~beam$^{-1}$).
\textit{Bottom}: Ratio between the flux density distribution and the result of a Gaussian fit.
}
\label{fig:pix_hist_combined}
\end{figure}
\subsection{Combined Map} \label{subsec:combined}
The archival data sets of HUDF \citep{dunl17} and GOODS-ALMA \citep{fran18} are combined to the original ASAGAO data to make a deeper map with the total effective frequency coverage of $\sim$27 GHz (table~\ref{fig:tunings} and figure~\ref{tab:tunings}).
Before combining the data sets, we relabel the coordinates of Cycle 1 and 2 data from J2000.0 to the International Celestial Reference System (ICRS) by using a CASA script offered by the ALMA project, because the position reference frame in ALMA $uv$ data and images is given as J2000 before Cycle 3 and as ICRS from Cycle 3.
The $uv$ data sets are averaged in frequency and time directions (32 channels and 10.08~sec) in the same manner as the original ASAGAO data.
Figure~\ref{fig:uv} shows the $uv$-plane coverage of the combined data.
The combined map was produced with CASA with the same parameters adopted in Sec~\ref{subsec:reduction}.
The representative frequency of the map is 243.047 GHz (1.23 mm).
We adopt an $uv$-taper of 250~k$\lambda$ to weight extended components, which gives a final synthesized beam size of $0\farcs59 \times 0\farcs53$.
Maps without $uv$-taper (synthesized beam size of $0\farcs30 \times 0\farcs24$) and with a $uv$-taper of 160~k$\lambda$ ($0\farcs83 \times 0\farcs72$) were also created to see whether detected sources are spatially resolved.
The signal map, the coverage map, and the rms noise map (corrected for primary beam attenuation) with a 250~k$\lambda$ taper are shown in figure~\ref{fig:combined_sigmap} and \ref{fig:combined_covmap}.
We use the same region as adopted in the original ASAGAO map (Sec.~\ref{subsec:reduction}).
The map has two layers, the central deeper area (the deepest region has $1\sigma \sim 26$~$\mu$Jy~beam$^{-1}$) and the rest, as can be seen in figure~\ref{fig:combined_covmap} and figure~\ref{fig:effective_area} of the cumulative area as a function of rms noise level.
Figure~\ref{fig:pix_hist_combined} shows the histogram of flux density of the signal map (before primary beam correction).
The dashed curve represents the result of a Gaussian fit, which gives $1\sigma$ of 34~$\mu$Jy~beam$^{-1}$.
The presence of real sources in the map makes excess of positive pixels.
This fit also deviates from the distribution of pixel values at high negative flux densities, which can be explained by the non-uniform noise distribution of the entire map.
\section{Source Catalog} \label{sec:source}
\subsection{Source Detection} \label{subsec:detection}
Source detection is conducted on the signal map before correcting for the primary beam attenuation.
We adopt the source-finding algorithm called {\sc Aegean} \citep{hanc12, hanc18}, which achieves high reliability and completeness performance for radio maps.
The background and noise estimation are done with the {\sc BANE} package in the same manner as described in Sec.~\ref{subsec:reduction}.
We find 25 (45) sources with a peak SN of $\ge$5$\sigma$ ($\ge$4.5$\sigma$).
The detected sources are fitted with a 2D elliptical Gaussian to estimate the source size and integrated flux density.
The integrated flux density ($S_{\rm int}$) is calculated as
\begin{eqnarray}
S_{\rm int} = S_{\rm peak} \frac{a b}{\theta_{\rm maj}\theta_{\rm min}},
\end{eqnarray}
where $S_{\rm peak}$ is the peak flux density, $a/b$ are the fitted major/minor axes, and $\theta_{\rm maj}/\theta_{\rm min}$ are the synthesized beam major/minor axes.
We adopt $S_{\rm int}$ as the source flux density.
When $S_{\rm int} < S_{\rm peak}$, we adopt $S_{\rm peak}$, since it is possible that the source fitting failed due to the low SN.
The source catalog for the 4.5$\sigma$ sources extracted in the combined signal map with a 250 k$\lambda$ taper is presented in table~\ref{tab:source}.
Hereafter we refer to these sources as ASAGAO sources, and adopt the integrated flux densities measured in the 250 k$\lambda$ tapered map.
The range of continuum flux densities is 0.16--2 mJy (after correcting for primary beam attenuation).
The integrated flux densities in the untapered map ($S_{\rm int}^{\rm untaper}$) and in the map with a 160 k$\lambda$ taper ($S_{\rm int}^{\rm 160k\lambda}$) measured in the same manner as in the 250 k$\lambda$ tapered map are also shown.
When a source is not detected with a peak SN $>$ 3 in these maps, the flux density is not listed in the source catalog.
ASAGAO ID31, 36, and 37 are not detected in the untapered map with a peak SN $>$ 3.
This can be due to the lack of sensitivity for spatially extended structures or clumpy structures and multiple peaks as can be seen in the postage-stamp images in figure~\ref{fig:source}, each having a peak SN less than 3.
The median ratio between integrated flux and peak flux is $S_{\rm int}/S_{\rm peak} = 1.3 \pm 0.8$.
The median ratio of integrated flux between 250~k$\lambda$-tapered map and 160~k$\lambda$-tapered map or the untapered map is $S_{\rm int}^{\rm 250k\lambda}/S_{\rm int}^{\rm 160k\lambda} = 0.86 \pm 0.24$, and $S_{\rm int}^{\rm 250k\lambda}/S_{\rm int}^{\rm untaper} = 1.3 \pm 3.0$.
These suggest that sources are resolved by the synthesized beam in the 250 k$\lambda$-tapered and the untapered maps.
In order to estimate the degree of contamination by spurious sources, we count the number of negative peaks as a function of SN threshold (figure~\ref{fig:fdr}).
The number of independent beams in the map is $2.7 \times 10^5$, and the expected number of $\ge$$4.5\sigma$ sources in a Gaussian statistics is $\sim$1.
However, it is reported that this estimation underestimates the negative peaks in previous studies based on ALMA images \citep{dunl17, vio16, vio17}.
The actual number of negative peaks in the combined map is 1 at $\ge$5$\sigma$ and 8 at 4.5--5$\sigma$.
The small number of negative peaks at $\ge$5$\sigma$ suggests the robustness of the 5$\sigma$ sources.
Actually, 22 out of the 25 5$\sigma$ sources (88\%) have counterparts at optical, {\sl Spitzer}/IRAC, radio, or ALMA 850~$\mu$m \citep{cowi18} (see \cite{yama18} for multi-wavelength identifications of ASAGAO sources).
\begin{table*}[t]
\tbl{Source catalog of $\ge$5$\sigma$ sources (ID1--25) and 4.5--5$\sigma$ sources (ID26--45)}{
\footnotesize
\begin{tabular}{ccccccccc}
\hline
ID&R.A.&Dec.&SN&$S_{\rm peak}$&$S_{\rm int}$&$S_{\rm int}^{\rm untaper}$&$S_{\rm int}^{\rm 160k\lambda}$&Note\\
ASAGAO&(J2000)&(J2000)& &($\mu$Jy)&($\mu$Jy)&($\mu$Jy)&($\mu$Jy)& \\
(1)&(2)&(3)&(4)&(5)&(6)&(7)&(8)&(9)\\
\hline
1 & 03:32:44.03 & $-$27:46:35.97 & 26.0 & $ 839 \pm 32$ & $ 990 \pm 36$ & $ 877 \pm 32$ & $1023 \pm 44$ & UDF1, AGS6, U3 \\
2 & 03:32:28.51 & $-$27:46:58.36 & 25.6 & $1851 \pm 72$ & $1983 \pm 75$ & $1996 \pm 57$ & $2251 \pm 112$ & AGS1, U1 \\
3 & 03:32:35.72 & $-$27:49:16.27 & 24.0 & $1656 \pm 69$ & $1758 \pm 70$ & $1816 \pm 54$ & $1709 \pm 100$ & AGS3, U2 \\
4 & 03:32:43.53 & $-$27:46:39.25 & 21.0 & $ 658 \pm 31$ & $ 914 \pm 41$ & $ 761 \pm 40$ & $1019 \pm 50$ & UDF2, AGS18, U6 \\
5 & 03:32:38.55 & $-$27:46:34.61 & 18.1 & $ 554 \pm 31$ & $ 745 \pm 39$ & $ 634 \pm 35$ & $ 791 \pm 45$ & UDF3, ASPECS/C1, AGS12, U8 \\
6 & 03:32:47.59 & $-$27:44:52.43 & 12.4 & $ 768 \pm 62$ & $ 954 \pm 74$ & $ 735 \pm 43$ & $1161 \pm 123$ & U4 \\
7 & 03:32:32.90 & $-$27:45:41.07 & 8.8 & $ 546 \pm 63$ & $ 829 \pm 86$ & $ 593 \pm 65$ & $ 835 \pm 104$ & U5 \\
8 & 03:32:31.48 & $-$27:46:23.50 & 8.7 & $ 576 \pm 66$ & $ 650 \pm 72$ & $ 618 \pm 57$ & $ 705 \pm 103$ & AGS13, U12 \\
9 & 03:32:47.18 & $-$27:45:25.48 & 8.6 & $ 495 \pm 57$ & $ 488 \pm 55$ & $ 945 \pm 106$ & $ 406 \pm 69$ & \\
10 & 03:32:41.02 & $-$27:46:31.59 & 8.6 & $ 255 \pm 30$ & $ 278 \pm 31$ & $ 350 \pm 39$ & $ 246 \pm 32$ & UDF4 \\
11 & 03:32:29.25 & $-$27:45:09.96 & 8.5 & $ 580 \pm 68$ & $ 678 \pm 78$ & $ 587 \pm 51$ & $ 855 \pm 124$ & \\
12 & 03:32:36.96 & $-$27:47:27.14 & 7.4 & $ 227 \pm 31$ & $ 408 \pm 49$ & $ 190 \pm 35$ & $ 484 \pm 61$ & UDF5 \\
13 & 03:32:34.44 & $-$27:46:59.86 & 7.2 & $ 224 \pm 31$ & $ 436 \pm 53$ & $ 227 \pm 34$ & $ 503 \pm 65$ & UDF6 \\
14 & 03:32:43.33 & $-$27:46:46.96 & 7.2 & $ 229 \pm 32$ & $ 259 \pm 35$ & $ 224 \pm 26$ & $ 281 \pm 48$ & UDF7, U7 \\
15 & 03:32:40.07 & $-$27:47:55.72 & 6.6 & $ 197 \pm 30$ & $ 458 \pm 64$ & $ 166 \pm 32$ & $ 490 \pm 69$ & UDF11 \\
16 & 03:32:39.75 & $-$27:46:11.67 & 6.5 & $ 192 \pm 29$ & $ 539 \pm 65$ & $ 106 \pm 22$ & $ 640 \pm 76$ & UDF8, ASPECS/C2 \\
17 & 03:32:49.45 & $-$27:49:09.00 & 6.1 & $ 516 \pm 83$ & $ 564 \pm 90$ & $ 485 \pm 55$ & $1286 \pm 289$ & U11 \\
18 & 03:32:48.57 & $-$27:49:34.62 & 5.8 & $ 749 \pm 130$ & $1091 \pm 172$ & $ 353 \pm 94$ & $1868 \pm 370$ & \\
19 & 03:32:44.61 & $-$27:48:36.13 & 5.7 & $ 375 \pm 67$ & $ 434 \pm 73$ & $ 345 \pm 51$ & $ 431 \pm 99$ & U10 \\
20 & 03:32:28.91 & $-$27:44:31.54 & 5.6 & $ 614 \pm 109$ & $ 653 \pm 110$ & $ 637 \pm 80$ & $ 749 \pm 187$ & \\
21 & 03:32:47.90 & $-$27:44:33.96 & 5.5 & $ 499 \pm 91$ & $1011 \pm 178$ & $ 356 \pm 57$ & $3116 \pm 536$ & \\
22 & 03:32:41.20 & $-$27:49:01.75 & 5.4 & $ 371 \pm 68$ & $ 612 \pm 101$ & $ 187 \pm 43$ & $ 797 \pm 153$ & \\
23 & 03:32:35.09 & $-$27:46:47.82 & 5.4 & $ 163 \pm 30$ & $ 206 \pm 37$ & $ 135 \pm 20$ & $ 202 \pm 44$ & UDF13 \\
24 & 03:32:43.99 & $-$27:45:18.74 & 5.0 & $ 299 \pm 60$ & $ 446 \pm 82$ & $ 65 \pm 120$ & $ 482 \pm 102$ & \\
25 & 03:32:48.24 & $-$27:47:22.14 & 5.0 & $ 385 \pm 77$ & $ 858 \pm 223$ & $ 186 \pm 46$ & $1168 \pm 186$ & \\
\hline
26 & 03:32:43.68 & $-$27:48:51.12 & 4.9 & $ 314 \pm 64$ & $ 254 \pm 52$ & $ 286 \pm 37$ & $ 364 \pm 91$ & \\
27 & 03:32:36.17 & $-$27:46:03.04 & 4.9 & $ 154 \pm 32$ & $ 226 \pm 45$ & $ 170 \pm 43$ & $ 374 \pm 80$ & \\
28 & 03:32:30.41 & $-$27:44:59.97 & 4.9 & $ 348 \pm 72$ & $ 716 \pm 154$ & $ 124 \pm 34$ & $ 888 \pm 184$ & \\
29 & 03:32:38.74 & $-$27:48:40.12 & 4.8 & $ 348 \pm 71$ & $ 227 \pm 46$ & $ 551 \pm 84$ & $ 677 \pm 185$ & \\
30 & 03:32:32.76 & $-$27:49:32.41 & 4.7 & $ 578 \pm 122$ & $ 886 \pm 180$ & $ 452 \pm 123$ & $1157 \pm 253$ & \\
31 & 03:32:34.02 & $-$27:49:00.11 & 4.7 & $ 339 \pm 72$ & $ 846 \pm 158$ & -- & $ 881 \pm 173$ & \\
32 & 03:32:43.68 & $-$27:44:29.66 & 4.7 & $ 461 \pm 98$ & $ 769 \pm 164$ & $ 299 \pm 56$ & $1140 \pm 249$ & \\
33 & 03:32:28.59 & $-$27:48:50.57 & 4.7 & $ 347 \pm 74$ & $ 366 \pm 79$ & $ 283 \pm 42$ & $ 425 \pm 108$ & \\
34 & 03:32:48.60 & $-$27:49:07.95 & 4.6 & $ 298 \pm 65$ & $ 313 \pm 67$ & $ 303 \pm 52$ & $ 328 \pm 80$ & \\
35 & 03:32:38.89 & $-$27:47:35.50 & 4.6 & $ 140 \pm 30$ & $ 180 \pm 39$ & $ 74 \pm 18$ & $ 169 \pm 44$ & \\
36 & 03:32:28.46 & $-$27:46:58.83 & 4.6 & $ 333 \pm 73$ & $ 635 \pm 134$ & -- & -- & \\
37 & 03:32:45.83 & $-$27:46:08.86 & 4.6 & $ 270 \pm 59$ & $ 362 \pm 76$ & -- & $1879 \pm 405$ & \\
38 & 03:32:36.74 & $-$27:44:38.73 & 4.6 & $ 334 \pm 73$ & $ 441 \pm 91$ & $ 175 \pm 50$ & $1496 \pm 309$ & \\
39 & 03:32:32.90 & $-$27:45:39.37 & 4.6 & $ 286 \pm 63$ & $ 529 \pm 107$ & $ 148 \pm 37$ & $ 833 \pm 194$ & \\
40 & 03:32:33.65 & $-$27:46:47.94 & 4.6 & $ 149 \pm 33$ & $ 198 \pm 42$ & $ 304 \pm 66$ & $ 202 \pm 51$ & \\
41 & 03:32:27.72 & $-$27:47:15.17 & 4.6 & $ 459 \pm 99$ & $ 629 \pm 133$ & $ 229 \pm 60$ & $ 805 \pm 188$ & \\
42 & 03:32:50.25 & $-$27:48:21.16 & 4.6 & $ 588 \pm 129$ & $ 621 \pm 137$ & $ 596 \pm 137$ & $ 885 \pm 224$ & \\
43 & 03:32:45.99 & $-$27:47:57.18 & 4.6 & $ 283 \pm 62$ & $ 495 \pm 106$ & $ 149 \pm 42$ & $ 589 \pm 128$ & \\
44 & 03:32:28.84 & $-$27:48:29.72 & 4.5 & $ 350 \pm 77$ & $2051 \pm 447$ & $ 107 \pm 25$ & $1679 \pm 362$ & \\
45 & 03:32:37.83 & $-$27:47:16.49 & 4.5 & $ 131 \pm 29$ & $ 157 \pm 33$ & $ 117 \pm 25$ & $ 128 \pm 33$ & \\
\hline
\end{tabular}}\label{tab:source}
\begin{tabnote}
Notes. -
(1) ASAGAO ID.
(2) Right ascension.
(3) Declination.
(4) Peak signal-to-noise ratio.
(5) Peak flux density (corrected for primary beam attenuation).
(6) Integrated flux density (corrected for primary beam attenuation).
(7) Integrated flux density (corrected for primary beam attenuation) measured in the untapered map when the peak SN is above 3.
(8) Integrated flux density (corrected for primary beam attenuation) measured in the 160-k$\lambda$ tapered map when the peak SN is above 3.
(9) Notes on source IDs of \citet{dunl17} (UDF), \citet{arav16} (ASPECS), \citet{fran18} (AGS), and \citet{ueda18} (U).
\end{tabnote}
\end{table*}
\begin{figure*}
\begin{center}
\includegraphics[width=.8\linewidth]{fig10.eps}
\end{center}
\caption{
Postage-stamp images of the ASAGAO $\ge$4.5$\sigma$ sources with no taper (left), a 250-k$\lambda$ taper (middle), and a 160-k$\lambda$ taper (right).
The image size is $4'' \times 4''$.
Contours are 3$\sigma$, 4$\sigma$, 5$\sigma$, and 5$\sigma$ steps subsequently (negative contours are shown as dashed lines).
The synthesized beam size is shown in the lower left corners of each panel.
\label{fig:source}}
\end{figure*}
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig11.eps}
\end{center}
\caption{
Cumulative number of positive and negative peaks as a function of peak SN threshold.
}
\label{fig:fdr}
\end{figure}
\subsection{Astrometry} \label{subsec:astrometry}
Calibration for astrometry is performed by interpolating the phase information of the phase calibrators over the target fields.
The astrometric accuracy of a source depends on statistical errors determined by the source SN and systematic errors such as the atmospheric phase stability, the proximity of an astrometric calibrator, and baseline errors.
The minimum obtainable astrometric accuracy with no systematic errors is determined by a source SN, observing frequency, and maximum baseline length, which gives $\sim$$0.15''$ for a 5$\sigma$ source with the observing frequency of 243.047 GHz and the maximum baseline of 3.2 km (see ALMA Technical Handbook).
To confirm the astrometry of ASAGAO sources,
the positions of the 5$\sigma$ sources are cross-matched with sources detected in the VLA 5-cm survey (\cite{rujo16}, Rujopakarn et al. in prep.).
The radio sources are more suitable for evaluating the astrometry of the ALMA sources compared to optical sources because
(i) the angular resolution and positional accuracy are comparable to those of the ALMA observations,
and (ii) the positions of submm/mm emission and optical emission, which typically trace dust obscured and unobscured parts, respectively, do not necessarily coincide within a galaxy, and radio observations can trace dust obscured parts.
The radio counterparts are found for 20 out of the 25 ASAGAO 5$\sigma$ sources within a 0\farcs5 search radius, and the positional offset between them is plotted in figure~\ref{fig:offset}.
The median offset is ($\Delta \alpha$, $\Delta \delta$) $= (+0\farcs03 \pm 0\farcs08, -0\farcs01 \pm 0\farcs06)$, which is within the expected positional uncertainty between the ALMA and the radio sources of $\sim$0\farcs1 as the square-root of sum of squares of both uncertainties ($\Delta \alpha = \Delta \delta \simeq 0.6$ (SN)$^{-1}$ FWHM; \cite{ivis07}).
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig12.eps}
\end{center}
\caption{
Positional offsets of the 5$\sigma$ ASAGAO sources from the VLA 5-cm radio sources (\cite{rujo16}, Rujopakarn et al. in prep.).
The errors are the square root of the sum of the squares of expected 1$\sigma$ positional uncertainties of the ASAGAO and VLA sources.
\label{fig:offset}}
\end{figure}
\subsection{Comparison with ALMA 1-mm Sources in GOODS-S}
We cross-matched the ASAGAO sources with the HUDF, GOODS-ALMA, and ASPECS sources (Table~\ref{tab:source}).
\citet{dunl17} listed 16 HUDF sources, and we confirmed that all the eight sources with SN $> 4.5$ out of 16 are detected in our map.
Two additional other sources are detected in our map, and the other 6 sources are not detected due to their lower SNs.
Among the 20 GOODS-ALMA sources presented in \citet{fran18}, we confirmed that all of the six sources inside the ASAGAO region are detected in our map.
A comparison of flux densities of sources common with these surveys shows that the median flux ratios are $S_{\rm 243GHz}^{\rm ASAGAO}/S_{\rm 221GHz}^{\rm HUDF} = 1.15 \pm 0.64$ and $S_{\rm 243GHz}^{\rm ASAGAO}/S_{\rm 265GHz}^{\rm GOODS-ALMA} = 0.89 \pm 0.13$, which are consistent with the flux ratios assuming a modified black body with a dust emissivity index of $\beta = 1.5$, a dust temperature of 35~K, and $z = 2$ ($S_{\rm 243GHz}/S_{\rm 221GHz} = 1.3$ and $S_{\rm 243GHz}/S_{\rm 265GHz} = 0.78$).
The two brightest sources ($S_{\rm 1.2mm} > 0.2$ mJy) of ASPECS, which are the highest SN sources (SN $>$ 10) in their source catalog, are also detected in our map.
The non-detection of lower SN ASPECS sources can be explained by their lower flux densities ($S_{\rm 1.2mm} < 0.15$ mJy).
The ASAGAO 5$\sigma$ sources without counterpart in the other surveys are outside the regions of ASPECS and HUDF, and have lower flux densities than the detection limit of GOODS-ALMA.
\subsection{Comparison with AzTEC Sources} \label{subsec:aztec}
The central 270~arcmin$^2$ area of the GOODS-S field was observed with AzTEC \citep{wils08}, mounted on the Atacama Submillimeter Telescope Experiment (ASTE; \cite{ezaw04, ezaw08}) at 1.1 mm (270 GHz) \citep{scot10}.
The beam size of AzTEC on ASTE is 30$''$ (FWHM).
Two AzTEC sources identified in \citet{scot10} (AzTEC/GS18 and 21) are located inside the ASAGAO region, and detected as multiple sources in our 4.5$\sigma$ source catalog.
AzTEC/GS18 is detected as three ASAGAO sources (ID1, 4, and 14), and the total flux of the three sources is $S_{\rm 1.2mm} = 2.16 \pm 0.06$~mJy, which is consistent with the flux density of the AzTEC source, $S_{\rm 1.1mm} = 3.2 \pm 0.6$~mJy \citep{down12} taking into account the flux ratio between 1.2 mm and 1.1 mm of $S_{\rm 1.2mm}/S_{\rm 1.1mm} \sim 0.73$.
\citet{yun12} studied the radio and {\sl Spitzer} counterparts of the AzTEC/GOODS-S sources.
They found three counterpart candidates for AzTEC/GS18, two of which are detected in the ASAGAO map.
The other is identified in the 1.3 mm source catalog of \citet{dunl17} as a 4.26$\sigma$ source (UDF9).
AzTEC/GS21 has an ASAGAO counterpart (ID6) within $15''$ from the AzTEC source position.
Another source (ID21) is located $\sim$$16''$ away from the AzTEC source position.
ASAGAO ID6 is identified as a radio and {\sl Spitzer} counterpart candidate of \citet{yun12}.
The total flux of the two ALMA sources is $S_{\rm 1.2mm} = 1.97 \pm 0.19$~mJy, which is also consistent with the flux density of the AzTEC source, $S_{\rm 1.1mm} = 2.7 \pm 0.6$~mJy \citep{down12} by considering the expected flux ratio between 1.2 mm and 1.1 mm emission.
\section{Number Counts}\label{sec:counts}
Number counts are constructed by using the 45 4.5$\sigma$ sources.
We correct for the effective area where sources are detected at SN $\ge$ 4.5, contribution of spurious sources, survey completeness, and flux boosting.
In this section, we present the methods of estimating survey completeness and flux boosting (Sec.~\ref{subsec:completeness}), and constructing number counts (Sec.~\ref{subsec:counts}).
Next we compare the obtained number counts with previous studies (Sec.~\ref{subsec:counts_comparison}) and estimate the contribution of the ASAGAO sources to the 1.2~mm EBL (Sec.~\ref{subsec:ebl}).
\subsection{Completeness and Flux Boosting}\label{subsec:completeness}
We calculate the completeness, which is the rate at which a source is expected to be detected in a map, to see the effect of noise fluctuations on the source detection.
The calculation is conducted on the signal map (corrected for primary beam attenuation).
An artificial source of an elliptical Gaussian with the synthesized beam size is injected into a position randomly selected in the map.
In order to take into account the effect of source size, the input source is convolved with another Gaussian function.
\citet{fran18} computed the completeness with different convolving Gaussian FWHM between 0\farcs2 and 0\farcs9, and found that the completeness is lower for a larger FWHM.
Recent ALMA measurements of source size of SMGs ($S_{\rm 1mm} > 1$~mJy) show that source sizes (FWHM) range from $0\farcs08$ to $0\farcs8$ (e.g., \cite{ikar15, simp15, hodg16, ikar17, umeh17}).
The median source sizes in these studies are $0\farcs20^{+0\farcs03}_{-0\farcs05}$ \citep{ikar15}, $0\farcs30 \pm 0\farcs04$ \citep{simp15}, and $0\farcs31 \pm 0\farcs03$ \citep{ikar17}.
\citet{fuji17} find a positive correlation between the effective radius in the rest-frame FIR wavelength and FIR luminosity by using a sample of 1034 ALMA sources, suggesting that the ASAGAO sources which have fainter flux densities ($S_{\rm 1mm} \lesssim 1$~mJy) may have smaller source sizes.
This is proved to be valid for the ASAGAO sources based on $uv$-visibility stacking analysis \citep{fuji18}.
In the completeness calculation, we take a convolving beam size to be uniformly distributed from 0\farcs01--0\farcs5.
We input 30000 artificial sources into the signal map one at a time, each with an integrated flux density randomly selected from 0.05--2~mJy by considering the flux range of detected sources.
The input sources are then extracted in the same manner as in Sec.~\ref{subsec:detection}.
When the input source is detected with a peak SN $\ge 4.5$, the source is considered to be recovered.
The completeness calculation is conducted separately for the central deeper region (coverage $> 0.6$) and the rest (coverage $< 0.6$) to see the effect of the survey depth.
The result is shown in figure~\ref{fig:completeness}.
The completeness calculated in regions with different coverage are consistent within errors and we do not find a significant difference.
The completeness is 60\% at SN $= 4.5$, and 100\% at SN $\gtrsim 7$.
When dealing with low SN sources, we need to consider the effect that flux densities are boosted by noise \citep{murd73, hogg98}.
In the course of the completeness simulation, we calculate the ratio between input and output integrated flux density to estimate the intrinsic flux density of the detected sources (figure~\ref{fig:boosting}, top panel).
The effect of flux boosting for the sources with SN $\ge 4.5$ is on average less than 15\%, and the deboosted flux densities range from 135~$\mu$Jy to 1.97~mJy.
As in the completeness calculation, we do not see any significant difference in the flux boosting for the different coverage regions.
The fraction of output peak SN and input peak SN is also calculated and shown in figure~\ref{fig:boosting} (bottom panel).
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig13.eps}
\end{center}
\caption{
Completeness calculated for the regions with coverage $> 0.6$ (red) and $< 0.6$ (blue) as a function of input peak SN.
The squares and error bars represent mean and 1$\sigma$ from the binomial distribution within a bin obtained by 30000 trials in each coverage region.
The dashed curve show the best-fit function of $f({\rm SN}) = [1 + {\rm erf}(({\rm SN} - a)/b)]/2$ for the entire region, where $(a, b) = (4.33, 1.50)$.
}
\label{fig:completeness}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig14.eps}
\end{center}
\caption{
The ratio between input flux ($S_{\rm in}$) and output flux ($S_{\rm out}$) (top) and the ratio between input peak SN (SN$_{\rm in}$) and output peak SN (SN$_{\rm out}$) (bottom) as a function of output peak SN calculated for for the regions with coverage $> 0.6$ (red) and $< 0.6$ (blue).
The 30000 trials in each coverage region are presented as dots.
The squares and error bars represent mean and 1$\sigma$ with an bin.
The dashed curves show the best-fit function of $f({\rm SN}) = 1 + \exp(a {\rm SN}^b)$ for the entire region, where $(a,b) = (-0.725, 0.612)$ and $(-0.360, 1.11)$ for the top and bottom panels, respectively.
}
\label{fig:boosting}
\end{figure}
\subsection{1.2mm Number Counts} \label{subsec:counts}
By using the 4.5$\sigma$ sources, we create differential and cumulative number counts.
To create number counts, we correct for the contamination of spurious sources, the effective area, and the completeness as follows:
\begin{eqnarray}
\frac{dN}{dS} = \frac{1}{\Delta S} \sum_{i} \frac{1 - f_{\rm neg}{({\rm SN}_i})}{A(S_i) C({\rm SN}_i)},
\end{eqnarray}
where $S_i$ is the observed source flux density, $f_{\rm neg}$ is the negative fraction accounting for spurious detections, $A$ is the effective area, $C$ is the completeness, and $\Delta S$ is the width of the flux bin.
Figure~\ref{fig:fdr_frac} shows the differential fraction of the number of negative peaks to positive peaks ($f_{\rm neg}$) as a function of SN.
The contamination of spurious sources to each source is estimated by using the best-fit function of the negative fraction and is subtracted from unity.
Then the counts are divided by the completeness by using the best-fit function as a function of SN (figure~\ref{fig:completeness}).
Here we use SNs corrected for the boosting effect presented in figure~\ref{fig:boosting} (bottom panel).
The effective area estimated for each flux density is used as the survey area for a source.
The effect of flux boosting on the source flux density is corrected by using the best-fit function shown in figure~\ref{fig:boosting} (top panel).
The uncertainties from Poisson fluctuations is estimated from Poisson confidence limits of 84.13\% \citep{gehr86}, which correspond to 1$\sigma$ for Gaussian statistics that can be applied to small number statistics.
The derived number counts are shown in figure~\ref{fig:counts} and table~\ref{tab:counts}.
The differential number counts obtained in this study and previous studies are fitted to a Schechter function of the form,
\begin{eqnarray}
\frac{dN}{dS} = \frac{N'}{S'} \left( \frac{S}{S'} \right)^{\alpha} \exp{\left( \frac{-S}{S'} \right)}.
\end{eqnarray}
In this fit, we use the ALMA number counts plotted in figure~\ref{fig:counts}, which are based on blank-field surveys and serendipitously-detected sources at 1.1--1.3~mm to constrain the faint flux range ($<$1 mJy), and the results of 870-$\mu$m follow-up observations of single dish sources \citep{kari13, stac18} for the bright end by scaling the flux densities from 870-$\mu$m to 1.2~mm.
Here we assume a modified black body with a dust emissivity index of $\beta = 1.5$, dust temperature of 35~K, and $z = 2$.
The best-fit parameters are summarized in table~\ref{tab:fit_counts}.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig15.eps}
\end{center}
\caption{
The differential fraction of negative peaks to positive peaks as a function of peak SN.
The dashed curve represents the best-fit function of $f({\rm SN}) = [1 + {\rm erf}(({\rm SN} - a)/b)]/2$, where $(a, b) = (4.60, 0.165)$.
}
\label{fig:fdr_frac}
\end{figure}
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig16.eps}
\end{center}
\caption{
Differential (left) and cumulative (right) number counts at 1.2~mm obtained for ASAGAO sources (red squares).
For comparison, we plot the results for the
ALMA blank field surveys of SXDF-ALMA at 1.1~mm \citep{hats16}, ASPECS at 1.2~mm \citep{arav16}, HUDF at 1.3~mm \citep{dunl17}, and GOODS-ALMA at 1.1~mm \citep{fran18}.
Number counts derived from serendipitously-detected ALMA sources by \citet{hats13}, \citet{ono14}, \citet{carn15}, \citet{fuji16}, and \citet{oteo16} are also presented.
For the bright end, ALMA 870~$\mu$m follow-up observations of single-dish sources by \citet{kari13}, \citet{simp15b}, and \citet{stac18} are presented.
The solid curve and shaded area represent the best-fitting functions in the form of Schechter function and 1$\sigma$ error fitted to the differential number counts.
The flux densities of the counts are scaled to the wavelength of ASAGAO by assuming a modified black body with a dust emissivity index of $\beta = 1.5$, dust temperature of 35~K, and $z = 2$.
}
\label{fig:counts}
\end{figure*}
\begin{table}
\tbl{Differential and cumulative number counts. \label{tab:counts}}
{
\footnotesize
\begin{tabular}{cccccc}
\hline
$S$ & $N$ & $dN/dS$ & $S$ & $N$ & $N$($>$$S$) \\
(1) & (2) & (3) & (4) & (5) & (6) \\
(mJy)& & ($10^2$ mJy$^{-1}$ deg$^{-2}$) & (mJy) & & ($10^2$ deg$^{-2}$) \\
\hline
0.180 & 6 & $924^{+552}_{-366}$ & 0.135 & 45 & $213^{+64}_{-43}$ \\
0.341 & 12 & $299^{+114}_{-85}$ & 0.240 & 39 & $116^{+26}_{-20}$ \\
0.568 & 17 & $132^{+40}_{-32}$ & 0.427 & 27 & $ 60^{+15}_{-12}$ \\
0.878 & 7 & $ 20.0^{+10.7}_{-7.4}$ & 0.759 & 10 & $ 16^{+7.5}_{-4.9}$ \\
1.828 & 3 & $ 4.0^{+3.9}_{-2.2}$ & 1.350 & 3 & $ 4.2^{+4.1}_{-2.3}$ \\
\hline
\end{tabular}}
\begin{tabnote}
(1) Weighted-mean flux density for bin center.
(2) Number of sources for differential number counts.
(3) Differential number counts.
(4) Flux density for bin minimum.
(5) Number of sources for cumulative number counts.
(6) Cumulative number counts.
\end{tabnote}
\end{table}
\subsection{Comparison with Previous ALMA Studies}\label{subsec:counts_comparison}
We compare the ASAGAO number counts with the previous results in the ALMA blank-field surveys.
The number counts of SXDF-ALMA are obtained by using 23 (4$\sigma$) sources detected in a 2~arcmin$^2$ area at 1.1~mm \citep{hats16}.
The ASPECS number counts are derived from 16 (3$\sigma$) sources detected in a deeper 1~arcmin$^2$ survey at 1.2~mm, covering a fainter flux range \citep{arav16}.
The HUDF number counts are obtained in a 4.5~arcmin$^2$ survey at 1.3~mm \citep{dunl17} by using 16 sources (3.5$\sigma$, $S_{\rm 1.3mm} > 120$~$\mu$Jy) with secure galaxy counterparts.
The GOODS-ALMA number counts are obtained from 20 sources (4.8$\sigma$) detected in a 69~arcmin$^2$ survey at 1.1~mm \citep{fran18}.
The ASAGAO number counts are constructed from the largest sample among the blank-field surveys, leading to the small uncertainty from Poisson statistics.
The flux range connects the fainter range probed by ALMA deep observations and the brighter range constrained by ALMA follow-up observations of single-dish detected sources.
We find that our number counts are consistent with those of the previous ALMA blank-field surveys.
The number counts obtained by using the ensemble of serendipitously-detected sources are also compared \citep{hats13, ono14, carn15, fuji16, oteo16}.
While the faintest bin of \citet{oteo16} is lower than the ASAGAO number counts, these number counts are overall consistent within errors.
Note that the lower SN thresholds ($\lesssim$4.5--5$\sigma$) adopted in previous studies might include a larger fraction of spurious sources and overestimate the number counts, although the number counts are corrected for the contamination of spurious sources (e.g., \cite{oteo16, hats16, umeh17, umeh18}).
\subsection{Contribution to Extragalactic Background Light} \label{subsec:ebl}
By using the derived differential number counts, we calculate the fraction of the EBL resolved into discrete sources in this survey.
The integration of the ASAGAO differential number counts yields $7.7^{+1.7}_{-1.2}$ Jy~deg$^{-2}$ ($S_{\rm 1.2mm} > 135$~$\mu$Jy).
The EBL at 1.2~mm (243~GHz) is estimated from the measurements by the {\sl Planck} satellite \citep{plan14} following \citet{arav16} and \citet{muno17}.
By interpolating the measurements at 217 and 353 GHz, the EBL at 1.2~mm is calculated to be $15.1 \pm 0.59$ Jy~deg$^{-2}$.
We find that $52^{+11}_{-8}$\% of the EBL at 1.2~mm is resolved into discrete sources in the ASAGAO map.
The integration of the best-fitting function in the form of Schechter function reaches 100\% at $S_{\rm 1.2mm} \sim 20$~$\mu$Jy, although we note that there is a large uncertainty to extend the function to the faint flux regime.
The flux density of $\sim$20~$\mu$Jy is comparable to the stacked ALMA 1.3~mm signal ($S_{\rm 1.3mm} = 20.1 \pm 4.6$~$\mu$Jy, corresponding to SFR of $6.0 \pm 1.4$~$M_{\odot}$~yr$^{-1}$) derived by \citet{dunl17} on the positions of 89 galaxies in the redshift range of $1 < z < 3$ and the stellar mass range of $9.3 < \log(M_*/M_{\odot}) < 10.3$.
This flux density is also comparable to the stacked flux density of 21 NIR sources with 3.6~$\mu$m magnitudes of $m_{\rm 3.6 \mu m} = 22$--$23$ ($S_{\rm 1.1mm} = 29 \pm 15$~$\mu$Jy, corresponding to SFR of several $M_{\odot}$~yr$^{-1}$) in SXDF-ALMA derived by \citet{wang16}, who found that $\sim$80\% of the EBL is recovered by $m_{\rm 3.6 \mu m} < 23$ sources.
To individually detect these faint submm sources, which significantly contribute to the EBL, it is essential to conduct much deeper observations than in existing deep surveys or use gravitational lensing effects.
\citet{fuji16} showed that nearly 100\% of the EBL can be explained by including gravitational lensed sources at the faint end ($S_{\rm 1.2mm} \sim 20$~$\mu$Jy).
On the other hand, \citet{muno17} argue that their 1$\sigma$ upper limits to differential counts derived from three galaxy clusters as part of the ALMA Frontier Fields Survey are lower than those of \citet{fuji16} by $\approx$0.5 dex and the resolved fraction is only 32\% down to $S_{\rm 1.1mm} = 13$~$\mu$Jy.
Since the faintest end of number counts derived from lensed sources depends on the lensing model, deeper surveys in blank fields are essential to resolve this discrepancy.
\begin{table}
\tbl{Best-fit parameters of parametric fit to differential number counts.$^*$ \label{tab:fit_counts}}{
\footnotesize
\begin{tabular}{ccc}
\hline
$N'$ & $S'$ & $\alpha$ \\
($10^2$ deg$^{-2}$) & (mJy) & \\
\hline
$31.3 \pm 16.6$ & $1.34 \pm 0.30$ & $-2.03 \pm 0.16$ \\
\hline
\end{tabular}}
\begin{tabnote}
$^*$The errors are 1$\sigma$.
\end{tabnote}
\end{table}
\section{Luminosity Function}\label{sec:lf}
While IR luminosity functions of submm sources have been extensively studied by {\sl Herschel} at wavelengths $\le$ 500~$\mu$m (e.g., grup13, magn13), the results are affected by source blending and sensitivity limit due to the large beam size.
Studies at 850~$\mu$m--1~mm wavelengths has been very limited \citep{kopr17}.
In this section, we present the methods of constructing IR LFs from the ASAGAO sources (Sec.~\ref{subsec:lf}), and compare the results with previous studies (Sec.~\ref{subsec:lf_comparison}).
We estimate the contribution of the ASAGAO sources to the cosmic SFR density (SFRD) at $z \sim 2$ by using the derived LFs (Sec.~\ref{subsec:sfrd}).
\subsection{IR Luminosity Function of ASAGAO Sources}\label{subsec:lf}
To estimate LFs, the redshifts of the ASAGAO sources are required.
We utilize spectroscopic or photometric redshifts of optical/NIR counterparts.
We identify $K_S$-band selected sources from the catalog of the {\tt FourStar} galaxy evolution survey (ZFOURGE; \cite{stra16}).
The ZFOURGE covers a total of 400 arcmin$^2$ including the ASAGAO region with a limiting 5$\sigma$ depth in $K_S$ of 26.0 and 26.3 AB mag for 80\% and 50\% completeness with masking, respectively.
The counterpart identification and SED fitting are described in detail in \citet{yama18}, and here we just give a brief explanation.
The ASAGAO sources are cross-matched with the ZFOURGE catalog.
For point-like $K_S$-band sources, we adopt a search radius of 0\farcs5, which is small enough to identify a counterpart.
For extended $K_S$-band sources, we adopt a larger radius, up to half-light radius.
By using ancillary multi-wavelength data (0.4--500 $\mu$m) and our ALMA photometry, SED fitting with the {\sc magphys} model \citep{dacu08, dacu15} is performed.
The SED templates of \citet{bruz03} and the dust extinction model of \citet{char00} are adopted.
The number of ASAGAO sources with ZFOURGE counterparts are 20 (80\%) and 25 (56\%) for 5$\sigma$ and 4.5$\sigma$ sources, respectively.
We use the 5$\sigma$ sources for constructing IR LFs by considering the completeness of the counterpart identification.
Note that the 5$\sigma$ sources without counterparts are likely to be at higher redshifts ($z \gtrsim$4--5) based on their optical--ratio SEDs \citep{yama18}, and therefore they do not affect the following discussion for the LFs at $z =$ 1--3 significantly.
The spectroscopic or photometric redshifts are available in the ZFOURGE catalog.
IR luminosities (measured in the rest-frame 8--1000 $\mu$m) are derived in the SED fitting.
The IR luminosities as a function of redshift are shown in figure~\ref{fig:z-lir}.
To construct the LFs, we adopt the $V_{\rm max}$ method \citep{schm68}.
This method uses the maximum observable volume of each source.
The LF gives the number of ALMA sources in a comoving volume per logarithm of luminosity and is obtained as
\begin{eqnarray}
\Phi(L,z) = \frac{1}{\Delta L} \sum_{i} \frac{1}{C({\rm SN}_i) V_{{\rm max},i}},
\end{eqnarray}
where $V_{{\rm max},i}$ is the maximum observable volume of the $i$th source, $C$ is the completeness, and $\Delta L$ is the width of the luminosity bin.
We adopt a luminosity bin width of $\Delta\log{(L)} = 0.6$.
Because the noise level in the map is not uniform, we need to take into account the effective solid angle where a source can be detected for calculating $V_{\rm max}$.
Following the description of \citet{nova17}, where they construct radio LFs taking into account a nonuniform noise in their radio maps, we calculate $V_{\rm max}$ as the integration of comoving volume spherical shells as
\begin{eqnarray}
V_{{\rm max},i} = \int_{z_{\rm min}}^{z_{\rm max}} \frac{\Omega(S_i(z))}{4\pi} \frac{dV}{dz} dz,
\end{eqnarray}
where $z_{\rm min}$ and $z_{\rm max}$ are maximum and minimum redshifts of a redshift bin, $S_i(z)$ is the flux density of source $i$ observed when it is located at $z$, and $\Omega$ is the solid angle where source $i$ with a flux density of $S_i(z)$ can be detected with SN $>$ 5.
$S_i(z)$ is estimated from the SED model of each source, and $\Omega(S_i(z))$ is derived from the effective area for $S_i(z)$.
Because the number of sources in each bin is small, the error of the LFs is estimated from Poisson confidence limits of 84.13\% (corresponding to Gaussian 1$\sigma$ errors) in \citet{gehr86}.
We derive IR LFs in the redshift ranges of $1.0 < z < 2.0$, $1.5 < z < 2.5$, and $2.0 < z < 3.0$ by using 6 (mean redshift of $z_{\rm mean} = 1.55$), 9 ($z_{\rm mean} = 2.12$), and 13 ($z_{\rm mean} = 2.49$) sources, respectively.
To increase the number of sources in each redshift bin, we adopt the bin width of 1.0, resulting in the overlap of the bins.
The derived IR LFs are presented in table~\ref{tab:lf} and figure~\ref{fig:lf}.
Our study constrains the faintest luminosity end of the LF at $2.0 < z < 3.0$ among other studies.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig17.eps}
\end{center}
\caption{
IR luminosity of the ASAGAO sources with $K_S$-band counterpart as a function of redshift.
The solid curve represents the luminosity limit in this study estimated from the average SED of the 4.5$\sigma$ sources and a detection limit of $\sigma_{\rm 1.2mm} = 26$~$\mu$Jy~beam$^{-1}$.
}
\label{fig:z-lir}
\end{figure}
\begin{table}
\tbl{IR Luminosity Functions. \label{tab:lf}}{
\footnotesize
\begin{tabular}{ccc}
\hline
$\log{(L_{\rm IR}/L_{\odot})}^{\dagger}$ & $N$ & $\log{(\Phi/{\rm Mpc^{-3} dex^{-1}})}$ \\
\hline\\
\multicolumn{3}{c}{$1.0 < z < 2.0$} \\
\hline
11.86 & 4 & $-3.89^{+0.25}_{-0.28}$ \\
12.46 & 2 & $-4.34^{+0.37}_{-0.45}$ \\
\hline\\
\multicolumn{3}{c}{$1.5 < z < 2.5$} \\
\hline
11.91 & 6 & $-3.66^{+0.20}_{-0.22}$ \\
12.44 & 3 & $-4.25^{+0.30}_{-0.34}$ \\
\hline\\
\multicolumn{3}{c}{$2.0 < z < 3.0$} \\
\hline
11.94 & 7 & $-3.05^{+0.19}_{-0.20}$ \\
12.57 & 6 & $-3.97^{+0.20}_{-0.22}$ \\
\hline
\end{tabular}}
\begin{tabnote}
$^{\dagger}$ Weighted-mean luminosity in each bin.
\end{tabnote}
\end{table}
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig18.eps}
\end{center}
\caption{
IR luminosity functions constructed from the ASAGAO sources at $1.0 < z < 2.0$ (left), $1.5 < z < 2.5$ (middle), and $2.0 < z < 3.0$ (right).
We plot luminosity functions obtained in \citet{kopr17} (K17) by using 1.3~mm sources from the ALMA HUDF survey and ~$\mu$m sources from the SCUBA-2 Cosmology Legacy Survey.
The dashed curve and shaded area represent the best-fitting functions and 1$\sigma$ error of \citet{kopr17}.
At $1.5 < z < 2.5$, we plot their data points derived from the $V_{\rm max}$ method and the best-fitting function.
At $1.0 < z < 2.0$ and $2.0 < z < 3.0$, we plot their functional form of the redshift evolution of the LF derived from the maximum-likelihood method, adopting the mean redshifts of the ASAGAO sources in each redshift bin ($z = 1.7$ and $z = 2.5$, respectively).
The results of {\sl Herschel} observations by \citet{magn13} (M13) and \citet{grup13} (G13) are also compared.
The solid curve and shaded area represent the the best-fitting Schechter function and 1$\sigma$ error fitted to the results of ASAGAO and \citet{kopr17} at $1.5 < z < 2.5$.
}
\label{fig:lf}
\end{figure*}
\subsection{Comparison with Previous Studies}\label{subsec:lf_comparison}
We compare the ASAGAO LFs with those derived from sources detected with ALMA, SCUBA2, and {\sl Herschel}.
\citet{kopr17} derived rest-frame 250~$\mu$m LFs and IR LFs up to $z \sim 5$ by using 16 1.3-mm sources detected in the ALMA HUDF survey \citep{dunl17} for constraining the faint end and 577 850-$\mu$m sources detected in the COSMOS and UDS fields as part of the SCUBA-2 Cosmology Legacy Survey (S2CLS; \cite{geac17, chen16, mich17}) for constraining the bright end.
The wide coverage of the luminosity range and the large sample for the bright end allowed them to examine the evolution of LFs derived for submm sources.
They derived LFs for four redshift bins $z =$ 0.5--1.5, 1.5--2.5, 2.5--3.5, and 3.5--4.5, by using the $V_{\rm max}$ method.
They determined the faint-end slope of $\alpha = -0.4$ in the Schechter form of
\begin{eqnarray}
\Phi(L) = \Phi_* \left( \frac{L}{L_*} \right)^{\alpha} \exp{\left( \frac{-L}{L_*} \right)}, \label{eq:lf}
\end{eqnarray}
by fitting to the data in the redshift bin of $1.5 < z < 2.5$, where ALMA sources are available for constraining the faint end.
The remaining Schechter-function parameters were determined by fixing the faint-end slope $\alpha$ to $-0.4$.
To estimate the continuous form of the redshift evolution of the LF, they used the maximum-likelihood method.
In figure~\ref{fig:lf}, we plot their data points and the best-fitting function determined in the redshift bin of $1.5 < z < 2.5$, and the LFs determined from the maximum-likelihood method for the redshift bins of $1.0 < z < 2.0$ and $2.0 < z < 3.0$.
They find that the LFs are well characterized by the number density/luminosity evolution of LFs with positive luminosity evolution coupled with negative density evolution with increasing redshift.
We find that the ASAGAO LFs are consistent with those of \citet{kopr17} within the errors, supporting the evolution of LFs derived in \citet{kopr17}, although the large uncertainties of our LFs due to the small sample size and the limited coverage of IR luminosity do not allow us to further discuss the density/luminosity evolution of submm sources.
The ASAGAO LFs at $2.0 < z < 3.0$ is above their results, while those results are consistent.
This may suggest a stronger luminosity evolution or weaker density evolution.
The fainter bin of the ASAGAO LFs at $1.0 < z < 2.0$ is about a factor of a few lower than that of \citet{kopr17}.
This may be due to the fact that they fixed the faint-end slope when deriving the LF evolution.
The results of {\sl Herschel} observations are also compared in figure~\ref{fig:lf}.
\citet{grup13} derived IR LFs up to $z \sim 4$ by using the data from the {\sl Herschel}-PEP survey in combination with the {\sl Herschel}-HerMES data.
\citet{magn13} presented IR LFs up to $z \sim 2$ obtained in the GOODS fields from the PEP and the GOODS-{\sl Herschel} programs.
\citet{kopr17} found that a discrepancy between the results based on submm sources and {\sl Herschel} sources at the bright end, and concluded that {\sl Herschel} results are contaminated and biased high by a mix of source blending, mis-identification of counterpart (and hence redshift) due to the large beam size of {\sl Herschel}/SPIRE.
Although the {\sl Herschel} results scattered and the redshift ranges are not exactly the same as in ours, we find that they are overall consistent with the ASAGAO LFs.
We fit the IR LFs at $1.5 < z < 2.5$ obtained from the ASAGAO sources and the results of \citet{kopr17} with a Schechter function of the form of equation~\ref{eq:lf}.
The best-fitting parameters are presented in table~\ref{tab:fit_lf}.
The derived spectral slope of $\alpha = -0.22 \pm 0.28$ is flatter than $\alpha = -0.4$ derived by \citet{kopr17}, but consistent within the errors.
In order to constrain the redshift evolution of LFs, it is essential to conduct wider-area surveys for obtaining a larger sample in a wide range of IR luminosity.
\subsection{Contribution to the Cosmic SFR Density}\label{subsec:sfrd}
By integrating the best-fit IR LF and converting it to SFRD, we estimate the contribution of ASAGAO sources to the cosmic SFRD at $z \sim 2$.
SFR is converted from IR luminosity by using the relation of \citet{kenn98} and corrected to a \citet{chab03} IMF.
The integration of the best-fitting luminosity function down to the lowest luminosity of the sources ($\log{(L_{\rm IR}/L_{\odot})} = 11.78$) gives a SFRD of $7.2^{+3.0}_{-1.9} \times 10^{-2}$~$M_{\odot}$~yr$^{-1}$~Mpc$^{-3}$.
This is consistent with the results of \citet{yama18}, where they derived the SFRD by counting the contribution from individual ASAGAO sources.
We compare the SFRD with the total SFRD (UV $+$ IR) at $z \sim 2$ estimated in previous studies:
0.13 $M_{\odot}$~yr$^{-1}$~Mpc$^{-3}$ at $z = 2$ by \citet{mada14},
or 0.11--0.12 $M_{\odot}$~yr$^{-1}$~Mpc$^{-3}$ at $z =$ 1.8--2.25 by \citet{burg13}.
The fraction of SFRD contributed by the ASAGAO sources is $\approx$60--90\% at $z \sim 2$, indicating that the major portion of SFRD at that redshift is composed of obscured star formation from sources with $\log{(L_{\rm IR}/L_{\odot})} \gtrsim 11.8$ (e.g., \cite{dunl17, kopr17}).
This is reasonable considering that the IR luminosity is somewhat lower than the turnover IR luminosity of the best-fit Schechter function.
\begin{table}
\tbl{Best-fit parameters of parametric fit to LF at $1.5 < z < 2.5$ by using the ASAGAO sources and the results of \citet{kopr17}.$^*$ \label{tab:fit_lf}}{
\footnotesize
\begin{tabular}{ccc}
\hline
$\log{(\Phi_*/{\rm Mpc^{-3} dex^{-1}})}$ & $\log{(L_*/L_{\odot})}$ & $\alpha$ \\
\hline
$-3.07 \pm 0.07$ & $12.12 \pm 0.05$ & $-0.22 \pm 0.28$ \\
\hline
\end{tabular}}
\begin{tabnote}
$^*$The errors are 1$\sigma$.
\end{tabnote}
\end{table}
\section{Conclusions}\label{sec:conclusions}
We performed the ALMA twenty-six arcmin$^2$ survey of GOODS-S at one-millimeter (ASAGAO).
The central 26~arcmin$^2$ area of the GOODS-S field was observed at 1.2~mm, providing a map with $1\sigma \sim 61$~$\mu$Jy~beam$^{-1}$ (250 k$\lambda$-taper) and a synthesized beam size of $0\farcs51 \times 0\farcs45$.
By combining the ALMA archival data available in the GOODS-S field (HUDF by \cite{dunl17} and GOODS-ALMA by \cite{fran18}), we obtained a deeper map for the 26~arcmin$^2$ area, which has a rms noise level of $1\sigma \sim 30$~$\mu$Jy~beam$^{-1}$ for the central region with a 250 k$\lambda$-taper and a synthesized beam size of $0\farcs59 \times 0\farcs53$.
We find 25 sources at 5$\sigma$ and 45 sources at 4.5$\sigma$ in the combined ASAGAO map, providing the largest source catalog among ALMA blank field surveys.
The flux densities are consistent with those estimated in the other ALMA GOODS-S surveys by considering the difference in observing wavelength.
The larger sample allow us to construct 1.2~mm number counts with smaller uncertainties from Poisson statistics.
The flux coverage of the number counts connects the fainter range probed by ALMA deep observations and the brighter range constrained by ALMA follow-up observations of single-dish detected sources.
We find that our number counts are consistent with previous ALMA studies.
By integrating the derived differential number counts, we find that $52^{+11}_{-8}$\% of the EBL at 1.2~mm is revolved into the discrete sources.
The integration of the best-fitting function reaches 100\% at $S_{\rm 1.2mm} \sim 20$~$\mu$Jy, although there is a large uncertainty to extend the function to the fainter flux range.
Deeper surveys are required to individually detect faint submm sources, which significantly contribute to the EBL.
By using the 5$\sigma$ sources, we construct IR LFs in the redshift ranges of $1.0 < z < 2.0$, $1.5 < z < 2.5$, and $2.0 < z < 3.0$.
Our study constrains the faintest luminosity end of the LF at $2.0 < z < 3.0$ among other studies.
We find that the ASAGAO LFs are consistent with those of \citet{kopr17}, supporting the evolution of LFs (positive luminosity evolution and negative density evolution with increasing redshift) derived in \citet{kopr17}.
The integration of the best-fitting LF down to the lowest luminosity of the sources ($\log{(L_{\rm IR}/L_{\odot})} = 11.78$) gives a SFRD of $7.2^{+3.0}_{-1.9} \times 10^{-2}$~$M_{\odot}$~yr$^{-1}$~Mpc$^{-3}$.
We find that the IR-based star formation of ASAGAO sources contribute to $\approx$60--90\% of the SFRD at $z \sim 2$ derived from UV--IR observation, indicating that the major portion of $z \sim 2$ SFRD is composed of sources with $\log{(L_{\rm IR}/L_{\odot})} \gtrsim 11.8$.
\begin{ack}
We are grateful to Maciej Koprowski for providing the scaling factor of their LFs.
BH, KK, YT, HU, and YU are supported by JSPS KAKENHI Grant Number 15K17616, 17H06130, and 17K14252.
RJI acknowledges support from ERC in the form of Advanced Investigator Programme, COSMICISM, 321302.
This study is supported by the NAOJ ALMA Scientific Research Grant Number 2017-06B and 2018-09B, and by the ALMA Japan Research Grant of NAOJ Chile Observatory, NAOJ-ALMA-190.
This paper makes use of the following ALMA data: ADS/JAO.ALMA\#2015.1.00098.S, \#2012.1.00173.S, and \#2015.1.00543.S.
ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada), NSC and ASIAA (Taiwan), and KASI (Republic of Korea), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO and NAOJ.
\end{ack}
|
\section{Preliminary lemmas}
Before proving the main result we need some lemmas.
\begin{lemma} \label{arit2p}
Let $p$ be an odd prime and let $n=2p$. Then the following are equivalent.
\begin{enumerate}
\item $S_n$ has primitive subgroups different from $A_n$ and $S_n$.
\item $A_n$ has primitive subgroups different from $A_n$.
\item $S_n$ has primitive maximal subgroups different from $A_n$.
\item $A_n$ has primitive maximal subgroups.
\item $n$ is equal to $22$ or of the form $q+1$ where $q$ is a prime power.
\end{enumerate}
\end{lemma}
\begin{proof}
$(2)$ and $(3)$ clearly imply $(1)$, and $(2) \Leftrightarrow (4)$ is clear as a permutation group containing a primitive one is primitive. We prove that $(5)$ implies $(1)$, $(2)$, $(3)$ and $(4)$. Suppose $(5)$ holds. If $n=22$ then $A_n$ contains a primitive Mathieu subgroup $\mbox{M}_{22}$ and this gives the result, now assume $n \neq 22$. Consider $G=\PGL_2(q) \leq S_n$ where $n=2p=q+1$, $q$ a prime power, and $G$ is acting in the usual way on the projective line. Since $p \geq 3$ we have $q \geq 5$ hence $G$ is an almost-simple group with socle $\PSL_2(q)$. Being $2$-transitive, $G$ is a primitive subgroup of $S_n$. We now prove that $G \cap A_n = \PSL_2(q)$. The inclusion $\supseteq$ is clear as $G \cap A_n$ is a normal subgroup of $G$ of index at most $2$. We are left to prove that $G$ is not contained in $A_n$. Let $g \in \PGL_2(q)$ be the element corresponding to the diagonal matrix having $u$ and $1$ on the diagonal, where $u$ is a generator of the multiplicative group $\mathbb{F}_q^{\ast}$. $g$ fixes exactly two $1$-dimensional subspaces and cyclically permutes the remaining $q-1=2p-2$, now since $2p-2$ is even $g$ is an odd permutation, therefore $G$ is a primitive subgroup of $S_n$ not contained in $A_n$.
\
We are left to prove that $(1) \Rightarrow (5)$. Let $G$ be a primitive group of degree $n=2p$. It was shown by Wielandt \cite{wie} that $G$ has rank $2$ or $3$, and later using the classification of finite simple groups it was shown that $G$ is $2$-transitive (that is, has rank $2$) with the only exception of $p=5$ (see \cite{cl2p}). We can use the known classification of $2$-transitive groups. Suppose $p \neq 5$. Since $G$ is not affine (because $n$ is not a prime power) $G$ belongs to \cite[Table 7.4]{cam} and it follows that $n$ is equal to $22$ or is of one of the following types.
\begin{itemize}
\item $(q^d-1)/(q-1)$, $d \geq 2$, $(d,q) \neq (2,2),(2,3)$, $q$ a prime power. We have $2p=(q^d-1)/(q-1)$, hence $d$ is a prime because if $d=ab$ with $a,b$ integers then $2p=(q^{ab}-1)/(q^b-1) \cdot (q^b-1)/(q-1)$, a contradiction. On the other hand $(q^d-1)/(q-1) = 1+q+q^2+\ldots+q^{d-1}$ is odd if $d$ is odd, therefore $d=2$ and $2p=q+1$.
\item $q^2+1$, $q=2^{2d+1}>2$. This cannot happen being $2p$ even and $q^2+1$ odd.
\item $q^3+1$, $q \geq 3$ a prime power. Then $2p=q^3+1=(q+1)(q^2-q+1)$ shows that this cannot happen.
\end{itemize}
If $p=5$ then $n=10$ and $S_n$ has the primitive maximal subgroup $\PGammaL_2(9)$, so the result holds in this case as $10=9+1$.
\end{proof}
Observe that from the above proof it follows that
\begin{cor}
Let $p$ be an odd prime and let $G \leq S_{2p}$ be a primitive group of degree $2p$ not containing the alternating group $A_{2p}$. Then one of the following holds.
\begin{itemize}
\item $p=11$ and $\soc(G) \cong \mbox{M}_{22}$ in its primitive action of degree $22$.
\item $2p-1=q \geq 5$ is a prime power and $\soc(G)$ is isomorphic to $\PSL_2(q)$ acting on the projective line.
\end{itemize}
\end{cor}
\section{Proof of Theorem \ref{main}}
\begin{lemma} \label{small}
$\Mindim(A_n) \leq 3$ for all $n \geq 4$ even and for $n=7$, $n=11$. Moreover $\Mindim(A_n)=3$ for $n \in \{6,7,8,11,12\}$.
\end{lemma}
\begin{proof}
Suppose $n=2m \geq 4$ is even. Consider the dihedral group $D$ of order $n$ embedded (via the regular representation) in $S_n$. This gives two subgroups, $H$ and $K$ (the left and right regular representations), of $S_n$ which commute pointwise. Since $H$ is transitive, $C_{S_n}(H)$ is semiregular, in particular its order is at most $n$, implying that $C_{S_n}(H)=K$ and similarly $C_{S_n}(K)=H$. In particular $H$ is an intersection of centralizers of two fixed-point-free involutions, which are maximal subgroups of the form $S_2 \wr S_m$. Since the nontrivial elements of $H$ are fixed-point-free the intersection between $H$ and a point stabilizer is trivial. Intersecting with $A_n$ we obtain $\Mindim(A_n) \leq 3$.
In the alternating group $A_7$ the intersection of the setwise stabilizers of $\{1,2,3,4\}$, $\{1,2,5\}$ and $\{2,5,6\}$ is trivial, and such stabilizers are maximal subgroups of $A_7$, therefore $\Mindim(A_7) \leq 3$. The first, second and fourth subgroups of the sixth conjugacy class of maximal subgroups of $A_{11}$ stored in \cite{gap} are isomorphic to $\mbox{M}_{11}$ and have trivial intersection, implying $\Mindim(A_{11}) \leq 3$. Let now $n \in \{6,7,8,11,12\}$. In order to prove that $\Mindim(A_n) \geq 3$, using \cite{gap} we defined a function that builds the list of pairs $(A,B)$ where $A$ is a fixed representative of every conjugacy class of maximal subgroups of $A_n$ and $B$ is any maximal subgroup of $A_n$ distinct from $A$, and it returns true if and only if for each such pair there exists a maximal subgroup $C$ of $A_n$ belonging to the first conjugacy class of maximal subgroups of $A_n$ and with the property that $\{A,B,C\}$ is irredundant.
\end{proof}
We will use \cite[Corollary 1.5]{basesym}, which we state here for convenience.
\begin{teor} \label{balt}
Let $G=A_n$ with $n \geq 5$ acting primitively on a set with point stabilizer $H$. Then either $b(G)=2$ or $(n,H)$ is one of the following:
\begin{enumerate}
\item $H=(S_k \times S_{n-k}) \cap G$ with $k < n/2$.
\item $H=(S_2 \wr S_l) \cap G$ and $n=2l$.
\item $H=(S_k \wr S_l) \cap G$, $n=kl$ with $k \geq 3$, and either $l < k+2$ or $l=k+2 \in \{5,6\}$.
\item $(n,H)=(12,\mbox{M}_{12})$, $(11,\mbox{M}_{11})$, $(9,\PGammaL_2(8))$, $(8,\AGL_3(2))$, $(7,PSL_2(7))$, $(6,PSL_2(5))$, $(5,D_{10})$.
\end{enumerate}
In particular if $H$ acts primitively on $\{1,\ldots,n\}$ then $b(G)=2$ with finitely many exceptions.
\end{teor}
We proceed to the proof of Theorem 1. Observe that $A_4$ has a base of size $2$, so $\Mindim(A_4) = 2$. Now assume $n \geq 5$. By Lemma \ref{small} we know that $\Mindim(A_n)=3$ for $n \in \{6,7,8,11,12\}$. Also, $\Mindim(A_5)=2$ because the intransitive subgroups stabilizing $\{1,2\}$ and $\{1,3\}$ have trivial intersection, $\Mindim(A_9)=2$ by Theorem \ref{balt} because $A_9$ contains the primitive maximal subgroup $\ASL(2,3)$, which is different from $\PGammaL_2(8)$, and $\Mindim(A_{10})=2$ by Theorem \ref{balt} because $A_{10}$ contains a primitive maximal subgroup by Lemma \ref{arit2p}. This means that we may assume $n \geq 13$.
\
Observe that, being $n \geq 13$, by Theorem \ref{balt} we may assume that for all factorizations $n = k \ell$ with $k \geq 3$ we have $\ell \leq k+1$ or $\ell = k+2 \in \{5,6\}$. Also, if $n=p$ is a prime then letting $H$ to be any maximal subgroup of $G$ containing a $p$-cycle, $H$ acts primitively on $\{1,\ldots,p\}$ so $\Mindim(G)=2$. If $\ell = k+2 \in \{5,6\}$ then $n \in \{15,24\}$, and $A_{15}$, $A_{24}$ have proper primitive subgroups (with respect to the natural action on $\{1,\ldots,n\}$), so $\Mindim(A_{15}) = \Mindim(A_{24}) = 2$. Now assume that for all such factorizations we have $\ell \leq k+1$. Suppose $n$ is a product $abc$ with $2 \leq a \leq b \leq c$. If $b \geq 3$ then $ac \leq b+1 \leq c+1$, a contradiction. If $a=b=2$ then $n=4c$ so $c \leq 4+1=5$ and $n \in \{16,20\}$. However $A_{16}$ has a proper primitive subgroup $\AGL_2(4)$, and $A_{20}$ has a proper primitive subgroup $\PSL_2(19)$. We are left to discuss the case in which $n$ is of the form $pq$ with $p \geq q$ primes. If $q \neq 2$ then $q \leq p \leq q+1$ hence $p=q$ and $n=p^2$, $A_n$ has a maximal subgroup containing $\ASL_2(p)$, which acts primitively, so the corresponding primitive action of $A_n$ has a base of size $2$. Assume now that $n=2p$ with $p$ an odd prime, so that $p \geq 7$ being $n \geq 13$. If $A_n$ has primitive maximal subgroups then $\Mindim(A_n)=2$, now assume this is not the case. By Lemma \ref{arit2p}, $n$ is of the form specified in the statement.
\
Assume $n$ has the form $2p$ with $p \geq 7$ an odd prime such that $2p$ is not of the form $q+1$ with $q$ a prime power and $p \neq 11$. We claim that in this case $\Mindim(A_n)=3$. By Lemma \ref{small} it is enough to prove that $\Mindim(A_n) \geq 3$. By Lemma \ref{arit2p} the maximal subgroups of $S_n$ distinct from $A_n$ and the maximal subgroups of $A_n$ act non-primitively on $\{1,\ldots,n\}$. Moreover the maximal subgroups of $A_n$ are precisely the intersections between $A_n$ and the maximal subgroups of $S_n$. We will show that if $A$ and $B$ are maximal subgroups of $S_n$ then there exists a maximal subgroup $C$ of $S_n$ such that $\{A \cap A_n,B \cap A_n,C \cap A_n\}$ is irredundant.
\begin{enumerate}
\item Suppose $A$ and $B$ are of the form $S_p \wr S_2$. Say $A$ preserves a base block $\mathscr{A}$ and $B$ preserves a base block $\mathscr{B}$ (both of size $p$). By replacing, if needed, $\mathscr{A}$ with its complement we may assume $|\mathscr{A} \cap \mathscr{B}| \geq (p+1)/2 \geq 4$ so that $|\mathscr{A} \cup \mathscr{B}| \leq 2p-4$. Let $a \in \mathscr{A}-\mathscr{B}$, $i,j \in \mathscr{A} \cap \mathscr{B}$, $b \in \mathscr{B}-\mathscr{A}$, $\ell,t,s \in \{1,\ldots,2p\}-(\mathscr{A} \cup \mathscr{B})$. Let $C$ be the maximal intransitive subgroup stabilizing $\{b,i,\ell\}$ and its complement.
\begin{itemize}
\item $(ij)(\ell t) \in (A \cap B \cap A_n)-C$.
\item $(aj)(ts) \in (A \cap C \cap A_n)-B$.
\item $(bi)(ts) \in (B \cap C \cap A_n)-A$.
\end{itemize}
\item Suppose $A$ is of the form $S_p \wr S_2$ and $B$ is of the form $S_k \times S_{n-k}$ where $1 \leq k < p$. Say $A$ stabilizes a base block $\mathscr{A}$ of size $p$ and $B$ stabilizes $\mathscr{B}$ and its complement, with $|\mathscr{B}|$ larger than $p$.
Assume first that $\mathscr{A} \subseteq \mathscr{B}$. Pick $a,x,r,s \in \mathscr{A}$, $b \in \mathscr{B}-\mathscr{A}$ and $c \in \{1,\ldots,2p\}-\mathscr{B}$, and let $C$ be the intransitive maximal subgroup stabilizing $\{a,b,c\}$ and its complement.
\begin{itemize}
\item $(ax)(rs) \in (A \cap B \cap A_n)-C$.
\item $(bc)(rs) \in (A \cap C \cap A_n)-B$.
\item $(ab)(rs) \in (B \cap C \cap A_n)-A$.
\end{itemize}
Now assume $\mathscr{A} \not \subseteq \mathscr{B}$, and let $a \in \mathscr{A}-\mathscr{B}$. Since $|\mathscr{B}| > p$ there exist two distinct elements $b,x$ in $\mathscr{B}-\mathscr{A}$. Since $|\mathscr{B}|>p$, up to replacing, if needed, $\mathscr{A}$ with its complement we may assume that $|\mathscr{A} \cap \mathscr{B}| \geq (p+1)/2$, so that there exist distinct elements $i,j,r \in \mathscr{A} \cap \mathscr{B}$. Let $C$ be the maximal intransitive subgroup preserving $\{a,b,i\}$ and its complement.
\begin{itemize}
\item $(ir)(bx) \in (A \cap B \cap A_n)-C$.
\item $(ai)(jr) \in (A \cap C \cap A_n)-B$.
\item $(bi)(jr) \in (B \cap C \cap A_n)-A$.
\end{itemize}
\item Suppose $A$ is of the form $S_p \wr S_2$ and $B$ is of the form $S_2 \wr S_p$. Let $\mathscr{A} = \{1,\ldots,p\}$ be a base block for $A$, and let $\mathscr{B}_1,\ldots,\mathscr{B}_p$ be the base blocks corresponding to $B$, each of size $2$. Choose them so that $|\mathscr{B}_1 \cap \mathscr{A}| = 1$.
Suppose $|\mathscr{B}_i \cap \mathscr{A}|=1$ for all $i=1,\ldots,p$, without loss of generality $\mathscr{B}_i = \{i,p+i\}$. Let $C$ be the maximal intransitive subgroup stabilizing $\{1,2,p+1\}$ and its complement.
\begin{itemize}
\item $(1 \ldots p)(p+1 \ldots 2p) \in (A \cap B \cap A_n)-C$.
\item $(1\ p+1) (3\ p+3) \in (B \cap C \cap A_n)-A$.
\item $(12)(34) \in (A \cap C \cap A_n)-B$.
\end{itemize}
We may now assume that $\mathscr{B}_2 = \{r,s\} \subseteq \mathscr{A}$ and $\mathscr{B}_3 = \{y,z\}$ is disjoint from $\mathscr{A}$. Let $x \in \mathscr{B}_1-\mathscr{A}$, $t \in \mathscr{B}_1 \cap \mathscr{A}$, and let $C$ be the maximal intransitive subgroup stabilizing $\{x,y,t\}$ and its complement.
\begin{itemize}
\item $(yz)(rs) \in (A \cap B \cap A_n)-C$.
\item $(xy)(rs) \in (A \cap C \cap A_n)-B$.
\item $(xt)(rs) \in (B \cap C \cap A_n)-A$.
\end{itemize}
\item Suppose $A$ and $B$ are of the form $S_2 \wr S_p$. Since $A \neq B$ there are a base block $\mathscr{A}$ of $A$ and a base block $\mathscr{B}$ of $B$ distinct and with a non-empty intersection, without loss of generality $\mathscr{A}=\{1,2\}$ and $\mathscr{B}=\{1,3\}$. Let $\{3,x\}$ be the base block of $A$ containing $3$ and let $\{2,y\}$ be the base block of $B$ containing $2$. Let $C$ be the stabilizer of the point $1$.
First we prove that $A \cap C \cap A_n \not \subseteq B$ and $B \cap C \cap A_n \not \subseteq A$. If $x \neq y$ let $\{y,z\}$, $\{x,w\}$ be blocks of $A$ and $B$ respectively, we have $(3x)(yz) \in (A \cap C)-B$, $(2y)(xw) \in (B \cap C)-A$ and they also belong to $A_n$, if $x=y=4$ let $\{\alpha,\beta\}$, $\{\gamma,\delta\}$ be new blocks of $A$ and $B$ respectively, then $(34)(\alpha \beta) \in (A \cap C)-B$, $(24)(\gamma \delta) \in (B \cap C)-A$.
We are left to show that $A \cap B \cap A_n$ is not contained in $C$. Construct $\gamma \in A_n$ in the following way. Set $a_1=b_1=1$, $a_2=2$, $b_2=3$ and write the two block systems (corresponding to $A$ and to $B$) as follows: $$\{a_1,a_2\},\ \{b_2,a_3\},\ \{b_3,a_4\},\ \{b_4,a_5\}, \ldots,\ \{b_k,a_{k+1}\}, \ldots$$ $$\{a_1,b_2\},\ \{a_2,b_3\},\ \{a_3,b_4\},\ \{a_4,b_5\},\ \ldots, \{a_k,b_{k+1}\}, \ldots$$where $a_{k+1}=b_{k+1}$.
We want to construct $\gamma \in A_n$ fixing all the elements not belonging to $\{a_1,\ldots,a_{k+1},b_2,\ldots,b_k\}$ and with the property that $\gamma \in A \cap B$ and $\gamma(a_1)=a_3$, so that $\gamma(1) \neq 1$, implying $\gamma \not \in C$. As it turns out there is only one way to do so. If $k=2$ let $\gamma=(a_1 a_3)(a_2 b_2)$, if $k=3$ let $\gamma=(a_1 a_3 b_3)(a_2 b_2 a_4)$. If $k>3$ is odd let $$\gamma := (a_1 a_3 a_5 \ldots a_k b_k b_{k-2} b_{k-4} \ldots b_3) (a_2 b_2 b_4 b_6 \ldots b_{k-1} a_{k+1} a_{k-1} a_{k-3} \ldots a_4),$$whereas if $k>3$ is even let $$\gamma := (a_1 a_3 a_5 \ldots a_{k+1} b_{k-1} b_{k-3} \ldots b_3) (a_2 b_2 b_4 b_6 \ldots b_k a_k a_{k-2} a_{k-4} \ldots a_4).$$ Clearly $\gamma \in A_n$, since it is a product of two $k$-cycles in both cases. It is easy to show that $\gamma \in (A \cap B \cap A_n)-C$.
\item Suppose $A$ is of the form $S_2 \wr S_p$ and $B$ is of the form $S_k \times S_{n-k}$. Say $B$ stabilizes $\mathscr{B}$ and its complement with $|\mathscr{B}| > p$. The cardinality condition implies that $\mathscr{B}$ contains a base block $\mathscr{A}_1=\{x,y\}$. Suppose $\mathscr{B}$ intersects two base blocks $\mathscr{A}_2,\mathscr{A}_3$ in exactly one element, say $\mathscr{A}_2 = \{a,b\}$, $\mathscr{A}_3=\{z,w\}$ with $a,z \in \mathscr{B}$, $b,w \not \in \mathscr{B}$. Let $C$ be the maximal intransitive subgroup stabilizing $\{y,z,w\}$ and its complement.
\begin{itemize}
\item $(az)(bw) \in (A \cap B \cap A_n)-C$.
\item $(ab)(wz) \in (A \cap C \cap A_n)-B$.
\item $(ax)(yz) \in (B \cap C \cap A_n)-A$.
\end{itemize}
Suppose $\mathscr{B}$ contains two base blocks $\mathscr{A}_1=\{x,y\}$, $\mathscr{A}_2=\{a,b\}$ and intersects another base block $\mathscr{A}_3=\{z,w\}$ in exactly one element, say $z \in \mathscr{B}$, $w \not \in \mathscr{B}$. Let $C$ be the maximal intransitive subgroup stabilizing $\{y,z,w\}$ and its complement.
\begin{itemize}
\item $(ax)(by) \in (A \cap B \cap A_n)-C$.
\item $(ab)(wz) \in (A \cap C \cap A_n)-B$.
\item $(abx) \in (B \cap C \cap A_n)-A$.
\end{itemize}
We are left to discuss the case in which $\mathscr{B}$ is a union of some $\mathscr{A}_i$'s (blocks of $A$). In this case let $\{x,y\}$, $\{a,b\}$ be two blocks contained in $\mathscr{B}$ and let $\{t,w\}$ be a block disjoint from $\mathscr{B}$. Let $C$ be the maximal intransitive subgroup stabilizing $\{x,t\}$ and its complement.
\begin{itemize}
\item $(xy)(tw) \in (A \cap B \cap A_n)-C$.
\item $(yw)(xt) \in (A \cap C \cap A_n)-B$.
\item $(yab) \in (B \cap C \cap A_n)-A$.
\end{itemize}
\item Suppose $A$ and $B$ are maximal intransitive. Say $A$ stabilizes $\mathscr{A}$ and its complement, and $B$ stabilizes $\mathscr{B}$ and its complement.
Suppose first that $\mathscr{A} \subseteq \mathscr{B}$ or $\mathscr{B} \subseteq \mathscr{A}$. Without loss of generality assume the former case occurs. Up to exchanging $\mathscr{A}$ and $\mathscr{B}$ we may assume that $\mathscr{A}$ contains at least three distinct elements $a,x,y$. Let $b \in \mathscr{B}-\mathscr{A}$ and $z \in \{1,\ldots,n\}-\mathscr{B}$. Let $C$ be the maximal intransitive subgroup stabilizing $\{b,y,z\}$ and its complement.
\begin{itemize}
\item $(axy) \in (A \cap B \cap A_n)-C$.
\item $(ax)(bz) \in (A \cap C \cap A_n)-B$.
\item $(ax)(by) \in (B \cap C \cap A_n)-A$.
\end{itemize}
We may therefore assume that $\mathscr{A} \not \subseteq \mathscr{B}$ and $\mathscr{B} \not \subseteq \mathscr{A}$. Let $a \in \mathscr{A}-\mathscr{B}$, $b \in \mathscr{B}-\mathscr{A}$. By replacing, if needed, $\mathscr{A}$ or $\mathscr{B}$ or both with their complements we may assume that $|\mathscr{A} \cap \mathscr{B}| \geq (p+1)/2 \geq 4$. Let $x,y,z \in \mathscr{A} \cap \mathscr{B}$. Let $C$ be the maximal intransitive subgroup stabilizing $\{a,b,x\}$ and its complement.
\begin{itemize}
\item $(xyz) \in (A \cap B \cap A_n)-C$.
\item $(ax)(yz) \in (A \cap C \cap A_n)-B$.
\item $(xb)(yz) \in (B \cap C \cap A_n)-A$.
\end{itemize}
\end{enumerate}
\section{Acknowledgements}
We would like to thank the referee for carefully reading a previous version of this paper and having very useful suggestions.
|
\section{Introduction}
There is now considerable evidence that a single fermionic field in
the fundamental of $U(N_F)$ minimally coupled to $U(N_F)$ Chern-Simons
gauge theory at level\footnote{In our conventions the level of a
Chern-Simons theory coupled to fermions is defined to be the level
of the low energy gauge group obtained after deforming the theory
with a fermion mass of the same sign as the fermion level.} $k_F$ is
dual to vector $SU(N_B)$ Wilson-Fisher scalars minimally coupled to
$SU(N_B)$ Chern-Simons gauge theory at level $ k_B=-{\rm sgn}(k_F)N_F$
with $N_B=|k_F|$ \cite{Choudhury:2018iwf}-\cite{Cordova:2017vab}
\footnote{See the introduction to the recent paper
\cite{Choudhury:2018iwf} for a more more detailed description of
earlier work.}. This (almost standard by now) duality asserts that
the two so-called \emph{quasi-fermionic} CFTs i.e.~Chern-Simons gauged
`regular fermions' (RF) and `critical bosons' (CB) - are secretly the same
theory.
It has also been conjectured (see \cite{Minwalla:2015sca} and
references therein) that the `quasi-fermionic' duality of the previous
paragraph follows as the infrared limit of a duality between pairs of
fermionic and bosonic RG flows. The fermionic RG flows are obtained by
starting in the ultraviolet with the Chern-Simons gauged Gross-Neveu
or `critical fermion' (CF) theory and deforming this theory with
relevant operators fine tuned to ensure that the IR end point of the
RG flow is the RF theory. In a similar manner the conjecturally dual
bosonic flows are obtained by starting in the ultraviolet with the
gauged `regular boson' (RB) theory deformed with the fine tuning that
ensures that the RG flow ends in the CB theory.
The UV starting points of the flows described above define dual pairs
of conformal field theories. These RB and CF theories - so-called
\emph{quasi-bosonic} theories - are conjectured to be dual to each
other \footnote{At leading order in large $N$ - the order to which we
work in this paper - the RB and CF theories appear as a line of
fixed points parametrized by the single parameter $x_6$, the
coefficient of the $\phi^6$ coupling of the bosonic theory (see
below for the dual statement in the fermionic theory). In other
words the one parameter set of RB and CF theories (and flows
originating therein) that we study in this paper are actually only
physical at three particular values of the parameter $x_6$. See the
very recent paper \cite{abcd} for a computation of the beta function
for $x_6$ that establishes this point.}. If valid, this conjecture
implies that the set of all RG flows that originate in the RB theory
are dual to the set of all RG flows that originate in the CF theory.
The duality of the pair of specially tuned RG flows of the last
paragraph- those that end in the IR in the quasi-fermionic conformal
field theories - is a special case of this general phenomenon.
Generic RG flows that originate at quasi-bosonic fixed points lead to
gapped phases, or more accurately, phases whose low energy behaviour
is governed by a topological field theory. There are two inequivalent
topological phases. In the unHiggsed phase the bosonic
(resp.~fermionic) theory is governed at long distances by pure
$SU(N_B)_{k_B}$ (resp.~$U(N_F)_{k_F}$) topological field theory with
the two theories being level-rank dual to each other. In the Higgsed
phase the bosonic (resp.~fermionic) theory is governed in the IR by a
pure $SU(N_B-1)_{k_B}$ (resp.~$U(N_F)_{{\rm sgn}\, k_F(|k_F|-1)}$)
Chern-Simons theory with the two topological field theories once again
being level-rank dual to each other. \footnote{As first explained in
\cite{Aharony:2012nh}, the reduction in rank of the bosonic
Chern-Simons theory compared to the unHiggsed phase is a consequence
of the Higgs mechanism in the bosonic field theory. The reduction in
level of the fermionic Chern-Simons theory is a consequence of the
switch in sign of the mass of the fermion - level of the pure
Chern-Simons theory obtained by integrating out a negative mass
fermion is one unit smaller than the level obtained by integrating
out a positive mass fermion. }
The most compelling evidence for the scenarios spelt out above comes
from explicit results of direct all-orders calculations that have been
performed separately in the fermionic and bosonic theories in the
large $N$ limit. In particular, the thermal partition function of
deformed RF and CB theories have both been computed in the unHiggsed
phases to all orders in the 't Hooft coupling, and have been shown to
match exactly with each other for all relevant deformations that end
up in this phase \cite{Minwalla:2015sca} \footnote{A similar
matching has also been performed for the S-matrix in the unHiggsed
phase \cite{Jain:2014nza, Yokoyama:2016sbx}. The generalisation of
this match to the Higgsed phase is also an interesting project, but
one that we will not consider in this paper.}. While impressive,
this matching is incomplete, as the restriction to the unHiggsed
phase covers only half of the phase diagram of these theories.
The authors of \cite{Minwalla:2015sca} (and references therein) were
also able to compute the thermal partition function of the CF theory
in the `Higgsed' phase. However, they were unable to perform the
analogous computation in the bosonic theory in this phase and so were
unable to verify the matching of thermal free energies in this phase. In this paper we fill the gap described above. We present an explicit
all-orders computation of the thermal free energy of the RB theory
in the Higgsed phase. Under duality our final results exactly match
the free energy of the fermionic theory in the Higgsed phase, completing
the large $N$ check of the conjectured duality in a satisfying manner.
At the technical level, the computation described in the previous
paragraph (and presented in detail in section \ref{higgsphase}) is a relatively
straightforward generalisation of the computations presented in the
recent paper \cite{Choudhury:2018iwf}. In \cite{Choudhury:2018iwf} the
large $N$ free energy of the Higgsed phase of the Chern-Simons gauged
Wilson-Fisher boson theory was computed for the first time. As we
describe in much more detail below, the computation of the free energy
in the Higgsed phase of the RB theory can be divided into two
steps. In the first step we compute the thermal free energy (or
equivalently, the gap equation) of the CB theory as a function of its
Higgs vev. We are able to import this computation directly from
\cite{Choudhury:2018iwf}. In the relatively simple second step
carried through in this paper, we derive a second gap equation
that determines the effective value of the Higgs vev.
The second step described at the end of the last paragraph had no
counterpart in \cite{Choudhury:2018iwf}. In the critical bosonic
theory the `classical' potential for the scalar field is infinitely
deep. This potential freezes the magnitude of the scalar field in the
Higgsed phase to its classical minimum even in the quantum theory. It
follows that the Higgs vev is independent of the temperature and has a
simple dependence on the 't Hooft coupling in the critical boson
theory. In the regular boson theory, on the other hand, the classical
potential for the scalar field is finite and receives nontrivial
quantum corrections. The value of the scalar condensate is determined
extremizing the quantum effective action for the scalar field. The
result of this minimisation yields a scalar vev that is a nontrivial
function of both the 't Hooft coupling and the temperature. It
follows that the computations of this paper give us a bonus: we are
able to compute the smooth `quantum effective potential' for the RB
theory as a function of the Higgs vev. More precisely we compute the
quantum effective potential for the composite field
$({\bar \phi \phi})$. In the Higgsed phase and in the unitary gauge
employed in the computations of this paper, this quantity reduces to a
potential - an exact Landau-Ginzburg effective potential - for the
Higgs vev. The extremization of this potential determines the Higgs
vev - and is an equivalent and intuitively satisfying way of obtaining
the gap equations - in the Higgsed phase. The later sections of this
paper - sections \ref{ose} and \ref{ost} are devoted to the study of
the exact quantum effective potential of the theory and its physical
consequences.
Let us denote the expectation value of $({\bar \phi \phi})$ by
$({\bar \phi \phi})_{\rm cl}$. At zero temperature it turns out that
the quantum effective action is non-analytic at
$({\bar \phi \phi})_{\rm cl}=0$. For this reason the domain of the
variable in our effective potential - namely ${\bar \phi \phi}$ -
naturally splits into two regions. We refer to the region
${\bar \phi \phi} >0$ as the Higgsed branch of our effective
potential. On the other hand the region ${\bar \phi \phi} <0$ is the
unHiggsed branch of our effective potential. On the Higgsed branch the
quantum effective potential for ${\bar \phi \phi}$ is simply a quantum
corrected version of the classical potential of the
theory. Classically ${\bar \phi \phi}$ always positive, and so the
potential for the theory on the unHiggsed branch (i.e.~at negative
${\bar \phi \phi}$) has no simple classical limit and is purely
quantum in nature. The extremization of the effective potential on
the Higgsed/unHiggsed branches exactly reproduces the gap equations in
the Higgsed/unHiggsed phases.
In both phases the extrema of this effective potential are of two
sorts; local maxima and local minima. Local maxima clearly describe
unstable `phases'. The instability of these phases has an obvious
semiclassical explanation in the Higgsed phase; it is a consequence of
the fact that we have chosen to expand about a maximum of the
potential for the Higgs vev. In this paper we find an analogous
physical explanation for the instability of the `maxima' in the
unHiggsed phase. In Section \ref{uhp} we use the results for exact
S-matrices in these theories \cite{Jain:2014nza, Yokoyama:2016sbx} to
demonstrate that the `phases' constructed about maxima in the
unHiggsed branch always have bound states of one fundamental particle
(created by $\phi$) and one antifundamental particle (created by
${\bar \phi}$) in the so called `singlet' channel. Moreover we
demonstrate in Section \ref{uhp} that these bound states are always
tachyonic (i.e. have negative squared mass). As a consequence, such
expansion points are maxima in the potential of the field that creates
these bound states (in this case ${\bar \phi } \phi$), explaining the
instability of the corresponding solutions of the gap equation.
It turns out that our effective potential is unbounded from below in
the limit $({\bar \phi \phi})_{\rm cl} \to +\infty$ when
$x_6< \phi_1$. Here $x_6$ is the parameter that governs the $\phi^6$
interaction of the theory defined precisely in \eqref{rst}, and
$\phi_1$ is a particular function of the 't Hooft coupling $\lambda_B$
of this theory listed in \eqref{phi12def}. When
$({\bar \phi \phi})_{\rm cl} \to -\infty$, on the other hand, the
potential turns out to be unbounded from below when $x_6> \phi_2$;
$\phi_2$ is given in \eqref{phi12def}. It follows that the RB theory
is unstable - i.e. does not have a stable vacuum state - if either of
the conditions above are met. Happily it turns out that
$\phi_1<\phi_2$ so that there is a range of values for $x_6$, namely
\begin{equation}
\label{stabrange}
\phi_1 \leq x_6 \leq \phi_2\ ,
\end{equation}
over which the regular boson theory is stable.
The zero temperature phase diagram of the RB theory was worked out in
great detail in the recent paper \cite{abcd}. In order to accomplish
this, the authors of \cite{abcd} evaluated every solution of the gap
equation of the RB theory and then compared their free energies. The
dominant phase at any given values of microscopic parameters is simply
the solution with the lowest free energy; this dominant solution was
determined in \cite{abcd} by performing detailed computations. In
Section \ref{phase} of this paper we demonstrate that the structure
of the phase diagrams presented in \cite{abcd} has a simple intuitive
explanation in terms of the exact Landau-Ginzburg effective potential
for $({\bar \phi} \phi)_{\rm cl}$ described above. As we explain in
Section \ref{phase} below, the general structure of the phase diagram
follows from qualitative curve plotting considerations and can be
deduced without performing any detailed computations. Moreover the
analysis of the current paper has an added advantage; it allows us to
distinguish regions of the phase diagram where the dominant phase is
merely metastable (this happens when $x_6>\phi_2$ or $x_6<\phi_1$)
from regions in the phase diagram in which the dominant solution of
the gap equation is truly stable (this happens in the range
\eqref{stabrange}).
\section{Review of known results}
\subsection{Theories and the conjectured duality map}
The RB theory is defined by the action
\begin{align}
S_B &= \int d^3 x \biggl[i \varepsilon^{\mu\nu\rho}{\kappa_B\over 4\pi}
\mathrm{Tr}( X_\mu\partial_\nu X_\rho -{2 i\over3} X_\mu X_\nu X_\rho)
+ D_\mu \bar \phi D^\mu\phi \nonumber\\
&\qquad\qquad\quad
+m_B^2 \bar\phi \phi + {4\pi { b}_4 \over \kappa_B}(\bar\phi \phi)^2
+ \frac{(2\pi)^2}{\kappa_B^2} \left( x_6^B+1\right) (\bar\phi \phi)^3\biggl],
\label{rst}
\end{align}
while the $\zeta_F$ and $\zeta_F^2$ deformed critical fermion (CF)
theory is defined by the Lagrangian
\begin{align}\label{csfnonlinear}
S_F &= \int d^3 x \bigg[ i \varepsilon^{\mu\nu\rho} {\kappa_F \over 4
\pi} \mathrm{Tr}( X_\mu\partial_\nu X_\rho -{2 i\over3} X_\mu X_\nu
X_\rho) + \bar{\psi} \gamma_\mu D^{\mu} \psi \nonumber\\
&\qquad\qquad\quad
-\frac{4\pi}{\kappa_F}\zeta_F (\bar\psi \psi - {\kappa_F
y_2^2\over4\pi} ) - { 4\pi y_4\over \kappa_F} \zeta_F^2 +
\frac{(2\pi)^2}{\kappa_F^2} x_6^F \zeta_F^3\bigg]\ .
\end{align}
In these formulae
\begin{equation}\label{kkmt}
\kappa_B={\text{sgn}}(k_B) \left( |k_B|+N_B \right), \quad\kappa_F
= {\text{sgn}}(k_F) \left( |k_F|+N_F \right)\ .
\end{equation}
The levels $k_F$ and $k_B$ are defined to be the levels of the WZW
theory dual to the pure Chern-Simons theory (throughout this paper we
work with the dimensional regularisation scheme). For concreteness, in
this paper we will assume that the bosonic theory gauge group
is $SU(N_B)$ while the fermionic gauge group is $U(N_F)$ with `equal'
levels $k_F$ for the $SU(N_F)$ and $U(1)$ parts of the gauge
group. The generalisation to $U(N_B) \leftrightarrow SU(N_F)$ and
$U(N_B) \leftrightarrow U(N_F)$ dualities is straightforward at large
$N$ and will not be explicitly considered in this paper.
In the
rest of this paper we will present our formulae in terms of the 't
Hooft couplings defined by
\begin{equation}\label{tcdef}
\lambda_B= \frac{N_B}{\kappa_B},~~~ \lambda_F= \frac{N_F}{\kappa_F}\ .
\end{equation}
We have already mentioned in the introduction that the two theories
above have been conjectured to be dual to each other under the
level-rank duality map
\begin{equation} \label{lgdm}
N_B=|\kappa_F|-N_F,~~~
\kappa_B= -\kappa_F\ .
\end{equation}
This implies that the bosonic 't Hooft coupling is given in terms of
its fermionic counterpart by
\begin{equation}\label{tcm}
\lambda_B=\lambda_F-{\rm sgn}(\lambda_F), \quad
\end{equation}
The relations \eqref{lgdm} and \eqref{tcm} are expected to hold even
at finite $N$. On the other hand the map between deformations of these
two theories is conjectured to be\footnote{Since $x_6^B = x_6^F$, we
drop the superscript ${}^B$ or ${}^F$ on $x_6$ often in the paper
when referring to this coupling.}
\begin{equation}
x_6^F =x_6^B\ , \quad y_4 = { b}_4\ , \quad
y_2^2 = m_B^2\ .
\label{dualitytransform}
\end{equation}
The above equation \eqref{dualitytransform} is known to hold only in
the large $N$ limit; this relationship may well receive corrections in
a power series expansion in $\frac{1}{N}$.
To end this subsection, let us note that under the field redefinition
\begin{equation}\label{psidef}
\phi= \sqrt{\kappa_B}\, \varphi\ ,
\end{equation}
the action \eqref{rst} turns into
\begin{align}
S_B &= \frac{N_B}{\lambda_B} \int d^3 x \biggl[i \varepsilon^{\mu\nu\rho}{1\over 4\pi}
\mathrm{Tr}( X_\mu\partial_\nu X_\rho -{2 i\over3} X_\mu X_\nu X_\rho) + D_\mu \bar \varphi D^\mu\varphi \nonumber\\
&\qquad\qquad\qquad\qquad + m_B^2 \bar\varphi \varphi + 4\pi { b}_4 (\bar\varphi \varphi)^2
+ (2\pi)^2 \left( x_6^B+1\right) (\bar\varphi \varphi)^3\biggl]\ .
\label{rstr}
\end{align}
It follows immediately that in the limit
\begin{equation}\label{class}
\lambda_B \to 0\ ,\quad m_B^2,~b_4,~x_6 ={\rm fixed}\ ,
\end{equation}
the theory \eqref{rst} should reduce to a nonlinear but classical
theory of the fields $\varphi$ and $X_\mu$. We will return to this
point below.
\subsection{Structure of the thermal partition function}
As explained in e.g.~\cite{Choudhury:2018iwf}, the large $N$ thermal
free energy of either of these theories on $S^2 \times S^1$ can be
obtained following a two step process. In the first step we compute
the free energy of the theory in question on
${\mathbb R}^2 \times S^1$, at a fixed value of the gauge
holonomies around $S^1$. The result is a functional of the holonomy
eigenvalue distribution function $\rho(\alpha)$ and is given by the
schematic equation
\begin{equation}\label{nseff}
e^{-\mc{V}_2 T^2 v[\rho] } = \int_{\mbb{R}^2 \times S^1} [d \phi]\ e^{-S[\phi, \rho]}\ .
\end{equation}
where $\mc{V}_2$ is the volume of two dimensional space and $T$ is the
temperature.
In order to complete the evaluation of the $S^2 \times S^1$ partition
function of interest, in the next step we are instructed to evaluate
the unitary matrix integral
\begin{equation}\label{nioev}
\mc{Z}_{S^2\times S^1} =\int [dU]_{\text{CS}} ~e^{-\mc{V}_2 T^2 v[\rho]}.
\end{equation}
where $[dU]_{\text{CS}}$ is the Chern-Simons modified Haar measure
over $U(N)$ described in \cite{Jain:2013py}.
It was demonstrated in \cite{Jain:2013py} that the thermal partition
functions \eqref{nioev} of the bosonic and fermionic theories agree
with each other in the large $N$ limit provided that under duality
\begin{equation}\label{mv}
v_F[\rho_F]= v_B[\rho_B]\ ,
\end{equation}
where the bosonic and fermionic eigenvalue distribution functions,
$\rho_B$ and $\rho_F$, are related via
\begin{equation}\label{nred}
|\lambda_B| \rho_B (\alpha)+ |\lambda_F| \rho_F(\pi - \alpha) = \frac{1}{2 \pi}.
\end{equation}
In this paper we will evaluate the free energy $v_B[\rho_B]$ of the
bosonic theory in the Higgsed phase and verify \eqref{mv}, thus
establishing the equality of thermal free energies of the RB and CF
theories in the Higgsed phase. We summarise the map between the
parameters \eqref{lgdm}, \eqref{tcm}, \eqref{dualitytransform} and the
holonomy distributions \eqref{nred}:
\begin{align}\label{dualitymapmain}
&N_F = |\kappa_B| - N_B\ ,\quad \kappa_F = - \kappa_B\ ,\quad \lambda_F = \lambda_B - \text{sgn}(\lambda_B)\ ,\nonumber\\
&x_6^F = x_6^B\ ,\quad y_4 = b_4\ ,\quad y_2^2 = m_B^2\ ,\quad |\lambda_B| \rho_B (\alpha)+ |\lambda_F| \rho_F(\pi - \alpha) = \frac{1}{2 \pi}.
\end{align}
In Appendix \ref{review} we provide a comprehensive review of
everything that is known about the large $N$ thermal free energies of
the CF and RB theories (the appendix also contains a formula for a
`three variable off-shell' free energy functional of the CF theory
that is valid in both phases \eqref{ofcrf12}). In the rest of this
section we only present those results that will be of relevance for
the computations in the paper.
The free energy $v_F[\rho_F]$ in the critical fermion theory has been
computed in both fermionic phases in \cite{Minwalla:2015sca}. The
result is given in terms of an auxiliary off-shell free
energy\footnote{ We put a hat over a particular quantity
(e.g.~${\hat c}_F$) to denote the dimensionless version of that
quantity (e.g.~$c_F$) obtained by multiplying by appropriate powers
of the temperature $T$.} (equation \eqref{ofcrf} in Appendix \ref{review})
\begin{align}\label{ofcrfintro}
F_F(c_F, \zeta_F) &=\frac{N_F }{6\pi} \bigg[ \frac{|\lambda_F| - \text{sgn}(\lambda_F)\text{sgn}(X_F)}{|\lambda_F|} \hat{c}_F^3 - \frac{3}{2\lambda_F}\left(\frac{4\pi}{\kappa_F} \hat\zeta_F\right) \hat{c}_F^2 \nonumber \\ &\qquad\qquad + \frac{1}{2\lambda_F}\left(\frac{4\pi}{\kappa_F} \hat\zeta_F\right)^3 + \frac{6\pi {\hat y}_2^2}{\kappa_F\lambda_F}{\hat \zeta}_F -\frac{24 \pi^2 {\hat y}_4}{\kappa_F^2\lambda_F}{\hat \zeta}_F^2 +\frac{24 \pi^3 x^F_6}{\kappa_F^3\lambda_F}{\hat \zeta}_F^3\nonumber \\ &\qquad\qquad - 3 \int_{-\pi}^{\pi}d\alpha \rho_{F}(\alpha)\int_{{\hat c}_F}^{\infty} dy ~y~\left( \ln\left(1+e^{-y-i\alpha}\right) + \ln\left(1+e^{-y+i\alpha}\right)\right)\bigg].
\end{align}
The above free energy is a function of two variables $c_F$ and
$\zeta_F$. Extremizing $F_F$ with respect to these variables and
plugging back in the extremum values gives us the free energy
$v_F[\rho_F]$. The physical interpretation of the variable $c_F$ is
that its value at the extremum of $F_F$ coincides with the pole mass
of the fermion.
The free energy \eqref{ofcrfintro} assumes two different analytic
expressions depending on the sign $\text{sgn}(\lambda_F) \text{sgn}(X_F)$ and
governs the dynamics of the two different phases. The phase in which
${\rm sgn}({ X}_F) \text{sgn}(\lambda_F) = \pm 1$ is referred to as the
unHiggsed phase and the Higgsed phase respectively. In equation
\eqref{ofcrf12} in Appendix \ref{review}, we give an off-shell free energy in terms
of three variables (which include $c_F$ and $\zeta_F$) which is
analytic in all three variables and encompasses the behaviour of both
phases.
The free energy \eqref{ofcrfintro} in the unHiggsed phase of the CF
theory matches the free energy of the regular boson theory in the
unHiggsed phase (equation \eqref{noffb12} in Appendix \ref{review}) computed in
\cite{Minwalla:2015sca} under the duality map
\eqref{dualitymapmain}. The free energy \eqref{ofcrfintro} with
$\text{sgn}(\lambda_F)\text{sgn}(X_F) = -1$ gives a prediction for the regular
boson theory in the Higgsed phase. Applying the duality transformation
\eqref{dualitymapmain} and making the following `field' redefinitions:
\begin{equation}
c_F = c_B\ ,\quad \frac{4\pi \zeta_F}{\kappa_F} = -2\lambda_B\sigma_B\ ,
\end{equation}
we get the following prediction for the free energy in the Higgsed
phase (equation \eqref{ferdualm}):
\begin{align}\label{ferdualmintro}
F_B(c_B,\sigma_B) & = \frac{N_B }{6\pi} \bigg[- \frac{\lambda_B - 2\text{sgn}(\lambda_B)}{\lambda_B} {\hat c}_B^3 - 3 {\hat \sigma}_B (\hat{c}_B^2 - \hat{m}_B^2) + 6{\hat b}_4 \lambda_B{\hat\sigma}_B^2 + (3 x^B_6 + 4)\lambda_B^2{\hat\sigma}_B^3\nonumber \\ &\qquad\qquad + 3 \int_{-\pi}^{\pi}d\alpha \rho_{B}(\alpha)\int_{{\hat c}_B}^{\infty} dy ~y\left( \ln\left(1-e^{-y-i\alpha}\right) + \ln\left(1-e^{-y+i\alpha}\right)\right)\bigg].
\end{align}
The extremum value of the variable $c_B$ corresponds to the pole mass
of the W boson excitation in the Higgsed phase. In the next section we
will independently compute the off-shell free energy of the RB theory,
and demonstrate that our answer agrees with \eqref{ferdualmintro} once
we identify the field $\sigma_B$ with
\begin{equation}\label{mdefnintro}
\sigma_B = 2\pi \frac{ {\bar \phi \phi}}{N_B}\ .
\end{equation}
where ${\bar \phi}$ and $\phi$ respectively stand for the saddle point
values of the corresponding fields denoted by the same letters (recall
these fields have nonzero saddle point values in the Higgsed phase).
\textbf{Note:} Here and in the rest of the paper, we define the
quantities $c_F$ and $c_B$ to be always positive. In other words,
$c_{F,B}$ is shorthand for $|c_{F,B}|$. This is the same
convention used in \cite{Minwalla:2015sca}.
\section{The Higgsed Phase of the regular boson theory}
\label{higgsphase}
\subsection{Lagrangian in Unitary gauge}
Consider the following action for the $SU(N_B)$ regular boson theory:
\begin{equation} \label{regboselag}
\begin{split}
S_{\text{E}}& =\int d^3x \Big[\ i \epsilon^{\mu \nu \rho} \frac{\kappa_B}{4 \pi}\, \mathrm{Tr}(X_\mu \partial_\nu X_\rho-\frac{2 i}{3} X_\mu X_\nu X_\rho) + D_\mu\bar{\phi}D^\mu \phi\\
&\qquad\qquad\quad + m_B^2 \bar\phi \phi+\frac{4\pi b_4}{\kappa_B}(\bar\phi
\phi)^2+\frac{(2\pi)^2}{\kappa_B^2}(x_6^B + 1)(\bar\phi
\phi)^3\Big]\ ,
\end{split}
\end{equation}
with $D_\mu = \partial_\mu - i X_\mu$. The above action can be
reorganised as follows in the Higgsed phase where we anticipate
$\langle \bar\phi \phi \rangle \neq 0$. Following
\cite{Choudhury:2018iwf} we work in the unitary gauge
\begin{equation}\label{unitary}
\phi^i(x) = \delta^{iN_B} \sqrt{|\kappa_B|}\, V(x)\ .
\end{equation}
For future reference we note also that \eqref{unitary} implies the following for the `classical' field $\varphi$ defined in \eqref{psidef}:
\begin{equation}\label{unitaryc}
\varphi^i(x) = \delta^{iN_B} \, \sqrt{{\rm sgn}(\kappa_B)}\, V(x)\ .
\end{equation}
The field $V(x)$ shall be termed the \emph{Higgs field}. The above
gauge choice lets us decompose the gauge field $X_\mu$ as
\begin{equation}
X_\mu = \begin{pmatrix} (A_\mu)^a{}_b - \tfrac{ \delta^a{}_b}{N_B-1} Z_\mu & \tfrac{1}{\sqrt{\kappa_B}}(W_\mu)^a \\ \tfrac{1}{\sqrt{\kappa_B}}(\bar{W}_\mu)_b & Z_\mu \end{pmatrix}\ ,
\end{equation}
where the indices $a, b$ run over $1,\ldots, N_B-1$. In terms of these
variables, the action can be rewritten as follows\footnote{The
notation $ABC$ stands for
$d^3 x \epsilon^{\mu\nu \rho} A_\mu B_\nu C_\rho$.}:
\begin{align}\label{asb}
S_{\text{E}}[A,W,Z,V]&=\frac{i \kappa_B}{4 \pi}\int \mathrm{Tr}(AdA-\tfrac{2i}{3}AAA) \nonumber\\
&\quad +\frac{i }{4 \pi}\int\left(2 \bar{W}_{a}d W^a + \kappa_B Zd Z - 2iZ\bar{W}_a W^a - 2 i \bar{W}_aA^a{}_b W^b\right) \nonumber\\
&\quad +\int d^3x\, (|\kappa_B| V^2 Z_\mu Z^\mu + {\text{sgn}}(\kappa_B) V^2\bar{W}_{a\mu} W^{a\mu}) \nonumber\\
&\quad+|\kappa_B|\int d^3x\left( \partial_\mu V \partial^\mu V + m_B^2 V^2 + 4\pi b_4\text{sgn}(\kappa_B) V^4 + 4\pi^2 (x^B_6 + 1) V^6\right)\ .
\end{align}
\subsection{An effective action for the Higgs field $V$}
We will now compute the thermal partition function of the
regular boson theory in the Higgsed phase, i.e.~we will compute
$v_B[\rho_B]$ defined by
\begin{equation}\label{nseffrb}
e^{-\mc{V}_2 T^2 v_B[\rho_B] } = \int_{\mbb{R}^2 \times S^1}
[dV dW dZ dA]\ e^{-S_{\text{E}}[A,W,Z,V]}\ ,
\end{equation}
where $S_{\text{E}}[A,W,Z,V]$ was defined in \eqref{asb}. For this
purpose it is convenient to break up the effective action
$S_{\text{E}}[A,W,Z,V]$ into two parts
\begin{equation}
\label{betp}
S_{\rm E}[A,W,Z,V]= S_1[A,W,Z,V] + S_2[V]
\end{equation}
where
\begin{align}\label{asb1}
S_1[A,W,Z,V]&=\frac{i \kappa_B}{4 \pi}\int d^3x \ \mathrm{Tr}(AdA-\tfrac{2i}{3}AAA) \nonumber\\
&\quad +\frac{i }{4 \pi}\int \left(2 \bar{W}_a d W^a + \kappa_B ZdZ - 2iZ\bar{W}_a W^a-2 i \bar{W}_aA^a{}_b W^b\right) \nonumber\\
&\quad +\int d^3x\, (|\kappa_B| V^2 Z_\mu Z^\mu + {\text{sgn}}(\kappa_B) V^2\bar{W}^\mu_{a} W^{a}_\mu)\ ,
\end{align}
and
\begin{align}\label{vpot}
S_2[V] &= \int d^3x \Big(|\kappa_B| \partial_\mu V \partial^\mu V + U_{\rm cl}(V)\Big)\ ,\nonumber\\
U_{\rm cl}(V) &= |\kappa_B| m_B^2 V^2 + 4\pi b_4\kappa_B V^4 + 4\pi^2 |\kappa_B| (x^B_6 + 1) V^6\ .
\end{align}
The path integral \eqref{nseff} can be rewritten as
\begin{equation}\label{nseffn}
e^{-\mc{V}_2 T^2 v_B[\rho_B] } =
\int [dV]\, e^{-S_2[V]} \int [dW dZ dA]\, e^{-S_1[A,W,Z,V]}
\end{equation}
Let us first study `inner' path integral i.e.
\begin{equation}
\label{pie}
e^{-\mc{V}_2 T^2 v_{\rm CB}[\rho_B, V]} \equiv \int [dW dZ dA]\, e^{-S_1[A,W,Z,V]}\ ,
\end{equation}
where the right hand side defines the quantity
$v_{\rm CB}[\rho_B, V]$. As far as the path integral in \eqref{pie} is
concerned, $V(x)$ is a background field. The path integral \eqref{pie}
is difficult to evaluate for arbitrary $V(x)$ even in the large $N_B$
limit\footnote{\eqref{pie} is effectively the generating function of
all correlation functions of the dimension two scalar $J_0$ in the
large $N$ critical boson theory, and so contains a great deal of
information.}. This problem simplifies, however, in the special case
that $V(x)$ is a constant. In fact, precisely in this limit, the path
integral \eqref{pie} has been evaluated in the recent paper
\cite{Choudhury:2018iwf}. Luckily, it will turn out that, in the large
$N$ limit, the integral over $V(x)$ in \eqref{nseffn} localises to a
saddle point at which $V(x)$ is constant (see below). As a consequence
we only need the result of the path integral \eqref{pie} for constant
$V(x)$; we are able to read off this result directly from
\cite{Choudhury:2018iwf} which we now pause to recall.
The authors of \cite{Choudhury:2018iwf} studied the critical boson
theory in its Higgsed phase. Working in unitary gauge and following
manipulations essentially identical to those outlined in the previous
subsection, they found that the CB theory in the Higgsed phase can be
rewritten as effective theory of interacting massive $W$ bosons, $Z$
bosons and $SU(N_B-1)$ gauge fields, whose action is given by
\begin{align}\label{oldact}
S_{\text{E}}[A,W,Z]&=\frac{i \kappa_B}{4 \pi}\int \mathrm{Tr}(AdA-\tfrac{2i}{3}AAA) \nonumber\\
&\quad +\frac{i }{4 \pi}\int\left(2 \bar{W}_a d W^a + \kappa_B ZdZ - 2iZ\bar{W}_a W^a-2 i \bar{W}_aA^a{}_b W^b\right) \nonumber\\
&\quad -\int d^3x\, \left(\frac{N_B}{4\pi} m_B^{\text{cri}} Z_\mu Z^\mu + \frac{\lambda_B}{4\pi} m_B^{\text{cri}} \bar{W}_{a\mu} W^{a\mu}\right)\ .
\end{align}
The authors of \cite{Choudhury:2018iwf} were then able to evaluate the
finite temperature partition function for the theory defined by
\eqref{oldact}. Their final result for $v_{\rm CB}[\rho_B]$ is given
as follows. One obtains $v_{\rm CB}[\rho_B]$ by extremizing an
off-shell free energy $F_{\rm CB}(c_B)$ given by
\begin{equation} \label{osfeeh}
\begin{split}
F_{\rm CB}(c_B) &=\frac{N_B}{6\pi} {\Bigg[} -
\frac{ \lambda_B-2{\rm sgn}(\lambda_B) }{\lambda_B}\hat{c}_B^3 +\frac{3}{2} {\hat m}_B^{\rm cri} \hat{c}_B^2 + \Lambda\left({\hat m}_B^{\rm cri}\right)^3\\
&+3 \int_{-\pi}^{\pi} d\alpha\rho_B(\alpha)
\int_{\hat{c}_B}^{\infty} dy\,
y\left(\ln\left(1-e^{-y-i\alpha}\right)+\ln\left(1-e^{-y+i\alpha}\right)
\right)
{\Bigg]},\\
\end{split}
\end{equation}
Here, $\Lambda$ is an undetermined constant; shifts in $\Lambda$
correspond to shifts in the cosmological constant counterterm in the
starting action for the CB theory (see \cite{Choudhury:2018iwf} for a
discussion).
Note that the action $S_1$ in
\eqref{asb} agrees precisely with the action \eqref{oldact} reported
in \cite{Choudhury:2018iwf} if we replace $m_B^{\rm cri}$ by the quantity
\begin{equation}\label{reprule}
m_B^{\text{cri}} = -\frac{4\pi}{|\lambda_B|} V^2\quad\text{with $V$ constant}.
\end{equation}
It follows that for the special case that $V(x)$ is constant, the path
integral \eqref{pie} is given by the extremum value of \eqref{osfeeh}
with the replacement \eqref{reprule} and the path integral over $V$ in
\eqref{nseffn} takes the form
\begin{equation}
\label{piv}
\int [dV] e^{-S_{\rm eff} [V]}\quad\text{with}\quad S_{\rm eff}[V] = S_2[V] + {\cal V}_2 T^2 v_{\rm CB}[\rho_B, V]\ .
\end{equation}
From the expressions for $S_2[V]$,\footnote{In
\cite{Choudhury:2018iwf}, the free energy $v_{{\rm CB}}[\rho_B]$
depended on $m_B^{\rm cri}$ which was a parameter in the
theory. After the replacement \eqref{reprule}, the dependence on the
parameter $m_B^{\rm cri}$ is replaced by a dependence on the field
$V(x)$. We have included an explicit $V$ in the notation for
$v_{\rm CB}[\rho_B, V]$ to highlight this dependence on the Higgs
field $V$. } and $F_B(c_B)$ in \eqref{vpot} and \eqref{osfeeh}, it
is clear that there is an overall factor of $N_B$ in front of the
effective action $S_{\rm eff}[V]$. In the large $N_B$ limit the path
integral over $V$ may be evaluated in the saddle-point
approximation. We expect the dominant minima of the effective action
to occur at constant values of $V$ since the kinetic term
$\partial_\mu V \partial^\mu V$ adds a positive definite piece to the
action. For this reason it is sufficient to have the expression
$v_{\rm CB}[\rho_B, V]$ only at constant $V$. As we have already
explained above, this result is given by extremizing \eqref{osfeeh}
w.r.t.~$c_B$ after making the replacement \eqref{reprule}. It follows
that the final result for $v_B[\rho_B]$ in \eqref{nseffrb} is obtained
by extremizing the regular boson off-shell free energy
$$F_B(c_B, V) = F_{\rm CB}(c_B) + \frac{1}{{\cal V}_2 T^2}S_2[V]\ ,$$ with respect
to both $c_B$ ${\it and }$ $V$.\footnote{It is understood that
$F_{\rm CB}[\rho_B, c_B]$ is evaluated after making the
replacement \eqref{reprule}. } Using the explicit expressions
\eqref{vpot}, \eqref{osfeeh} and \eqref{reprule} we find the following
explicit result for the off-shell free energy of the RB theory:
\begin{align}\label{Feff}
F_B(c_B, V) &= \frac{N_B}{6\pi} {\Bigg[} -
\frac{ \left(\lambda_B-2{\rm sgn}(\lambda_B) \right) }{\lambda_B}\hat{c}_B^3 - \frac{8\Lambda}{|\lambda_B|^3}(2\pi {\hat V}^2)^3\nonumber\\
&\qquad -\frac{3}{|\lambda_B|} (\hat{c}_B^2 - {\hat m}_B^2)2\pi {\hat V}^2 + \frac{6 {\hat b}_4}{\lambda_B}(2\pi{\hat V}^2)^2 + \frac{ ( 3x^B_6 + 3)}{|\lambda_B|}(2\pi {\hat V}^2)^3\nonumber\\
&\qquad+3 \int_{-\pi}^{\pi} \rho_B(\alpha) d\alpha\int_{\hat{c}_B}^{\infty}dy y\left(\ln\left(1-e^{-y-i\alpha}\right)+\ln\left(1-e^{-y+i\alpha}\right) \right)
{\Bigg]}\ .
\end{align}
We compare the result \eqref{Feff} with the prediction
\eqref{ferdualmintro} of duality for the Higgsed phase free energy
given by
\begin{align}\label{Feff1}
F_B(c_B, V)
& = \frac{N_B }{6\pi} \Bigg[- \frac{\lambda_B - 2\text{sgn}(\lambda_B)}{\lambda_B} {\hat c}_B^3 \nonumber\\
&\qquad - \frac{3 }{|\lambda_B|} (\hat{c}_B^2 - \hat{m}_B^2) 2\pi \hat{V}^2 +\frac{6 \hat b_4}{\lambda_B}(2\pi \hat{V}^2)^2 + \frac{(3 x^B_6 + 4)}{|\lambda_B|} (2 \pi \hat{V}^2)^3 \nonumber \\
&\qquad + 3 \int_{-\pi}^{\pi}d\alpha \rho_{B}(\alpha)\int_{{\hat c}_B}^{\infty} dy ~y~\left( \ln\left(1-e^{-y-i\alpha}\right) + \ln\left(1-e^{-y+i\alpha}\right)\right)\Bigg]\ ,
\end{align}
where, to get the above expression, we have used \eqref{unitary} and
\eqref{mdefnintro} to write the field $\sigma_B$ in
\eqref{ferdualmintro} as
\begin{equation}\label{mdef}
\sigma_B = \frac{2\pi V^2}{|\lambda_B|}\ .
\end{equation}
We see that \eqref{Feff} agrees precisely with \eqref{Feff1}
provided we choose the as yet undetermined parameter $\Lambda$ as
\begin{equation}\label{alphapred}
\Lambda = -\frac{1}{8}\lambda_B^2\ .
\end{equation}
In the next subsection we will verify that the result
\eqref{alphapred} - which is so far just a prediction of duality - can
also be obtained by direct computation within the Higgsed CB
theory. The strategy we employ is the following. We first note that
the gap equation corresponding to stationarity of \eqref{Feff} with
respect to ${\hat V}^2$ is
\begin{equation}\label{gappred}
{\hat c}_B^2 - {\hat m}_B^2 + 8\Lambda {\hat \sigma}_B^2 - 4\hat b_4 \lambda_B {\hat \sigma}_B - (3 x_6^B + 3)\lambda_B^2 {\hat \sigma}_B^2 = 0\ .
\end{equation}
where $\sigma_B$ is given in terms of $V^2$ by \eqref{mdef}. The
equation \eqref{gappred} merely simply expresses the condition that
the tadpole of the fluctuation of the scalar field $V$ vanishes when
the field $V$ is expanded around its true solution $v$. In the next
subsection we directly evaluate this `tadpole vanishing condition' in
the RB theory in the Higgsed phase and thereby determine $\Lambda$ by
comparison with \eqref{gappred}.
\subsection{Tadpole cancellation for $V$}
As we have explained above, in the Higgsed phase our scalar field $V$
gets the expectation value $v$. It is useful to define
\begin{equation} \label{deffh}
V(x)= v + H(x)\ .
\end{equation}
The condition that $v$ is the correct vacuum expectation value of
$V(x)$ is equivalent to the condition that the expectation value
(i.e. one point function, i.e. tadpole) of the fluctuation $H(x)$
vanishes. In other words we require that
\begin{equation}\label{nseffH}
\int_{\mbb{R}^2 \times S^1}
[dV dW dZ dA]\ H(x)\ e^{-S_{\text{E}}[A,W,Z,V]} =0\ .
\end{equation}
Using the explicit form of $S_{\text{E}}[A,W,Z,V]$ in \eqref{asb},
equation \eqref{nseffH} can be rewritten as
\begin{equation} \label{exptc} {\rm sgn}(\kappa_B) \langle
\bar{W}_{a\mu}(x) W^{a\mu}(x) \rangle + |\kappa_B| \langle Z_\mu(x)
Z^\mu(x) \rangle \ + \frac{\partial } {\partial (v^2)}U_{\rm cl}(v^2) =0\ .
\end{equation}
where $U_{\rm cl}(V)$ is the potential for the Higgs field $V$ given in
\eqref{vpot} and all expectation values are evaluated about the
`vacuum' where $V(x)=v$. While the first and third terms in
\eqref{exptc} above are both of order $N_B$, it is easily verified
that the second term in this equation - the term proportional to
$\langle Z_\mu (x)^2 \rangle $ is of order unity\footnote{This follows
from the observation that the $Z$ propagator scales like $1/N_B$.}
and so can be dropped in the large $N_B$ limit. At leading order in
the large $N_B$ limit, it follows that the tadpole cancellation
condition \eqref{nseffH} can be rewritten as
\begin{equation} \label{exptcn}
\frac{\lambda_B}{2 \pi {\cal V}_3} \int d^3 x \langle \bar{W}_{a\mu}(x) W^{a\mu}(x) \rangle + \frac{\partial U_{\rm cl}(\sigma_B)} {\partial \sigma_B} =0\ ,
\end{equation}
where we have integrated the equation \eqref{exptc} over spacetime and
have divided the resulting expression by the volume of spacetime
${\cal V}_3$. We have also changed variables to
$\sigma_B = 2\pi v^2 / |\lambda_B|$ defined in \eqref{mdef}. Moving to
momentum space, \eqref{exptcn} turns into
\begin{equation} \label{vgap}
\frac{\lambda_B}{2 \pi} \int \frac{\mathcal{D}^3 p}{(2\pi)^3} \eta^{\mu\nu}\,G^a_{a\mu\nu}(p) + \frac{\partial U_{\rm cl}(\sigma_B)} {\partial \sigma_B} =0\ ,
\end{equation}
where
\begin{equation} \label{propmomn}
\langle \bar{W}_{a\mu}(-p) W^b_\nu (p')\rangle=
G^b_{a\mu\nu}(p)\, (2 \pi)^{3} \delta^{(3)}(p-p') = \delta_a{}^b\ G_{\mu\nu}(p)\, (2 \pi)^{3} \delta^{(3)}(p-p')\ ,
\end{equation}
and the measure $\mathcal{D}$ is the natural measure in momentum space at
finite temperature.\footnote{Explicitly, the notation $\mathcal{D}^3 p$
signifies that we work at finite temperature i.e. on the spacetime
$\mbb{R}^2 \times S^1_\beta$ where the third direction $x^3$ is a
circle of circumference $\beta$. The measure $\mathcal{D}^3 p$ is then given
by
\begin{equation}
\int \frac{\mathcal{D}^3 p}{(2\pi)^3} f(p) = \int \frac{dp_1 dp_2}{(2\pi)^2} \int_{-\pi}^\pi d\alpha\, \rho_B(\alpha) \frac{1}{\beta} \sum_{n=-\infty}^\infty f\left(\frac{2\pi n + \alpha}{\beta}\right)\ ,
\end{equation}
where $\rho_B(\alpha)$ is the distribution of the eigenvalues
$\alpha$ of the gauge field holonomy around $S^1_\beta$.} Happily,
the exact all-orders formula for the propagator $G_{\mu\nu}$ was
computed in \cite{Choudhury:2018iwf}. In Appendix \ref{app} we proceed
to plug the explicit expression for $G_{\mu\nu}$ and evaluate the
first term in \eqref{vgap}. We are able to evaluate all the relevant
summations and integrals analytically, and demonstrate that in our
choice of regularisation scheme \eqref{exptcn} takes the explicit form
\begin{equation}\label{vgapsim}
-\frac{N_B}{2\pi} \left({ c}_B^2-\lambda_B^2{ \sigma}_B^2 \right) + \frac{\partial U_{\rm cl}(\sigma_B)}{\partial { \sigma}_B} = 0\ .
\end{equation}
Recall the expression for $U(\sigma_B)$ from \eqref{vpot}:
\begin{align}
U_{\rm cl}(\sigma_B) &= \frac{N_B}{2\pi} \left( m_B^2 \sigma_B + 2 b_4 \lambda_B \sigma_B^2 + ( x^B_6+1) \lambda_B^2 \sigma_B^3\right)\ .
\end{align}
Plugging this back into \eqref{vgapsim} we find
\begin{equation}\label{vgapfinal}
{ c}_B^2 - { m}_B^2 - 4 {b}_4 \lambda_B {\sigma}_B - (3x_6^B + 4) \lambda_B^2 {\sigma}_B^2= 0\ ,
\end{equation}
Comparing this with the gap equation obtained earlier in
\eqref{gappred}, we see that \eqref{gappred} matches \eqref{vgapfinal}
for the predicted value of $\Lambda = -\lambda_B^2 / 8$ in
\eqref{alphapred} as expected.
\section{A three variable off-shell free energy} \label{ose}
The finite temperature unHiggsed phase is governed by the two-variable
off-shell free energy (equation \eqref{bosoff} in Appendix
\ref{review})
\begin{align}\label{bosoffmain}
F_B(c_B, \tl{\cal S}) &= \frac{N_B}{6\pi} \Bigg[ -{\hat c}_B^3+ 3 {\tilde {\cal S}} \left({\hat c}_B^2-{\hat m}_B^2 \right) + 6 {\hat b_4} \lambda_B {\tilde {\cal S}}^2 - (4+3 x_6^B)\lambda_B^2 {\tilde {\cal S}}^3\nonumber \\
&\qquad+ 3 \int_{-\pi}^{\pi} d\alpha \rho_{B}(\alpha)\int_{{\hat c}_B}^{\infty} dy ~y~\left( \ln\left(1-e^{-y-i\alpha}\right) + \ln\left(1-e^{-y+i\alpha}\right)\right)\Bigg]\ ,
\end{align}
On the other hand, we have demonstrated in this paper that the finite
temperature Higgsed phase is governed by the two-variable off-shell
free energy \eqref{ferdualmintro}. As these are two separate `phases'
of the same theory it is somewhat unsatisfying that the off-shell
`Landau-Ginzburg' free energies used to describe them are
different. The reader may wonder whether there exists a single master
off-shell free energy functional - analytic in all `fields' - which
encompasses the physics of both \eqref{bosoffmain} and
\eqref{ferdualmintro}. At least at the algebraic level there is a simple
affirmative answer to this question as we now describe.
Consider the off-shell free energy
\begin{align}\label{Feffofsh}
F(c_B, \sigma_B, \tl{\cal S}) & = \frac{N_B }{6\pi} \Bigg[- {\hat c}_B^3- 4 {\tilde{\cal S}}^3 \lambda_B^2-3 {\hat c}_B^2 {\hat \sigma}_B-12{\tilde{\cal S}}^2 \lambda_B^2 {\hat \sigma}_B -12{\tilde{\cal S}} \lambda_B^2 {\hat \sigma}_B^2 \nonumber\\
&\qquad\qquad + 6 {\hat c}_B |\lambda_B| ({\tilde{\cal S}} + {\hat \sigma}_B)^2 + 3 \left({\hat m}_B^2 {\hat \sigma}_B + 2\lambda_B {\hat b}_4 {\hat \sigma}_B^2 + \lambda_B^2 x_6^B {\hat \sigma}_B^3\right) \nonumber\\
&\qquad\qquad + 3 \int_{-\pi}^{\pi}d\alpha \rho_{B}(\alpha)\int_{{\hat c}_B}^{\infty} dy ~y~\left( \ln\left(1-e^{-y-i\alpha}\right) + \ln\left(1-e^{-y+i\alpha}\right)\right)\Bigg].
\end{align}
Note that \eqref{Feffofsh} is a function of three `field' variables,
namely $c_B$, ${\tilde{\cal S}}$ and $\sigma_B$. Extremizing
\eqref{Feffofsh} w.r.t. ${\tilde{\cal S}}$, $c_B$ and $\sigma_B$
respectively yields the equations
\begin{align} \label{foll}
&({\tilde{\cal S}}+{\hat \sigma}_B) (-{\hat c}_B + |\lambda_B| ({\tilde{\cal S}}+{\hat \sigma}_B))=0\ ,\nonumber\\
&{\hat c}_B ({\cal S}(c_B)+{\hat \sigma}_B)-|\lambda_B| ({\tilde{\cal S}}+{\hat \sigma}_B)^2=0\ ,\nonumber\\
&{\hat c}_B^2 - {\hat m}_B^2-4 {\hat c}_B |\lambda_B| ({\tilde{\cal S}}+{\hat \sigma}_B) +\lambda_B \left(4 {\tilde{\cal S}}^2 \lambda_B- 4 {\hat b}_4 {\hat \sigma}_B+ 8 \lambda_B {\hat \sigma}_B {\tilde{\cal S}}-3 \lambda_B {\hat \sigma}_B^2 x_6^B \right)=0\ .
\end{align}
The quantity ${\cal S}(c_B)$ that appears in the second of \eqref{foll} is
defined in \eqref{nss}. Off-shell, the objects ${\cal S}(c_B)$ and
${\tilde {\cal S}}$ are completely distinct. ${\cal S}(c_B)$ is a
function of $c_B$ while ${\tilde {\cal S}}$ is an independent
variable. However it is easy to see (by subtracting the first two
equations in \eqref{foll}) that these two quantities are, in fact,
equal on-shell.
Note in particular that the first of \eqref{foll} - the equation that
follows upon extremizing \eqref{Feffofsh} w.r.t. ${\tilde{\cal S}}$ - is
the product of two factors. This equation is satisfied either if
\begin{equation}
\label{foee}
{\tilde{\cal S}}+{\hat \sigma}_B=0\ ,
\end{equation}
or if
\begin{equation}
\label{soee}
-{\hat c}_B + |\lambda_B| ({\tilde{\cal S}}+{\hat \sigma}_B)=0\ .
\end{equation}
(clearly \eqref{foee} and \eqref{soee} cannot simultaneously be obeyed
unless $c_B=0$). Let us first suppose that \eqref{foee} is
obeyed. Using \eqref{foee} to eliminate ${ \sigma}_B$ from
\eqref{Feffofsh} yields an off-shell free energy that now depends only
on ${\tilde{\cal S}}$ and $c_B$. It is easily verified that the resultant
free energy agrees exactly with the two-variable free energy
\eqref{bosoffmain} in the unHiggsed phase. It follows that solutions
of \eqref{foee} parametrize - and govern the physics of - the
unHiggsed phase of the RB theory.
In a similar manner let us now suppose that \eqref{soee} is obeyed in
which case we use it to eliminate ${\tilde{\cal S}}$. It is easily verified
that the resultant two-variable free energy - which depends on $c_B$
and ${ \sigma}_B$ - agrees exactly with \eqref{Feff1} with the
identification \eqref{mdef}:
\begin{equation}\label{frd}
\sigma_B = \frac{2\pi v^2}{|\lambda_B|}\ .
\end{equation}
It follows that solutions of \eqref{soee} parametrize - and govern the
physics of - the Higgsed phase of the RB theory.
The identification \eqref{frd} has a simple explanation. Recall that
the bare mass $m_B^2$ appeared in the action \eqref{rst} as the
coefficient of ${\bar \phi} \phi$. It follows that the Legendre
transform of the free energy of our theory w.r.t.~$m_B^2$
yields the exact quantum corrected effective potential of our theory
as a function of the composite field
$( {\bar \phi} \phi )_{\rm cl}$. This Legendre transform may
be computed by adding the term
$$-m_B^2 ( {\bar \phi} \phi )_{\rm cl}$$ to \eqref{Feffofsh} and
then treating $m_B^2$ as a new dynamical field w.r.t.~which
\eqref{Feffofsh} has to be extremized (of course we also continue to
extremize \eqref{Feffofsh} w.r.t. $c_B$, $\sigma_B$ and ${\tilde {\cal S}}$
as before). Note that the dependence of \eqref{Feffofsh} on $m_B^2$ is
extremely simple; it occurs entirely through the term
$\tfrac{N_B}{2 \pi} \sigma_B m_B^2$. As a consequence, extremizing
w.r.t. $m_B^2$ sets
\begin{equation} \label{frdnn}
{ \sigma}_B= \frac{2 \pi ( {\bar \phi} \phi )_{\rm cl} }{N_B}\ .
\end{equation}
In the Higgsed phase it follows from
$\phi^i = \delta^{i N_B} \sqrt{|\kappa_B|}\, v$ (equation
\eqref{unitary}) that
\begin{equation}
\label{neh}
( {\bar \phi} \phi )_{\rm cl}= |\kappa_B| v^2\ .
\end{equation}
(in obtaining \eqref{neh} we use the fact that the Higgs field $V$ is
effectively classical in the large $N_B$ limit). Inserting
\eqref{neh} into \eqref{frdnn} yields \eqref{frd}. We note, however,
that \eqref{frdnn} is more general than \eqref{frd} because it applies
even in the unHiggsed phase. We will make use of this fact in the
next section.
We have thus found a simple single off-shell free energy - namely
\eqref{Feffofsh} - that captures the physics of both the Higgsed and
the unHiggsed phases. We have also explained that one of the three
variables that appears in this free energy - namely $\sigma_B $ - has
a simple direct physical interpretation given by \eqref{frdnn}. It
follows, in particular, that if we integrate $c_B$ and ${\tilde {\cal S}}$
out from \eqref{Feffofsh}, the resultant free energy (which is a
function of $\sigma_B$) can be reinterpreted as the quantum effective
potential of the theory as a function of $ ({\bar \phi} \phi )_{\rm cl}$.
In the next section we will explicitly undertake this exercise in the
zero temperature limit.
It is easily verified that the duality map \eqref{dualitymapmain}
between parameters together with the field redefinitions
\begin{equation} \begin{split}
\label{dmfr}
&\lambda_B {\tilde {\cal S}} = \lambda_F {\tilde {\cal C}} - \frac{\text{sgn}(\lambda_F)}{2}{\hat c}_F\ ,\quad \lambda_B{ \sigma}_B= -\frac{2 \pi \zeta_F}{\kappa_F }\ ,\quad c_B= c_F\ .
\end{split}
\end{equation}
turns the bosonic off-shell free energy \eqref{Feffofsh} into the
fermionic off-shell free energy \eqref{ofcrf12}. This match captures
the Bose Fermi duality between RB and CF theories at the level of the
complete thermal off-shell free energies of the two theories; note
that each of these off-shell free energies is analytic in all
`fields'. In Appendix \ref{cbl} we investigate the behaviour of our
three-variable off-shell free energy in the so called critical boson
scaling limit of the RB theory.
\section{The exact Landau-Ginzburg effective potential} \label{ost} In
this section we integrate out the variables ${\tilde {\cal S}}$ and $c_B$
out from the effective action \eqref{bosoffmain} and obtain an
off-shell free energy for the field ${\sigma}_B$. We work at zero
temperature throughout this section. In this simple - and
physically especially important - limit we obtain a simple analytic
expression for the resultant free energy as a function of
$\sigma_B$. As we have explained in the previous section, this free
energy is simply related to the quantum effective potential of the RB
theory as a function of the field $({\bar \phi}\phi)_{\rm cl}$.
After having obtained this exact Landau-Ginzburg potential we study
and use it in various ways. First we note that this effective
potential has extrema of two sorts - local maxima and local
minima. Local maxima represent unstable saddle point solutions of the
theory. In the case of the unHiggsed branch (see below) we present an
interpretation of the resultant instability in terms of the tachyonic
bound states of the system. We also use the exact Landau-Ginzburg
effective action that we obtain to understand the zero temperature
phase diagram of the RB theory (as a function of its microscopic
parameters) in a simple and intuitive way. Finally we also make a
prediction for the range of the parameter $x_6$ over which the RB
theory is stable, i.e. has a stable vacuum.
\subsection{An effective potential for ${ \sigma}_B$}
In the zero temperature limit the three-variable off-shell free energy
\eqref{Feffofsh} simplifies to
\begin{align}
F(c_B, \sigma_B, \tl{\cal S}) & = \frac{N_B }{6\pi} \Big[- {\hat c}_B^3- 4 {\tilde{\cal S}}^3 \lambda_B^2-3 {\hat c}_B^2 {\hat \sigma}_B-12{\tilde{\cal S}}^2 \lambda_B^2 {\hat \sigma}_B -12{\tilde{\cal S}} \lambda_B^2 {\hat \sigma}_B^2 \nonumber\\
&\qquad\qquad +6 {\hat c}_B |\lambda_B| ({\tilde{\cal S}}+{\hat \sigma}_B)^2 + 3 \left({\hat m}_B^2 {\hat \sigma}_B +2\lambda_B {\hat b}_4 {\hat \sigma}_B^2 + x_6^B \lambda_B^2 {\hat \sigma}_B^3\right) \Big]\ .
\end{align}
Varying this free energy w.r.t. ${\tilde {\cal S}}$ produces the first of
the gap equations in \eqref{foll} which we repeat here for convenience
\begin{align}
({\tilde{\cal S}}+{\hat \sigma}_B) (-{\hat c}_B + |\lambda_B| ({\tilde{\cal S}}+{\hat \sigma}_B))=0\ .
\end{align}
As we have discussed above, this equation has two solutions
corresponding to the unHiggsed and Higgsed branches:
\begin{align}
\text{unHiggsed}:&\quad \tl{\cal S} = -{\hat \sigma}_B\ ,\qquad \text{Higgsed}:\quad \tl{\cal S} = -{\hat \sigma}_B + \frac{{\hat c}_B}{|\lambda_B|}\ .
\end{align}
Plugging these solutions back into the expression for the free energy,
we have, in the unHiggsed phase,
\begin{align}\label{Fuhig}
F^{({\rm uH})}(c_B, \sigma_B) & = \frac{N_B }{6\pi T^3} \left(-{ c}_B^3 + 4 \lambda_B^2{ \sigma}_B^3 - 3({ c}_B^2 - { m}_B^2) { \sigma}_B + 6 { b}_4 \lambda_B { \sigma}_B^2 + 3 x_6^B \lambda_B^2{ \sigma}_B^3\right)\ ,
\end{align}
and in the Higgsed phase,
\begin{align}\label{Fhig}
F^{({\rm H})}(c_B, \sigma_B) & = \frac{N_B }{6\pi T^3} \left(\tfrac{2 - |\lambda_B|}{|\lambda_B|} { c}_B^3 + 4\lambda_B^2{ \sigma}_B^3 - 3({ c}_B^2 -{ m}_B^2) { \sigma}_B + 6 { b}_4 \lambda_B { \sigma}_B^2 + 3 x_6^B \lambda_B^2{ \sigma}_B^3 \right)\ .
\end{align}
We then extremize the above free energies with respect to $c_B$ to get
\begin{align} \label{solssig}
\text{unHiggsed}:&\quad { c}_B =
-2{ \sigma}_B\ ,\qquad \text{Higgsed}:\quad { c}_B =
\frac{|\lambda_B|}{2 - |\lambda_B|} 2{ \sigma}_B\ .
\end{align}
Recall that $c_B$ is positive by definition. It follows that the
solutions \eqref{solssig} exist only when ${\sigma}_B$ is positive
(negative) in the Higgsed (unHiggsed) phase respectively. Plugging
back the above expressions into the free energies in \eqref{Fuhig} and
\eqref{Fhig}, we get
\begin{align}\label{Fuhigfin}
F^{({\rm uH})}(\sigma_B) & = \frac{N_B }{2\pi T^3} \left[\left((1 + x_6^B) - \frac{4-\lambda_B^2}{3\lambda_B^2} \right) \lambda_B^2 { \sigma}_B^3 + 2 { b}_4 \lambda_B { \sigma}_B^2 + { m}_B^2 { \sigma}_B \right]\ ,
\end{align}
and in the Higgsed phase,
\begin{align}\label{Fhigfin}
F^{({\rm H})}(\sigma_B) & = \frac{N_B }{2\pi T^3} \left[\left((1 + x_6^B) - \frac{|\lambda_B|(4 - |\lambda_B|)}{3(2-|\lambda_B|)^2}\right)\lambda_B^2 { \sigma}_B^3 + 2 { b}_4 \lambda_B { \sigma}_B^2 + { m}_B^2 { \sigma}_B \right]\ .
\end{align}
The quantum effective potential for the field
$(\bar\phi \phi)_{\rm cl}$ is related to the above free energies as
\begin{equation}
U_{\rm eff}((\bar\phi\phi)_{\rm cl}) = T^3 F(\sigma_B)\quad\text{with the replacement}\quad \sigma_B \to \frac{2\pi (\bar\phi\phi)_{\rm cl}}{N_B}\ .
\end{equation}
We continue to use the variable $\sigma_B$ as the argument of the
effective potential $U_{\rm eff}$ to avoid clutter, with the
understanding that all instances of $\sigma_B$ in $U_{\rm eff}$ are to
be replaced with $2\pi (\bar\phi \phi)_{\rm cl} / N_B$. Explicitly, we
have
\begin{empheq}[box=\fbox]{align}\label{LGpot}
U_{\rm eff}(\sigma_B) &= \left\{\arraycolsep=1.4pt\def\arraystretch{2.2}\begin{array}{cl} \frac{N_B}{2 \pi}
\left[ (x_6 - \phi_2) \lambda_B^2 { \sigma}_B^3 + 2 \lambda_B b_4 { \sigma}_B^2 + m_B^2 { \sigma}_B \right] & \quad\text{for }\sigma_B < 0\ , \\ \frac{N_B}{2\pi}\left[( x_6 - \phi_1)\lambda_B^2 { \sigma}_B^3 + 2 \lambda_B{ b}_4 { \sigma}_B^2 + { m}_B^2 { \sigma}_B\right] & \quad\text{for }\sigma_B > 0\ , \end{array}\right.\nonumber\\
&\qquad\ \text{with the replacement}\quad \sigma_B \to \frac{2\pi (\bar\phi \phi)_{\rm cl}}{N_B}\ .
\end{empheq}
The constants $\phi_1$ and $\phi_2$ are given by
\begin{equation}\label{phi12def}
\phi_1 = \frac{4}{3}\left(\frac{1}{(2-|\lambda_B|)^2} - 1\right)\ ,\quad \phi_2 = \frac{4}{3}\left(\frac{1}{\lambda_B^2} - 1\right)\ .
\end{equation}
Observe that the effective potential \eqref{LGpot} is bounded from
below for positive values of $\sigma_B$ if the coefficient of
$\sigma_B^3$ is positive in the second of \eqref{LGpot}, i.e. when
\begin{equation}\label{phi1def}
x_6 > \phi_1\ .
\end{equation}
Similarly, the effective potential is bounded from below for negative
values of $\sigma_B$ if the coefficient of the $\sigma_B^3$ term is
negative in the first of \eqref{LGpot}, i.e. when
\begin{equation}\label{phi2def}
x_6 < \phi_2\ .
\end{equation}
Note that $\phi_1 < \phi_2$.
Note that the terms proportional to $\sigma_B^2$ and $\sigma_B$ are
identical for the two ranges of $\sigma_B$ but the coefficients of the
$\sigma_B^3$ terms are different: this non-analyticity in the cubic
term is what gives a sharp distinction between the Higgsed and
unHiggsed branches of the effective potential at zero temperature.
When we turn on temperature we expect this non-analyticity to be
smoothed out. \footnote{This should be easy to verify - and seems to
follow from the fact that the finite temperature free energy is an
analytic function of its variables - but we have not verified it in
detail.}
We also give a slightly different expression for the Landau-Ginzburg
potential in terms of the variable $c_B$ which is useful for the
analysis of the gap equations as performed in Section 4 of
\cite{abcd}. For this purpose, we substitute back the expressions for
$\sigma_B$ in terms of $c_B$ from \eqref{solssig}:
\begin{align}\label{gappot}
&U_{\rm eff}(c_B) = \left\{\arraycolsep=1.4pt\def\arraystretch{2.2}\begin{array}{cl} \frac{N_B}{2 \pi}
\left[ A_u \frac{ { c}_B^3}{6} + B_{4, u} \frac{{ c}_B^2}{2} - m_B^2 \frac{{ c}_B}{2} \right] & \quad\text{unHiggsed}\ , \\ \frac{N_B(2-|\lambda_B|)}{2\pi|\lambda_B|}\left[-A_h \frac{ c_B^3}{6} - B_{4,h} \frac{c_B^2}{2} + m_B^2 \frac{c_B}{2}\right] & \quad\text{Higgsed }\ . \end{array}\right.
\end{align}
Here, $A_u$, $B_{4,u}$ and
$A_h$, $B_{4,h}$ are constants defined by
\begin{alignat}{2}\label{abdef}
&A_u= 1-\left( 1+ \frac{3 x_6}{4} \right) \lambda_B^2\ ,\quad
&& B_{4,u}= \lambda_B b_4 \ ,\nonumber\\
&A_h=1- \left( 1+\frac{3x_6}{4}\right) (2-|\lambda_B|)^2\ ,\quad &&{B}_{4,h} = - \text{sgn}(\lambda_B)(2-|\lambda_B|) b_4\ .
\end{alignat}
The off-shell variable $c_B$ has the following advantage; its on-shell
value coincides with the pole mass (the \emph{gap}) of the fundamental
excitation in the corresponding phase. We record the gap equations
that follow from extremizing \eqref{gappot} w.r.t~the variable $c_B$:
\begin{align}\label{exactgap}
\text{unHiggsed}:&\quad A_u c_B^2 + 2 B_{4,u} c_B - m_B^2 = 0\ ,\nonumber\\
\text{Higgsed}:&\quad A_h c_B^2 + 2 B_{4,h} c_B - m_B^2 = 0\ .
\end{align}
Solutions to the quadratic equations in \eqref{exactgap} above
correspond to candidates for the Higgsed or unHiggsed phases of the
theory. The very recent paper \cite{abcd} analysed the solutions of
\eqref{exactgap} in detail and used information about the free energy
at these solutions to obtain the phase structure as a function of the
parameters $x_6$, $\lambda_B b_4$ and $m_B^2$. In the next subsection
we will extract the same physical information from the exact
Landau-Ginzburg effective potential \eqref{LGpot}.
\subsection{Higgsed branch} \label{hp}
In this subsection we study the effective potential \eqref{LGpot}
in more detail on the Higgsed branch.
\subsubsection{Potential for the Higgs vev}
We have already noted around \eqref{frd} that the variable $\sigma_B$
has a simple interpretation in terms of the Higgs vev on the Higgsed
branch. Making the replacement \eqref{frd}, i.e.
$\sigma_B = 2\pi v^2 / |\lambda_B|$ in the second line of
\eqref{LGpot} we find
\begin{align}\label{Feff1mm}
U_{\rm eff}(v
& = \frac{N_B}{|\lambda_B|}\left( m_B^2 v^2 + 4\pi \text{sgn}(\lambda_B) b_4 v^4 + \left( (1 + x_6^B) - \tfrac{|\lambda_B|(4-|\lambda_B|)}{3(2 - |\lambda_B|)^2}\right) 4 \pi^2 v^6\right)\ ,
\end{align}
It is easily verified that \eqref{Feff1mm} may also be obtained by
taking the zero temperature limit of \eqref{Feff1} (i.e. dropping the
last line in that formula) and integrating $c_B$ out of that
equation. It follows that the true value of the Higgs vev in the
vacuum is obtained by extremizing \eqref{Feff1mm}.
Note that \eqref{Feff1mm} manifestly reduces to the classical potential
\begin{equation}\label{classpot}
U_{\rm cl}(v) = \frac{N_B}{|\lambda_B|}
\left( m_B^2 v^2 + 4 \pi\text{sgn}(\lambda_B) b_4 v^4 + (2 \pi)^2 (x_6+1) v^6 \right)\ ,
\end{equation}
in the classical limit \eqref{class} (i.e.~$\lambda_B \to 0$ with
$m_B^2$, $b_4$ and $x_6$ fixed) as expected on general
grounds.
\subsubsection{Graphs of $U_{\rm eff}(\sigma_B)$ in various cases}
In this subsubsection we will study the graphs of
$U_{\rm eff}(\sigma_B)$ on the Higgsed branch for various ranges of
values of microscopic parameters. The results of this subsubsection
will prove useful in sketching the phase diagram of the RB theory in
later subsections.
Recall that, on the Higgsed branch, $U_{\rm eff}(\sigma_B)$ is given
by the expression
\begin{equation}\label{ueffsb}
U_{\rm eff}(\sigma_B) = \frac{ N_B }{2\pi}\left(- \tfrac{|\lambda_B|^2}{(2-|\lambda_B|)^2} {A_h}\frac{4 \sigma_B^3}{3} - \tfrac{|\lambda_B|}{(2-|\lambda_B|)} 2{B}_{4, h} \sigma_B^2 + m_B^2 \sigma_B \right)\ ,\quad \sigma_B > 0\ ,
\end{equation}
($A_h$, $B_{4,h}$ were defined in the second line of \eqref{abdef}).
The extremization of $U_{\rm eff}(\sigma_B)$ produces the gap equation
in the second of \eqref{exactgap} which we reproduce here:
\begin{equation}\label{hatc}
{A_h} c_B^2 + 2 { B_{4,h}} c_B -m_B^2=0\ .
\end{equation}
The structure of the curves for $U_{\rm eff}$ above turn out to depend
sensitively on discriminant $D_h$ of this gap equation \eqref{hatc}:
\begin{equation}\label{disch}
D_h = 4( B_{4,h}^2 + m_B^2 A_h)\ .
\end{equation}
As discussed below \eqref{LGpot}, the above effective
potential is bounded below when $x_6> \phi_1$ where $\phi_1$ was
defined in \eqref{phi12def}. In other words $A_h<0$ when $x_6 > \phi_1$ and $A_h>0 $ when $x_6< \phi_1$.
\footnote{Note that $\phi_1$ is an increasing function of
$\lambda_B$. In particular $\phi_1= -1$ for the free theory
($\lambda_B=0$), whereas for the strongly coupled case
($|\lambda_B|=1$) we have $\phi_1=0$. The fact that $\phi_1$ increases
as we increase $|\lambda_B|$ indicates that coupling effects increase
the propensity of our theory to develop a runaway instability along the
$v$ direction.}
We have the following cases as depicted in Figures \ref{higneg} and \ref{higpos}:
\begin{enumerate}
\item \emph{$A_{h}$ is negative}: the potential $U_{\rm eff}$
increases at large $\sigma_B$.
\begin{figure}[htbp]
\begin{center}
\input{higneg.pdf_t}
\caption{Effective potential in the Higgsed phase for $A_h$ negative.}
\label{higneg}
\end{center}
\end{figure}
\begin{enumerate}
\item \emph{$m_B^2$ is positive}:
\begin{enumerate}
\item \emph{$B_{4, h}$ is negative, or $B_{4,h}$ is positive
such $D_h$ is negative }: the potential rises monotonically as
$\sigma_B$ increases from zero to infinity, and there are no nontrivial
positive solutions of the gap equation \eqref{hatc}.
\item \emph{$B_{4, h}$ is positive such that $D_h$ is
positive}: As $\sigma_B$ is increased from zero, $U_{\rm eff}$
initially increases, reaches a local maximum and then decreases,
reaches a minimum and then increases without bound. In this case
the gap equation has two solutions; the larger of which is the
candidate for a stable phase (the smaller solution presumably
describes unstable dynamics since it occurs at a local maximum
of the effective potential).
\end{enumerate}
\item \emph{$m_B^2$ is negative}: For either sign of $B_{4, h}$,
$U_{\rm eff}$ initially decreases, reaches a minimum and then
turns and increases indefinitely. The gap equation has exactly one
legal solution (i.e.~a solution for $c_B$ which is positive) which
is the candidate for a stable phase.
\end{enumerate}
\item \emph{$A_{h}$ is positive}: the potential $U_{\rm eff}$
decreases at large $\sigma_B$.
\begin{figure}[htbp]
\begin{center}
\input{higpos.pdf_t}
\caption{Effective potential in the Higgsed phase for $A_h$ negative.}
\label{higpos}
\end{center}
\end{figure}
\begin{enumerate}
\item \emph{$m_B^2$ is negative}
\begin{enumerate}
\item \emph{$B_{4, h}$ is positive, or $B_{4, h}$ is negative
such that $D_h$ is negative}: The potential decreases
monotonically as $\sigma_B$ increases from zero to infinity, and
there are no nontrivial positive solutions of the gap equation
\eqref{hatc}.
\item \emph{$B_{4, h}$ is negative such that $D_h$ is
positive}: As $\sigma_B$ is increased from zero $U_{\rm eff}$
initially decreases, reaches a local minimum and then increases
till it reaches a local maximum after which it decreases without
bound. In this case the gap equation has two solutions; the
smaller of which is the candidate metastable phase (the larger
solution presumably describes unstable dynamics again since it
occurs at a local maximum of the potential).
\end{enumerate}
\item \emph{$m_B^2$ is positive}: For either sign of $B_{4, h}$,
$U_{\rm eff}$ initially increases, reaches a maximum and then
turns and decreases indefinitely. The gap equation has exactly one
legal solution; this is a local maximum and so presumably
describes an unstable `phase'.
\end{enumerate}
\end{enumerate}
In the last two paragraphs above we have encountered three examples of
solutions (1.a.ii, 2.a.ii, 2.b) to the gap equations that describe
unstable `phases'. For future use we note that these three unstable
solutions are all given by the following root of \eqref{hatc}:
\begin{equation}\label{roeeuns}
c_B= \frac{-B_{4, h}+\sqrt{B_{4,h}^2+A_h m_B^2}}{A_{h}}\ .
\end{equation}
It is also easy to check that the three cases described above are the
only three legal roots of the form \eqref{roeeuns}. In other words
every local maximum of the potential \eqref{Feff1mm} is a root of the
form \eqref{roeeuns}, and a legal root (i.e.~a root for which the RHS
is positive) of the form \eqref{roeeuns} is one of the three `local
maxima' situations described above \footnote{To repeat, these three
cases are as follows. First when $A_{h}$ is negative, $m_B^2$
positive and $B_{4, h}$ positive such that $D_h$ is positive. Second
when $A_h$ is positive, $m_B^2$ is negative and $B_{4, h}$ negative
such that $D_h$ is positive. Lastly when $A_h$ is positive and
$m_B^2$ positive for either sign of $B_{4, h}$. }.
\subsection{unHiggsed branch} \label{uhp}
\subsubsection{Graphs of $U_{\rm eff}(\sigma_B)$ in various cases}
In this subsubsection we plot $U_{\rm eff}(\sigma_B)$ on the
unHiggsed branch for various ranges of microscopic parameters.
We start with the following expression for the Landau-Ginzburg
potential in the unHiggsed branch in terms of the constants $A_u$ and
$B_{4,u}$:
\begin{equation}\label{fuh}
U_{\rm eff}(\sigma_B) = \frac{N_B}{2 \pi}
\left( - A_u \frac{ 4 { \sigma}_B^3}{3} + 2 B_{4, u} { \sigma}_B^2 + m_B^2 { \sigma}_B \right)\ ,
\end{equation}
where $A_u$ and $B_{4,u}$ are as in the first line of \eqref{abdef}.
Now recall that ${\sigma}_B$ is necessarily negative in the unHiggsed
phase. As a consequence \eqref{fuh} may be rewritten as
\begin{equation}\label{fuhn}
U_{\rm eff}(\sigma_B) = \frac{N_B}{2 \pi}
\left( A_u \frac{ 4 |{ \sigma}_B|^3}{3} + 2 B_{4, u}
|{ \sigma}_B|^2 - m_B^2 |{\sigma}_B| \right) \ .
\end{equation}
The gap equation that follows by varying \eqref{fuhn} w.r.t.
${\sigma}_B$ is given in \eqref{exactgap} and is reproduced below:
\begin{equation}\label{hatcu}
{A_u} c_B^2 + 2 B_{4,u} c_B -m_B^2=0\ ,
\end{equation}
where $c_B=2 |{\sigma}_B|$. Note the formal and notational similarity
with the analogous equation \eqref{hatc} in the Higgsed phase. As in
the previous subsection we briefly analyse the behaviour of
\eqref{fuhn} as a function of ${\sigma}_B$ in all the various
cases. We define the discriminant $D_u$ of \eqref{hatcu}:
\begin{equation}\label{discuh}
D_u = 4(B_{4,u}^2 + m_B^2 A_u)\ .
\end{equation}
We then have the following cases as depicted in Figures \ref{unhigpos}
and \ref{unhigneg}:
\begin{enumerate}
\item \emph{$A_{u}$ is positive}: the potential $U_{\rm eff}$
increases at large $|\sigma_B|$.
\begin{figure}[htbp]
\begin{center}
\input{unhigpos.pdf_t}
\caption{Effective potential in the unHiggsed phase for $A_u$ positive.}
\label{unhigpos}
\end{center}
\end{figure}
\begin{enumerate}
\item \emph{$m_B^2$ is negative}:
\begin{enumerate}
\item \emph{$B_{4,u}$ is positive, or $B_{4,u}$ is negative
such that $D_u$ is negative}: the potential rises
monotonically as $|\sigma_B|$ increases from zero to infinity,
and there are no nontrivial positive solutions of the gap
equation \eqref{hatcu}.
\item \emph{$B_{4,u}$ is negative such that $D_u$ is
positive}: As $|\sigma_B|$ is increased from zero,
$U_{\rm eff}$ initially increases up to a local maximum and then
decreases down to a local minimum and then increases without
bound. The gap equation has two solutions the larger of which is
the dominant stable phase (the smaller solution presumably
describes unstable dynamics since it occurs at a local maximum
of the effective potential).
\end{enumerate}
\item \emph{$m_B^2$ is positive}: For either sign of $B_{4,u}$,
$U_{\rm eff}$ initially decreases, reaches a minimum and then
turns and increases indefinitely. The gap equation has exactly one
legal solution which is the dominant stable phase.
\end{enumerate}
\item \emph{$A_{u}$ is negative}: the potential $U_{\rm eff}$
decreases at large $|\sigma_B|$.
\begin{figure}[htbp]
\begin{center}
\input{unhigneg.pdf_t}
\caption{Effective potential in the unHiggsed phase for $A_u$ negative.}
\label{unhigneg}
\end{center}
\end{figure}
\begin{enumerate}
\item \emph{$m_B^2$ is positive}
\begin{enumerate}
\item \emph{$B_{4,u}$ is negative, or $B_{4,u}$ is positive
such that $D_u$ is negative}: The potential decreases
monotonically as $|\sigma_B|$ increases from zero to infinity,
and there are no nontrivial positive solutions of the gap
equation \eqref{hatcu}.
\item \emph{$B_{4,u}$ is positive such that $D_u$ is
positive}: As $|\sigma_B|$ is increased from zero
$U_{\rm eff}$ initially decreases down to a local minimum and
then increases up to a local maximum after which it decreases
without bound. The smaller of the two solutions to the gap
equation \eqref{hatcu} is the `dominant' metastable phase (the
larger solution presumably describes unstable dynamics again
since it occurs at a local maximum of the potential).
\end{enumerate}
\item \emph{$m_B^2$ is negative}: For either sign of $B_{4,u}$,
$U_{\rm eff}$ initially increases, reaches a maximum and then
turns and decreases indefinitely. The gap equation has exactly one
legal solution; this is a local maximum and so presumably
describes an unstable `phase'.
\end{enumerate}
\end{enumerate}
In the paragraphs above we have encountered three examples of
`unstable phases' (1.a.ii, 2.a.ii, 2.b). The solution of the gap
equation \eqref{hatcu} associated with each of these phases is easily
verified to be
\begin{equation}\label{roee}
c_B = \frac{-B_{4, u}-\sqrt{B_{4,u}^2+A_u m_B^2}}{A_{u}}\ .
\end{equation}
Moreover it is also easy to check that every legal (i.e. positive)
solution of the form \eqref{roee} is one of the three local maxima of
the paragraphs described above.
\subsubsection{Explanation for the instability of local maxima}
We have already explained above that
\begin{equation}\label{signi} |{
\sigma}_B|= -\frac{2\pi ({\bar \phi} \phi)_{\rm cl}}{N_B}\
\end{equation}
where the RHS of this equation should be interpreted in quantum rather
than semiclassical terms since semiclassically the variable
$(\bar\phi \phi)_{\rm cl}$ is given by
$\langle \bar\phi\rangle\langle \phi\rangle$ and is positive. The
operator ${\bar \phi} \phi$ however is not necessarily positive (this
follows because the subtraction that is used to give this operator
meaning is not positive). Equation \eqref{signi} effectively asserts
that the unHiggsed branch explores only negative values of the
operator $\bar\phi \phi$. We have noted above the effective potential
as a function of $({\bar \phi} \phi)_{\rm cl}$ has unstable `phases'
that sit at local maxima of the effective potential. In the rest of
this subsection we will present an explanation of these instabilities.
Our proposal for the mechanism of the instability of the `local
maxima' phases is that it is the tachyonic instability of a bound
state of a single fundamental and antifundamental field in the singlet
channel. We claim that the solutions \eqref{roee} are all unstable in
this sense, while none of the stable phases - i.e. the phases that
occur at legal values of
\begin{equation}\label{roeen}
c_B = \frac{-B_{4, u}+\sqrt{B_{4,u}^2+A_u m_B^2}}{A_{u}}\ ,
\end{equation}
suffer from such an instability.
In order to see that this is indeed the case let us recall that bound
states do occur as poles in the S-matrix of a fundamental $\phi$ field
scattering off an antifundamental ${\bar \phi}$ field. Moreover these
poles do sometimes go tachyonic (i.e. their squared mass sometimes
goes below zero). The condition for this to happen can be worked out
by following discussion in Section $4.5$ of \cite{Jain:2014nza} and
Appendix C of \cite{Yokoyama:2016sbx}. The particle-antiparticle
scattering S-matrix has a pole with positive squared mass when
\begin{equation}
4\lambda_B \le \lambda_B^2 (3 x_6^B+4)-4\frac{b_4 \lambda_B}{c_B}\le 4\ .
\end{equation}
When
$$ 4\lambda_B = \lambda_B^2 (3 x_6^B+4)-4\frac{b_4 \lambda_B}{c_B}\ , $$
the pole is at threshold, i.e. $\sqrt{s}=2 c_B$.
On the other hand when
\begin{equation}\label{poth}
\lambda_B^2 (3 x_6^B+4)-4\frac{b_4 \lambda_B}{c_B}=4\ ,
\end{equation}
the pole lies at $\sqrt{s}=0$. Using the definitions \eqref{abdef} for
$A_u$ and $B_{4,u}$ it is easy to see that the condition \eqref{poth}
can be rewritten as
\begin{equation} \label{pothn}
B_{4,u}= - A_u c_B\ .
\end{equation}
(recall that the quantity $c_B$ is positive by definition). When
\begin{equation}\label{unpoth}
\lambda_B^2 (3 x_6^B+4)-4\frac{b_4 \lambda_B}{c_B}>4\ ,
\end{equation}
we have a bound state with negative squared mass, i.e.~a tachyonic
bound state. The condition for the existence of this tachyonic pole is
\begin{equation}\label{ctbs}
A_u c_B \leq -B_{4,u}\ .
\end{equation}
Of course the quantity $c_B$ is not independent of $A_u$ and
$B_{4,u}$ but is determined in terms of these quantities by
the gap equation. The solutions to the gap equation are given by
\begin{equation}\label{roeege}
c_B = \frac{-B_{4, u} \pm \sqrt{B_{4,u}^2+A_u m_B^2}}{A_{u}}
\end{equation}
Inserting these solutions into the condition \eqref{ctbs}, we
find that the condition \eqref{ctbs} is met whenever
\begin{equation}\label{roeege}
-B_{4, u} \pm \sqrt{B_{4,u}^2+A_u m_B^2} \leq -B_{4, u}
\end{equation}
This condition is obeyed by the `minus' branch of solutions
\eqref{roee} but not by the `plus' branch of solutions
\eqref{roeen}. But we have seen above that this is precisely the split
between the local maxima (solutions \eqref{roee}) and local minima
(solutions \eqref{roeen}) of the effective action \eqref{fuhn}. It is
thus natural to identify the tachyonic bound states as the explanation
for the instability of the `minus' branch of solutions \eqref{roee}.
It follows, in other words, that the instabilities in the unHiggsed
phase occur for the same reason as the instabilities in the Higgsed
phase, but for a different field. Unstable Higgsed `phases' occurred
when our solution to the gap equations was at a maximum of the
potential for the field $\phi$. We propose that instabilities in the
unHiggsed phase occur for solutions to the gap equation around maxima
for the field $(\bar\phi \phi)_{\rm cl}$ that is very ostensibly
related to the bound state of ${\bar \phi}$ and $\phi$.
In the case of the Higgsed theory in the $\lambda_B \to 0$ limit
\eqref{class} we obtained a classical theory (with an overall factor
of $\lambda_B^{-1}$ outside the action) in terms of the variable
$\varphi$. In the current unHiggsed context the effective potential
does not have a clear classical limit as $\lambda_B\to 0$. On
solutions to the gap equation that follows from varying \eqref{fuhn}
w.r.t. $|{ \sigma}_B|$, it turns out that ${ \sigma}_B$ and
${\bar \phi} \phi$ (rather than $\lambda_B { \sigma}_B$ and
${\bar \varphi } \varphi$ as in the Higgsed phase) are finite as
$\lambda_B \to 0$. The field $\varphi$ which was the natural classical
variable at weak coupling in the Higgsed phase does not seem to be
useful in the analysis of the unHiggsed branch (this is probably a
reflection of the fact that dynamics is always quantum on this
branch).
\subsection[Landau-Ginzburg Analysis of the zero temperature phase
diagram]{Landau-Ginzburg Analysis of the zero temperature phase
diagram\protect\footnote{This subsection was
worked out in collaboration with O. Aharony.}}\label{phase}
In subsections \ref{hp} and \ref{uhp} we have already explored the
qualitative structure of the Landau-Ginzburg potential \eqref{LGpot},
plotted as a function of $\sigma_B$, separately for $\sigma_B>0$ and
$\sigma_B<0$. In this section we will simply put the analyses of
subsections \ref{hp} and \ref{uhp} together to obtain a global picture
of the Landau-Ginzburg potential as a function of $\sigma_B$ over all
possible ranges of parameters $x_6$, $\lambda_B b_4$ and $m_B^2$. We
reproduce the potential below. Recall that our exact Landau-Ginzburg
potential as a function of $\sigma_B$ (or equivalently, using
\eqref{frdnn}, a function of $(\bar\phi\phi)_{\rm cl}$) is given by
\begin{align}\label{LGpot1}
&U_{\rm eff}(\sigma_B) = \left\{\arraycolsep=1.4pt\def\arraystretch{2.2}\begin{array}{cl} \frac{N_B}{2 \pi}
\left[ (x_6 - \phi_2) \lambda_B^2 { \sigma}_B^3 + 2 \lambda_B b_4 { \sigma}_B^2 + m_B^2 { \sigma}_B \right] & \quad\text{for }\sigma_B < 0\ , \\ \frac{N_B}{2\pi}\left[( x_6 - \phi_1) \lambda_B^2 { \sigma}_B^3 + 2 \lambda_B{ b}_4 { \sigma}_B^2 + { m}_B^2 { \sigma}_B\right] & \quad\text{for }\sigma_B > 0\ . \end{array}\right.
\end{align}
Recall from \eqref{phi12def} that $\phi_1 < \phi_2$. As we have
explored in detail above, the plots of the effective potential are
qualitatively different when $x_6<\phi_1$, $\phi_1 <x_6< \phi_2$ and
$x_6 > \phi_2$. For this reason we analyse these three ranges of
parameters separately. It is also useful to recall the formulae for
the discriminants \eqref{disch} and \eqref{discuh} of the gap
equations \eqref{exactgap} in either phase of the theory:
\begin{align}
D_h &= \frac{4(2-|\lambda_B|)^2}{\lambda_B^2} \left[ (\lambda_B b_4)^2 - \tfrac{3 \lambda_B^2}{4}(x_6 - \phi_1)m_B^2\right]\ ,\nonumber\\
D_u &= 4\left[(\lambda_B b_4)^2 - \tfrac{3\lambda_B^2}{4}(x_6 - \phi_2) m_B^2\right]\ .
\end{align}
\subsubsection{Case I: $x_6> \phi_2$} \label{sso} In this case the
coefficient of $\sigma_B^3$ in the effective potential \eqref{LGpot1}
is positive both when $\sigma_B>0$ and when $\sigma_B<0$. It follows
that $U_{\rm eff}(\sigma_B)$ is an increasing function in the limits
$\sigma_B \to \pm \infty$. Note, in particular, that
$U_{\rm eff}(\sigma_B)$ is unbounded from below at large negative
$\sigma_B$ presumably indicating a runaway instability of the
theory. In other words, the theory has no truly stable phase in this
range of $x_6$. In this subsection we will sketch the `phase diagram'
of the theory, defined as the diagram that tracks the dominant
metastable phase as a function of the relevant parameters. \footnote{We
emphasise that this phase diagram is formal; no phase - not even the
dominant one - is stable. The theory always has a run away
instability to tunnel to large negative values of $\sigma_B$.} In
order to do this we simply plot $U_{\rm eff}(\sigma_B)$ as a function
of $\sigma_B$. The detailed behaviour of the curve
$U_{\rm eff}(\sigma_B)$ at finite values of $\sigma_B$ depends on the
signs and values of $m_B^2$ and $\lambda_B b_4$. We have the
following sub cases.
\begin{figure}[htbp]
\begin{center}
\input{caseI.pdf_t}
\caption{Effective potential for $x_6 > \phi_2$.}
\label{Case I}
\end{center}
\end{figure}
\begin{enumerate}
\item \emph{$m_B^2$ positive.}
\begin{enumerate}
\item \emph{$\lambda_B b_4$ positive with $D_u$ negative} or
\emph{$\lambda_B b_4$ negative with $D_h$ negative}: In this
case $U_{\rm eff}(\sigma_B)$ is a monotonically increasing
function of $\sigma_B$ as depicted in Fig \ref{Case I}(a).
$U_{\rm eff}(\sigma_B)$ has no extrema and so the gap equation has
no solutions.
\item \emph{$\lambda_B b_4$ positive with $D_u$ positive}: In
this case the curve of $U_{\rm eff}(\sigma_B)$ takes the schematic
form depicted in Fig \ref{Case I}(b).
$U_{\rm eff}(\sigma_B)$ has two extrema; a local minimum and a
local maximum both for $\sigma_B<0$, so both in the unHiggsed
branch. The local minimum is the only metastable phase of the
theory (the maximum is unstable) and so is the dominant `phase'.
\item \emph{$\lambda_B b_4$ negative with $D_h$ positive:} The
graph of $U_{\rm eff}(\sigma_B)$ takes the schematic form depicted
in Fig \ref{Case I}(c). $U_{\rm eff}(\sigma_B)$ has two extrema;
a local minimum and a local maximum both for positive $\sigma_B$
so in the Higgsed branch. The local minimum is the only metastable
phase of the theory (the maximum is unstable) and so is the
dominant `phase'.
\end{enumerate}
\item \emph{$m_B^2$ negative, $\lambda_B b_4$ arbitrary:} In this
case the graph of $U_{\rm eff}(\sigma_B)$ versus $\sigma_B$ takes
the schematic form depicted in Fig \ref{Case I}(d). We
have a local maximum at negative $\sigma_B$ (so in the unHiggsed
phase) and a local minimum - so a metastable phase - at positive
$\sigma_B$, so in the Higgsed branch. This local minimum is the
dominant (metastable) phase.
\end{enumerate}
Putting all this together we conclude that our theory has the
(metastable) phase structure depicted in Fig. 8 of \cite{abcd}
and redrawn here for convenience in Fig. \ref{PhaseI} of this paper.
\begin{figure}[htbp]
\begin{center}
\resizebox{0.24\linewidth}{!}{\input{phaseI.pdf_t}}
\caption{Phase diagram for $x_6 > \phi_2$. The positive
$\lambda_B b_4$ axis (shown in blue) corresponds to a
second-order phase transition.}
\label{PhaseI}
\end{center}
\end{figure}
Notice that whenever a metastable phase exists, a subdominant unstable
local maximum of $U_{\rm eff}(\sigma_B)$ also exists in the
vicinity. These are the subdominant `phases' that appear in Fig. 22(a)
of \cite{abcd}.
To end this subsubsection let us study what happens in the limit in
which $x_6 \to \phi_2$ from above. First, nothing special happens to
$U_{\rm eff}(\sigma_B)$ for positive $\sigma_B$. At negative
$\sigma_B$, however, the coefficient of the $\sigma_B^3$ term tends to
zero when $\sigma_B<0$. In this limit $D_u$ is always positive, so the
top half of the red curve in Fig.~\ref{PhaseI} tends to a horizontal
line (the $m_B^2 > 0$ axis). Moreover, when $m_B^2$ and
$\lambda_B b_4$ are both positive (i.e. the case of Fig. \ref{Case
I}(b)) the local minimum (which can be thought of as arising due to
a competition between the linear and quadratic terms in the action)
continues to occur at a fixed value of $\sigma_B$ and the
$U_{\rm eff}(\sigma_B)$ evaluated at this minimum also remains
fixed. But the local maximum of this diagram (which is a result of the
competition between the cubic and quadratic terms in the action) now
occurs at a value of $\sigma_B$ that tends to $-\infty$. Moreover the
value of $U_{\rm eff}(\sigma_B)$ at this maximum also tends to
$\infty$. For $x_6\leq \phi_2$ this local maximum simply does not
exist any more.
\subsubsection{Case II: $\phi_1 < x_6 < \phi_2$} \label{sstw} In this
case, the coefficient of $\sigma_B^3$ is positive for $\sigma_B > 0$
and negative for $\sigma_B < 0$ which implies that the potential is
bounded below for all values of $\sigma_B$, so the theory is stable.
$U_{\rm eff}(\sigma_B)$ is a decreasing function of $\sigma_B$ for
large negative $\sigma_B$, but is an increasing function of $\sigma_B$
for large positive $\sigma_B$.
\begin{figure}[htbp]
\begin{center}
\input{caseII.pdf_t}
\caption{Effective potential for $\phi_1 < x_6 < \phi_2$.}
\label{Case II}
\end{center}
\end{figure}
\begin{enumerate}
\item\emph{$m_B^2$ positive}: \emph{$\lambda_B b_4$ positive or
$\lambda_B b_4$ negative with $D_h$ negative:} In this case the
graph of $U_{\rm eff}(\sigma_B)$ versus $\sigma_B$ takes the form
depicted in Fig \ref{Case II}(a). The global minimum in the
unHiggsed phase (negative $\sigma_B$) is the only extremum of
$U_{\rm eff}(\sigma_B)$; this phase dominates the phase diagram.
\item \emph{$m_B^2$ positive and $\lambda_B b_4$ negative with
$D_h$ positive} or \\ \emph{$m_B^2$ negative and
$\lambda_B b_4$ negative with $D_u$ positive}: In this case the
graph of $U_{\rm eff}(\sigma_B)$ versus $\sigma_B$ takes the form
depicted in Fig \ref{Case II}(b) when $m_B^2$ is positive and of the
form depicted in Fig \ref{Case II}(d) when $m_B^2$ is negative. In
either case the graph has a local minimum in the unHiggsed branch
(negative $\sigma_B$) and a local minimum in the Higgsed branch
(positive $\sigma_B$) separated by a local maximum. The maximum
occurs in the Higgsed branch when $m_B^2>0$ but in the unHiggsed
branch when $m_B^2<0$. \footnote{In fact at $m_B^2=0$ the local
maximum goes through $\sigma_B=0$; this maximum undergoes a
`second order phase transition' at this point from the Higgsed to
the unHiggsed phase. This point is depicted in Figure
\ref{InterII}.} The dominant phase is the local minimum with the
smaller free energy. Which phase dominates depends on the precise
values of $m_B^2$, $\lambda_B b_4$ and $x_6$. A detailed analysis
has been performed in \cite{abcd} and we summarise the results
here. When $x_6$ is strictly between $\phi_1$ and $\phi_2$ the
theory has a first order phase transition line along the curve
\begin{equation}
D_\nu = m_B^2 - \nu_c(x_6) {(\lambda_B b_{4})^2} = 0\ .
\end{equation}
The function $\nu_c(x_6)$ was studied in detail in \cite{abcd}; see
around Fig. 25 and Fig. 26. The function $\nu_c(x_6)$ is monotonically
decreasing as a function of $x_6$ with $x_6 \in (\phi_1 , \phi_2)$.
The function $\nu_c(x_6)$ is negative when $x_6$ is near $\phi_2$
and hence the first order transition line is in the third quadrant
(corresponding to Fig. \ref{PhaseII}(a)). When $x_6$ is near
$\phi_1$, the function $\nu_c(x_6)$ is positive and hence the first
order transition line is in the fourth quadrant
(Fig. \ref{PhaseII}(b)). The phase transition line crosses over to
the fourth quadrant from the third quadrant (equivalently,
$\nu_c(x_6)$ goes from being negative to positive) at some
intermediate value of $x_6$. This intermediate value occurs at
$x_6 = \tfrac{1}{2}(\phi_1 + \phi_2)$ and the phase transition line
coincides with the negative $\lambda_B b_4$ axis. We plot the phase
diagram for this case and also the corresponding Landau-Ginzburg
potential on the phase transition line in Figure
\ref{InterII}. Clearly, we have two exactly equal minima and hence
the onset of a first order phase transition.
\begin{figure}[htbp]
\begin{center}
\resizebox{0.8\linewidth}{!}{\input{phaseII.pdf_t}}
\caption{The phase diagram for $\phi_1 < x_6 < \phi_2$. There is a
second order phase transition (shown in blue) along the positive
$\lambda_B b_4$ axis. The first order phase transition line is
the curve (shown in green) between the two dashed curves. The
precise location of this phase transition curve varies as we
change $x_6$. Two possible locations of this curve have been
sketched in the two figures above. The first figure corresponds
to $x_6$ near $\phi_2$ and the second figure corresponds to
$x_6$ near $\phi_1$.}
\label{PhaseII}
\end{center}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\resizebox{0.8\linewidth}{!}{\input{interII.pdf_t}}
\caption{The first figure is the phase diagram for
$x_6 = \tfrac{1}{2}(\phi_1 + \phi_2)$. The second figure is
the Landau-Ginzburg potential at the same value of $x_6$ for a
point on the first order phase transition line (green)
corresponding to $m_B^2 = 0$ and some $\lambda_B b_4 < 0$.}
\label{InterII}
\end{center}
\end{figure}
Let us study the behaviour in the limits $x_6 \to \phi_1$ and
$x_6 \to \phi_2$. In the limit $x_6 \to \phi_2$ from below, the
unHiggsed branch minimum occurs at $\sigma_B \rightarrow -\infty$
and the potential $U_{\rm eff}(\sigma_B)$ evaluated on this solution
tends to $-\infty$. In this limit the unHiggsed branch local minimum
is the dominant phase for every value of $\lambda_Bb_{4} < 0$ and
$m_B^2$. In the opposite limit $x_6 \to \phi_1$ the Higgsed branch
local minimum occurs at very large values of $\sigma_B$ and
$U_{\rm eff}(\sigma_B)$ evaluated on this solution tends to
$-\infty$. In this limit the Higgsed branch local minimum is the
dominant phase for every value of $\lambda_B b_{4} < 0$ and $m_B^2$.
\item\emph{$m_B^2$ negative:} \emph{$\lambda_B b_4$ positive
or $\lambda_B b_4$ negative with $D_u$ negative:} The graph of
$U_{\rm eff}(\sigma_B)$ versus $\sigma_B$ takes the form depicted in
Fig. \ref{Case II}(c). The global minimum in the Higgsed branch
(positive $\sigma_B$) is the only extremum of
$U_{\rm eff}(\sigma_B)$; this phase dominates the phase diagram.
\end{enumerate}
Putting all this together we arrive at the phase diagram
presented in Fig. 7 of \cite{abcd}. This phase diagram is
resketched in Fig \ref{PhaseII} for convenience.
\subsubsection{Case III: $x_6 < \phi_1$} \label{ssth} In this case,
the coefficient of $\sigma_B^3$ is negative for both $\sigma_B > 0$
and $\sigma_B < 0$. It follows that $U_{\rm eff}(\sigma_B)$ is a
increasing function in the limits $\sigma_B \to \pm \infty$. Note, in
particular, that $U_{\rm eff}(\sigma_B)$ is unbounded from below at
large positive $\sigma_B$ presumably implying a runaway instability of
the theory. In this case the instability is easy to understand as it
is present even in the classical theory at sufficiently negative
values of $x_6$. Just as in Section \ref{sso}, in range of parameters
the RB the theory has no truly stable phases. As in Section \ref{sso},
in this subsubsection we will sketch the `phase diagram' of the theory,
defined as the diagram that tracks the dominant metastable phase
as a function of relevant parameters. As in Section \ref{sso} we read
off our results from plots of $U_{\rm eff}(\sigma_B)$ as a function of
$\sigma_B$. We have the following subcases.
\begin{figure}[htbp]
\begin{center}
\input{caseIII.pdf_t}
\caption{Effective potential for $x_6 < \phi_1$.}
\label{Case 3}
\end{center}
\end{figure}
\begin{enumerate}
\item \emph{$m_B^2$ negative.}
\begin{enumerate}
\item \emph{ $\lambda_B b_4$ negative with $D_u$ negative or
$\lambda_B b_4$ positive with $D_h$ negative:}
$U_{\rm eff}(\sigma_B)$ is a monotonically decreasing function of
$\sigma_B$ as depicted in Fig \ref{Case
3}(a). $U_{\rm eff}(\sigma_B)$ has no extrema, and so the gap
equation has no solutions.
\item \emph{$\lambda_B b_4$ negative with $D_u$ positive:} In
this case the curve of $U_{\rm eff}(\sigma_B)$ takes the schematic
form depicted in Fig. \ref{Case 3}(b).
$U_{\rm eff}(\sigma_B)$ has two extrema; a local minimum and a
local maximum both for $\sigma_B<0$, so both in the unHiggsed
phase. The local minimum is the only metastable phase of the
theory (the maximum is unstable) and so is the dominant `phase'.
\item \emph{ $\lambda_B b_4$ positive with $D_h$ positive:} The
graph of $U_{\rm eff}(\sigma_B)$ takes the schematic form depicted
in Fig \ref{Case 3}(c). $U_{\rm eff}(\sigma_B)$
has two extrema; a local minimum and a local maximum both for
positive $\sigma_B$ so in the Higgsed phase. The local minimum is
the only metastable phase of the theory (the maximum is unstable)
and so is the dominant `phase'.
\end{enumerate}
\item \emph{$m_B^2$ positive, $\lambda_B b_4$ arbitrary:} In this
case the graph of $U_{\rm eff}(\sigma_B)$ versus $\sigma_B$ takes
the form depicted in Fig \ref{Case 3}(d). We have a
local minimum in at negative $\sigma_B$ (so in the unHiggsed phase)
and a local maximum at positive $\sigma_B$, so in the Higgsed
phase. This local minimum is the dominant (metastable) phase.
\end{enumerate}
Putting all this together we conclude that our theory has the
(metastable) phase structure depicted in Fig. 9 of \cite{abcd}
and redrawn here for convenience in Fig. \ref{Phase3} of this paper.
\begin{figure}[htbp]
\begin{center}
\resizebox{0.25\linewidth}{!}{\input{phaseIII.pdf_t}}
\caption{Phase diagram for $x_6 < \phi_1$. The
positive $\lambda_B b_4$ axis (shown in blue)
corresponds to a second-order phase transition.}
\label{Phase3}
\end{center}
\end{figure}
Notice that whenever a metastable phase exists, a subdominant unstable
local maximum of $U_{\rm eff}(\sigma_B)$ also exists in the
vicinity. These are the subdominant `phases' that appear in Fig. 27(b)
of \cite{abcd}.
To end this subsubsection let us study what happens in the limit in
which $x_6 \to \phi_1$ from above. For negative values of $\sigma_B$,
nothing special happens to $U_{\rm eff}(\sigma_B)$. At positive
$\sigma_B$, however, the coefficient of the $\sigma_B^3$ term tends to
zero. In this limit $D_h$ is always positive, so the top half of the
red curve in Fig. \ref{Phase3} tends to a horizontal line (the
negative $m_B^2$ axis). Moreover, when $m_B^2$ and $\lambda_B b_4$
are both positive (i.e. the case of sub Fig. \ref{Case 3}(d)) the
local minimum (which can be thought of as arising due to a competition
between the linear and quadratic terms in the action) continues to
occur at a fixed value of $\sigma_B$ and the $U_{\rm eff}(\sigma_B)$
evaluated at this minimum also remains fixed. But the local maximum of
this diagram (which is a result of the competition between the cubic
and quadratic terms in the action) now occurs at a value of $\sigma_B$
that tends to $\infty$. Moreover the value of $U_{\rm eff}(\sigma_B)$
at this maximum also tends to $\infty$. For $x_6 \geq \phi_1$ this
local maximum simply does not exist any more.
\section{Discussion}
The results of this paper suggest several questions for future
work. First, it would be interesting to generalise the computation of
S-matrices presented in \cite{Jain:2014nza, Yokoyama:2016sbx} to the
Higgsed phase of the Regular Boson theory. The fact that this (and
related) computations may throw light on the dual fermionic
interpretation of the $Z$ boson - as discussed in detail in
\cite{Choudhury:2018iwf} - make it particularly interesting.
One of the most interesting results of this paper is the off-shell
effective action \eqref{LGpot}. It would be interesting to generalise
this result to finite values of temperature and chemical potential and
explicitly observe the smoothing-out of the non-analyticity which was
present at zero temperature. It would be also interesting to compute a
similar action for the theory of one fundamental boson and one
fundamental fermion studied in \cite{Jain:2013gza} and to use this
action to unravel the phase structure of the deformed ${\cal N}=2$
supersymmetric matter Chern-Simons theory with a single chiral
multiplet in the fundamental representation. It is possible that such
an investigation will have interesting interplays with supersymmetry:
for example it may be possible to find a superspace version of
\eqref{LGpot}.
In Section \ref{ose} we have presented a three-variable off-shell free
energy that reproduces the gap equation and thermal free energy of the
regular boson theory. We have also presented a physical interpretation
of the variable $\sigma_B$ that enters this action. It would be
interesting to investigate whether there are interesting off-shell
interpretations of the other dynamical variables - $c_B$ and
${\tilde {\cal S}}$ - that appear in Section \ref{ose}.
Above, we have found a preferred value for the cosmological constant
counterterm $\Lambda$ of the CB theory - one that correctly reproduces
the tadpole condition for the regular boson theory (see
\eqref{alphapred}). It would be interesting to derive
\eqref{alphapred} from a more fundamental physical principle. It would
also be interesting to investigate if this result makes any physical
predictions for the critical boson theory: is it correct, for
instance, to interpret the Legendre transform of \eqref{osfeeh} w.r.t
$m_B^{\rm cri}$ (with $\Lambda$ chosen to have the value
\eqref{alphapred}) as the Coleman-Weinberg potential of the CB theory
w.r.t its dimension two scalar operator $J_0$? It would be interesting
to further investigate this and similar questions, and their
implications.
Finally it would be interesting to generalise the considerations of
this paper - even qualitatively - to finite values of $N_B$. We leave
all these questions for future work.
\acknowledgments
We would like to thank F.~Benini and D.~Radicevic for useful
discussions. We would especially like to thank O. Aharony for
collaboration on section \ref{phase}, for several very useful
discussions over the course of many months and for very useful
comments on a preliminary version of this manuscript. The work of
A.~D., I.~H., L.~J., S.~M., and N.~P. was supported by the Infosys
Endowment for the study of the Quantum Structure of Spacetime.
S.~J. would like to thank TIFR, Mumbai for hospitality during the
completion of the work. The work of S.~J. is supported by the
Ramanujan Fellowship. Finally we would all like to acknowledge our
debt to the steady support of the people of India for research in the
basic sciences.
|
\section{Introduction}
Acoustic reciprocity is a fundamental physical principle stating that sound propagation between two points is independent of the choice of source and receiver \cite{Strutt1871, Fleury2014,Achenbach2003}, and is generally obeyed except for certain specific scenarios. Breaking acoustic reciprocity allows waves to be tailored differently in different directions, including the possibility of one-way sound propagation \cite{Fleury2014,Haberman2016acoustic,Cummer2016}, and could lead to the design of direction-dependent acoustic devices with the potential to aid in numerous acoustical applications, such as vibration isolation, signal processing, acoustic communication, and energy harvesting.
One way to realize acoustic non-reciprocity is by applying a bias that is oddly-symmetric upon time reversal, which has been achieved in moving media \cite{Godin1997,Godin2006}, gyroscopic phononic crystals \cite{Nash2015,Wang2015}, and piezophononic media \cite{Merkel2018}, for example, and effectively establishes `up-stream' and `down-stream' directions for propagating waves. Another means to break reciprocity is nonlinearity, which has been used to create one-way sound propagation via harmonic generation \cite{Boechler2011a,Zhang2016,Bunyan2018}. A third mechanism, which is the subject of the present study, is spatiotemporal modulation of material properties \cite{Cassedy1963,Cassedy1967,Trainiti2016,Vila2017,Nassar2017,Nassar2017a,Nassar2018}. Past studies have demonstrated that effective mechanical properties can be modified using electromagnetic effects in piezoeletric materials \cite{Casadei2012,Chen2014,Chen2016}, magnetorheological elastomers \cite{Danas2012}, and phononic crystals containing electromagnets \cite{Wang2018}.
Of particular interest to the present work is periodic, wave-like modulation caused by purely mechanical, nonlinear deformation, which has the effect of altering the linearized stiffness and/or mass properties of small disturbances propagating in superposition. This behavior is often referred to as small-on-large propagation, and has been of interest for ultrasonic, non-destructive testing \cite{Renaud2009,Zhang2013} and mechanical metamaterials \cite{Bertoldi2008, Goldsberry2018, Amendola2018}. Non-reciprocal elastic wave propagation via nonlinear deformation has previously been achieved in chains of cylinders in Hertzian contact \cite{Chaunsali2016}, where the effective stiffness of waves propagating transversely to the cylinder axes was modulated by dynamically changing the angles between them. In this case, the nonlinear deformation and overlaid nonreciprocal propagation occurred in degrees of freedom that were naturally decoupled (torsional and longitudinal displacements, respectively); that is, relative torsional displacements between the cylinders altered the effective stiffness for longitudinal waves, but did not directly generate noticeable longitudinal displacements on their own.
In this work, we present a discrete spring-mass chain model that achieves modulated elastic properties via small-on-large propagation, where the small and large deformations may occur in the same degrees of freedom. The small and large scales are analyzed via multiple-scale perturbation analysis, which provides more specific details about the accuracy of the small-on-large approximation and its range of validity. We find that where the approximation is valid, the linearized equation describing the small-amplitude signal has the same form as a monatomic spring-mass chain with time-dependent on-site stiffness, and this stiffness can be tuned significantly by varying the geometric and stiffness parameters of the unit cell. The linearized chain model is suitable for theoretical analysis using techniques from prior works \cite{Trainiti2016,Vila2017,Attarzadeh2018}. Finally, we demonstrate the effectiveness of the small-on-large approximation by comparing theoretical results to direct numerical simulations of the fully nonlinear equations of motion. The methods presented herein should aid in the design of more complex and/or continuous structures for manipulating mechanical material properties.
\vfill
\section{Theoretical Model}
For simplicity, we consider a lumped-element model, as wave propagation with geometric nonlinearity can be easily modeled by systems of ordinary differential equations. While the analysis presented herein could be applied to continuous structures in the future, this would likely require more complex computational techniques (e.g. finite element methods). Thus, a discrete system is sufficient to convey our main results in a straightforward manner.
Specifically, we consider longitudinal wave propagation in a periodic, monatomic chain of springs and masses with the unit cell shown in Fig. \ref{fig:schematic}(b). The mass $ m $ at the $ n^{th} $ site is coupled to its nearest neighbors and the left and right movable nodes by linear springs with stiffnesses $ k $, $ k_a $, and $ k_b $, respectively, as shown in Fig. \ref{fig:schematic}(a). The horizontal displacement of the $ n^{th} $ mass is denoted $ u_n(t) $ and the vertical displacement of the node to the right is denoted $ y_n(t) $. The modulation is achieved by prescribing $ y_n(t) $ and allowing displacements on the order of the unit cell height $ h $; this causes significant geometric nonlinearity, which alters the incremental stiffness of the structure.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{ModulatedChain_Schematic_v2}
\caption{Schematics of representative unit cells of the spring-mass chain. (a) Two unit cells with labeled displacements and lumped element parameters. (b) One unit cell with labeled dimensions.}
\label{fig:schematic}
\end{figure}
\subsection{Equation of Motion}
To derive the equation of motion of the unit cell, we form the Lagrangian $ \mathcal{L} = \mathcal{T} - \mathcal{V} $, where $ \mathcal{T} $ and $ \mathcal{V} $ represent the kinetic and potential energies, respectively, and are given by the relations
\begin{eqnarray}\label{eq:TV}
\mathcal{T} &=& \frac{1}{2} m \dot{u}_n^2, \\
\mathcal{V} &=& \frac{1}{2} k \left(u_{n+1} - u_n \right)^2 + \frac{1}{2} k \left(u_{n} - u_{n-1} \right)^2 \nonumber \\
&+& \frac{1}{2} k_b \left(\delta_b - l_b \right)^2 + \frac{1}{2} k_a \left(\delta_a - l_a \right)^2.
\end{eqnarray}
\noindent
Here, $ \delta_a = \sqrt{ (h+y_{n-1})^2 + (a + u_n)^2 } $ and $ \delta_b = \sqrt{ (h + y_{n})^2 + (b - u_n)^2 } $ are the instantaneous lengths of springs $ k_a $ and $ k_b $, and the corresponding un-stretched spring lengths are given by $ l_a = \sqrt{a^2 + h^2} $ and $ l_b = \sqrt{b^2 + h^2} $. Substituting Eq. \eqref{eq:TV} into Lagrange's equation,
\begin{equation}\label{eq:lagrange}
\td{}{t} \left[\pd{\mathcal{L}}{\dot{u}_n}\right] - \pd{\mathcal{L}}{u_n} = 0,
\end{equation}
\noindent
we obtain the equation of motion for the displacement $ u_n $:
\begin{eqnarray}\label{eq:EOM}
m \ddot{u}_n &+& k \left( -u_{n-1} + 2 u_n - u_{n+1} \right) \nonumber \\
&+& k_b \left(\delta_b - l_b\right) \frac{u_n - b}{\delta_b} \nonumber \\
&+& k_a \left(\delta_a - l_a\right) \frac{u_n + a}{\delta_a} = 0.
\end{eqnarray}
\subsection{Small-on-Large Approximation}
We seek to model small-amplitude `signal' waves in $ u_n $ propagating in the presence of a large, slowly-varying `pump' wave generated by nodal displacements $ y_n $. This is achieved using multiple-scale perturbation analysis \cite{Nayfeh1979}, which allows for separation of the pump and signal wave dynamics.
To begin, we non-dimensionalize Eq. \eqref{eq:EOM} by defining the dimensionless variables $ U_n = u_n/h $, $ Y_n = y_n/h $, and $ T = \omega_0 t $, where $ \omega_0 = \sqrt{k/m} $ is a characteristic frequency. Substitution of these expressions into Eq. \eqref{eq:EOM} gives the following dimensionless equation of motion:
\begin{eqnarray}\label{eq:EOM_ND}
\frac{d^2 U_n}{dT^2} &+& \left( -U_{n-1} + 2 U_n - U_{n+1} \right) \nonumber \\
&+& \kappa_b \left(\Delta_b - \lambda_b \right) \frac{U_n - \beta}{\Delta_b} \nonumber \\
&+& \kappa_a \left(\Delta_a - \lambda_a \right) \frac{U_n + \alpha}{\Delta_a} = 0,
\end{eqnarray}
\noindent
where $ \kappa_{(a,b)} = k_{(a,b)}/k $, $ \lambda_{(a,b)} = l_{(a,b)}/h $, $ \alpha =a/h $, $ \beta = b/h $, $ \Delta_a = \sqrt{ (1+Y_{n-1})^2 + (\alpha + U_n)^2 } $, and $ \Delta_b = \sqrt{ (1 + Y_{n})^2 + (\beta - U_n)^2 } $. Next, we vary the nodal displacements $ Y_n $ in a periodic, wave-like fashion with frequency $ \wm $ and wave number $ \qm $, i.e.
\begin{equation}
Y_n = Y_0 \cos{(\theta_n(t))},
\label{eq:Modulation}
\end{equation}
\noindent
where $ \theta_n(t) = \qm nD - \wm t $ is the traveling wave phase and $ Y_0 $ is a constant amplitude. To ensure that $ Y_n $ varies slowly in space and time, we define a small, dimensionless parameter $ \eps << 1 $, and let $ \qm D \propto \eps $ and $ \wm /\omega_0 \propto \eps $. We also define fast and slow time scales $ T_0 = T $ and $ T_1 = \eps T $, respectively, so that
\begin{equation}\label{eq:slowDDT}
\frac{d}{dT} = \frac{d}{dT_0} + \eps \frac{d}{dT_1},
\end{equation}
\noindent
as well as the slow, dimensionless spatial variable $ X_1 = \eps n $. To solve Eq. \eqref{eq:EOM_ND}, we seek a solution of the form
\begin{eqnarray}
U_n &=& P(D \eps n, T_1) + \eps S_n(T_0) \nonumber\\
&=& P(D X_1, T_1) + \eps S_n(T_0), \label{eq:Ansatz1}\\
U_{n \pm 1} &=& P(D \eps(n \pm 1), T_1) + \eps S_{n \pm 1}(T_0) \nonumber \\
&=& P(D(X_1 \pm \eps), T_1) + \eps S_{n \pm 1}(T_0), \label{eq:Ansatz2}
\end{eqnarray}
\noindent
where $ P(X_1, T_1) $ and $ S_n(T_0) $ have the roles of pump and signal waves, respectively, and depend on separate time scales. In choosing this trial solution, we have assumed \textit{a priori} that $ P $ varies slowly, with the same length and time scales as $ Y_n $. To ensure the validity of the approximate solutions that follow, this assumption must be checked for specific parameter sets after the complete solution is obtained.
Substituting Eqs. \eqref{eq:slowDDT} - \eqref{eq:Ansatz2} into Eq. \eqref{eq:EOM_ND} results in the following relationship:
\begin{eqnarray}
&\eps &\left[ \td{^2S_n}{T_0^2} - S_{n+1} + 2S_n - S_{n-1} \right] \nonumber \\
+ &\eps^2 &\left[ \pd{^2P}{T_1^2} - \pd{^2 P}{X_1^2}\right] \nonumber \\
+&\kappa_b &\frac{\Dist{1+Y_n}{\beta - P - \eps S_n} - \lambda_b}{\Dist{1+Y_n}{\beta - P - \eps S_n}} (P + \eps S_n - \beta) \nonumber \\
+&\kappa_a &\frac{\Dist{1+Y_{n-1}}{\alpha + P + \eps S_n} - \lambda_a}{\Dist{1+Y_{n-1}}{\alpha + P + \eps S_n}} (P + \eps S_n +\alpha) \nonumber \\
= &0&, \label{eq:Expansion1}
\end{eqnarray}
\noindent
where we have used the Taylor expansion
\begin{equation}\label{key}
P(D(X_1 \pm \eps), T_1) \approx P(DX_1,T_1) \pm \eps \pd{P}{X_1} + \frac{\eps^2}{2} \pd{^2P}{X_1^2}.
\end{equation}
\noindent
Finally, we expand the nonlinear terms of Eq. \eqref{eq:Expansion1} as Taylor series in $ \eps S_n $ about $ \eps S_n = 0 $, collect terms proportional to each power of $ \eps $, and find the following equations describing the dynamics of the pump and signal waves (accurate to $ \mathcal{O}(\eps) $):
\hfill \\
\noindent
$ \mathcal{O}(1) $:
\begin{eqnarray}
\kappa_b(P_n-\beta)\frac{\tDelta_b-\lambda_b}{\tDelta_b}
+ \kappa_a(P_n+\alpha)\frac{\tDelta_a-\lambda_a}{\tDelta_a} = 0, \label{eq:Order1}
\end{eqnarray}
\noindent
$ \mathcal{O}(\eps) $:
\begin{eqnarray}
\frac{d^2 S_n}{dT_0^2} &+& \left( -S_{n-1} + 2 S_n - S_{n+1} \right) + K_n(P_n)S_n = 0, \nonumber \\
&& \label{eq:OrderEps}\\
K_n(P_n) &=& \kappa_b \frac{\tDelta_b^2(\tDelta_b-\lambda_b) + \lambda_b(P_n - \beta)^2}{\tDelta_b^3} \nonumber \\
&+& \kappa_a \frac{\tDelta_a^2(\tDelta_a-\lambda_a) + \lambda_a(P_n + \alpha)^2}{\tDelta_a^3} \label{eq:Kn},
\end{eqnarray}
\noindent
where $ \tDelta_a = \sqrt{(1+Y_{n-1})^2 + (\alpha+P_n)^2} $ and $ \tDelta_b = \sqrt{(1+Y_n)^2 + (\beta-P_n)^2} $.
\subsection{Effective Linear Chain}
Equation \eqref{eq:OrderEps} has the form of a linear, monatomic spring-mass chain with space- and time-dependent on-site stiffness $ K_n $, as shown schematically in Fig. \ref{fig:linSchematic}. As is evident from Eq. \eqref{eq:Kn}, this on-site stiffness depends on the pump wave $ P_n $, which can calculated for prescribed $ Y_n $ using Eq. \eqref{eq:Order1}.
By varying the geometric and stiffness parameters of the unit cell, we find a wide range of tunability in $ K_n $. The variation of $ K_n $ over one period of the traveling wave phase $ \theta_n $ is shown in Fig. \ref{fig:effStiffness}, where we have numerically solved for $ P_n $ using Eq. \eqref{eq:Order1} with $ Y_n $ given by Eq. \eqref{eq:Modulation}, and substituted the results into Eq. \eqref{eq:Kn}. Cases of varied modulation amplitude, unit cell shape, and unit cell stiffness distribution are shown in Fig. \ref{fig:effStiffness}(a), Fig. \ref{fig:effStiffness}(b), and Fig. \ref{fig:effStiffness}(c), respectively. We note that care must be taken to validate the assumption of a slowly-varying pump wave. Specifically, while the prescribed modulation $ Y_n $ can always be made to vary slowly by enforcing $ \qm D \propto \eps $, the pump wave $ P_n $ and stiffness $ K_n $ depend nonlinearly on $ Y_n $, and therefore may contain harmonics that do not vary slowly in comparison to the signal wave. In situations where significant harmonics are present (e.g. the dark red curves in all three panels of Fig. \ref{fig:effStiffness}, which have sharp changes as a function of the traveling wave phase), $ \qm $ and $ \wm $ must be made sufficiently small for the harmonics to be considered slow as well.
Finally, we remark that strong stiffness modulations can be found in the presence of mechanical instabilities, i.e. buckling. The dark red curves in Fig. \ref{fig:effStiffness}(a) and Fig. \ref{fig:effStiffness}(c) approach zero stiffness at a few points in the cycle due to proximity to a point of instability. While the parameters considered in this work were chosen to keep the effective stiffness everywhere positive, the points of instability can be reached by further increasing the respective parameter variations. These instabilities, while interesting in their own right (see \cite{Kochmann2017} and references therein), would violate the assumptions of a slowly-varying pump wave and are outside the scope of the present study. Nevertheless, buckling structures may be useful in future studies to achieve strong and highly tunable modulations by operating near, but not fully reaching, instability.
\begin{figure}[t]
\centering
\includegraphics[width=0.7\linewidth]{ModulatedChain_Linearized_v1}
\caption{Schematic of one unit cell of a linear, time-dependent, monatomic chain describing the dynamics of the signal wave $ S_n(T_0) $.}
\label{fig:linSchematic}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{Fig_Effstiffness_v1}
\caption{Effective on-site stiffness $ K_n $ as a function of traveling wave phase $ \theta_n $, for (a) variation of nodal displacement amplitude $ Y_0 $ in a symmetric unit cell; (b) variation of unit cell shape $ a/D $ with $ D = 1 $, $ \kappa_a = \kappa_b = 1 $, and $ Y_0 = 0.2 $; and (c) variation of stiffness $ \kappa_a $ with $ \kappa_a+\kappa_b = 2 $, $ a/D = 0.5 $, and $ Y_0 = 0.2 $. Colormaps indicate the value of the varied parameter for each curve.}
\label{fig:effStiffness}
\end{figure}
\section{Non-reciprocal Traveling Waves}
\subsection{Periodic Traveling Wave Solutions}
To find periodic, traveling wave solutions of Eq. \eqref{eq:OrderEps}, we use a Bloch wave expansion approach similar to the one used in Ref. \cite{Trainiti2016}. We assume solutions of the form
\begin{equation}\label{eq:ansatz}
S_n = \ee^{\ii \left(\xi n-\Omega T_0 \right)} \sum_{j=-\infty}^{+\infty} \uh_j \ee^{\ii j\Theta_n \left(T_0 \right)},
\end{equation}
\noindent
where $ \xi $ and $ \Omega $ are a dimensionless wave number and frequency, respectively, and $ \Theta_n(T_0) = \xim n - \Wm T_0 $ is the phase of the traveling wave modulation in terms of the dimensionless parameters $ \xim = \qm D $ and $ \Wm = \wm/\omega_0$. Equation \eqref{eq:ansatz} is an infinite sum of plane waves with wave numbers $ \xi \pm j \xim $ and frequencies $ \Omega \pm j \Wm $, with corresponding complex amplitude coefficients $ \uh_j $. Thus, when the on-site stiffness is modulated in space and time, the $ (\xi, \Omega) $ spectrum is not a `dispersion relation' in the traditional sense, because the presence of one plane wave necessitates the existence of others. The modulated stiffness $ K_n $ must also be represented as a summation of plane waves:
\begin{equation}\label{eq:Kfourier}
K_n(\Theta_n(T_0)) = \sum_{p=-\infty}^{+\infty} \hat{K}_p \ee^{\ii p\Theta_n(T_0)},
\end{equation}
\noindent
where the amplitude coefficients $ \hat{K}_p $ are calculated from the well-known formula for a complex Fourier series:
\begin{eqnarray}\label{Kcoeffs}
\hat{K}_{\pm p} = \frac{1}{2 \pi} \int_{0}^{2 \pi} K(\Theta_n)\ee^{-\ii(\pm p)\Theta_n} d\Theta_n.
\end{eqnarray}
\noindent
Substituting Eqs. \eqref{eq:ansatz} - \eqref{Kcoeffs} into Eq. \eqref{eq:OrderEps} and utilizing the orthogonality of the complex exponential functions to eliminate one of the summations, we find the hierarchy of equations
\begin{eqnarray}\label{eq:Spectrum}
-\left( \Omega + l \Wm \right)^2 \uh_l &+& 4 \sin^2 \left(\frac{\xi + l \xim}{2}\right) \uh_l \nonumber \\
&+& \sum_{j=-\infty}^{+\infty} \hat{K}_{l-j} \uh_j = 0,
\end{eqnarray}
\noindent
where the free index $ l = p+j $ arises from orthogonality. To solve Eq. \eqref{eq:Spectrum}, we must truncate the infinite series in the final term of Eq. \eqref{eq:Spectrum} at some value $ \pm J $. Then, the indicies $ j $ and $ l $ take on integer values $ [-J, -(J-1), ... , 0 , J-1, J] $, and for fixed wavenumber $ \xi $, Eq. \eqref{eq:Spectrum} may be cast as a quadratic eigenvalue problem with eigenvalues $ \Omega $ and eigenvectors $ [\uh_{-J}, ... 0, ..., \uh_J]^T $ \cite{Trainiti2016}. The quadratic eigenvalue problem can be solved numerically using standard computational software (in this work, we have used the {\it polyeig} function in MATLAB).
Example solutions of Eq. \eqref{eq:Spectrum} for $ J = 3 $ and the parameter values $ \alpha = \beta = \kappa_a = \kappa_b = 1 $, $ Y_0 = 0.2 $, $ \xim = 2\pi/10 $, and $ \Wm = 0.1 $, as well as the solution for an un-modulated chain with equivalent mean on-site stiffness, are shown in Fig. \ref{fig:nonRepBands}(a), where non-reciprocity is evident from asymmetry about the $ \xi=0 $ axis. This $ (\xi, \Omega) $ spectrum contains $ 2J+1 $ bands, which appear similar to the single band of the un-modulated chain, tiled along a line of slope $ \cm = \Wm/\xim $ \cite{Nassar2018}. However, Fig. \ref{fig:nonRepBands}(a) does not show the relative amplitudes of the plane waves in the solution, nor does it indicate which plane waves belong to each mode; thus, a more intuitive representation of the spectrum can be found by 1) coloring each $ (\xi, \Omega) $ point according to the magnitude of its eigenvector component, and 2) removing the `tilt' due to the temporal modulation, i.e. plotting $ \Omega - \cm \xi $ on the ordinate axis instead of $ \Omega. $ This representation is shown in Fig. \ref{fig:nonRepBands}(b), from which it can be seen that most of the energy in each mode (a `mode' now corresponding to points on any horizontal line) is concentrated near the un-modulated branch.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{Fig_Theory_v2}
\caption{Theoretical band structure for parameters $ \alpha = \beta = \kappa_a = \kappa_b = 1 $, $ Y_0 = 0.2 $, $ \xim = 2\pi/10 $, and $ \Wm = 0.1 $. (a) Frequency-wave number spectra of the modulated (blue solid curves) and non-modulated (red dashed curves) linear chains. (b) Frequency-wave number spectrum of the modulated linear chain with amplitude visible and tilt removed (color map: normalized eigenvector components $ \uh_l / |\uh| $, in decibels).}
\label{fig:nonRepBands}
\end{figure}
\section{Numerical Results}
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{Fig_Numerical_v3}
\caption{(a) Two-dimensional Fourier transform of the simulated spatiotemporal data (colormap: magnitude of each Fourier coefficient, normalized by the maximum magnitude, in decibels). (b) Fourier transform data from panel (a) with theoretical curves overlaid. (c) Theoretical amplitude-colored spectrum (same data as in Fig. \ref{fig:nonRepBands}(b), shown with tilt restored). }
\label{fig:FFT}
\end{figure}
To demonstrate the effectiveness of the linearized model developed using the small-on-large approximation, we simulate Eq. \eqref{eq:EOM_ND} (the fully nonlinear equations of motion) with the same parameters used in the previous section, using a standard fourth-order Runge-Kutta scheme, and compare the results to the linear theoretical model. The simulation is performed with a chain length of 600 masses and a dimensionless time step of $ \Delta T_0 = 0.01 $, for a duration of $ 8 \times 10^4 $ time steps. With the pump wave present, a broad-band signal wave is imparted to the chain by giving one of the masses an initial dimensionless velocity of 0.01. To isolate the signal wave, we perform a second simulation with the pump wave only, and subtract the result from the first simulation. We note that while the subtraction of the pump wave does not yield the signal wave exactly (due to nonlinearity), it gives an accurate representation of the signal wave as long as the small-on-large approximation is valid. Finally, we perform a two-dimensional Fast Fourier Transform on the resulting spatiotemporal data, giving a numerical frequency-wave number spectrum, which is shown in Fig. \ref{fig:FFT}(a). For comparison, we repeat this spectrum with overlaid theoretical curves from Fig. \ref{fig:nonRepBands}(a) (modulated case) in Fig. \ref{fig:FFT}(b), and include the theoretical, color-coded band structure (i.e. the data from Fig. \ref{fig:nonRepBands}(b) with the tilt restored) in Fig. \ref{fig:FFT}(c). Overall, we find excellent agreement between the fully nonlinear, numerical results and the linearized, small-on-large theoretical model.
\section{Conclusion}
We have developed a model for non-reciprocal elastic wave propagation via modulated stiffness in a discrete spring-mass chain, where the modulation is achieved by applying a slow, nonlinear deformation that alters the effective on-site stiffness in a quasi-static manner. By applying multiple-scale perturbation analysis, we have shown that in the presence of a sufficiently slow, nonlinear deformation (the `pump' wave), a small-amplitude disturbance (the `signal' wave) can be accurately modeled by a linear spring-mass chain with time-dependent properties. This effective linear chain model may be analyzed using existing methods from recent works. By tuning the material and geometric parameters of our unit cell, we have shown that the modulation depth of the effective stiffness is highly variable. In particular, it can be made large (e.g. on the order of its mean value) when operating near mechanical instabilities. Finally, we have demonstrated the effectiveness of the linearized model by comparing the theoretical results to direct numerical simulations of the fully-nonlinear chain, and found excellent agreement.
Opportunities for future studies include applying similar analyses to chains in which the effective inter-site stiffness is also modulated by nonlinearity (e.g. by allowing the displacement-controlled nodes to move horizontally) and to continuous structures (for example, negative-stiffness honeycomb lattices, which have been shown to exhibit large effective property changes under an applied strain \cite{Correa2015,Goldsberry2018}). Methods for optimizing these structures to obtain a targeted non-reciprocal response should be explored. It would also be of interest to revisit the application of homogenization techniques to modulated media in the context of Willis constitutive equations \cite{Willis1981,Nassar2015willis,Muhlestein2017,Sieck2017}, which have been discussed briefly in some prior specific cases \cite{Nassar2017a,Torrent2018}. Finally, experimental realizations of mechanically-modulated structures are also within reach, as highly deformable elastic lattices can be fabricated using additive manufacturing techniques \cite{Raney2015,Bertoldi2017,Kochmann2017,Correa2015}.
\section*{Acknowledgments}
This work supported by National Science Foundation EFRI award no. 1641078 and the postdoctoral fellowship program at Applied Research Laboratories at The University of Texas at Austin.
\vfill
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
Intelligent assistants like Siri and Alexa are increasingly becoming an important part of our daily lives, be it in the household, the workplace or in public places. As these systems become more advanced, we will have them interacting with each other to achieve a particular goal \cite{leviathan2018}. We want these conversations to be interpretable to humans for the sake of transparency and ease of debugging. Having the agents communicate in natural language is one of the most universal ways of ensuring interpretability. This motivates our work on goal-driven agents which interact in coherent language understandable to humans. \gs{Most prior work on visual dialog \cite{visdial,vdialog_old} has approached the problem using supervised learning where}, conditioned on the question - answer pair dialog history, a caption $c$ and the image $I$, the agent is required to answer a given question $q$. The model is trained in a supervised learning framework using ground truth supervision from a human-human dialog dataset.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{flowchart.pdf}
\caption{Multi-Agent (with 1 Q-Bot, 3 A-Bots) Dialog Framework}
\label{fig:diagram}
\end{figure}
\gs{Some recent work \cite{vdialog} has tried to approach the problem using reinforcement learning, with two agents, namely the Question (Q-) Bot and the Answer (A-) Bot. }While the A-Bot still has the image, caption and the dialog history to answer any question, the Question Bot only has access to the caption and the dialog history. \ak{The two agents are initially trained with supervision using the VisDial v0.9 dataset \cite{visdial}, which consists of 80k images, each with a caption and 10 human generated question-answer pairs discussing the image. Under supervision, the agents are trained in an isolated manner to maximize the likelihood of generating the ground truth answers.\gs{ The agents are then made to interact and talk to each other, with a common goal of trying to improve the Q-Bot's understanding of the image.} The agents learn from their conversation with each other via reinforcement learning. While the supervised training \textit{in isolation} helps the agents to learn to interpret the images and communicate information, it is the \textit{interactive} training phase which leads to richer dialog with more informative questions and answers as the agents learn to adapt to each others' strengths and weaknesses. However, it is important to note that the optimization problem in this conversational setting does not make the agents stick to the domain of grammatically correct and coherent natural language.\gs{ Indeed, if the two agents are allowed to communicate and learn from each other for too long, they quickly start generating non-grammatical and semantically meaningless sentences}. While the generated sentences stop making sense\gs{ to human observers, the two agents are able to understand each other much better, and the Q-Bot's understanding of the image improves. This is similar to how twins often develop a private language \cite{rutter2003twins}, an idiosyncratic and} exclusive form of communication understandable only to them. This, however, reduces transparency of the agents' dialog to any observer (human or AI), and is hence undesirable. Prior work \cite{visdial,vdialog} which has focused on improving performance as measured by the Q-Bot's image retrieval rank has suffered from incoherent dialog. We address this problem of improving the agents' performance while increasing dialog quality by taking inspiration from humans. We observe that humans continue to speak in commonly spoken languages, \akr{and hypothesize that this is} \textit{because they need to communicate with an entire community}, \akr{and having a private language for each person would be extremely inefficient}. With this idea, we let our agents learn in a similar setting, by making them talk to (ask questions of, get answers from) multiple agents, one by one.We call this \textit{Community Regularization}.
In the subsequent sections we describe the Visual Dialog task and the neural network architectures of our Q-Bots and A-Bots in detail. We then describe the training process of the agents sequentially: (a) in isolation (via supervision), (b) while interacting with one partner agent (via reinforcement), and (c) our proposed multi-agent setup where each agent interacts with multiple other agents (via reinforcement). We compare the performance of the agents trained in these different settings, both quantitatively using image retrieval ranks, and qualitatively evaluating the overall coherence, grammar and relevance of the dialog generated, as judged by impartial human evaluators. We make the following contributions: we show that community regularization resulting from our multi-agent dialog setup ensures that the interactions between the agents remain grounded in the rules and grammar of natural language, are coherent and human-interpretable \akr{without compromising on task performance}}. We make our code available as open-source\footnote{\url{https://github.com/agakshat/visualdialog-pytorch}}.
\section{Problem Statement}
We begin by defining the problem of Visually Grounded Dialog for the co-operative image guessing game on the VisDial dataset.
\textbf{Players and Roles}: The game involves two collaborative agents – a question bot (Q-bot) and an answer bot (A-bot). \ak{The A-bot has access to an image and caption, while the Q-bot has access to the image's caption, but not the image itself. Both the agents share a common objective, which is for the Q-bot to form a good ``mental representation" of the unseen image which can be used to retrieve, rank or generate that image. This is facilitated by the exchange of 10 pairs of questions and answers between the two agents, using a shared vocabulary, where the Q-bot asks the A-bot a question about the image, and the A-bot answers the question, hence improving the Q-Bot's understanding of the image scene.}
\textbf{General Game Objective}: At each round, in addition to communicating with the A-bot, the Q-bot also provides the learning algorithm with its best estimate $y_t$ of the unknown image $I$ based only on the dialog history and caption. Both agents receive a common reward from the environment which is inversely proportional to the error in this description under some metric $L(y_t, y_{gt})$. We note that this is a general setting where the `description’ $y_t$ can take on varying levels of specificity – from \ak{image feature embeddings extracted by deep neural networks to textual descriptions and pixel-level image re-generations.}
\textbf{Specific Instantiation}: \gs{In our experiments, we focus on the setting where the Q-bot is tasked with estimating a vector embedding of the image I, which is later used to retrieve a similar image from the dataset}. \ak{Given a feature extractor (say, a pretrained CNN model like VGG \cite{vgg}), the target `description’ $y_{gt}$ of the image, can be obtained by simply forward propagating through the VGG model, without the requirement of any human annotation. Reward/error can be measured by the Euclidean distance between the target description $y_{gt}$ and the predicted description $y_t$, and any image may be used as the visual grounding for a dialog. Thus, an unlimited number of games may be simulated without human supervision, motivating the use of reinforcement learning in this framework.
Our primary focus for this work is to ensure that the agents' dialog remains coherent and understandable while also being informative and improving task performance. For concreteness, consider an example of dialog that is informative yet incoherent: \textbf{question}: "do you recognize the guy and age is the adult?", \textbf{answered with}: "you couldn't be late teens, his".} \gs{The example shows that the bots try to extract and convey as much information as possible in a single question/answer (sometimes by incorporating multiple questions or answers into a single statement). But in doing so they lose basic semantic and syntactic structure. We also provide a sample of the dialogs in Figure \ref{fig:humaneval}.}
\begin{figure*}
\centering
\includegraphics[height=0.35\textwidth]{fig_vd_shrunk.pdf}
\caption{A randomly selected image from the VisDial dataset followed by the ground truth (human) and generated dialog about that image for each of our 4 systems (SL, RL-1Q,1A, RL-1Q,3A, RL-3Q,1A). This example was one of 102 used in the human evaluation results shown in Table \ref{table:humaneval}}
\label{fig:humaneval}
\end{figure*}
\section{Related Work}
Most of the major works which combine vision and language have traditionally focused on the problem of image captioning ((\cite{kiros}, \cite{caps1}, \cite{caps2}, \cite{caps3}, \cite{caps4}, \cite{caps5}) and visual question answering (\cite{vqa}, \cite{vqa2}, \cite{vqa3}).
The problem of visual dialog is relatively new and was first introduced by \citet{visdial} who also created the VisDial dataset to advance the research on visually grounded dialog. The dataset was collected by pairing two annotators on Amazon Mechanical Turk to chat about an image. They formulated the task as a `multi-round' VQA task and evaluated individual responses at each round in an image guessing setup. In subsequent work by \citet{vdialog} they proposed a reinforcement learning based setup where they allowed the Question bot and the Answer bot to have a dialog with each other with the goal of correctly predicting the image unseen to the Question bot. However, in their work they noticed that the reinforcement learning based training quickly led the bots to diverge from natural language. In fact \citet{NLemergence} recently showed that language emerging from two agents interacting with each other might not even be interpretable or compositional. We use community regularization to alleviate this problem.
Recent work has also proposed using such goal driven dialog agents for other related tasks including negotiation \cite{deal} and collaborative drawing \cite{codraw}. We believe that our work can easily extend to those settings as well.
\citet{vdialog2} proposed a generative-discriminative framework for visual dialog where they train only an answer bot to generate informative answers for ground truth questions. These answers were then fed to a discriminator, which was trained to rank the generated answer among a set of candidate answers. This is a major restriction of their model as it can only be trained when this additional information of candidate answers is available, which restricts it to a supervised setting. Furthermore, since they train only the answer bot and have no question bot, they cannot simulate an entire dialog which also prevents them from learning by self-play via reinforcement. \citet{vdialog3} further improved upon this generative-discriminative framework by formulating the discriminator as a more traditional GAN \cite{gan}, where the adversarial discriminator is tasked to distinguish between human generated and machine generated dialogs.
\section{Agent Architectures}
\label{sec:arch}
We briefly describe the agent architectures in this section and leave the details for the appendix.
\subsection{Question Bot Architecture}
\ak{\gs{The question bot architecture we use is inspired by the answer bot architecture in \citet{vdialog,vdialog2} but the individual units have been modified to provide more useful representations. Similar to the original architecture, our Q-Bot, shown in Fig. \ref{im:im2}, also consists of 4 parts, (a) fact encoder, (b) state-history encoder, (c) question decoder and (d) image regression network. The fact encoder is modelled using a Long-Short Term Memory (LSTM) network, which encodes the previous question-answer pair into a fact embedding $F_t$. We modify the state-history encoder to incorporate a two-level hierarchical encoding of the dialog. It uses the fact embedding $F_t$ at each time step to compute attention over the history of dialog, $(F_1, F_2, F_3... F_{t-1})$ and produce a history encoding $H_t^Q$. The key modification (compared to \citet{vdialog2}) in our framework is the addition of a separate LSTM to compute a caption embedding $C$. This is key to ensuring that the hierarchical encoding does not exclusively attend on the caption while generating the history embedding. }The caption embedding is then concatenated with $F_t$ and $H_t^Q$, to obtain $S_t^Q$. $S_t^Q$ is then passed through multiple fully connected layers to compute the state-history encoder embedding $e_t^Q$ and the predicted image feature embedding $y_t = f(S_{t}^Q)$.} The encoder embedding, $e_t^Q$ is fed to the question decoder (not pictured), another LSTM, which generates the question, $q_t$. For all LSTMs and fully connected layers in the model we use a hidden layer size of 512. The image feature vector is 4096 dimensional. The word embeddings and the encoder embeddings are 300 dimensional.
\begin{figure}[ht]
\centering
\begin{subfigure}{0.46\textwidth}
\centering
\includegraphics[width=1\textwidth]{qbot_encoder.pdf}
\caption{Encoder architecture for Q-Bot}
\label{im:im2}
\end{subfigure}
\begin{subfigure}{0.46\textwidth}
\centering
\includegraphics[width=1\textwidth]{abot_encoder.pdf}
\caption{Encoder architecture for A-Bot}
\label{im:im1}
\end{subfigure}
\label{fig:encoders}
\caption{}
\end{figure}
\subsection{Answer Bot Architecture}
\gs{The architecture for A-Bot, also inspired from \citet{vdialog2}, shown in Fig. \ref{im:im1}, is similar to that of the Q-Bot. It has 3 components: (a) question encoder, (b) state-history encoder and (c) answer decoder.} The question encoder computes an embedding, $Q_t$ for the question to be answered, $q_t$. The history encoding $(F_1, F_2, F_3... F_{t}) \rightarrow H_t^A$ uses a similar two-level hierarchical encoder, where the attention is computed using the question embedding $Q_t$. The caption is passed on to the A-Bot as the first element of the history, which is why we do not use a separate caption encoder. Instead, we use the fc7 feature embedding of a pretrained VGG-16 \cite{vgg} model to compute the image embedding $I$. The three embeddings $H_t^A, Q_t, y_{gt}$ are concatenated and passed through another fully connected layer to extract the encoder embedding $e_t^A$. The answer decoder (not visualized), which is another LSTM, uses this embedding $e_t^A$ to generate the answer $a_t$. Similar to the Q-Bot, we use a hidden layer size of 512 for all LSTMs and fully connected layers. The image feature vector coming from the CNN is 4096 dimensional (FC7 features from VGG16). The word embeddings and the encoder embeddings are 300 dimensional.
\section{Training}
\label{sec:training}
\gs{We follow the training process proposed in \citet{vdialog}. Two agents, a Q-Bot and an A-Bot are first trained in isolation, by supervision from the VisDial dataset. After this supervised pretraining for 15 epochs over the data, we smoothly transition the agents to learn from each other via reinforcement learning. The individual phases of training will be described in more detail below with some details in the appendix.}
\subsection{Supervised pre-training}
\label{sec:supervised}
In the supervised part of training, both the Q-Bot and A-Bot are trained separately, using a \ak{Maximum Likelihood Estimation (MLE) loss computed using the ground truth questions and answers, respectively, for every round of dialog. The Q-Bot simultaneously optimizes another objective: minimizing the Mean Squared Error (MSE) loss between the true ($y_{gt}$) and predicted ($y_t$) image embeddings. The ground truth dialogs and image embeddings are from the VisDial dataset.}
Given the true dialog history, image features and a question from the dataset, the A-Bot generates an answer to that question. Given the true dialog history and the previous question-answer pair from the dataset, the Q-Bot is made to generate the next question to ask the A-Bot. Both agents receive only ground truth questions and answers, never what the other agent generated - so the two agents never actually interact during this phase of training. \gs{However, there are multiple problems with this training scheme. First, MLE is known to result in models that generate repetitive dialogs and often produce generic responses. Second, since the agents are never allowed to interact during training, they end up facing out of distribution questions and answers when made to interact during evaluation, \gs{which reduces the task performance. This can be observed in Figure \ref{fig:percentile}. The performance of the agents trained via supervised learning dips after each successive dialog round.}}
\subsection{Reinforcement Learning Setup}
\label{sec:rl}
To alleviate the issues pointed out with supervised training, we let the two bots interact with each other via self-play (no ground-truth except images and captions). This interaction is also in the form of questions asked by the Q-Bot, and answered in turn by the A-Bot, using a shared vocabulary. The state space is partially observed and asymmetric, with the Q-Bot observing $\{c,q_{1},a_{1}\ldots q_{10},a_{10} \}$ and the A-Bot observing the same, plus the image $I$. Here, $c $ is the caption, and $q_{i},a_{i}$ is the $i^{th}$ dialog pair exchanged where $i={1\ldots 10}$. The action space for both bots consists of all possible output sequences of a specified maximum length (Q-Bot: 16, A-Bot: 9) under a fixed vocabulary (size 8645). Each action involves predicting words sequentially until a stop token is predicted, or the generated statement reaches the maximum length. Additionally, Q-Bot also produces a guess of the visual representation of the input image (VGG fc-7 embedding of size 4096). Both Q-Bot and A-Bot share the same objective and get the same reward to encourage cooperation. Information gain in each round of dialog is incentivized by setting the reward as the \textbf{change in distance} of the predicted image embedding to the ground-truth image representation. This means that a QA pair is of high quality only if it helps the Q-Bot make a better prediction of the image representation. Both policies are modeled by neural networks, as discussed in Section \ref{sec:arch}.
A dialog round at time $t$ consists of the following steps: 1) the Q-Bot, conditioned on the state encoding, generates a question $q_{t}$, 2) A-Bot updates its state encoding with $q_{t}$ and then generates an answer $a_{t}$, 3) Both Q-Bot and A-Bot encode the completed exchange as a fact embedding, 4) Q-Bot updates its state encoding to incorporate this fact and finally 5) Q-Bot predicts the image representation for the unseen image conditioned on its updated state encoding.
Similar to Das et al. \cite{visdial}, we use the REINFORCE \cite{williams1992simple} algorithm that updates policy parameters in response to experienced rewards. The per-round rewards that are used to calculate the discounted returns follow:
\begin{equation}
r_{t}(s_{t}^{Q},(q_{t},a_{t},y_{t})) = l(y_{t-1},y^{gt}) - l(y_{t},y^{gt})
\label{eq:rewardfn}
\end{equation}
This reward is positive if the distance between image representation generated at time $t$ is closer to the ground truth than the representation at time $t-1$, hence incentivizing information gain at each round of dialog. The REINFORCE update rule ensures that a $(q_{t},a_{t})$ exchange that is informative has its probabilities pushed up. Do note that the image feature regression network $f$ is trained directly via supervised gradient updates on the L-2 loss.
\ak{However, as noted above, this RL optimization problem is ill-posed, since rewarding the agents for information exchange does not motivate them to stick to the rules and conventions of the English language. Thus, we follow an elaborate curriculum scheme described in \cite{visdial}. Specifically, for the first K rounds of dialog for each image, the agents are trained using supervision from the VisDial dataset. For the remaining 10-K rounds, however, they are trained via reinforcement learning. K starts at 9 and is linearly annealed to 0 over 10 epochs. Despite these modifications the bots are still observed to diverge from natural language and produce non-grammatical and incoherent dialog. Thus, we propose a multi bot architecture to help the agents interact in diverse and coherent, yet informative, dialog.}
\subsection{Multi-Agent Dialog Framework (MADF)}
\label{sec:madf}
\gs{In this section we describe our proposed Multi-Agent Dialog architecture in detail. We claim that if, instead of allowing a single pair of agents to interact, we were to make the agents more social, and make them \textit{interact and learn from multiple other agents}, they would be disincentivized to develop a private language, and would have to conform to the common language.} We call this Community Regularization.
\ak{In particular, we create either multiple Q-bots to interact with a single A-bot, or multiple A-bots to interact with a single Q-bot. All these agents are initialized with the learned parameters from the supervised pretraining as described in Section \ref{sec:supervised}. Then, for each batch of images from the VisDial dataset, we randomly choose a Q-bot to interact with the A-bot, or randomly choose an A-bot to interact with the Q-bot, as the case may be. The two chosen agents then have a complete dialog consisting of 10 question-answer pairs about each of those images, and update their respective weights based on the rewards received (as per Equation \ref{eq:rewardfn}) during the conversation, using the REINFORCE algorithm. This process is repeated for each batch of images, over the entire VisDial dataset. It is important to note that histories are \textit{not shared} across batches.} MADF can be understood in detail using the pseudocode in Algorithm \ref{alg:multibot}.
\begin{algorithm*}
\small
\caption{Multi-Agent Dialog Framework (MADF)}\label{alg:multibot}
\begin{algorithmic}[1]
\Procedure{MultiBotTrain}{}
\While{train\_iter < max\_train\_iter}\Comment{Main Training loop over batches}
\State $Qbot \gets \textit{random\_select }(Q_1, Q_2, Q_3.... Q_q)$
\State $Abot \gets \textit{random\_select }(A_1, A_2, A_3.... A_a)$\Comment{Either $q$ or $a$ is equal to 1}
\State $history \gets (0,0,...0)$\Comment{History initialized with zeros}
\State $fact \gets (0,0,...0)$\Comment{Fact encoding initialized with zeros}
\State $\Delta image\_pred \gets 0$ \Comment{Tracks change in Image Embedding}
\State $Qz_{1} \gets Ques\_enc(Qbot,fact,history,caption)$
\For{t in 1:10}\Comment{Have 10 rounds of dialog}
\State $ques_{t} \gets Ques\_gen(Qbot, Qz_{t})$
\State $Az_{t} \gets Ans\_enc(Abot,fact,history,image,ques_{t},caption)$
\State $ans_{t} \gets Ans\_gen(Abot, Az_{t})$
\State $fact \gets [ques_{t}, ans_{t}]$ \Comment{Fact encoder stores the last dialog pair}
\State $history \gets concat(history, fact)$ \Comment{History stores all previous dialog pairs}
\State $Qz_{t} \gets Ques\_enc(Qbot,fact,history,caption)$
\State $image\_pred \gets image\_regress(Qbot, fact, history, caption)$
\State $ R_{t} \gets (target\_image - image\_pred)^2 - \Delta image\_pred$
\State $\Delta image\_pred \gets (target\_image - image\_pred)^2$
\EndFor
\State $\Delta(w_{Qbot}) \gets \frac{1}{10}\sum_{t=1}^{10} \nabla_{\theta_{Qbot}} \left[G_t \log p(ques_{t},\theta_{Qbot}) - \Delta image\_pred \right]$
\State $\Delta(w_{Abot}) \gets \frac{1}{10} \sum_{t=1}^{10} G_t \nabla_{\theta_{Abot}} \log p(ans_{t},\theta_{Abot}) $
\State $w_{Qbot} \gets w_{Qbot} + \Delta(w_{Qbot})$ \Comment{REINFORCE and Image Loss update for Qbot}
\State $w_{Abot} \gets w_{Abot} + \Delta(w_{Abot})$ \Comment{REINFORCE update for Abot}
\EndWhile\label{euclidendwhile}
\EndProcedure
\end{algorithmic}
\end{algorithm*}
\begin{table*}[t]
\small
\centering
\caption{Comparison of answer retrieval metrics with previously published work}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
\textbf{Model} & \textbf{MRR} & \textbf{Mean Rank} & \textbf{R@10} \\
\hline \hline
Answer Prior \cite{visdial} & 0.3735 & 26.50 & 53.23 \\
MN-QIH-G \cite{visdial} & 0.5259 & 17.06 & 68.88 \\
HCIAE-G-DIS \cite{vdialog2} & 0.547 & 14.23 & 71.55\\
Frozen-Q-Multi \cite{vdialog} & 0.437 & 21.13 & 60.48\\
CoAtt-GAN \cite{vdialog3} & 0.5578 & 14.4 & 71.74\\
\hline
SL(Ours) & \textbf{0.610} & \textbf{5.323} &\textbf{ 72.68}\\
RL - 1Q,1A(Ours) & 0.459 & 7.097 & 72.34\\
RL - 1Q,3A(Ours) & 0.601 & 5.495 & 72.48\\
RL - 3Q,1A(Ours) & 0.590 & 5.56 & 72.61\\
\hline
\end{tabular}
\label{tab:metrics}
\end{table*}
\section{Experiments and Results}
\subsection{Dataset description}
We use the VisDial 0.9 dataset for our task introduced by Das et al. \cite{visdial}. The data is collected using Amazon Mechanical Turk by pairing 2 annotators and asking them to chat about the image as a multi round VQA setup. One of the annotators acts as the questioner and has access to only the caption of the image and has to ask questions from the other annotator who acts as the `answerer' and must answer the questions based on the visual information from the actual image. This dialog repeats for 10 rounds at the end of which the questioner has to guess what the image was.
We perform our experiments on VisDial v0.9 (the latest available release) containing 83k dialogs on COCO-train and 40k on COCO-val images, for a total of 1.2M dialog question-answer pairs. We split the 83k into 82k for train, 1k for validation, and use the 40k as test, in a manner consistent with \cite{visdial}. The caption is considered to be the first round in the dialog history.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{percentile.pdf}
\caption{Comparison of Task Performance: Image Retrieval Percentile scores. This refers to the percentile scores of the ground truth image compared to the entire test set of 40k images, as ranked by distance from the Q-Bot's estimate of the image. The X-axis denotes the dialog round number (from 1 to 10), while the Y-axis denotes the image retrieval percentile score.}
\label{fig:percentile}
\end{figure}
\begin{table*}
\centering
\caption{Human Evaluation Results - Mean Rank (Lower is better)}
\label{table:humaneval}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
\textbf{} & \textbf{Metric} & \textbf{N} & \textbf{Supervised} & \textbf{RL 1Q,1A} & \textbf{RL 1Q,3A} & \textbf{RL 3Q,1A} \\
\hline \hline
1 & Question Relevance & 49 & \textbf{1.97} & 3.57 & 2.20 & 2.24 \\
2 & Question Grammar & 49 & 2.16 & 3.67 & 2.24 & \textbf{1.91} \\
3 & Overall Dialog Coherence: Q& 49 & 2.08 & 3.73 & 2.34 & \textbf{1.83} \\
\hline
4 & Answer Relevance & 53 & 2.09 & 3.77 & 2.28 & \textbf{1.84} \\
5 & Answer Grammar & 53 & 2.20 & 3.75 & 2.05 & \textbf{1.98} \\
6 & Overall Dialog Coherence: A & 53 & 2.09 & 3.64 & 2.35 & \textbf{1.90} \\
\hline
\end{tabular}
\end{table*}
\subsection{Evaluation Metrics}
We evaluate the performance of our model's individual responses by using 4 metrics, proposed by \cite{vdialog}: \textbf{1) Mean Reciprocal Rank (MRR)}, \textbf{2) Mean Rank}, \textbf{3) Recall@10} and \textbf{4) Image Retrieval Percentile}. Mean Rank and MRR compute the average rank (and its reciprocal, respectively) assigned to the \akr{ground truth} answer, over a set of 100 candidate answers for each question (also averaged over all the 10 rounds). Recall@10 computes the percentage of answers with rank less (better) than 10. Image Retrieval percentile is a measure of how close the image prediction generated by the Q-bot is to the ground truth. All the images in the test set are ranked according to their distance from the predicted image embedding, and the rank of the ground truth embedding is used to calculate the image retrieval percentile. \akr{All results for RL-1Q,1A, RL-1Q,3A and RL-3Q,1A are reported after 15 epochs of supervised learning and 10 epochs of curriculum learning as described in Section \ref{sec:training}. Consequently, the training time of all 3 systems are equal.}
\ak{Table \ref{tab:metrics} compares the Mean Rank, MRR, and Recall@10 of our agent architecture and dialog framework (below the horizontal line) with previously proposed architectures (above the line). SL refers to the agents after only the isolated, supervised training of Section \ref{sec:supervised}. RL-1Q,1A refers to a single, idiosyncratic pair of agents trained via reinforcement as in Section \ref{sec:rl}. RL-1Q,3A and RL-3Q,1A refer to social agents trained via our Multi-Agent Dialog framework in Section \ref{sec:madf}, with 1Q,3A referring to 1 Q-Bot and 3 A-Bots, and 3Q,1A referring to 3 Q-Bots and 1 A-Bot.
It can be seen that our agent architectures clearly outperform all previously published results using generative architectures in MRR, Mean Rank and R@10. This indicates that our approach produces consistently good answers (as measured by MRR, Mean Rank and R@10). But it is important to note that the point here is not to demonstrate the superiority of our architecture compared to other architectures. The point here is instead to show that the MADF framework is able to recover the language quality of the supervised agent. In fact, community regularization (in the form of the proposed MADF setup) can be integrated with any of the visual dialog algorithms in Table \ref{tab:metrics}.
Notice that SL has the best scores, which drops drastically in RL-1Q,1A. But the agents trained by MADF \akr{recover the scores obtained by SL. This shows that the agents trained by MADF are able to recover the language quality of SL agents without sacrificing much on the task performance (image retrieval percentile).} Fig. \ref{fig:percentile} shows the change in image retrieval percentile scores over dialog rounds. The percentile score decreases monotonically for SL, but is stable for all versions using RL.} \akr{The decrease in image retrieval score over dialog rounds for SL is because the test set questions and answers are not perfectly in-distribution (compared to the training set), and the SL system can't adapt to these samples as well as the systems trained with RL. As the dialog rounds increase, the out-of-distribution nature of dialog exchange increases, hence leading to a decrease in SL scores. Interestingly, despite having strictly more information in later rounds, the scores of RL agents do not increase - which we think is because of the nature of recurrent networks to forget.
The results in Fig. \ref{fig:percentile} and Table \ref{tab:metrics} show that the MADF setup allows the agents to achieve consistent task performance without sacrificing on language quality. We further support this claim in the next section where we show that human evaluators rank the language quality of MADF agents to be much better than the agents trained via reinforcement without community regularization.}
\subsection{Human Evaluation}
\label{sec:humaneval}
There are no quantitative metrics to comprehensively evaluate dialog quality, hence we do a human evaluation of the generated dialog. There are 6 metrics we evaluate on: 1) Q-Bot Relevance, 2) Q-Bot Grammar, 3)A-Bot Relevance, 4) A-Bot Grammar, 5) Q-Bot Overall Dialog Coherence and 6) A-Bot Overall Dialog Coherence. We evaluate 4 Visual Dialog systems, trained via: 1) \textbf{Supervised Learning (SL)}, 2) \textbf{Reinforce for 1 Q-Bot, 1 A-Bot (RL-1Q,1A)}, 3) \textbf{Reinforce for 1 Q-Bot, 3 A-Bots (RL-1Q,3A)} and 4) \textbf{Reinforce for 3 Q-Bots, 1 A-Bot (RL-3Q,1A)}. We asked a total of 61 people to evaluate the 10 QA-pairs generated by each system for a total of 102 randomly chosen images, requiring them to give an ordinal ranking (from 1 to 4) for each metric. All the evaluators were provided with the caption from the dataset.
Evaluators taking the perspective of the A-Bot were provided with the image and asked to evaluate answer relevance and grammar, while those taking the perspective of the Q-Bot evaluated question relevance and grammar. Both groups rated dialogs for overall coherence.
Table \ref{table:humaneval} contains the average ranks obtained on each metric (lower is better).
\ak{The results convincingly validate our hypothesis that having multiple A-Bots/Q-Bots improves the language quality as compared with single Q-Bot and A-Bot.
Kruskal-Wallis tests found strong differences among rankings (p< .0001) across all measures. Pairwise comparisons using the Mann-Whitney U test found a consistent pattern in which RL 1Q,1A performed substantially worse than other methods across all measures: for \textbf{Q-relevance}: SL: U=348, p<.0001; RL-1Q3A: U=2235, p< .0001; RL-3Q1A U=2209, p< .0001, \textbf{Q-grammar}: SL: U=319, p< .0001; RL-1Q3A U=2280, p < .0001; RL-3Q1A U=2221, p < .0001; \textbf{A-relevance}: SL U=256, p < .0001; RL-1Q3A U=2741, p < .0001; RL-3Q1A U-2909, p < .0001; \textbf{A-grammar}: SL U=305, p < .0001; RL-1Q3A U=2857, p < .0001; RL-3Q1A U=2673, p < .0001; \textbf{Overall (both groups)}: SL U=1206, p < .0001; RL-1Q3A U= 9458, p < .0001; RL-3Q1A U=10052, P < .0001. Slight differences favoring RL 3Q,1A over RL 1Q,3A were found for A-relevance U=1889, p < .02 and overall coherence U=6543, p < .006 but otherwise SL, RL-1Q,3A, and RL-3Q,1A showed equivalent performance indicating that community regularization can effectively eliminate any losses to human intelligibility introduced through reinforcement learning.
These results further support the claims made in the previous section that the MADF setup allows the agents to show consistent task performance (image retrieval percentile) while maintaining the language quality of the supervised agents.
We show a randomly chosen example from the set shown to the human evaluators in Fig. \ref{fig:humaneval}. The trends observed in the scores given by human evaluators are also clearly visible in this example. MADF agents are able to model the human responses much better than RL 1Q,1A and are about as well as (if not better) than SL trained agents. It can also be seen that although the RL-1Q,1A system has greater diversity in its responses, the quality of those responses is greatly degraded, with the A-Bot's answers especially being both non-grammatical and irrelevant.}
\section{Discussion and Conclusion}
\ak{In this paper we propose a novel community regularization technique, the Multi-Agent Dialog Framework (MADF), to improve the dialog quality of artificial agents. We show that training 2 agents with supervised learning does not ensure good task performance (measured by the image retrieval percentile scores) at test time, and it only deteriorates as the agents exchange more information about the image. We hypothesize that this is because the agents were trained in isolation and never allowed to interact during supervised learning, which leads to failure during testing when they encounter out of distribution samples (generated by the other agent, instead of ground truth) for the first time. We show how allowing a single pair of agents to interact and learn from each other via reinforcement learning dramatically improves their percentile scores, which additionally does not deteriorate over multiple rounds of dialog, since the agents have interacted with one another and been exposed to the other's generated questions or answers. However, in an attempt to improve task performance, the agents end up developing their own private language which does not adhere to the rules and conventions of human languages. As a result, the dialog system loses interpretability and sociability. To alleviate this issue, we propose a multi-agent dialog framework to provide regularization. In this framework, a single A-Bot is allowed to interact with multiple Q-Bots and vice versa. Through a human evaluation study, we show that this leads to significant improvements in dialog quality measured by relevance, grammar and overall coherence, \akr{without compromising the task performance}.}
\section{Future Work}
There are several possible extensions to this work. We plan to explore several other multi bot architectural settings and perform a more thorough human evaluation for qualitative analysis of our dialog. \ak{We also plan on incorporating MADF into other architectures and models proposed by more recent work and test how well MADF generalizes to other models. }
Another avenue for future exploration is to use a richer image feature embedding to regress on. Currently, we use a regression network to compute the estimated image embedding which represents the Q-Bot's understanding of the image. However, a GAN which uses this embedding as a latent code to generate an image is an interesting possibility.
\section*{Acknowledgments}
\ak{This research was sponsored in part by AFOSR Grant FA9550-15-1-0442. We would like to thank William Guss, Abhishek Das, Satwik Kottur and Fei Fang for their insightful and fruitful discussions and feedback. }
|
\section{Introduction}
\begin{figure}
\begin{subfigure}[b]{1.0\textwidth}
\centering
\includegraphics[height=0.4in,width=0.8\textwidth]{cover.jpg}
\caption{Occlusion}
\end{subfigure}
\begin{subfigure}[b]{1.0\textwidth}
\centering
\includegraphics[height=0.4in, width=0.8\textwidth]{blur.jpg}
\caption{Large motion}
\end{subfigure}
\begin{subfigure}[b]{1.0\textwidth}
\centering
\includegraphics[height=0.4in, width=0.8\textwidth]{cut.jpg}
\caption{Cut transition}
\end{subfigure}
\centering
\begin{subfigure}[b]{1.0\textwidth}
\centering
\includegraphics[height=0.4in, width=0.8\textwidth]{dissolve.jpg}
\caption{Gradual transition}
\end{subfigure}
\caption{Challenge of shot boundary detection}\label{fig:1}
\end{figure}
Shot transition detector is a necessary component in many video recognition tasks. The goal of shot transition detection is to find semantic breaks in videos. Cut transitions are defined as abrupt transitions from one sequence to another while gradual transitions are almost the same but in a gradual manner. They share one common attribute, the start of a transition and the end of a transition are semantically different. Previous methods focus on finding both cut transitions and gradual transitions with one similarity function.\cite{yusoff2000video,yuan2005unified} Such methods have shown a great success in cut transition detection in the aspects of both speed and accuracy. However, when applied to gradual transition detection, it is not effective in the detection of gradual transitions. As Figure \ref{fig:1} shows, it is widely recognized that many large motions or occlusion, e.g. camera movement, are detected as positive when only measuring similarity. In order to overcome this shortcoming, recent research\cite{lu2013fast,hassanien2017large} begins to explore the temporal pattern of gradual transitions. Therefore, in \cite{hassanien2017large}, the C3D ConvNet is adopted to classify segments into three classes (cut, gradual and background), which achieves state-of-the-art performance. Yet C3D ConvNet not only consumes too much computing resources, but is also not an effective architecture for handling both cut and gradual transitions, i.e. the lengths of gradual transitions are varying but C3D ConvNet is not designed for multi-scale detection. Inspired by this method and previous similarity measurement method, we present a cascade framework, consisting of a targeted cut transition detector and a targeted gradual transition detector. The cut transition detector, for measuring the image similarity, is fast and accurate while the gradual transition detector is capable of capturing the temporal pattern of gradual transitions in multi-scale level. In addition, compared to deepSBD, our framework can locate both cut transitions and gradual transitions accurately.
In this work, we present a new cascade framework, a fast and accurate approach for shot boundary detection. The first stage applies a ridiculously fast method to initially filter the whole video and selects the candidate segments. This stage is for accelerating the framework (up to 2 times faster than not) and facilitate the training for the cut/gradual detector. In the second stage, we use a well designed 2D ConvNet learning the similarity function between two images to locate the cut transitions. The third stage utilizes a novel C3D ConvNet model to locate positions of gradual transitions. Typically, we use the notation of default boxes introduced in \cite{liu2016ssd} and propose a novel single shot boundary detector (SSBD).
In sum, our framework is fast and accurate for shot boundary detection and achieves state-of-the-art performance on many public databases running at 700FPS without any bells and whistles.
Current datasets, i.e. TRECVID and RAI, are
not sufficient for training deep neural net due to limited dataset size. Besides, the training set is various in different work when evaluating supervised methods on TRECVID and RAI databases. For training a high performance neural network and a fair comparison between different methods, we contribute a new large-scale video shot database ClipShots consisting of different types of videos collected from Youtube and Weibo. ClipShots is the first large-scale database for shot boundary detection and will be released.
Aspects of novelty of our work include:
\begin{itemize}
\item We separate cut transition detection and gradual transition detection, designing targeted network structures with different purposes.
\item We design a cascade framework for accelerating the processing speed.
\item We collect the first large-scale database for shot boundary detection training and evaluation.
\end{itemize}
\section{RELATED WORK}
In this section, we introduce the work related to our proposed framework.
\textbf{Unsupervised shot boundary detection method} In decades, many researchers explore to design similarity function finding transitions with hand-crafted features. In \cite{yusoff2000video}, average Intensity Measurement(AIM), Histogram Comparison(HC), Likelihood Ratio(LR) is used as the feature extractor. It is observed that similarities often vary gradually within a shot but abruptly in shot boundaries so the paper proposes an adaptive threshold should be applied when selecting positive samples. This method greatly improves the gradual transition performance compared to methods that only use static thresholds. Besides, another benefit is that it runs very fast so we integrate it in our framework to select potential shot boundaries. Yuan et al.\cite{yuan2005unified} proposes a graph partition model to perform temporal data segmentation. It treats every frame as a node and calculate the similarity metrix and the scores of the cuts, selecting feasible cuts whose scores are the local minima of the corresponding neighborhoods. These two methods all rely on well designed hand-crafted features to calculate the similarity of two images.
\textbf{Supervised shot boundary detection method} Due to the shortcoming of unsupervised methods, Yuan et al.\cite{yuan2007formal} adopts a supervised way, a support vector machine trained to classify different shot boundaries with extracted features. In \cite{liu2007t}, shot boundaries are classified into 6 categories (cut, fast dissolve, fade in, fade out, dissolve, wipe). Different features are used to train different SVMs targeting at different shot boundaries. Researchers explore which features can most effectively classify the shot boundaries.
\textbf{Shot boundary detection with deep learning} Hassanien et al.\cite{hassanien2017large} introduces a simple C3D network that takes a segments of fixed length as input and classify it into 3 categories (cut, gradual, background). This method shows the effectiveness of ConvNet in this task. However, this method deals with gradual transitions of different scales in the same way and cannot locate the accurate`boundaries. Gygli\cite{gygli2017ridiculously} also adopts fully convolutional network. It takes the whole video sequence as input and assigns the positive label to the frames in transitions.
\textbf{Image similarity comparison} Image similarity computation is a necessary component in shot boundary detection. Deep learning has been successful on image similarity comparison task. In \cite{zagoruyko2015learning}, three architectures are proposed to compute image similarities, siamese net, image concatenation net, pseudo-siamese net. Empirical experiments show the image concatenation network and its variants obtain the best performance. In \cite{wang2014learning}, a ranking model that employs deep learning techniques to learn similarity metric directly from images. We apply the similarity measurement only for the cut transition detection.
\textbf{Object detection} State-of-the-art methods for general object detection are mainly based on deep ConvNet to extract rich semantic features from images. Liu et al.\cite{liu2016ssd} introduces single shot detector(SSD) using default boxes to match the feature to ground truth and achieve the speed of 19-46 fps. Our gradual detection model design share the same spirit with SSD.
\textbf{Action recognition} Recently, researchers have paid more attention on video recognition and temporal detection. Carreira and Zisserman\cite{kay2017kinetics} has released the kinectics database for large-scale action classification. I3D\cite{carreira2017quo} shows a good weights initialization is necessary to train the C3D network. Qiu et al.\cite{qiu2017learning} proposes a fast network architecture based a spatial convolution kernel and temporal kernel to explore the temporal information. Action recognition is closely related to our work because we want to use temporal information to distinguish large motions and the gradual transitions.
\textbf{Action detection} This task focuses on learning how to detect action instances in untrimmed videos where the boundaries and categories of action instances have been annotated. Recently, many approaches adopt 'detection by classification' framework. Xu et al.\cite{xu2017r} builds faster-RCNN style architecture for fast classifying and locating actions. It first selects potential segments with region proposal network and proposes the ROI 3D pooling layer to extract rich features for further classification. In \cite{lin2017single}, the single shot detector locates action on feature map extracted from well trained action classification ConvNets. Escorcia et al.\cite{escorcia2016daps} proposes to generate a set of proposals based on the RNN network. Zhao et al.\cite{zhao2017temporal} models the temporal structure of each action instance via a structured temporal pyramid. Although some of the methods can be applied to gradual detection directly, these methods rely on extracting rich spatial-temporal features from a heavy ConvNet body, so these methods are far slower than our proposed methods.
\begin{figure}
\centering
\includegraphics[height=3.4in,width=0.8\textwidth]{shot_detection_pipeline}
\caption{An overview of our framework}\label{fig:2}
\end{figure}
\section{OUR APPROACH}
In this section, we will introduce our approach in details. The framework of our approach is shown in Figure \ref{fig:2}.
\subsection{An Overview}
The framework takes a video as input and predicts the locations of transitions. The proposed method, as shown in Figure \ref{fig:2}, is composed of three modules, including initial filtering, cut transition detector and gradual transition detector, implemented with three stages. 1) Adaptive thresholding produces a set of transition candidates. Each candidate comes with a center frame index indicating whether the content in frames has drastic changes. These positions may be transitions or caused by large motion, e.g. camera movement. 2) The candidate transitions are further feed into a strong cut transition detector to filter out false cut transitions. 3) For the remaining center frames which have negative responses to the cut detector, we expand them by \(x\) frames on both forward and backward temporal directions to form candidate segments. The gradual transition detector processes all these segments, locating the gradual transitions. The whole framework is designed in a cascade way and computation is light except the detection of gradual transition in the stage three. However, considering that most of the candidates have been filtered by the cut model, the number of candidates left for the gradual model is quite small and the computation at this stage is subtle.
\subsection{Initial Filtering}
As most of the consecutive video frames are highly similar to each other, a trivial unsupervised algorithm can be applied to reduce the candidate regions for further processing. A fast method, adaptive thresholding, is chosen as the initial filtering step.
Let \(I_n\) and \(I_{n+1}\) be the potential transition candidates and \(F_{n-a+1}\), \(F_{n-a+2}\), ..., \(F_{n+a}\) be a set of features extracted from consecutive video frames in a sliding window of length \(2a\) centered at frame \(n\). In practice, we use the feature extracted from SqueezeNet\cite{iandola2016squeezenet} trained on Imagenet\cite{deng2009imagenet}. The computation cost in this step is subtle. We calculate the similarity metric of each frame \(S_i\), which is represented as the cosine distance between the current frame feature and its neighboring frame feature. Given the similarity metric of these frames as \(S_{n-a+1}\), \(S_{n-a+2}\), \(S_{n-a+3}\), ..., \(S_{n+a-1}\), the threshold of a window is calculated as
\begin{equation}
T=t+\frac{\sigma}{n}\sum_{i=n-a+1}^{n+a-1}(1-S_i)
\end{equation}
The hyper-parameter \(\sigma\) is the dynamic threshold ratio and \(t\) is the static threshold. In practice, we set \(\sigma\) to 0.05 and \(t\) to 0.5. The frame is selected as a candidate center if \(1-S_n\) is larger than \(T\). Lengths of gradual transitions vary greatly. In order not to miss any gradual transition, we down-sample frames with multiple temporal scales. At scale \(\omega\), we sample one video frame every \(\omega\) frames and do the above thresholding operations on these down-sampled frames. Finally, results of different scales are merged together. If two candidates on different scales are too close, i.e., within a distance of 5 frames. The candidate with a lower scale will be kept. In practice, we use scales of 1, 2, 4, 8, 16, and 32.
\subsection{Cut Model}\label{sec:1}
Some image pairs are semantically similar even when they are cut transitions, i.e. images containing the same object but the backgrounds are different. Therefore, a stronger cut transition detector is needed for filtering out these negative cut candidates from the candidates selected by adaptive threholding. Zagoruyko and Komodakis\cite{zagoruyko2015learning} show CNN can learn the similarity function directly from image pairs. We design a ConvNet to determine whether a image pair is a cut transition or not. In this paper, we compare four models, including siamese, image concatenation, feature concatenation and C3D ConvNet. In contrast to deepSBD, where the position of the cut transition is unknown in one segment, adaptive thresholding can find the cut transition position accurately since it selects the pair of adjacent frames with the largest dissimilarity as the center, facilitating the learning task for our cut detector.
\textbf{Siamese} A siamese neural network consists of twin networks that accept distinct images and output their features. The parameters are shared between the twin networks and each network computes the same function, so two extremely similar images could not be mapped to very different location in the feature space. An energy loss function is added to the top for optimization. Besides, the network is symmetric, so that whenever we present two distinct images to the twin networks, the top conjoining layer will compute the same metric as if we were to present the same two images but to the opposite twins. In our problem, we choose contrastive loss as the top energy function. The siamese net outputs a similarity score. At inference, we select the score above some threshold.
\textbf{Feature concatenation} This network can be seen as a variant of siamese network. More specifically, it has the structure of the siamese net described above, computing the feature using the same network architecture and weights. The loss energy function is not applied directly to the features. Instead, we concatenate features from both images and add cross entropy loss function to the top.
\textbf{Image concatenation} We simply consider the two patches of an RGB image pairs as a 6-channel image and feed it to a generic network. This network provides greater flexibility compared to the above models as it starts by processing the two patches jointly. It is fast to train and infer. Further more, it allows to concatenate multiple images as a input. We find the performance is much improved when using more images.
\textbf{C3D ConvNet} Hassanien et al.\cite{hassanien2017large} shows the C3D ConvNet is capable of classifying cut transitions. Therefore, we also test this structure for comparison. However, the C3D ConvNet is more complex than 2D ConvNet, which requires much computation resources.
\begin{figure}
\centering
\includegraphics[height=3.4in, width=0.8\textwidth]{ssd}
\caption{An overview of gradual detector}\label{fig:3}
\end{figure}
\subsection{Gradual Model}
Inspired by region proposal network\cite{renNIPS15fasterrcnn} and single shot detector\cite{liu2016ssd}, we propose a single shot boundary network, a novel network to locate gradual transitions in a continuous video stream. The network, illustrated in Figure \ref{fig:3}, consists of 2 components, a shared C3D ConvNet feature extractor and subnets for classification and localization.
\textbf{Feature hierarchies} Innovated by deepSBD, the C3D ConvNet shows impressive performance in this task. Therefore, we use a C3D ConvNet to extract rich temporal feature hierarchies from a given input video buffer. The input to our model is a sequence of RGB video frames with a dimension of \(3\times L\times H\times W\) and we use ResNet-18 proposed in \cite{hara3dcnns} as the backbone network. However, unlike \cite{hara3dcnns}, the input to our model can be of variable lengths. We modify all the temporal strides to 1 in ResNet-18 so that the length of the final feature map is also L. The number of frames L can be arbitrary and is only limited by memory.
\textbf{Subnets for Classification and Location} Since the lengths of gradual transitions are various, we use the same notion default boxes introduced in \cite{liu2016ssd}. In our task, we call it default segments. Default segments are predefined multi-scale windows centered at a location. we put one default segment every \(l\times (1-a)\) frames where \(l\) is the length of the default segment and \(a\) is the positive IOU threshold. Therefore, each ground truth whose length is between \(l/a\) and \(l\times a\) can be matched to a default segment. The total number of default segments is \(L/(l\times(1-a))\). The default segments serve as reference segments for ground truth matching. To get features for predicting gradual transitions, we first apply a spatial global average pooling layer to reduce the spatial dimension to \(1\times1\). At each location which has \(k\) default segments, we apply a \(2k\times3\times1\times1\) filter \(A\) for binary classification, and a \(2k\times3\times1\times1\) filter \(B\) for location refinement. For both \(A\) and \(B\), 3 is the size of the temporal convolution kernel. For \(A\), 2 corresponds to binary classification of a gradual transition or not. For B, 2 corresponds to two relative offsets of \(\{\delta c_i ,\delta l_i\}\) to the center location and the length of each default segment respectively, where the ground truth of \(\{\delta c_i ,\delta l_i\}\) is defined as
\begin{align}
\delta c_i&=(c-c_i)/l_i \\
\delta l_i&=log(l/l_i)
\end{align}
The mark \(c_i\) and \(l_i\) are the center location and the length of default segments while c and l is the ground truth position and length.
\textbf{Optimization strategy} In training, positive/negative labels are assigned to default segments. Following the same protocol in object detection, positive labels are assigned if default segments are overlapped with some ground truth if intersection of union \(IOU>a\) and negative labels are assigned for default segments if \(IOU<b\). Segments with IOU between \(a\) and \(b\) are ignored during training. In practice, we set \(a\) to 0.5 and \(b\) to 0.1, which achieves the best performance. As the length of the gradual transitions in our training data ranges in 3 to 40, we use 2 default segments of length 6 and 20 to cover all true transitions. Similar to single shot detector, we implement hard negative example mining and dynamically balance the positive and negative examples with a ratio of \(1:1\) during training. To utilize the GPU efficiently, we fixed the length of each segment, consisting of L consecutive frames, i.e., L is 64 in our experiment.
We train the network by optimizing the classification and the regression losses jointly with a fixed learning rate of 0.001 for 5 epochs. We adopt softmax loss for classification and smooth L1 loss for regression. The loss function is given in (\ref{Eq:locationloss}). The hyper-parameter \(\lambda\) is set to 1 in practice. \(Y_i^1\) is the predicted score and \(T_i^1\) is the assigned label. \(Y_i^2 =\{ \delta c_i ,\delta l_i\}\) is the predicted relative offset to the default segments and \(T_i^2\) is the target location. The loss function is the same as \cite{liu2016ssd}, which is
\begin{equation}\label{Eq:locationloss}
Loss=\frac{1}{N_{cls}} \sum_{i} L_{cls}(Y_i^1,T_i^1)+\lambda \frac{1}{N_{loc}} \sum_{i} L_{loc}(Y_i^2,T_i^2)
\end{equation}
\textbf{Inference} At inference, the framework processes input videos of varying lengths. However, in order not to exceed the limit of memory, a video will be divided into segments of length \(T_{seg}\) with a overlap of \(\frac{1}{2}T_{seg}\) such that transitions won't be missed due to the division. After predicting one video, we apply non maximum suppression(NMS) to all the predictions. If two predicted gradual transitions are overlapped, we remove the one with lower classification score.
\section{ClipShots}
Current datasets, i.e. TRECVID and RAI, are not sufficient for training deep neural network due to a limited size. In addition, previous work utilized different training sets when evaluating their supervised methods on TRECVID and RAI. Therefore, a benchmark is made for comparing different methods fairly. ClipShots is the first large-scale dataset for shot boundary detection collected from Youtube and Weibo covering more than 20 categories, including sports, TV shows, animals, etc. In contrast to TRECVID2007 and RAI, which only consist of documentaries or talk shows where the frames are relatively static, we construct a database containing 4039 short videos from Youtube and Weibo. Many short videos are home-made, with more challenges, e.g. hand-held vibrations and large occlusion. The training set consists of 3539 videos, 122760 cut transitions, and 35698 gradual transitions while the evaluation set consists of 500 videos, 5876 cut transitions, and 2422 gradual transitions. The types of these videos are various, including movie spotlights, competition highlights, family videos recorded by mobile phones etc. Each video has a length of 1-20 minutes. The gradual transitions in our database include dissolve, fade in fade out, and sliding in sliding out. In order to annotate such a large dataset, we design an annotation tool allowing annotators to watch multiple frames on a single page and select the begin frame and the end frame of transitions. More details are given in the appendix. Every video is double annotated for quality assurance.
\section{Experiments}
\subsection{Databases and Evaluation Metrics}
We will introduce the databases and evaluation metrics in this section.
\textbf{Training and evaluation set} The proposed framework is trained and tested on ClipShots.
In order to illustrate the effectiveness of our approach and ClipShots, we also evaluated them on two public databases (TRECVID2007, RAI).
\textbf{Evaluation metrics} For all 3 databases, we use the standard TRECVID evaluation metrics: one-to-one match if the predicted boundary has at least 1 frame overlapped with the ground truth. For our testing set, we add an additional criterion using IOU to measure the localization performance. We assess performance quantitatively using precision (P), recall (R) and F-score (F).
\subsection{Experiments Configuration}
We adopt adaptive thresholding to find candidate segments and adjust the parameters to make sure it achieves nearly 100\% recall for both cut and gradual transitions. For cut detector, 122760 positive examples and 224312 negative examples are used for training. For gradual detector, the training set contains 35698 ground truths. The potential segments filtered by adaptive thresholding are divided into subsegments of fixed length 64, with overlapped length of 32 between 2 consecutive segments. We choose ResNet-18 3D- ConvNet as the backbone, setting all the strides in the temporal dimension to 1 so that the temporal length of the output feature is identical with the input length. The weights of 3D ResNet-18 are initialized with model pretrained on kinectics database, as the inflated 3D-Conv \cite{carreira2017quo}. For both cut and gradual model, the positive examples and negative examples are highly unbalanced so the positive and negative samples are dynamically balanced with ratio 1:1 in each mini-batch.
\subsection{Experiments on ClipShots}
\begin{table}
\centering
\begin{tabular} {c|c|c|c}
\hline Model&P&R&F\\
\hline Image concat(2 frames)&0.771&0.793&0.782\\
\hline Image concat(4 frames)&0.775&0.862&0.816\\
\hline Image concat(6 frames)&0.776&0.934&0.848\\
\hline Feature concat(2 frames)&0.231&0.574&0.329\\
\hline Siamese(2 frames)&0.638&0.852&0.729\\
\hline C3D(16 frames)&0.760&0.910&0.831\\
\hline
\end{tabular}
\caption[Comparison of cut models]{Comparison of cut models. Image concat(6 frames) obtains the best performance.}\label{table:1}
\end{table}
\textbf{Cut detector comparison} In this section, we choose four potential models introduced in Sec. \ref{sec:1} and test their performance. We use ResNet-50 as the backbone for all models and a fixed learning rate of 0.0001, We train each model for 5 epochs. For C3D, we adopt the same configuration as deepSBD. For image concatenation model, we evaluated it with different number of images. We expand \(\emph{x}\) frames to the forward and backward in the temporal direction. As Table \ref{table:1} shows, the image concatenation model obtains best performance among these four models when using 4 or more frames. Siamese net performs worse than image concatenation (2 frames) and C3D network. Given the fact that siamese net cannot explore information on multiple frames and its computation cost is much larger than image concatenation, this architecture is not adopted in our framework. C3D network (16 frames) is a little better than image concatenation (2 frames), but much worse than image concatenation (4 frames or 6 frames). Feature concatenation is not a working architecture, but we still list it here. For image concatenation, we also study the relationship between the number of input images and performance. More input frames can improve performance. The model gains improvements when increasing the frame number from 2 to 6 and saturates around 6. Therefore, we use an input of 6 frames in our method considering both performance and the processing speed.
\begin{table}
\centering
\begin{tabular} {c|c|c|c}
\hline method&initial filtering&cut&gradual\\
\hline (1)&No&DeepSBD&DeepSBD\\
\hline (2)&Yes&DeepSBD&DeepSBD\\
\hline (3)&Yes&image concat(6 frames)&DeepSBD\\
\hline (4)&Yes&image concat(6 frames)&SSBD\\
\hline
\end{tabular}
\caption[general comparison]{All methods under a unified viewpoint. Different cut models and gradual models are compared.}\label{table:8}
\end{table}
\vspace{-5em}
\begin{table}
\centering
\small
\begin{tabular} {c|c|c|c|c|c|c}
\hline \multicolumn{1}{c|}{methods}&\multicolumn{3}{|c|}{cut}&\multicolumn{3}{|c}{gradual}\\
\hline &P&R&F&P&R&F\\
\hline (1)&0.765&0.910&0.831&0.770&0.622&0.688\\
\hline (2)&0.757&0.902&0.823&0.699&0.810&0.750\\
\hline (3)&0.776&0.934&0.848&0.711&0.830&0.766\\
\hline (4)&0.776&0.934&0.848&0.840&0.904&0.870\\
\hline
\end{tabular}
\caption[general comparison]{Performance of different methods. Our method (4) obtains the best performance in both cut transition detection and gradual transitions detection.}\label{table:2}
\end{table}
\vspace{-2em}
\begin{table}
\centering
\begin{tabular} {c|c|c|c|c|c|c}
\hline \multicolumn{1}{c|}{methods}&\multicolumn{3}{|c|}{cut}&\multicolumn{3}{|c}{gradual}\\
\hline &P&R&F&P&R&F\\
\hline deepSBD (Original)&0.731&0.921&0.815&0.837&0.386&0.528\\
\hline deepSBD (ResNet-18)&0.765&0.910&0.831&0.770&0.622&0.688\\
\hline FCN\cite{gygli2017ridiculously}&0.410&0.093&0.151&0.393&0.053&0.093\\
\hline DSM (ours)&0.776&0.934&0.848&0.840&0.904&0.870\\
\hline
\end{tabular}
\caption[general comparison]{Benchmark in ClipShots}\label{table:5}
\end{table}
\textbf{Ablation study} We conduct ablation study with different options. The detailed setting is shown in Table \ref{table:8}. The difference is mainly at cut models, gradual models, and whether initial filtering is used. We also implement deepSBD but the post processing technology introduced in \cite{hassanien2017large} is abandoned for a fair comparison. We adopt 3D ResNet-18 as the backbone for both deepSBD and our single shot boundary detector.
\textit{Method (1)} The model classifies segments directly into 3 categories (cut, gradual, and background).
\textit{Method (2)} Compared to method (1), initial filtering is utilized to find candidate segments for deepSBD. As is shown, the performance of gradual transition is higher than the original deepSBD. It is implied that the initial filtering can also improve performance of deepSBD.
\textit{Method (3)} For gradual transitions, the deepSBD model only classifies the segments into 2 categories (gradual transition and background) so cut transitions are treated as negative samples. For cut detector, we use image concatenation model. Compared to method(2), the results show the single shot boundary detector is better than deepSBD by a large margin given that the cut detector are the same.
\textit{Method (4)} The results reveals that our single shot boundary detector is far better than deepSBD. We attribute the performance gain to the following reasons: 1) The receptive field of our model is much bigger than deepSBD, hence the detector can exploit more temporal information. 2) Our default segment design is effective for dealing with gradual transitions of multi scales.
\textbf{Benchmark in ClipShots}
We implement \cite{hassanien2017large,gygli2017ridiculously} and evaluate them in ClipShots. Table \ref{table:5} summaries performance of different methods. DeepSBD with 3D ResNet-18 is significantly better than the original network (3D Alexnet alike). It is also noted that the method introduced in \cite{gygli2017ridiculously} obtains worst performance on our dataset. Its network only consists of 3 convolutional layers. We suppose it is too shallow to fit large-scale data.
\begin{table}
\centering
\begin{tabular} {c|c}
\hline model&speed(FPS)\\
\hline deepSBD&382\\
\hline adaptive thresholding+deepSBD&680\\
\hline ours&700\\
\hline
\end{tabular}
\caption[Comparison of cut models]{Comparison of speed}\label{table:speed}
\end{table}
\textbf{Speed Comparison}
In this section, we compare the speeds of different models as shown in Table \ref{table:speed}. The code is implemented using PyTorch and tested with one TITAN XP GPU. Our method is nearly 2 times faster than the original deepSBD on account of adaptive thresholding based initial filter.
\begin{table}
\centering
\begin{tabular} {c|c|c|c}
\hline IOU&P&R&F\\
\hline 0.1&0.813&0.865&0.839\\
\hline 0.5&0.726&0.772&0.748\\
\hline 0.75&0.599&0.637&0.618\\
\hline
\end{tabular}
\caption[localization]{Localization performance. We calculate the F1-score at different IOU threshold.}\label{table:3}
\end{table}
\vspace{-1em}
\textbf{Gradual Model Localization Performance} An accurate localization of gradual transitions is important in many video recognition task. Therefore, we also evaluate performance of the gradual transition localization using the proposed framework. F1 scores are measured at different IOU level \((0.1,0.5,0.75)\). A predicted gradual transition is considered as correct only if its \(IOU>a\), otherwise it's considered wrong. When the IOU is 0.75, we can still obtain a F1 score of 0.618, indicating the proposed gradual detector is able to accurately locate gradual transitions.
\begin{table}
\centering
\begin{tabular} {c|c|c|c|c|c|c}
\hline \multicolumn{1}{c|}{methods}&\multicolumn{3}{|c|}{cut}&\multicolumn{3}{|c}{gradual}\\
\hline &P&R&F1&P&R&F1\\
\hline ATT\cite{liu2007t}&0.996&0.979&0.972&0.802&0.709&0.753\\
\hline THU11\cite{yuan2007formal}&0.982&0.968&0.975&0.733&0.718&0.725\\
\hline Marburg\cite{muhling2007university}&0.942&0.945&0.944&0.595&0.766&0.670\\
\hline NHT\cite{kawai2007shot}&0.975&0.816&0.945&0.768&0.578&0.66\\
\hline Priya\cite{domnic2014walsh}&0.972&0.976&0.974&0.869&0.719&0.78\\
\hline DeepSBD\cite{hassanien2017large}&0.978&0.968&0.973&0.826&0.731&0.776\\
\hline Ours&0.971&0.988&0.980&0.813&0.806&0.810\\
\hline Ours (correct label)&0.981&0.997&0.989&0.838&0.845&0.841\\
\hline
\end{tabular}
\caption[Trecvid comparison]{Trecvid07 top performers.}\label{table:4}
\end{table}
\vspace{-3em}
\subsection{Experiments on TRECVID07}
TRECVID07 contains a total of 17 videos, including 2236 cut transitions and 225 gradual transitions. They are all color and black/white documentaries. The videos include cases such as global illumination variation, smoke, fire, and fast non-rigid motion. We take the ground truth from TRECVID07 SBD task. In addition, the experimental results of the proposed method over this database are compared to the top performers of TRECVID07 SBD task. We find some of the ground truths are wrong, so we correct these labels. Evaluation results using original labels and corrected labels are both reported. The cut and gradual models are trained with the same training setting described in Section 5.2.
In Table \ref{table:4}, we present a comparative evaluation of the shot boundary detection performance with existing state-of-the-art approaches in terms of F1-score and report the results using both the original ground truth and the corrected ground truth. We evaluate cut transitions and gradual transitions separately. Cut transitions are the most part of all transitions in a video so it plays a dominate role in the overall performance. For cut transitions, we improve the-state-of-art by 0.6\%, which is a huge improvement considering there is no much space for improvement. In fact, the errors concentrate in black/white videos due to the lack of similar ones in the training set. Further improvement can be achieved through adding more black/white videos into the training set. For gradual transitions, we achieve 2.9\% improvement comparing to the state-of-the-art when using the original ground truth and 6.4\% improvement when using the corrected ground-truth.
\vspace{-2em}
\begin{table}
\centering
\begin{tabular} {c|c}
\hline &F1 score \\
\hline Apostolidis et al.\cite{apostolidis2014fast}&0.84\\
\hline Baraldi et al.\cite{baraldi2015shot}&0.84\\
\hline Song et al.\cite{song2016click}&0.68\\
\hline Michael et al.\cite{gygli2017ridiculously}&0.88\\
\hline Hassanien et al.\cite{hassanien2017large}&0.934\\
\hline Ours&0.935\\
\hline
\end{tabular}
\caption[RAI comparison]{RAI comparison}
\end{table}
\vspace{-3em}
\subsection{Experiments on RAI}
RAI database is a collection of ten randomly selected broadcasting videos from the Rai Scuola video archive 1, which are mainly documentaries and talk shows. This database includes 722 cut transitions and 263 gradual transitions. Shots have been manually annotated by a set of human experts. The proposed method achieves a competitive results compared to deepSBD. It is noted that DeepSBD adopts posting-processing technology, i.e. filtering the segments whose HSV similarity under a threshold, which is not used in our methods. We perform evaluations on TRECVID and RAI using the same models, weights, and hyper-parameters, which indicates the proposed framework are robust on different databases.
\section{Conclusion}
We propose a cascade shot transition detection framework and annotate the first large-scale shot boundary database. Adaptive threshoding is adopted to find candidate regions for acceleration. The cut and gradual transition detector are designed separately. The cut transition detector is for measuring similarity while the gradual transition detector is for capturing temporal patterns. Especially, the gradual detector is able to locate gradual transitions of multi-scales. We outperform state-of-the-art methods on both TRECVID and RAI databases. In addition, our framework is very fast, achieving a 30\(\times\) real-time speed.
\bibliographystyle{splncs}
\section{Annotation tool}
\begin{figure}
\centering
\includegraphics[height=3.4in,width=1.0\textwidth]{tools.png}
\caption{Annotation tool interface}\label{fig:1}
\end{figure}
The Graphical User Interface (GUI) is shown in \ref{fig:1}. This tool allows the annotator to watch 45 consecutive frames of a video segment simultaneously. This visualization is helpful for annotators to find a transition with a single look. On each page, the annotator can annotate at most one transition, with the selection of a begin frame and a end frame. The selected begin frame and end frame are asked to cross the center frame on a page. This setting is for assuring an annotator can obtain enough frames to determine the ends of a transition. When a page covers several ground truths, the others can be made to the center frame by clicking the button 'Make Center'. In the top-right corner, there are two tags, 'Gradual/Cut' or 'Too hard'. 'Gradual/Cut' represents the user wants the selected segments to be the ground truth while 'Too hard' represents the segments is difficult to be determined. We exclude 'Too hard' segments in the training stage. The user can watch next or previous page by clicking the buttons in the bottom or selecting the page number in the top.
\end{document}
|
\section{Introduction}
Non-classical correlations has been identified as a resource for different communication tasks from quantum teleportation \cite{bennett} to remote state preparation \cite{dakic}. Most of the communication tasks uses entanglement as a resource like teleportation, but only a few tasks use quantum discord as a resource. Essentially the use of quantum discord as a resource is not very well explored or is controversial. For instance, let's take the example of remote state preparation. It was shown that separable states can outperform entangled states for remote state preparation. However it was later shown that, its validity is restricted to some special conditions \cite{horodecki}. Quantum discord has also been shown to be generated in the state conversion process even when there is no entanglement \cite{Tame,Koashi}.
A two qubit mixed state, is called classically correlated \cite{oppenheim} if it can be written as
\begin{equation}\label{classically correlated state}
\rho_{cc}=\sum_{i,j}p_{ij}\ketbra{i}{i}^A\otimes\ketbra{j}{j}^B,
\end{equation}
where $\ket{i}^A$ are the orthogonal states over the hilbert space $H_A$ and $\ket{j}^B$ are the orthogonal states over the hilbert space $H_B$. However if $\ket{i}^A$ or $\ket{j}^B$ are not orthogonal then the state $AB$ might be non-classically correlated. This led to the idea that mixed states might posses some non-classical correlations which is different from entanglement as the state could be separable. The first non-classical correlations which differ from entanglement is Quantum discord. However, it should be noted that the term ``non-classicality" may have different meanings in quantum information and quantum optics, as was shown in Ref. \cite{Paris,Vogelr} that in the context of quantum optics non-zero quantum discord might have some classical explanations and as well as in Ref. \cite{Sanders} it was shown that classical state may posses some nonzero discord if the measurements are noisy and can be represented by a stochastic channel. In our manuscript the term ``non-classicality" has been used in the language of quantum information where the measurements are not noisy. The notion of Quantum discord was first introduced by Zurek \cite{zurek} and independently by Vedral \cite{vedral} and Horodecki \cite{oppenheim}. Oliver and Zurek then went on to give a measure for Quantum discord \cite{zurek}. Quantum discord was first identified as a resource for computation by Datta {\it et al.}\cite{datta}, where they took a separable state as the resource and showed that the computation was better than using a classical state but not better than using an entangled state. Later Dakic {\it et al.} showed the role of quantum discord in remote state preparation \cite{dakic}, provided that Alice and Bob don't share a reference frame or the operations are bio-stochastic \cite{horodecki}. Many other applications of quantum discord have been proposed whose descriptions can be found in \cite{adesso,streltsov}.
Olliver and Zurek \cite{zurek} proposed a measure of quantum discord in terms of the mutual information. In classical information theory, one can express the mutual information between two random variables $X$ and $Y$ in two different ways -
\begin{eqnarray}
&&I(X:Y)=H(X)+H(Y)-H(X,Y) \quad \mbox{and}\nonumber\\
&&J(X:Y)=H(X)-H(X|Y),
\end{eqnarray}
where $H(X)=-\sum_xp_x\log_2p_x$ is the Shannon entropy of $X$, $p_x$ is the probability that $X$ takes the value $x$ and $H(X,Y)$ is the joint Shannon entropy of $X$ and $Y$. $H(X|Y)$ is the conditional entropy and defined as $H(X|Y)=\sum_yp_yH(X|y)$, where $p_y$ is the probability of $Y$ taking value $y$ and $H(X|y)$ is the conditional entropy of $X$, such that $Y$ take the value $y$. Unlike classical domain these two expression are different in quantum domain and their difference serve as quantum discord. The generalization of $I(X:Y)$ in quantum theory is
\begin{equation}
I(\rho^{AB})=S(\rho^A)+S(\rho^B)-S(\rho^{AB}),
\end{equation}
where $I(\rho^{AB})$ is the mutual information between $A$ and $B$ for the state $\rho^{AB}$, $S$ is the von Neumann entropy and $\rho^A$, $\rho^B$ are the reduced density matrix. However, the generalization of $J(X:Y)$ is not that straight forward. Olliver and Zurek \cite{zurek} extended this to quantum realm as
\begin{equation}
J(\rho^{AB})_{{\Pi_{i}^{B}}}=S(\rho^A)-S(A|{\Pi_{i}^{B}})
\end{equation}
and $S(A|{\Pi_{i}^{B}})$ is given by
\begin{equation}
S(A|{\Pi_{i}^{B}})=\sum_i p_iS(\rho_i),
\end{equation}
where $\rho_i=Tr_B[\Pi_{i}^{B}\rho^{AB}]/p_i$, $p_i=Tr[\Pi_{i}^{B}\rho^{AB}]$ and $\Pi_{i}^{B}$ are the measurement operators on the subsystem $B$. Therefore, the measure of Quantum discord as proposed in \cite{zurek}
\begin{equation}\label{quantum discord}
\delta^{B|A}= \mbox{min}_{{\Pi_{i}^{B}}} [I(\rho^{AB})-J(\rho^{AB})_{{\Pi_{i}^{B}}}],
\end{equation}
where the quantity $I$ represents mutual information and the quantity $J$ represents the amount of information gained about the subsystem $A$ by measuring the subsystem $B$. The minimization occurs over the set of measurement operators such that the quantum discord is measurement independent. For two-qubit states a partial analytic approach was given by Girolami and Adesso \cite{girolami}, however this also needed numerical minimization scheme.
In Quantum teleportation \cite{Liu,Miranowicz} two spatially separated parties Alice and Bob share a quantum channel. Alice has an unknown qubit which she wants to prepare at Bob's end without physically sending it. Then optimizations are done over the measurement basis and channel such that Alice's and Bob's state have greatest degree of overlap (essentially this is maximizing the fidelity). We look at an almost similar scenario where Alice and Bob are spatially separated and have a shared quantum channel. Alice wants to prepare a two qubit state, at Bob's end such that the quantum discord of Bob's state is higher than Alice's state. Moreover, we have shown that when the average quantum discord of Bob's state is higher than Alice's initial state and if we use Bob's state as a resource to teleport a single qubit, instead of Alice's state then the teleportation fidelity increases. As, Quantum discord is very difficult to calculate analytically we used Mathematica (Qdensity~\cite{diaz}) for numerically finding the results.
The paper is organized as follows. Section \ref{sec:sec2} contains the description of the protocol and contains a detailed worked out example with the protocol. A possible application of our result is discussed in section \ref{sec:sec3} and finally we conclude in section \ref{sec:sec4}.
\section{Description of the protocol}
\label{sec:sec2}
Alice and Bob share a quantum channel which is a pure state of dimension more than $4$ (for eg. three-qubit $W$ state, or four-qubit cluster state, etc.).
Alice has an unknown quantum state (a two-qubit Werner state $\rho_{A} = \lambda \ket{\phi^+} \bra{\phi^+} + (1-\lambda) I/4$, where $\ket{\phi^+}=\frac{1}{\sqrt{2}}(\ket{00}+\ket{11})$ and $I$ is a four-dimensional identity matrix). Then, Alice jointly measures her two-qubit state and the channel in such a way that Bob receives a two-qubit state. For this joint measurement Alice can choose a basis arbitrarily. As a result she would find different outcomes corresponding to the basis elements. Depending on Alice's outcomes, the two-qubits with Bob will collapse to different states. Alice classically communicates her outcome to Bob. Now Bob measures the quantum discord for each of the separate outcomes and then averages it over all the outcomes. It should be noted here that Bob doesn't need to apply any specific unitary measurement ($U_1\otimes U_2$), as discord remains conserved under a unitary transformation. Also this protocol is different from the remote state preparation in the sense that Alice doesn't know about the state she wants to send to Bob.
We compare the average quantum discord of the states with Alice and Bob and find that average quantum discord of Bob's state is higher than Alice's initial Werner state for most of the range of $\lambda$. We keep the basis chosen by Alice fixed, and vary the parameter $\lambda$ of the Werner state.
\subsection{Example 1}
\label{sec:sec2.1}
\begin{figure}[t!]
\includegraphics[width=\linewidth]{discord_werner_state_shared_4-qubit_cluster_state.pdf}
\caption{The red line shows the quantum discord as function of parameter $\lambda$ for the initial Alice's Werner state. The green line shows the average discord of the final states with variation in $\lambda$ (using the cluster state (\ref{shared state 1}) as the channel).}
\label{fig1:}
\end{figure}
Let's take a four-qubit Cluster state
\begin{equation}\label{shared state 1}
\ket{\Psi}_{C} = 1/2 \Big(\ket{00}_A\ket{00}_B + \ket{01}_A\ket{10}_B + \ket{10}_A\ket{01}_B - \ket{11}_A\ket{11}_B\Big),
\end{equation}
where first two qubits are with Alice and rest two qubits are with Bob. Alice has a two-qubit Werner state $\rho_{A} = \lambda \ket{\phi^+}\bra{\phi^+} + (1-\lambda) I/4$.
So the total state of the $2^6$ dimensional system is
\begin{equation}
\rho_{T} = \rho_{A} \otimes \ket{\Psi}_C\bra{\Psi}.
\end{equation}
Alice performs joint measurements on the qubits $1,2,3,4$ in the basis which is chosen arbitrarily (we want to project the state of those 4-qubits onto an entangled basis similar to teleportation) and sends 4 bits of classical information to Bob. We are interested in the properties (specifically discord) of the state of the qubits $5,6$ (which is with Bob) after Alice performs her measurement. Alice chooses a complete orthonormal measurement basis as
\begin{eqnarray}\label{four dim basis}
&&\ket{b_1} = \frac{1}{2}\Big(\ket{0001}+\ket{0010}+\ket{0100}+\ket{1000}\Big),\nonumber\\
&&\ket{b_2} = \frac{1}{2}\Big(-\ket{0001}+\ket{0010}+\ket{0100}-\ket{1000}\Big),\nonumber
\end{eqnarray}
\begin{eqnarray}
&&\ket{b_3} = \frac{1}{2}\Big(-\ket{0001}+\ket{0010}-\ket{0100}+\ket{1000}\Big),\nonumber\\
&&\ket{b_4} = \frac{1}{2}\Big(-\ket{0001}-\ket{0010}+\ket{0100}+\ket{1000}\Big),\nonumber\\
&&\ket{b_5} = \frac{1}{2}\Big(\ket{1110}+\ket{1101}+\ket{1011}+\ket{0111}\Big),\nonumber\\
&&\ket{b_6} = \frac{1}{2}\Big(-\ket{1110}+\ket{1101}+\ket{1011}-\ket{0111}\Big),\nonumber\\
&&\ket{b_7} = \frac{1}{2}\Big(-\ket{1110}+\ket{1101}-\ket{1011}+\ket{0111}\Big),\nonumber\\
&&\ket{b_8} = \frac{1}{2}\Big(-\ket{1110}-\ket{1101}+\ket{1011}+\ket{0111}\Big),\nonumber\\
&&\ket{b_9} = \frac{1}{2}\Big(\ket{0000}+\ket{0011}+\ket{1100}+\ket{1111}\Big),\nonumber\\
&&\ket{b_{10}} = \frac{1}{2}\Big(\ket{0000}+\ket{0011}-\ket{1100}-\ket{1111}\Big),\nonumber\\
&&\ket{b_{11}} = \frac{1}{2}\Big(\ket{0000}-\ket{0011}+\ket{1100}-\ket{1111}\Big),\nonumber\\
&&\ket{b_{12}} = \frac{1}{2}\Big(-\ket{0000}+\ket{0011}+\ket{1100}-\ket{1111}\Big),\nonumber\\
&&\ket{b_{13}} = \frac{1}{2}\Big(\ket{0101}+\ket{0110}+\ket{1010}+\ket{1001}\Big),\nonumber\\
&&\ket{b_{14}} = \frac{1}{2}\Big(-\ket{0101}-\ket{0110}+\ket{1010}+\ket{1001}\Big),\nonumber\\
&&\ket{b_{15}} = \frac{1}{2}\Big(\ket{0101}-\ket{0110}-\ket{1010}+\ket{1001}\Big)\,\mbox{and}\nonumber\\
&&\ket{b_{16}} = \frac{1}{2}\Big(\ket{0101}-\ket{0110}+\ket{1010}-\ket{1001}\Big).
\end{eqnarray}
The measurement is mathematically defined as
\begin{equation}\label{measurement alice}
\hat{M}_{i} = \ketbra{b_i}{b_i} \otimes I,
\end{equation}
where $I$ is the four dimensional identity matrix. Alice sends her outcome to Bob using a classical channel. Then Bob's state would collapse to
\begin{equation}\label{conditional state bob}
\rho_{B}^i = \mbox{Tr}_{1234}[ \hat{M}_{i} \rho_{T}\hat{M}_ {i}^\dagger],
\end{equation}
upto some normalization constant $N_i$ which is the probability of occurrence of outcome $\rho^B_i$, is
\begin{equation}\label{normalization constant}
N_i = \mbox{Tr}[\hat{M}_{i} \rho_{T}\hat{M}_{i}^\dagger].
\end{equation}
Then we calculate the quantum discord of the state $ \rho_{B}^i $ as given by \cite{zurek}. Since there are sixteen different outcomes possible for Alice, therefore there are sixteen different $ \rho_{B}^i$'s. Note that as Alice is communicating her results to Bob, Bob's state would collapse to one of the $\rho_B^i$'s. Thus, we compute the discord of each such state and then average it over (which means multiplying the probability $(N_i)$ of the occurrence of state to the quantum discord of the state $\rho_{B}^i$) to find the average discord $\overline{\delta}$.
\begin{equation}\label{average discord}
\overline{\delta} = \frac{\sum_{i} N_i \delta(\rho_{B}^i)}{\sum_{i} N_i}.
\end{equation}
Interestingly, from the Fig. \ref{fig1:}, we find that upto $\lambda\approx 0.82047$ the average discord of Bob's state is more than the initial Werner state posses by Alice.
\begin{figure}
\includegraphics[width=\linewidth]{discord_werner_state_shared_4-qubit_omega_state.pdf}
\caption{The red line shows the quantum discord as function of parameter $\lambda$ for the initial Alice's Werner state. The green line shows the average discord of the final states with variation in $\lambda$ (using the state (\ref{shared state 2}) as the channel).}
\label{fig3:}
\end{figure}
Note that for some of the post-selected outcomes $\rho_B^i$'s have a higher Quantum discord than the initial Werner state. Hence, we calculate the average discord and it comes out to be more than that of Alice's initial state for most of the range of $\lambda$. However, here we are not interested in the individual states which have a higher quantum discord
and focus on the quantum discord that Bob would find after averaging over all the outcomes.
\subsection{Example 2}
\label{sec:sec2.2}
We took another 4-qubit channel,
\begin{eqnarray}
\ket{\Psi}_{\Omega} =&& \frac{1}{\sqrt{6}} \Big(\ket{00}_A\ket{11}_B + \ket{01}_A\ket{01}_B + \ket{01}_A\ket{10}_B+\ket{10}_A\ket{01}_B \nonumber\\&&\quad\quad+ \ket{10}_A\ket{10}_B + \ket{11}_A\ket{00}_B\Big).\label{shared state 2}
\end{eqnarray}
We choose the measurement basis same as in Eq. (\ref{four dim basis}). Here also we get a similar kind of result. From Fig. \ref{fig3:}, it is clear that upto $\lambda\approx0.8876$ the average discord of final state is larger than the initial state.
\subsection{Example 3}
\label{sec:sec2.3}
We can get similar kind of result if we take a 3-qubit state as a channel instead of the 4-qubit channel. However, Alice needs to send only 3 bits of classical information for this case. Let's take a three qubit W state
\begin{equation}\label{shared state 3}
\ket{\Psi}_{W} = \frac{1}{\sqrt{3}} \Big(\ket{0}_A\ket{01}_B + \ket{0}_A\ket{10}_B + \ket{1}_A\ket{00}_B\Big).
\end{equation}
We choose the measurement basis as
\begin{eqnarray}\label{three dim basis}
&&\ket{b_1} = \frac{1}{2}\Big(\ket{000}+\ket{100}+\ket{011}+\ket{111}\Big),\nonumber\\
&&\ket{b_2} = \frac{1}{2}\Big(\ket{000}+\ket{100}-\ket{011}-\ket{111}\Big),\nonumber\\
&&\ket{b_3} = \frac{1}{2}\Big(\ket{000}-\ket{100}+\ket{011}-\ket{111}\Big),\nonumber\\
&&\ket{b_4} = \frac{1}{2}\Big(-\ket{000}+\ket{100}+\ket{011}-\ket{111}\Big),\nonumber\\
&&\ket{b_5} = \frac{1}{2}\Big(\ket{001}+\ket{010}+\ket{101}+\ket{110}\Big),\nonumber\\
&&\ket{b_6} = \frac{1}{2}\Big(\ket{001}-\ket{010}-\ket{101}+\ket{110}\Big),\nonumber\\
&&\ket{b_7} = \frac{1}{2}\Big(\ket{001}+\ket{010}-\ket{101}-\ket{110}\Big)\,\mbox{and}\nonumber\\
&&\ket{b_8} = \frac{1}{2}\Big(-\ket{001}+\ket{010}-\ket{101}+\ket{110}\Big).
\end{eqnarray}
We get similar kind of behavior like two previous examples. From Fig. \ref{fig5:}, it is clear that upto $\lambda\approx0.7722$ the average discord of final state is larger than the initial state.
\begin{figure}[t!]
\includegraphics[width=\linewidth]{discord_werner_state_shared_3-qubit_W_state.pdf}
\caption{The red line shows the quantum discord as function of parameter $\lambda$ for the initial Alice's Werner state. The green line shows the average discord of the final states with variation in $\lambda$ (using the W state (\ref{shared state 3}) as the channel).}
\label{fig5:}
\end{figure}
\begin{figure}
\includegraphics[width=\linewidth]{telefidclusterstate2.pdf}
\caption{The red line shows the teleportation fidelity for the Werner state as variation in $\lambda$, the green line shows the upper bound or the maximum average fidelity possible for the final states res. as $\lambda$ is varied (using the cluster state given in \ref{sec:sec2.1} as the channel).}
\label{fig2:}
\end{figure}
\section{Application}
\label{sec:sec3}
Essentially Werner state is maximally entangled state with some isotropic noise added. Consider a scenario where Alice wants to teleport a 1-qubit state using the Werner state as the channel. We ask the question that if the above protocol (see section~\ref{sec:sec2}) is applied and then the final states are used as channels instead of the initial Werner state, can one increase the teleportation fidelity?
Verstraete and Verschelde in \cite{verstraete} have found the upper and lower bounds of fidelity for any teleportation scheme. Given the channel $\rho$ the fidelity $\mathcal{F}^*(\rho)$ is
\begin{equation}
\frac{1}{2}\bigg(1 + \frac{\mathcal{N}(\rho)}{1+\sqrt{1-(\frac{\mathcal{N}(\rho)}{\mathcal{C}(\rho)})^2}}\bigg) \leq \mathcal{F}^*(\rho) \leq \frac{1}{2}(1 + \mathcal{N}(\rho)),
\end{equation}
where $\mathcal{N}(\rho)$ is the negativity \cite{Vidal} of the state $\rho$ and $\mathcal{C}(\rho)$ is the concurrence \cite{Wootters}. We calculate these two quantities as given in \cite{Vidal,Wootters} for the initial state and then find the bounds. For Werner state these two bounds are same. We do the same calculation for the final states (which can be obtained by applying our protocol as described in section \ref{sec:sec2}) to find the upper and lower bounds and then average it over. The average teleportation fidelity $\overline{\mathcal{F}^*(\rho)}$ will belong to this range specified by lower and upper bound. We will compare our results graphically for those three examples described in the previous section. In the figures we have only showed the upper bound as this is the maximum achievable fidelity.
First we consider the example in subsection \ref{sec:sec2.1}. We find (as shown in Fig. \ref{fig2:}) that when the final states are used as the resource state instead of the Werner state, the upper bound or the maximum average teleportation fidelity is more than the fidelity of the initial Werner state (for most of the range of $\lambda$). Interestingly, we find that the upper bound of the average fidelity and discord both are decreasing with $\lambda$. This can be easily checked by looking at Fig.~\ref{fig1:} and Fig.~\ref{fig2:}. Moreover, we find that the $\lambda$ for which the average quantum discord of the final states is equal to the initial Werner state is different from the $\lambda$ where the teleportation fidelity using the Werner state is equal to the fidelity when using the final states. In this region one can see that average quantum discord for the final states is more, but the teleportation fidelity is less than the initial Werner state. From Fig.~\ref{fig1:} and Fig.~\ref{fig2:}, this region corresponds to the values from $\lambda \approx 0.678$ to $\lambda \approx 0.82047$. This difference may be due to the fact that quantum discord and quantum teleportation are not comparable. Proper optimization of the measurement basis may reduce this gap. Nonetheless, one important fact is that even when the initial Werner state is separable, we can use the final state as a channel for quantum teleportation.
Now we find the teleportation fidelity of the final state for the example given in subsection \ref{sec:sec2.2}. We see a similar nature as the 4-qubit cluster state. In this example as we can see from Fig~\ref{fig3:} and Fig~\ref{fig4:}, the region of $\lambda$ (for which quantum discord for the final states are more in this region, but the teleportation fidelity is less than the initial Werner state) is from $\lambda \approx 0.7582$ to $\lambda \approx 0.8876$.
\begin{figure}[t!]
\includegraphics[width=\linewidth]{telefidomegastate.pdf}
\caption{The red line shows the teleportation fidelity for the Werner state as variation in $\lambda$, the green line shows the upper bound or the maximum average fidelity possible for the final states res. as $\lambda$ is varied (using the state given in \ref{sec:sec2.2} as the channel).}
\label{fig4:}
\end{figure}
\begin{figure}[t!]
\centering
\includegraphics[width=\linewidth]{telefid3Wstate.pdf}
\caption{The red line shows the teleportation fidelity for the Werner state as variation in $\lambda$, the green line shows the upper bound or the maximum average fidelity possible for the final states res. as $\lambda$ is varied (using the state given in \ref{sec:sec2.3} as the channel).}
\label{fig6:}
\end{figure}
Again we find the teleportation fidelity using the 3-qubit W state as the resource state. The measurement basis is given by Eq.~(\ref{three dim basis}). For the 3-qubit state the nature of teleportation fidelity is different from the 4-qubit states. From Fig.~\ref{fig5:} and Fig.~\ref{fig6:} we see that the teleportation fidelity behaves completely opposite of the quantum discord. The quantum discord decreases, but the teleportation fidelity increases as $\lambda$ increases. However, in this case the region (for which quantum discord for the final states are more in this region, but the teleportation fidelity is less than the initial Werner state) is from $\lambda \approx 0.6858$ to $\lambda \approx 0.7722$. Although we see for the examples considered above that when quantum discord is more than the initial state discord, the average teleportation fidelity is also more than the fidelity of the initial state (for some range of $\lambda$) but a direct connection can't be established from these results. These above examples give us some insight about the usefulness of discord in such a protocol. But whether it will be true for any general case needs to be further explored.
\section{Conclusion}
\label{sec:sec4}
We have shown that it is possible for Alice to prepare a state at Bob with a higher amount of quantum discord by sharing an entangled channel, local operations and classical communications. We also see that when the quantum discord of Alice's state increases then the average quantum discord of Bob's state decreases. Moreover, we showed above that the increment of the quantum discord in our protocol may be used as a way to increase fidelity of one qubit teleportation. Our results could also be applied to other quantum communication tasks like remote state preparation, entanglement distribution, etc. However our results are not optimized, so it might be possible that there could exist some basis for the given quantum channel such that the quantum discord of Alice's Werner state is always less than Bob's prepared state.
\section{Acknowledgment}
We would like to thank Satyabrata Adhikari and Prasanta Kumar Panigrahi for useful discussions.
|
\section{$F$-subharmonic functions}
\renewcommand{Pos}{\operatorname{Pos}}
\newcommand{\mathcal P}{\mathcal P}
\newcommand{\operatorname{Int}}{\operatorname{Int}}
\setcounter{tocdepth}{1}
\section{Introduction}\label{sec:introduction}
The first part of this paper is a proof of the following elementary statement about regularity of certain argmin functions.
\begin{theorem}\label{thm:intoargmin1} \
Suppose $f:\mathbb R^{n}\times \mathbb R^{m}\to \mathbb R$ is such that
\begin{enumerate}
\item There are $\kappa\ge 0$ and $\sigma>0$ so $$f(x,y) + \frac{\kappa}{2} \|x\|^2-\frac{\sigma}{2}\|y\|^2\text{ is convex.}$$
\item For each $x\in \mathbb R^{n}$ there is a unique $\gamma(x)$ such that
$$\inf_y f(x,y) = f(x,\gamma(x)).$$
\end{enumerate}
Then the function $\gamma$ is differentiable almost everywhere.
\end{theorem}
Our motivation for this is the following. Following Harvey-Lawson \cite{HL_Dirichletduality,HL_Dirichletdualitymanifolds}, by a \emph{(primitive) subequation} on an open $X\subset \mathbb R^n$ we mean a subset $F\subset J^2(X)$ of the space of $2$-jets on $X$ with certain properties. Given such an $F$ and a $\mathcal C^2$ function $f$, we say that $f$ is \emph{$F$-subharmonic} if every $2$-jet of $f$ lies in $F$. Moreover, using the so-called viscosity technique it is possible to extend the notion of $F$-subharmonicity to any upper-semicontinuous function (details and precise definitions will be given in \S\ref{sec:Fsubharmonics}).
In our previous work \cite{MP1} we introduced a notion of ``product subequation" $F\#\mathcal P$ on $X\times \mathbb R^m$ and show (under suitable hypothesis) that if $F$ is convex and $f$ is $F\#\mathcal P$-subharmonic then its marginal function
$$ g(x) := \inf_y f(x,y)$$
is $F$-subharmonic. This statement generalises the classical statement that the marginal function of a convex function is again convex. We will use Theorem \ref{thm:intoargmin1} to prove a similar minimum principle that does not require $F$ to be convex:
\begin{theorem}\label{thm:introminimum}
Let $X\subset \mathbb R^n$ be open and $F\subset J^2(X)$ be a constant-coefficient primitive subequation that depends only on the Hessian part. Suppose
$$f:X\times \mathbb R^m\to \mathbb R$$
is locally semiconcave, bounded from below, and $F\#\mathcal P$-subharmonic. Then the marginal function
$$ g(x) = \inf_{y} f(x,y)$$
is $F$-subharmonic on $X$.
\end{theorem}\vspace{1mm}
\noindent A few remarks are in order.\vspace{2mm}
\begin{asparaenum}\setlength{\itemsep}{3mm}
\item The semiconcavity assumption on $f$ is rather unnatural, since one would expect a subsolution to have some kind of convexity rather than concavity, but it captures what we are able to prove. Observe that $f$ is certainly locally semiconcave if it is $\mathcal C^{1,1}_{loc}$.
\item The assumption that $f$ is $F\#\mathcal P$-subharmonic implies that for each $x$ the function $y\mapsto f(x,y)$ is convex. This along with the semiconcavity assumption implies that $y\mapsto f(x,y)$ is $\mathcal C^{1,1}_{loc}$.
\item Theorem \ref{thm:introminimum} can be proved rather easily when $f$ is $\mathcal C^2$ (see \cite[Prop. 7.5]{MP1} for a stronger statement). To do so, we first approximate $f$ by adding a small multiple of the function $(x,y)\mapsto \|y\|^2$, so there is no loss in assuming $f$ is strictly convex in $y$ and that for each fixed $x$ the function $y\mapsto f(x,y)$ attains its unique minimum at some point $\gamma(x)$. Said another way, $\gamma(x)$ is the unique point such that
$$ \frac{\partial f}{\partial y}|_{(x,\gamma(x))} = 0.$$
If we assume $f$ is $\mathcal C^2$ we can then:
\begin{enumerate}[(a)]
\item Use the implicit function theorem to deduce that $\gamma$ is $\mathcal C^1$.
\item Use the chain rule to compute the Hessian of $g$ at a point $x$ in terms of the Hessian of $f$ at the point $(x,\gamma(x))$ and the derivative of $\gamma$ at $x$.
\end{enumerate}
The combination of (b) and assumption that $f$ is $F\#\mathcal P$-subharmonic yields that $g$ is $F$-subharmonic as claimed.
\item If we assume furthermore that $F$ is convex, then using smooth mollification to approximate any upper-semicontinuous $F\#\mathcal P$-subharmonic function by those that are $\mathcal C^2$, we can deduce a much more general minimum principle -- this is the approach taken in \cite{MP1}.
\item If instead we assume that $f$ is merely $\mathcal C^{1,1}_{loc}$ then it is of course twice differentiable almost everywhere. However it may well be that $f$ is not twice differentiable at any point of the form $(x,\gamma(x))$ so part (b) of the above argument does not apply.\medskip
\end{asparaenum}
To prove Theorem \ref{thm:introminimum} we will first use a partial-sup convolution to approximate $f$ by $F\#\mathcal P$-subharmonic functions $f_{\epsilon}$ such that
$$ f_{\epsilon}(x,y) + \frac{1}{2\epsilon} \|x\|^2 - \frac{\epsilon}{2}\|y\|^2 \text{ is convex}.$$
In particular for fixed $x$ the function $y\mapsto f_{\epsilon}(x,y)$ is strongly convex, and we will further arrange so the argmin of ${f_{\epsilon}}$ is a well-defined single-valued function $\gamma$. Having done so we can apply Theorem \ref{thm:intoargmin1} to deduce that $\gamma$ is differentiable almost everywhere, which will act in lieu of the implicit function argument used in (a). From this one can prove, essentially from the definition, that at almost every point $x$ the Hessian of $g$ is contained in $F$. As $g$ is semiconvex, this is known by the Almost-Everywhere Theorem of Harvey-Lawson \cite{HL_AE} to be enough to conclude that $g$ is $F$-subharmonic.
\subsection*{Comparison with other work: }
The authors do not have sufficient expertise to properly survey all previously known regularity results that are related to Theorem \ref{thm:intoargmin1}. Suffice to say there has been much interest in studying regularity of marginal functions (by which we mean functions of the form $\inf_y f(x,y)$ or $\sup_y f(x,y)$ for some function $f$ which also go under the name ``performance function") due to its relevance for optimization problems (see for instance \cite{Aubin1,Quasidifferentiabilityandrelatedtopics,non-smooth,Globaloptimzationinaction} and the references therein). For example, various regularity properties of marginal functions have been shown when $f$ has some convexity property (see for example \cite[Theorems 23.4 and 24.5]{Rockafellar_Convex_analysis}) and without this convexity hypothesis (e.g. \cite{Clarke_generalizedgradients,Penot1982a,Penot04,Penot04b,ward1} to list just a few).
Much less appears to have been written about regularity of the argmin function itself. We remark that in general the argmin function will be multi-valued, and so regularity must be phrased in terms of set-valued functions \cite{Clarke_generalizedgradients}. The only previous such results we have found relate to continuity rather than differentiability (for example \cite[Theorem 2.10]{wets_Lipschitzcontinuity}, which is taken from \cite[Theorems 1.17 and 7.41]{Rockafellar_Wets_VariationalAnalysis}, gives conditions under which the argmin function is outer semicontinuous).
Regarding the minimum principle, the fact that the marginal function of a convex function is again convex is a basic property in convex analysis. In the complex case this has an analog for plurisubharmonic functions due to Kiselman \cite{KiselmanInvent,Kiselman}. Both convexity and plurisubharmonicity are massively generalized through the notion of $F$-subharmonic functions which uses the viscosity technique that arose in the study of fully non-linear degenerate second-order differential equations (in particular the work of Caffarelli--Nirenberg--Spruck \cite{CaffarelliNirenbergSpruckIII} and Lions--Crandall--Ishii \cite{CILUserGuide}, who often refer to such functions as \emph{subsolutions}).
Our motivation for introducing the product $F\#\mathcal P$ came from a desire to generalise this minimum principle to general subequations, which we do in \cite{MP1} under the assumption that $F$ is convex. As discussed above, this assumption is needed only to be able to approximate $F\#\mathcal P$-subharmonic functions by smooth ones, and thus suggests that it is a facet of the proof rather than an essential requirement. Theorem \ref{thm:introminimum} is, as far as we know, the first such minimum principle that does not require any convexity hypothesis on the subequation in question. For further background in this area the reader is referred to \cite{MP1}.\\
\noindent {\bf Organization: } Section \ref{sec:lipschitzargmin} is devoted to the proof of Theorem \ref{thm:intoargmin1}. In \S\ref{sec:statement} we recall some standard terminology and notation concerning semiconvex functions, and use this to give a refined statement (Theorem \ref{thm:argmin1}) about calmness of the argmin function. Theorem \ref{thm:intoargmin1} then follows immediately from this by Stepanov's Theorem (see Corollary \ref{cor:wisdiffae}). In \S\ref{sec:propertiessemiconvex} we collect some further properties of semiconvex functions, in preparation for \S\ref{sec:proofargminthm} in which we give a functional equation for the argmin function. Then the proof of Theorem \ref{thm:argmin1} is given in \S\ref{sec:proofargmin1} using the Implicit Function Theorem for Lipschitz maps (which for completeness is proved in Appendix \ref{sec:appendiximplicit}).
In Section \ref{sec:Fsubharmonics} we summarize the basics of $F$-subharmonic functions in a way suited to our needs, including the idea of product subequations in \S\ref{sec:productsubequations}. In Section \ref{sec:partialsup} we describe the partial sup-convolution, which is used in Section \ref{sec:fsubharmonicityofmarginals} to complete the proof of Theorem \ref{thm:introminimum}.
\section{Differentiability of the Argmin Function}\label{sec:lipschitzargmin}
\subsection{Statement}\label{sec:statement}
In this section we prove that the argmin function of a certain kind of semiconvex functions is differentiable (resp.\ calm) almost everywhere. Suppose $\Omega\subset \mathbb R^{n+m}$ is open and let $\pi:\mathbb R^{n+m}\to \mathbb R^n$ be the projection, and write $$\Omega_x = \{ y\in \mathbb R^m : (x,y)\in \Omega\}.$$ We will assume throughout that $\Omega$ is convex and that each $\Omega_x$ is connected. Now suppose
$$f:\Omega \to \mathbb R$$ and set
$$ g(x) : = \inf_{y\in \Omega_x} f(x,y) \text{ for } x\in \pi(\Omega).$$
\begin{definition}[Argmin]
The \emph{argmin function} is the set-valued function
$$ \operatorname{argmin}_f(x) : = \{ \gamma\in \Omega_x : \inf_{y\in \Omega_x} f(x,y) = f(x,\gamma)\}$$
where we allow the possibility that $\operatorname{argmin}_f(x)$ is empty.
\end{definition}
Below we shall make assumptions on $f$ that ensure that $\operatorname{argmin}(x)$ is everywhere defined and single-valued. In such cases we shall write
$$ \gamma(x) = \operatorname{argmin}_f(x)$$
so
$$ f(x,\gamma(x)) = \inf_{y\in \Omega_x} f(x,y) =g(x) \text{ for all } x\in \pi(\Omega).$$
The precise statement we will prove requires some terminology concerning subdifferentials. Let $X\subset \mathbb R^n$ be open.
\begin{definition}
Suppose $g:X\to \mathbb R$. For each $x_0\in X$ define
$$\nabla_{x_0} g = \{ u\in \mathbb R^n: g(x) - g(x_0) \ge u.(x-x_0) \text{ for all }x \text{ sufficiently near } x_0 \}$$
which may be empty. We call any $u\in \nabla_{x_0}g$ \emph{a lower support vector} for $g$ at $x_0$. Similarly if $\kappa\in \mathbb R$ we let
$$ \nabla^\kappa_{x_0} g = \{u\in \mathbb R^n : g(x) - g(x_0) \ge u\cdot(x-x_0) - \frac{\kappa}{2} \|x-x_0\|^2 \text{ for all } x \text{ sufficiently near } x_0\}.$$
\end{definition}
\begin{definition}[Semiconvexity and Semiconcavity]
Let $\kappa\ge 0$. We say $g:X\to \mathbb R$ is \emph{$\kappa$-semiconvex} (resp. \emph{$\kappa$-semiconcave}) if $g(x) + \frac{\kappa}{2} \|x\|^2$ is convex (resp. $g(x) - \frac{\kappa}{2} \|x\|^2$ is concave). If $g$ is $\kappa$-semiconvex/semiconcave for some $\kappa \ge 0$ then we say simply $g$ is \emph{semiconvex/semiconcave}.
\end{definition}
\begin{remark}
In the literature one will also find the term \emph{weakly-convex/concave} also used for semiconvex/semiconcave.
\end{remark}
One can check that $g:X\to \mathbb R$ is locally $\kappa$-semiconvex if and only if $\nabla_{x_0}^{\kappa}g$ is non-empty for all $x_0$. Moreover $g$ is differentiable at $x_0$ if and only if $\nabla_{x_0}g$ is a singleton, in which case its unique element is the derivative of $g$ at $x_0$. Finally if $g$ is a convex function on a convex set $X$ then
$$ \nabla^\kappa_{x_0} g = \{u\in \mathbb R^n : g(x) - g(x_0) \ge u\cdot(x-x_0) - \frac{\kappa}{2} \|x-x_0\|^2 \text{ for all } x \}.$$
We now give a refined statement of Theorem \ref{thm:intoargmin1} that will be proved in section \S\ref{sec:proofargminthm}.
\begin{theorem}[Argmin is calm almost everywhere]\label{thm:argmin1} \
Let $\Omega\subset \mathbb R^{n+m}$ be open, convex and so that $\Omega_x$ is connected for all $x$. Also let $f:\Omega\to \mathbb R$ and suppose there are $\kappa\ge 0$ and $\sigma>0$ so that
\begin{equation}\label{eq:condagmin1}f(x,y) + \frac{\kappa}{2} \|x\|^2-\sigma \|y\|^2\text{ is convex and}\end{equation}
\begin{equation}\label{eq:condagmin2}
\operatorname{argmin}_f(x) \text{ is non-empty for all } x\in \pi(\Omega).
\end{equation}
Then
\begin{enumerate}[(i)]
\item The function $$g(x):=\inf_y f(x,y) = f(x,\gamma(x)) \text{ for }x\in\pi(\Omega)$$ is $\kappa$-semiconvex and $\gamma(x) := \operatorname{argmin}_f(x)$ is single valued for all $x\in \pi(\Omega)$.
\item Given any $x_0\in \pi(\Omega)$ and $u_0\in \nabla_{x_0}^{\kappa} g$ there exists a Lipschitz function $$\phi:V\to \mathbb R$$ defined on a neighbourhood $V$ of $(x_0,u_0)$ in $\mathbb R^{2n}$ such that
$$ \gamma(x) = \phi(x,u) \text{ for all } (x,u)\in V \text{ with } u\in \nabla^\kappa_{x} g.$$
\item The function $\gamma$ is calm almost everywhere. That is, for almost all $x_0\in \pi(\Omega)$ there are $C$ and $\delta>0$ such that
\begin{equation} \|\gamma(x) - \gamma(x_0)\| \le C \| x-x_0\|\text{ for } \|x-x_0\|<\delta.\label{eq:wlocallylipschitzae}\end{equation}
\end{enumerate}
\end{theorem}
\begin{corollary}[The Argmin is Differentiable Almost Everywhere]\label{cor:wisdiffae}
Under the hypothesis of the Theorem the argmin function $\gamma$ is differentiable almost everywhere.
\end{corollary}
\begin{proof}
This follows from \eqref{eq:wlocallylipschitzae} and Stepanov's Theorem \cite{Stepanoff} (see also \cite[Theorem 3.4] {Heinonen_lectures}).
\end{proof}
The strategy of the proof of Theorem \ref{thm:argmin1} is to construct a functional equation satisfied by the argmin function, and then apply the implicit function theorem for Lipschitz functions. In the next section we setup the necessary machinery to do so.
\subsection{Properties of semiconvex functions}\label{sec:propertiessemiconvex}
We collect a few basic statements about convex and semiconvex functions. As above $\Omega\subset \mathbb R^{n+m}$ is open, convex and $\Omega_x$ is connected for all $x$.
\begin{lemma}\label{lem:semiconvexgrad}
Suppose $f:\Omega\to \mathbb R$ and $\tilde{f}(x,y) = f(x,y) + \kappa\frac{\|x\|^2}{2}$. Then
$$ \nabla_{(x_0,y_0)} f = \nabla_{(x_0,y_0)} \tilde{f} + \kappa x_0$$
as sets.
\end{lemma}
\begin{proof}
This is immediate from the definition, and left to the reader.
\end{proof}
\begin{lemma}\label{lem:basicconvexityI}
Suppose $f: \Omega\to \mathbb R$ and $h:\mathbb R^m\to \mathbb R$ and set $$\hat{f}(x,y) = f(x,y) + h(x) \text{ for } (x,y)\in \Omega.$$ Then
\begin{equation}\operatorname{argmin}_{\hat{f}} (x) = \operatorname{argmin}_{f}(x) + h(x).\label{eq:argminshift}\end{equation}
In particular $\operatorname{argmin}_{\hat{f}}$ is single-valued if and only if $\operatorname{argmin}_f$ is single-valued.
\end{lemma}
\begin{proof}
Clearly
$$\operatorname{argmin}_{\hat{f}}(x) = \{ \gamma: \gamma = \inf_y \hat{f}(x,y) \} = \{ \gamma :\gamma= \inf_{y} f(x,y) + h(x)\} = \operatorname{argmin}_f(x) + h(x)$$
giving \eqref{eq:argminshift}. The last statement follows immediately.
\end{proof}
\begin{lemma}\label{lem:basicconvexityII}
Let $f:\Omega\to \mathbb R$ and suppose $f(x,y) + \frac{\kappa}{2}\|x\|^2$ is convex. Then $g(x) = \inf_y f(x,y)$ is $\kappa$-semiconvex.
\end{lemma}
\begin{proof}
Write $\tilde{f}(x,y) = f(x,y) + \frac{\kappa}{2} \|x\|^2$ so
$$ g(x) +\frac{\kappa}{2} \|x\|^2 = \inf_y \tilde{f}(x,y)$$
which is the marginal function of convex function defined on $\Omega$ (which is assumed to be convex and $\Omega_x$ is connected for each $x$). Thus $g(x) + \frac{\kappa}{2} \|x\|^2$ is convex.
\end{proof}
\begin{lemma}[Gradient at argmin]\label{lem:gradientatargmax}
Suppose that $f:\Omega\to \mathbb R$ is convex and set $g(x) = \inf_y f(x,y)$. Then for all $x\in \pi(\Omega)$ and $\gamma\in \operatorname{argmin}_f(x)$
$$u\in \nabla_{x} g \Rightarrow (u,0)\in \nabla_{(x,\gamma)} f.$$
\end{lemma}
\begin{proof}
Suppose $\gamma\in \operatorname{argmin}_f(x)$ so $g(x) = f(x,\gamma)$. Let $u\in \nabla_{x} g$. Then for any $(x',y')\in \mathbb R^{n}\times \mathbb R^m$,
\begin{align}
f(x',y') - f(x,\gamma)& \ge g(x') - g(x) \\ &\ge u.(x'-x) = (u,0). ((x',y') - (x,\gamma))
\end{align}
so $(u,0)\in \nabla_{(x,\gamma)}f$ as claimed.
\end{proof}
The next statement is a slight modification of \cite[Proposition 6.4]{Howard}.
\begin{proposition}\label{prop:thefunctionsGH}
Let $\sigma>0$ and suppose $f:\mathbb R^{n+m}\to \mathbb R$ is such that $f(x,y) - \frac{\sigma}{2} \|y\|^2$ is convex. Define the set-valued function
$$ G(p) = p + \nabla_{p} f \text{ for } p=(x,y)\in \mathbb R^{n+m}.$$
Then
\begin{enumerate}[(i)]
\item $G$ is non-contractive. That is, if $\zeta_i \in G(p_i)$ for $i=1,2$ then
\begin{equation}\label{eq:Fnoncontract}\| \zeta_1-\zeta_2\| \ge \|p_1-p_2\|.\end{equation}
\item There exist a single-valued function $H:\mathbb R^{n+m} \to \mathbb R^{n+m}$ that is inverse to $G$, by which we mean
\begin{equation} H(\zeta) = p \Longleftrightarrow \zeta \in G(p).\label{eq:inverse}\end{equation}
\item The function $H$ is Lipschitz with Lipschitz constant $1$. Moreover there is a $\mu<1$ such that letting $\pi_2:\mathbb R^{n+m}\to \mathbb R^m$ denote the second projection,
\begin{equation}
\| \pi_2 H(\zeta_1)-\pi_2 H(\zeta_2)\| \le \mu \| \zeta_1-\zeta_2\| \text{ for all } \zeta_1,\zeta_2\in\mathbb R^{n+m}.\label{eq:inequalitylambda}
\end{equation}
\end{enumerate}
\end{proposition}
\begin{proof}
Let $p_i:= (x_i,y_i)\in \mathbb R^{n+m}$ for $i=1,2$. We first claim
\begin{equation}\label{eq:thefunctionsGH1}
(\nabla_{p_2} f- \nabla_{p_1}f ).(p_2-p_1) \ge \sigma \|y_2-y_1\|^2 \text{ for all } (x_i,y_i)\in \mathbb R^{n+m}.
\end{equation}
To see this, let $\tilde{f}(x,y) = f(x,y) - \frac{\sigma}{2} \|y\|^2$ which by assumption is convex and $\nabla_{(x,y)} \tilde{f} = \nabla_{(x,y)} f - (0,\sigma y)$. Then
\begin{equation}\label{eq:thefunctionsGH2}
\tilde{f}(p_1) - \tilde{f}(p_2) \ge \nabla_{p_2}\tilde{f}.(p_1-p_2) = \nabla_{p_2} f.(p_1-p_2) - \sigma y_2.(y_1-y_2).
\end{equation}
Swapping the indices we also have
\begin{equation}\label{eq:thefunctionsGH3}
\tilde{f}(p_2) - \tilde{f}(p_1) \ge \nabla_{p_1}\tilde{f}.(p_2-p_1) = \nabla_{p_1} f.(p_2-p_1) - \sigma y_1.(y_2-y_1).
\end{equation}
Adding \eqref{eq:thefunctionsGH2} and \eqref{eq:thefunctionsGH3} and rearranging gives \eqref{eq:thefunctionsGH1}.
Now from Cauchy-Schwarz and \eqref{eq:thefunctionsGH1}
\begin{align}
\| G(p_1) - G(p_2)\| \|p_1-p_2\| &\ge ( G(p_1) - G(p_2)).(p_1-p_2) \nonumber \\
&= (p_1 -p_2 + \nabla_{p_1} f- \nabla_{p_2}f).(p_1-p_2) \nonumber\\
&\ge \|p_1-p_2\|^2 + \sigma \|y_1-y_2\|^2\label{eq:strict}\\
&\ge \|p_1-p_2\|^2 \nonumber\end{align}
which in particular implies (i).
We claim next that $G$ is surjective, by which we mean for all $\zeta\in \mathbb R^{n+m}$ there is an $p\in \mathbb R^{n+m}$ such that $\zeta\in G(p)$. To see this let
$$\phi(p):= \frac{1}{2} \|p\|^2 + f(p) -p.\zeta.$$
The function $p\mapsto \frac{1}{2}\|p\|^2 -p.\zeta$ is convex, and hence so is $\phi$ and
$$\nabla_{p_0} \phi = \nabla_{p_0} f+ p_0 - \zeta = G(p_0)-\zeta.$$
Similarly the function
$$\psi(p): = \frac{1}{4} \|p\|^2 + f(p) -p.\zeta = \phi(p) - \frac{1}{4}\|p\|^2$$ is convex. Pick $b\in \nabla_{0} \psi$ so $\psi(p) -\psi(0) \ge b.p$ giving
$$\phi(p) \ge \phi(0) + \frac{1}{4}\|p\|^2.$$
As $\phi$ is continuous this implies $\phi$ has a global minimum at some $p_0\in \mathbb R^{n+m}$, and so $0$ is a lower support vector for $\phi$ at $0$. Thus $0\in \nabla_{p_0}\phi = G(p_0)-\zeta$ implying that $\zeta\in G(p_0)$. Thus $G$ is surjective as claimed.
In particular the inverse $H$ to $G$ defined by
$$ H(\zeta) = \{ p\in \mathbb R^{n+m}: \zeta\in G(p)\}$$
is non-empty, and $G$ being non-contractive implies that it is single-valued. That $H$ has Lipshitz constant $1$ follows from (i).
Finally given $\zeta_1,\zeta_2$ set $p_i:=(x_i,y_i):= H(\zeta_i)$ so by definition $\zeta_i\in G(p_i)$ and $y_i=\pi_2 H(\zeta_i)$ .
To ease notation let $\alpha: = \|x_1-x_2\|$ and $\beta: = \|y_1-y_2\| = \| \pi_2 H(\zeta_1) - \pi_2 H(\zeta_2)\|$. Then dividing \eqref{eq:strict} by $\|p_1-p_2\|$ gives
$$ \|\zeta_1-\zeta_2\| \ge (\alpha^2 + \beta^2)^{1/2}+ \sigma \frac{\beta^2}{(\alpha^2 + \beta^2)^{1/2}}.$$
If $\alpha \ge \sigma \beta$ then $\|\zeta_1-\zeta_2\|\ge (1+\sigma^2)^{1/2} \beta$. If $\alpha\le \sigma \beta$ then
\begin{align*} \|\zeta_1-\zeta_2\| &\ge \beta + \sigma \frac{\beta^2}{(\sigma^2\beta^2+ \beta^2)^{1/2}}\\
&= (1 + \frac{\sigma}{(1+\sigma^2)^{1/2}}) \beta.
\end{align*}
Hence \eqref{eq:inequalitylambda} holds with $\mu: = \min\{ (1+\sigma^2)^{1/2}, (1 + \frac{\sigma}{(1+\sigma^2)^{1/2}})\}^{-1}<1$.
\end{proof}
We will also need the following simpler corollary (which is proved in the same way, or follows formally from Proposition \ref{prop:thefunctionsGH} upon taking $m=0$).
\begin{corollary}\label{cor:thefunctionsGH}
Suppose $g:\mathbb R^n\to \mathbb R$ is convex and define the set-valued function
$$ G_1(x) = x + \nabla_{x} g \text{ for } x\in \mathbb R^n.$$
Then
\begin{enumerate}
\item $G_1$ is non-contractive, that is
\begin{equation}\|G_1(x_1) - G_1(x_2)\| \ge \|x_1 -x_2\| \text{ for all } x_1,x_2\in X.\label{eq:hefunctionsGH:Gnoncontractive:repeat}\end{equation}
\item There exist a single-valued function $H_1:\mathbb R^{n} \to \mathbb R^{n}$ that is inverse to $G_1$, and $H_1$ is Lipschitz with Lipschitz constant 1.
\end{enumerate}
\end{corollary}
\subsection{Functional Equation for argmin}\label{sec:proofargminthm}
Suppose now that $f:\mathbb R^{n+m} \to \mathbb R$ is convex and as usual let $g(x) = \inf_y f(x,y)$ which is also convex. Consider the set-valued functions
\begin{align*} G_1(x) &= x + \nabla_x g,\\
G(x,y) &= (x,y) + \nabla_{(x,y)} f.\end{align*}
By Proposition \ref{prop:thefunctionsGH} and Corollary \ref{cor:thefunctionsGH} these have single-valued inverses $H_1:\mathbb R^{n}\to \mathbb R^{n}$ and $H:\mathbb R^{n+m}\to \mathbb R^{n+m}$. That is
\begin{align} H_1(u) = x & \Leftrightarrow u\in G_1(x) \text{ for } x,u\in \mathbb R^{n}\label{eq:Hinverse1}\\
H(u,v) = (x,y) &\Leftrightarrow (u,v)\in G(x,y) \text{ for } (x,y), (u,v) \in \mathbb R^{n+m}.\label{eq:Hinverse2}
\end{align}
We use these to define a functional equation for $\operatorname{argmin}_f$. Let
$$J:\mathbb R^{n}\times \mathbb R^{n}\times \mathbb R^{m} \to \mathbb R^{m}$$
\begin{equation}\label{def:J}J(x,u,y) := y - \pi_2 H(H_1(x+u)+u,y).\end{equation}
\begin{proposition}[Functional Equation for argmin]\label{prop:Jargmin}
Suppose that $f(x,y)$ is convex and let $g(x) = \inf_y f(x,y)$.
Then
$$ J(x,\nabla_x g,\operatorname{argmin}_f(x)) =0 \text{ for all } x\in \mathbb R^n.$$
That is,
$$ J(x,u,\gamma) =0 \text{ for all } x\in \mathbb R^n \text{ and }\gamma\in \operatorname{argmin}_f(x) \text{ and }u\in \nabla_x g.$$
\end{proposition}
\begin{proof}
Let $x\in\mathbb R^{n}$, $\gamma\in \operatorname{argmin}_f(x)$ and $u\in \nabla_x g$. Then $x+u\in G_1(x)$ so \eqref{eq:Hinverse1} gives $H_1(x+u)=x$. On the other hand since $\gamma\in \operatorname{argmin}_f(x)$ we have by Lemma \ref{lem:gradientatargmax}, $$(u,0) \in \nabla_{(x,\gamma)} f.$$
Thus $$ (x,\gamma) + (u,0)= (u+x,\gamma)\in G(x,\gamma)$$ so \eqref{eq:Hinverse1} gives $H(u+x,\gamma) = (x,\gamma)$. So
\begin{align*} J(x,u,\gamma) &= \gamma - \pi_2 H(H_1(x+u)+u,\gamma)\\
&= \gamma - \pi_2 H(x+u,\gamma) = \gamma - \pi_2(x,\gamma) \\&= 0\end{align*}
as claimed.
\end{proof}
We next collect two basic properties of $J$:
\begin{lemma}[Properties of $J$]\label{lem:propertiesJ}
The function $J$ is Lipschitz in $(x,u,y)$. Moreover if $f(x,y) - \frac{\sigma}{2}\|y\|^2$ is convex for some $\sigma>0$ then there is a $\lambda>0$ such that for fixed $x,u$
$$ \|J(x,u,y_1) - J(x,u,y_2)\| \ge \lambda \|y_1-y_2\| \text{ for all } y_1,y_2.$$
\end{lemma}
\begin{proof}
Clearly $J$ is Lipschitz in all variables since both $H$ and $H_1$ are. For the second statement, suppose $f(x,y) - \frac{\sigma}{2}\|y\|^2$ is convex and let $\pi_2:\mathbb R^{n}\times \mathbb R^{m}\to \mathbb R^{m}$ be the second projection. We know from Proposition \ref{prop:thefunctionsGH}(iii) that there is a $\mu<1$ such that
\begin{align}
\| \pi_2H(v,y_1) - \pi_2H(v,y_2)\| &\le \mu \|y_1-y_2\| \text{for all } v,y_1.\end{align}
Now fix $x,u$ and let $v:=H_1(x+u)+u$. Then if $y_1,y_2\in \mathbb R^{m}$,
\begin{align*}
\|J(x,u,y_2) - J(x,u,y_1) \| &= \| y_2-y_1 - \pi_2H(v,y_2) + \pi_2 H(v,y_1)\|\\
&\ge \| y_2 -y_1\| - \| \pi_2H(v,y_2) - \pi_2 H(v,y_1)\|\\
&\ge (1-\mu) \|y_2-y_1\|.
\end{align*}
\end{proof}
\subsection{Statement of Alexandrov's Theorem}
Let $X\subset \mathbb R^n$ be open. The following is a precise version of Alexandrov's Theorem:
\begin{theorem}[Alexandrov's Theorem]\label{thm:alexandrov:repeat}
Let $g:X\to \mathbb R$ be locally convex. Then the set-valued function $$x\mapsto \nabla_{x} g$$ is differentiable at $x_0$ for almost all $x_0$ in $X$. That is, for almost all $x_0$ there is an $L\in \operatorname{Hom}(\mathbb R^{n},\mathbb R^{m})$ such that for all $\epsilon>0$ there is a $\delta>0$ such that for $\|x-x_0\|<\delta$ we have
\begin{equation}\label{eq:twicediffgradient} \| u-u_0 - \frac{L}{2}(x-x_0)|\| \le \epsilon \|x-x_0\| \text{ for all }u\in \nabla_{x}g \text{ and } u_0\in \nabla_{x_0}g.
\end{equation}
Moreover for almost all $x_0$ the function $g$ is twice differentiable at $x_0$ and $\operatorname{Hess}_{x_0}(g)=L$. That is, for any $\epsilon>0$ there is a $\delta>0$ such that
\begin{equation}\label{eq:twicediffhessian} |g(x) - g(x_0) - \nabla g|_{x_0}.(x-x_0) - \frac{1}{2}(x-x_0)^t \operatorname{Hess}_{x_0}(g) (x-x_0)\rangle | \le \epsilon \|x-x_0\|^2\end{equation}
for all $\|x-x_0\|<\delta.$
\end{theorem}
\begin{proof}
This originates in \cite{Alexandrov} and for an exposition the reader is referred to \cite[Theorems 6.1,7.1]{Howard}. (We remark that the latter cited work requires the function to be convex and defined on all of $\mathbb R^n$; but the statement we want is local, and being locally convex, $g$ is also locally Lipschitz \cite{WayneState}, and so using \cite[Theorem 4.1]{MinYan} we know that $X$ is covered by small open sets $U$ such that $g|_U$ extends to a convex function on $\mathbb R^n$ so the cited work applies.)
\end{proof}
\subsection{Proof of Theorem \ref{thm:argmin1}}\label{sec:proofargmin1}
\begin{lemma}[Continuity of argmin]\label{lem:continuityargmin}
Let $\Omega\subset X\times \mathbb R$ be convex and such that $\Omega_x$ is connected for each $x\in X$. Let $f:\Omega\to \mathbb R$ be continuous, and suppose that for each $x\in X$ the function $y\mapsto f(x,y)$ is strongly convex and attains its minimum at some point. Then $\gamma(x) = \operatorname{argmin}_f(x)$ is single valued and continuous.
\end{lemma}
\begin{proof}
For fixed $x$ the hypothesis imply that $y\mapsto f(x,y)$ is a strongly convex function on the connected set $\Omega_x$ that attains its minimum, and thus this minimum $\gamma(x)$ must be a unique. We first claim that $\gamma$ is locally bounded. Fix $x_0\in X$ and let $a:= \gamma(x_0)$. Then by strong convexity there is an $\epsilon>0$ and $c>0$ such that $f(x_0,y)>a+\epsilon$ if $\|y-\gamma(x_0)\|\ge c$. By continuity we may take $\delta>0$ small so if $\|x-x_0\|<\delta$ and $\|y-\gamma(x_0)\|=c$ then $f(x,y)>a+\epsilon$ and, and furthermore that $f(x,\gamma(x_0))<a+\epsilon$. But by strict convexity of $y\mapsto f(x,y)$ this implies $\gamma(x) \in [\gamma(x_0)-c,\gamma(x_0)+c]$ for all $\|x-x_0\|<\delta$, and thus $\gamma$ is locally bounded.
Now suppose $(x_n)$ is a sequence in $X$ converging to $x$ as $n\to \infty$. By the above we may assume $S:=\{\gamma(x_n)\}$ is bounded. Let $b$ be a cluster point of $S$, so there is a subsequence $x_{n_r}$ with $\gamma(x_{n_r})\to b$ as $r\to \infty$. By continuity of $f$ for any $y\in \mathbb R^m$,
$$ f(x,b)= \lim_{r\to \infty} f(x_{n_r},\gamma(x_{n_r})) \le \lim_{r\to \infty} f(x_{n_r}, y) = f(x,y).$$
Hence $b=\gamma(x)$. As this holds for all cluster points of $S$ we deduce $\gamma(x_n)\to \gamma(x)$ as $n\to \infty$, proving continuity of $\gamma$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:argmin1}]
We first claim that there is no loss in generality in assuming that $\Omega= \mathbb R^{n+m}$. To see this, suppose $f:\Omega\to \mathbb R$ has properties \eqref{eq:condagmin1} and \eqref{eq:condagmin2}. Then $\gamma=\operatorname{argmin}_f$ is single-valued and continuous (Lemma \ref{lem:continuityargmin}). So given $x_0\in \pi(\Omega)$ there are small balls $x_0\in U\subset \pi(\Omega)$ and $\gamma(x_0)\in V\subset \mathbb R^m$ so that $U\times V\subset \Omega$ and $\gamma(U)\subset V$. Moreover as $f$ is semiconvex, by shrinking $U,V$ we may assume that $f|_{U\times V}$ is Lipschitz (all convex functions are Lipschitz, see e.g.\ \cite{WayneState}). Let $\tilde{f}(x,y) : = f(x,y) + \frac{\kappa}{2} \|x\|^2 - \frac{\sigma}{2} \|y\|^2$ which we are assuming is convex on $\Omega$. Then \cite[Theorem 4.1]{MinYan} we know $\tilde{f}|_{U\times V}$ extends to a convex function $\tilde{h}$ on all of $\mathbb R^{n+m}$. Now let $$h(x,y): = \tilde{h}(x,y) - \frac{\kappa}{2}\|x\|^2 + \frac{\sigma}{2} \|y\|^2.$$
For fixed $x$ the convex function $y\mapsto h(x,y)$ agrees with the function $y\mapsto f(x,y)$ when $y\in V$. Since $V$ contains $\gamma(x)=\operatorname{argmin}_f(x)$, this implies $\operatorname{argmin}_h(x) = \operatorname{argmin}_f(x)=\gamma(x)$. Hence $h$ satisfies the hypothesis of the Theorem with $\Omega = \mathbb R^{n+m}$, and so $\gamma|_U$ has the properties in the conclusion of the theorem (which are all local), which proves the claim.
So from now on assume $f:\mathbb R^{n+m} \to \mathbb R$ satisfies \eqref{eq:condagmin1} and \eqref{eq:condagmin2}. Consider first the case $\kappa=0$, so $(x,y)\mapsto f(x,y) -\frac{\sigma}{2} \|y\|^2$ is convex. Then in particular $f$ is convex, and so $g(x) = \inf_y f(x,y)$ is also convex. Moreover for fixed $x$ the function $y\mapsto f(x,y)$ is strictly convex, and so $\operatorname{argmin}_f$ (which is assumed to be non-empty) must be single valued. Consider the functional $J$ from \eqref{def:J} so by Proposition \ref{prop:Jargmin}
\begin{equation}J(x,u,\gamma(x)) =0 \text{ for all } x \text{ and } u\in \nabla_x g.\label{eq:functionalJrepeat}\end{equation}
Fix $x_0\in \mathbb R^n$ and $u_0\in \nabla_{x_0}g$ so $J(x_0,u_0,\gamma(x_0)) =0$. The properties of $J$ proved in Lemma \ref{lem:propertiesJ} mean we can apply the Inverse-function Theorem for Lipschitz maps (for convenience of the reader we give a proof of this in Appendix \ref{sec:appendiximplicit}, and apply it here with $r$ replaced with $2n$ and $s$ replaced with $m$). This yields a Lipschitz function $\phi:V\to \mathbb R^{m}$ defined on a neighbourhood $V$ of $(x_0,u_0)$ such that
$$ J(x,u,y)=0 \Leftrightarrow y = \phi(x,u).$$
This combined with \eqref{eq:functionalJrepeat} gives
$$\gamma(x) = \phi(x,u) \text{ for all } (x,u)\in V \text{ with } u\in \nabla_x g.$$
We next prove $\gamma$ is calm almost everywhere. As $g$ is convex we have by Alexandrov's Theorem \eqref{eq:twicediffgradient} that for almost all $x_0$ there are $\delta_1>0$ and linear $L:\mathbb R^{n}\to \mathbb R^{n}$ such that for $\|x-x_0\|<\delta_1$
\begin{equation}\| u - u_{0} \| \le (1 + \|L\|) \|x-x_0\| \text{ for all } u\in \nabla_{x}g \text{ and } u_0\in \nabla_{x_0} g.\label{eq:argmax2eq1}
\end{equation}
Pick $u_0\in \nabla_{x_0} g$, and let $\phi:V\to \mathbb R$ be the Lipschitz function constructed above. For concreteness say that $V$ contains the set $\|x-x_0\|<\delta_2$ and $\|u-u_0\|<\delta_2$ and that $\phi$ has Lipschitz constant $C'$ there. Thus
$$ \gamma(x) = \phi(x,u) \text{ for } \|x-x_0\|<\delta_2, \|u-u_0\|<\delta_2 \text{ and } u\in \nabla_x g.$$
Set $$\delta :=\min\{\delta_1, \frac{\delta_2}{1+\|L\|}\}$$ and suppose $\|x-x_0\|<\delta$. Picking any $u\in \nabla_x g$, by \eqref{eq:argmax2eq1} $\|u-u_0\|<\delta_2$ and so
$$ \|\gamma(x) -\gamma(x_0)\| = \|\phi(x,u) - \phi(x_0,u_0)\| \le C' (\|x-x_0\| + \|u-u_0\|) \le C'(2 + \|L\|) \|x-x_0\|.$$
Thus $\gamma$ is calm $x_0$.
The case of general $\kappa$ is easily reduced to the case $\kappa=0$. For suppose $f(x,y) + \frac{\kappa}{2} \|x\|^2 -\frac{\sigma}{2}|y|^2$ is convex and $\operatorname{argmin}_f$ is single-valued. Set
$$\tilde{f}(x,y) = f(x,y) + \frac{\kappa}{2} \|x\|^2$$
Then $\tilde{f}(x,y) - \frac{\sigma}{2} \|y\|^2$ is convex, and by \eqref{eq:argminshift}
$$\operatorname{argmin}_{\tilde{f}}(x) = \operatorname{argmin}_f (x).$$
Thus $\operatorname{argmin}_{\tilde{f}}$ is also single-valued, so by the above the Theorem can be applied to $\tilde{f}$. Let
$$ \gamma(x) := \operatorname{argmin}_{\tilde{f}}(x) = \operatorname{argmin}_f (x)$$
Setting $\tilde{g}(x):=\inf_y \tilde{f}(x,y)$, given $x_0$ and $u_0\in \nabla_{x_0} \tilde{g}$ we know that there is a locally Lipschitz function $\tilde{\phi}:\tilde{V}\to \mathbb R$ defined on a neighbourhood $\tilde{V}$ of $(x_0,u_0)$ such that
$$ \gamma(x) = \tilde{\phi}(x,u) \text{ for } (x,u)\in \tilde{V} \text{ with } u\in \nabla_x \tilde{g}.$$
Set $\phi(x,u) = \tilde{\phi}(x,u+\kappa x)$ which is locally Lipschitz around $(x_0,u_0 + \kappa x_0)$. And if $u\in \nabla^{\kappa}g$ then $u-\kappa x_0 \in \nabla_x\tilde{g}$ so $\gamma(x) = \tilde{\phi}(x,u-\kappa x_0) = \phi(x,u)$. Thus the conclusion of the Theorem also holds for $f$ and we are done.
\end{proof}
\section{F-subharmonic functions}\label{sec:Fsubharmonics}
\subsection{Basic definitions}\label{sec:basics} We summarise some basic properties of F-subharmonic functions from the work of Harvey-Lawson. We refer the reader to \cite{MP1} for a more detailed summary, or the original papers \cite{HL_Dirichletduality,HL_Dirichletdualitymanifolds}. Let $X\subset \mathbb R^{n}$ be open and
$$J^2(X):=X\times \mathbb R\times \mathbb R^n \times \operatorname{Sym}^2_{n} = X\times J^2_n$$
be the jet-bundle over $X$. For $F\subset J^2(X)$ and $x\in X$ we write
$$ F_x= \{ (r,p,A)\in J^2_n : (x,r,p,A)\in F\}.$$
\begin{definition}[Primitive Subequations]
We say that $F\subset J^2(X)$ is a \emph{primitive subequation} if
\begin{enumerate}
\item (Closedness) $F$ is closed.
\item (Positivity)
\begin{equation}\label{eq:positivity}
(r,p,A)\in F_x \text{ and } P\in Pos_{n} \Rightarrow (r,p,A+P)\in F_x.\end{equation}
\end{enumerate}
We say that $ F\subset J^2(X)$ has the \emph{Negativity Property} if
\begin{enumerate}\setcounter{enumi}{2}
\item (Negativity) \begin{equation}\label{eq:negativity}
(r,p,A)\in F_x \text{ and } r'\le r \Rightarrow (r',p,A)\in F_x.\end{equation}
\end{enumerate}
\end{definition}
\begin{definition}[Upper contact points, Upper contact jets]
Let $$f: X\to \mathbb R\cup\{-\infty\}.$$ We say that $x\in X$ is an \emph{upper contact point} of $f$ if $f(x)\neq -\infty$ and there exists $(p,A)\in \mathbb R^n\times \operatorname{Sym}^2_n$ such that
$$f(y)\le f(x) + p.(y-x) + \frac{1}{2} (y-x)^tA(y-x) \text{ for all } y \text{ sufficiently near } x.$$
When this holds we refer to both $(f(x),p,A)$ and $(p,A)$ as an \emph{upper contact jet} of $f$ at $x$.
\end{definition}
\begin{definition}[F-subharmonic function]
Suppose $F\subset J^2(X)$. We say that an upper-semicontinuous function $f:X\to \mathbb R\cup \{-\infty\}$ is \emph{F-subharmonic} if $$ (f(x),p,A)\in F_x \text{ for all upper contact jets } (p,A) \text{ of }f \text{ at } x.$$
We let $F(X)$ denote the set of $F$-subharmonic functions on $X$.
\end{definition}
Clearly being $F$-subharmonic is a local condition on $X$.
\begin{proposition}\label{prop:basicproperties}
Let $F\subset J^2(X)$ be closed. Then
\begin{enumerate}
\item (Maximum Property) If $f,g\in F(X)$ then $\max\{f,g\}\in F(X)$.
\item (Decreasing Sequences) If $f_j$ is decreasing sequence of functions in $F(X)$ (so $f_{j+1}\le f_j$ over $X$) then $f:=\lim_j f_j$ is in $F(X)$.
\item (Uniform limits) If $f_j$ is a sequence of functions on $F(X)$ that converge locally uniformly to $f$ then $f\in F(X)$.
\item (Families locally bounded above) Suppose $\mathcal F\subset F(X)$ is a family of $F$-subharmonic functions locally uniformally bounded from above. Then the upper-semicontinuous regularisation of the supremum
$$ f:= {\sup}^*_{f\in \mathcal F} f$$
is in $F(X)$.
\item If $F$ is constant coefficient and $f$ is $F$-subharmonic on $X$ and $x_0\in \mathbb R^{n}$ is fixed, then the function $x\mapsto f(x-x_0)$ is $F$-subharmonic on $X-x_0$.
\end{enumerate}
\end{proposition}
\begin{proof}
See \cite[Theorem 2.6]{HL_Dirichletdualitymanifolds} for (1-4). Item (5) is immediate.
\end{proof}
\begin{definition}
Let $F\subset J^2(X)$.
\begin{enumerate}
\item We say $F$ is \emph{constant coefficient} if $F_x$ is independent of $x$, i.e.\
$$ (x,r,p,A) \in F_x \Leftrightarrow (x',r,p,A)\in F_{x'} \text{ for all }x,x',r,p,A.$$
\item We say $F$ \emph{depends only on the Hessian part} if each $F_x$ is independent of $(r,p)$, i.e.
$$ (r,p,A) \in F_x \Leftrightarrow (r',p',A)\in F_{x} \text{ for all }x,r,r',p,p',A.$$
\end{enumerate}
\end{definition}
An important example is $$\mathcal P:= \{ (x,r,p,A)\in J^2(X): A \text{ is semipositive}\}$$ which is a constant-coefficient primitive subequation that depends only on the Hessian part. Then $\mathcal P$-subharmonic functions are precisely those that are locally convex \cite[Example 14.2]{HL_Dirichletdualitymanifolds}.
\begin{lemma}[Sums of $F$-subharmonic and convex functions]\label{lem:sumsubharmoniconvex}
Suppose $F\subset J^2(X)$ is a constant coefficient primitive subequation that depends only on the Hessian part. If $f$ is $F$-subharmonic on $X$ and $g$ is a convex quadratic function on $X$, then $f+g$ is $F$-subharmonic.
\end{lemma}
\begin{proof}
The hypothesis is that $g(x) = a + b.x + \frac{1}{2} x^tC x$ for some $a,b\in \mathbb R^n$ and some semipositive symmetric matrix $C$. One can check that if $(p,A)$ is an upper-contact point of $f+g$ at $x$ then $(x^tC +p-b, A-C)$ is an upper-contact jet for $f$ at $x$. As $f$ is $F$-subharmonic this implies $(f(x), x^tC +p-b, A-C)\in F$. Since $F$ depends only on the Hessian part, and satisfies the Positivity property, this in turn implies $(f(x)+g(x),p,A)\in F$ proving that $f+g$ is $F$-subharmonic as required.
\end{proof}
\subsection{Product Subequations}\label{sec:productsubequations}
For $\Gamma\in \operatorname{Hom}(\mathbb R^{n},\mathbb R^{m})=M_{m\times n}(\mathbb R)$ consider
\begin{align}i_\Gamma:\mathbb R^{n} &\to \mathbb R^{n+m} \quad i_\Gamma(x) = (x,\Gamma x)\\
j : \mathbb R^{m} &\to \mathbb R^{n + m} \quad j(y) = (0,y).\end{align}
which induce natural pullback maps
\begin{equation}
i_\Gamma^*: J^2_{n+m}\to J^2_{n} \text{ and }
j^* : J^2_{n+m}\to J^2_{m}.\end{equation}
We can write these explicitly. Suppose
$$ p := \left(\begin{array}{c} p_1 \\ p_2 \end{array} \right)\in \mathbb R^{n+ m} \text{ and } A := \left(\begin{array}{cc} B & C \\ C^t & D \end{array}\right)\in \operatorname{Sym}^2_{n+m}$$
where the latter is in block form, so $B\in \operatorname{Sym}^2_{n}$ and $D\in \operatorname{Sym}^2_{m}$. Then
\begin{align}\label{eq:defofistar}
i_\Gamma^*(r,p,A) &= \left(r, p_1 + \Gamma^t p_2, B + C\Gamma + \Gamma^t C^t + \Gamma^tD\Gamma\right)\\
j^*(r,p,A) &= (r,p_2,D).
\end{align}
\begin{comment}
\begin{definition}[Slices]
Given ${x_0}\in X$ we call
$$ i_{x_0} : \{ y\in Y : (x_0,y)\in X\times Y \} \to X\times Y \quad i_{x_0}(y) = (x,y)$$
a \emph{vertical slice} of $X\times Y$.
Given $y_0\in Y$ and $\Gamma\in Hom(\mathbb R^{n},\mathbb R^{m})$ we call
$$ i_{y_0,\Gamma} : \{ x : (x,y_0 + \Gammax)\in X\times Y\} \to X\times Y \quad i_{y_0,\Gamma}(x) = (x,y_0 + \Gammax)$$
a \emph{non-vertical slice} of $X\times Y$.
\end{definition}
Vertical and non-vertical slices induce restriction maps
$$ i_{y_0,\Gamma}^* : i_{y_0,U}^*J^2(X\times Y) \to J^2(X).$$
$$ i_{x_0}^*: i_{x_0}^*J^2(X\times Y) \to J^2(Y)$$
\end{comment}
\begin{definition}[Products]
Let $X\subset\mathbb R^n$ and $Y\subset \mathbb R^m$ be open, and $F\subset J^2(X)$ and $G\subset J^2(Y)$. Define $$F\#G \subset J^2(X\times Y)$$ by
$$(F\#G)_{(x,y)} = \left \{ \alpha\in J^2_{n+m} : \begin{array}{l} i_{\Gamma}^*\alpha \in F_{x} \text{ and } j^*\alpha \in G_y \\ \text{for all } \Gamma\in \operatorname{Hom}(\mathbb R^{n},\mathbb R^{m})\end{array}\right\}.$$
\end{definition}
\begin{lemma}
\begin{enumerate}
\item If $F$ and $G$ are primitive subequations then so is $F\#G$. Moreover if $F$ and $G$ both have the Negativity Property then so does $F\#G$.
\item Let $F$ be a constant-coefficient primitive subequation on $X$. Suppose and $f$ is $F\#\mathcal P$-subharmonic on some open $\Omega\subset X\times Y$. The for each $x\in X$ the function $y\mapsto f(x,y)$ is locally convex.
\end{enumerate}
\end{lemma}
\begin{proof}
The reader will easily prove these straight from the definition, or otherwise find the proofs in \cite{MP1}.
\end{proof}
\subsection{The almost everywhere theorem}
We will rely on a very useful theorem of Harvey-Lawson that characterizes $F$-subharmonic semiconvex functions in terms of second order jets almost everywhere.
\begin{definition}[Twice differentiability at a point]
We say that a function $f:X\to \mathbb R$ is \emph{twice differentiable} at $x_0\in X$ if there exists a $p\in \mathbb R^n$ and an $L\in \operatorname{Sym}_n^2$ such that for all $\epsilon>0$ there is a $\delta>0$ such that for $\|x-x_0\|<\delta$ we get
\begin{equation}\label{eq:twicediff} |f(x) - f(x_0) - p.(x-x_0) - \frac{1}{2} (x-x_0)^tL (x-x_0) | \le \epsilon \|x-x_0\|^2.
\end{equation}
\end{definition}
When $f$ is twice differentiable at $x_0$ then the $p,L$ in \eqref{eq:twicediff} are unique, and moreover in this case $f$ is differentiable at $x_0$ and
$$p = \nabla f|_{x_0}= \left( \begin{array}{c} \frac{\partial f}{\partial x_1} \\ \frac{\partial f}{\partial x_2} \\ \vdots \\ \frac{\partial f}{\partial x_n} \end{array}\right)|_{x_0}\in \mathbb R^n.$$
When $f$ is twice differentiable at $x_0$ we shall refer to $L$ as the \emph{Hessian} of $f$ at $x_0$ and denote it by $\operatorname{Hess}(f)|_{x_0}$. Of course when $f$ is $\mathcal C^{2}$ in a neighbourhood of $x_0$ then $\operatorname{Hess}_{x}(f)$ is the matrix with entries
$$(\operatorname{Hess}(f)_{x_0} )_{ij}: = \frac{\partial^2 f}{\partial x_i\partial x_j}|_{x_0}.$$
\begin{definition}[Second order jet]
Suppose that $f: X\to \mathbb R$ is twice differentiable at $x_0$. We denote the \emph{second order jet} of $f$ at $x_0$ by
\begin{equation}J^2_{x_0}(f):= (f(x_0), \nabla f|_{x_0}, \operatorname{Hess}(f)|_{x_0}) \in J^2_{n} = \mathbb R\times \mathbb R^n\times \operatorname{Sym}_n^2.\label{eq:secondorderjet}
\end{equation}
\end{definition}
We have seen in Alexandrov's Theorem (Theorem \ref{thm:alexandrov:repeat}) that if $f$ is locally semiconvex then $J^2_x(f)$ exists for almost all $x$.
\begin{theorem}[The Almost Everywhere Theorem]\label{thm:ae}
Assume that $F\subset J^2(X)$ is a primitive subequation and let $f:X\to \mathbb R$ be locally semiconvex. Then
$$ f\in F(X) \Leftrightarrow J^2_{x}(f)\in F_x \text{ for almost all } x\in X.$$
\end{theorem}
\begin{proof}
See \cite[Theorem 4.1]{HL_AE}.
\end{proof}
\section{Partial sup-convolutions}\label{sec:partialsup}
Fix open $U\subset \mathbb R^n$ and $V\subset \mathbb R^m$, and suppose $f:U\times V\to \mathbb R$ is upper-semicontinuous and bounded.
\begin{definition}[Partial-Sup-Convolutions]\label{def:partialsupconvolution} For $\epsilon>0$ the \emph{partial sup-convolution} of $f$ is
\begin{equation}\label{eq:defpartialsupconvolution}
f^{\epsilon,p}(x,y): = \sup_{z\in U} \{ f(z,y) - \frac{1}{2\epsilon} \| z-x\|^2\} \text{ for } (x,y)\in U\times V.
\end{equation}
\end{definition}
For $\delta>0$ let
$$U(\delta) = \{ x\in \mathbb R^n : B_{\delta}(x)\subset U\}.$$
\begin{lemma}[Basic Properties of Partial-Sup-Convolutions]\label{lem:partialsupconvolutionbasicproperties}\
\begin{enumerate}[(i)]
\item (Strong Semiconvexity) Assume that for each fixed $x$ the function $y\mapsto f(x,y)$ is convex. Then
$$(x,y)\mapsto f^{\epsilon,p}(x,y) + \frac{1}{2\epsilon} \|x\|^2$$ is convex.
\item (Monotonicity) For $0<\epsilon'\le \epsilon$ we have \begin{equation}\label{eq:orderf}f \le f^{\epsilon',p} \le f^{\epsilon,p}.\end{equation}
\item Let $\delta: =2 (\epsilon \|f\|_{\infty})^{1/2}$. Then
$$f^{\epsilon,p}(x,y) = \sup_{\|\tau\|<\delta} \{ f(x+\tau,y) - \frac{1}{2\epsilon} \|\tau\|^2\} \text{ for } (x,y)\in U(\delta)\times V.$$
\item (Pointwise convergence)
$$\lim_{\epsilon\to 0^+} f^{\epsilon,p}(x,y) = f(x,y) \text{ for } (x,y)\in U\times V.$$
\item (Magic-Property)
Suppose that $F$ is a constant-coefficient primitive subequation on $U$ that has the Negativity Property and $f$ is $F\#\mathcal P$-subharmonic. Then $f^{\epsilon,p}$ is $F\#\mathcal P$-subharmonic on $U(\delta)\times V$.
\end{enumerate}
\end{lemma}
\begin{proof}
\begin{align*}f^{\epsilon,p}(x,y) + \frac{1}{2\epsilon} \|x\|^2& = \sup_{z\in U} \{ f(z,y) - \frac{1}{2\epsilon} \|z-x\|^2 + \frac{1}{2\epsilon} \|x\|^2\} \\
&=\sup_{z\in U} \{ f(z,y) + \frac{1}{\epsilon} x.z - \frac{1}{2\epsilon} \|z\|^2\}.
\end{align*}
Now for fixed $z$ the function $y\mapsto f(z,y)$ is assumed to be convex in $y$, and so the function $(x,y) \mapsto f(z,y)$ is convex in $(x,y)$. Thus, again for $z$ fixed, $(x,y) \mapsto f(z,y) + \frac{1}{\epsilon} x.z + \frac{1}{2\epsilon} \|z\|^2$ is convex in $(x,y)$, and hence so is $f^{\epsilon,p}(x,y) + \frac{1}{2\epsilon} \|x\|^2$ proving (i).
Item (ii) is immediate. For (iii) we claim that
\begin{equation}\label{eq:partialsupconvolutionslocalises}f^{\epsilon,p}(x,y) = \sup_{z\in U : \|z-x\|<\delta}\{ f(z,y) - \frac{1}{2\epsilon} \|z-x\|^2\} \text{ for } (x,y)\in U\times V.\end{equation}
To see this let $M: = \|f\|_{\infty}$. Then for $z\in U$ with $\|z-x\|\ge \delta = \sqrt{4\epsilon M}$,
$$f(z,y) - \frac{1}{2\epsilon} \|z-x\|^2 \le M - \frac{1}{2\epsilon} \delta^2 = -M \le f(x,y)\le f^{\epsilon,p}(x,y)$$
which proves \eqref{eq:partialsupconvolutionslocalises}. Then (iii) follows upon making the change of variables $\tau: = z-x$. For the pointwise convergence fix $(x,y)\in U\times V$ and let $a>f(x,y)$. Then $f<a$ on some open neighbourhood of $(x,y)$. Let $\epsilon$ be small enough so that $B_\delta(x)$ is contained in this neighbourhood. Then \eqref{eq:partialsupconvolutionslocalises} implies $f^{\epsilon,p}(x,y) \le a$, proving (iv). For the final statement, since $F$ is constant coefficient for any fixed $\tau$ the function $f(x+\tau,y)$ is $F\#\mathcal P$-subharmonic (where defined), and hence (iii) shows $f^{\epsilon,p}$ as a supremum of $F\#\mathcal P$-subharmonic functions. Now being $F\#\mathcal P$-subharmonic implies that $y\mapsto f(x,y)$ is convex, and so by (i) $f^{\epsilon,p}$ is certainly continuous and hence equal to its upper semicontinuous regularisation. Thus $f^{\epsilon,p}$ is $F\#\mathcal P$-subharmonic on $U(\delta)\times V$ as claimed in (v).
\end{proof}
The next lemma reveals a surprising property of the above construction, namely that the partial sup-convolution of a semiconcave function is fibrewise semiconcave.
\begin{lemma}\label{lem:preservationofconcavity}
Suppose that $f$ is $\kappa$-semiconcave for some $\kappa>0$. Then for $\epsilon<\kappa^{-1}$ and fixed $x\in U$ the function
$$ y\mapsto f^{\epsilon,p}(x,y)-\frac{\kappa}{2} \|y\|^2$$
is concave.
\end{lemma}
\begin{proof}
Let $x$ be fixed. Then
\begin{align*}
f^{\epsilon,p}(x,y) - \frac{\kappa}{2} \|y\|^2 &=
\sup_{z\in U} \{f(z,y) - \frac{\kappa}{2} \|y\|^2 - \frac{1}{2\epsilon} \|z-x\|^2\} \\
&= \sup_{z\in U} \{f(z,y) - \frac{\kappa}{2} \|z\|^2 - \frac{\kappa}{2} \|y\|^2 +\frac{\kappa-\epsilon^{-1}}{2} \|z\|^2 + \frac{1}{\epsilon} z.x - \frac{1}{2\epsilon} \|x\|^2\}.
\end{align*}
Observe that (since $x$ is fixed and $\kappa-\epsilon^{-1}<0$) the function $(z,y)\mapsto \frac{\kappa-\epsilon^{-1}}{2} \|z\|^2 + \frac{1}{\epsilon} z.x - \frac{1}{2\epsilon} \|x\|^2$ is convex as a function of $(z,y)$. Furthermore by hypothesis $f(z,y) - \frac{\kappa}{2} \|z\|^2 - \frac{\kappa}{2} \|y\|^2$ is concave in $(z,y)$. Hence $y\mapsto f^{\epsilon,p}(x,y) - \frac{\kappa}{2} \|y\|^2$ is a supremum of functions concave in two variables, and thus is concave.
\end{proof}
\section{$F$-subharmonicity of marginal functions}\label{sec:fsubharmonicityofmarginals}
Let $\Omega\subset \mathbb R^{n+m}$ be open, convex and such that $\Omega_x$ is connected for all $x$.
\begin{proposition}\label{prop:semiI}
Let $F\subset J^2(\mathbb R^{n})$ be a primitive subequation. Let $f:\Omega\to \mathbb R$ be $F\#\mathcal P$-subharmonic, and suppose that for some $\sigma,\kappa_1,\kappa_2>0$ the function
\begin{equation}\label{eq:semiIH1} f(x,y) + \frac{\kappa_1}{2} \|x\|^2-\frac{\sigma}{2} \|y\|^2 \text{ is convex}
\text{ and }\end{equation}
and for each fixed $x$ the function
\begin{equation} \label{eq:semiIH2} y\mapsto f(x,y) - \frac{\kappa_2}{2} \|y\|^2 \text{ is concave}\end{equation}
and that $\gamma(x)=\operatorname{argmin}_f(x)$ is single valued. Then
$$ g(x): = \inf_{y\in \Omega_x} f(x,y)$$
is $F$-subharmonic.
\end{proposition}
\begin{proof}
By hypothesis $$g(x) = f(x,\gamma(x)).$$
Now $g$ is $\kappa$-semiconvex (Lemma \ref{lem:basicconvexityII}) so by Alexandrov's Theorem (Theorem \ref{thm:alexandrov:repeat}) $g$ is twice differentiable almost everywhere. Furthermore \eqref{eq:semiIH1} allows us to invoke our results on the argmin function, so by Corollary \ref{cor:wisdiffae} $\gamma$ is differentiable almost everywhere. Let $x_0$ be a point where $g$ is twice differentiable and $\gamma$ is differentiable, and we will show
\begin{equation}J^2_{x_0} g = (g(x_0), \nabla g|_{x_0}, \operatorname{Hess}_{x_0}(g)) \in F_{x_0}\label{eq:J2g}.\end{equation}
By the Almost Everywhere Theorem (Theorem \ref{thm:ae}) this implies that $g$ is $F$-subharmonic.
Actually we will show that for any $\epsilon>0$ it holds that
\begin{equation}(g(x_0), \nabla g|_{x_0}, \operatorname{Hess}_{x_0}(g)+\epsilon \operatorname{Id}_{n}) \in F_{x_0}.\label{eq:j2gepsilon}\end{equation}
Letting $\epsilon\to 0$ and using that $F_{x_0}$ is closed yields \eqref{eq:J2g}.
To this end set $y_0: = \gamma(x_0)$ and
$$\Gamma:= D\gamma|_{x_0}\in \operatorname{Hom}(\mathbb R^{n},\mathbb R^{m})$$
and $d(x,y)$ be the vertical distance between $(x,y)\in \mathbb R^{n}\times \mathbb R^{m}$ and the tangent to the graph of $\gamma$ at $(x_0,y_0)$, so
$$ d(x,y) := \|y - y_0 - \Gamma (x-x_0)\| \text{ for } (x,y)\in \mathbb R^{n}\times \mathbb R^{m}.$$
Consider the quadratic
$$ q(x,y) = g(x_0) + \nabla g|_{x_0}.(x-x_0) +\frac{1}{2} (x-x_0)^t\operatorname{Hess}_{x_0}(g) (x-x_0)+ \frac{\epsilon}{2}\|x-x_0\|^2 + \kappa_2 d(x,y)^2$$
for
$(x,y)\in \mathbb R^{n}\times \mathbb R^{m}$. By construction $$q(x_0,y_0) = g(x_0) = f(x_0,\gamma(x_0)) = f(x_0,y_0),$$ and in Lemma \ref{lem:semiI} below we show that $q\ge f$ sufficiently near $(x_0,y_0)$.
Hence $(x_0,y_0)$ is an upper contact point for $f$ and
\begin{align}J^2_{(x_0,y_0)}(q) &= \left( q(x_0,y_0), \nabla q|_{(x_0,y_0}, \operatorname{Hess}_{(x_0,y_0}(q)\right)\\
&=\left(f(x_0,y_0),\left(\begin{array}{c} \nabla g|_{x_0} \\0 \end{array}\right),\left(\begin{array}{cc} \operatorname{Hess}_{x_0}(g) +\epsilon \operatorname{Id}_{n} + 2\kappa_2\Gamma^t\Gamma& -2\kappa_2\Gamma^t \\ -2\kappa_2\Gamma & 2\kappa_2\operatorname{Id}_{m}\end{array}\right)\right)
\end{align}
is an upper-contact jet of $f$ at $(x_0,y_0)$. So as $f$ is $F\#\mathcal P$-subharmonic we have
$$ J^2_{(x_0,y_0)}(q) \in (F\#\mathcal P)_{(x_0,y_0)}.$$
And from the definition of $i_{\Gamma}^*$,
$$ i_{\Gamma}^* (J^2_{(x_0,y_0)}(q))=\operatorname{Hess}_{x_0}(g) +\epsilon \operatorname{Id}_{n}$$
which must lie in $F_{x_0}$. This gives \eqref{eq:j2gepsilon} and completes the proof.
\end{proof}
\begin{lemma}\label{lem:semiI}
With the notation as in the proof of Theorem \ref{prop:semiI} we have
\begin{equation}\label{eq:qlarger} q(x,y) \ge f(x,y) \text{ for } (x,y) \text{ sufficiently near } (x_0,y_0).\end{equation}
\end{lemma}
\begin{proof}
Fix $\epsilon'>0$ small enough so $\epsilon' + \kappa_2\epsilon'^2 <\epsilon/2$. That $\Gamma=D\gamma|_{x_0}$ means there is a $\delta>0$ such that for all $\|x-x_0\|<\delta$
$$ \|\gamma(x) - y_0 - \Gamma(x-x_0)\| \le \epsilon' \|x-x_0\|.$$
Shrinking $\delta$ is necessary, the definition of $g$ being twice differentiable at $x_0$ means \eqref{eq:twicediff} that for $\|x-x_0\|<\delta$ we also have
$$ | g(x) - g(x_0) - \nabla g|_{x_0}.(x-x_0) -\frac{1}{2} (x-x_0)^t \operatorname{Hess}_{x_0}g(x-x_0) | \le \epsilon' \|x-x_0\|^2.$$
Consider now a point $(x,y)$ with $\|x-x_0\|<\delta$ and $\|y-y_0\|<\delta$. Then
\begin{align}
\| y-\gamma(x)\| &\le \|y-y_0 - \Gamma (x-x_0) \| + \|y_0+\Gamma (x-x_0)-\gamma (x)\|\\
&\le d(x,y) + \epsilon' \|x-x_0\|. \end{align}
So
$$\| y-\gamma (x)\|^2 \le 2\epsilon'^2 \|x-x_0\|^2 + 2d(x,y)^2.$$
Now we use (in an essential way) hypothesis \eqref{eq:semiIH2}. Since $\gamma (x)$ is the minimum of the function $y'\mapsto f(x,y')$ \eqref{eq:semiIH2} implies
$$ f(x,y) \le f(x,\gamma(x)) + \frac{\kappa_2}{2} \|y-\gamma(x)\|^2.$$
Thus
\begin{align*}
f(x,y) &\le g(x) + \kappa_2(\epsilon'^2 \|x-x_0\|^2 + d(x,y)^2)\\
&\le g(x_0) + \nabla g|_{x_0} (x-x_0) + \frac{1}{2} (x-x_0)^t\operatorname{Hess}_{x_0}g(x-x_0) \\&\quad + (\kappa_2\epsilon'^2 + \epsilon')\|x-x_0\|^2 + \kappa_2 d(x,y)^2\\
&\le q(x,y)
\end{align*}
as required.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:introminimum}]
Let $f:X\times \mathbb R\to \mathbb R$ be locally semiconcave, bounded from below and $F\#\mathcal P$-subharmonic. We are to show that $g(x):= \inf_y f(x,y)$ is $F$-subharmonic.
We first claim that without loss of generality we may assume in addition that for each $x$ it holds that $\operatorname{argmin}_f(x)$ is non-empty and single valued. To prove this, for $j\ge 1$ let
$$ f_j(x,y) = f(x,y) + \frac{1}{j} \|y\|^2.$$
As $F$ depends only on the Hessian part, $f_j$ is still $F\#\mathcal P$-subharmonic, and is still bounded from below and semiconcave. Moreover since $f$ is bounded from below, for each fixed $x$ the function $y\mapsto f(x,y)$ is strictly convex and tends to infinity as $|y|$ tends to infinity, implying that it has a unique global minimum. By assumption the theorem applies to $f_j$ meaning that letting $g_j(x):= \inf_y f_j(x,y)$ the function $g_j$ is $F$-subharmonic. But $g_j\searrow g$ pointwise as $j\to \infty$, and thus $g$ will be $F$-subharmonic as well, proving the claim.\medskip
So from now on assume that $\gamma(x)=\operatorname{argmin}_f(x)$ is single valued. Fix $x_0\in \mathbb R^n$. As $\gamma$ is continuous, there exist small balls $x_0\in U\subset X$ and $\gamma(x_0)\in V\subset \mathbb R^m$ such that $\gamma(U)\subset V$ and $f$ is semiconcave on $U\times V$. For $\epsilon>0$ consider the function
$$f_\epsilon(x,y) := f^{\epsilon,p}(x,y) + \frac{\epsilon}{2} \|y\|^2 = \sup_{z\in U} \{ f(z,y) - \frac{1}{2\epsilon} \|z-x\|^2\} + \frac{\epsilon}{2} \|y\|^2.$$
We claim that for $\epsilon$ sufficiently small the following all hold:
\begin{enumerate}[(i)]
\item $f_{\epsilon}(x,y) + \frac{1}{2\epsilon} \|x\|^2 - \frac{\epsilon}{2} \|y\|^2$ is convex.
\item $f_{\epsilon}$ is $F\#\mathcal P$-subharmonic on $U'\times V$ for some smaller ball $x_0\in U'\subset U$.
\item $f^{\epsilon}\searrow f$ pointwise on $U\times V$ as $\epsilon\to 0^+$.
\item There is a $\kappa_2>0$ such that for each $x\in U$ the function $y\mapsto f_{\epsilon}(x,y) - \frac{\kappa_2}{2} \|y\|^2$ is concave.
\end{enumerate}
Items (i,ii,iii) follow from Lemma \ref{lem:partialsupconvolutionbasicproperties} (we have used here the hypothesis that $F$ depends only on the Hessian part so adding a multiple of $\|y\|^2$ preserves the property of being $F\#\mathcal P$-subharmonic by Lemma \ref{lem:sumsubharmoniconvex}). The statement (iv) follows from Lemma \ref{lem:preservationofconcavity} (observing that the addition of $\frac{\epsilon}{2}\|y\|^2$ to the partial sup-convolution only means we may need to increase the value of $\kappa_2$)
Thus we are in a position to apply Proposition \ref{prop:semiI} to $f_{\epsilon}$ to conclude that if
$$g_{\epsilon}(x): = \inf_{y\in V} f_{\epsilon}(x,y)$$
then $g_{\epsilon}$ is $F$-subharmonic on $U'$. But by (iii) if $x\in U'$ then
$$ g_{\epsilon}(x) \searrow \inf_{y\in V} f(x,y) = f(x,\gamma(x)) = g(x) \text{ as } \epsilon\to 0^+$$
and hence $g$ is also $F$-subharmonic on $U'$. Since $x_0$ was arbitrary we conclude $g$ is $F$-subharmonic on all of $\mathbb R^n$ as required.
\end{proof}
|
\section{Introduction}\label{Intro}
The stochastic heat equation with delta initial measure at the origin is given as
\begin{align}\label{eq:SHEDef}
\frac{\partial }{\partial T}\mathcal{Z}(T,X) &= \frac{1}{2} \frac{\partial^2}{\partial^2 X} \mathcal{Z}(T,X) +\frac{1}{2} \mathcal{Z}(T,X) \mathcal{W}(T,X) \qquad X\in \mathbb{R}, t\in \mathbb{R}_{+}\\
\mathcal{Z}(0,X) & = \delta_{X=0}.
\end{align}
Here, $\mathcal{W}(T,X)$ is space time white noise which is a distribution valued Gaussian random field with the following correlation structure
\begin{align}\label{eq:WNCorr}
\mathbb{E}\Big[\mathcal{W}(T_1,X_1)\mathcal{W}(T_2, X_2)\Big] = \delta_{T_1=T_2} \delta_{X_1= X_2}.
\end{align}
For the solution theory of the SHE, we refer to \cite{Walsh86,Corwin12,Quastel12}. Applications of the SHE range from modeling the density of the particles diffusing under the space-time random environment or random drifts \cite{Mol96, BarraquandCorwin15, CG17} to the contnuous directed random polymers \cite{AKQ1}. The logarithm of the SHE is called the \emph{Cole-Hopf} solution of the \emph{Kardar-Parisi-Zhang} (KPZ) equation which is a prototype for random growth processes and a testing ground for the study of nonlinear stochastic PDEs. Moments of the SHE are important ingredients in extracting information about intermittency \cite{Amir11, Bertini1995, BC14, CK10, CK12, CJKS}, regularity \cite{SS00, SS02}, tail decay \cite{CDR10, CJK} etc. In this article, we prove one-point moment formulas of $\mathcal{Z}(T,X)$. Our main result is stated as follows.
\begin{thm}\label{MainTheo}
Let $\mathcal{Z}(T,X)$ be the unique solution of \eqref{eq:SHEDef}. Then, for any $k\in \mathbb{N}$
\begin{align}\label{eq:SHEMom}
\mathbb{E}_{\mathrm{SHE}}\left[(\mathcal{Z}(T, X))^{k}\right] = \frac{1}{(2\pi \mathbf{i})^{k}} \int_{\mathcal{C}_1}\ldots \int_{\mathcal{C}_k} \prod_{1\leq A < B\leq k} \frac{z_{A} - z_{B}}{z_{A} - z_{B}-1} e^{\frac{T}{2}\sum_{j=1}^{k} z^2_j + X\sum_{j=1}^{k}z_j} \prod_{j=1}^{k} dz_j &&&&
\end{align}
where $\mathcal{C}_j$ is the line $\alpha_j +\mathbf{i}\mathbb{R}$ such that $\alpha_1>\alpha_2+2> \ldots >\alpha_k+(k-1)$
\end{thm}
Recently, \cite[Theorem~2]{BorGor16} established an identity between the integer moments of $\mathcal{Z}(T,0)$ and the complete homogeneous symmetric functional moments of the \emph{Airy point process}. Their proof relies on the moment formulas of \cite[Proposition~5.4.8]{BC14} which are same as in Theorem~\ref{MainTheo}. Here, we demonstrate an alternative way deriving those formulas. The main technical innovation of this paper is to prove the correspondence (see Proposition~\ref{AiryToSHE}) between the moments of $\mathcal{Z}(T,0)$ and the Airy point process without using the formula of \cite[Proposition~5.4.8]{BC14}. In combination with a rigorous derivation of the Airy moments (see Lemma~\ref{ExpKSumLem}), this correspondence will provide a rigorous proof of Theorem~\ref{MainTheo}. Here, we would like to stress that the derivation of the Airy moments were known from the work of \cite{BorGor16}. For the sake of completeness, we will give a brief outline of their proof avoiding much details. In \cite[Proposition~6.2.3]{BC14}, the authors showed that the right hand side of \eqref{eq:SHEMom} solves the delta-Bose gas with delta potential. Combining our main theorem with \cite[Proposition~6.2.3]{BC14} yields the proof of the conjecture that the one-point moments of $\mathcal{Z}(T,X)$ coincides with the solution of the delta-Bose gas with delta potential. However, the uniqueness of the solution is not clear yet.
As a main tool, we use identity $(1)$ of \cite[Theorem~1]{BorGor16} which relates the Laplace transform of $\mathcal{Z}(T,0)$ with some multiplicative functional of the Airy point process. More specifically, they showed
\begin{align}\label{eq:identity}
\mathbb{E}_{\mathrm{Airy}}\Big[\prod_{p=1}^{\infty}\frac{1}{1+u\exp(C\mathbf{a}_p)}\Big]=\mathbb{E}_{\mathrm{SHE}}\Big[\exp\big(-u\mathcal{Z}(T,0)\exp(T/24)\big)\Big] , \quad \forall u\in \mathbb{R}_{+}
\end{align}
where $C=(T/2)^{1/3}$ and $\mathbf{a}_1\geq \mathbf{a}_2\geq \ldots $ are the ordered points of the Airy point process (see Section~\ref{SHEMoment} for its definition). This identity comes from some determinantal manipulation of the formulas in \cite[Theorem~1.1]{Amir11} and does not require the moment formula of $\mathcal{Z}(T,X)$ as an input. One should note that similar identities have also been observed in other models, namely, between the stochastic higher spin vertex model and the Macdonald measure \cite[Theorem~4.2]{B16Aug}, the asymmetric simple exclusion process (ASEP) and the discrete Laguerre ensemble \cite[Theorem]{BO16} etc. In fact, \eqref{eq:identity} follows from those identities by taking appropriate limits. From Taylor expanding the right hand side of \eqref{eq:identity}, one sees that coefficient of $(-u)^{k}$ is the $k$th moment of $\mathcal{Z}(T,0)$ upto some constant. However, the series obtained after Taylor expanding the the left hand side of \eqref{eq:identity} is divergent in nature. Because of this, one faces difficulties in interchanging derivatives and expectation on the right hand side \eqref{eq:identity}. We overcome these difficulties by considering only finitely many terms in the expansion of $[1+u\exp(C\mathbf{a}_p)]^{-1}$ for $p=1,2, \ldots $ and bounding the contributions of the remainder terms (see Section~\ref{ASHE} for more details) appropriately.
The moment formulas of \eqref{eq:SHEMom} were formally derived in \cite{BC14} in two different ways. One of those two ways hinges upon the folklore that the moments of $\mathcal{Z}(T,X)$ solves the attractive \emph{delta-Bose gas}. In \cite[Proposition~6.2.3]{BC14}, the right hand side of \eqref{eq:SHEMom} is shown to be a solution of the attractive delta-Bose gas with delta potential. Under the assumption that the folklore is true, their result indicates the moment formulas of Theorem~\ref{MainTheo}. However, a rigorous proof of this claim is missing except when $k=2$ (see \cite{AGHH}). Proof for the case $k=2$ relies on the observation that the moments of the SHE with smoothed out (in space) white noise solve the Lieb-Liniger many body problem with smoothed delta potential. It then suffices to show that the smoothed moments converges to the moments without smoothing and likewise, the solution of the Lieb-Liniger system with smooth potential converges to the solution of the delta-Bose gas with delta interaction. However, showing this twofold convergence is nontrivial and becomes complicated for higher order moments. To our knowledge, there is no follow up work where this approach is made to work for the moments of order greater than $2$.
The second approach of \cite{BC14} is based on the observation that the discrete and semi-discrete directed polymers converges in distribution to the SHE under the intermediate disorder scaling. See \cite{alberts2014, DR02} for the discrete case and \cite{Nica16} for the semi-discrete case. Moments of the semi-discrete polymer measure are given by some integral formulas of the Whittaker measure which is a degeneration of the Macdonald measure. By taking limit of the integral formulas of the Whittaker measure, \cite[Section~5.4.2]{BC14} obtained the multipoint moment formulas of the SHE. However, the convergence in distribution is not good enough to support the convergence of moments. To show the moments also converge, one needs uniform tail bounds of the pre-limiting object.
Prof. Vadim Gorin kindly informed us that it is indeed possible to obtain such bounds via Markov's inequality and analysis of known contour integral formulas for the semi-discrete polymer. Combining these tail bounds with the known weak convergence provides an alternative derivation of \eqref{eq:SHEMom}. For the sake of completeness, we include some details of this approach in Section~\ref{AlrMethod}. We should also point out that using this approach, it is possible to to obtain multi-point moment formulas of the SHE as conjectured in \cite[Proposition~5.4.6]{BC14}. For more details, we refer to \cite[Corollary~1.14]{Nica16}.
There is yet another approach to derive the moment formulas of Theorem~\ref{MainTheo}\footnote{Communicated to us by Guillaume Barraquand}. The main idea of this approach relies upon an identity between the distribution of $\mathcal{Z}(0,T)$ and a random infinite series involving the Airy point process and an infinite sequence of independent chi-squared random variables. This identity is proved in \cite[Theorem~1.3]{GS18} and is stated as follows
\begin{align}\label{eq:RandomSeries}
\mathcal{Z}(T,0)\exp(T/24) \stackrel{d}{=} 2\sum_{p=1}^{\infty} u^2_p\exp(C\mathbf{a}_p)
\end{align}
where $\{u^2_p\}_{p\in \mathbb{N}}$ is a sequence of independent
chi-squared random variables with $1$ degree of freedom. Hence, the $k$-th moment of $\mathcal{Z}(0,T)$ is given by the $k$-th moment of the random infinite series of the right hand side of \eqref{eq:RandomSeries} modulo the constant $\exp(-kT/24)$. Now, the contour integral formula of Theorem~\ref{MainTheo} can be obtained by expanding the $k$-th moments of the random infinite series and applying the moment formulas of the chi-squared random variables. However, we do not pursue the details of these computations in this paper.
Recently, identity \eqref{eq:identity} played a key role in \cite{CG18} to find the crossover behavior of the lower tail probability of the Cole-Hopf solution of the KPZ equation. In \cite{CGKLT18}, this identity was at the focal point in deriving the lower tail large deviation rate function of the KPZ equation. In an upcoming work \cite{CG18b}, the main result of this paper will be used to estimate the upper tail fluctuation of the KPZ equation. We should also point out that \cite[Proposition~2.3]{Bertini1995} obtained some moment formulas of the SHE under the general initial conditions. It is not clear whether those moment formulas (even for $k=2$) coincide with the right hand side of \eqref{eq:SHEMom} under the delta initial measure. Those formulas of \cite{Bertini1995} were later used (for instance in \cite{CJK, ChD15}) to get the upper tail estimates of the SHE. However, for our purpose in \cite{CG18b}, we need more refined estimates which we obtain from \eqref{eq:SHEMom}. This was our primary motivation to give a rigorous proof of Theorem~\ref{MainTheo}.
The rest of the paper is organized as follows. We prove Theorem~\ref{MainTheo} in Section~\ref{ProofOfMT}. In Section~\ref{AlrMethod}, we demonstrate one of the other alternative ways of proving Theorem~\ref{MainTheo}.
\section{Proof of Theorem~\ref{MainTheo}}\label{ProofOfMT}
Substituting $\tilde{z}_j = z_j+ \frac{X}{T}$ on the right hand side of \eqref{eq:SHEMom}, we get
\begin{align}
\text{r.h.s of \eqref{eq:SHEMom}} = \frac{e^{-\frac{kX^2}{2T}}}{(2\pi \mathbf{i})^{k}} \int_{\mathcal{C}_1}\ldots \int_{\mathcal{C}_k} \prod_{1\leq A<B \leq k} \frac{\tilde{z}_A - \tilde{z}_{B}}{\tilde{z}_A - \tilde{z}_{B}-1} e^{\frac{T}{2}\sum_{j=1}^{k}\tilde{z}^2_j} \prod_{j=1}^{k} d\tilde{z}_j.\label{eq:Subs}
\end{align}
Using spatial stationarity of the process $\mathcal{Z}(T,X)\exp\big(\frac{X^2}{2T}\big)$ (see \cite[Theorem~1.4]{Amir11}, \cite[Proposition~1.17(1)]{CorHam16}), it suffices to show
\begin{align}\label{eq:ObsMomeq}
\mathbb{E}_{\mathrm{SHE}}\left[(\mathcal{Z}(T, 0))^{k}\right] = \frac{1}{(2\pi \mathbf{i})^{k}} \int_{\mathcal{C}_1}\ldots \int_{\mathcal{C}_k} \prod_{1\leq A < B\leq k} \frac{z_{A} - z_{B}}{z_{A} - z_{B}-1} e^{\frac{T}{2}\sum_{j=1}^{k} z^2_j} \prod_{j=1}^{k} dz_j.
\end{align}
By residue expansion, \cite[Proposition~3.2.1]{BC14} (see also \cite[Proposition~5.1]{BBC16}) gave an alternative form of the right hand side of \eqref{eq:ObsMomeq}. Our next result shows that the moments of $\mathcal{Z}(T,0)$ matches with their alternative formula.
\begin{thm}\label{SHEMoment}
Continuing with the notation of Theorem~\ref{MainTheo}, one has
\begin{align}
\mathbb{E}_{\mathrm{SHE}}\Big[(\mathcal{Z}(T,0))^k\Big] = \sum_{\substack{\lambda\vdash k\\ \lambda = 1^{m_1} 2^{m_2}\ldots }} &\frac{k!}{m_1!m_2!\ldots } \int^{\mathbf{i}\infty}_{-\mathbf{i}\infty} \frac{dw_1}{2\pi \mathbf{i}}\ldots \int^{\mathbf{i}\infty}_{-\mathbf{i}\infty} \frac{dw_{\ell(\lambda)}}{2\pi \mathbf{i}} \mathrm{det}\left[\frac{1}{w_i+\lambda_i- w_j}\right]^{\ell(\lambda)}_{i,j=1}\\
&\times \prod_{j=1}^{\ell(\lambda)}\exp\left(\frac{T}{2}\left[w^2_j + (w_j+1)^2+ \ldots (w_j+\lambda_j-1)^2\right]\right). \label{eq:Moment}
\end{align}
Here, $\lambda \vdash k$ denotes a partition $k$. For any partition $\lambda =(\lambda_1 >\lambda_2>\ldots)$, the number of nonzero elements of $\lambda$, i.e., $|\{k: \lambda_k\neq 0\}|$ is called the length of the partition $\lambda$ and denoted by $\ell(\lambda)$.
\end{thm}
\textsc{Final step of proof of Theorem~\ref{MainTheo}:}
Combining \eqref{eq:Moment} with \cite[Proposition~3.2.1]{BC14} yields \eqref{eq:SHEMom}. This completes the proof.
\subsection{Proof of Theorem~\ref{SHEMoment}}
We prove Theorem~\ref{SHEMoment} using identity \eqref{eq:identity}. In order to do so, our first step is to give an alternative derivation of the equivalence between the moments of $\mathcal{Z}(T,X)$ and the complete homogeneous symmetric functional moments of the Airy point process (see Proposition~\ref{AiryToSHE}). This equivalence was first proved in \cite[Theorem~1.2]{BorGor16} subjected to the fact that the moments $\mathcal{Z}(T,X)$ solves the delta-Bose gas with delta potential. Unlike \cite{BorGor16}, we take the route of Taylor expanding both sides of \eqref{eq:identity} with respect $u$. At this point, it is not hard to guess that equating the coefficients of $(-u)^{k}$ on both sides of \eqref{eq:identity} one may get the desired identity of \cite[Theorem~1.2]{BorGor16}. Although, it sounds very straightforward, the exact details is convoluted due to the divergent nature of the power series obtained after Taylor's expansion. The main difficulty lies in exchanging the derivatives and the expectation which we deal in Section~\ref{AiryToSHE}. Our second step is to show that the right hand side of \eqref{eq:Moment} coincides with the complete homogeneous symmetric polynomial moment (of order $k$) of the Airy point process upto the factor $\exp\big(\frac{kT}{24}\big)$. This is proved in Lemma~\ref{ExpKSumLem}. We state our results after a brief overview of the Airy point process.
\begin{defn}
\emph{Airy point process} is a \emph{determinantal point process} on $\mathbb{R}$. A \emph{point process} on $\mathbb{R}$ is an integer-valued measure on the point configurations of the real line. Any point process $\chi$ is characterized by its correlation functions $\{\rho^{\chi}_k\}_{k\in \mathbb{N}}$. For any $k\in \mathbb{N}$, $\rho^{\chi}_k:\mathbb{R}^k\to \mathbb{R}_{\geq 0}$ is a locally integrable function such that for any Borel sets $B_1, \ldots ,B_k\in \mathcal{B}(\mathbb{R})$
\begin{align}\label{eq:Expectation}
\mathbb{E}\Big[\prod_{i=1}^{k} \mathbbm{1}_{\chi}(B_i)\Big]= \int_{\prod_{i=1}^kB_i} \rho_k(x_1, \ldots ,x_k) dx_1 \cdots dx_k
\end{align}
A point process $\chi$ is called \emph{determinantal} when $\rho^{\chi}(x_1, \ldots ,x_k)= \mathrm{det}[K_{\chi}(x_i,x_j)]_{1\leq i\leq j\leq k}$ for some $K_{\chi}:\mathbb{R}^2\to \mathbb{R}$, namely, the kernel of the point process $\chi$. The kernel of the Airy point process is given as
\begin{align}\label{eq:Airy}
K_{\mathrm{Ai}}(x,y) = \int^{\infty}_{-\infty}\mathrm{Ai}(x+t)\mathrm{Ai}(y+t) dt.
\end{align}
\end{defn}
\begin{prop}\label{AiryToSHE}
Fix $T>0$ and set $C= (T/2)^{\frac{1}{3}}$. Let $\mathbf{a}_1 \geq \mathbf{a}_2 \geq \ldots $ be the ordered points of the Airy point process. Let $h_k(x_1, x_2, \ldots )= \sum_{i_1\leq i_2\leq \ldots \leq i_k}x_{i_1}x_{i_2}\ldots x_{i_k}$ be the complete homogeneous symmetric functions in variables $x_1, x_2, \ldots $. Then, for any $k\in \mathbb{N}$,
\begin{align}\label{eq:kMomEq}
\mathbb{E}_{\mathrm{Airy}}\big[h_k(\exp(C\mathbf{a}_1),\exp(C\mathbf{a}_2), \ldots )\big] = \mathbb{E}_{\mathrm{SHE}}\Big[ \frac{(\mathcal{Z}(T,0))^k}{k!}\exp\big(\frac{kT}{24}\big)\Big]
\end{align}
\end{prop}
Proof of Proposition~\ref{AiryToSHE} is deferred to Section~\ref{ASHE}.
\begin{lemma}\label{ExpKSumLem}
Continuing with the notations of Theorem~\ref{SHEMoment} and Proposition~\ref{AiryToSHE}, one has
\begin{align}
\mathbb{E}_{\mathrm{Airy}}&\left[h_{k}\big(\exp(C\mathbf{a}_1, C\mathbf{a}_2, \ldots )\big)\right] = \sum_{\substack{\lambda\vdash k\\ \lambda = 1^{m_1}2^{m_2}\ldots }} \frac{\exp( \frac{k^3 T}{24})}{m_1! m_2!\ldots } \int^{\mathbf{i}\infty}_{-\mathbf{i}\infty}\frac{dw_1}{2\pi \mathbf{i}}\cdots \int^{\mathbf{i}\infty}_{-\mathbf{i}\infty} \frac{dw_{\ell(\lambda)}}{2\pi \mathbf{i}} \\\times &\mathrm{det}\left[\frac{1}{w_i+\lambda_i-w_j}\right]^{\ell(\lambda)}_{i,j=1}
\times \prod_{j=1}^{\ell(\lambda)} \exp\left(T/2\left[w^2_j+ (w_j+1)^2+\ldots +(w_j+\lambda_j-1)^2\right]\right).\label{eq:CIntFor}
\end{align}
\end{lemma}
\begin{proof}
Result of this lemma is implicitly present in the proof of Theorem~1.2 of \cite{BorGor16}. To avoid repetition, we present here a sketch of the proof and refer \cite{BorGor16} for more details. To begin, we define the Laplace transform of the Airy kernel as follows. For any vector $c=(c_1, \ldots ,c_n)$, we define
\begin{align}\label{eq:LaplCor1}
R(c_1, \ldots ,c_n) = \int^{\infty}_{-\infty} e^{\sum_{i=1}^n c_ix_i} \mathrm{det}[K_{\mathrm{Ai}}(x_i,x_j)]^{n}_{i,j=1} \prod_{i=1}^{n}dx_i.
\end{align}
One may now note that
\begin{align}\label{eq:LaplCor2}
\mathbb{E}[h_{k}(\exp(C\mathbf{a}_1),\exp(C\mathbf{a}_2), \ldots )] = \sum_{\substack{\lambda\vdash k\\ \lambda= 1^{m_1}2^{m_2}\ldots }} \frac{1}{m_1!m_2!\ldots } R(C\lambda_1, \ldots , C\lambda_{\ell(\lambda)}).
\end{align}
To complete the proof, it suffices to show that the contour integral formula on the right hand side of \eqref{eq:CIntFor} is same as $R(C\lambda_1, C\lambda_2, \ldots )$. Using the following identity (see \cite[Lemma~2.6]{Oku02})
\begin{align}
\int^{\infty}_{-\infty} e^{xz} \mathrm{Ai}(z+a) \mathrm{Ai}(z+b) dz = \frac{1}{2\sqrt{\pi x}} \exp\left(\frac{x^3}{12} - \frac{a+b}{2}x- \frac{(a-b)^2}{4x}\right)
\end{align}
we write
\begin{align}
R(c_1, \ldots ,c_n) = \frac{e^{\sum_{i=1}^n \frac{c^3_i}{12}}}{(2\pi)^n} \int_{z_1\in \mathbb{R}}dz_1\ldots \int_{z_n\in \mathbb{R}}dz_n &\exp\left(-\sum_{i=1}^{n}c_iz^2_i\right)\\
&\times\mathrm{det}\left[\frac{1}{(-\mathbf{i}z_i+\frac{c_i}{2})+ (\mathbf{i}z_j +\frac{c_j}{2})}\right]^{n}_{i,j=1}. \label{eq:RForm1}
\end{align}
To show that $R(C\lambda_1, \ldots ,C\lambda_{\ell(\lambda)})$ equals to the contour integral on the right hand side of \eqref{eq:CIntFor}, we set $c_i=C\lambda_i$ and substitute $\mathbf{i}z = Cw_j+C\frac{\lambda_j}{2} - \frac{C}{2}$ for $1\leq i\leq \ell(\lambda)$in \eqref{eq:RForm}. This entails to write
\begin{align}
R(C\lambda_n, \ldots ,C\lambda_{\ell(\lambda)}) =& \frac{e^{C^3\sum_{i=1}^{\ell(\lambda)} \frac{\lambda^3_i}{12}}}{(2\pi)^n} \int^{\mathbf{i}\infty}_{-\mathbf{i}\infty}dz_1\ldots \int^{\mathbf{i}\infty}_{-\mathbf{i}\infty}dz_{\ell(\lambda)} \\
&\times\exp\left(C^3\sum_{i=1}^{\ell(\lambda)}\lambda_i\left(w_i+\frac{\lambda_i}{2}-\frac{1}{2}\right)^2\right)\mathrm{det}\left[\frac{1}{w_j+\lambda_j-w_i}\right]^{\ell(\lambda)}_{i,j=1}. \label{eq:RForm}
\end{align}
Applying the following identity
\begin{align}\label{eq:Simple1}
\exp\left(C^3\sum_{i=1}^{\ell(\lambda)}\left(\frac{\lambda^3_i}{12}+\lambda_i\Big(w_i+\frac{\lambda_i}{2}-\frac{1}{2}\Big)^2\right)\right)
=\exp\left(C^3\sum_{i=1}^{\ell(\lambda)}\left(w_i^2+\ldots +(w_i+\lambda_i-1)^2\right)+\frac{(Ck)^3}{12}\right)
\end{align}
into the right hand side of \eqref{eq:RForm} indeed shows that $R(C\lambda_1, \ldots ,C\lambda_{\ell(\lambda)})$ coincides with the contour integral formula on the right hand side of \eqref{eq:CIntFor}. This completes the proof.
\end{proof}
\medskip
\textsc{Final step in the Proof of Theorem~\ref{SHEMoment}:} Proof of Theorem~\ref{SHEMoment} follows by combining the identity \eqref{eq:kMomEq} with the formula of $\mathbb{E}_{\mathrm{Airy}}\left[h_{k}\big(\exp(C\mathbf{a}_1, C\mathbf{a}_2, \ldots )\big)\right]$ in Lemma~\ref{ExpKSumLem}.
\subsection{Proof of Proposition~\ref{AiryToSHE}}\label{ASHE}
To prove Proposition~\ref{AiryToSHE}, we need two new inputs which are listed in Proposition~\ref{InfProdDeiv} and Proposition~\ref{SHEMomentExtract}.
\begin{prop}\label{InfProdDeiv}
Continuing with the notations of Proposition~\ref{AiryToSHE}, we have
\begin{align}\label{eq:AiryDeri}
\lim_{u\downarrow 0}\frac{1}{u^k}\mathbb{E}_{\mathrm{Airy}}&\left[\prod_{p=1}^{\infty}\frac{1}{1+u\exp(C\mathbf{a}_{p})}-\sum_{m=0}^{k-1} h_{m}(\exp(C\mathbf{a}_1), \exp(C\mathbf{a}_2), \ldots )(-u)^{m}\right] \\& = (-1)^{k} \mathbb{E}_{\mathrm{Airy}}\Big[h_k\big(\exp(C\mathbf{a}_1), \exp(C\mathbf{a}_2), \ldots \big)\Big]
\end{align}
for any $k\in \mathbb{N}$.
\end{prop}
Proof of Proposition~\ref{InfProdDeiv} contains the core of the technical part of this paper. For clarity, we defer its proof to Section~\ref{LongProp}.
\begin{prop}\label{SHEMomentExtract}
Recall that $\mathcal{Z}(T,X)$ is the unique solution of the SHE started from the delta initial data. Then, we have
\begin{align}\label{eq:sheMGFDeri}
\lim_{u \downarrow 0} \frac{1}{u^k}\mathbb{E}_{\mathrm{SHE}}&\left[\exp\Big(- u\mathcal{Z}(T,0)\exp(\frac{T}{24})\Big)- \sum_{\ell=0}^{k-1}(-u)^{t}\frac{[\mathcal{Z}(T,0)]^{\ell}\exp(\frac{\ell T}{24})}{\ell!}\right] \\& = (-1)^{k}\mathbb{E}_{\mathrm{SHE}}\Big[\frac{(\mathcal{Z}(T,0))^{k}\exp(\frac{kT}{24})}{k!}\Big].
\end{align}
\end{prop}
\begin{proof}
Applying mean value theorem, we note
\begin{align}\label{eq:MVTappl}
\left|\exp\big(-u \theta\big) - \sum_{\ell=0}^{k} (-u)^{\ell} \frac{\theta^{\ell}}{\ell!}\right|\leq u^{k}\frac{\theta^{k+1}}{(k+1)!} \exp(-\bar{u}\theta), \quad \text{ for some }\bar{u}= \bar{u}(\theta)\in (0,u).
\end{align}
Owing to \eqref{eq:MVTappl}, we get
\begin{align}\label{eq:ExpUnifBd}
\frac{1}{u^{k}} \mathbb{E}&\left|\exp\big(-u \mathcal{Z}(T,0)\exp(\frac{T}{24})\big) - \sum_{\ell=0}^{k} (-u)^{\ell} \frac{(\mathcal{Z}(T,0))^{\ell}\exp(\frac{\ell T}{24})}{\ell!}\right|\\&\leq u\mathbb{E}\left[\frac{(\mathcal{Z}(T,0))^{k+1}}{(k+1)!}\exp\big(\frac{(k+1)T}{24}\big)\right]
\end{align}
where the right hand side of \eqref{eq:ExpUnifBd} is finite (see \cite[Example~2.10]{ChD15} for its upper bound). Now, letting $u$ to go to $0$ on both sides of \eqref{eq:ExpUnifBd}, we arrive at \eqref{eq:sheMGFDeri}. This completes the proof.
\end{proof}
\subsubsection{Final step in the proof of Proposition~\ref{AiryToSHE}}
We prove \eqref{eq:kMomEq} by induction. Let us first prove \eqref{eq:kMomEq} for $k=1$. Subtracting $1$ from sides of identity \eqref{eq:identity}, diving both sides by $u$, letting $u$ to go to $0$ and applying Proposition~\ref{InfProdDeiv} and Proposition~\ref{SHEMomentExtract} yield \eqref{eq:kMomEq} for $k=1$. Now, we assume that \eqref{eq:kMomEq} holds for all $k\in \mathbb{Z}_{[1,k_0]}$ where $k_0\in \mathbb{N}$. To prove \eqref{eq:kMomEq} for $k=k_0+1$, using \eqref{eq:identity} and \eqref{eq:kMomEq} for $k=1,2, \ldots , k_0$, we note
\begin{align}\label{eq:RecentIdenty}
\frac{1}{u^{k_0+1}}\mathbb{E}_{\mathrm{Airy}}&\left[\prod_{\ell=1}^{\infty}\frac{1}{1+u\exp(C\mathbf{a}_{\ell})}-\sum_{m=0}^{k_0} h_{m}(\exp(C\mathbf{a}_1), \exp(C\mathbf{a}_2), \ldots )(-u)^{m}\right]\\ & = \frac{1}{u^{k_0+1}}\mathbb{E}_{\mathrm{SHE}}\left[\exp\Big(- u\mathcal{Z}(T,0)\exp(\frac{T}{24})\Big)- \sum_{\ell=0}^{k_0}(-u)^{\ell}\frac{[\mathcal{Z}(T,0)]^\ell\exp(\frac{\ell T}{24})}{\ell!}\right].
\end{align}
Now, taking $u$ to $0$ on both sides of \eqref{eq:RecentIdenty} and applying Proposition~\ref{InfProdDeiv} and Proposition~\ref{SHEMomentExtract}, we arrive at \eqref{eq:kMomEq} for $k=k_0+1$. This completes the proof.
\subsubsection{Proof of Proposition~\ref{InfProdDeiv}}\label{LongProp}
We first prove \eqref{eq:AiryDeri} for $k=1$. Fixing $n\in \mathbb{N}$, we notice
\begin{align}\label{eq:Expand}
\prod_{p=1}^{n}\frac{1}{1+u\exp(C\mathbf{a}_{p})}-1 = -\sum_{p=1}^{n} \frac{u\exp(C\mathbf{a}_{p})}{1+u\exp(C\mathbf{a}_{p})}\prod_{m=p+1}^{n}\frac{1}{1+u\exp(C\mathbf{a}_{m})}
\end{align}
which shows that
\begin{align}\label{eq:BaseIneq}
\left|\prod_{p=1}^{n}\frac{1}{1+u\exp(C\mathbf{a}_{p})}-1\right| \leq u\sum_{p=1}^{n} \exp(C\mathbf{a}_{p}).
\end{align}
Taking $n$ to $\infty$ on both sides of \eqref{eq:BaseIneq}, we get
\begin{align}\label{eq:InftyIneq}
\left|\prod_{p=1}^{\infty}\frac{1}{1+u\exp(C\mathbf{a}_{p})}-1\right| \leq u\sum_{p=1}^{\infty} \exp(C\mathbf{a}_{p})= uh_1(\exp(C\mathbf{a}_1), \exp(C\mathbf{a}_2), \ldots ).
\end{align}
Since, the right hand side of \eqref{eq:InftyIneq} is in $L^{1}$, therefore, by the dominated convergence theorem and monotonocity of both sides of \eqref{eq:Expand}, we write
\begin{align}\label{eq:InfProdMinus1}
\prod_{p=1}^{\infty}\frac{1}{1+u\exp(C\mathbf{a}_{p})}-1 = -\sum_{p=1}^{\infty} \frac{u\exp(C\mathbf{a}_{p})}{1+u\exp(C\mathbf{a}_{p})}\prod_{m=p+1}^{\infty}\frac{1}{1+u\exp(C\mathbf{a}_{m})}.
\end{align}
Furthermore, owing to \eqref{eq:InftyIneq}, one observes
\begin{align}\label{eq:Ineq2}
\frac{1}{u}\left|\mathbb{E}\big[\prod_{p=1}^{\infty}\frac{1}{1+u\exp(C\mathbf{a}_p)}\big]-1\right|\leq \frac{1}{u}\mathbb{E}\left|\prod_{p=1}^{\infty}\frac{1}{1+u\exp(C\mathbf{a}_p)}-1\right|
\end{align}
where the right hand side is bounded above by $\mathbb{E}\big[h_1(\exp(C\mathbf{a}_1), \exp(C\mathbf{a}_2), \ldots)\big].$ Thanks to \eqref{eq:Ineq2}, one can do the following interchange of limit and integral
\begin{align}\label{eq:Interchange}
\lim_{u\to 0}\frac{1}{u}\Big[\mathbb{E}\Big[\prod_{p=1}^{\infty}\frac{1}{1+u\exp(C\mathbf{a}_{p})}\Big]-1\Big] = \mathbb{E}\Big[\lim_{u\to 0}\frac{1}{u}\Big[\prod_{p=1}^{\infty}\frac{1}{1+u\exp(C\mathbf{a}_{p})}-1\Big]\Big]
\end{align}
if the limit exists. To see why the limit on the right hand side of \eqref{eq:Interchange} exists, we divide both sides of \eqref{eq:InfProdMinus1} by $u$. As $u$ goes to $0$, owing to \eqref{eq:InftyIneq} and the dominated convergence theorem, one has
\begin{align}\label{eq:RHLimit}
\lim_{u\to 0}\frac{1}{u}\left[\prod_{p=1}^{\infty}\frac{1}{1+u \exp(C\mathbf{a}_{p})}-1\right]= -\sum_{\ell=1}^{\infty}\exp(C\mathbf{a}_{p}) \quad \text{in }L^{1}.
\end{align}
Plugging the limit of \eqref{eq:RHLimit} into the right hand side of \eqref{eq:Interchange}, we arrive at \eqref{eq:AiryDeri} for $k=1$.
Now, we prove \eqref{eq:AiryDeri} for any $k\in \mathbb{N}_{>1}$. For this, we first notice the following.
\medskip
\textbf{Claim:} For any $k\in \mathbb{N}$, we have
\begin{align}\label{eq:Expj}
\prod_{p=1}^{\infty} \frac{1}{1+u\exp(C\mathbf{a}_{p})} &= \prod_{p=1}^{\infty}\Big[\sum_{m=0}^{k}(-u\exp(C\mathbf{a}_{p}))^{m}+ \frac{(-u\exp(C\mathbf{a}_p))^{k+1}}{1+u\exp(C\mathbf{a}_{p})} \Big]
\end{align}
for all $u< \exp(-C\mathbf{a}_1)$.
\medskip
\textsc{Proof of Claim:}
Proof of this claim directly follows by noting that one can write
\begin{align}
\frac{1}{1+u\exp(C\mathbf{a}_{p})} = \sum_{m=0}^{k} (- u \exp(C\mathbf{a}_{p}))^{m} + \frac{(-u \exp(C \mathbf{a}_m))^{k+1}}{1+u \exp(C\mathbf{a}_{p})}\quad \forall k,p \geq 1
\end{align}
whenever $u < \exp(- C \mathbf{a}_1)$.
\medskip
\textbf{Claim:} Fix any $k\in \mathbb{N}$. Then, for any $u\in ( 0, \min\{e^{\frac{-Tk^3}{24}}, \exp(-C\mathbf{a}_1)\})$,
\begin{align}\label{eq:TrunPSer}
\prod_{p=1}^{\infty}\big(\sum_{m=0}^{k} (- u\exp(C\mathbf{a}_p))^{m}\big) = \sum_{n=0}^{\infty} h^{(k)}_n(\exp(C\mathbf{a}_1), \exp(C\mathbf{a}_2), \ldots )(-u)^{t}
\end{align}
where $h^{(k)}_n$ is defined as
\begin{align}\label{eq:htkdef}
h^{(k)}_n\big(x_1, x_2, \ldots\big) := \sum_{(i_1, i_2, \ldots ,i_n)\in \mathcal{S}^{(k)}_n} x_{i_1}\cdot x_{i_2}\cdots x_{i_n}
\end{align}
such that
\[
\mathcal{S}^{(k)}_n = \{(i_1,i_2, \ldots ,i_n): i_1\leq i_2\leq \ldots \leq i_n, \text{ s.t. }|\{i_1, i_2, \ldots ,i_n\}\cap \{m\}|\leq k,\quad \forall m\in \mathbb{N}\}.\]
The right hand side of \eqref{eq:TrunPSer} is a power series absolutely convergent in $L^{1}$ for all $u\in (0,e^{\frac{-Tk^3}{24}})$. Moreover, for all $t\in \mathbb{Z}_{[0,k]}$, one has $h^{(k)}_t(x_1, x_2, \ldots )= h_t(x_1, x_2, \ldots )$.
\medskip
\textsc{Proof of Claim:}
Define the following power series
\begin{align}\label{eq:DefHk}
H_k(u):=\sum_{n=0}^{\infty} h^{(k)}_n\big(\exp(C\mathbf{a}_1), \exp(C \mathbf{a}_2), \ldots\big)u^{n}
\end{align}
We first show that $H_k(u)$ is absolutely convergent in $L^{1}$ for all $u\in (0, \exp(-\frac{k^3T}{24}))$. Appealing to \eqref{eq:htkdef}, we note
\begin{align}\label{eq:hkExp}
\mathbb{E}\Big[h^{(k)}_n\big(\exp(C\mathbf{a}_1), \exp(C \mathbf{a}_2), \ldots\big)\Big] = \sum_{\substack{\lambda\vdash n\\\lambda= 1^{m_1} 2^{m_2}\ldots \\\lambda_1\leq k \ldots}} \frac{1}{m_1!m_2!\ldots } R(C\lambda_1, \ldots , C\lambda_{\ell(\lambda)}).
\end{align}
Using \eqref{eq:RForm1} and Cauchy's determinantal formula \cite{C95}, we arrive at
\begin{align}\label{eq:RFormula}
R(C\lambda_1, \ldots , C\lambda_{\ell(\lambda)}) = \frac{e^{\sum_{i=1}^{\ell(\lambda)}\frac{C\lambda^3_i}{12}}}{(2\pi)^{\frac{n}{2}}} \prod_{i=1}^{\ell(\lambda)}\frac{1}{(2C\lambda_i)^{\frac{3}{2}}} \mathbb{E}\left[\prod_{1\leq i<j\leq \ell(\lambda)}\frac{(Z_i-Z_j)^2+ \frac{1}{4}(\lambda_i-\lambda_j)^2}{(Z_i+Z_j)^2+ \frac{1}{4}(\lambda_i-\lambda_j)^2}\right] &&
\end{align}
where $Z_1, Z_2, \ldots , Z_{\ell(\lambda)} $ are i.i.d $N(0,1)$ random variables. Plugging \eqref{eq:RFormula} into the right hand side of \eqref{eq:hkExp}, noting that that each term of the sum is bounded above by $\exp(\frac{nTk^3}{24})$ and applying Siegel's bound (see \cite[pp. 316-318]{Apostol76}, \cite[pp. 88-90]{Knopp70}) to the number of partitions of $n$, we arrive at
\begin{align}
\text{l.h.s of \eqref{eq:hkExp}} \leq e^{\sqrt{n}+\frac{nTk^3}{24}}.
\end{align}
This entails to write
\begin{align}\label{eq:ExpOfSumBd}
\mathbb{E}\left[\sum_{t=0}^{\infty}h^{(k)}_n\big(\exp(C\mathbf{a}_1), \exp(C\mathbf{a}_2), \ldots \big)u^n\right]\leq \sum_{n=0}^{\infty} u^{n}e^{\sqrt{n}+\frac{nTk^3}{24}}
\end{align}
where the right hand side is finite for all $u\in (0, \exp(-\frac{Tk^3}{24}))$. This proves that $H_k(u)$ is absolutely convergent for all $u\in (0, \exp(-\frac{Tk^3}{24}))$ in $L^{1}$.
It remains to show that the left hand side of \eqref{eq:TrunPSer} is indeed same as $H_k(u)$ for all $u\in (0, \min\{\exp(-\frac{Tk^3}{24}), \exp(-C\mathbf{a}_1)\})$. To prove this, for any $Q\in \mathbb{N}$, we see that
\begin{align}\label{eq:nTruncPSer}
\prod_{p=1}^{Q}\Big(\sum_{m=0}^{\infty}\big(-u\exp(C\mathbf{a}_p)\big)^m\Big) = \sum_{n=0}^{\infty}h_n^{(k)}\big(\exp(C\mathbf{a}_1), \exp(C \mathbf{a}_2), \ldots , \exp(C\mathbf{a}_Q)\big)(-u)^{n}.&&
\end{align}
Thanks to $h^{(k)}_n(x_1, x_2, \ldots,x_Q)\leq h^{(k)}_n(x_1,x_2, \ldots )$ for all $Q\in \mathbb{N}$ and $x_1,x_2,\ldots \geq 0$, one observes
\begin{align}
|\text{l.h.s of \eqref{eq:nTruncPSer}}|\leq H_k(u).
\end{align}
To this end, using the dominated convergence theorem and absolute convergence of $H_k(u)$, we obtain \eqref{eq:TrunPSer} for all $u\in (0, \min\{\exp(-\frac{Tk^3}{24}), \exp(-C\mathbf{a}_1)\})$.
\medskip
\begin{lemma}
Continuing with the notation of Proposition~\ref{ExpKSumLem}, for any $k\in 2\mathbb{N}$ and $u\in (0, \exp(- \frac{Tk^3}{3}))$, we define
\begin{align}
\mathbf{X}_k:= \Big|\prod_{p=1}^{\infty} &\frac{1}{1+u\exp(C\mathbf{a}_{p})} - \sum_{t=0}^{k}h_n(\exp(C\mathbf{a}_1), \exp(C\mathbf{a}_2), \ldots ) (-u)^{n}\Big|.
\end{align}
Then $\mathbf{X}_k$ satisfies the following.
\begin{enumerate}[(i)]
\item For $u<\exp(-C\mathbf{a}_1)$,
\begin{align}\label{eq:L1ExpBd}
\mathbf{X}_k \leq \sum_{n=k+1}^{\infty} h^{(k)}_n(\exp(C\mathbf{a}_1), \exp(C\mathbf{a}_2), \ldots )u^{n} + u^{k+1}H_k(u)\sum_{p=1}^{\infty}\exp(C(k+1)\mathbf{a}_p) . &&&
\end{align}
See \eqref{eq:DefHk} for the definition of $H_k(.)$.
\item The right hand side of \eqref{eq:L1ExpBd} is in $L^{1}$ for all $u\in (0, \exp(-\frac{Tk^3}{3}))$.
\item We have
\begin{align}\label{eq:XkLimit}
\lim_{u\to 0}\frac{1}{u^k}\mathbb{E}_{\mathrm{Airy}}\mathbf{X}_k=0.
\end{align}
\end{enumerate}
\end{lemma}
\begin{proof}
\begin{enumerate}[(i)]
\item Owing to \eqref{eq:Expj}, for all $u< \exp(-C\mathbf{a}_1)$ and $k\in 2\mathbb{N}$, we observe
\begin{align}\label{eq:InfProdRep}
\prod_{p=1}^{\infty} \frac{1}{1+u \exp(C\mathbf{a}_{p})} = \prod_{p=1}^{\infty}\left(\sum_{m=0}^{k} (-u \exp(C\mathbf{a}_{p}))^m\right)\prod_{p=1}^{\infty} \frac{1}{1+(u\exp(C\mathbf{a}_p))^{k+1}}.
\end{align}
In the same way as in \eqref{eq:InfProdMinus1}, one can now write
\begin{align}\label{eq:InfProdBreak}
\prod_{p=1}^{\infty}\frac{1}{1+(u\exp(C\mathbf{a}_{p}))^{k+1}} =1 - \sum_{p=1}^{\infty} \frac{(u\exp(C\mathbf{a}_{p}))^{k+1}}{1+(u\exp(C\mathbf{a}_{p}))^{k+1}} \prod_{m=p+1}^{\infty}\frac{1}{1+(u\exp(C\mathbf{a}_{m}))^{k+1}}. &&&
\end{align}
Plugging \eqref{eq:InfProdBreak} into the right hand side of \eqref{eq:InfProdRep}, applying \eqref{eq:TrunPSer} and noting that after subtracting $1$ from the right hand side of \eqref{eq:InfProdBreak}, the rest is bounded above by $u^{k+1}\sum_{p=1}^{\infty}\exp(C(k+1)\mathbf{a}_p)$, we arrive at \eqref{eq:L1ExpBd}. This completes showing \eqref{eq:L1ExpBd}.
\medskip
\item
The first term on the right hand side of \eqref{eq:L1ExpBd} is bounded above by $H_k(u)$, thus, belongs to $L^{1}$ for all $u\in (0,\exp(-\frac{Tk^3}{12}))$. To complete proving $(ii)$, one now needs to show that
\begin{align}\label{eq:Ek}
\mathbb{E}H_k(u)\sum_{p=1}^{\infty} \exp\big(C(k+1)\mathbf{a}_{p}\big)<\infty
\end{align}
for all $u\in (0, \exp(-\frac{Tk^3}{3}))$. To show this, we observe
\begin{align}
H_k(u)\sum_{p=1}^{\infty}\exp(C(k+1)\mathbf{a}_p)\leq \sum_{n=0}^{\infty}h^{(2k+1)}_{n+k+1}(\exp(C\mathbf{a}_1), \exp(C\mathbf{a}_2), \ldots )u^{n}.\label{eq:CSIineqApp}
\end{align}
which holds because of the following inequalities \[h^{(k)}_{n}(x_1,x_2, \ldots )\times \sum_{p=1}^{\infty}x^{k+1}_p\leq h^{(2k+1)}_{n+k+1}(x_1,x_2, \ldots ),\quad n\in\mathbb{Z}_{\geq 0}, k\in \mathbb{N}, x_1,x_2,\ldots \in\mathbb{R}_{+}\]
Expectation of the right hand side of \eqref{eq:CSIineqApp} is finite for all $u\in (0, \exp(-\frac{T(2k)^3}{24}))$. This implies \eqref{eq:Ek}.
\item Write
\begin{align}\label{eq:SplitX}
\frac{1}{u^{k}}\mathbf{E}\mathbf{X}_k = \frac{1}{u^k} \mathbb{E}\mathbf{X}_k \mathbbm{1}(u<\exp(-C\mathbf{a}_1)) + \frac{1}{u^k} \mathbb{E}\mathbf{X}_k \mathbbm{1}(u\geq\exp(-C\mathbf{a}_1)).
\end{align}
Thanks to \eqref{eq:L1ExpBd} and part $(ii)$ of this proposition, the first term on the right hand side of \eqref{eq:SplitX} converges to $0$. Note that $\mathbf{X}_k\in L^2$ and furthermore, $\mathbb{E}\mathbf{X}^2_k$ is bounded by a constant for all $u\in(0, \exp(-\frac{Tk^3}{24}))$. Applying Cauchy-Schwarz inequality, one sees
\begin{align}\label{eq:CSIneq}
\mathbb{E}\mathbf{X}_k \mathbbm{1}(u\geq\exp(-C\mathbf{a}_1))\leq \sqrt{\mathbb{E}[\mathbf{X}^2_k]} \mathbb{P}(\mathbf{a}_1\geq -C^{-1}\log u)^{\frac{1}{2}}.
\end{align}
The right hand side of \eqref{eq:CSIneq} is bounded above by $C\exp(-\frac{4}{3}T^{-1/2}(\log u^{-1})^{3/2})$ (thanks for instance to \cite[Theorem~1.3]{RRV11}) which if divided by $u^{k}$ converges to $0$ as $u\downarrow 0$. This shows that the second term on the right hand side of \eqref{eq:SplitX} is also converging to $0$ as $u$ nears $0$. This completes proving \eqref{eq:XkLimit}.
\end{enumerate}
\end{proof}
\medskip
\textsc{Final step in the proof of Proposition~\ref{InfProdDeiv}:} From \eqref{eq:XkLimit}, it immediately follows that
\begin{align}
\lim_{u\to 0}\frac{1}{u^k} \mathbb{E}&\left[\prod_{p=1}^{\infty} \frac{1}{1+u\exp(C\mathbf{a}_{p})} - \sum_{n=0}^{k-1}h_n(\exp(C\mathbf{a}_1), \exp(C\mathbf{a}_2), \ldots ) (-u)^{n}\right] \\&= \mathbb{E}\Big[(-1)^{k}h_k(\exp(C\mathbf{a}_1), \exp(C\mathbf{a}_2), \ldots )\Big], \quad k\in 2\mathbb{N}.
\end{align}
Therefore, to complete the proof, it suffices to show \eqref{eq:XkLimit} holds when $k$ is odd. However, using triangle inequality, one may note that
\begin{align}\label{eq:kodd}
\mathbf{X}_k\leq \mathbf{X}_{k+1}+ u^{k+1}h_{k+1}(\exp(C\mathbf{a}_1), \exp(C \mathbf{a}_2), \ldots).
\end{align}
Dividing both sides of \eqref{eq:kodd} by $u^{k}$ and letting $u$ to go to $0$, we see that right hand side of \eqref{eq:kodd} converges to $0$ in $L^{1}$ (thanks to \eqref{eq:XkLimit} for $k+1$). This proves \eqref{eq:XkLimit} when $k$ is odd.
\section{Alternative Method}\label{AlrMethod}
In this section, we give a proof of Theorem~\ref{SHEMoment} following a suggestion of Prof. Vadim Gorin \cite{VGPrivate}. This approach is based on the computation of the limiting moment formulas of the semi-discrete directed polymers. In what follows, we recall the definition of the semi-discrete directed random polymers and its moment formulas. Subsequently, we finish proving Theorem~\ref{MainTheo} in Section~\ref{Last}.
\subsection{Semi-discrete directed random polymers}
\begin{defn}
A path of length $N$ in $\mathbb{R}\times \mathbb{Z}$ is a sequence of $N$ points from $\mathbb{R}\times \mathbb{Z}$. We call a path up-right and increasing if it either proceeds to the right or jumps by one unit above. For each sequence $0<s_1<\ldots <s_{N-1}<t$, we associate a path $\phi$ from $(0,1)$to $(t, N)$ such that $\phi$ jumps between the points $(s_{i},i)$ and $(s_{i+1}, i)$ for $i=1, \ldots , N-1$ and remain continuous otherwise. Let $B_1, \ldots ,B_N$ are $N$ independent Brownian motions. We define energy of the path $\phi$ as
\begin{align}
E(\phi) = B_1(s_1) + \big[B_2(s_2) - B_2(s_1)\big]+ \ldots + \big[B_{N}(t) - B_{N}(s_{N-1})\big].
\end{align}
Then, the \emph{semi-discrete polymer measure} is defined as follows
\begin{align}
\mathbf{Z}^{N}(t) = \int e^{E(\phi)} d\phi.
\end{align}
where the integral is defined with respect to the Lebesgue measure on the set of all up-right and increasing paths.
\end{defn}
\begin{prop}[Proposition~5.2.8 of \cite{BC14}]\label{MomProp}
For any $k\geq 1$, one has
\begin{align}
\mathbb{E}\Big[ (\mathbf{Z}^{N}(t))^k\Big] = \frac{1}{(2\pi \mathbf{i})^{k}} \oint \ldots \oint \prod_{1\leq A< B \leq k} \frac{w_{A} - w_{B}}{w_{A} - w_{B}-1}\prod_{j=1}^{k}z^{-N}_j e^{t w_j} dw_j
\end{align}
where the $w_A$ contour contains only the poles at $\{w_{A+1}+1, \ldots ,w_{k}+1, 0\}$.
\end{prop}
In \cite{Nica16}, it was shown that the semi-discrete directed polymer measure converges to the solution of the SHE started from delta initial measure. To illustrate, let us define
\begin{align}
C(N,T,X) = \exp\Big(N+ \frac{\sqrt{NT}+ X}{2} + X T^{-\frac{1}{2}}N^{\frac{1}{2}}\Big) (T^{\frac{1}{2}} N^{-\frac{1}{2}})^{N}.
\end{align}
\begin{prop}[Corollary~1.6 of \cite{Nica16}]\label{CDRPlim}
Fix $T\in \mathbb{R}_{>0}$ and $X\in \mathbb{R}$. Then,
\begin{align}\label{eq:CDRPLim}
\mathcal{Z}^{N}(T,X) := \frac{\mathbf{Z}^{N}\big(\sqrt{TN}+X\big)}{C(N,T,X)} \Rightarrow \mathcal{Z}(T,X) \quad \text{as } N \to \infty
\end{align}
where $\mathcal{Z}(T, X)$ is the unique solution of \eqref{eq:SHEDef} and '$\Rightarrow$' denotes the convergence in distribution.
\end{prop}
The following proposition shows that the moments of $\mathcal{Z}^{N}(T,X)$ converge as $N$ goes to $\infty$.
\begin{prop}
Fix $T>0$, $X\in \mathbb{R}$. Then, for $k\geq 1$, we have
\begin{align}\label{eq:MomLim}
\lim_{N\to \infty} \mathbb{E}[(\mathcal{Z}^{N}(T,X))^k] = \frac{1}{(2\pi \mathbf{i})^{k}} \int \ldots \int \prod_{1\leq A<B\leq k} \frac{z_A-z_{B}}{z_{A} -z_{B}-1} \prod_{j=1}^{k} e^{\frac{T}{2}z^2_j + Xz_j} dz_j.
\end{align}
where the $z_{A}$-contour is along $C_{A}+\mathbf{i}\mathbb{R}$ where $C_1>C_2+1>C_3+2>\ldots >C_k+(k-1)$.
\end{prop}
\begin{proof}
Set $t = \sqrt{TN}+X$. Owing to Proposition~\ref{MomProp}, we see
\begin{align}\label{eq:PreMomForm}
\mathbb{E}[(\mathbf{Z}^{N}(t))^k]= \frac{1}{(2\pi \mathbf{i})^{k}} \oint \ldots \oint \prod_{1\leq A< B \leq k}\frac{z_{A} - z_{B}}{z_{A} - z_{B}-1} \prod^{k}_{j=1}z^{-N}_{j} e^{t z_{j}} dz_j
\end{align}
where the contours contain all the singularities of the integrand. Let us examine the factor $e^{tz}/ z^{N}$. Note that $tz - N\log z$ has a critical point at $z_c= N/t$. When $N$ is large, then,
\begin{align}\label{eq:CritForm}
z_c = N/(\sqrt{TN}+X) = N^{\frac{1}{2}} T^{-\frac{1}{2}} (1- X/\sqrt{TN}).
\end{align}
Letting $z= z_c+w-\frac{X}{T}$ and Taylor expanding $e^{tz -N\log z}$ around the point $z_c$, we get
\begin{align}
e^{tz- N\log z} &= e^{N +\sqrt{TN}(w-X/T)+ N \log (\sqrt{N/T}) -w\sqrt{TN} + O(1/N^{1/2})} e^{Xw+\frac{1}{2}T w^2}\\
& = C(N,T, X)e^{Xw+\frac{1}{2}Tw^2 + O(1/\sqrt{N})}. \label{eq:MainIntgLimit}
\end{align}
One gets $\mathbb{E}[(\mathcal{Z}^{N}(T,X))^k]$ by dividing the right hand side of \eqref{eq:PreMomForm} by $C(N,T,X)^k$. This cancels out $C(N,T,X)^k$ which are obtained by factoring $e^{tz_j- N\log z_j}$ (see \eqref{eq:MainIntgLimit}) for $j=1, \ldots, k$. Substituting $z_j= z_c+w_j-X/T$, we see
\begin{align}\label{eq:IntActLimit}
\prod_{1\leq A< B\leq k}\frac{z_{A} - z_{B}}{z_{A} -z_{B}-1} = \prod_{1\leq A<B\leq k} \frac{w_{A} -w_{B}}{w_{A} - w_{B}-1}.
\end{align}
Due to the exponential decay of the integrand in the region far away from the critical point, the contours transform to a sequence of bi-infinite lines which are ordered among themselves. Combining this with \eqref{eq:MainIntgLimit} and \eqref{eq:IntActLimit} and applying those to \eqref{eq:PreMomForm}, we get \eqref{eq:MomLim}.
\end{proof}
\medskip
\subsection{Final steps of the proof of Theorem~\ref{MainTheo}}\label{Last}
Note that the limiting value in \eqref{eq:MomLim} matches with the right hand side of \eqref{eq:SHEMom}. However, it is not enough to show that the moments of the SHE are finite. As we have pointed out in Section~\ref{Intro}, one also need some uniform bound on the tail probability. To achieve this, using Markov's inequality, we see
\begin{align}\label{eq:Markov}
\mathbb{P}\Big(\mathcal{Z}^{N}(T,X)\geq M\Big) &\leq \frac{1}{M^{k+1}} \mathbb{E}\big[(\mathcal{Z}^{N}(T,X))^{k+1}\big].
\end{align}
Thanks to \eqref{eq:MomLim}, we see that the right hand side of \eqref{eq:Markov} is bounded above by $C/M^{k+1}$ for all large $M$ where $C$ does not depend on $N$. Taking $N$ to $\infty$ on both sides of \eqref{eq:Markov} and combining with \eqref{eq:CDRPLim}, one obtains
\begin{align}
\mathbb{P}\Big(\mathcal{Z}(T,X)\geq M\Big)\leq \frac{C}{M^{k+1}}, \quad \forall M>0
\end{align}
which shows $\mathbb{E}[(\mathcal{Z}(T,X))^k]<\infty$. Owing to the uniform integrability of $\{\mathcal{Z}^{N}(T,X)\}_{N\in \mathbb{N}}$ and the dominated convergence theorem, the left hand side of \eqref{eq:MomLim} is equal to $\mathbb{E}[(\mathcal{Z}(T,X))^k]$. This completes the proof.
\section*{Acknowledgment}
This work would not have been possible without numerous inputs and valuable suggestions of Prof. Ivan Corwin to whom the author likes to express his earnest gratitude. The author is thankful to Guillaume Barraquand for several insightful discussions and suggestions about the proof techniques. The author also likes to thank Prof. Alexei Borodin and Prof. Vadim Gorin for their comments on the early version of this paper and pointing out other methods to derive the main result of this paper.
\bibliographystyle{alpha}
|
\section{Introduction}
The Hubbard model, since its formulation \cite{Anderson1959, Hubbard1963, Gutzwiller1963, Kanamori1963}, has been intensively studied by the solid state physicists.
Being the first model capable to describe the metal-insulator (Mott) transition, it has also been studied in relation to such problems as the magnetic phase transitions, high-temperature superconductivity, optical lattices and graphene properties \cite{Chen1979, Ho1979, Hirsch1980, Robaszkiewicz1981a, Robaszkiewicz1981, Hirsch1985, Lieb1968, Lieb2003, Hirsch1989, Sorella1992, Pelizzola1993, Janis1993, Staudt2000, Peres2004, Kent2005, Zaleski2008,
Joura2015, Li2015, McKenzie1998, Fuchs2011, Rohringer2011, Yamada2014, Kozik2013, Karchev2013, Claveau2014, Lieb1989, Shastry1986, Su1992, Mancini2009, Tocchio2013, Dang2015, Mermin1966, Nolting2009a, Dombrowsky1996, Feldner2010, Weymann2015, Chao1977, Yosida1998, Tasaki1998, Micnas1990, Georges1996, Hirschmeier2015, Mielke2015}.
Despite numerous theoretical efforts, the rigorous solutions to the Hubbard model for infinite systems have been obtained in very few cases only. The exact results include, for instance, the solution for one-dimensional (1D) system \cite{Lieb1968, Lieb2003, Shastry1986, Su1992} as well as several rigorous theorems, to mention Mermin-Wagner theorem in 2D systems \cite{Mermin1966, Nolting2009a, Tasaki1998} or Lieb theorems for the ground state \cite{Lieb1989}.
At the same time, it has been noticed that the exact solutions to the model can be obtained for small clusters, consisting of several lattice sites \cite{Schumann2008,Schumann2007, Cisarova2014, Cencarikova2016, Galisova2015, Galisova2015a, Harris1967, Silantev2015, Hasegawa2005, Hasegawa2011, Spalek1979, Longhi2011, Juliano2016, Kozlov1996, Alvarez-Fernandez2002, Avella2003, Szalowski2015, Szalowski2017, Balcerzak2017, Balcerzak2018,Wortis2011,Wortis2017, Cheng1976, Ohta1994, Noce1997, Iglesias1997,Amendola2015, Becca2000, Ricardo-Chavez2001, Yang2009, Ovchinnikova2013, Carrascal2015, Fuks2014, Kamil2016, Souza2016, Balasubramanian2017}. Intensive investigations of such systems have been carried out both from the point of view of static properties
\cite{Cheng1976, Spalek1979, Ohta1994, Noce1997, Amendola2015,Becca2000, Ricardo-Chavez2001, Alvarez-Fernandez2002, Schumann2007, Schumann2008, Yang2009, Schumann2010, Ovchinnikova2013, Cisarova2014, Hancock2002,Hancock2005,Hancock2014,Galisova2015, Szalowski2015, Carrascal2015, Kamil2016, Souza2016, Balcerzak2017, Balcerzak2018,Ullrich2018}, as well as for dynamical description \cite{Harris1967, Kozlov1996, Longhi2011, Fuks2014, Wortis2017, Balasubramanian2017}.
In case of very small atomic clusters, exact results for the Hubbard model have been obtained by analytical methods \cite{Harris1967, Cheng1976, Noce1997,Iglesias1997, Amendola2015,Alvarez-Fernandez2002, Schumann2007, Schumann2008, Schumann2010, Ovchinnikova2013, Cisarova2014, Galisova2015, Carrascal2015, Kamil2016, Balcerzak2017, Balcerzak2018, Balasubramanian2017}. However, for larger clusters the numerical techniques turned out to be indispensable \cite{Spalek1979, Ohta1994, Becca2000, Ricardo-Chavez2001, Yang2009, Szalowski2015, Souza2016, Szalowski2017}.
It is worth mentioning that theoretical studies of finite clusters are becoming increasingly important for the development of experimental nanophysics and nanotechnology.
The simplest system, for which the Hubbard model can be solved analytically, is a two-site atomic cluster (dimer). Despite many theoretical works, the system has not been fully examined yet. For instance, this concerns the case when the two-site cluster is simultaneously embedded in two external fields: magnetic and electric one, and is able to exchange the electrons with its environment. Such a system can model a physical situation where the atomic dimer is deposited on the surface and interacts both with the surface and the external fields. The influence of the electric field, acting as a control factor, on the magnetic properties of the cluster constitutes a manifestation of magnetoelectric effect and is very interesting from the point of view of possible application, for instance, in spintronics and/or memory devices. Some examples can be recalled here, mainly to mention the molecular dimer systems. Among them molecular mixed-valence dimers \cite{Soncini2010,Bosch-Serrano2012,Suzuki2014,Palii2014,Dudnik2015,Pansini2018} or $\kappa$-(BEDT-TTF) \cite{Naka2016} focus particular attention and appear highly promising; however, also some non-molecular systems such as dimers on graphene surface \cite{Hu2014} also attract the interest.
The theoretical studies of two-atomic Hubbard cluster, treated as a thermodynamic open system and placed simultaneously in two external fields, have been initiated in the papers \cite{Balcerzak2017} and \cite{Balcerzak2018}. In Ref.~\cite{Balcerzak2017} the main formalism has been presented and thorough investigations of the chemical potential have been carried out. On this basis, in the paper \cite{Balcerzak2018} the studies have been extended to magnetic properties, concentrating mainly on the phase diagrams, cluster magnetization, spin-spin correlation functions and mean hopping energy.
The aim of the present work is a continuation of these studies, basing on the formalism developed in Ref.~\cite{Balcerzak2017}, towards elucidation of interesting interrelations between magnetic and electric properties for the Hubbard dimer exhibiting a non-trivial magnetoelectric behaviour. In particular, the electric polarization of the cluster, as well as the electric susceptibility in the external fields will be studied. Simultaneously, the magnetic polarization and magnetic susceptibility will be analysed. A comparison of the magnetic and electric properties will be done, which seems interesting not only from the purely theoretical point of view for this model. In our opinion, the magnetoelectric correlations existing between the described measurable quantities may be also of practical interest, giving the possibility of controlling the magnetic state of the cluster by the electric potential.
The paper is organized as follows: In the theoretical Section \ref{theory} the model is briefly presented and the basic quantities, important for numerical calculations, are defined. In the successive Section \ref{num} the numerical results are illustrated in figures and discussed.
An extensive comparison of magnetic and electric properties is performed there.
The last Section \ref{conc} is devoted to a brief summary of the results and concluding remarks. The \ref{app} collects the expressions for the eigenenergies corresponding to the quantum states with two electrons per dimer and shows the behaviour of these states as a function of the electric and magnetic field.
\section{\label{theory}Theoretical model}
The Hamiltonian of the Hubbard pair-cluster (dimer) consisting of $(a,b)$ atoms and interacting with the external fields is assumed in the form:
\begin {eqnarray}
\mathcal{H}_{a,b}&=&-t\sum_{\sigma=\uparrow,\downarrow}\left( c_{a,\sigma}^+c_{b,\sigma}+c_{b,\sigma}^+c_{a,\sigma} \right)+U\left(n_{a,\uparrow}n_{a,\downarrow}+n_{b,\uparrow}n_{b,\downarrow}\right)\nonumber\\
&&-H\left(S_a^z+S_b^z\right) -V\left(n_{a}-n_{b}\right),
\label{eq1}
\end {eqnarray}
where $t>0$ is the hopping integral and $U\ge 0$ is the on-site Coulomb repulsion parameter. The symbol $H=-g\mu_{\rm B}H^z$ stands for an external uniform magnetic field $H^z$ oriented along $z$-direction. The term with $V$ introduces the potential energy of the atoms $a$ and $b$ in the electric field. For such potential distribution the external electric field $E$ is oriented along the pair and is equal to $E=2V/\left(|e|d\right)$ with $d$ being the interatomic distance, whereas $e$ is the electron charge. For the sake of simplicity, we assume that the hopping integral is a constant parameter, independent on the external fields.
In Hamiltonian (\ref{eq1}), $c_{\gamma,\sigma}^+$ and $c_{\gamma,\sigma}$ are the electron creation and annihilation operators, respectively, and $\sigma$ denotes the spin state. The on-site occupation number operators for given spin, $n_{\gamma,\sigma}$, are expressed by $n_{\gamma,\sigma}=c_{\gamma,\sigma}^+c_{\gamma,\sigma}$. The $z$-component of the electron spin on given atom, $S_{\gamma}^z$, is then defined as $S_{\gamma}^z=\left(n_{\gamma,\uparrow}-n_{\gamma,\downarrow}\right)/2$. In turn, the total occupation number operators $n_{\gamma}$ for site $\gamma = a,b$, are defined as a sum of occupation operators for given spin, $n_{\gamma}=n_{\gamma,\uparrow}+n_{\gamma,\downarrow}$.
Beause of treating the pair-cluster as an open electron system within the formalism of grand canonical ensemble, the Hamiltonian should be extended by the chemical potential term, i.e., $\mathcal{H}_{a,b}-\mu \left(n_a+n_b\right)$ is considered, where $\mu$ is the chemical potential. The exact analytical diagonalization of the extended Hamiltonian has been performed in Ref. \cite{Balcerzak2017}. As a result, not only the statistical, but also thermodynamic properties can be calculated exactly. In particular, the grand thermodynamic potential $\Omega_{a,b}$ has been obtained in the form:
\begin {equation}
\Omega_{a,b}=-k_{\rm B}T \ln \mathcal{Z}_{a,b}=-k_{\rm B}T \ln \{ {\rm Tr}_{a,b} \,\exp \lbrack -\beta \left(\mathcal{H}_{a,b}-\mu\left(n_a+n_b\right)\right)\rbrack \},
\label{eq2}
\end {equation}
where $\mathcal{Z}_{a,b}$ is the grand partition function.
The self-consistent calculations of the chemical potential have also been performed in Ref. \cite{Balcerzak2017}. It is worth mentioning here that $\mu$ can be found from the relationship
\begin {equation}
-\left(\frac{\partial \Omega}{\partial \mu}\right)_{T,H,E}=\left(\left<n_a\right>+\left<n_b\right>\right)=2x
\label{eq3}
\end {equation}
where $\left<n_a\right>$ and $\left<n_b\right>$ are the thermodynamic mean values of the total occupation number operators for $\gamma =a,b$ sites, respectively. The parameter $x$, where $0\le x\le 2$ for open system in equilibrium, denotes the mean number of electrons per lattice site, i.e., the electron concentration. The partial derivative in Eq.(\ref{eq3}) is performed at constant temperature $T$ and external fields $H$ and $E$. The statistical averages of the on-site occupation number operators in Eq. (\ref{eq3}) are independently calculated from the formula:
\begin {equation}
\left<n_{\gamma} \right>={\rm Tr}_{a,b} \left[ \left(n_{\gamma,\uparrow}+n_{\gamma,\downarrow}\right)\; \hat \rho_{a,b} \right],
\label{eq4}
\end {equation}
where $\hat \rho_{a,b}$ is the statistical operator for the grand canonical ensemble given by:
\begin {equation}
\hat \rho_{a,b}=\frac{1}{\mathcal{Z}_{a,b}}\, \exp \lbrack -\beta \left(\mathcal{H}_{a,b}-\mu\left(n_a+n_b\right)\right)\rbrack.
\label{eq5}
\end {equation}
By the same token, the on-site magnetization, $m_{\gamma}$, can be calculated as the statistical average of $z$-component of the spins, namely $m_{\gamma}=\left<S_{\gamma}^z\right>$, where
\begin {equation}
\left<S_{\gamma}^z \right>={\rm Tr}_{a,b} \left[ \frac{1}{2}\left(n_{\gamma,\uparrow}-n_{\gamma,\downarrow}\right)\; \hat \rho_{a,b} \right].
\label{eq6}
\end {equation}
Having calculated these averages, the mean magnetic polarization per one atom, $m$, is defined as:
\begin {equation}
m=\frac{1}{2}\left(m_{a}+m_{b}\right),
\label{eq7}
\end {equation}
from which the magnetic susceptibility $\chi_{H}$ can be directly obtained:
\begin {equation}
\chi_{H}=\left(\frac{\partial m}{\partial H}\right)_{T,E}.
\label{eq8}
\end {equation}
On the other hand, the electric field $E$ induces the dipolar electric moment on the Hubbard dimer. The absolute value of the electric polarization, $P$, which is proportional to the mean charge displacement and interatomic distance $d$, can be found from the formula:
\begin {equation}
P=d \;|e|\;|\left<n_a\right>-x|.
\label{eq9}
\end {equation}
On this basis, the electric response function, i.e., electric susceptibility
$\chi_{E}$, can be found:
\begin {equation}
\chi_{E}=\left(\frac{\partial P}{\partial E}\right)_{T,H}.
\label{eq10}
\end {equation}
The above formalism will be employed as a basis for the numerical calculations discussed in the next Section \ref{num}.\\
\section{\label{num}Numerical results and discussion}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\textwidth]{f1}
\vspace{2mm}
\caption{\label{Fig1} The electric polarization, $P$, plotted in dimensionless units $P/\left( d |e|\right)$ as a function of the potential difference $E|e|d/t$, for $U/t=1$ and $x=1$. Different curves correspond to various dimensionless temperatures $k_{\rm B}T/t$ and magnetic fields $H/t$.}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\textwidth]{f2}
\vspace{2mm}
\caption{\label{Fig2} The electric susceptibility, $\chi_E$, plotted in dimensionless units $\left(t\chi_E\right)/\left( d^2 |e|^2\right)$ as a function of the potential difference $E|e|d/t$, for $U/t=1$ and $x=1$. Different curves correspond to various dimensionless temperatures $k_{\rm B}T/t$ and magnetic fields $H/t$.}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\textwidth]{f3}
\vspace{2mm}
\caption{\label{Fig3} The magnetic polarization (dimensionless magnetization) per atom, $m$, and magnetic susceptibility, $\chi_H$, plotted in dimensionless units $t\chi_H$ vs. potential difference $E|e|d/t$, for $U/t=1$ and $x=1$. Different curves correspond to various dimensionless temperatures $k_{\rm B}T/t$ and magnetic fields $H/t$.}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\textwidth]{f4}
\vspace{2mm}
\caption{\label{Fig4} The electric polarization, $P$, plotted in dimensionless units $P/\left( d |e|\right)$ as a function of the dimensionless temperature $k_{\rm B}T/t$, for $U/t=1$, $E|e|d/t=2$ and $x=1$. Different curves correspond to various external magnetic fields $H/t$.}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\textwidth]{f5}
\vspace{2mm}
\caption{\label{Fig5} The electric susceptibility, $\chi_E$, plotted in dimensionless units $\left(t\chi_E\right)/\left( d^2 |e|^2\right)$ as a function of the dimensionless temperature $k_{\rm B}T/t$, for $U/t=1$, $E|e|d/t=2$ and $x=1$. Different curves correspond to various external magnetic fields $H/t$.}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\textwidth]{f6}
\vspace{2mm}
\caption{\label{Fig6} The dimensionless magnetization per atom, $m$, as a function of the dimensionless temperature $k_{\rm B}T/t$, for $U/t=1$, $E|e|d/t=2$ and $x=1$. Different curves correspond to various external magnetic fields $H/t$.}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\textwidth]{f7}
\vspace{2mm}
\caption{\label{Fig7} The dimensionless magnetic susceptibility, $t\chi_H$, as a function of the dimensionless temperature $k_{\rm B}T/t$, for $U/t=1$, $E|e|d/t=2$ and $x=1$. Different curves correspond to various external magnetic fields $H/t$.}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\textwidth]{f8}
\vspace{2mm}
\caption{\label{Fig8} The dimensionless electric polarization, $P/\left( d |e|\right)$, vs. magnetic field $H/t$, for $U/t=1$, $E|e|d/t=3$ and $x=1$. Different curves correspond to several selected temperatures $k_{\rm B}T/t$.}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\textwidth]{f9}
\vspace{2mm}
\caption{\label{Fig9} The dimensionless electric susceptibility, $\left(t\chi_E\right)/\left( d^2 |e|^2\right)$, vs. magnetic field $H/t$, for $U/t=1$, $E|e|d/t=3$ and $x=1$. Different curves correspond to several selected temperatures $k_{\rm B}T/t$.}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\textwidth]{f10}
\vspace{2mm}
\caption{\label{Fig10} The dimensionless magnetization per atom, $m$, vs. magnetic field $H/t$, for $U/t=1$, $E|e|d/t=3$ and $x=1$. Different curves correspond to several selected temperatures $k_{\rm B}T/t$.}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\textwidth]{f11}
\vspace{2mm}
\caption{\label{Fig11} The dimensionless magnetic susceptibility, $t\chi_H$, vs. magnetic field $H/t$, for $U/t=1$, $E|e|d/t=3$ and $x=1$. Different curves correspond to several selected temperatures $k_{\rm B}T/t$.}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\textwidth]{f12}
\vspace{2mm}
\caption{\label{Fig12} The dimensionless electric polarization, $P/\left( d |e|\right)$, as a function of the potential difference $E|e|d/t$, for $H/t=2$, $k_{\rm B}T/t=0.2$ and $x=1$. Different curves correspond to various Coulomb repulsion $U/t$-parameter.}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\textwidth]{f13}
\vspace{2mm}
\caption{\label{Fig13} The dimensionless electric susceptibility, $\left(t\chi_E\right)/\left( d^2 |e|^2\right)$, as a function of the potential difference $E|e|d/t$, for $H/t=2$, $k_{\rm B}T/t=0.2$ and $x=1$. Different curves correspond to various Coulomb repulsion $U/t$-parameters.}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\textwidth]{f14}
\vspace{2mm}
\caption{\label{Fig14} The dimensionless magnetization per atom, $m$, as a function of the potential difference $E|e|d/t$, for $H/t=2$, $k_{\rm B}T/t=0.2$ and $x=1$. Different curves correspond to various Coulomb repulsion $U/t$-parameters.}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\textwidth]{f15}
\vspace{2mm}
\caption{\label{Fig15} The dimensionless magnetic susceptibility, $t\chi_H$, as a function of the potential difference $E|e|d/t$, for $H/t=2$, $k_{\rm B}T/t=0.2$ and $x=1$. Different curves correspond to various Coulomb repulsion $U/t$-parameters.}
\end{center}
\end{figure}
In this Section, the results of the rigorous calculations of the electric and magnetic polarization as well as the electric and magnetic susceptibilities of the Hubbard pair embedded in the external magnetic and electric fields are presented. The calculations are restricted to the most interesting case when the orbital states of Hubbard dimer are half-filled ($x=1$).
The behaviour of both order parameters (electric polarization $P$ and magnetization $m$) as well as the behaviour of the response functions (electric and magnetic susceptibility) stems from the behaviour of the energy states of the dimer. Therefore, some microscopic insight into the physics of the studied model regarding the effect of the external fields would be valuable. For this purpose we provide the \ref{app}, which presents the energy states corresponding to half-filling of the dimer and discusses their behaviour as a function of the external electric and magnetic field. It serves as a reference point for the ground state discussion, as for $T\to 0$ only the states with two electrons per dimer are important, because the charge density fluctuations vanish.
In Fig.~\ref{Fig1} the dimensionless electric polarization $P/\left( d |e|\right)$ is plotted vs. difference of the electric potential energies $E|e|d/t$, for the electron concentration $x=1$ and Coulomb parameter $U/t=1$. Different curves correspond to two reduced temperatures: $k_{\rm B}T/t=0.001$ (i.e. system close to the ground state) and $k_{\rm B}T/t=1$, as well as two values of the magnetic fields: $H/t=0$ and $H/t=4$. Thus, the curves present the polarization process from the initial state, when the charge of two electrons is distributed equally among two atomic sites ($a$ and $b$), up to the final state, when both electrons are localized on $a$-atom (then the site $b$ is empty) and saturation of the electric polarization takes place. In particular, the curve for $k_{\rm B}T/t$=0.001 and $H/t$=4 shows a discontinuous change of electric polarization at very low temperatures when the electric field $E = 2V/\left(|e|d\right)$ exceeds some critical value. This critical value amounts to $E_c|e|d/t=4.4721$ and corresponds to the change of the ground state from a triplet one to the singlet one, as demonstrated in Fig.~\ref{Fig17} and discussed in \ref{app}. This interesting property is correlated with the magnetic polarization behaviour, as it will be explained in the discussion of Fig.~\ref{Fig3}. Comparing curves for $H/t$ = 0.0 and 4.0, plotted for the same temperature, we see that the magnetic field suppresses electric polarization. This is a kind of "spin blockade", which arises when two spins are parallel aligned in the magnetic field and they are residing on different atoms. Then, the shift of both electrons by the electric field to the same atom requires additionally the spin reversal of one electron.
The corresponding response function, electric susceptibility $\chi_E= \left(\partial P/ \partial E\right)_{T,H}$, is presented in Fig.~\ref{Fig2} in dimensionless units vs. the electric field $E|e|d/t$. The rest of parameters are the same as in Fig.~\ref{Fig1}. Speaking about the low-temperature behaviour ($k_{\rm B}T/t$=0.001), it is seen that for the magnetic field $H/t=4$ the electric susceptibility jumps from zero value up to the value corresponding to absence of the magnetic field, and this jump takes place at the same critical electric field as seen in Fig.~\ref{Fig1} and discussed above. Thus, the behaviour of electric susceptibility is closely correlated with the behaviour of electric polarization from Fig.~\ref{Fig1}. Again, comparing the curves plotted in Fig.~\ref{Fig2} for $H/t$=0.0 and 4.0, for the same temperature $k_{\rm B}T/t=1$, one can note that the external magnetic field enforces some smooth maximum of the electric susceptibility. This maximum corresponds to the point of inflection of the curve for the same parameters ($k_{\rm B} T/t=1.0$ and $H/t$=4.0) in Fig.~\ref{Fig1}. The maximum occurring on the curve for $k_{\rm B}T/t$=1.0 and $H/t$=4.0 can also be treated as a high-temperature vestige of the low-temperature sharp maximum at the discontinuous transition seen on curve for the same $H$ and $k_{\rm B}T/t$=0.001. For very large electric fields $E\propto V$, when the saturation of electric polarization takes place, the electric susceptibility approaches the zero value.
In Fig.~\ref{Fig3} the magnetic polarization (i.e., averge magnetization) per one atom, $m=\left(\left<S_a^z\right>+\left<S_b^z\right>\right)/2$, as well as the magnetic susceptibility, $\chi_H= \left(\partial m/ \partial H\right)_{T,E}$, are simultaneously presented as the functions of $E|e|d/t$. The parameters $U/t=1$ and $x=1$ are the same as in Figs.~\ref{Fig1} and \ref{Fig2}. The curve plotted in this figure for $k_{\rm B}T/t=0.001$ and $H/t=4$ shows the first order magnetic transition caused by the electric field (and corresponds to the curves for the same parameters in Figs.~\ref{Fig1} and ~\ref{Fig2} also exhibiting the discontinuous behaviour). For low values of $E|e|d/t$, the state of the system is a triplet one, as discussed in the \ref{app}, with the spins of both electrons aligned along the magnetic field $H$ if its magnitude exceeds the critical value (see discussion of this effect in our previous paper \cite{Balcerzak2018}). The electrons are then localized on $a$ and $b$ atoms and the electric polarization is zero (see Fig.~\ref{Fig1}). When the critical electric field is reached in low temperatures, both electrons are forced to localize on the same atom, which results in rapid increase of polarization (see Fig.~\ref{Fig1}), but at the same time the spin of one electron must be reversed and the magnetization discontinuously drops to zero. On the other hand, for curve plotted near the ground state ($k_{\rm B}T/t$=0.001) and for $H/t$=0, the lack of magnetic polarization is observed in the whole range of electric field $E$. Since in this case the magnetic field $H$ is absent, thus the state of the system is a singlet and the electrons with opposite spins can freely occupy both atoms with the same probability. Increasing the temperature will not change this nonmagnetic state. For instance, the magnetic susceptibility, illustrated in the case of $k_{\rm B}T/t=1$ and $H/t=0$ shows typical paramagnetic behaviour. However, when the strong magnetic field $H/t=4$ is applied at high temperatures, for instance, at $k_{\rm B}T/t=1$, the behaviours of magnetization and magnetic susceptibility are very different from those at the ground state. Namely, magnetic polarization decreases as a function of the field $E$, whereas magnetic susceptibility reveals a broad maximum, whose position is correlated with the magnetic transition observed in the ground state. We note that diminishing of magnetic polarization vs. electric field, as seen in curve for $k_{\rm B}T/t$=1.0 and $H/t$=4.0, is also correlated with the increase of electric polarization illustrated in Fig.~\ref{Fig1}. The continuous (and smooth) curves, observed for high temperatures, are connected with the statistical averaging of occupancy of all electronic states. The electric field-dependence of the magnetization demonstrated in Fig.~\ref{Fig3} is a clear example of magnetoelectric effect, allowing the control of magnetization by electric means.
Figs.~\ref{Fig4}, \ref{Fig5}, \ref{Fig6} and \ref{Fig7} present the electric polarization, electric susceptibility, magnetization and magnetic susceptibility, respectively. The plots show the temperature dependence of the mentioned quantities and are prepared for the same remaining parameters: $x=1$, $U/t=1$ and $E|e|d/t=2$, as well as for the magnetic fields: $H/t=1$, $H/t=2.1$, $H/t=2.2$ and $H/t=3$. For the above set of parameters the critical magnetic field has been found as $H_c/t=2.1411$ (see the discussion in the \ref{app} concenrning the critical field for the transition between the singlet and triplet state). Thus, two of the curves fall into the range below the critical field (i.e. correspond to the singlet at the ground state) and two other are shown for $H>H_c$ (and correspond to a triplet at the ground state). It can be mentioned here that, as discussed in the paper \cite{Balcerzak2018}, the critical magnetic field $H_c$ is the field above which the system presents ferromagnetic ordering in the ground state, whereas for $H<H_c$ the paramagnetic (singlet) ground state exists.
Analysing the Fig.~\ref{Fig4}, which shows the dimensionless polarization $P/\left( d |e|\right)$, we see that below $H_c$ (curves for $H/t$ =1.0 and 2.1) the electric polarization in the field $E$ reaches a maximum for $T=0$, whereas above the critical magnetic field $H_c$ (curves plotted for $H/t$ = 2.2 and 3.0) it takes the zero value at the ground state. The jump of electric polarization, when $H_c$ is crossed, is evidently connected with the change of spin state and jump of the magnetization $m$, as seen in Fig.~\ref{Fig6}. When temperature increases, all the curves in Fig.~\ref{Fig4} converge and the influence of magnetic field becomes negligible. The same can be said about the magnetization curves in Fig.~\ref{Fig6}, where for $T\to \infty$ the magnetic polarization tends to zero, however, the convergence in that case is much slower. The fact that both electric and magnetic polarizations are very slowly tending to zero when $T$ is large, will have consequences to the existence of non-vanishing corresponding susceptibilities.
The behaviour of dimensionless electric susceptibility, $\chi_E$, vs. temperature, is illustrated in Fig.~\ref{Fig5}. First of all, in the ground state the electric susceptibility is different in two ranges: $H<H_c$ and $H>H_c$. When temperature slightly increases, narrow maxima appear for the curves plotted in the vicinity of the critical magnetic field (i.e. for $H/t$=2.1 and 2.2). During further increase in temperature all the curves become mutually convergent, independently on the magnetic field strength. It can be noted that the jump of the curves in Figs.~\ref{Fig4} and \ref{Fig5} from non-zero value (for $H<H_{c}$) to zero value (for $H>H_{c}$) when $T \to 0$ means that both the electric polarization and the electric susceptibility are vanishing in the ferromagnetic (triplet) ground state. This reflects the fact that two electrons with the same spin cannot be localized at the same atom. At the ground state there is no contribution from other states than these with two electrons per dimer, as charge density fluctuations vanish.
Regarding Fig.~\ref{Fig6}, in which the mean magnetic polarization $m=\left(\left<S_a^z\right>+\left<S_b^z\right>\right)/2$ is shown, it is worth noticing that the curves plotted for $H<H_{c}$ (i.e. for $H/t$=1.0 and 2.1) show an anomalous behaviour vs. temperature in a form of a broad maximum. It can also be noted that Fig.~\ref{Fig6} is qualitatively similar to Fig.~\ref{Fig4}, however, the curves with the same $H$-parameters are arranged in the inverse order. It should be mentioned that similar anomalous behaviour of the magnetization vs. temperature has been found in Ref. \cite{Ricardo-Chavez2001} (Fig.2 in \cite{Ricardo-Chavez2001} for $U/t$=16) in case of 6-site cluster, but without external fields.
The magnetic susceptibility, $\chi_H$, for the system embedded in the electric field $E$, is plotted vs. temperature in Fig.~\ref{Fig7}. Also in this case the pronounced maxima for the curves near the critical magnetic field $H_c$ ($H/t$ = 2.1 and 2.2) can be observed in the low-temperature region. The positions of these maxima are correlated with the most rapid changes of the magnetization presented in Fig.~\ref{Fig6} which manifest themselves for the magnetic field close to the critical field for the transition from singlet to triplet state. Similarly to Fig.~\ref{Fig5}, the magnetic susceptibility curves presented here show also the mutual convergence when temperature increases. Moreover, it should be noted that in the ground state the magnetic susceptibility always goes to zero, irrespective of the magnetic field.
In Figs.~\ref{Fig8}-\ref{Fig11} the same quantities are presented as in Figs.~\ref{Fig4}-\ref{Fig7}, but now the dependencies are shown vs. external magnetic field $H/t$, whereas the electric field is constant. The remaining parameters are: $x=1$, $U/t=1$ and $E|e|d/t=3$. For this set of parameters the critical magnetic field amounts to $H_c/t=2.7986$ (for discussion see \ref{app}). Different curves in Figs.~\ref{Fig8}-\ref{Fig11} correspond to four selected temperatures: $k_{\rm B}T/t$=0.1, 0.2, 0.5, and 1.0.
In Fig.~\ref{Fig8} the electric polarization, $P/\left( d |e|\right)$, is shown as a function of the magnetic field, $H/t$, for the temperatures specified above. The electric polarization diminishes with the increase of the magnetic field, and the most rapid changes occur in the vicinity of the critical field $H_c$. This fact confirms our previous observation that the change of magnetic ordering inevitably influences the electric polarization so that the system shows clear magnetoelectric properties. When the temperature increases, the curves flatten, showing a decrease of electric polarization for low magnetic fields, however, in the region of strong fields and $H>H_c$ some increase of the electric polarization with increasing temperature can be observed.
The behaviour of electric susceptibility, $\chi_E$, vs. $H/t$ can be analysed on the basis of Fig.~\ref{Fig9}. For low temperatures ($k_{\rm B}T/t$=0.1 and 0.2) a strong peak appears at the critical magnetic field $H_c$. For higher temperatures the peak diminishes and becomes increasingly diffused. It can be noted that for $H=0$ the electric susceptibility takes a non-zero value and hardly depends on the temperature. This fact is connected with a strong electric polarization in this region, as shown in Fig.~\ref{Fig8}. On the other hand, for $H/t \to \infty$ the electric polarization tends to zero value, since the system approaches the magnetic saturation state, in which the electrons are spatially separated and localized on different atoms.
The above discussion concerning electric properties can be extended by the analysis of magnetization and magnetic susceptibility. In Fig.~\ref{Fig10} the mean magnetic polarization, $m=\left(\left<S_a^z\right>+\left<S_b^z\right>\right)/2$, is shown as a function of $H/t$. An increasing $m$ as a function of $H$ describes the magnetization process at different temperatures. It is seen that the most rapid changes in magnetization occur near the critical magnetic field $H_c/t=2.7986$. Moreover, the changes are mostly evident for low temperatures ($k_{\rm B}T/t$=0.1), whereas for high temperatures ($k_{\rm B}T/t$=1.0) the dependence becomes weaker and tends to be linear in the presented range of magnetic fields. In the range where $H<H_c$ an increase of the magnetization with temperature can be predicted, which supports the anomalous behaviour discussed previously and shown in Fig.~\ref{Fig6}. On the other hand, for $H>H_c$ the magnetic polarization tends to saturation value, $m=0.5$, although an increasing temperature makes this process slower.
The behaviour of magnetic susceptibility, $\chi_H$, vs. field $H/t$ is shown in Fig.~\ref{Fig11} and it resembles, to some extent, the electric susceptibility from Fig.~\ref{Fig9}. Again, the most pronounced maxima occur near the critical magnetic field, $H_c$, and in the range of low temperatures. For $H=0$ the magnetic susceptibility approaches zero value when $T\to 0$. On the other hand, for $H\to \infty$, the magnetic susceptibility tends to zero at any finite temperature, since the system reaches the magnetically saturated (triplet) state.
In up to now presentation of the results we have assumed that the Coulomb repulsion parameter is equal to $U/t=1$. Now we will study how the variation of this parameter influences the discussed properties. In Figs.~\ref{Fig12}-\ref{Fig15} the same quantities as in previous figures are plotted vs. external electric field. The constant parameters are: $x=1$, $H/t=2$ and $k_{\rm B}T/t=0.2$. Different curves in these figures correspond to 6 values of the Coulomb parameter: $U/t$ = 0.01, 1, 2, 3, 5 and 8.
Starting from the electric polarization, $P$, which is presented in Fig.~\ref{Fig12} in dimensionless units as a function of $E|e|d/t$, we see that with an increase in $U$-parameter the polarization curves become markedly shifted towards larger electric fields. However, the characteristic shape of these curves remains the same. Namely, for $E|e|d/t=0$ all the curves start from zero value, then, for $E|e|d/t>0$ they are increasing functions of the electric field, and eventually all of them tend to saturation polarization whereas $E|e|d/t\to \infty$. It is worth noticing that similar behaviour of the ground-state occupation difference vs. $E$, plotted for various $U$-parameters, has been found in Ref. \cite{Carrascal2015} (Fig.~4 in \cite{Carrascal2015}), but without magnetic field. The most rapid changes of polarization in Fig.~\ref{Fig12} occur for the electric fields $E$ corresponding to the critical field for singlet-triplet transition reached for the given field $H$ (see the discussion in the \ref{app}). It should be mentioned that the dependence of $H_c$ on the electric field for various $U$-parameters has been discussed in our previous paper \cite{Balcerzak2018}. It has been shown there that, for increasing $U$-parameter and constant $H=H_c$, the electric field corresponding to the transition is always shifted towards larger values. This phenomenon, resulting from the dependence of the eigenenergies of singlet and triplet states on the external fields, is nicely confirmed in the present figure. From Fig.~\ref{Fig12} we can conclude that the role of $U$-parameter consists in countering the electric polarization, as it can be expected from the model Hamiltonian, since this parameter prevents the charge localization on a single atomic site.
In Fig.~\ref{Fig13} the electric susceptibility, $\chi_E$, is plotted in dimensionless units for the same parameters as in Fig.~\ref{Fig12}. The pronounced peaks of the susceptibility can be seen for the electric fields corresponding to the singlet-triplet transitions. At the same time, the positions of these peaks reflect the points of inflection seen on the curves in Fig.~\ref{Fig12}. It can be noted that with increase of $U$-parameter the maxima of $\chi_E$ become higher and sharper. For the electric fields far from the maxima the electric susceptibilities tend to zero value, provided the $U$-parameter is high enough. On the other hand, for the range of small $U$ (like $U/t$= 0.01) and $E|e|d/t\to 0$, the susceptibility is non-zero, because of the dominating role of temperature, which here amounts to $k_{\rm B}T/t=0.2$ (and the energy of the thermal fluctuations exceeds $U$-energy).
The mean magnetic polarization of the Hubbard pair, $m=\left(\left<S_a^z\right>+\left<S_b^z\right>\right)/2$, is presented in Fig.~\ref{Fig14} as a function of the electric field $E$. Similarly to electric polarization, the $U$-parameter has also pronounced influence on the $m$ vs. $E|e|d/t$ curves. For given $U$, magnetization diminishes from its maximal value at $E|e|d/t=0$ towards zero value when $E|e|d/t\to \infty$. The steepest decrease is observed for the same values of $E|e|d/t$ for which the electric polarization showed the steepest increase (in Fig.~\ref{Fig12}). When $U$-parameter is strong enough, the magnetization for $E|e|d/t\to 0$ is in saturated state. However, for small $U$, for example $U/t$=0.01, the magnetization cannot reach the saturation, because it is disordered by the thermal fluctuations. In general, an increase in $U$-parameter extends the range of $E|e|d/t$ in which the magnetic polarization remains in saturated state and only weakly depends on the electric field.
Finally, in the Fig.~\ref{Fig15} the magnetic susceptibility, $\chi_H$, is presented for the same parameters as in Figs.~\ref{Fig12}-\ref{Fig14}. For most of the curves, when $U$-parameters are strong enough, the distinct maxima are shown, whose character is similar to those seen in the electric susceptibility curves (Fig.~\ref{Fig13}). As discussed previously, these maxima can be attributed to the singlet-triplet transitions in the ground state, and far from these regions the magnetic susceptibility practically does not depend on the electric fields and eventually vanishes. The vanishing of $\chi_H$ occurs both in the magnetic saturation region (for $E|e|d/t\to 0$) and in the electric polarization saturation region (for $E|e|d/t\to \infty$). However, when $U$ parameter is small, for instance, for $U/t$ = 0.01 or 1, the magnetic saturation is not reached for $E|e|d/t\to 0$ at present temperature ($k_{\rm B}T/t=0.2$) and magnetic field ($H/t=2$). This fact results in a non-zero value of magnetic susceptibility for $E|e|d/t\to 0$. Moreover, in this region the positions of the magnetic susceptibility maxima, seen on curves plotted for $U/t$ = 0.01 or 1, are not coincident with the maxima of electric susceptibility (seen in Fig.~\ref{Fig13}). Thus, one can conclude that for the temperatures $T$ far from the ground state and small $U$-parameters, i.e., when the energy of the thermal fluctuations dominates, a maximum of the magnetic susceptibility is shifted with respect to the corresponding maximum of the electric susceptibility. It should be emphasized that at finite temperatures also the states higher in energy than the ground state contribute to the behaviour of the system and no sharp transition between singlet and triplet state is seen when the fields are varied. Therefore, the maxima of both susceptibilities need not to coincide exactly.
\section{\label{conc}Summary and conclusion}
In the paper the electric polarization and electric susceptibility, as well as the magnetic polarization and magnetic susceptibility have been studied for the Hubbard pair-cluster (dimer) embedded simultaneously in the electric and magnetic fields. The electron concentration has been selected at the half-filling level, i.e., with one electron per atom, since in this case the magnetic properties are most sensitive to the external fields \cite{Balcerzak2018}.
The analytical method developed in \cite{Balcerzak2017} was utilized, which enabled exact calculations for the model Hamiltonian. The present paper is a continuation of our previous work \cite{Balcerzak2018} which has been focused solely on the studies of the magnetic properties. In this work we extend the studies to include electric polarization, as well as the electric and magnetic response functions, i.e., susceptibilities. A special attention has been drawn to search of mutual correlations between the electric and magnetic properties (manifestations of the magnetoelectric effect) for the quantities mentioned above.
It has been found that the electric and magnetic polarizations, as well as the electric and magnetic susceptibilities, are strictly interrelated. For instance, an increase in the electric polarization is accompanied by corresponding decrease in total magnetization. This sort of behaviour can be generally traced back to the behaviour of singlet and triplet quantum states, which are the only states involved at the ground state (see \ref{app}). Namely, the nonmagnetic singlet state, in which the spin average vaue is zero at both sites, allows the charge redistribution under the action of the electric field and emergence of nonzero eletric polarization. On the contrary, for triplet state which exhibits ferromagnetic polarization, both electrons have parallel spins, so that they have to occupy strictly different sites and no redistribution is allowed in the electric field, preventing the electric polarization. It has also been found that the discontinuous transition (traced back to the transition between singlet and triplet state) can be registered not only by the magnetic quantities, but also as the jumps of electric polarization or electric susceptibility. The phenomena mentioned above have been widely studied in various external fields $H$ and $E$, as well as vs. temperature $T$ for different Coulomb on-site repulsion parameter $U$. In particular, an anomalous behaviour of the electric and magnetic polarization has been found in some region of model parameters, showing the wide maxima of these quantities as a function of the temperature. The existence of the critical magnetic field $H_c$ enforcing the singlet-triplet transitions has been confirmed, and its dependence on the electric field $E$ and $U$-parameter is in accordance with our previous results \cite{Balcerzak2018}.
Simultaneous application of the electric and magnetic fields on the system turned out to be fruitful, since it revealed a competing character of these fields. We hope that the present simple model, which has been solved exactly, can serve not only as a theoretical toy model, but it will also enable better understanding of the Hubbard model itself, and can elucidate competing interrelations between the electric and magnetic properties in strongly correlated systems. Moreover, demonstrating a clear magnetoelectric effect, it might show high potential for applications.
Further studies of the pair-cluster can be done within the so-called extended Hubbard model, when the Coulomb repulsion between electrons residing on different atomic sites is taken into account. In addition, the magnetic exchange interaction between nearest-neighbour spins can be considered in the presence of an arbitrary electron concentration. Such generalizations of the model in question might significantly extend the range of observed phenomena.\\
|
\section{Introduction}
Since its introduction in the electromagnetic theory, gauge invariance has been a paradigm in physics, and constitutes one of the main properties of successful theories such as the Standard Model of particle interactions \cite{book_Quigg13}.
On the one side, the gauge principle can be used as a guiding principle to define new theories, where the development of the Electroweak interaction theory is just an example.
On the other side, the symmetry predicts the existence of a conserved current, which constitutes a powerful tool in the analysis of dynamical phenomena.
In this paper, we discuss the manifestation of U(1) gauge invariance within the context of a discrete-time quantum walk (DTQW) in a two-dimensional (2D) lattice, which could be generalized to 3D lattices.
The dynamics of such DTQWs is driven by the action of unitary operators that act both on the spatial and internal degrees of freedom \cite{Meyer96a}.
A particular interest in this gauge-invariant dynamical scheme arises from the possibility of describing with it, artificially, i.e.\ by engineering an appropriate spacetime dependence of the walker's phase, the effect of a magnetic field, or even a combination of electric and magnetic fields, on charged matter.
By itself, the magnetic field gives rise to interesting phenomena such as localization or controlled spreading \cite{Yalcinkaya2015} and Landau levels \cite{AD15}.
The magnetic field is also one of the main ingredients of the quantum Hall effect, with associated topological effects \cite{Kitagawa2012,Kitagawa2012b} and edge currents \cite{Verga2017}.
On the other hand, the combination of both a magnetic and an electric field exhibits richer features, like Bloch oscillations and the $\vec{E}\times\vec{B}$ drift \cite{AD16}.
The observation of these effects with discrete-time schemes as we study here may be available in the future using internal-state- dependent transport of atoms in 2D optical lattices \cite{Groh2016, Brakhane16, SAMWA2018}, or of photons in 3D integrated-photonics circuits \cite{ONSY17}. In continuous-time schemes, atoms in optical lattices are also a promising platform \cite{JAKSCH200552, dalibard10a, Bloch2012, Dalibar2015} to observe such effects.
In order to consistently describe these effects with DTQWs, one needs to understand how U(1) gauge invariance can be incorporated within this framework, which differs notably from the electromagnetic theory in the continuum (i.e., in continuous spacetime).
In fact, this is a general (serious) problem in physics, since going from the continuum to a lattice formulation is plagued with difficulties and new features \cite{book_montvay_munster_1994, Munster2000, book_smit_2002}.
Moreover, the way of implementing gauge invariance in lattice models is usually not unique, with different approaches leading to the same limit in the continuum.
Our proposal to achieve U(1) gauge invariance on the lattice exhibits close analogies both with the method used in quantum field theory \cite{Wilson74} and with recent works exhibiting similar but different U(1) lattice gauge invariances, in DTQWs \cite{DMD14, AD16, Arnault17} or in reversible cellular automata \cite{arrighi2018gauge}. We comment on the similarities and differences with these recent works.
This paper is organized as follows.
In Sec.\ \ref{sec:model}, we define a new family of DTQWs on a line, which satisfy a U(1) gauge invariance on the (1+1)D lattice.
The discrete derivatives which intervene in this lattice gauge invariance treat time and space on the same footing, and are very much like those used in standard LGTs, in contrast with those of Refs.\ \cite{Arnault17,AD16}.
This is achieved by applying the gauge-field exponentials either before or after the spatial shift, depending on whether the internal state of the walker is, say, up or down, respectively.
We formally compute the continuum limit of these DTQWs, which concides, as desired and as in Refs.\ \cite{DMD14,Arnault17}, with the dynamics of a Dirac fermion in (1+1)D spacetime, coupled to a U(1), i.e., electro(magnetic) gauge field.
In Sec.\ \ref{sec:model2D}, we extend the previous results to 2D walks, constructed by alternating 1D walks in the $x$ and $y$ directions of the spatial lattice.
The way we ensure the U(1) lattice gauge invariance of this 2D scheme is by requiring it \emph{for each one-dimensional substep}, in contrast with the gauge invariance of Ref.\ \cite{AD16}.
This ensures that time and space are still treated on the same footing at the level of the discrete derivatives, up to the fact that there are now, in 2D, two discrete derivatives in time, one for the even discrete-time coordinates, corresponding to the motion in the, say, $x$ direction and another one for the odd ones, corresponding to the motion in the $y$ direction.
In Sec.\ \ref{sec:current}, finally, we derive analytically a lattice continuity equation, stating the conservation of a certain current on the lattice which is computed exactly. We comment on the differences between this continuity equation and that of Ref.\ \cite{AD16}.
\section{A new $\mathrm{U(1)}$ lattice gauge invariance for the DTQW on the line}\label{sec:model}
\subsection{Defining the 1D walk}
The state $\ket{\psi_{j}}$ of the walker at some arbitrary discrete time $j\in{\mathbb{N}}$, belongs to a Hilbert space $\mathcal{H}=\mathcal{H}_{\text{coin}} \otimes \mathcal{H}_{\text{position}}$.
The Hilbert space $\mathcal{H}_{\text{position}}$ describes the external, spatial degree of freedom of the walker, and is spanned by the basis states $\{ \ket{x= p \epsilon} \}_{p \in \mathbb{Z}}$, where $\epsilon$ is the lattice spacing.
The two-dimensional Hilbert space $\mathcal{H}_{\text{coin}}=\text{Span}\{\ket R,\ket L\}$ describes the internal, so-called coin degree of freedom of the walker, where `$R$' and `$L$' stand for `right' and `left'.
The projection of the walker's state on the position state $\ket{x= p \epsilon}$ at time $j$ is $\psi_{j,p} \equiv \langle x = p\epsilon | \psi_j\rangle$.
We identify $\ket R = (1,0)^{\top}$ and $\ket L = (0,1)^{\top}$, where $\top$ denotes matrix transposition.
The dynamics of the DTQW is defined by its one- time-step evolution operator $U_j$, which is unitary and may depend on $j$,
\begin{equation}
\ket{\psi_{j+1}}=U_{j+1}\ket{\psi_{j}} \, .
\label{eq:UQW}
\end{equation}
As usual for DTQWs, the dynamics alternates between (i) rotations, $C$, of the coin degree of freedom, and (ii) spatial coin-state- dependent shifts, $S$:
\begin{equation}
U = S C \, ,
\label{eq:U_op}
\end{equation}
where, to lighten notations, the multiplication of $C$ by the identity tensor factor of the position Hilbert space has been, and will be, in similar cases, omitted.
We choose, for the coin rotation, the following one,
\begin{equation}
C(\theta) = e^{i\sigma^1\frac{\theta}{2}} = \begin{bmatrix}
\cos \frac{\theta}{2} & i \sin \frac{\theta}{2} \vspace{0.1cm} \\ i \sin \frac{\theta}{2} & \cos \frac{\theta}{2} \end{bmatrix} \, ,
\end{equation}
where $\sigma^n$ is the $n$th Pauli matrix, and $\theta$ is some angle, constant in time and uniform in position.
Now, one of the novelties of the present work, is the way we gauge our walk.
In Refs.\ \cite{DMD14,Arnault17,AD16}, gauging the walk amounts to gauge the standard coin-state- dependent shift, $S_{\text{free}} = e^{-i \sigma^3 \mathcal{K}}$, where $\mathcal{K}$ is the quasimomentum operator, as $S_{\text{free}} \rightarrow e^{i\alpha_{j}} S_{\text{free}} \, e^{-i \sigma^3 \xi_{j}}$, where $\alpha_{j,p}$ and $\xi_{j,p}$ are lattice counterparts of the temporal and spatial components of an electric potential of the continuum, $(A^0, A^1)$, with which they coincide in the continuum limit of the DTQW.
We have used the notation $\varphi_j : p \mapsto \varphi_{j,p}$ for diagonal operators in the position basis, such as $\alpha_j$ and $\xi_j$.
In the present work, we gauge the shift as follows: the relative order in which the shift and the gauge-field exponentials are applied, depend on the coin state, that is,
\begin{subequations}
\begin{align}
S({\alpha}_{j},{\xi}_{j}) &= \begin{bmatrix}
e^{-i \mathcal{K}} e^{i( \xi_{j} -\alpha_{j} )} & 0 \\
0 & e^{-i( \xi_{j} + \alpha_{j} )} e^{i \mathcal{K}}
\end{bmatrix} \\
&= T e^{i (\beta_-)_{j}} \Lambda_{R} +e^{-i (\beta_+)_{j}}T^{\dagger} \Lambda_{L} \, ,
\end{align}
\end{subequations}
where $\dag$ denotes Hermitian conjugation.
We have introduced the following objects: (i) the translation operator by one lattice site to the right,
\begin{equation}
T = e^{-i\mathcal{K}} \, ,
\end{equation}
(ii) the two projectors associated to the coin space,
\begin{equation}
\Lambda_s = \ket s \! \! \bra s \, , \ \ \ \ s = R, L \, ,
\end{equation}
and (iii) the difference and sum of $\xi $ and $\alpha$,
\begin{subequations}
\begin{align}
\beta_- &= \xi - \alpha \\
\beta_+ &= \xi + \alpha \, .
\end{align}
\end{subequations}
The non-gauged coin-state- dependent shift is of course $S_{\text{free}} = S(0,0)$.
We have chosen the superscripts $R$ and $L$ for, respectively, the upper and lower components of the wavefunction, because $S_{\text{free}}$ shifts the upper one to the right, and the lower one to the left.
To make notations clear, we introduce an auxiliary notation $\tilde{U}$ for the evolution operator, such that
\begin{subequations}
\begin{align}
U_{j} &\equiv \tilde{U}(\alpha_{j},\xi_{j},\theta) \\
&\equiv S(\alpha_{j},\xi_{j}) \, C(\theta)\, .
\end{align}
\end{subequations}
\subsection{Continuum limit of the 1D walk}
A first fact to mention is that this new way of gauging the walk does not change the continuum limit $\epsilon \rightarrow 0$.
Indeed, the fact that $e^{i\mathcal{K}}$ and $e^{i {f}(\mathcal{P})}$, where $\mathcal{P}$ is the position operator and $f$ an arbitrary function, do not commute, does make an important difference between the gauge procedure of the present work and that of Refs.\ \cite{DMD14,Arnault17,AD16} at the level of the DTQW, i.e.\ for a finite spacetime-lattice spacing. However, this becomes irrelevant in the continuum limit, since the latter is obtained by Taylor expanding all exponentials in their argument, and keeping only the first-order terms: in other words, at first order in their arguments, the exponentials always commute.
Let us now recall this continuum limit $\epsilon \rightarrow 0$.
Assume that, for a given quantity $Q$ defined on the spacetime lattice, $Q_{j,p}$ coincides with the value $Q{(t=j\epsilon,x=p\epsilon)}$ of some continuous function $Q$ of $t$ and $x$.
First, rotate the coin state by a small amount at each time step, that is, set
\begin{equation}
\theta = - 2 \epsilon_{m} m \, ,
\end{equation}
with $\epsilon_m$ going to zero with $\epsilon$, which is the necessary condition for the continuum limit to exist; now, when going to the continuum, we will actually choose $\epsilon_m = \epsilon$, and the parameter $m$ will be identified as the mass of the walker.
Second, consider small gauge fields, that is, set
\begin{subequations}
\begin{align}
\alpha_{j,p} &= \epsilon_A q A^0_{j,p} \\
\xi_{j,p} &= \epsilon_A q A^1_{j,p} \, ,
\end{align}
\end{subequations}
with $\epsilon_A$ going to zero with $\epsilon$, which is also a necessary condition for the continuum limit to exist; again, when going to the continuum, we will actually choose $\epsilon_A = \epsilon$, and the parameter $q$ will be identified as the electromagnetic charge of the walker.
Assuming now that all $Q$'s are twice differentiable in both $t$ and $x$, and Taylor expanding the dynamics of the walker, Eq.\ (\ref{eq:UQW}), at first order in $\epsilon$, delivers (i) zeroth-order terms that, by construction of our walk, cancel each other, which is a necessary condition for the continuum limit to exist, and (ii) first-order terms, which deliver a Hamiltonian equation that can be identified as the Dirac equation in (1+1)D spacetime, with a coupling to a U(1) (and thus Abelian) gauge field.
This equation reads, in manifestly-covariant form,
\begin{equation}
\left(i\gamma^{\mu}_{\text{1D}} D_{\mu}-m\right)\psi=0 \, ,
\end{equation}
with $\mu=0,1$, the covariant derivative $D_{\mu}= \partial_{\mu} + iqA_{\mu}$, where
\begin{equation}
A_{0} =A^0 \, , \ \ \ \ A_{1} =- A^1 \, ,
\end{equation}
are the covariant components of the electric potential, and with the following gamma matrices,
\begin{equation}
\gamma^0_{\text{1D}} = \sigma^1 \, , \ \ \ \ \gamma^1_{\text{1D}} = -i\sigma^2 \, .
\end{equation}
As announced, we obtain, in the limit of small coin-rotation angles and small phases, the same continuum limit as if we had used the gauge procedure of Refs.\ \cite{DMD14,Arnault17,AD16}.
\iffalse
Let us recall that this approach is notably different from that of continuous-time QWs (CTQWs). With DTQWs, one follows a digital approach \cite{Johnson2014, Lloyd1996, Wiese2013, Berry2015, Martinez16, Zohar_2017},
since these are defined in discrete time: one constructs the evolution as a sequence of unitary operators, the gates, each of them corresponding to a desired evolution of the system. A key feature of DTQWs, which makes it particularly simple to construct them as one does, is that each unitary gate is local, i.e.\ the state of the walker, after applying the gate, at position $p$, is determined solely by the state of the walker, before applying the gate, \emph{within a certain bounded spatial neighborhood around $p$}. This implies that the walker evolves within a certain `light cone'. CTQWs, in contrast, are suitable for analog approaches \cite{Johnson2014, JAKSCH200552, Wiese2013, Sarovar2017, Zohar2015},
since they are defined in continuous time: one defines, on a spatial lattice or more generally a graph, a Hamiltonian having local hoppings (this is usually implied by the terminology `CTQWs'). Two reasons for requiring a locality property, either that of CTQWs or that of DTQWs, are simplicity and gain in resources. That being said, that the Hamiltonian has local hoppings does not imply that the corresponding evolution operator will be local in the sense defined above -- actually, this does usually not happen, since, for a lattice Hamiltonian to have a non-trivial dynamics, it must be non- block-diagonal, otherwise there is no interblock dynamics, and a non- block-diagonal Hamiltonian does usually not yield a local evolution operator \cite{Schmitz2016, Korsch2003}. Hence, DTQWs and CTQWs are structurally very different.
Notice that, although there is, in CTQWs, no `light cone' such as that of DTQWs \footnote{That being said, it is shown in Ref.\ \cite{Strauch2007} that the locality of the hoppings in CTQWs can also give rise to relativistic-like effects, i.e. to propagation with bounded group velocity -- that being said, note that the continuous-space limit of the CTQW considered in Ref.\ \cite{Strauch2007} is a \emph{non-relativistic} Schr{\"o}dinger equation. However, this bounding velocity is model-dependent and arbitrary, i.e. not bounded by any universal speed of light -- one would anyways have to precise what one means by `relativistic' in a discrete-spacetime setting, a question considered in Refs.\ \cite{AFF14a, ADMF16, D18}, in the specific framework of DTQWs and quantum cellular automata. The relativistic-like effects of Ref.\ \cite{Strauch2007} should be analyzed under the light \cite{Caruso2016} of the general Lieb-Robinson theorem \cite{Lieb1972}, which also proves the existence of a light cone in the propagation of information in spin Hamiltonians with local hoppings; the Lieb-Robinson velocity is also model-dependent and arbitrary.}, one can still recover, by taking -- whenever possible -- their continuous-space limit, relativistic wave equations \cite{Berry2015, Schmitz2016}, such as the Klein-Gordon \cite{CSO18} or the Dirac equation \cite{APAF18}. This is done by choosing the appropriate lattice Hamiltonian. For other choices of this lattice Hamiltonian, one obtains, in the continuous-space limit, non-relativistic wave equations such as Schr{\"o}dinger's \cite{Strauch06b, Strauch2007}.
\fi
\subsection{A new $\mathrm{U(1)}$ lattice gauge invariance}
Our DTQW, Eq.\ (\ref{eq:UQW}), exhibits a remarkable U(1) lattice gauge invariance: it is invariant under local phase shifts of the form $\psi_{j,p} \rightarrow \psi'_{j,p}= e^{iq\chi_{j,p}} \psi_{j,p}$, where $\chi_{j,p}$ is an arbitrary space- and time-dependent quantity, provided the gauge fields become
\begin{align}
(A_{\mu}')_{j,p} &= (A_{\mu})_{j,p} - (d_{\mu} \chi)_{j,p} \, ,
\end{align}
for $\mu=0,1$, with
\begin{subequations} \label{eq:discrete_derivatives}
\begin{align}
d_0 &= \frac{1}{\epsilon_A} \Delta_0 \Sigma_1 \\
d_1 &= \frac{1}{\epsilon_A} \Delta_1 \Sigma_0 \, ,
\end{align}
\end{subequations}
where the $\Sigma$'s and $\Delta$'s act on sequences $Q_{j,p}$ of time and space as
\begin{subequations}
\begin{align}
(\Sigma_{\mu} Q)_{p_{\mu}} &= Q_{p_{\mu}+1} + Q_{p_{\mu}} \\
(\Delta_{\mu} Q)_{p_{\mu}} &= Q_{p_{\mu}+1} - Q_{p_{\mu}} \, ,
\end{align}
\end{subequations}
having introduced $p_0 \equiv j$ and $p_1 \equiv p$ for a more compact notation.
The discrete derivatives, Eqs.\ (\ref{eq:discrete_derivatives}), treat time and space on the same footing, on the contrary to those of Ref.\ \cite{AD16,Arnault17}. Morever, the $\Sigma$'s and $\Delta$'s defined here are sums and differences over one lattice spacing, or link between two sites, while in Ref.\ \cite{AD16,Arnault17} they were over two links. Notice that the $\Delta$'s are nothing but standard finite differences over one link. The fact that, here, one has to apply the $\Sigma$'s in addition to the $\Delta$'s, underlines that it may be appropriate that the gauge variables, that is, both the gauge fields and the local phase change, be defined on the links rather than on the sites, as in standard LGTs. We leave this matter to future work. Up to these extra $\Sigma$'s, the discrete derivatives involved in Eqs. (\ref{eq:discrete_derivatives}) are the same as those used in standard LGTs, that is, standard finite differences.
As done in Ref.\ \cite{Arnault17} for the 1D case, Ref.\ \cite{AD16} for the 2D case, and Ref.\ \cite{ADMDB16} for the non-Abelian 1D case, one can define a lattice counterpart to the electromagnetic tensor in the continuum,
\begin{equation} \label{eq:electric_tensor}
(F_{\mu\nu})_{j,p} = (d_{\mu} A_{\nu})_{j,p} - (d_{\nu} A_{\mu})_{j,p} \, ,
\end{equation}
which is antisymmetric by construction. Since we are in 1D space, the only non-vanishing components are $(F_{01})_{j,p}=-(F_{10})_{j,p}$, which encode a lattice counterpart to the electric field, and there is no magnetic field. This quantity, $(F_{\mu\nu})_{j,p}$, is, as in the continuum, gauge-invariant by construction (on the spacetime lattice, obviously), since the $d_{\mu}$'s commute with each other.
In the continuum limit, $d_{\mu}$
tends towards the partial derivative $\partial_{\mu}$, the gauge transformation of Eq.\ (\ref{eq:discrete_derivatives}) becomes the standard one of the continuum, and the lattice counterpart to the electro(magnetic) tensor, Eq.\ (\ref{eq:electric_tensor}), becomes that electro(magnetic) tensor.
\section{2D generalization by alternating 1D walks along the $x$ and $y$ directions}\label{sec:model2D}
\subsection{Defining the 2D walk}
The walker can now move on a 2D lattice, and has spatial coordinates $x=p\epsilon$ and $y=q\epsilon$, where $p,q \in \mathbb{Z}$.
We will also use the notation $p=p_1$ and $q=p_2$.
Now, the 1D walk defined in the previous section admits a 2D generalization via a walk which alternates 1D walks in the $x$ and $y$ directions of the 2D lattice. This generalization reads
\begin{subequations} \label{eq:2D_main}
\begin{align}
\ket{\psi_{2l}}&=U^{(1)}_{2l} \ket{\psi_{2l-1}} \label{eq:2D_main_even}\\
\ket{\psi_{2l+1}}&=U^{(2)}_{2l+1} \ket{\psi_{2l}}\, ,
\label{eq:2D_main_odd}
\end{align}
\end{subequations}
with $l\in \mathbb{N}$ and where, for $i=1,2$,
\begin{subequations}
\begin{align}
U^{(i)}_{j} &\equiv \tilde{U}^{(i)}(\tfrac{1}{2} \alpha_{j}, \xi^i_{j},\theta^i) \\
&\equiv S^{(i)}(\tfrac{1}{2} \alpha_{j}, \xi^i_{j}) \, C(\theta^i) \, ,
\end{align}
\end{subequations}
and
\begin{equation}
S^{(i)}(\tfrac{1}{2} \alpha_{j}, \xi^i_{j}) = T_{i} e^{i (\beta_-^{(i)})_{j}} \Lambda_{R} +e^{-i (\beta_+^{(i)})_{j}}T^{\dagger}_{i} \Lambda_{L} \, ,
\end{equation}
where
\begin{equation}
T_i = e^{-i \mathcal{K}_i} \, ,
\end{equation}
$\mathcal{K}_i$ being the quasimomentum operator along direction $i$, and
\begin{subequations}
\begin{align}
(\beta_-^{(i)})_{j,p,q} &= \xi^i_{j,p,q} - \frac{1}{2} \alpha_{j,p,q} \\
(\beta_+^{(i)})_{j,p,q} &= \xi^i_{j,p,q} + \frac{1}{2} \alpha_{j,p,q} \, .
\end{align}
\end{subequations}
When the gauge fields, $\alpha_j$, $\xi^1_j$ and $\xi^2_j$, vanish, the alternate walk is translationally invariant in both time and space every two time steps.
We will thus sometimes use the wording `substep' for the time evolutions $2l-1 \rightarrow 2l$ and $2l \rightarrow 2l+1$, and the wording `step' for $2l-1 \rightarrow 2l+1$.
We also introduce the two-substep walk,
\begin{align}
\ket{\psi_{2l+1}}=U^{\text{2D}}_{2l+1}\ket{\psi_{2l-1}} \, ,
\label{eq:2D_main_2}
\end{align}
where
\begin{align}
U^{\text{2D}}_{2l+1} = U^{(2)}_{2l+1} \, U^{(1)}_{2l} \, .
\end{align}
\subsection{Continuum limit for the 2D walk}
We perform the continuum limit of the two-substep walk.
Adapting the 1D-case procedure, we write
\begin{subequations}
\begin{align}
\alpha_{j,p,q} &= \epsilon_A q A^0_{j,p,q} \\
\xi^{i}_{j,p,q} &= \epsilon_A q A^i_{j,p,q} \, ,
\end{align}
\end{subequations}
for $i=1,2$.
Moreover, we choose
\begin{subequations}
\begin{align}
\theta^1 &= \frac{\pi}{2} - \epsilon_m m \\
\theta^2 &= - \frac{\pi}{2} - \epsilon_m m \, .
\end{align}
\end{subequations}
Assume now that, for a quantity defined on the spacetime lattice, $Q_{j,p,q}$ coincides with the value $Q(t=j\epsilon/2,x=p\epsilon,y=q \epsilon)$ of some continuous function $Q(t,x,y)$. The factor $1/2$ in the time variable is necessary to make the continuum limit of this two-substep DTQW match with the standard form of the Dirac equation.
Taking the continuum limit, $\epsilon\rightarrow 0$, of Eq.\ (\ref{eq:2D_main}), with $\epsilon_A = \epsilon_m = \epsilon$, we obtain
\begin{equation}
\left(i\gamma^{\mu}D_{\mu}-m\right)\psi=0 \, ,
\end{equation}
with
\begin{equation}
\gamma^0 = \sigma^1 \, , \ \ \ \gamma^1 = -i\sigma^3 \, ,
\ \ \ \gamma^2 = - i \sigma^2 \, .
\end{equation}
\subsection{Two-substep $\mathrm{U(1)}$ lattice gauge invariance}
By construction from 1D gauge-invariant walks, the 2D walk we have introduced, Eq.\ (\ref{eq:2D_main}), is invariant under the local phase shift $\psi_{j,p,q} \rightarrow \psi'_{j,p,q} = e^{iq\chi_{j,p,q}} \psi_{j,p,q}$, provided the gauge fields become
\begin{subequations}
\label{eq:2D_gauge_transfo}
\begin{align}
(A_{0}')_{j,p,q} &= \left\{
\arraycolsep=1.4pt\def1.5{1.5}
\begin{array}{ll}
(A_{0})_{j,p,q} - (d_{0}^1 \chi)_{j,p,q} & \ \ \ \text{for} \ j \ \text{even} \\
(A_{0})_{j,p,q} - (d_{0}^2 \chi)_{j,p,q} &\ \ \ \text{for} \ j \ \text{odd}
\end{array}
\right.
\\
(A'_{i})_{j,p,q} &= (A_{i})_{j,p,q} - (d_{i} \chi)_{j,p,q} \, ,
\end{align}
\end{subequations}
for $i=1,2$, with
\begin{subequations} \label{eq:discrete_derivatives_2D}
\begin{align}
d_0^k &= \frac{1}{\epsilon_A} \Delta_0 \Sigma_k \label{eq:discrete_derivatives_2D_time} \\
d_i &= \frac{1}{\epsilon_A} \Delta_i \Sigma_0 \, .
\end{align}
\end{subequations}
A first comment to make is that this 2D U(1) lattice gauge invariance differs from that of Ref.\ \cite{AD16} in the following: we have required, here, the gauge invariance \emph{for each one-dimensional substep}; we thus call this 2D U(1) lattice gauge invariance a two-substep gauge invariance.
In such a two-substep gauge invariance, $A_0$ transforms, by construction, differently at even and odd times: indeed, we have two different difference operators in time, $d^1_0$ for even times, and $d^2_0$ for odd times, which manifests the alternate construction of the walk.
Apart from this, the difference operators of Eq.\ (\ref{eq:discrete_derivatives_2D}) are a straightforward generalization of those used above in the 1D case, Eq.\ (\ref{eq:discrete_derivatives}).
As in the 1D case, the difference operators of Eq.\ (\ref{eq:discrete_derivatives_2D}) treat space and time on the same footing (up to the two discrete derivatives in time), in constrast with Ref.\ \cite{AD16}.
Additionally, in the present 2D case, these difference operators also treat the two directions of the lattice on the same footing, which is also in contrast with Ref.\ \cite{AD16}.
Finally, one can define a lattice counterpart to the electromagnetic tensor, by generalizing the 1D lattice tensor, Eq.\ (\ref{eq:electric_tensor}), to the present 2D setting. Notice that one has to use the discrete temporal derivative $d^i_0$ in the definition of $F_{0i}$, so that $F_{01}$ and $F_{02}$ involve \emph{different} discrete temporal derivatives. This 2D lattice tensor has by construction the same properties of antisymmetry and of U(1) lattice gauge invariance as in the 1D case, and contains, additionally, a `lattice magnetic field' orthogonal to the 2D plane, namely, $(F_{12})_{j,p,q}$, for which there is no room in the 1D case.
In Ref.\ \cite{JWD18}, a three-substep U(1) lattice gauge invariance is suggested for a 2D DTQW on an equilateral triangular lattice, which is likely to be generalizable to the other DTQWs presented in this reference (isocele triangular and honeycomb lattices).
There are two main differences between this work and the present one.
First, the correspondence between the spatial components of the lattice gauge field and those of the continuum is, in Ref.\ \cite{JWD18}, not one to one: three such components are needed on the lattice -- one for each substep --, while, the scheme being in 2D space, only two such components are needed in the continuum, which can be expressed as linear combinations of the three former ones, see Eqs.\ (18) of that reference.
In the present work, in contrast, the spatial components of the lattice gauge field match exactly those of the continuum.
This difference between Ref.\ \cite{JWD18} and the present work reflects the connectivity of the lattice, and calls for an understanding, in \emph{arbitrary} $n$D lattices, $n\in\mathbb{N}$, of the coupling of DTQWs to lattice counterparts of the electric and magnetic fields of the $n$D continuum.
Reference \cite{D18} opens the way to such an understanding.
\iffalse
This is because the lattice difference operators (LDOs) which appear in lattice gauge invariance such as that of Ref.\ \cite{JWD18}, given the geometry of the lattice, one still hopes to generate only two spatial components for a lattice gauge field
This gap between the two works is not a small one, which could simply be bridged by taking linear combinations as mentioned above: it is a structural gap which reflects at the level of the lattice difference operators (LDOs).
One has indeed, in Ref.\ \cite{JWD18}, three spatial LDOs on the lattice, the $\delta_i$'s of Eqs.\ (17) of that reference, while the scheme is 2D. This is because the number of LDOs in this scheme reflects, not the dimensionality of space, as it is often wished for differential operators in , but the geometry of the lattice, and in the coordinates used.
These three LDOs are thus not good candidates to be used independently of the geometry
In the present work we have, in contrast, two spatial LDOs, which, although naturally `produced' by the structure of our walk, are good candidates to be used independently of the structure of the 2D scheme.
Of course, a corrollary of this difference is that in the present work, there is a simple one to one correspondence between the difference operators used on the lattice and the partial derivatives of the continuum, while this is not the case in Ref.\ \cite{JWD18}.
\fi
The second main difference between Ref.\ \cite{JWD18} and the present work is that in the former, the relative order in which one applies the gauge field and the shift is not coin-state dependent, in contrast with ours.
As a consequence, the difference operators appearing in Ref.\ \cite{JWD18} do not treat time and space on the same footing as we do.
More precisely, the temporal difference operator of Ref.\ \cite{JWD18}, see, e.g., the first equation of Eqs.\ (17) of that reference, is the same as that of the earlier work already mentioned previously, namely, Ref.\ \cite{AD16}\footnote{This is of course true up to the fact that only one `spatial symmetrization' $\sigma_i$, $i$ labelling the direction taken by the walker at each sub time step, appears in Ref.\ \cite{JWD18}, instead of the double one, $\Sigma_2\Sigma_1$, of the earlier reference \cite{AD16}, see the first equation of Eqs.\ (9) in that reference. This difference is simply due to the fact that in the earlier reference, the lattice gauge invariance is not multi-substep.}, and we already mentioned above that, in that earlier work, time and space are not treated on the same footing at the level of the difference operators, in contrast with the present work.
Eventually, notice the two following facts.
If one tries to impose to the 2D walk of Ref.\ \cite{AD16} a U(1) lattice gauge invariance for each one-dimensional substep, one needs at least to choose, for the corresponding gauge fields, linear combinations of those introduced in that reference -- for the no-substep gauge invariance --, but at \emph{different} spacetime-lattice sites.
The same thing happens when trying to impose, conversely, a no-substep U(1) lattice gauge invariance to the present 2D walk.
\section{Continuity equation and conserved current for the 2D walk}\label{sec:current}
In this section, we derive a lattice continuity equation from the dynamics of the DTQW, allowing us to introduce a current density which is both conserved and gauge invariant.
In the whole section, we work on the spacetime lattice, and use the notations $t=j\epsilon$, $x=p\epsilon$ and $y=q\epsilon$, already introduced previously.
By construction, the probability density at time $t$ and point $(x,y)$ is
\begin{equation}
J^{0}(t,x,y)=\bra{\psi_{t}}\Lambda_{x,y}\ket{\psi_{t}} \, ,
\end{equation}
where $\Lambda_{x,y}=\Ket{x,y} \! \! \Bra{x,y}$ is the projector on state $\Ket{x,y}$.
Now, another novelty of the present work with respect to Ref.\ \cite{AD16}, apart from the way we gauge our walk, discussed in the previous sections, is that we are going to derive our continuity equation and define the current density over \emph{two} time steps, i.e.\ $2\epsilon$, of Evolution (\ref{eq:2D_main_2}), i.e.\ four integers steps in the discrete-time variable $j$, since $t=j\epsilon/2$, while Ref.\ \cite{AD16} considers a single time step of this evolution to define the current density. As the reader shall see, this -- i.e.\ considering two time steps to derive the continuity equation, instead of a single one -- will lead to the appearance of the standard (symmetric) finite difference as discrete derivatives, both in time and space, while the discrete derivatives of Ref.\ \cite{AD16} are more complicated, in particular the temporal one.
So, from this evolution over two time steps, one can easily derive a formula for the difference $J^{0}(t+\epsilon,x,y)-J^{0}(t-\epsilon,x,y)$, which can be written as
\begin{align} \label{eq:differentesint}
&[\Delta^{\text{sym.}}_{0}J^{0}](t,x,y)= \\
& \ \ \ \ \frac{1}{2}\Bra{\psi_{t}}\left({U_{t+\epsilon}^{\text{2D}}}^{\dagger}\Lambda_{x,y}U_{t+\epsilon}^{\text{2D}}-U_{t-\epsilon}^{\text{2D}}\Lambda_{x,y}{U_{t-\epsilon}^{\text{2D}}}^{\dagger}\right)\Ket{\psi_{t}} \, , \nonumber
\end{align}
where, pay attention, we have used the following notation of the Hermitean conjugate for the backwards evolution, ${U^{\text{2D}}}^{\dagger}_{t-\epsilon} \equiv {U^{(1)}}^{\dagger}_{t-\epsilon/2} \, {U^{(2)}}^{\dagger}_{t}$, and where $\left[ \Delta_{0}^{\text{sym.}}f\right] (t,x,y)\equiv\frac{1}{2}\left[f(t+\epsilon,x,y)-f(t-\epsilon,x,y)\right]$, which defines a symmetric finite difference in time.
We compute this quantity in App.\ \ref{sec:Continuity-equation}, and the result is given by Eq.\ (\ref{eq:continuityeq}).
We can then recast Eq.\ (\ref{eq:continuityeq}), i.e. Eq.\ (\ref{eq:differentesint}), as
\begin{equation} \label{eq:compactcontinuity}
\Delta_{\mu}^{\text{sym.}}J^{\mu}= 0 \, ,
\end{equation}
with implicit sum over $\mu=0,1,2$, and where we have introduced the symmetric finite differences in the $x$ and $y$ directions, $\left[\Delta_{1}^{\text{sym.}}f\right](t,x,y)\equiv\frac{1}{2}\left[f(t,x+\epsilon,y)-f(t,x-\epsilon,y)\right]$ and $\left[\Delta_{2}^{\text{sym.}}f\right](t,x,y)\equiv\frac{1}{2}\left[f(t,x,y+\epsilon)-f(t,x,y-\epsilon)\right]$. Equation (\ref{eq:compactcontinuity}) has the form of a continuity equation on the lattice. $J^{1} =J^{x}$ and $J^{2} = J^{y}$, appearing naturally as the current densities along the $x$ and $y$ directions, respectively, are defined by
\begin{align} \label{eq:Jx}
&J^{x}(t,x,y)=\bra{\psi_{t}} \Lambda_{x}\Big[ e^{i\beta^y_-(t,x,y)}e^{i\beta^y_+(t,x,y-\epsilon)}D_{2}M_{x}^{(1)} \nonumber \\
& \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ +e^{-i\beta^y_+(t,x,y-\epsilon)}e^{-i\beta^y_-(t,x,y)}D_{2}^{\dagger}M_{x}^{(2)} \nonumber \\
& \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ +\Lambda_{y-\epsilon}M_{x}^{(3)} \nonumber \\
& \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ +\Lambda_{y+\epsilon} M_{x}^{(4)}
\Big]\ket{\psi_{t}} \, ,
\end{align}
and
\begin{align} \label{eq:Jy}
&J^{y}(t,x,y)=\bra{\psi_{t}} \Lambda_{y} \Big[ e^{i\beta^x_+(t+\frac{\epsilon}{2},x,y)} e^{i\beta^x_-(t+\frac{\epsilon}{2},x-\epsilon,y)}D_{1}M_{y}^{(1)} \nonumber \\
& \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ +e^{-i\beta^x_-(t+\frac{\epsilon}{2},x-\epsilon,y)}e^{-i\beta^x_+(t+\frac{\epsilon}{2},x,y)}D_{1}^{\dagger}M_{y}^{(2)} \nonumber \\
& \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ +\Lambda_{x+\epsilon}M_{y}^{(3)}\nonumber \\
& \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ + \Lambda_{x-\epsilon}M_{y}^{(4)} \Big]\ket{\psi_{t}} \, ,
\end{align}
where we have used the following notations:
\begin{subequations}
\begin{align}
& \beta^{x}_{\pm} =\beta^{(1)}_{\pm} \, , \ \ \ \ \ \ \beta^{y}_{\pm} = \beta^{(2)}_{\pm} \, , \\
&S_x = S^{(1)} \, , \ \ \ \ \ \ S_y = S^{(2)} \, , \\
&C_x = C(\theta^1) \, , \ \ \ \, C_y = C(\theta^2) \, .
\end{align}
\end{subequations}
The rest of the notations we have introduced are defined in App.\ \ref{sec:Continuity-equation}.
Both the time and space differences are symmetric, which implies that they can be used to approximate true derivatives with a truncation error $O(\epsilon^{3})$, in contrast with the difference schemes over one time step, as that in Ref.\ \cite{AD16}, where the error is $O(\epsilon^{2})$.
There is a price to pay for this at the level of the discrete-spacetime scheme: the current is only defined at times $t$ which are even multiples of the time step $\Delta t$, while the walk is defined at all times -- less importantly, one needs in practice, in order to compute the current dynamics over a given area on a finite-size 2D lattice, more sites on the edges of that area with a two-step current than with a single-step one.
In terms of formal simplicity and connection to standard lattice gauge theories, notable advantages of the present continuity equation, Eq.\ (\ref{eq:compactcontinuity}), with respect to that of Ref.\ \cite{AD16}, is that the difference operators involved in it, i.e the $\Delta^{\text{sym.}}$'s, not only (i) treat all three spacetime coordinates on the same footing (while \emph{all} three are treated differently in Ref.\ \cite{AD16}), but (ii) correspond, in addition, to standard symmetric finite differences, while more complicated operators are used in Ref.\ \cite{AD16}.
As in Ref.\ \cite{AD16}, however, the present difference operators intervening in the continuity equation are still different from those intervening in the gauge invariance.
It is easy to check (i) that the current densities defined above are gauge invariant under the transformations of Eq.\ (\ref{eq:2D_gauge_transfo}), and (ii) that Eq.\ (\ref{eq:compactcontinuity}) ensures the conservation of the total probability, i.e.\ $\sum_{x,y}J^{0}(t,x,y)$ does not change with time.
Eventually, notice the following. On the one hand, one can check that the present 2D DTQW, defined by Eqs. (\ref{eq:2D_main}), satisfies, in addition to the present two-step lattice continuity equation, a single-step one -- obtained by comparing the probability densities between two consecutive instants --, which has the same structure as that of Ref.\ \cite{AD16}, and involves, in particular, the same discrete derivatives -- the corresponding current is gauge invariant under the gauge transformations defined in the present work. On the other hand, one can also check that the 2D DTQW defined in Ref.\ \cite{AD16} satisfies, in addition to the single-step lattice continuity equation presented in that reference, a two-step one, which has the same structure as that of the present work, and involves, in particular, \emph{also symmetric finite differences as discrete derivatives} -- the corresponding current is gauge invariant under the gauge invariance of Ref.\ \cite{AD16}, which, recall, is different from the present one. These two combined results indicate that the `symmetrization' of the discrete derivatives when going from single-step to two-step continuity equations is independent from the way one gauges the walk, and is solely due to the alternate construction of the 2D walk.
\iffalse
This is actually a feature that one may wish to have for one's DTQW. Indeed, one typically wishes that a continuity equation exist, first, for the free, i.e. non-gauged walk, as a reflect of probability conservation (implied by unitarity) at the local level, which exist for usual continuum dynamics, such as the Schrödinger or the Dirac equation. Now, if the `symmetrization' we mention had something to do with the way of gauging the walk, it would mean that the continuity equation of the gauged walk involves discrete derivatives that are different from those appearing in the continuity equation of the non-gauged walk. The interpretation of such a feature could prove difficult since it would differ from usual continuum theories.
\fi
\section{Conclusion}
In this paper we have discussed some of the subtleties related to gauge invariance on discrete-time quantum walks that include the interaction with external, synthetic electromagnetic fields, appearing as additional phases related to those fields. As in standard lattice gauge theories, the way to introduce such interactions is not unique, and can lead to interesting new features. We introduce these additional phases in a way that differs from previous works in the literature. We have first described how this definition works for one-dimensional discrete-time quantum walks. This procedure has the advantage that the discrete derivatives which intervene in this lattice gauge invariance treat time and space on the same footing, and are very much like those used in standard LGTs, in contrast with those of Refs.\ \cite{Arnault17,AD16}.
We extended the above dynamics to 2D lattices, by alternating 1D walks in the $x$ and $y$ directions of the spatial lattice, where we ensure the U(1) lattice gauge invariance of this 2D scheme by requiring it \emph{for each one-dimensional substep}, in contrast with the gauge invariance of Ref.\ \cite{AD16}. Also here, time and space are treated on the same footing at the level of the discrete derivatives -- up to the fact that there are now, in 2D space, two discrete derivatives in time, one for the even discrete-time coordinates, corresponding to the motion in the, say, $x$ direction, and another one for the odd ones, corresponding to the motion in the $y$ direction.
By taking two time steps of the alternate walk, we introduced a density current which is both conserved and gauge invariant.
Both in the 1D and in the 2D cases, we have computed the continuum limit of these DTQWs. They coincide, as desired and as in Refs.\ \cite{DMD14,Arnault17} and \cite{Arnault17,AD16}, respectively, with the dynamics of a Dirac fermion in (1+1)D and (1+2)D spacetime, respectively, coupled to a U(1), i.e., electromagnetic gauge field. We also showed that, in two dimensions, the current conservation reproduces, in the continuum, that corresponding to the Dirac field. The procedure discussed here could be easily extended to the case of 3D lattices.
In our opinion, this work represents a sensible step on the way to quantum simulating the dynamics of a Dirac particle coupled to an external electromagnetic field. In addition to this, the quantum walk, as a dynamical process taking place on a lattice, introduces by itself new interesting phenomena, which are still to be fully explored even in the case of two dimensional lattices, which is the minimum dimensionality allowing for the description of both an electric and a magnetic field.
Let us finally mention that a recent work \cite{CGWW2018} presents a unified framework to understand U(1) gauge invariance in discrete-time quantum walks on lattices and with coin spaces of arbitrary dimensions. This work should at the very least enlighten this field.
\begin{acknowledgments}
P.\ Arnault acknowledges inspiring discussions with Christopher Cedzich, Terry Farrelly and Reinhard F.\ Werner, and is grateful for their one-week hosting in the Quantum Information Group, Institut f\"ur Theoretische Physik, Leibniz Universit\"at Hannover, Germany.
This work has been funded by the ANR-12-BS02-007-01 TARMAC grant,
the STICAmSud project 16STIC05 FoQCoSS, CSIC PIC2017FR6, the Spanish Ministerio
de Econom{\'i}a, Industria y Competitividad, MINECO-FEDER project FPA2017-84543-P,
SEV-2014-0398, and the Generalitat Valenciana grant GVPROMETEOII2014-087.
\end{acknowledgments}
|
\section{Introduction}
Matching two pieces of text is a common pattern in many natural language processing tasks. For example, in the textual entailment task, given a pair of premise and hypothesis sentences, the task is to classify them into one of three labels \{entailment, contradiction, neutral\} \cite{Bowman2015}. In paraphrase detection, a pair of sentences need to be classified according to whether they are paraphrases of each other \cite{Dolan2005AutomaticallyCA}. In the semantic relatedness task, a pair of sentences need to be scored based on how closely related they are semantically. Other problems like question answering can also be reduced to textual matching by scoring each question-answer pair and picking the answer with the highest score \cite{Wang2017BilateralMM}. In this paper we assume the texts are a pair of sentences $S_1$ and $S_2$.
There has been a large amount of work on building machine learning models to solve each of these specific problems or text matching in general. In recent years, neural network models have been able to achieve impressive performance on several benchmark datasets related to these problems. The neural models can be divided into roughly two categories. In the sentence encoder based models, each sentence is encoded into a fixed length distributed representation using a sequence encoder like a BiLSTM \cite{Hochreiter1997} acting on the embeddings of the words in the sentence. The two sentence representations are then composed into a single representation by using heuristic matching features like element-wise difference and element-wise product \cite{Mou2016NaturalLI}, which is then passed through a classification layer. These so-called Siamese architectures are simple, but do not take into account the dependencies between words in the two sentences.
\begin{figure*}[t]
{
\centering
\includegraphics[width=0.9\linewidth]{modelfig.pdf}
\caption{Schematic diagram of REGMAPR. The original sentence encodings are used for textual entailment and paraphrase detection and the exponential function is used for semantic relatedness.}
\label{fig:model}
}
\end{figure*}
The other category of neural models incorporate dependencies between words in the two sentences, typically by using an attention mechanism \cite{rocktaschel2016reasoning}. The contextual representation of each word in $S_1$, obtained from the intermediate states of a BiLSTM for example, is composed with the representations of words in $S_2$ using attention and then compared. This produces a series of representations for words in $S_1$ dependent on the words in $S_2$, which can then be encoded further before being used for classification. Many of the best results in text matching are achieved by architectures that use some form of inter-sentence attention e.g. the ESIM model for textual entailment \cite{Chen2017}, the BiMPM model for paraphrase detection \cite{Wang2017BilateralMM} and DIIN for both \cite{Gong2018}. While more expressive, these models are quite complex with a large number of parameters. The question remains whether such complex models are absolutely necessary to achieve good performance in text matching problems. In fact, recent work in language modeling has shown that properly regularized vanilla LSTM networks can achieve results that are comparable to the state-of-the-art \cite{Melis2018} without the need for more complex architectures.
In this paper, we take a middle path and propose a simple Siamese architecture for text matching problems. Each sentence is encoded using a BiLSTM and the representations are composed by computing the element-wise absolute difference and product. Optionally, we also concatenate the original sentence encodings before passing the vector to the classification layer. As mentioned above, inter-sentence dependence information is crucial for good performance in text matching. To avoid the use of complex attention mechanisms, we augment the embeddings of the words to incorporate inter-sentence information. For each word $t \in S_1$, we add a \emph{matching feature} that indicates whether $t$ appears in $S_2$ to the embedding of $t$. Similarly, for each word $t \in S_2$, we add a matching feature that indicates whether $t$ appears in $S_1$. Such matching features have been successfully used in neural models for information retrieval \cite{Guo2016ADR}.
While the matching feature provides important syntactic information to the model, it is too restrictive. If two words in $S_1$ and $S_2$ that are not exactly same but semantically related, there is a good change that this influences the fact that $S_1$ and $S_2$ are related through an entailment, paraphrase or semantic relationship. In fact, inter-sentence attention mechanisms try to capture some form of semantic dependency by using the contextual representations derived from a BiLSTM. We take a different approach. We use an external database of paraphrase or semantically related words \cite{Pavlick2015PPDB2B} to capture dependence. For each word $t \in S_1$, we add a \emph{paraphrase feature} to its embedding that indicates whether a paraphrase of $t$ appears in $S_2$. Similarly, for each word $t \in S_2$, we add a paraphrase feature that indicates whether a paraphrase of $t$ appears in $S_1$.
The matching and the paraphrase features add only two dimensions to the embeddings of each word but capture important syntactic and semantic interaction between the words of the two sentences.
The importance of regularization in obtaining good generalization performance is a well established fact in deep learning. Several types of regularization specific to recurrent neural networks have been shown to improve performance of LSTM based models e.g. variational dropout \cite{Gal2016ATG} and DropConnect \cite{Merity2018}. We use three types of regularization to train our models in order to achieve good generalization performance.
The base Siamese architecture augmented with the matching and paraphrase features that capture inter-sentence word interaction and regularization define our model -- REGMAPR. We evaluate its performance on six benchmark datasets on textual entailment, paraphrase detection and semantic relatedness. Despite its simplicity, REGMAPR improves upon several existing models which either use complex inter-sentence attention mechanisms or a large number of handcrafted features across all the datasets. It achieves a new state-of-the-art on the SICK dataset for semantic relatedness and on the SNLI dataset for textual entailment among models that do not use inter-sentence attention.
\section{REGMAPR - The Model}
\label{sec:model}
We describe our model by starting from a basic Siamese architecture and augmenting it with additional features. The input to the model is a pair of sentences $S_1$ and $S_2$, with each word mapped to its corresponding distributed representation or word embedding. In this paper we use GloVe embeddings \cite{Pennington2014}). We denote the set of words of $S_i$ by $T(S_i)$ for $i\in\{1,2\}$.
\subsection{BASE}
The basic model uses a standard Siamese architecture. Each sentence is encoded into a single vector using a BiLSTM. As the encoder, we use a max-pooling of the intermediate states of the BiLSTM operating on the sentence. Our choice is inspired by the success of such an encoder in learning general sentence representations \cite{infersent}. In our experiments, we tried other sentence encoders but a max-pooled BiLSTM consistently gave the best results. The encodings of the two sentences $\mathbf{h}_{S_1}$ and $\mathbf{h}_{S_2}$ are composed by concatenating the element-wise absolute difference and element-wise product with the original vectors to form the following feature vector for textual entailment and paraphrase detection.
\begin{equation}
\mathbf{h}_{S_1,S_2} = \left[ \mathbf{h}_{S_1}; \mathbf{h}_{S_2}; |\mathbf{h}_{S_1}-\mathbf{h}_{S_2}|; \mathbf{h}_{S_1}\bigcdot \mathbf{h}_{S_2}\right]
\end{equation}
where \textbf{;} denotes concatenation. Such matching features have been used successfully for textual entailment in the past \cite{Mou2016NaturalLI}. For semantic relatedness, we only use the absolute difference and product, as follows.
\begin{equation}
\mathbf{h}_{S_1,S_2} = \left[ |\mathbf{h}_{S_1}-\mathbf{h}_{S_2}|; \mathbf{h}_{S_1}\bigcdot\mathbf{h}_{S_2}\right]
\end{equation}
This feature vector is passed through a fully connected layer, followed by ReLU activation, followed by a classification or scoring layer.
For semantic relatedness, we produce a single number which is then passed through an exponential function and clamped to 1 to constrain it in the range $[0,1]$ \cite{Mueller2016}.
\subsection{BASE+REG}
Although regularization is strictly not a part of the architecture, we emphasize its importance. Based on the work of \cite{Merity2018}, we apply three types of regularization.
\begin{enumerate}
\item \emph{Locked Dropout} ($d_e$) after the word embedding layer. In this case, a single dropout mask, where each dimension is dropped with probability $d_e$, is selected for a sentence and applied to all the words in the sentence. Also known as variational dropout, its effectiveness has been demonstrated in sequence processing using recurrent neural networks \cite{Gal2016ATG}.
\item \emph{Dropout} ($d_f$) after the ReLU activation. This is classical dropout proposed in \cite{Srivastava2014DropoutAS}.
\item \emph{Recurrent Dropout} ($d_w$) on the recurrent weights in the BiLSTM encoder. First proposed in \cite{Wan2013RegularizationON}, this regularization helps reduce overfitting of the hidden-to-hidden weight matrices in the LSTM.
\end{enumerate}
Note that the three types of regularization have been chosen carefully to prevent overfitting in each of the main components of our model. As shown later in the paper, they help significantly in getting good generalization performance.
\subsection{BASE+REG+MA}
In this model, in addition to the regularization, we augment the word embeddings with a \emph{matching feature}, denoted by MA henceforth. That is, for each word $t$ in $S_i$, we augment the embedding of $t$ as
\begin{equation}
\mathbf{E}_{S_i}^{\text{MA}}(t) = \left[ \text{GloVe}(t); \mathbbm{1}\{t\in T(S_{3-i})\} \right]
\end{equation}
where $\mathbbm{1}$ is the indicator function. Note that this is a binary feature and provides basic syntactic information to the sentence encoder that the same word is present in both the sentences.
\subsection{BASE+REG+PR}
In this model, we augment the word embedding with information about the presence of semantically related words in $S_1$ and $S_2$. We use the paraphrase database (PPDB) \cite{Pavlick2015PPDB2B} as the source dictionary of semantically related words.
For each word $t$ in PPDB, we compute the following set
\begin{equation}
P(t) = \{ t' | t' \text{ is a paraphrase of } t \text{ in PPDB}\}
\end{equation}
We augment the embedding of $t\in S_i$ using $P(t)$ as follows.
\begin{equation}
\mathbf{E}_{S_i}^{\text{PR}}(t) = \left[ \text{GloVe}(t); \mathbbm{1}\{|P(t)\cap T(S_{3-i})|>0\} \right]
\end{equation}
We call this the \emph{paraphrase feature} or PR henceforth. This again is a binary feature and is easy to compute once $P(t)$ has been precomputed.
Depending on the criteria used for selecting the paraphrase database, this feature can be slightly noisy and yet provides valuable semantic information which is not easily obtainable from the surface forms or word embeddings directly. To the best of our knowledge, this is the first use of PPDB in neural models for text matching.
\subsection{BASE+REG+MA+PR}
The full REGMAPR model combines regularization with the matching and paraphrase features. The word embedding then becomes
\begin{eqnarray}
\mathbf{E}_{S_i}^{\text{MA+PR}}(t) = & \text{[} \text{GloVe}(t); \mathbbm{1}\{t\in T(S_{3-i})\}; \\ \notag
& \mathbbm{1}\{|P(t)\cap T(S_{3-i})|>0\} \text{]}
\end{eqnarray}
Note that the full REGMAPR model increases the word embedding by only two dimensions and uses a Siamese architecture for matching the representations of the two sentences. We completely avoid inter-sentence attention mechanisms and instead encode the inter-sentence interaction using the very simple MA and PR features. Crucially, the MA and PR features provide important syntactic and semantic clues that the BiLSTM can exploit. A schematic diagram of the model is shown in Fig. \ref{fig:model}.
REGMAPR is a general architecture which can be applied to any text matching problem. As we show in the next sections, despite its simplicity, it is highly effective in achieving results comparable or better than more complex models.
\section{Training and Evaluation}
\begin{table}[t]
\centering
\begin{tabular}{llll}
\toprule
Dataset & Train & Dev & Test \\
\midrule
Textual Entailment \\
SNLI & 549.4K& 98.4K& 98.2K\\
SICK-E & 4.5K & 0.5K & 4.9K\\
\midrule
Paraphrase Detection \\
MSRP & 3.7K & 0.4k & 1.7K\\
QUORA & 384.4K & 10K & 10K \\
\midrule
Semantic Relatedness \\
SICK & 4.5K & 0.5K & 4.9K\\
STSB & 5.8K &1.5K &1.4K \\
\bottomrule
\end{tabular}%
\caption{Datasets and the sizes of the respective train, dev and test sets.}
\label{tab:datasets}
\end{table}
\subsection{Datasets}
\pgfplotsset{cycle list/Set1-4}
\begin{figure}[t]
\begin{tikzpicture}
\usetikzlibrary{patterns}
\begin{loglogaxis}[ybar, bar width=2,
xtick={ 50, 100, 200, 400, 1000, 2000, 4000},
xticklabels={ 1, 0.1K, 0.2K, 0.4K , 1K, 2K, 4K},
x tick label style={rotate=0},
xlabel={Number of paraphrases},
ylabel={Number of Words},
y label style={at={(0.05,0.5)}},
ymin=1,
legend cell align={left},
every axis plot/.append style={fill},
]
\addplot[black!60!green]
coordinates
{(50,84640)(101,5776)(201,2709)(301,1717)(401,1113)(501,749)(601,586)(701,447)(801,360)(901,275)(1001,231)(1101,185)(1201,145)(1301,116)(1401,103)(1501,91)(1601,63)(1701,54)(1801,37)(1901,41)(2001,22)(2101,23)(2201,20)(2301,17)(2401,18)(2501,9)(2601,10)(2701,4)(2801,5)(2901,3)(3001,4)(3101,1)(3201,3)(3301,1)(3401,0)(3501,1)(3601,1)(3701,0)(3801,0)(3901,0)(4001,0)(4101,1)(4201,0)(4301,0)(4401,0)(4501,0)(4601,0)(4701,0)(4801,0)(4901,0)};
\end{loglogaxis}
\end{tikzpicture}
\caption{Histogram of words based on the number of paraphrases in PPDB. Both axes are log scaled. The gap between the first two bars has been truncated for ease of visualization.}
\label{fig:ppdb_hist}
\end{figure}
We evaluate our models on six diverse datasets related to three tasks - textual entailment, paraphrase detection and semantic relatedness. For textual entailment, we use the SNLI dataset \cite{Bowman2015} and the SICK-E dataset \cite{Marelli2014ASC}. Each sentence pair in SNLI and SICK-E has a label from the set \{entailment, contradiction, neutral\}.
For paraphrase detection we use the MSRP \cite{Dolan2005AutomaticallyCA} and QUORA \cite{Quora} datasets . Each sentence pair in these two datasets has a binary label indicating whether they are paraphrases of each other. For semantic relatedness we use the SICK \cite{Marelli2014ASC} and STS Benchmark (STSB) \cite{Cer2017SemEval2017T1} datasets, in which each pair of sentences has a semantic relatedness score. The sizes of the train, dev and test sets for each of these datasets is shown in Table \ref{tab:datasets}.
To compute the PR feature, we use the lexical subset of the PPDB paraphrase dataset (specifically the ppdb\_xxl set). There are about 3.7 million pairs of words with an associated score, which we ignore. There are about 99.6K unique words, with more than 50\% having less than 11 paraphrases. A histogram of the frequency of words according to the number of paraphrases is shown in Fig. \ref{fig:ppdb_hist}.
\subsection{Training}
For all experiments, we set the LSTM hidden dimension and the dimension of the fully connected layer to 600. We use 300 dimensional GloVe \cite{Pennington2014} word embeddings. The word embeddings are not updated during training. For the textual entailment and paraphrase detection tasks, a cross-entropy loss function is used, while for the semantic relatedness task the mean squared error (MSE) between the predicted and ground-truth score is used as the loss function. We optimize the weights of the network using Adam \cite{Kingma2015} with a learning rate of 1e-3, which is decayed by 0.5 when the validation performance drops.
For each word in PPDB, we construct a one to many map representing its paraphrases. To create the PR feature for a word in a sentence, we lookup this map and check whether there are any words common with the other sentence.
Hyperparameter search for regularization is done over the following ranges -- locked dropout $d_e \in \{0.0,0.1,0.2,0.3,0.4\}$, dropout $d_f \in \{0.0,0.1,0.2,0.3,0.4\}$ and recurrent dropout $d_w \in \{0,0.1,0.2\}$.
For the SICK dataset, the relatedness scores are linearly scaled from $[1,5]$ to $[0,1]$. For the STSB dataset, the scores are scaled from $[0,5]$ to $[0,1]$.
\section{Results}
We report results for each of the models defined in Section 2
i.e. BASE, BASE+REG, BASE+REG+MA, BASE+REG+PR and BASE+REG+MA+PR
for all the six datasets. In all the tables, the best among these is highlighted by bold fonts and the overall best by an underline.
\subsection{Textual Entailment}
\begin{table}[t]
\centering
\begin{tabular}{lll}
\toprule
Model & Acc. \\
\midrule
Shortcut Stacked Encoder \cite{Nie2017ShortcutStackedSE} & 86.0 \\
Reinforced Self Attention \cite{Shen2018ReinforcedSN} & 86.3 \\
Generalized Pooling \cite{Chen2018EnhancingSE} & 86.6 \\
\midrule
\midrule
BASE & 85.2 \\
BASE+REG & 85.9 \\
BASE+REG+PR &86.1 \\
BASE+REG+MA & 86.1\\
BASE+REG+MA+PR & \underline{\bf{86.8}} \\
\bottomrule
\end{tabular}%
\caption{Accuracy results on SNLI test set. Previous results are for models without inter-sentence attention.}
\label{tab:snli_results}
\end{table}
\begin{table}[t]
\centering
\begin{tabular}{lll}
\toprule
Model & Acc. \\
\midrule
Denotational Approach & 84.6 \\
\cite{Lai2014IllinoisLHAD} & \\
ABCNN \cite{Yin2016ABCNNAC} & 86.2 \\
Attention Pooling \cite{Yin2017TaskSpecificAP} & \underline{87.2} \\
\midrule
\midrule
BASE & 86.4 \\
BASE+REG & 86.6 \\
BASE+REG+PR &86.6 \\
BASE+REG+MA & 86.7\\
BASE+REG+MA+PR & \bf{87.0} \\
\bottomrule
\end{tabular}%
\caption{Accuracy results on SICK-E test set. }
\label{tab:sicke_results}
\end{table}
The results of REGMAPR on SNLI are shown in Table \ref{tab:snli_results}. Regularization helps push the performance of the base model by 0.7\%. Both word matching and paraphrase matching help further, but with a smaller boost of 0.2\%. The combination of MA and PR improves model performance by a much larger 0.9\% compared to the regularized model only. We compare our results with existing models that do not use inter-sentence attention. REGMAPR sets a new state-of-the-art accuracy of 86.8\% for this class of models. Although REGMAPR is not strictly a sentence encoding based model, we do not use sophisticated attention mechanisms like those in ESIM \cite{Chen2017}.
For the SICK-E dataset, REGMAPR achieves 87.0\% accuracy on the test set, at par with more complex models that use task specific inter-sentence attention mechanisms \cite{Yin2017TaskSpecificAP}. Interestingly, the base model itself achieves 86.4\% accuracy. This points to the fact that simple models with an appropriate number of parameters can sometimes achieve similar or better performance than more complex models, even without regularization. The gain from using regularization and the MA and PR features is modest, maxing out at 0.6\%. One possible reason for this is the relatively small size of the dataset and the class skew in the training set (56\% sentence pairs have neutral labels), as compared to the almost uniform class distribution in the SNLI training set.
\begin{table}[t]
\centering
\begin{tabular}{lll}
\toprule
Model
& Acc. & F1 \\
\midrule
TF-KLD \cite{Ji2013DiscriminativeIT} & 78.6 & 84.6 \\
TF-KLD + Fine-Grained & \underline{79.9} & \underline{85.4} \\
\cite{Ji2013DiscriminativeIT} & & \\
Relation Learn & 79.1 & 85.2 \\
\cite{Filice2015StructuralRF} & & \\
\midrule
\midrule
BASE & 77.2 & 84.0 \\
BASE+REG & 77.9 & 84.3 \\
BASE+REG+PR & 78.1 & 84.5 \\
BASE+REG+MA & {78.6} & {84.7} \\
BASE+REG+MA+PR & \bf{79.1} & \bf{85.3} \\
\bottomrule
\end{tabular}%
\caption{Accuracy and F1 results on MSRP test set. }
\label{tab:msrp_results}
\end{table}
\begin{table}[t]
\centering
\begin{tabular}{lll}
\toprule
Model & Dev & Test \\
\midrule
pt-DECATTchar \cite{Tomar2017NeuralPI} & 88.89 & 88.40\\
DIIN \cite{Gong2018} & {89.44} & {89.06}\\
MwAN \cite{Tan2018MultiwayAN} & 89.60 & \underline{89.12} \\
\midrule
\midrule
BASE & 88.17 & 87.56 \\
BASE+REG & 88.06 & 88.02 \\
BASE+REG+PR & 88.37 & 88.12 \\
BASE+REG+MA & 88.74 & 88.39 \\
BASE+REG+MA+PR & {89.05} & \bf{88.64} \\
\bottomrule
\end{tabular}%
\caption{Accuracy results on the QUORA dev and test set. }
\label{tab:quora_results}
\end{table}
\begin{table}[t]
\centering
\begin{tabular}{llll}
\toprule
Model
& $r$ & $\rho$ & MSE \\
\midrule
Atten. Tree-LSTM & 0.8730 & 0.8117 & 0.2426 \\
\multicolumn{3}{l}{\cite{Zhou2016ModellingSP} } \\
Match-LSTM & 0.8822 &0.8345 & 0.2286 \\
\multicolumn{3}{l}{\cite{Mueller2016} }\\
IWAN-skip & 0.8833 & 0.8263 & 0.2236 \\
\multicolumn{3}{l}{\cite{Shen2017InterWeightedAN}} \\
\midrule
\midrule
BASE & 0.8842 & 0.8294 & 0.2224 \\
BASE+REG & 0.8850 & 0.8311 & 0.2240 \\
BASE+REG+PR & 0.8857 & 0.8335 & 0.2208 \\
BASE+REG+MA & 0.8857 & 0.8298& 0.2192 \\
BASE+REG+MA+PR & \underline{\bf{0.8864}} & \underline{\bf{0.8308}} & \underline{\bf{0.2192}} \\
\bottomrule
\end{tabular}%
\caption{Pearson correlation ($r$), Spearman correlation ($\rho$) and MSE w.r.t the ground-truth scores on SICK test set.}
\label{tab:sick_results}
\end{table}
\subsection{Paraphrase Detection}
The results for the MSRP dataset are shown in Table \ref{tab:msrp_results}.
The trend here is also quite clear with increasing performance as regularization, MA and PR features are added.
REGMAPR achieves an accuracy of 79.1\%, at par with the results obtained by \cite{Filice2015StructuralRF}, where the authors use, among other things, a combination of more than five handcrafted features (including the MA feature) in a non-neural model. Our model is surpassed only by \cite{Ji2013DiscriminativeIT}, where the authors use a combination of a term frequency based model and 10 fine grained features.
The results on the QUORA dataset are shown in Table \ref{tab:quora_results}. The full REGMAPR model achieves a test accuracy of 88.64\%. This is better than the BiMPM model of \cite{Wang2017BilateralMM} and pt-DECATTchar model of \cite{Tomar2017NeuralPI}, both of which use attention mechanisms and, in the latter case, heavy data augmentation. For both the paraphrase datasets, the combination of MA and PR features works the best, reflecting their complementary strengths.
We emphasize the good performance of REGMAPR on two of the largest datasets (SNLI and QUORA) considered in this paper. More complex models like DIIN \cite{Gong2018} do eventually perform better on both but at the cost of significantly increased model complexity and training time.
\begin{table}[t]
\centering
\begin{tabular}{lll}
\toprule
Model
& Dev $r$ & Test $r$ \\
\midrule
CNN (HCTI) \cite{Shao2017HCTIAS} & 0.834 & 0.784\\
Conversations \cite{YangSTS2018} & 0.835 & 0.808\\
ECNU \cite{Tian2017ECNUAS} & {0.847} & \underline{0.810} \\
\midrule
\midrule
BASE & 0.797 & 0.786 \\
BASE+REG & 0.798 & 0.786 \\
BASE+REG+PR & 0.801 & 0.789 \\
BASE+REG+MA & 0.802 & 0.793 \\
BASE+REG+MA+PR & {0.805} & \bf{0.795} \\
\bottomrule
\end{tabular
\caption{Dev and test Pearson correlation ($r$) w.r.t the ground-truth scores on STSB test set.}
\label{tab:sts_results}
\end{table}
\subsection{Semantic Relatedness}
For the SICK dataset, REGMAPR sets a new state-of-the-art performance of 0.8864 Pearson correlation, without using any of the additional features or post-processing used by \cite{Mueller2016}. In fact, their base Siamese model built using an LSTM achieves a Pearson correlation which is about 0.07 less than our base model which achieves 0.8842 Pearson correlation. This points to the significance of using a good sentence encoder. The improvements obtained from the MA and PR features are significantly lesser than the improvements seen in textual entailment and paraphrase detection. This can be partially explained by the fact that capturing semantic relatedness is a more complex problem.
The results on the STS Benchmark are shown in Table \ref{tab:sts_results}. Here again, the full REGMAPR model performs better than previous neural models like \cite{Shao2017HCTIAS} by almost 1.1\% in Pearson correlation on the test set. Compared to REGMAPR, better performing models like ECNU \cite{Tian2017ECNUAS} use a large number of handcrafted features (66 to be precise) and \cite{YangSTS2018} use transfer learning after training on much larger datasets.
\pgfplotsset{cycle list/Set1-4}
\begin{figure}[t]
\begin{tikzpicture}
\usetikzlibrary{patterns}
\begin{axis}[ybar, bar width=4,
enlargelimits=0.1,
xtick={1, 2,3,4,5,6},
xticklabels={SNLI, SICK-E, MSRP, QUORA, SICK, STSB},
x tick label style={rotate=0},
ylabel={Percentage},
y label style={at={(0.09,0.5)}},
ymin=0,
legend cell align={left},
legend entries={+REG,+REG+PR,+REG+MA,+REG+MA+PR},
legend style={at={(0.5,1.15)}, anchor=north,legend columns=2},
every axis plot/.append style={fill},
cycle list name=Set1-4
]
\addplot
coordinates
{(1,0.7) (2,0.2) (3,0.7) (4,0.46) (5,0.08) (6,0) };
\addplot
coordinates
{(1,0.9) (2,0.2) (3,0.9) (4,0.56) (5,0.15) (6,0.3) };
\addplot
coordinates
{(1,0.9) (2,0.3) (3,1.4) (4,0.83) (5,0.15) (6,0.7) };
\addplot
coordinates
{(1,1.6) (2,0.6) (3,1.9) (4,1.08) (5,0.22) (6,0.9) };
\end{axis}
\end{tikzpicture}
\caption{Gains in test set performance by using REG, MA and PR over the BASE model. We use $100\times$Pearson correlation for SICK and STSB and accuracy percentage for others.}
\label{fig:gains}
\end{figure}
\begin{figure}
\begin{tikzpicture}
\begin{axis}[
width=0.99\linewidth,
height=0.99\linewidth,
xlabel={Feature},
y label style={at={(axis description cs:0.1,.5)},anchor=south},
ylabel={Relative difference in proportion $R^X$},
xmin=0.5, xmax=3.1,
ymin=-0.2, ymax=0.7,
xtick={1,2,3},
xticklabels={$\text{X=PR}$ , X=MA, X=MA+PR},
legend pos=north west,
ymajorgrids=true,
grid style=dashed,
legend cell align={left},
legend entries={SNLI, SICK-E, MSRP, QUORA,SICK,STSB}
]
\addplot[
color=blue,
mark=*,
]
coordinates {
(1, 0.049)(2,0.233)(3,0.333)
};
\addplot[
color=magenta,
mark=*,
]
coordinates {
(1, 0.111)(2,-0.038)(3,-0.167)
};
\addplot[
color=cyan,
mark=*,
]
coordinates {
(1,0.141)(2,0.158)(3,0.333)
};
\addplot[
color=orange,
mark=*,
]
coordinates {
(1,0.0)(2,0.321)(3,0.4)
};
\addplot[
color=black,
mark=*,
]
coordinates {
(1,-0.044)(2,0.323)(3,0.444)
};
\addplot[
color=black!30!green,
mark=*,
]
coordinates {
(1,0.258)(2,0.305)(3,0.667)
};
\end{axis}
\end{tikzpicture}
\captionof{figure}{Estimate of predictive power of MA and PR.}
\label{fig:prop}
\end{figure}
\begin{figure}[h!]
\begin{tikzpicture}
\usetikzlibrary{patterns}
\begin{axis}[ybar, bar width=4,
xtick={1, 2,3,4,5,6},
xticklabels={SNLI, SICK-E, MSRP, QUORA, SICK, STSB},
x tick label style={rotate=0},
ylabel={Percentage},
y label style={at={(0.09,0.5)}},
ymin=0,
legend cell align={left},
legend entries={BASE,BASE+PR,BASE+MA,BASE+MA+PR},
legend style={at={(0.5,1.13)}, anchor=north,legend columns=2},
every axis plot/.append style={fill},
cycle list name=Set1-4
]
\addplot
coordinates
{(1,0.7) (2,0.2) (3,0.7) (4,0.46) (5,0.08) (6,0) };
\addplot
coordinates
{(1,1.0) (2,0.7) (3,1.8) (4,1.19) (5,0.67) (6,0.2) };
\addplot
coordinates
{(1,0.7) (2,0.4) (3,1.4) (4,1.09) (5,0.57) (6,0.8) };
\addplot
coordinates
{(1,1.1) (2,0.7) (3,1.1) (4,0.79) (5,0.01) (6,0.5) };
\end{axis}
\end{tikzpicture}
\caption{Gains in test set performance by using REGularization over BASE, BASE+MA, BASE+PR and BASE+MA+PR.}
\label{fig:reg_gains}
\end{figure}
\section{Analysis of Results}
In this section, we analyze the results presented in the previous section by comparing the contribution of each of the main components of REGMAPR. In Fig.~\ref{fig:gains}, we summarize the gains in performance due to each component of REGMAPR over the base model for all the six datasets. Some trends can be ascertained from the plot. In all the cases, the MA feature helps equally or more than the PR feature. The combination of MA and PR consistently produces the highest gains, which justifies our choice of modeling inter-sentence word interaction using both of these features.
To further illustrate the effect of the two features, we investigate the correlation of the the presence of these features with the class labels or scores. For each of the datasets, we partition the training set into two classes - positive ($P$) and negative ($N$). For MSRP and QUORA, pairs of sentences that are paraphrases are positive and pairs that are not paraphrases are negative. For SNLI and SICK-E, pairs of sentences that have the entailment label are positive and those that have the contradiction label are negative. For SICK and STSB, we compute the average semantic relatedness score in the training set and mark all pairs that have score greater than equal to the average as positive and the remaining as negative. Next, over all the pairs in $P$, we compute the proportion of words that have the MA feature on as follows.
\begin{equation}
R_P^{\text{MA}} = \frac{{\sum_{(S_1,S_2)\in P}} |\{t | \mathbf{E}_{S_1}^{\text{MA}}(t)=1\}| + |\{t | \mathbf{E}_{S_2}^{\text{MA}}(t)=1\}| }{{\sum_{(S_1,S_2)\in P}} |S_1|+|S_2|}
\end{equation}
Similarly, we compute the proportions $R_N^{\text{MA}}, R_P^{\text{PR}}, R_N^{\text{PR}}$ and $R_P^{\text{MA+PR}}, R_N^{\text{MA+PR}}$. For $\text{MA+PR}$, we count the words that have both the features on. These proportions may be interpreted as estimates of how likely a word occurring in a sentence pair with a positive (or negative) label has the MA, PR or both features on. Finally, to estimate the predictive power of the features, we condense the ratios for the positive and negative classes into one number as follows.
\begin{equation}
R^{\text{MA}} = \frac{R_P^{\text{MA}}- R_N^{\text{MA}}}{(R_P^{\text{MA}}+R_N^{\text{MA}})/2}
\end{equation}
The ratios $R^{\text{MA}}$ and $R^{\text{MA+PR}}$ are defined similarly. A higher positive value of this ratio means that it is more likely that the corresponding feature is present in words occurring in sentences in $P$ as compared to $N$ and hence can have higher predictive power.
We plot the values of $R^{\text{MA}}$, $R^{\text{PR}}$ and $R^{\text{MA+PR}}$ for each of the six datasets in Fig. \ref{fig:prop}. Except for the SICK-E dataset, there is a consistent increase as we move from $R^{\text{PR}}$, to $R^{\text{MA}}$ and finally to $R^{\text{MA+PR}}$. This partially explains the relative gains shown in Fig. \ref{fig:gains}. Since we used the PPDB database without any filtering, there may be noisy paraphrases resulting in values of $R^{\text{PR}}$ that are close to zero for some datasets.
Finally, to evaluate the effect of regularization (REG), we plot the gains obtained by including it in the presence or absence of the MA and PR features in Fig. \ref{fig:reg_gains}. It is clear that proper regularization helps a great deal in improving generalization and hence test set performance. The trends across the features here are less clear, although regularization tends to help most when only the PR feature is used.
\section{Related Work}
Text matching is a general problem in NLP and the three specific tasks considered in this paper have a long history of their own. For textual entailment,
most of the state-of-the-art architectures for the SNLI dataset use inter-sentence attention mechanisms, canonical examples being ESIM \cite{Chen2017} and BiMPM \cite{Wang2017BilateralMM}. The attention in these models is typically computed using the intermediate states of a BiLSTM. On the other hand, we encode the word interaction features in the word embeddings and are not parameter free. Purely Siamese architectures have been used in the past for textual entailment \cite{Mou2016NaturalLI}, but they have lower accuracy.
For paraphrase detection, many of the best models for the MSRP dataset use a combination of handcrafted features in a non-neural setting. These include KL-Divergence based features in \cite{Ji2013DiscriminativeIT} and syntactic and semantic matching features in \cite{Filice2015StructuralRF}. The latter work uses at least five features, including the MA feature. REGMAPR uses fewer and simpler features and incorporates them in a neural model to achieve comparable performance. On the larger QUORA dataset, most existing models use inter-sentence attention mechanisms.
Previous work on semantic relatedness has centered largely on combining neural models with handcrafted features. The state-of-the-art model for STSB \cite{Tian2017ECNUAS} combines a total of 66 features with a neural model. Our model achieves comparable performance with far lesser complexity. The use of the exponential function for semantic relatedness scoring was first introduced by \cite{Mueller2016} who apply it directly to the L1 distance between the encodings of the two sentences. Our model uses a more complex matching function followed by fully connected layers before applying the exponential function.
In a related work, the authors in \cite{Tymoshenko2015AssessingTI} devise a number of syntactic and semantic matching features for the answer passage reranking task from information retrieval in a non-neural setting. The syntactic features include the MA feature and the semantic matching features are derived from external resources like YAGO, DBPedia and Wikipedia. In our model, we use PPDB as the only external resource of semantic matching information and let the neural network learn the features that are most suitable for a particular task.
Finally, there has been some work on exploring general architectures for text matching problems. The work of \cite{Wang2016ACM} explores various techniques for estimating word interaction through an attention mechanism. Similar approaches were explored in \cite{He2016PairwiseWI}, \cite{Parikh2016ADA} and \cite{Wang2016MachineCU}. Among more recent models, \cite{Gong2018} and \cite{Kim2018SemanticSM} have achieved state-of-the-art performance on the SNLI and QUORA datasets using multi-layered and highly complex attention mechanisms.
The use of dropout regularization for better generalization is a well established principle in deep learning. Our work adopts a subset of the suite of regularizations successfully used in language modeling by \cite{Merity2018}. Of particular significance is the use of recurrent dropout or DropConnect \cite{Wan2013RegularizationON} for the weights of the BiLSTM. To the best of our knowledge, ours is the first work to explore its use in text matching problems.
\section{Conclusion}
In this paper, we propose REGMAPR -- a neural model for text matching that incorporates simple word interaction features in a Siamese architecture and train it with three different types of regularization. Our model performs comparably or better than many existing models that use complex inter-sentence attention mechanisms or many handcrafted features on six diverse datasets. In future work, we plan to explore further ways to infuse inter-sentence semantic information in the word embeddings and matching heuristics over multi-layered sentence representations.
\bibliographystyle{aaai}
|
\section{Introduction}
\label{sec:introduction}
The accreting neutron star binary IGR J17062$-$6143 (hereafter, J1706)
is one of the most recently identified accreting millisecond X-ray
pulsars (AMXP). First observed in outburst in 2006
\citep{2007A&A...467..529C, 2008ATel.1840....1R, 2008ATel.1853....1R},
it has since then been persistently accreting at luminosities in the
range 5.8 - $7.5 \times 10^{35}$ erg s$^{-1}$ (2 - 20 keV), assuming a
distance of 7.3 kpc \citep{2013ApJ...767L..37D, 2017ApJ...836..111K,
2017ApJ...836L..23S, 2018MNRAS.475.2027V}. The object's neutron star
nature was first revealed by the detection of thermonuclear X-ray
bursts. The first of these was observed by {\it Swift} in 2012
\citep{2013ApJ...767L..37D}, and most recently
\citet{2017ApJ...836..111K} reported on {\it Swift} observations of a
long duration burst first detected with {\it MAXI}
\citep{2015ATel.8241....1N} that was likely powered by burning of a
deep helium layer. The properties of these long duration (tens of
minutes) thermonuclear X-ray bursts are consistent with the
accumulation of helium rich material on the neutron star surface,
which could be accommodated by accretion from a degenerate helium
dwarf in an ultracompact system. However, accretion of hydrogen-rich
fuel under certain conditions can also lead to thick, combustible
helium layers, so the observation of apparently helium-powered nuclear
flashes is not necessarily a definitive indication of an ultracompact
system \citep{1981ApJ...247..267F,
2006ApJ...652..559G}. \citet{2017ApJ...836L..23S}, hereafter SK17,
reported the detection ($4.3 \sigma$) of 163.656 Hz pulsations in a
single $\approx 1200$ s observation with the {\it Rossi X-ray Timing
Explorer} ({\it RXTE}; \cite{1993A&AS...97..355B}). They found a
fractional pulsed amplitude (after background subtraction) of $9.4 \pm
1.1 \%$, but could not determine the orbital period of the system due
to the single, short {\it RXTE} observation. They were able to place a
lower limit on the orbital period of about 17 minutes.
The source has recently been studied extensively with {\it Swift},
{\it NuSTAR}, {\it Chandra} and {\it XMM-Newton}. For example,
\citet{2017MNRAS.464..398D} reported the presence of Fe K$\alpha$
reflection features in {\it NuSTAR} data, modeling of which suggested
an inner disk that may be truncated out to $\approx 100 R_g$, where
$R_g = GM/c^2$. Most recently, \citet{2018MNRAS.475.2027V} presents
results of simultaneous {\it NuSTAR} and {\it XMM-Newton}
observations. They report the presence of reflection features as well,
and suggest a similarly truncated disk as in
\citet{2017MNRAS.464..398D}. They note, however, that a disk
extending down to the neutron star cannot be excluded if the binary
inclination is very low. Based on analysis of {\it XMM-Newton}
Reflection Grating Spectrometer (RGS) data they also suggest the
system may have an oxygen-rich circumbinary environment, perhaps due
to an outflow. Interestingly, they also searched for pulsations using
the {\it XMM-Newton} EPIC timing mode data but did not detect them.
They placed an upper limit on the pulsed fraction in those data of
$5.4 \%$ ($0.5 - 10$ keV). They concluded that the persistently faint
X-ray luminosity could be indicative of either an ultracompact binary
system or perhaps magnetic truncation, but the spectroscopic data
alone were not decisive between these two
possibilities. \citet{2018arXiv180103006H} have recently reported on
broad-band optical to near-infrared (NIR) photometry of J1706{} that
they modeled as emission from an irradiated accretion disk. Their
modeling indicates an accretion disk size consistent with an
ultracompact orbit, and they argued for an orbital period in the range
from 0.4 - 1 hr. Additionally, their optical spectroscopy showed no
H-$\alpha$ emission, consistent with a hydrogen-deficient donor and an
ultracompact system. Thus, sensitive, new timing observations to
determine the binary orbital parameters, and the nature of the system,
were clearly warranted.
In this paper we report results of recent {\it Neutron Star Interior
Composition Explorer} (NICER) observations of J1706{}. The
principal goals of the {\it NICER} observing campaign were to confirm
(or not) the {\it RXTE} detection of 163.656 Hz pulsations and, if
pulsations could be detected, to determine the system's orbital
parameters. We show below that the new {\it NICER} data confirm that
J1706{} is an 163.656 Hz pulsar, and also reveal an ultracompact
orbit with similarities to other ultracompact AMXPs
\citep{2012arXiv1206.2727P}. The plan of the paper is as follows. We
first describe the observations, data selection, and our initial
pulsation search and detection, confirming that J1706{} is a 163.656
Hz pulsar. We next discuss our orbit search and detection, and we
summarize the properties of the system given the orbit solution. We
conclude with a brief summary and discussion of the implications of
our findings for the nature of J1706{}.
\section{NICER Observations and Pulsation Search}
\label{sec:observations}
{\it NICER} was installed on the International Space Station ({\it
ISS}) in 2017 June, and began full science operations after a one
month checkout and verification period. {\it NICER} is optimized for
low background, high throughput, fast timing observations across the
$0.2 - 12$ keV band \citep{2012SPIE.8443E..13G}, achieving an absolute
timing precision of $\approx 100$ ns with the aid of a GPS
receiver. We obtained with {\it NICER} $26$ ks of good exposure on
J1706{} in the time window spanning 2017 August 9 - 15. Additional
observations were obtained in October and November, but we focused on
the August data for our initial pulsation search. We processed and
analyzed the data using HEASOFT version 6.22 and NICERDAS
2017-09-06\_V002. We barycentered the data using the tool {\it
barycorr} employing the {\it DE200} solar system ephemeris and
source coordinates {\it R.A.} $=256.5677 ^{\circ}$, decl. $=-61.7113
^{\circ}$ \citep{2008ATel.1840....1R}. After data processing and
selection we identified 58 good time intervals (GTIs) of at least 50 s
duration in the August data, for a total of $26$ ks of on-source
exposure. The on-source dwell times with {\it NICER} tend to be
somewhat shorter than for free-flying, low-Earth orbit observatories.
Figure 1 shows the resulting light curve accumulated in 16 s bins, and
including events with energies in the range from 0.3 to 5 keV. The
average count rate was $\approx 31$ s$^{-1}$, which is consistent with
the expected rate estimated using source flux and spectra from recent
observations \citep{2017MNRAS.464..398D, 2017ApJ...836..111K}. We do
see evidence for variations in the background counting rates,
particularly in the August data. This is most evident in the band
above $\approx 5$ keV. Based on the average count rate spectrum we
estimate that the {\it NICER} background in the 0.3 to 5 keV band is,
on average, less than the source count rate by a factor of $\approx
15$. We note that we did not observe any thermonuclear X-ray bursts
from the source, but given the long recurrence time, this is not very
surprising \citep{2017ApJ...836..111K}.
For our pulsation search we further limited the upper end of the
energy band to 3.2 keV due to the higher backgrounds present in some
dwells. Our choice here reflected a trade-off between either removing
a substantial number of dwells completely or reducing the upper energy
threshold somewhat, and thereby allowing us to utilize most of the
dwells. To search for pulsations we computed a light curve (0.3 - 3.2
keV) of the full August dataset sampled at 4096 Hz. This light curve
spans 500 ksec ($\approx 5.8$ days) and has $(5\times 10^5)\times 4096
= 2.048\times 10^9$ bins. We computed a FFT power spectrum of this
light curve and searched in the frequency range in which pulsations
were reported by SK17. Figure 2 shows the resulting power spectrum,
normalized as in \cite{1983ApJ...266..160L}, in the vicinity of the
163.656 Hz pulse frequency. The red, vertical dashed lines denote the
approximate range of pulse frequency detected in the {\it RXTE}
observations. An excess of signal power consistent with this
frequency range is clearly evident. The highest power in the plotted
frequency range has a value of 56.3. The expected noise power
distribution is a $\chi^2$ distribution with 2 degrees of freedom.
The probability to exceed this value (56.3) in a single trial is
$6\times 10^{-13}$. There are 15,000 frequency bins in the range from
163.64 to 163.67 Hz. This gives a chance occurrence probability of $9
\times 10^{-9}$ for the highest observed power only. Since additional
excess powers are present as well, this is an extremely conservative
significance estimate.
In Figure 2 one can see that the pulsar signal is comprised of two
main sidebands, each of which is modulated by a number of finely
spaced peaks. The first, most significant, sideband is that near the
center of the frequency band denoted by the red, vertical lines, and
the second is near the high end of the band. There is likely a third
sideband mid-way between these two, but it is weak enough that it is
harder to discern above the noise. The presence of such sidebands in
the power spectrum is a strong indication of the presence of phase
modulation due to orbital motion of the pulsar
\citep{2003ApJ...589..911R}. We do not detect any excess power at
higher harmonics of the pulsar frequency. We note that the finely
spaced peaks modulating each sideband are consistent with the {\it
ISS} orbit period, thus these result from the incomplete sampling
(gaps) in the time series, that is, {\it NICER}'s window function.
As a further test we also computed power spectra of some of the long
individual on-source dwells, and then averaged these. We used 21 of
the longest dwells with exposures ranging from 1294 s (longest) to 545
s (shortest). These intervals are indicated in Figure 1 by vertical
dashed lines drawn at the center of each interval. For each of these
intervals we computed a light curve sampled at 4096 Hz and with a
duration of 2048 s. Since this is longer than each of the individual
exposures we padded the light curves to 2048 s using the mean value
determined from the good exposure in each dwell. This procedure
ensures that the same Fourier frequency spacing is used for each
interval and facilitates simple averaging of the resulting power
spectra. We then computed FFT power spectra for each dwell and
averaged them. The resulting power spectrum is shown in Figure 3, and
clearly shows an excess of power at the same frequency as is evident
in Figure 2. Indeed, we see the same basic signal structure of two
dominant sidebands. Thus, {\it NICER} clearly detects pulsations from
J1706{} in a frequency range consistent with the earlier {\it RXTE}
detection, and moreover, the sidebands in the power spectra are
strongly indicative of accelerated motion of the pulsar.
\section{Searching for the Orbit}
Having recovered pulsations from J1706{} with {\it NICER} we next
began a search for the orbital parameters. The combination of a weak
pulsed signal and relatively short uninterrupted exposures means that
it is not really possible to closely track the pulse frequency
variations with time around the orbit, particularly if the orbital
period is short compared to the typical gap in exposure. It is
therefore not possible to directly ``see'' the orbital frequency
evolution with time in, for example, a dynamic power spectrum. This
makes it more challenging to deduce the orbit. Nevertheless, orbital
motion of the pulsar introduces periodic light travel time
delays/advances that depend on its orbital phase. As noted above,
these produce a characteristic sideband structure in power spectra
computed from a light curve that samples at least several orbital
periods \citep{2003ApJ...589..911R}. In principle, one can measure
the orbital period by detecting this sideband structure in the power
spectrum, as the frequency spacing of the phase modulation sidebands
is set by the orbital period. However, the complex window function
(due to the data gaps) associated with {\it NICER}'s observing windows
makes it challenging in the present case to directly infer
the orbital period in this way.
Because of the challenges outlined above we employed a coherent, grid
search for the orbital parameters using the $Z^2_n$ statistic
\citep{1983A&A...128..245B, 2002ApJ...577..337S}, and since there is
no evidence for harmonic signals we began by restricting the analysis
to $n=1$. The observed population of AMXPs are all in highly circular
orbits, and indeed there are good theoretical arguments why this
should be the case, so we began our search with circular orbit models.
We also start with the assumption that the pulsar frequency does not
vary significantly across the August data epoch. We used the Blandford
\& Teukolsky (1976) relativistic orbit model to parameterize the time
delays, and with the assumption of circularity, we have a four
parameter search space; the pulsar frequency, $\nu_0$, projected
semi-major axis, $a_x \sin i/c$, orbital period, $P_{orb}$, and the
epoch of mean longitude equal to zero, $T_0$. Mean longitude is the
orbit phase angle measured from the ascending node. For a circular
orbit, the pulsar is ``behind'' the companion star at mean longitude
of $90^{\circ}$. We evaluate the statistic,
\begin{equation}
Z_1^2 = \left ( \sum_{j=1}^N \cos\phi_j \right )^2 +
\left ( \sum_{j=1}^N \sin\phi_j \right )^2 \; ,
\end{equation}
where, $\phi_j = \nu_0(t_j + \Delta t_{BT} (t_j, a_x\sin i/c, P_{orb},
T_0))$, $\Delta t_{BT}$ is the binary time delay model
\citep{1976ApJ...205..580B} as a function of orbital parameters, and
$t_j$ are the barycentric photon arrival times. Since such ``brute
force'' searches can be computationally expensive, we began by
searching a subset of the full August dataset. For this we used the
more densely sampled portion of the light curve, the portion of Figure
1 between about 1.5 and 2.2 days. We set up grids of values spanning
the relevant ranges for each parameter. We used the recovered signal
in the power spectra (Figures 2 and 3), as well as prior results to
guide these choices. For example, SK17 used the {\it RXTE} data to
place a lower limit on the orbital period of about 17 min. As noted
above, the sideband structure in the power spectrum is suggestive of a
compact orbit. Based on this we confined our initial search to orbital
periods between 10 and 90 min. We used the power spectral results to
bound both $\nu_0$ and $a_x\sin i/c$. Finally, we employed a sampling
in $T_0$ equivalent to $2^{\circ}$ of orbital phase. We then computed
$Z_1^2$ for all combinations of parameters to find candidate solutions
with large $Z_1^2$. This procedure yielded a candidate orbit solution
with $Z_1^2 = 77.1$, an orbital period of $P_{orb} = 2278$ s, $a_x\sin
i/c = 0.00393$ lt-sec, and $\nu_0 = 163.65611055$ Hz. We did not find
any other comparable $Z^2_1$ maxima within the range of parameter
space searched.
Since this result was obtained from a subset of the August data, we
next attempted to coherently add all the additional August data
segments. We did this by adding data segments one at a time into the
total $Z_1^2$ sum and then used the {\it IDL} function
minimizer/maximizer {\it tnmin} to optimize the solution
\citep{2009ASPC..411..251M}. In each case $Z_1^2$ increased in a
monotonic fashion, and the orbit parameters remained consistent.
Phasing up all the August data in this way resulted in a peak value of
$Z_1^2 = 196.1$. Recall that in the case of pure Poisson noise, the
$Z_1^2$ statistic is distributed as $\chi^2$ with 2 degrees of
freedom, and a value this high has a chance probability (single trial)
of $2.6 \times 10^{-43}$ (13.8 $\sigma$). The pulsed signal after
accounting for the orbital phase delays is dramatically stronger than
with no orbit correction, as expected. Figure 4 compares these two
signals. The curves show $Z_1^2$ evaluated on a grid of $\nu_0$ with
orbit phase delays included (black), and without (red). We note that,
as in Figure 2, the modulating ``comb'' of finely spaced sub-peaks
results from the observing window function. Using the full August data
set we find the best orbital solution has $P_{orb} = 2277.89 \pm 0.48
$ s, $a_x\sin i/c = 0.00395 \pm 0.0003$ lt-sec, $\nu_0 = 163.65611058
\pm 2.7 \times 10^{-7}$ Hz, and $T_0 =$ MJD $57974.82835 \pm 0.0007$
(TDB), where we quote nominal $3\sigma$ errors for a single parameter
by finding the values for each parameter at which $\Delta Z_1^2 = 9$
\citep{2002ApJ...575L..21M}.
We then carried out similar analyses on the October and November data
segments, treated separately, and found consistent results. Finally,
we combined data across all epochs, and found a peak $Z_1^2 =
355.4$. We again determined confidence regions using $\Delta Z_1^2 =
9$, and found the best solution has $P_{orb} = 2278.208 \pm 0.012 $ s
(37.97 min), $a_x\sin i/c = 0.00389 \pm 0.0002$ lt-sec, $\nu_0 =
163.656110496 \pm 9 \times 10^{-9}$ Hz, and $T_0 =$ MJD $57974.82795
\pm 0.00028$ (TDB). The timing solution is summarized in Table 1.
Next, we allowed the eccentricity, $e$, to be non-zero, but this did
not result in a significant increase in $Z_1^2$, and we placed an
upper limit on it of $e < 0.03$ ($1\sigma$, $\Delta Z_1^2 = 1$).
Using our best orbit solution we phase-folded all the data, and fit
the resulting pulse profile (0.3 - 3.2 keV) with a sinusoid, $A +
B\sin(\phi - \phi_0)$. The fit is excellent, with a minimum $\chi^2 =
8.6$ (13 degrees of freedom), and the implied fractional pulsed
amplitude is $B/A = 2.04 \pm 0.11 \%$. Figure 5 shows the resulting
pulse profile and fitted model. We did not detect any harmonic
signals. We note that the pulsed amplitude measured with {\it NICER}
is comfortably below the upper limits reached in the recent pulsation
search with {\it XMM-Newton} reported by \citet{2018MNRAS.475.2027V},
which likely explains why they did not detect the pulsations. We also
computed pulse phase residuals using the best orbit solution. We show
these in Figure 6, where we have added in the orbit-predicted phase
delay in order to visually show the size of the delays. We did not
find any statistically significant long-term trends in these
residuals. Finally, we allowed for a constant pulsar spin frequency
derivative, $\dot\nu$, in the timing model, and recomputed $Z_1^2$ on
a grid of $\nu_0$ and $\dot\nu$ values. We found no significant
increase in $Z_1^2$, and we derived the following limits, $-6 \times
10^{-15} < \dot\nu < 4 \times 10^{-15}$ Hz s$^{-1}$ ($1\sigma$).
\begin{table*}
\renewcommand{\arraystretch}{1.5}
\caption{Timing Parameters for IGR J17062-6143}
\scalebox{0.95}{
\begin{tabular}{lc}
\tableline\tableline
Parameter & Value \\
\tableline
Right ascension, $\alpha$ (J2000) & $256.5677^{\circ}$ \\
Declination, $\delta$ (J2000) & $-61.7113^{\circ}$ \\
Barycentric pulse frequency, $\nu_0$ (Hz) & 163.656110049(9) \\
Pulsar frequency derivative, $\dot\nu$ (Hz s$^{-1}$) & $-6\times 10^{-15} < \dot\nu < 4 \times 10^{-15}$ \\
Projected semi-major axis, $a_x \sin(i) / c$ (lt - s) & 0.0039(2) \\
Binary orbital period, $P_{orb}$ (s) & 2278.208(12) \\
Time of mean longitude equal to zero, $T_0$ & MJD 57974.82795(28) (TDB) \\
Orbital eccentricity, $e$ & $< 0.03$ \\
Pulsar mass function, $f_x$ ($10^{-8} M_{\odot}$) & 9.12(2) \\
Minimum donor mass, $m_d$ ($M_{\odot}$) & 0.005 - 0.007 \\
Maximum $Z_1^2$ power & 355.4 \\
\tableline
\end{tabular}}
\tablecomments{Parameter uncertainties for $\nu_0$, $a_x\sin(i)/c$,
$P_{orb}$ and $T_0$ are $3\sigma$ ($\Delta Z_1^2 = 9$) values, given
in the last quoted digits. Limits for $e$, and $\dot\nu$ are
$1\sigma$. The minimum donor mass range is for neutron star masses
of 1.2 and 2 $M_{\odot}$.}
\end{table*}
\section{Discussion and Summary}
Our analysis of observations of J1706{} with {\it NICER} obtained in
2017 August, October and November confirms the discovery by SK17 that
it is a 163.656 Hz AMXP, and allowed us to derive the orbital
parameters of the system for the first time. The 37.97 min orbital
period of J1706{} is the shortest currently known for an AMXP, and
our measurement confirms several previous indirect indications that
the system is an ultracompact binary \citep{2018arXiv180103006H,
2018MNRAS.475.2027V}. We measure a mass function, $f_x = (m_d
\sin i)^3 / (m_{ns} + m_d)^2 = ( (a_x\sin i)^3 \omega_{orb}^2 )/G =
9.12 \times 10^{-8} M_{\odot}$, which is also the smallest among
stellar binaries. The mass function defines a lower limit to the mass
of the donor star, $m_d$. For a neutron star mass in the range from
$m_{ns} = 1.2 - 2 M_{\odot}$ we find a minimum donor mass in the range
from $0.005 - 0.007 M_{\odot}$. Given the orbital period and a
plausible range of total system mass, the separation between the
components is of order $300,000$ km, and would fit within the Earth -
Moon distance.
The reasonable assumption that the donor star fills its Roche lobe
provides a constraint on the mean density of the donor. This can be
expressed as a constraint on its radius in units of the component
separation, $a$, that depends principally on the system's mass ratio,
$q = m_d / m_{ns}$ \citep{1983ApJ...268..368E}. We combine this
constraint with that from the measured mass function to explore the
implications for the nature of the donor star and the system's orbital
inclination.
Our results are summarized in Figure 7 which shows constraints on the
donor mass and radius. We plot the Roche lobe constraint for three
different neutron star masses, $1.2$ (green), $1.4$ (black), and $1.8$
$M_{\odot}$ (red). The closeness of the three curves is a visual
demonstration of how insensitive this constraint is to the assumed
neutron star mass. The different symbols along the curve mark
inferred donor masses from the mass function constraint for different
assumed orbital inclinations, $i$, and for two values of the neutron
star mass at each inclination. For each pair of symbols the left- and
right-most correspond to a neutron star mass of $1.2$ and $1.8
M_{\odot}$, respectively. The left-most symbol for $i = 90^{\circ}$
marks the minimum donor mass for a $1.2 M_{\odot}$ neutron star.
First, we note that the constraints require hydrogen-deficient donors,
as is expected for systems with an orbital period less than about 80
min \citep{1984ApJ...283..232R, 2002ApJ...577L..27B}. For additional
context we show mass - radius relations obtained from the literature
for several donor types. The dashed curve is the mass - radius
relation for low mass, cold, pure helium white dwarfs from
\cite{1969ApJ...158..809Z}, as corrected by
\cite{1984ApJ...283..232R}. Here we have plotted it using the fitting
formula of \cite{2001A&A...368..939N}. The dotted curves denote a
range of mass - radius values from the binary evolutionary
calculations of \citet{2007MNRAS.381..525D} for the helium donors of
AM CVn systems. The region between the upper and lower dotted curves
gives an indication of the allowed range in mass and radius for donors
at different evolutionary stages, and with different values of central
degeneracy at the onset of mass transfer (see Deloye et al. 2007 for
details). Lastly, the dash-dotted curves show mass - radius relations
for carbon white dwarfs with central temperatures of $10^4$ (lower)
and $3 \times 10^6$ K (upper), from \citet{2003ApJ...598.1217D}. Thus,
J1706{} appears to be a somewhat more extreme example of the
currently known ultracompact AMXPs \citep{2007ApJ...668L.147K,
2012arXiv1206.2727P}.
For a random distribution of inclination angles the chance probability
to observe a system with an inclination less than or equal to $i$ is
$1 - \cos(i)$. The probability to observe an inclination angle less
than or equal to $\approx 18.2^{\circ}$ is $5\%$. From this, and
assuming a $1.8 M_{\odot}$ neutron star mass, we deduce a $95\%$
confidence upper limit to the donor mass of $0.0216 M_{\odot}$, with a
corresponding radius of $0.05 R_{\odot}$, substantially less than that
of any hydrogen-rich brown dwarfs \citep{2000ApJ...542..464C}. We use
this upper limit on the donor radius to place an upper limit on the
inclination of $\approx 84^{\circ}$, since no eclipses are seen in the
light curve.
Additional insight is provided by estimates of the long term mass
accretion rate, $\dot M$, and the realization that the mass transfer
in such systems is driven by angular momentum loss due to
gravitational radiation \citep{1984ApJ...283..232R,
2001ApJ...557..292B}. Interestingly, for J1706{} we have $\dot M$
estimates from both persistent X-ray flux measurements and modeling of
its thermonuclear X-ray bursts that are in substantial agreement
\citep{2017ApJ...836..111K}, and suggest $\dot M \approx 2.5 \times
10^{-11}$ $M_{\odot}$ yr$^{-1}$. This value for J1706{} is also in
general agreement with the calculations of $\dot M$ versus $P_{orb}$
reported by Deloye et al. (2007, see their Figure 15). Based on this,
and the reasonable assumption that the donor responds to mass loss
like a degenerate star \citep{2001ApJ...557..292B}, we can estimate
the donor mass as $m_d = 0.0175 (m_{ns}/ 1.4 M_{\odot})^{-1/3} \;
M_{\odot}$. Using this result we additionally show in Figure 7 (with
blue ``$+$'' symbols) the donor masses and corresponding radii for
neutron star masses of 1.4 and 2 $M_{\odot}$, where the higher donor
mass estimate corresponds to the lower mass neutron star (1.4
$M_{\odot}$). These mass estimates further imply constraints on the
binary inclination angle of $19^{\circ} < i < 27.5^{\circ}$, where the
lower and upper bounds correspond to neutron star masses of 1.4 and 2
$M_{\odot}$, respectively.
The constraints summarized in Figure 7 suggest that J1706{} is
observed at relatively low inclination, and the donor mass - radius
constraints appear to be consistent with the helium donors of AM CVn
systems explored by \citep{2007MNRAS.381..525D} (the dotted curves in
Figure 7). We note that these authors also provide estimates of the
expected orbital period evolution, $\dot P_{orb}$, for these systems.
Given the observed orbital period of J1706{}, the predicted values
are in the range $\dot P_{orb} \approx 1 - 3 \times 10^{-6}$ s
yr$^{-1}$, which can be probed with additional {\it NICER} timing
observations.
Clues to the donor composition in a neutron star X-ray binary can in
principle be provided by the properties of its thermonuclear flashes.
The energetic, long duration burst events seen to date would appear to
be consistent with deep ignition of a helium-rich layer
\citep{2017ApJ...836..111K}. As noted previously, under certain
conditions the stable burning of accreted hydrogen into helium can
result in helium-powered thermonuclear flashes
\citep{1981ApJ...247..267F, 2006ApJ...652..559G}. Our measurements
confirm the previous indications for a hydrogen-deficient donor in
J1706{} \citep{2018arXiv180103006H, 2018MNRAS.475.2027V}, and
definitively rule out this option, since the accreted fuel cannot
contain a significant fraction of hydrogen.
While a helium donor in J1706{} appears quite plausible given the
measurements presented here, as well as its bursting properties, prior
X-ray spectroscopy results have suggested J1706{} may have an
oxygen-rich circumbinary environment, perhaps associated with an
outflow \citep{2018MNRAS.475.2027V}. In addition, spectral modeling of
the Fe K$\alpha$ reflection feature appears to favor a higher
inclination than suggested by our constraint derived from the
assumption of gravitational radiation driven mass loss
\citep{2017MNRAS.464..398D, 2018MNRAS.475.2027V, 2017ApJ...836..111K}.
Based on these indications, \cite{2018MNRAS.475.2027V} favor a CO or
O-Ne-Mg white dwarf donor. Given the constraints on the donor
summarized in Figure 7 this remains a viable option, particularly in
the case of non-conservative mass transfer, as would occur in the
presence of an outflow. However, such a conclusion would also open up
additional questions, such as the nature of the fuel for the observed
X-ray bursts, which is presumably helium \citep{2017ApJ...836..111K},
though we note that \cite{2018arXiv180103006H} did not detect helium
in their optical spectra of J1706{}. Further to this final point,
the bursting low mass X-ray binary 4U 0614$+$091 is another source
with apparently helium-powered X-ray bursts
\citep{2010A&A...514A..65K}, but whose optical spectra are suggestive
of a CO donor with little to no helium \citep{2006A&A...450..725W,
2004MNRAS.348L...7N}. Additional observations will likely be needed
to definitively pin down the nature of the donor in J1706{}.
While most AMXPs are transient systems, J1706{} is distinctive in
that it has been in outburst now for about a decade. This provides
an exciting opportunity to study the long-term spin and orbital
evolution with additional {\it NICER} observations. Moreover, we now
have detections of pulsations from J1706{} at two widely spaced
epochs, in 2008 May with {\it RXTE}, and the present 2017 August,
October and November observations with {\it NICER}. Interestingly,
the source shows some indications of a significant change in pulsed
amplitude in that time-frame. The estimated source pulsed amplitude
measured by SK17 with {\it RXTE} was $9.4 \pm 1.1\%$ ($2 - 12$ keV),
whereas we find $2.04 \pm 0.11 \%$ ($0.3 - 3.2$ keV) with {\it
NICER}. We note that given the current uncertainties associated
with modeling the {\it NICER} background, combined with the fact that
the source count rate is dropping steadily above $\approx 5$ keV, it
is presently challenging to accurately determine the pulsed amplitude
above this energy. Nevertheless, with the present data we can measure
the pulsed amplitude in the $2 - 5$ keV band with reasonable
precision, and we find a value of $3.2 \pm 0.3 \%$. Based on this we
think it likely that the smaller amplitude measured by {\it NICER} is
a real effect and likely represents some secular change within the
system, perhaps associated with the effect of accretion on the
magnetic field, as, for example, suggested by
\cite{2012ApJ...753L..12P}. More definitive conclusions in this
regard should become feasible as the {\it NICER} background
calibration improves. We will pursue this, as well as searches for
energy dependent phase lags and a detailed spectroscopic study in
subsequent work.
\label{sec:conclusions}
\acknowledgments
This work was supported by NASA through the {\it NICER} mission and
the Astrophysics Explorers Program. This research also made use of
data and/or software provided by the High Energy Astrophysics Science
Archive Research Center (HEASARC), which is a service of the
Astrophysics Science Division at NASA/GSFC and the High Energy
Astrophysics Division of the Smithsonian Astrophysical Observatory. SG
acknowledges the support of the Centre National d’Etudes Spatiales
(CNES). We thank the anonymous referee for a helpful report.
\facility{NICER, ADS, HEASARC}
\software{HEAsoft (v6.22; Arnaud 1996), mpfit (Markwardt 2009)}
\bibliographystyle{aasjournal}
|
\section{Introduction and Notation}
\label{1}
\subsection{Introduction}
\label{1.1}
The paper is devoted to presenting the number of Rumer diagrams with
$n$ atoms which form a molecule with $m$ valence bonds. We follow
the classical article~\cite{[50]} where the atoms are represented by
vectors
\begin{equation}
x^{\left(1\right)}=\left(
\begin{array}{cccccccccccc}
x_1^{\left(1\right)}\\
x_2^{\left(1\right)}\\
\end{array}
\right), \ldots,x^{\left(n\right)}=\left(
\begin{array}{cccccccccccc}
x_1^{\left(n\right)}\\
x_2^{\left(n\right)}\\
\end{array}
\right),\label{1.1.1}
\end{equation}
in the complex plane $\mathbb{C}^2$. When $n\geq 2$ each valence
bond between two of them, $x^{\left(i\right)}$ and
$x^{\left(j\right)}$, is modeled by the determinant
\begin{equation}
[x^{\left(i\right)},x^{\left(j\right)}]=\left|
\begin{array}{cccccccccccc}
x_1^{\left(i\right)}& x_1^{\left(j\right)}\\
x_2^{\left(i\right)}& x_2^{\left(j\right)}\\
\end{array}
\right|,\label{1.1.5}
\end{equation}
viewed as a polynomial in the four variables
$x_1^{\left(i\right)},x_2^{\left(i\right)},
x_1^{\left(j\right)},x_2^{\left(j\right)}$. These determinants,
called \emph{brackets}, considered as complex-valued functions in
two vector arguments $x^{\left(i\right)},x^{\left(j\right)}$,
defined on the plane $\mathbb{C}^2$, are the building blocks of the
set of all such functions
$f(x^{\left(1\right)},\ldots,x^{\left(n\right)})$ which have the
same value (are \emph{invariant}) when changing the coordinate
system in $\mathbb{C}^2$ via transformations with determinant $1$.
All such transformations of $\mathbb{C}^2$ form a group called
\emph{special linear group} and denoted $\sl(\mathbb{C}^2)$, or,
simply $\sl(2)$. More precisely, in accord with the first main
theorem on invariants, homogeneous degree $k$ invariant under the
group $\sl(2)$ polynomial $f$ in
$x_1^{\left(1\right)},x_2^{\left(1\right)},\ldots,
x_1^{\left(n\right)},x_2^{\left(n\right)}$ exists exactly when $k$
is even, $k=2m$, and in this case $f$ is a linear combinations of
products (called \emph{bracket monomials}) of the form
\begin{equation}
[x^{\left(i_1\right)},x^{\left(j_1\right)}]\ldots
[x^{\left(i_m\right)},x^{\left(j_m\right)}],\hbox{\ }1\leq
i_s<j_s\leq n,\hbox{\ }s\in [1,m].\label{1.1.10}
\end{equation}
The expression of any bracket monomial~\eqref{1.1.10} produces a
graph $G$ on $n$ vertices $1,\ldots,n$ and $m$ edges
$(i_1,j_1),\ldots,(i_m,j_m)$ without loops, where the vertices are
placed clockwise on a circle in this order and any edge $(i_s,j_s)$
is represented by the segment with end points $i_s$ and $j_s$.
Conversely, each bracket monomial's expression~\eqref{1.1.10} can be
restituted from its graph $G$ which thus is called its \emph{valence
scheme}. In case there are no intersecting edges, the valence scheme
$G$ is said to be \emph{Rumer diagram}. In quantum chemistry, each
Rumer diagram, or, what is the same, bracket monomial, represents a
pure valence state of a molecule with atoms $1,\ldots,n$ and any
invariant (linear combination of bracket monomials) represents a
valence state (mixture of pure valence states).
This idea goes back to the providential paper~\cite{[55]} of
J.~J.~Sylvester, where he, in particular, wrote: "An invariant of a
form or system of algebraic forms must thus represent a saturated
system of atoms in which the rays of all the atoms are connected
into bonds." A meticulous historic analysis of Sylvester's
contributions, written in a modern language, can be found
in~\cite{[25]}. The analogy between invariant theory and chemistry
was completely ignored by the chemists till the end of 19-th century
and was galvanized by the physicist G.~Rumer in his pioneering
paper~\cite{[45]} and in its popular continuation~\cite{[50]}.
In order for the exposition to be self-contained, in section 2 we
gather the necessary definitions and results from different
mathematical disciplines in three subsections: Algebra, Invariant
theory, and Algebraic geometry. From the content of the first
subsection we note the introduction of a special grading (via
graphical multidegree) of a polynomial ring used later. In the last
two subsections we discuss the isomorphism between the graded ring
consisting of $\sl(2)$-invariant polynomials in
$x_1^{\left(1\right)},x_2^{\left(1\right)},\ldots,
x_1^{\left(n\right)},x_2^{\left(n\right)}$ and the graded
homogeneous coordinate ring of the Grassmannian $G(n,2)$. As a
consequence, we use the famous postulation formula of W.~V.~D.~Hodge
and J.~E.~Littlewood in order to find the dimensions of the
homogeneous components of the graded ring of invariants in
Theorem~\ref{5.10.10}.
Section 3 is devoted to meticulous proofs of the three theorems
from~\cite{[50]} and to the enumeration of Rumer diagrams with given
numbers of atoms and valence bonds presented in
Theorem~\ref{10.5.1}.
\subsection{Notation}
\label{1.5}
Below we introduce some notation.
$\mathbb{N}$: The set of positive integers.
$\mathbb{N}_+$: The set of nonnegative integers.
$\mathbb{C}$: The field of complex numbers.
$[1,k]=\{1,\ldots,k\}\subset\mathbb{N}$.
$\binom{k}{\ell}$ for $k\in \mathbb{N}$, $\ell\in \mathbb{N}_+$: The
binomial coefficient.
$n\in\mathbb{N}$, $n\geq 2$: The number of atoms.
$m\in\mathbb{N}_+$: The number of valence bonds.
\section{Mathematical Background}
\label{5}
\subsection{Algebra}
\label{5.1}
Some elementary notions and statements from algebra are supposed to
be known. We suggest~\cite{[1]} and~\cite{[5]} as universal
references.
By a \emph{monoid} we mean a set $\Delta$ together with an
associative binary operation on $\Delta$, written additively, which
has zero element $0$. Examples are the set $\mathbb{N}_+$ with
addition of integers and any Cartesian power $\mathbb{N}_+^\ell$
with componentwise addition of $\ell$-tuples.
We denote by $A=\mathbb{C}[z_1,\ldots,z_r]$ the ring of polynomials
in the variables $z_1,\ldots,z_r$ over the field $\mathbb{C}$ of
complex numbers. Given a monoid $\Delta$, we say that $A$ is a
\emph{graded of type $\Delta$ ring} if there exists a family
$(A_\lambda)_{\lambda\in\Delta}$ of subgroups of the additive group
$A$ such that $A=\oplus_{\lambda\in\Delta} A_\lambda$ (any
polynomial $f\in A$ can be presented uniquely as
$f=\sum_{\lambda\in\Delta}f_\lambda$ , $f_\lambda\in A_\lambda$, and
only a finite number of $f_\lambda$ are nonzero) and $A_\lambda
A_\mu\subset A_{\lambda+\mu}$ for $\lambda,\mu\in\Delta$. A subring
$I\subset A$ is called \emph{graded subring} if
$I=\oplus_{\lambda\in\Delta}I_\lambda$ where $I_\lambda=I\cap
A_\lambda$ for all $\lambda\in\Delta$. A nonempty subset $I\subset
A$ is called \emph{ideal} of $A$ if difference of two polynomials in
$I$ is again in $I$ and if the product of any polynomial with a
member of $I$ is again a member of $I$. The ideal $I$ in $A$ is said
to be \emph{homogeneous} if $I$ is a graded subring. We note that an
ideal is homogeneous if and only if it can be generated by
homogeneous elements.
Any ideal $I$ in the ring $A$ produces a ring $A/I$, called
\emph{factor-ring}, in the following way: The members of the set
$A/I$ are all subsets of $A$ of the form $f+I$, $f\in A$, and we
have $f+I=g+I$ if and only if $f-g\in I$. Addition and
multiplication in $A/I$ are defined naturally: $(f+I)+(g+I)=f+g+I$,
$(f+I)(g+I)=fg+I$, $0+I=I$ is the zero element, and $1+I$ is the
unit.
In general, a map from a ring (graded type $\Delta$ ring) to another
ring (graded type $\Delta$ ring) is called \emph{homomorphism}
(\emph{homogeneous homomorphism}) if it maps sums and products onto
the sums and products of the corresponding images, respectively (and
maps any homogeneous component into the corresponding homogeneous
component). \emph{Kernel} of this homomorphism is the ideal
consisting of members of the domain which are mapped onto zero. When
the homomorphism is homogeneous, then its kernel turns out to be
homogeneous ideal. The map
\[
c\colon A\to A/I,\hbox{\ }f\mapsto f+I,
\]
is said to be the \emph{canonical homomorphism} and its kernel is
$I$. If the ring $A$ is a graded of type $\Delta$,
$A=\oplus_{\lambda\in\Delta} A_\lambda$, and if the ideal $I$ is
homogeneous, then $A/I$ inherits this grading via $c$:
$A/I=\oplus_{\lambda\in\Delta}c(A_\lambda)$. In this case the
canonical homomorphism is homogeneous. In general, any surjective
homomorphism $\varphi\colon A\to B$ of rings is, up to a unique
isomorphism produced by $\varphi$, canonical: In the commutative
diagram below $\bar{\varphi}$ is an isomorphism.
\begin{equation}
\begin{diagram}
A & & \\
\dTo_{c} & \rdTo^{\varphi}& \\
A/I & \rTo^{\bar{\varphi}} & B.\\\label{5.1.15}
\end{diagram}
\end{equation}
Here the kernel of $\varphi$ is $I$ and commutativity means
$\varphi=\bar{\varphi}\circ c$.
The \emph{natural grading} of $A$ can be defined as the grading of
type $\mathbb{N}_+$, where $A_\lambda$ is the subgroup of $A$,
consisting of all homogeneous degree $\lambda$ polynomials,
$\lambda\in\mathbb{N}_+$, plus the zero polynomial (which has degree
$-\infty$). Since $A_0=\mathbb{C}$, we obtain that any
\emph{homogeneous component} $A_\lambda$ is a $\mathbb{C}$-linear
space.
Let $y_{ij}$, $1\leq i<j\leq n$, be $\binom{n}{2}$ in number
variables, and let $\mathbb{C}[y_{ij}]=\mathbb{C}[y_{ij},\hbox{\
}1\leq i<j\leq n]$ be the corresponding polynomial ring. Let us set
$B=\mathbb{C}[y_{ij}]$ for short and let
$B=\oplus_{m\in\mathbb{N}_+}B_m$ be the natural grading. For any
monomial
\begin{equation}
g=y_{i_1j_1}\ldots y_{i_mj_m}\in B_m,\hbox{\ } 1\leq i_s<j_s\leq
n,\hbox{\ } s\in [1,m],\label{5.1.1}
\end{equation}
a graph $G$ on $n$ vertices $1,\ldots,n$ and $m$ edges
$(i_1,j_1),\ldots,(i_m,j_m)$ without loops can be constructed
exactly in the same way lake the graph of a bracket
monomial~\eqref{1.1.10}. Conversely, each such monomial $g$ can be
restituted from its graph $G$ and the correspondence is one-to-one.
We define $\deg_v(g)$ to be the degree of the vertex $v$ of the
graph $G$, $v\in [1,n]$.
Given $m,m_1,\ldots,m_n\in\mathbb{N}_+$ with $m_1+\cdots+m_n=2m$, we
denote by $B_{m_1,\ldots,m_n}$ the $\mathbb{C}$-linear subspace of
$B_m$, spanned by all monomials~\eqref{5.1.1} such that
$\deg_v(g)=m_v$, $v\in [1,n]$. The $n$-tuple $m_1,\ldots,m_n$ is
called \emph{graphical multidegree} of the monomial~\eqref{5.1.1}.
Let us denote by $\Delta_2^n$ the monoid consisting of all
$n$-tuples $(m_1,\ldots,m_n)\in \mathbb{N}_+^n$ with even sum. Thus,
we obtain a finer grading of type $\Delta_2^n$ of the polynomial
ring $B=\mathbb{C}[y_{ij}]$:
\begin{equation}
B=\oplus_{m\in\mathbb{N}_+}\oplus_{m_1+\cdots+m_n=2m}
B_{m_1,\ldots,m_n},\label{5.1.5}
\end{equation}
and, moreover, for any $m\in\mathbb{N}_+$ we have
\begin{equation}
B_m=\oplus_{m_1+\cdots+m_n=2m} B_{m_1,\ldots,m_n}.\label{5.1.10}
\end{equation}
\subsection{Invariant Theory}
\label{5.5}
The vectors in $\mathbb{C}^2$ have the form~\eqref{1.1.1} (that is,
are $2\times 1$-matrices) and the invertible linear transformations
of $\mathbb{C}^2$ (changes of coordinates) are $2\times 2$-matrices
$\sigma$ with $\det\sigma\neq 0$. The latter form a group $\gl(2)$
(\emph{general linear group}) and the special linear group $\sl(2)$
is the subgroup of $\gl(2)$, consisting of all $\sigma\in\gl(2)$
with $\det\sigma=1$.
For general vectors~\eqref{1.1.1} we set
\[
\mathbb{C}[x^{\left(1\right)},\ldots, x^{\left(n\right)}]=
\mathbb{C}[x_1^{\left(1\right)},x_2^{\left(1\right)},\ldots,
x_1^{\left(n\right)},x_2^{\left(n\right)}]
\]
to be the polynomial ring in $2n$ variables furnished with the
natural grading:
\begin{equation}
\mathbb{C}[x^{\left(1\right)},\ldots,
x^{\left(n\right)}]=\oplus_{k\in\mathbb{N}_+}\mathbb{C}[x^{\left(1\right)},\ldots,
x^{\left(n\right)}]_k. \label{1.5.1}
\end{equation}
The group $\gl(2)$ (and hence $\sl(2)$) acts on the plane
$\mathbb{C}^2$ by matrix multiplication, called \emph{left
multiplication}: If
\[
\sigma=\left(
\begin{array}{cccccccccccc}
\sigma_{11} & \sigma_{12} \\
\sigma_{21} & \sigma_{22} \\
\end{array}
\right)\hbox{\rm\ and\ }x=\left(
\begin{array}{cccccccccccc}
x_1\\
x_2\\
\end{array}
\right),
\]
then $\sigma\cdot x=\sigma x$. This action can be extended naturally
on the Cartesian product
$\mathbb{C}^2\times\cdots\times\mathbb{C}^2$ ($n$ times) via the
rule $\sigma\cdot (x^{\left(1\right)},\ldots,
x^{\left(n\right)})=(\sigma\cdot x^{\left(1\right)},\ldots,
\sigma\cdot x^{\left(n\right)})$. Thus, we obtain an action of the
group $\gl(2)$ on the polynomial ring
$\mathbb{C}[x^{\left(1\right)},\ldots, x^{\left(n\right)}]$ via the
formula
\begin{equation}
(\sigma\cdot f) (x^{\left(1\right)},\ldots,
x^{\left(n\right)})=f(\sigma^{-1}\cdot x^{\left(1\right)},\ldots,
\sigma^{-1}\cdot x^{\left(n\right)}).\label{5.5.1}
\end{equation}
All of the above actions are \emph{linear} ones: The left
multiplication by a fixed $\sigma\in\gl(2)$ induces an invertible
linear transformation of the corresponding $\mathbb{C}$-linear
space. In general, any linear action of a group $\Gamma$ on a
finite-dimensional linear space $W$ is called
\emph{finite-dimensional representation} of this group. In case $W$
can not be represented as a direct sum of two nonzero subspaces on
which this action of $\Gamma$ induces actions on these subspaces,
the representation $W$ of $\Gamma$ is said to be \emph{irreducible}.
When for any $\sigma\in\Gamma$, the corresponding left
multiplication by $\sigma$ in $W$ is the identity linear
transformation, the representation is called \emph{identical}.
Any finite-dimensional representation of $\gl(2)$ or $\sl(2)$ can be
represented as a finite direct sum of irreducible representations of
this group.
In order to explain the relation of the ring
$\mathbb{C}[x^{\left(1\right)},\ldots, x^{\left(n\right)}]$ with the
main statements in the classical invariant theory, we need to
furnish it with a finer structure --- this of a graded
ring~\eqref{1.5.1}. The action of the group $\gl(2)$ preserves the
grading: If $f\in \mathbb{C}[x^{\left(1\right)},\ldots,
x^{\left(n\right)}]_k$, then $\sigma\cdot f\in
\mathbb{C}[x^{\left(1\right)},\ldots, x^{\left(n\right)}]_k$.
Moreover, let $f\neq 0$, $f=\sum_{n\geq 0}f_k$. The polynomial $f$
is invariant if and only if each homogeneous polynomial $f_k$ is
invariant (see~\cite[Ch. 1, Sect. 2, Proposition 2]{[10]}).
The determinants (brackets)
$p_{ij}=[x^{\left(i\right)},x^{\left(j\right)}]$ from~\eqref{1.1.5}
for $1\leq i<j\leq n$ are members of the polynomial ring
$\mathbb{C}[x^{\left(1\right)},\ldots, x^{\left(n\right)}]$. Let us
define $p_{ij}=-p_{ji}$ for $i>j$. The homogeneous degree $2$
polynomials $p_{ij}$, $1\leq i<j\leq n$, satisfy the following
quadratic identities:
\begin{equation}
p_{i_1j_1}p_{j_2j_3}-p_{i_1j_2}p_{j_1j_3}+p_{i_1j_3}p_{j_1j_2}=0
\label{5.5.10}
\end{equation}
for all subsets $\{i_1,j_1,j_2,j_3\}\subset [1,n]$, consisting of
four elements. Let $\inv^{\left(n\right)}$ be the subring of
$\mathbb{C}[x^{\left(1\right)},\ldots, x^{\left(n\right)}]$,
generated by $p_{ij}$, $1\leq i<j\leq n$. The members of the ring
$\inv^{\left(n\right)}$ are constructed in the following way: First,
we define the map
\begin{equation}
\varphi\colon\mathbb{C}[y_{ij}]
\to\mathbb{C}[x^{\left(1\right)},\ldots, x^{\left(n\right)}],\hbox{\
}f(y_{ij})\mapsto f(p_{ij}).\label{5.5.15}
\end{equation}
This map is a homomorphism of rings and the ring
$\inv^{\left(n\right)}$ is the image of $\varphi$ (that is,
$\inv^{\left(n\right)}=\{f(p_{ij})\mid f\in \mathbb{C}[y_{ij}]\}$).
Moreover, $\inv^{\left(n\right)}$ inherits the natural structure of
graded ring on the polynomial ring $\mathbb{C}[y_{ij}]$:
$\inv^{\left(n\right)}=\oplus_{m\in\mathbb{N}_+}\inv_m^{\left(n\right)}$,
where $\inv_m^{\left(n\right)}=\varphi(\mathbb{C}[y_{ij}]_m)$. Thus,
for any $m\in\mathbb{N}_+$ the $\mathbb{C}$-linear space
$\inv_m^{\left(n\right)}$ is spanned by all bracket monomials
\begin{equation}
p_{i_1j_1}\ldots p_{i_mj_m}, 1\leq i_s<j_s\leq n,\hbox{\ } s\in
[1,m].\label{5.5.20}
\end{equation}
The first main theorem of invariants (see, for example~\cite[Ch. 1,
Sect. 2, Proposition 3 and Ch. 2, Sect. 5, Theorem]{[10]}, or,
\cite[1. Fundamentalsatz]{[50]}) yields that the $\mathbb{C}$-linear
space $\inv_m^{\left(n\right)}$ consists of all homogeneous
$\sl(2)$-invariant polynomials of degree $k=2m$ in
$\mathbb{C}[x^{\left(1\right)},\ldots, x^{\left(n\right)}]$ and
there are no homogeneous invariants in
$\mathbb{C}[x^{\left(1\right)},\ldots, x^{\left(n\right)}]_{2m+1}$,
$m\in\mathbb{N}_+$. Thus, the subring
$\inv^{\left(n\right)}\subset\mathbb{C}[x^{\left(1\right)},\ldots,
x^{\left(n\right)}]$ consists of all $\sl(2)$-invariant polynomials.
In case $n=1$, let us denote by $\inv^{\left(1\right)}$ the subring
of $\mathbb{C}[x^{\left(1\right)}]$, consisting of all
$\sl(2)$-invariant polynomials in two variables
$x_1^{\left(1\right)},x_2^{\left(1\right)}$. We have
$\inv^{\left(1\right)}=\mathbb{C}$, and, more precisely,
$\inv_0^{\left(1\right)}=\mathbb{C}$, $\inv_m^{\left(1\right)}=0$
for $m\in\mathbb{N}$.
\subsection{Algebraic Geometry: Grassmanians}
\label{5.10}
Let us consider the homogeneous ideal $Q$ in the polynomial ring
$\mathbb{C}[y_{ij}]$, generated by the homogeneous degree $2$
polynomials
\begin{equation}
Q_{i_1;j_1,j_2,j_3}=y_{i_1j_1}y_{j_2j_3}-y_{i_1j_2}y_{j_1j_3}+
y_{i_1j_3}y_{j_1j_2}\label{5.10.1}
\end{equation}
for all subsets $\{i_1,j_1,j_2,j_3\}\subset [1,n]$, consisting of
four elements. Because of~\eqref{5.5.10} and of the second main
theorem of invariants (\cite[2. Fundamentalsatz]{[50]}), the kernel
of the (graded) homomorphism~\eqref{5.5.15} is the homogeneous ideal
$Q$ of $\mathbb{C}[y_{ij}]$. Thus, in accord with~\eqref{5.1.15},
the graded ring
$\inv^{\left(n\right)}=\oplus_{m\in\mathbb{N}_+}\inv_m^{\left(n\right)}$
and the graded factor-ring $C=\mathbb{C}[y_{ij}]/Q$,
$C=\oplus_{m\in\mathbb{N}_+}C_m$, are isomorphic. In particular, for
any $m\in\mathbb{N}_+$ the $\mathbb{C}$-linear spaces
$\inv_m^{\left(n\right)}$ and $C_m$ are isomorphic.
Let us consider the finer grading of the polynomial ring
$\mathbb{C}[y_{ij}]$ from~\eqref{5.1.5} via graphical multidegrees:
\[
\mathbb{C}[y_{ij}]\oplus_{m\in\mathbb{N}_+}\oplus_{m_1+\cdots+m_n=2m}
\mathbb{C}[y_{ij}]_{m_1,\ldots,m_n}.
\]
Since any polynomial $Q_{i_1;j_1,j_2,j_3}\in \mathbb{C}[y_{ij}]_2$
has graphical multidegree $m_1,\ldots,m_n$ where $m_s=1$ for $s\in
\{i_1,j_1,j_2,j_3\}$ and $m_s=0$ otherwise, the ideal $Q$ is also
homogeneous relative to this finer grading. Therefore the
factor-ring $C=\mathbb{C}[y_{ij}]/Q$ inherits also this grading. In
other words, according to~\eqref{5.1.10}, we can write
\[
C=\oplus_{m\in\mathbb{N}_+}\oplus_{m_1+\cdots+m_n=2m}C_{m_1,\ldots,m_n},\hbox{\
} C_m=\oplus_{m_1+\cdots+m_n=2m}C_{m_1,\ldots,m_n},
\]
where $C_m$ is the image of $\mathbb{C}[y_{ij}]_m$ and
$C_{m_1,\ldots,m_n}$ is the image of
$\mathbb{C}[y_{ij}]_{m_1,\ldots,m_n}$ via the canonical homomorphism
$\mathbb{C}[y_{ij}]\to \mathbb{C}[y_{ij}]/Q$.
\begin{remark} \label{5.10.5} {\rm Let $X$ be the
projective algebraic variety in the projective space
$\mathbb{P}^{\binom{n}{2}-1}$, defined by the equations
$Q_{i_1;j_1,j_2,j_3}=0$ for all polynomials~\eqref{5.10.1}. The
structure of $X$ can be recovered completely by the structure of its
\emph{homogeneous coordinate ring} $C$, that is, by the graded ring
$\inv^{\left(n\right)}=\oplus_{m\in\mathbb{N}_+}\inv_m^{\left(n\right)}$
of $\sl(2)$-invariants in $\mathbb{C}[x^{\left(1\right)},\ldots,
x^{\left(n\right)}]$.
On the other hand, let $G(2,n)$ be the set of all planes in the
$\mathbb{C}$-linear space $C^n$ through the origin. There exists an
one-to-one map between $X$ and $G(2,n)$ via which we can furnish the
set $G(2,n)$ with a structure of projective algebraic variety called
\emph{Grassmann variety}, or, \emph{Grassmannian} for short. It
turns out that $G(2,n)$ is an irreducible smooth algebraic variety
of dimension $2n-4$ with homogeneous coordinate ring
$C=\oplus_{m\in\mathbb{N}_+}C_m$. }
\end{remark}
The dimension of the $\mathbb{C}$-linear space $C_m$,
$m\in\mathbb{N}$, is evaluated as a polynomial of degree $2n-4$ in
one variable $m$ with rational coefficients by W.~V.~D.~Hodge and
J.~E.~Littlewood (a particular case of their so called
\emph{postulation formula} which can be found, for example, in the
book~\cite[vol. 2, Ch. XIV,\P 9]{[30]}). Since $C_m$ and
$\inv_m^{\left(n\right)}$ are isomorphic $\mathbb{C}$-linear spaces
for all $m\in\mathbb{N}_+$, we obtain the following:
\begin{theorem}\label{5.10.10} For any $m\in\mathbb{N}$ one has
\[
\dim(\inv_m^{\left(n\right)})=\left|\begin{array}{ccc}
\binom{m+n-1}{n-1} & \binom{m+n-2}{n-1} \\
\binom{m+n-1}{n-2} & \binom{m+n-2}{n-2} \\
\end{array}\right|.
\]
\end{theorem}
\begin{corollary}\label{5.10.15} If $n=2$, then
$\dim(\inv_m^{\left(n\right)})=1$ for all $m\in\mathbb{N}$.
\end{corollary}
\begin{corollary}\label{5.10.20} If $n\geq 3$, then
\[
\dim(\inv_m^{\left(n\right)})=
\frac{1}{(n-1)!(n-2)!}(m+1)(m+n-1)\prod_{\iota=2}^{n-2}(m+\iota)^2
\]
for all $m\in\mathbb{N}$.
\end{corollary}
\begin{remark}\label{5.10.35} {\rm The author learned about the
isomorphism between the graded ring $\inv^{\left(n\right)}$ of
invariants of the special linear group $\sl(2)$ and the homogeneous
coordinate ring $C$ of the Grassmannian $G(2,n)$ from the comments
of V~L.~Popov to the Russian translation of~\cite{[50]} in the
book~\cite{[40]}.
}
\end{remark}
\section{Rumer Diagrams}
\label{10}
\subsection{The Results of G.~Rumer, E.~Teller, and H.~Weyl}
\label{10.1}
Let $R(m)$ be the set of all Rumer diagrams produced by a fixed
set~\eqref{1.1.1} of $n$ atoms connected with $m$ valence bonds.
Now, let us attach, in addition, to each atom $x^{\left(i\right)}$,
$1\in [1,n]$, the number $m_i$, $m_i\in\mathbb{N}_+$, of valence
bonds connecting this atom. Given a sequence of non-negative
integers $m_1,\ldots,m_n$, let $R(m_1,\ldots,m_n)$ be the set of all
Rumer diagrams with these prescribed valences and let
$\rho(m_1,\ldots,m_n)$ be their number. Let
$\inv_{m_1,\ldots,m_n}^{\left(n\right)}$ be the $\mathbb{C}$-linear
subspace of $\inv_m^{\left(n\right)}$ spanned by all bracket
monomials corresponding to Rumer diagrams from $R(m_1,\ldots,m_n)$,
where $m_1+\cdots+m_n=2m$. We have
\[
R(m)=\cup_{m_1+\cdots+m_n=2m} R(m_1,\ldots,m_n)
\]
--- a union of pair-wise disjoint sets $R(m_1,\ldots,m_n)$, where for
given $m$ the sequence $m_1,\ldots,m_n$ runs through all
non-negative solutions of the Diophantine equation
$m_1+\cdots+m_n=2m$.
In their paper~\cite{[50]}, G.~Rumer, E.~Teller, and H.~Weyl
obtained the theorems below. For completeness, for each one we give
an outline of its proof. Below we use the following notions from
graph theory: For any two vertices $i$,$j$ of a valence scheme $G$
there are two \emph{arcs} on the circle with these endpoints. We
define \emph{length} of an arc as the number of its internal
non-isolated vertices plus one.
\begin{theorem}\label{10.1.1} The bracket monomials, corresponding
to the Rumer diagrams from $R(m)$ span the $\mathbb{C}$-linear space
$\inv_m^{\left(n\right)}$.
\end{theorem}
\begin{proof} We proceed by induction on $m$. The case $m=1$ holds
because the $\mathbb{C}$-linear space $\inv_1^{\left(n\right)}$ is
spanned by the brackets $p_{ij}$, $1\leq i<j\leq n$, whose valence
schemes are exactly the Rumer diagrams from $R(1)$. Now, assume that
the statement is true for $m-1$. Let
\begin{equation}
p_{i_1j_1}\ldots p_{i_mj_m}\label{10.1.5}
\end{equation}
be the expression of the bracket monomial with valence scheme $G$
and without loss of generality we can assume that one of the arcs of
the edge $(i,j)$, $i=i_1$, $j=j_1$, is (one of) the arcs of $G$ with
minimal length. In case this minimal length is $1$, we represent the
bracket monomial $p_{i_2j_2}\ldots
p_{i_mj_m}\in\inv_{m-1}^{\left(n\right)}$ as a linear combination of
bracket monomials in $\inv_{m-1}^{\left(n\right)}$, whose valence
schemes $G,G',\ldots$ are Rumer diagrams from $R(m-1)$. Multiplying
the last equality by $p_{ij}$, we obtain that $p_{i_1j_1}\ldots
p_{i_mj_m}$ is a linear combination of bracket monomials in
$\inv_m^{\left(n\right)}$, whose valence schemes are Rumer diagrams
from $R(m)$ because the attaching the edge $(i,j)$ to the Rumer
diagrams $G,G',\ldots$ produces Rumer diagrams. Now, let this
minimal length $g$ be at least $2$ and let $k$ be an internal
non-isolated point. Since there is no edge of $G$, which connects
$k$ to $i$ and $k$ to $j$, the edge of $G$ that passes through $k$
intersects $(i,j)$. Let $\ell$ be the other endpoint. We can suppose
$\{k,\ell\}=\{i_2,j_2\}$ and then, using the
relation~\eqref{5.5.10}, we obtain
\[
\pm p_{i_1j_1}\ldots p_{i_mj_m}=p_{ij}p_{k\ell}p_{i_3j_3}\ldots
p_{i_mj_m}=
\]
\[
p_{ik}p_{j\ell}p_{i_3j_3}\ldots
p_{i_mj_m}-p_{i\ell}p_{jk}p_{i_3j_3}\ldots p_{i_mj_m}.
\]
Thus, the bracket monomial~\eqref{10.1.5} is a linear combination of
two bracket monomials such that their valence schemes have arcs with
length strictly less than $g$. By repeating this procedure a finite
number of times on the summands, we finish the proof.
\end{proof}
In the next two theorems we permit $n$ to assume also value $1$.
\begin{theorem}\label{10.1.10} The dimensions $N(m_1,\ldots,m_n)=\dim
\inv_{m_1,\ldots,m_n}^{\left(n\right)}$ satisfies the recurrence
relation
\[
N(m_1,\ldots,m_n,m_{n+1})=\sum_{\mu_n}N(m_1,\ldots,m_{n-1},\mu_n),
\]
\[
N(m_1)=\left\{\begin{array}{ccc}
1 & \hbox{\rm\ if\ } & m_1=0 \\
0 & \hbox{\rm\ if\ } & m_1\geq 1, \\
\end{array}\right.
\]
where $\mu_n\in\mathbb{N}_+$ runs through those members of the
sequence $m_n+m_{n+1},m_n+m_{n+1}-2,\ldots,|m_n-m_{n+1}|$, such that
$m_n,m_{n+1},\mu_n$ form an \emph{even triangle}, that is,
$\mu_n\leq m_n+m_{n+1}$, $m_n\leq \mu_n+m_{n+1}$, $m_{n+1}\leq
\mu_n+m_n$, and the sum $m_n+m_{n+1}+\mu_n$ is an even number.
\end{theorem}
\begin{proof} The group $\sl(2)$ acts on the polynomial ring
$\mathbb{C}[x^{\left(1\right)}]=
\mathbb{C}[x_1^{\left(1\right)},x_2^{\left(1\right)}]$ via the
rule~\eqref{5.5.1} and for any $k\in\mathbb{N}_+$ the homogeneous
components $\mathbb{C}[x^{\left(1\right)}]_k$ are stable. Therefore
we obtain an action of $\sl(2)$ on any $\mathbb{C}$-linear space
$\mathbb{C}[x^{\left(1\right)}]_k$ of dimension $k+1$, or, what is
the same, a representation of the group $\sl(2)$. According to, for
example,~\cite[Sec. 2, 2.3.1]{[20]}, the last representation is
irreducible. In particular,the group $\sl(2)$ acts trivially on
$\mathbb{C}[x^{\left(1\right)}]_0=\mathbb{C}$:
$\sigma\cdot\alpha=\alpha$ for all $\alpha\in\mathbb{C}$ and the
corresponding representation is the identical representation.
Given $m_1,m_2\in\mathbb{N}_+$, in accord with Clebsch-Gordan
formula (see, for example,~\cite[Sec. 7, 7.1.4, Ex. 2]{[20]}) we
have
\[
\mathbb{C}[x^{\left(1\right)}]_{m_1}\otimes
\mathbb{C}[x^{\left(1\right)}]_{m_2}=\oplus_{\mu}
\mathbb{C}[x^{\left(1\right)}]_\mu,
\]
where the index $\mu\in\mathbb{N}_+$ runs through all integers
$m_1+m_2,m_1+m_2-2,\ldots, |m_1-m_2|$. In other words, $m_1,m_2,\mu$
are side lengths of a triangle and, moreover, the perimeter of this
triangle is an even number (\emph{even triangle}). We denote this
symmetric ternary relation by $2\bigtriangleup(m_1,m_2,\mu)$ and can
write down for short
\begin{equation}
\mathbb{C}[x^{\left(1\right)}]_{m_1}\otimes
\mathbb{C}[x^{\left(1\right)}]_{m_2}=
\oplus_{2\bigtriangleup\left(m_1,m_2,\mu\right)}
\mathbb{C}[x^{\left(1\right)}]_\mu.\label{10.1.15}
\end{equation}
Note that the right hand side of decomposition~\eqref{10.1.15}
contains the identical representation if and only if $m_1=m_2$, and,
in this case all other summands are different from it.
Tensoring the equality~\eqref{10.1.15} by
$\mathbb{C}[x^{\left(1\right)}]_{m_3}$, the equality thus obtained
by $\mathbb{C}[x^{\left(1\right)}]_{m_4}$, etc., we obtain
inductively on $n$ that
\begin{equation}
\mathbb{C}[x^{\left(1\right)}]_{m_1}\otimes\cdots\otimes
\mathbb{C}[x^{\left(1\right)}]_{m_n}=\oplus_{\mu\geq 0}
N_\mu(m_1,\ldots,m_n)\mathbb{C}[x^{\left(1\right)}]_\mu
\label{10.1.20}
\end{equation}
where $N_\mu(m_1,\ldots,m_n)\geq 0$ and only for a finite number of
indices $\mu$ we have $N_\mu(m_1,\ldots,m_n)>0$. The corresponding
$\mathbb{C}[x^{\left(1\right)}]_\mu$ are the irreducible components
of the representation
$\mathbb{C}[x^{\left(1\right)}]_{m_1}\otimes\cdots\otimes
\mathbb{C}[x^{\left(1\right)}]_{m_n}$ and $N_\mu(m_1,\ldots,m_n)>0$
are their multiplicities.
Now let $n\geq 2$, let us write down equality~\eqref{10.1.20} for
$n-1$ and tensor it by $\mathbb{C}[x^{\left(1\right)}]_{m_n}$:
\begin{equation}
\mathbb{C}[x^{\left(1\right)}]_{m_1}\otimes\cdots\otimes
\mathbb{C}[x^{\left(1\right)}]_{m_n}=\oplus_{\mu\geq 0}
N_\mu(m_1,\ldots,m_{n-1})\mathbb{C}[x^{\left(1\right)}]_\mu\otimes
\mathbb{C}[x^{\left(1\right)}]_{m_n}. \label{10.1.25}
\end{equation}
The identical representation is contained in the right hand side
of~\eqref{10.1.25} if and only if $\mu=m_n$, and in this case we
have
\begin{equation}
N_{m_n}(m_1,\ldots,m_{n-1})=N_0(m_1,\ldots,m_n)=N(m_1,\ldots,m_n)
\label{10.1.30}
\end{equation}
for all $n\geq 2$.
Further, let us plug the equality
\[
\mathbb{C}[x^{\left(1\right)}]_\mu\otimes
\mathbb{C}[x^{\left(1\right)}]_{m_n}=
\oplus_{2\bigtriangleup\left(\mu,m_n,\kappa\right)}
\mathbb{C}[x^{\left(1\right)}]_\kappa
\]
in~\eqref{10.1.25}:
\[
\mathbb{C}[x^{\left(1\right)}]_{m_1}\otimes\cdots\otimes
\mathbb{C}[x^{\left(1\right)}]_{m_n}= \oplus_{\nu\geq 0}
N_\nu(m_1,\ldots,m_{n-1})
(\oplus_{2\bigtriangleup\left(\nu,m_n,\kappa\right)}
\mathbb{C}[x^{\left(1\right)}]_\kappa).
\]
Let us fix $\mu$ and set $\kappa=\mu$. Then
\[
\mathbb{C}[x^{\left(1\right)}]_{m_1}\otimes\cdots\otimes
\mathbb{C}[x^{\left(1\right)}]_{m_n}=
\]
\begin{equation}
(\sum_{2\bigtriangleup\left(\nu,m_n,\mu\right)}N_\nu(m_1,\ldots,m_{n-1}))
\mathbb{C}[x^{\left(1\right)}]_\mu
+\oplus_{\nu\neq\mu}\mathbb{C}[x^{\left(1\right)}]_\nu.\label{10.1.35}
\end{equation}
Comparing the equalities~\eqref{10.1.20} and~\eqref{10.1.35}, we
have
\[
N_\mu(m_1,\ldots,m_n)=
\sum_{2\bigtriangleup\left(m_n,\mu,\nu\right)}N_\nu(m_1,\ldots,m_{n-1}).
\]
Taking into account~\eqref{10.1.30}, we obtain
\[
N(m_1,\ldots,m_n,m_{n+1})=
\sum_{2\bigtriangleup\left(m_n,m_{n+1},\mu_n\right)}N(m_1,\ldots,m_{n-1},\mu_n).
\]
In case $n=1$ we have
\[
N(m_1)=\left\{\begin{array}{ccc}
1 & \hbox{\rm\ if\ } & m_1=0 \\
0 & \hbox{\rm\ if\ } & m_1\geq 1. \\
\end{array}\right.
\]
The proof is done.
\end{proof}
\begin{remark}\label{10.1.40} {\rm In quantum mechanics the
irreducible representation of the special linear group $\sl(2)$,
given by the formula
\[
\sl(2)\to\gl(\mathbb{C}[x_1,x_2]_m),\hbox{\
}\sigma\mapsto(f(x)\mapsto f(\sigma^{-1}\cdot x)),
\]
where $f\in\mathbb{C}[x_1,x_2]_m$, $x=\left(
\begin{array}{cccccccccccc}
x_1\\
x_2\\
\end{array}
\right)$, models the spin momentum $\frac{m}{2}$ of the electron
configuration of an atom of valence $m$.
Clebsch-Gordan formula
\[
\mathbb{C}[x_1,x_2]_{m_1}\otimes
\mathbb{C}[x_1,x_2]_{m_2}=\oplus_{\mu} \mathbb{C}[x_1,x_2]_\mu,
\]
where the index $\mu\geq 0$ runs through all integers
$m_1+m_2,m_1+m_2-2,\ldots, |m_1-m_2|$, reflects the following fact
from quantum theory: If the electron configuration of an atom of
valence $m_1$ interacts with the electron configuration of an atom
of valence $m_2$, then the resulting spin $\frac{\mu}{2}$ can assume
all values between $\frac{|m_1-m_2|}{2}$ and $\frac{m_1+m_2}{2}$.
}
\end{remark}
\begin{theorem}\label{10.1.45} {\rm (i)} The function
$\rho(m_1,\ldots,m_n)$ satisfies the recurrence relation from
Theorem~\ref{10.1.10}.
{\rm (ii)} One has $\rho(m_1,\ldots,m_n)=N(m_1,\ldots,m_n)$ for all
$m_1,\ldots,m_n\in\mathbb{N}_+$.
\end{theorem}
\begin{proof} {\rm (i)} Given a valence scheme
in $V(m_1,\ldots,m_n,m_{n+1})$, let $m_{n,n+1}$ be the number of
valence bonds that connect $n$ and $n+1$. We construct a valence
scheme in $V(m_1,\ldots,m_{n-1},\mu_n)$, where
$\mu_n=m_n+m_{n+1}-2m_{n,n+1}$ (so
$2\bigtriangleup(m_n,m_{n+1},\mu_n)$ holds) on the vertices
$1,\ldots,a$ in the following way:
(a) Omit the valence bonds that connect $n$ and $n+1$.
(b) Omit the vertex $n+1$ and attach its $m_{n+1}-m_{n,n+1}$
vertices left, to the vertex $n$ without changing their other
endpoints.
Thus, we obtain a map
\[
\psi\colon V(m_1,\ldots,m_n,m_{n+1})\to
\cup_{2\bigtriangleup\left(m_n,m_{n+1},
\mu_n\right)}V(m_1,\ldots,m_{n-1},\mu_n),
\]
which turns out to be surjective.
Indeed, given a valence scheme in $V(m_1,\ldots,m_{n-1},\mu_n)$ we
can find a valence scheme in $V(m_1,\ldots,m_{n-1},m_n,m_{n+1})$,
which is mapped onto the former, in the following way:
(A) Add an additional point $n+1$ to $1,\ldots,n$.
(B) Distribute the $\mu_n=m_n+m_{n+1}-2r$ valence bonds through $n$
where $0\leq r\leq\min(m_n,m_{n+1})$ (that is,
$2\bigtriangleup(m_n,m_{n+1},\mu_n)$) as follows: Leave $m_n-r$ in
number valence bonds attached to $n$ and attach $m_{n+1}-r$ in
number valence bonds to $n+1$.
(C) Add $r$ valence bonds which connect $n$ with $n+1$.
The inverse image $\psi^{-1}(G)$ of a valence scheme $G\in
V(m_1,\ldots,m_{n-1},\mu_n)$, where
$2\bigtriangleup(m_n,m_{n+1},\mu_n)$, contains at most
$\binom{\mu_n}{m_{n+1}-r}$ valence schemes from
$V(m_1,\ldots,m_n,m_{n+1})$.
Since $n$ and $n+1$ are neighbours (the arc $\widehat{n\hbox{\
}n+1}$ has length $1$), $\psi$ maps Rumer diagrams onto Rumer
diagrams (the construction in (b) does not produce intersecting
edges). Now, let $G\in R(m_1,\ldots,m_{n-1},\mu_n)$, where
$2\bigtriangleup(m_n,m_{n+1},\mu_n)$. The intersection
$\psi^{-1}(G)\cap R(m_1,\ldots,m_n,m_{n+1})$ has exactly one member
which can be constructed via the distribution process from (B) as
follows: Order the other endpoints of the $\mu_n$ valence bonds
through $n$ clockwise and attach the first $m_{n+1}-r$ of them to
$n+1$. Thus, the restriction
\[
\psi\colon R(m_1,\ldots,m_n,m_{n+1})\to
\cup_{2\bigtriangleup\left(m_n,m_{n+1},
\mu_n\right)}R(m_1,\ldots,m_{n-1},\mu_n)
\]
is a bijection, hence
\[
\rho(m_1,\ldots,m_n,m_{n+1})=
\sum_{2\bigtriangleup\left(m_n,m_{n+1},
\mu_n\right)}\rho(m_1,\ldots,m_{n-1},\mu_n).
\]
Moreover, we have
\[
\rho(m_1)=\left\{\begin{array}{ccc}
1 & \hbox{\rm\ if\ } & m_1=0 \\
0 & \hbox{\rm\ if\ } & m_1\in\mathbb{N}. \\
\end{array}\right.
\]
{\rm (ii)} This follows from part {\rm (i)}.
\end{proof}
\subsection{The number of Rumer digrams}
\label{10.5}
Let $\rho(n,m)$ be the number of Rumer diagrams on $n$ atoms with
$m$ valence bonds.
\begin{theorem}\label{10.5.1} If $m\in\mathbb{N}$, then one has
\[
\rho(n,m)=\left|\begin{array}{ccc}
\binom{m+n-1}{n-1} & \binom{m+n-2}{n-1} \\
\binom{m+n-1}{n-2} & \binom{m+n-2}{n-2} \\
\end{array}\right|.
\]
\end{theorem}
\begin{proof} In accord with Theorem~\ref{10.1.10} and
Theorem~\ref{10.1.45}, we have
\[
\rho(n,m)=\sum_{m_1+\cdots+m_n=2m}\rho(m_1,\ldots,m_n)=
\sum_{m_1+\cdots+m_n=2m}N(m_1,\ldots,m_n)=
\]
\[
\sum_{m_1+\cdots+m_n=2m}\dim \inv_{m_1,\ldots,m_n}^{\left(n\right)}=
\dim\inv_m^{\left(n\right)},
\]
therefore
\begin{equation}
\rho(n,m)=\dim\inv_m^{\left(n\right)}\label{10.5.5}
\end{equation}
and Theorem~\ref{5.10.10} finishes the proof.
\end{proof}
\begin{corollary} \label{10.5.10} The monomials that correspond to
the Rumer diagrams from $R(m)$ form a basis for the
$\mathbb{C}$-linear space $\inv_m^{\left(n\right)}$ for any
$m\in\mathbb{N}_+$.
\end{corollary}
\begin{proof} Because of Theorem~\ref{10.1.1}, the
monomials~\eqref{5.5.20} that correspond to the Rumer diagrams from
$R(m)$ span the $\mathbb{C}$-linear space $\inv_m^{\left(n\right)}$.
On the other hand,~\eqref{10.5.5} yields that their number is equal
to the dimension of $\inv_m^{\left(n\right)}$ for $m\in\mathbb{N}$.
Since $\inv_0^{\left(n\right)}=\mathbb{C}$ and $\rho(n,0)=1$, we are
done.
\end{proof}
|
\section{Introduction}
High-speed photodiodes are key components of a wide range of photonic systems such as microwave-photonic links \cite{Urick2011,Cliche2006}, opto-electronic oscillators (OEOs) \cite{Yao1996,Eliyahu2008}, photonic-based radar \cite{Ghelfi2014}, photonic signal processing \cite{Torres-Company2014,Valley2007}, high speed waveform generation \cite{Li2014}, and the generation of low-noise microwaves via optical frequency division (OFD) \cite{Fortier2011,Xie2017}. These systems take advantage of the low loss, low noise, and large bandwidth inherent in optical systems to provide improved performance over their purely electronic counterparts. As the optical-to-electrical converter, high fidelity operation of the high-speed photodiode is critical in limiting excess noise and signal distortion that would otherwise largely undermine the advantages of employing photonic techniques in microwave signal generation, dissemination, and processing.
In many cases, the noise and linearity of a photodetector cannot be considered separately. For example, the output microwave power of a photodetector will saturate with high optical power illumination, limiting the detector's operating range. This can in turn constrain the achievable signal-to-noise ratio (SNR) of a link, and can limit the photonic gain of microwave photonics systems. With the development of high saturation power photodiodes capable of generating microwave power approaching 2~W \cite{Xie2014}, increased SNR and all photonic gain \cite{Devgan2009} have become more feasible to a wider range of microwave photonic systems. Prior to the onset of saturation, nonlinearity is present in the form of frequency mixing, distorting the microwave output with the generation of harmonic and intermodulation distortion (IMD) products. While this can be used advantageously for frequency up- and down-conversion \cite{Cabon2010,Jaro2003,Fushimi2004}, nonlinear mixing generally reduces the spur free dynamic range of photonic links, and, in the case of ultrashort optical pulse detection, result in amplitude-to-phase conversion (\mbox{AM-to-PM}). The \mbox{AM-to-PM} nonlinearity has drawn considerable interest in low noise microwave generation by OFD, where the mode-locked laser's relative intensity noise (RIN) is converted to phase noise \cite{Bouchand2017,Taylor2011,Ivanov2005,Hu2017}, degrading the timing jitter of the microwave signal of interest. As the RIN of a train of ultrashort optical pulses can exceed the pulse-to-pulse timing jitter by orders of magnitude and \mbox{AM-to-PM} conversion coefficients can approach unity, \mbox{AM-to-PM} conversion can easily dominate all other noise sources in the generated microwave signal. As discussed in more detail below, \mbox{AM-to-PM} may be considered a special case of IMD where all of the microwave signals are harmonically related. Thus, studies of and solutions to the problem of \mbox{AM-to-PM} conversion apply to systems beyond those involving ultrashort pulse detection.
Existing solutions to the \mbox{AM-to-PM} problem in high-speed photodetectors can be separated into two general categories. One is to avoid direct detection of the ultrashort pulses, and instead synchronize a microwave oscillator to the optical pulse train \cite{Kim2006,Lessing2013,Peng2014}. This can be accomplished by way of a Sagnac interferometer containing an electro-optic modulator that compares the optical pulse time of arrival to the zero-crossings of the microwave oscillator. While this technique has demonstrated AM rejection as large as 60~dB \cite{Lessing2013}, the output frequency is limited to the microwave oscillator's capture range, and it hasn't demonstrated the extremely low phase noise capabilities of direct optical pulse detection \cite{Quinlan2013}. The other solution is to directly detect the optical pulses, but operate at an \mbox{AM-to-PM} ``null" point where the coefficient is near zero \cite{Xie2017,Taylor2011,Baynes2015}. By careful adjustments to the illuminating optical power and the photodetector bias voltage, 50~dB rejection of the amplitude onto the phase can be achieved. Here it is the extremely fine tuning, sometimes necessitating active feedback mechanisms \cite{Bouchand2017}, that makes relying on a null point operationally challenging. In this work, we show that charge-compensated modified uni-traveling carrier (\mbox{CC-MUTC}) photodiodes under short pulse illumination support 10~GHz generation with \mbox{AM-to-PM} coefficients below -50~dB over a photocurrent range of 40~mA, an improvement of several orders of magnitude over standard and advanced p-i-n photodiode designs \cite{Bouchand2017,Taylor2011}. This level of \mbox{AM-to-PM} rejection enables the utilization of these photodiodes without additional stabilization schemes and at higher microwave output powers than previously reported. Achieving these results required manipulating the nonlinear interplay between the photodetector and the external microwave circuit. By modifying the broadband complex impedance of the circuit following the photodiode, the system's \mbox{AM-to-PM} is drastically altered. Using such a scheme, we are able to generate 15~dBm (30~mW) of microwave power at 10~GHz with 53~dB rejection of amplitude noise onto the microwave phase. Importantly, this scheme can be readily adopted by other systems using high linearity photodiodes, including those that do not utilize ultrashort optical pulses
\section{Photodiode nonlinearity and cancellation}
In contrast with p-i-n photodiodes, MUTC photodiodes have demonstrated large saturation powers in part by separating the absorber into depleted and undepleted sections \cite{Li2004,Li2010,Nagatsuma2009}. \mbox{CC-MUTC} detectors further improve power handling by pre-distorting the electric field in the depletion region, resulting in microwave power generation exceeding 1~W under modulated-CW light illumination \cite{Xie2014,Li2011,Beling2016,Zhou2013}. This design is also beneficial to short pulse detection, where large peak photocurrents can shape the electric field and alter the detector's response. In p-i-n photodiodes, the induced electric field due to the photocarriers tends to broaden the photocurrent pulse as the average optical power is increased \cite{Diddams2009,Zhang2012}. However, as a consequence of the field pre-distortion and the photocarrier-induced fields in the undepleted regions in \mbox{CC-MUTCs}, the width of the current pulse slightly decreases at moderate average photocurrents, prior to the onset of saturation (where the current pulse width begins to significantly increase). This is depicted in Fig. 1(a). The changing impulse response impacts the \mbox{AM-to-PM} nonlinearity because the pulse ``center of mass" is photocurrent-dependent, leading to a coupling between amplitude and timing. Usually, we are concerned with how this impacts the phase of the microwave signal of interest, represented by a single harmonic of the pulse train repetition frequency. Since each harmonic individually carries the information of the optical pulse train timing \cite{vonderlinde1986}, measuring the \mbox{AM-to-PM} (and the timing jitter) on a single harmonic obviates the more difficult task of handling the full photocurrent spectrum. As illustrated in Figs. 1(a) and 1(b), the changing duration of the current pulse leads to the phase of a harmonic initially decreasing before increasing as the impulse width begins to broaden. The change in delay for a fractional change in the photocurrent is the \mbox{AM-to-PM} coefficient, illustrated in Fig. 1(d). Important parameters in the \mbox{AM-to-PM} curve are the coefficient's value at its peak, the average photocurrent at which the \mbox{AM-to-PM} crosses zero (the null point), and its slope at the zero crossing. These values determine how close to the minimum the photodiode should be operated for a given application, whether the \mbox{AM-to-PM} null is experimentally accessible, and the amount of microwave power that can be achieved with low \mbox{AM-to-PM} conversion. Evidently, the desired condition is a low coefficient value for the largest possible photocurrent and – when the intention is to operate at the null – a smaller slope at the zero crossing such that small changes in the photocurrent don't spoil the \mbox{AM-to-PM} coefficient.
\begin{figure}[htbp]
\centering
\includegraphics[width=4.5 in]{Fig1.pdf}
\caption{The inherent nonlinearity of the photodiode. The carrier dynamics are affected by the electric fields in the device, which change as a function of average photocurrent due to the photocarrier-generated fields. This leads to (a) average photocurrent dependent impulse response which translates into (b) photocurrent dependent microwave phase. (c) Shows the microwave delay as a function of photocurrent, converting (d) amplitude modulation to phase modulation. The width of the impulse response of the photodiode is depicted in (a) to have a minimum at a certain photocurrent, causing the existence of a turn-around point in the phase of the extracted microwave signal and a zero crossing of the \mbox{AM-to-PM} coefficient. This is emphasized by the circles in (c) and (d). (e) Frequency domain picture of the nonlinearity in a diode illuminated by a periodic optical pulse-train. All the tones generated are harmonically related, leading to a myriad of mixing products feeding into the same harmonic of the repetition frequency. The combined nonlinear phase-shifts to the harmonic of interest produce the \mbox{AM-to-PM} conversion shown in (d).}
\label{fig:false-color}
\end{figure}
Physics-based models of photocarrier transport of \mbox{CC-MUTCs} underpin the description above \cite{Hu2017,Xie2017a}. Of course, the particular carrier dynamics of different photodiode structures will vary. However, all photodiodes (MUTC and p-i-n) have displayed some photocurrent value where there is a sign change and a zero crossing in the \mbox{AM-to-PM} conversion value \cite{Taylor2011,Xie2017a}. Typically, the magnitude of the \mbox{AM-to-PM} coefficient varies by 30 to 40~dB as the average photocurrent is changed \cite{Taylor2011,Zhang2012}. The large \mbox{AM-to-PM} coefficient away from the null has led to feedback schemes \cite{Bouchand2017} where the optical power is kept at a constant level to maintain the photodiode's operation close to the null, as well as combining the outputs of two photodetectors with opposite sign nonlinearities \cite{Zhang2014}. As shown here, \mbox{CC-MUTC} detectors with optimized doping concentrations and layer thicknesses can reduce the variations in the carrier velocities, providing exceptionally low \mbox{AM-to-PM} over an extended photocurrent range.
For any photonic application, the photodetector is only one piece that will impact the overall system nonlinearity \cite{Urick2011}. We therefore also explore the importance of the microwave circuit following the detector. To this end, a frequency domain description, as shown in Fig. 1(e), is useful. Nonlinear distortion within the photodiode results in the generation of new frequencies as well as the mixing of microwave frequencies (collectively referred to as intermodulation distortion \cite{Lui1990,Pedro2003,Maas}). In the detection of ultrashort optical pulses, all the generated microwave signals are harmonically related, and \mbox{AM-to-PM} conversion may be viewed as the result of nonlinear mixing of the pulse train harmonics. As the photocurrent is varied, the amplitude and phase of the multitude of mixing products changes, altering the phase (and amplitude) of the microwave harmonic of interest. Additionally, like all diodes, photodiodes respond to changes in their bias voltage according to the exponential voltage-current relation. If microwave power is reflected back into the diode, the bias voltage is modulated thereby mixing the microwave fields directly generated by the optical illumination with the reflected signal \cite{Cabon2010,Jaro2003,Fushimi2004}. The newly generated microwave signal then adds to the original microwave from the optical field, resulting in a microwave phase shift at the output of the circuit. The strength of this effect must depend on the average photocurrent to modify the overall nonlinearity. See supplementary information for more details.
The significance of the nonlinear phase shift due to a reflected microwave signal depends on the strength of the photodiode's inherent nonlinearity and the size of the reflected signal. Experimentally, we find there is always some reflection of the repetition rate harmonics due to the difficulty in achieving perfect broadband impedance matching. The importance of impedance matching has been noted previously \cite{Bouchand2017,Ivanov2005}, although small unintentional reflections combined with the large native nonlinearity of the photodiodes resulted only in small shifts in the nonlinearity near an \mbox{AM-to-PM} null. However, in photodiodes with exceptionally high linearity such as \mbox{CC-MUTCs}, this effect can be much more prominent and great care should be taken with the complex impedance of the circuit following the photodiode. Here we use the additional nonlinearity to our advantage by purposefully reflecting microwave power back to the diode to modify its behavior, analogous to harmonic load pull optimization in high power transistors \cite{Stancliff1979,Ferrero2013}.
\section{Experiments and Results}
To measure the \mbox{AM-to-PM} coefficient of the photodiodes, we use the pulse-train from a mode-locked Er:fiber laser with a repetition rate of 208.33 MHz. This pulse-train is sent through four stages of fiberized asymmetric Mach-Zehnder interleavers (AMZI) with an arm on each stage having an additional half-period delay for the input pulse-train repetition frequency \cite{Haboucha2011,Jiang2011,Quinlan2014}. After four-stage pulse interleaving, this results in a 3.33~GHz pulse-train which is then sent through a final AMZI with a 100~ps (third-period) asymmetry, producing 3.33~GHz pulse-pairs separated by 100~ps. The final stage boosts the achievable power in the 10~GHz harmonic, which is the microwave frequency of interest. The use of optical pulse interleavers has the additional benefit of reducing the number of harmonics in the photocurrent spectrum to be manipulated, such that the predicted additional nonlinearity is more likely to be observed and the results are easier to interpret. The spectrum of the resulting optical pulse-train is filtered to 1 to 2 THz bandwidth and compressed to nearly Fourier-transform-limited pulse-width of $\sim$1~ps before being photodetected. The pulse-width is measured directly in front of the photodiode using a flip mirror to redirect the light to an autocorrelator. All \mbox{AM-to-PM} measurements presented here are performed on the pulse train harmonic at 10~GHz.
\begin{figure}[htbp]
\centering
\includegraphics[width=3.3 in]{Fig2.pdf}
\caption{Amplitude-to-phase conversion measurement setup. By using a microwave demultiplexer it is possible to have improved control over the reflections of each harmonic present in the circuit. The 3.33~GHz and 6.67~GHz harmonics are reflected using a controllable delay and a short circuit. In the case of the 10~GHz harmonic a fraction of it can also be reflected by using a circulator as shown in the dashed box. The signal is re-amplified in the loop, but the overall gain remains below unity to avoid oscillation. The pulse-train shown in this figure corresponds to the interleaved mode-locked laser pulse-train and is amplitude modulated by an acousto-optic modulator before impinging on the photodiode. Three separate demodulation systems of different type were used in these experiments to verify our results.}
\label{fig:expsetup}
\end{figure}
Just prior to photodetection, the optical pulse-train is amplitude modulated at a low frequency (200 kHz) with an acousto-optic modulator (AOM). The nonlinearity of the photodiode converts some of this amplitude modulation to phase modulation, as described above. To calculate the \mbox{AM-to-PM} coefficient we perform AM and PM demodulation of the 10~GHz microwave and compute the ratio of the measured sidebands. We report the amplitude to single-sideband phase (L(f)) conversion. Explicitly, if the depth of AM is measured in dBam${}^2$ and the PM depth in dBrad${}^2$, then the \mbox{AM-to-PM} coefficient reported here is given by $\textrm{PM} - \textrm{AM} - 3~\textrm{dB}$.
The photodiode's \mbox{AM-to-PM} coefficient was measured as a function of photocurrent both for the case when all the microwave reflections were terminated to the best of our ability, as well as for different chosen values of the reflected phase when one or more microwave signals are intentionally reflected. Observed changes in the \mbox{AM-to-PM} coefficient were consistent with our simple model (further elucidated in the Supplement). We observed similar results across multiple photodiodes and under varied conditions.
\begin{figure}[htbp]
\centering
\includegraphics[width=3.3 in]{Fig3.pdf}
\caption{Measured \mbox{AM-to-PM} coefficient with the mode-locked pulse-train. In this case the 3.33~GHz and 6.67~GHz harmonics are terminated to 50 $\Omega$ and a small amount of 10~GHz signal is reflected to the diode with a delay via a microwave circulator as shown in the setup in Fig. 2. Notice the shift of the \mbox{AM-to-PM} null to higher photocurrents for phases $>0$ in the reflected signal. Also notice that for phases $<0$ the \mbox{AM-to-PM} coefficient may show a minimum without going through zero or no minimum at all. Each of these cases can be qualitatively explained by our simple single diode mixing model in the Supplement.}
\label{fig:results1}
\end{figure}
To verify the simple model described in the Supplement we reflect a small amount of the 10~GHz harmonic back to the diode via a circulator as shown in the dashed box in Fig. 2, where it can mix with the 20~GHz harmonic and modify the nonlinear phase of the original 10~GHz tone. We then record the \mbox{AM-to-PM} coefficient as a function of photocurrent for several delays of the reflected signal. The introduced relative delays have been carefully calibrated up to an arbitrary overall offset. This overall phase offset has been estimated from a combination of S21 measurements and residual coplanar waveguide lengths. The results are shown in Fig. 3. For clarity, we chose a few representative curves which illustrate the kinds of effects that we expect to observe. For example, for a certain range of phases of the reflected signal, the photocurrent null is expected to appear at progressively higher photocurrents. In Fig. 3 this is represented in the 2$^\circ$ (squares), 30$^\circ$ (circles) and 58$^\circ$ (hourglasses) respectively. On the other hand, when the additional nonlinearity has the same sign as the photodiode's native nonlinearity, they add constructively, and the photodiode performance deteriorates. This is clearly seen in the case of the -54$^\circ$ data shown with diamonds in Fig. 3. Finally, there is the possibility that the additional nonlinearity improves the performance over a small photocurrent range, but it does not allow the existence of a zero crossing. The triangles in Fig. 3 illustrate this case.
\begin{figure}[htbp]
\centering
\includegraphics[width=3.3 in]{Fig4.pdf}
\caption{\mbox{AM-to-PM} coefficient obtained by ``best effort" termination and by optimizing the reflection phase of the 3.33~GHz and 6.67~GHz harmonics to obtain the lowest AM-to-PM coefficient at 24 mA.}
\label{fig:results2}
\end{figure}
More importantly, we are interested in optimizing the reflection phase to obtain a minimum in the \mbox{AM-to-PM} coefficient at higher photocurrent than under the case with minimized microwave reflections. To show the power of the technique, we set the bias voltage at -12~V and choose 25~mA average photocurrent for optimization. We then proceeded to optimize the reflected phase of both the 6.67~GHz and 3.33~GHz harmonics to minimize the \mbox{AM-to-PM} coefficient. The resulting \mbox{AM-to-PM} curve is shown in Fig. 4. The red circles show the case when we terminate each harmonic and the blue squares show the case when we control the reflection of the 6.67~GHz and 3.33~GHz harmonics. Important features to note are: 1) in the case with terminated harmonics the \mbox{AM-to-PM} coefficient remains below -48~dB up to $\sim$16~mA of photocurrent – already an unprecedented result which showcases the high linearity of the \mbox{CC-MUTC}, 2) relative to the terminated case the \mbox{AM-to-PM} coefficient at 24~mA is improved by $\sim$25~dB due to the additional nonlinearity induced by the microwave reflections, and 3) the photocurrent range with \mbox{AM-to-PM} below -40~dB is extended from 20~mA in the terminated case to 30~mA when the reflection phase of the 3.33~GHz and 6.67~GHz harmonics is optimized. As mentioned above, the effect of reflected microwaves modifying the \mbox{AM-to-PM} coefficients has been observed before, but typically it consisted of a small shift in the location of the null. We observe more significant changes in the coefficient because the reflections have an effect when the inherent \mbox{AM-to-PM} coefficient of the photodetector is sufficiently close to zero. In \mbox{CC-MUTC} diodes, there is a wide photocurrent range where this condition is met.
\begin{figure}[htbp]
\centering
\includegraphics[width=3.3 in]{Fig5.pdf}
\caption{\mbox{AM-to-PM} coefficient and microwave power using a 10~GHz pulse-train. When a small amount of 10~GHz reflection is allowed, we can reduce the \mbox{AM-to-PM} coefficient and maintain it below -50~dB out to 40~mA of photocurrent (blue empty squares). The power measurement shown in red (right axis) is calibrated to the connector at the bias tee.}
\label{fig:results3}
\end{figure}
To test the photodiode under the most favorable conditions, we use a second optical pulse source consisting of a 10~GHz frequency comb generated via electro-optic modulation of continuous-wave light \cite{Beha2017,Metcalf2013}. The spectral width of this comb is $\sim$1 THz and the pulses are also compressed to nearly transform limited pulse-width of $< 2$~ps before being sent to the photodiode. Using a pulse-train directly at
10~GHz has the advantage of having a reduced energy-per-pulse, which extends the linear regime of the photodiode to higher photocurrents \cite{Hu2017}. When using the 10~GHz electro-optic frequency comb to measure the \mbox{CC-MUTC's} \mbox{AM-to-PM} coefficient, we find that the 20~GHz harmonic (uncontrolled in our current setup) plays a more prominent role. The importance of the uncontrolled 20~GHz harmonic rejection is indicated by the shape of the \mbox{AM-to-PM} curve in solid squares in Fig. 5 where the curve does not present a clear null. Similar to Fig. 3, the lack of a null in the \mbox{AM-to-PM} curve signifies the presence of nonlinear distortion due to microwave reflection. By controlling the reflection of a small fraction of the 10~GHz output, we obtain \mbox{AM-to-PM} coefficients below -50~dB out to 40~mA, as shown in Fig. 5. At such photocurrent, the microwave power in the 10~GHz harmonic at the bias tee is $\sim$15~dBm and we estimate that the microwave power at the photodiode surpasses 17~dBm. The possibility of producing such high microwave powers while operating with sufficient linearity enables interesting possibilities in microwave beam forming with fiber-fed remote antennas \cite{Frankel1997}.
\section{Comparisons and Discussion}
To place the extremely low nonlinearity level of \mbox{CC-MUTC} photodiodes in context, we have compared their performance to that of p-i-n photodiodes with both standard and high-linearity designs. Two representative cases \cite{Taylor2011,Joshi2008} are shown in Fig. 6. We emphasize the fact that not only are \mbox{CC-MUTC} photodiodes capable of operating with extremely low \mbox{AM-to-PM} coefficients, but they also do so over an extended photocurrent range, allowing their application in varied conditions without the need to fine tune or stabilize the optical power. Furthermore, the \mbox{AM-to-PM} can stay at low values out to $> 40~\textrm{mA}$ of photocurrent, an unprecedented number even compared to other MUTC designs.
\begin{figure}[htbp]
\centering
\includegraphics[width=3.3 in]{Fig6.pdf}
\caption{Nonlinear performance comparison between leading photodiode structures. The blue range shows values of \mbox{AM-to-PM} typically considered to be acceptable. Notice how in standard photodiodes this range is only reached in a narrow photocurrent range.}
\label{fig:comparison}
\end{figure}
One of the applications with the most stringent requirements on the photodetector linearity under ultrashort pulse illumination is the generation of low noise microwave signals using OFD. To assess the performance of these diodes, we must determine a reasonable requirement for the \mbox{AM-to-PM} coefficient. Taking as typical an Er:fiber based mode-locked laser as a frequency divider with a RIN plateau at $\sim$-140~dBc/Hz \cite{Chen2007}, a RIN-to-phase noise rejection of 40~dB is sufficient to support the division of state-of-the-art cavity stabilized lasers. With a RIN rejection of $> 50$~dB as demonstrated here, the \mbox{AM-to-PM} limits the phase noise performance only at -190~dBc/Hz, a level not yet demonstrated in any photodetected signal, thereby supporting next-generation experiments in microwave generation via OFD. After suppression of RIN-induced phase noise, other sources such as shot noise \cite{Quinlan2013} and photocarrier scattering in the photodiode \cite{Sun2014} limit the overall performance.
In the more general case of microwave photonic links, the nonlinear impairments inherent in the optical-to-electrical conversion can be improved with the methods demonstrated here. We expect that microwave reflections to the photodiode will impact all IMD products in microwave photonic links and not only when all signals are harmonically related. For example, in many microwave photonic links it is the third-order nonlinearity that is most important, since the second-order sum or difference frequencies are generally out of band. However, controlled reflection of the second-order nonlinear products may help suppress third-order nonlinearities, improving link performance. Although we have focused on reducing the nonlinearity, reflecting back microwave signals can also increase the nonlinear response. This can be useful in optoelectronic mixing applications where a large photodetector nonlinearity is desired to perform up- or down-conversion of photonically encoded microwave signals \cite{Pan2010}. Finally, in OEOs, the \mbox{AM-to-PM} coefficient has been shown to degrade the OEO performance \cite{Eliyahu2008}. We expect that the high-linearity photodiodes and broadband impedance control demonstrated here will also impact the noise performance of OEOs.
In conclusion, we have shown that high linearity \mbox{CC-MUTC} photodiodes in conjunction with broadband impedance control of the external circuit can perform optical-to-electrical conversion with better than 50~dB rejection of \mbox{AM-to-PM} conversion over a very large range of photocurrents. When illuminated at a pulse repetition rate of 10~GHz,
15~dBm of microwave power is generated with an \mbox{AM-to-PM} coefficient of -53~dB. The generation of still higher power microwaves with high linearity appears feasible. This performance level is likely to impact photonic generation of low noise microwaves as well as microwave photonic links.
\textbf{Funding.} Defense Advanced Projects Agency (DARPA) (PULSE program); Naval Research Laboratory (NRL); National Institute of Standards and Technology (NIST)
\textbf{Acknowledgments.} We thank N. Orloff and J. Booth for assistance in the early stage of this work, D. Carlson for the use of the 10~GHz comb source, and D. Nicolodi and A. Kowligy for useful comments on the manuscript. This work is a contribution of an agency of the US government and not subject to copyright in the USA.
\bibliographystyle{IEEE}
|
\section{Introduction}
In this paper, we consider the decomposition of spectral flow for a path of self-adjoint Fredholm operators. Let ${\cal H}$ be a separable Hilbert space, and we denote by
$\cal{FS} ({\cal H})$ be the set of all densely defined self-adjoint Fredholm operator on ${\cal H} $. We always equipped $\cal{FS} ({\cal H})$ with the gap topology.
For a continuous path $A(s)\in \mathcal{FS}({\cal H})$, $t\in[a,b]$. The spectral flow $sf(A(t); t\in[a,b])$ is an integer that counts the net number of eigenvalues that change sign. This notation is first introduced by Atiyah-Patodi-Singer \cite{APS76} in their study of index theory on manifolds with boundary, since then it had found many significant applications, see \cite{ZL99,BLP05} and reference therein.
Some basic property of spectral flow such as homotopy invariant, path additivity, direct sum e.t. are well known, please refer the Appendix. Our first result is another basic property which is called cogredient invariant property of spectral flow. For convenience, we first introduce some notations.
Let ${\cal H}_1, {\cal H}_2$ be separable Hilbert space, we denote by $\cal{L}({\cal H}_1, {\cal H}_2)$ and $\cal{C}({\cal H}_1, {\cal H}_2)$ the set of bounded and closed operators from ${\cal H}_1\to {\cal H}_2$. We let $\cal{S}({\cal H})$ be the set of self-adjoint operators on ${\cal H}$.
For convenience, we denote by $\cal{L}^*({\cal H}_1, {\cal H}_2), \cal{C}^*({\cal H}_1, {\cal H}_2), \cal{S}^*({\cal H}), \cal{FS}^* ({\cal H})$ be the invertible subsets.
\begin{thm}\label{th:coginva} Let
$M_s\in C([a,b], \cal{L}^*({\cal H}_1, {\cal H}_2))$, $A_s\in C([a,b], \mathcal{FS}({\cal H}_2)) $, then
\begin{eqnarray} M_s^*A_sM_s\in C([a,b], \mathcal{FS}({\cal H}_1)), \label{th1f1}\end{eqnarray} and we have
\begin{eqnarray} sf(A_s; s\in[a,b])=sf(M_s^*A_sM_s; [a,b]). \label{sfcog} \end{eqnarray}
\end{thm}
In a preprint paper \cite{FPS06}, Fitzpatrick-Stuar-Pejsachowicz proved \eqref{sfcog} in the case that $M_s$ is constant, the domain of $A_s$ is fixed and
both $A_a, A_b$ are invertible. Theorem \ref{th:coginva} can be consider as a generalization of their result.
Our second main result is the decomposition formula based on the cogredient invariant property. Let ${\cal H}_i$ be closed subspace of ${\cal H}$ for $i=1,\cdots,m$, then we define $$ \sum_{1\leq i\leq m} {\cal H}_i={\cal H}_1+\cdots {\cal H}_m $$
which is the subspace spanned by ${\cal H}_i$, $i=1,\cdots,m$.
Suppose $g\in\cal{L}({\cal H})$, we call $g$ is a matrix-like operator if
$\sigma(g)=\{\lambda_1,\cdots,\lambda_n\}$ is finite and there exist $m>0$, such that
\begin{eqnarray} {\cal H}=\sum_{1\leq i\leq n}\ker(g-\lambda_i)^m. \end{eqnarray}
We denote by $\mathcal{M}({\cal H})$ be the set of matrix-like operators. For $g\in\cal{M}({\cal H})$, $\lambda\in \sigma(g) $, we set $${\cal H}_\lambda:=\ker(g-\lambda)^m, $$
and denote \begin{eqnarray} F_\lambda=\left\{\begin{array}{cc} {\cal H}_\lambda, \quad if\quad \lambda\in \mathbb{U};
\\ {\cal H}_\lambda+{\cal H}_{\bar{\lambda}^{-1}}, \quad \quad if \quad \lambda \notin \mathbb{U}. \end{array}\right.\end{eqnarray}
Then we have
\begin{eqnarray} {\cal H}=\sum_{1\leq i\leq k}F_{\lambda_i}. \nonumber \end{eqnarray}
Moreover, let $\hat{F}=span\{F_\lambda, \lambda\in\sigma(g)\cap\mathbb{U}^c \} $, we have the next theorem.
\begin{thm}\label{th:decom} Suppose $g\in \cal{M}({\cal H}) $ is invertible and preserve the domain of $A$, $\sigma(g)\cap \mathbb{U}=\{\lambda_1,\cdots,\lambda_j \}$. Assume \begin{eqnarray} g^* A_sg=A_s,\quad for \quad s\in[0,1], \label{coinvariant} \end{eqnarray} then we have
\begin{eqnarray} sf(A_s)=sf(A_s|_{F_{\lambda_1}})+\cdots+sf(A_s|_{F_{\lambda_j}})+\frac{1}{2}(dim\ker (A_1|_{\hat{F}}) -dim\ker (A_0|_{\hat{F}})) . \label{decomf} \end{eqnarray}
\end{thm}
In \cite{HS09}, by assume $g$ is unitary and $\sigma(g)$ is finite, Hu-Sun proved the decomposition formula
\begin{eqnarray} sf(A_s)=sf(A_s|_{\ker(g-\lambda_1})+\cdots+sf(A_s|_{\ker(g-\lambda_j)}) \label{decomfHS} \end{eqnarray} under the condition \begin{eqnarray} A_sg=gA_s. \end{eqnarray}
Obviously, we give a generalization of \eqref{decomfHS}. In fact, there is a significant difference is we are not assume $g$ is unitary in Theorem \ref{th:decom}, hence the subspace ${\cal H}_\lambda$ is not orthogonal. To overcome this difficulty, we develop a new technique (Lemma \ref{abstract_decomposition}) to prove the equality of spectral flow.
As an applications of Theorem \ref{th:decom}, we give generalization for the Bott-type iteration formula which is a powerful tool in study the multiplicity and stability of periodic orbits in Hamiltonian systems.
In 1956, Bott got his celebrated iteration formula for the Morse index of closed geodesics \cite{Bot56}, and it was generalized by \cite{BTZ82, CD77, CZ84, Eke90}. The precise iteration formula of the general Hamiltonian system was established by Long \cite{Lon99,Lon02}. In fact, the iteration could be regarded as a unitary group action. Motivated by the symmetry orbits in $n$-body problem \cite{FT04}, Hu-Sun \cite{HS09} use this opinion to give generalization of Bott-type iteration formula to the system under a circle-type symmetry or brake symmetry group action, and prove the stability of Figure Eight orbit \cite{CM00}. The case of the brake symmetry was deeply studied in \cite{LZZ06,LZ14a,LZ14b,HPY17b}.
Based on Theorem \ref{th:decom}, we prove the Bott-type iteration formula which cover all the previous cases and moreover give some new generalizations. Our generalized formula could be applied to the closed geodesics on Semi-Riemanian manifold and heteroclinic orbits with brake symmetry.
Now we consider the linear Hamiltonian system \begin{eqnarray} \dot{x}(t)=JB(t)x(t),\quad t\in\mathcal{I}, \label{ham1}\end{eqnarray}
where $J=\begin{pmatrix}
0 & -I_n\\
I_n & 0
\end{pmatrix}$, $\mathcal{I}\subset \mathbb{R}$ is a connected subinterval, $B(t)\in C(\mathcal{I}, \mathcal{S}(\mathbb{R}^{2n}))$.
In the case $\mathcal{I}$ is finite, the boundary conditions is given by the Lagrangian subspaces. Let $(\mathbb{C}^{2n},\omega)$ be the standard symplectic
space with $\omega(x,y)=(Jx,y)$. A Lagrangian subspace $V$ is a $n$-dimensional subspace with $\omega|_V=0$. We denote the set of Lagrangian subspace by $Lag(2n)$. It is obvious $(\mathbb{C}^{2n}\oplus\mathbb{C}^{2n},-\omega\oplus\omega)$ is a $4n$-dimensional symplectic space, then for $\mathcal{I}=[a,b]$, the boundary condition is given by
\begin{eqnarray} (x(a),x(b))\in \Lambda\in Lag(4n). \end{eqnarray}
In the case $\mathcal{I}=\mathbb{R}$, we always assume $B(\pm\infty)=\lim_{t\to\pm\infty}B(t)$ exist and $JB(\pm\infty)$ is hyperbolic, that is \begin{eqnarray} \sigma(JB(\pm\infty))\cap i{\mathbb R}=\emptyset. \end{eqnarray}
Let ${\cal H}=L^2(\mathcal{I}, \mathbb{C}^{2n})$ and $E$ is $W^{1,2}(\mathcal{I}, \mathbb{C}^{2n})$ which satisfied some boundary conditions. We denote by $$
A:= -J\frac{d}{dt} : E \subset {\cal H}\to {\cal H}.
$$
and $B\in\mathcal{{\cal H}}$ be the multiplicity operator of $B(t)$. Let $A_s=A-sB$ for $s\in{\mathbb R}$, then $A_s\in\mathcal{FS}({\cal H})$.
For $g\in\mathcal{{\cal H}}$, $gE=E$ and satisfied \begin{eqnarray} g^*Ag=A, \quad g^*Bg=B, \end{eqnarray} then $g^*A_sg=A_s$. By construct $g$, we get the spectral flow decomposition of $sf(A_s)$. We list $6$-cases which common in applications of Hamiltonian systems. Our results generalization all the previous results, especially for the brake symmetry of heteroclinic orbits (Case $5$ and $6$), our result is new. Please see Section 4 for the detail.
It is well known that spectral flow is equal to Maslov index, and this is also true for the unbounded domain, see \cite{CLM94,ZL99,RS95, CH07, HP17} and reference therein. The Maslov index is associated integer to a pair of continuous path $f(t)=(L_1(t), L_2(t))$, $t\in\mathcal{I}$, in $Lag(2n)\times Lag(2n)$ \cite{CLM94}. From the decomposition of spectral flow, we get the decomposition of Mslov index, please refer Section 5 for the detail. For reader's convenience, we give a brief describe for the Maslov index and spectral flow in the Appendix.
This paper is organized as follows. We proved Theorem \ref{th:coginva} in Section 2 and Theorem \ref{th:decom} in Section 3. In Section 4, we list $6$-cases of decompositions in Hamiltonian systems. In Section 5, we give the some case of the Bott-type iteration formulas. At last, we briefly review the basic property of spectral flow and Maslov index in the Section 6.
\section{Spectral flow is preserved under cogredient }
Let $V$ be a closed space of ${\cal H}$, and
$P_V$ be the orthogonal projection from ${\cal H}$ to $V$.
For $A\in\cal{C}({\cal H})$, we denote the operator $P_VAP_V:V\to V$ by $A_V$. Then $A_V\in \cal{C}(V)$. Obviously, if $A\in \cal{S}({\cal H})$ then $A_V\in \cal{S}(V)$.
\begin{defi}
Let $A:[a,b]\rightarrow \cal{FS} ({\cal H})$ be a continuous curve. We call $A(t)$ is a positive curve if
$\{t, \ker A(t)\neq0\}$ is a distinct set and
\begin{eqnarray} sf(A(t); [0,1])=\sum_{a< t\leq b} \hbox{\rm dim$\,$} \ker(A(t)).\end{eqnarray}
\end{defi}
Let $A\in \cal{FS}({\cal H})$ and $B\in \cal{L}({\cal H})\cap \cal{S}({\cal H})$,
then $A+tB\in \cal{FS}({\cal H})$ with $t\in {\mathbb R}$.
Note that it is positive if $B|_{\ker(A+tB)>0}$ for any
$t\in \{t|\ker(A+tB)\neq 0\}$.
For example, $A+t I$ is a positive curve with $A\in \cal{FS}(H)$ for $t\in{\mathbb R}$.
Let $S\subset\cal{FS}({\cal H})$ be a path connected subset.
We assume there exist $K\in\cal{L}({\cal H})$ such that $(Kx,x)>0,\, \forall x\in {\cal H}$ and for any $A\in S$, there is a neighborhood $U$ of $A$, and $\epsilon>0$ such that $B+tK\in S, \, t\in [0,\epsilon]$ for each $B\in U$.
Then $A+tK,\, t\in [0, \epsilon] $ is a positive curve in $\cal{FS}({\cal H})$.
Let $\{{\cal H}_k\}, 1\leq k\leq n$ be a family of Hilbert spaces and $f_k: S\rightarrow \cal{FS}({\cal H}_k)$ be a family of continuous maps.
We assume that
(a)
For any $ A\in S$, $f_k(A+tK),t\in [0,\epsilon]$ is a positive path in $\cal{FS}({\cal H}_k)$.
(b)
For any $A\in S$, $\sum_{1\leq k\leq n}\hbox{\rm dim$\,$}\ker f_k(A)=\hbox{\rm dim$\,$}\ker A $.
\\
Then we have the following lemma.
\begin{lem}\label{abstract_decomposition}
Let $A \in C([0,1], \mathcal{FS}({\cal H}))$ and satisfied condition (a) and (b),
we have
\begin{equation}
sf(A(t); t\in [0,1])=\sum_{1\leq k\leq n} sf (f_k(A(t)); t\in [0,1]). \label{eq2.2}
\end{equation}
\end{lem}
\begin{proof}
Since the spectral flow satisfied the Path additivity property,
we only need to prove \eqref{eq2.2} locally.
Let $h_k(s,t)=f_k(A(t)+s K)$, $t\in [0,1],s\in [0,\epsilon]$,
then for any $t\in [0,1]$, $h_k(s,t)$ is a positive curve with $1\leq k\leq n$. Let $t_0\in [0,1]$, since $(Kx,x)>0$ for $x\in \ker A(t_0)$, there is $\delta>0$ such that
\begin{equation*}
\hbox{\rm dim$\,$}\ker(A(t_0)+\delta K)=0.
\end{equation*}
It follows that $\hbox{\rm dim$\,$}\ker (h_k(\delta,t_0))=0$ for $1\leq k\leq n$.
Note that $A(t_0)+\delta K$ is a Fredholm operator,
so there is $\delta_1>0$ such that
\begin{eqnarray} \hbox{\rm dim$\,$} \ker( A(t)+\delta K) =0,\, \forall t\in [t_0-\delta_1,t_0+\delta_1].\nonumber\end{eqnarray}
It follows that $\hbox{\rm dim$\,$}\ker (h_k(\delta,t))=0$ for $t\in[t_0-\delta_1,t_0+\delta_1]$, $1\leq k\leq n$.
Then we have
\begin{equation*}
\begin{cases}
sf(A(t)+\delta K, t\in [t_0-\delta_1,t_0+\delta_1])=0\\
sf(h_k(\delta,t), t\in [t_0-\delta_1,t_0+\delta_1])=0
\end{cases}.
\end{equation*}
By homotopy invariance of spectral flow, we have
\begin{eqnarray}
sf(A(t); t\in [t_0-\delta_1,t_0+\delta_1])=sf(A(t_0-\delta_1+sK); s\in [0,\delta])-sf(A(t_0+\delta_1+sK); s\in [0,\delta]) \nonumber \end{eqnarray}
and
\begin{eqnarray}
sf(h_k(0,t); t\in [t_0-\delta_1,t_0+\delta_1])=sf(h_k(s,t_0-\delta_1); s\in [0,\delta])-sf(h_k(s,t_0+\delta_1); s\in [0,\delta]).\nonumber
\end{eqnarray}
Note that $A(t_0\pm\delta_1)+sK$, $h_k(s,t_0\pm\delta_1),1\leq k\leq n$ are positive paths.
It follows that
\begin{eqnarray}
sf(A(t_0\pm \delta_1)+sK; s\in [0,\delta])=\sum_{0<s\leq \delta }\hbox{\rm dim$\,$}\ker (A(t_0\pm \delta_1)+sK)\nonumber\\
=\sum_{0<s\leq \delta } \sum_{1\leq k\leq n} \hbox{\rm dim$\,$}\ker (h_k(s,t_0\pm\delta_1)\nonumber)\\
=\sum_{1\leq k\leq n}sf (h_k(s,t_0\pm\delta_1); s\in [0,\delta]). \nonumber\end{eqnarray}
This complete the proof.
\end{proof}
Please note that Lemma \ref{abstract_decomposition} can be consider as a generalization of Direct sum property of spectral flow.
In the next, we will prove the spectral flow is invariant under the cogredient. The next Lemma is contained in \cite{FPS06}, but for reader's convenience, we give details here.
\begin{lem}
Let $E$ be the domain of $A\in\cal{FS}({\cal H}_2)$.
If $M\in \cal{L}({\cal H}_1,{\cal H}_2)$ is invertible, then $M^*AM\in \cal{FS}({\cal H}_1)$ with domain $M^{-1}(E)$.
\end{lem}
\begin{proof}
Since $A\in\cal{FS}({\cal H}_2)$ with domain $E$, we have $\hbox{\rm dim$\,$} \ker A,\, \hbox{\rm dim$\,$} ({\cal H}_2/\image A) <+\infty$.
Since $M$ is invertible, we have $\ker(M^*AM)= \ker(AM)=M^{-1}\ker A$.
Then $M^{-1}$ induce an isomorphism from $\ker A$ to $\ker (M^*AM)$.
Note that $\image(M^*A M)=M^*\image(A)$.
Then $M^*$ induce an isomorphism from ${\cal H}_2/\image(A)$ to ${\cal H}_1/\image(M^*AM)$.
So $M^*AM$ is a Fredholm operator.
Since $A \in \cal{FS}({\cal H}_2)$ with domain $E$, we see that for each $x\in M^{-1}$, $(AMx,My)=(Mx,AMy)$ if and only if $y\in M^{-1}E$.
It follows that $(M^*AM)^*=M^*AM$ with domain $M^{-1}E$.
Then we can conclude that $M^*AM\in \cal{FS}({\cal H}_1)$.
\end{proof}
Recall that the gap topology can be induced by the gap distance $\hat{\delta}$.
Let $X$ be a Banach space.
Let $M,N$ be two closed linear spaces of $X$.
Denote by $S_M$ the unit sphere of $M$.
Then gap distance is defined as
\begin{equation}
\hat{\delta}(M,N)=\max \{\delta(M,N),\delta(N,M)\},
\end{equation} where
\begin{equation*}
\delta\{M,N\}:=\begin{cases}
\sup_{u\in S_M}\Dist(u,N), &\text{if} ~ M\neq \{0\}\\
0, & \text{if}~ M=\{0\}
\end{cases}.
\end{equation*}
The gap distance have the following properties.
\begin{lem}\label{lm:gap_distance_compare}
Let $X,Y$ be two Hilbert spaces.
Let $M,N$ be two closed linear subspaces of $X$.
Let $P,Q\in \cal{L}^*(X,Y)$.
Then
$\hat\delta(PM,QN)\le \hat\delta(M,N)\max\{\|P\|,\|Q\|\}+\|P-Q\|\max\{\|P^{-1}\|,\|Q^{-1}\|\}$.
\end{lem}
\begin{proof}
Without loss of generality, we assume that $M,N\neq \{0\}$, and
let $d_1=\hat\delta(M,N)$.
Let $x\in PM$ with $\|x\|=1$, we choose
$y\in N$ such that $\|P^{-1}x-y\|=\Dist(P^{-1}x,N)$,
then we have $\|y\|\leq \|P^{-1}x\|\leq \|P^{-1}\|$.
Note that
\begin{eqnarray*}
\|x-Qy\|\leq \|x-Py\|+\|Qy-Py\| &&\leq \|P\|\|P^{-1}x-y\|+\|Q-P\|\|y\|\\
&&\leq \|P\|\Dist(P^{-1}x,N)+\|Q-P\|\|P^{-1}\|\\
&&\leq \|P\|\delta(M,N)+\|Q-P\|\|P^{-1}\|.
\end{eqnarray*}
It follows that
$\delta(PM,QN)\leq \|P\|\delta(M,N)+\|Q-P\|\|P^{-1}\|$.
Similarly, we have
$\delta(QN,PM)\leq \|Q\|\delta(N,M)+\|Q-P\|\|Q^{-1}\|$.
This conclude the proof.
\end{proof}
\begin{lem}\label{lem: continuous} Suppose $M_s\in C([0,1], \cal{L}^*({\cal H}_1,{\cal H}_2))$, $A_s\in C([0,1],\cal{FS}({\cal H}_2))$, then $M_s^*A_sM_s\in C([0,1],\cal{FS}({\cal H}_1)) $. \end{lem}
\begin{proof}
We only need to show that $M_s^*A_sM_s$ is a continuous curve with the gap topology.
Let $E_s$ be the domain of $A_s$.
Note that
$$
\Graph(M_s^*A_sM_s)=\{(M_s^* A_s x,M_s^{-1}x)|x\in E_s\}.
$$
Let $Q_s:{\cal H}_2\oplus {\cal H}_2 \to {\cal H}_1\oplus {\cal H}_1$ be
$\begin{pmatrix}
M_s^* & 0\\
0& M_s^{-1}\\
\end{pmatrix} $,
then $Q_s \in C([0,1], \cal{L}^*({\cal H}_2\oplus {\cal H}_2,{\cal H}_1\oplus{\cal H}_1))$, and
we also have $\Graph(M_s^*A_sM_s)=Q_s\Graph(A_s)$.
Since $\|Q_s\|$ and $\|Q_s^{-1}\|$ are continuous functions on $[0,1]$,
we have $\|Q_s\|>0$, $\|Q_s^{-1}\|>0$ for $s\in [0,1]$.
Let $C_1=\sup(\|Q_s\|)$, $C_2=\sup(\|Q_s^{-1}\|)$.
For $s_0,s\in [0,1]$,
by Lemma \ref{lm:gap_distance_compare}, we have
\begin{eqnarray*}
\hat\delta(\Graph(M_{s_0}^*A_{s_0}M_{s_0},M_{s}^*A_{s}M_{s})&&=
\hat\delta(Q_{s_0}\Graph(A_{s_0}),Q_s\Graph(A_s))\\
&&\leq C_1\hat\delta(A_{s_0},A_{s})+C_2\|Q_{s}-Q_{s_0}\|.
\end{eqnarray*}
By the continuity of $A_s$ and $Q_s$, we see that
for any $\epsilon >0$ there is $\delta_1>0$ such that for any $s\in (s_0-\epsilon,s_0-\epsilon)$, we have
$\hat\delta(\Graph(A_{s_0}),\Graph(A_s))<\epsilon/(2C_1)$ and
$\|Q_s-Q_{s_0}\|<\epsilon/(2C_2)$.
Then we have $\hat\delta(\Graph(M_{s_0}^*A_{s_0}M_{s_0},M_{s}^*A_{s}M_{s})<\epsilon.$
This complete the proof.
\end{proof}
Now we give the proof of Theorem \ref{th:coginva}.
\begin{proof}[Proof of Theorem \ref{th:coginva}]
Please note that \eqref{th1f1} is from Lemma \ref{lem: continuous}.
We first prove the case $M_s\equiv M$.
Let $S=\cal{FS}({\cal H}_2)$, $K=I$, $f(A)=M^*AM$. Please note that
$$\hbox{\rm dim$\,$}\ker (M^*AM) =\hbox{\rm dim$\,$} (M^{-1}\ker A)=\hbox{\rm dim$\,$}\ker A$$ for each
$A\in \cal{FS}({\cal H}_2) $.
Furthermore, we have $\frac{d}{dt} M^*(A+tI)M|=M^*M>0$,
so $M^*(A+tI)M$ is a positive curve.
Then by Lemma \ref{abstract_decomposition}, we have
$$sf(A_s; s\in[a,b])=sf(M^*A_sM; s\in[a,b])\quad for \quad M\in \cal{L}^*({\cal H}_1,{\cal H}_2).$$
Now we consider the two family $M_{a+t(s-a)}^*A_sM_{a+t(s-a)}$, $(t,s)\in [0,1]\times[a, b]$. By the homotopy invariant property of spectral flow, we have
\begin{equation*}
sf(M_a^*A_sM_a)=sf(M_s^*A_s M_s)-sf(M_{a+t(b-a)}^*A_bM_{a+t(b-a)}).
\end{equation*}
Note that $\hbox{\rm dim$\,$} \ker M_{a+t(b-a)}^*A_bM_{a+t(b-a)}$ is a constant
which implies $sf(M_{a+t(b-a)}^*A_bM_{a+t(b-a)})=0$.
It follows that
$$sf(M_a^*A_sM_a)=sf(M_s^*A_s M_s).$$ This complete the proof.
\end{proof}
As an example, we consider the one parameter family of linear Hamiltonian systems
\begin{eqnarray} \dot{z}(t)= J B_s(t)z(t), (s,t)\in[0,1]\times[0,T], \label{ham2.1}
\end{eqnarray}
where $B(t)\in C([0,1]\times[0,T], \cal{S}({\mathbb R}^{2n}))$.
The boundary condition is given by
\begin{eqnarray} (x_s(0),x_s(T))\in \Lambda_s \in Lag(4n), \nonumber \end{eqnarray}
where we assume $\Lambda_s$ is continuous depend on $s$.
Let $A_s=-J{d\over dt}|_{E(\Lambda_s)}$ which is path of self adjoint operator on
${\cal H}:=L^2([0,T],\mathbb C^{2n})$with domain $$E(\Lambda_s)=\{x\in W^{1,2}([0,T], \mathbb C^{2n}), (x(0),x(T))\in\Lambda_s \}.$$
We define $B_s$ by $(B_sx)(t)=B_s(t)x(t)$. It is well known that
$A_s, A_s-B_s\in \cal{FS}({\cal H})$ with domain $E_s$. Let ${\gamma}_s(t)$ be the fundamental solution of \eqref{ham2.1}, i.e. \begin{eqnarray} \dot{{\gamma}}_s(t)=JB_s(t){\gamma}_s(t),\end{eqnarray}
then $${\gamma}_s(t)\in{\mathrm {Sp}}(2n):=\{P\in \cal{L}^*({\mathbb R}^{2n}), P^*JP=J \}, $$ which implies $Gr({\gamma}_s(T))\in Lag(4n ) $.
The following formula which gives the relation of spectral flow and Maslov index (please refer Theorem \ref{th:6.1})
\begin{eqnarray} -sf(A_s-B_s)=\mu(\Lambda_s, Gr(\gamma_s(T)) ). \nonumber \end{eqnarray}
Let $P_s(t)\in C^1([0,1]\times[0,T], {\mathrm {Sp}}(2n))$, then $P_s\in C^1([0,1], \cal{L}^*({\cal H}))$, hence $(P^{*}_s)^{-1}(A_s-B_s)P^{-1}_s\in\cal{FS}({\cal H})$ with domain
\begin{eqnarray} P_sE_s= \{x\in W^{1,2}([0,T], \mathbb C^{2n}), (x(0),x(T))\in\hat{P}_s(T)\Lambda_s \} , \nonumber \end{eqnarray} where
$\hat{P}_s(t)=diag(I_n, P_s(t))$.
Direct compute show that \begin{eqnarray} (P^{*}_s)^{-1}(-J{d\over dt}|_{E(\hat{P}_s(T)\Lambda_s)}-B_s)P^{-1}_s=A_s-\hat{B}_s, \nonumber \end{eqnarray}
where $\hat{B}_s(t)= -J\dot{P}_s(t)P^{-1}_s(t)+(P^{*}_s(t))^{-1}B(t)P^{-1}_s(t) $. From Theorem \ref{th:coginva}, we have
\begin{eqnarray} sf(-J{d\over dt}|_{E(\hat{P}_s(T)\Lambda_s)}-\hat{B}_s)= sf((P^{*}_s)^{-1}(A_s-B_s)P^{-1}_s) . \label{coninva1} \end{eqnarray}
From \eqref{th:morma}, we can express the left of \eqref{coninva1} as Maslov index. In fact,
the fundamental solution is $P_s(t){\gamma}_s(t)$, and the boundary conditions is given by $(\hat{P}_s(T)\Lambda_s $. Hence we have
\begin{eqnarray} sf(-J{d\over dt}|_{E(\hat{P}_s(T)\Lambda_s)}-\hat{B}_s)= \mu(\hat{P}_s(T)\Lambda_s, \hat{P}_s(T)Gr(\gamma_s(T)) ). \nonumber \end{eqnarray}
Formula \eqref{coninva1} implies that
\begin{eqnarray} \mu(\Lambda_s, Gr(\gamma_s(T)) ) =\mu(\hat{P}_s(T)\Lambda_s, \hat{P}_s(T)Gr(\gamma_s(T)) ), \nonumber \end{eqnarray}
which is just the symplectic invariant property (\ref{adp1.4}) of Maslov index.
\section{Decomposition of Spectral flow under cogredient invariant }
In this section, we will prove the decomposition formula for spectral flow. Suppose $g\in\cal{M}({\cal H})$ with
$\sigma(g)=\{\lambda_1,\cdots,\lambda_n\}$, then
\begin{eqnarray} {\cal H}=\sum_{1\leq i\leq n}{\cal H}_{\lambda_i}, \label{3.1a}\end{eqnarray}
where ${\cal H}_{\lambda_i}:=\ker(g-\lambda_i)^m $ for large enough $m$.
Note that $(\lambda-\lambda_1)^m$ and $(\lambda-\lambda_2)^m$ are coprime,
then there are polynomials $p_1,p_2$ such that $p_1(\lambda)(\lambda-\lambda_1)^m+p_2(\lambda)(\lambda-\lambda_2)^m=1$.
For each $x\in {\cal H}_{\lambda_1}\cap {\cal H}_{\lambda_2}$, we have
$$
x=p_1(g)(g-\lambda_1)^m x+p_2(g)(g-\lambda_2)^m x=0.
$$
Similarly we have ${\cal H}_{\lambda_i}\cap {\cal H}_{\lambda_j}=0$ with $i\neq j$. So the decomposition \eqref{3.1a} is a inner direct sum.
\begin{lem}
$g\in\mathcal{M}({\cal H})$ if and only if there exist $\lambda_1,\cdots, \lambda_n\in\mathbb{C}$ such that $\Pi_{i=1}^n (g-\lambda_i)^m=0$.
\end{lem}
\begin{proof}
We only need to show that $g\in\mathcal{M}({\cal H})$ if $\Pi_{i=1}^n (g-\lambda_i)^m=0$.
Let $G_l(\lambda) $ be the polynomial $\Pi_{i=1}^{l-1} (\lambda-\lambda_i)^m\Pi_{i=l+1}^n(\lambda-\lambda_i)^m$.
Then $G_1,G_2,\cdots,G_n$ are coprime polynomials.
It follows that there are polynomials $a_i(\lambda),(1\le i\le n)$, such that
$$
\sum_{i=1}^n a_i(\lambda)G_i(\lambda)=1.
$$
It follows that $\sum_{i=1}^n a_i(g)G_i(g)=\Id$.
Then we can conclude that
$${\cal H} =\sum_{1\leq i\leq n}G_i(g){\cal H}.$$
We also have $(g-\lambda_i)^mG_i(g){\cal H}=\Pi_{i=1}^n (g-\lambda_i)^m{\cal H}=0$, which implies \eqref{3.1a}.
\end{proof}
We have the following lemmas.
\begin{lem} \label{lm:A_orthogonal}
Let $A\in\cal{FS}({\cal H})$ with domain $E$. Suppose $g\in\cal{M}({\cal H})$, which satisfied $$g^*Ag=A,\quad gE=E, $$ then ${\cal H}_\lambda, {\cal H}_\mu$ is $A$-orthogonal if $\lambda\bar{\mu}\neq1$, i.e.
\begin{eqnarray} (Ax,y)=0,\quad if \quad x\in {\cal H}_\lambda\cap E,\quad y\in {\cal H}_\mu\cap E. \end{eqnarray}
\end{lem}
\begin{proof}
Let $x\in \ker(g-\lambda)^m \cap E$, $y\in \ker(g-\mu)^n\cap E$ with $m,n\ge 1$.
We see that $(Ax,y)=0$ if $m+n=2$. In fact, $$
(Ax,y)=(Agx,gy)=\lambda\bar\mu(Ax,y)
$$
implies $(Ax,y)=0$ since $\lambda\bar\mu\neq1$.
Assume that $(Ax,y)=0$ if $m+n\le k$.
Note that $(g-\lambda)x\in \ker(g-\lambda)^{m-1}\cap E$, $(g-\mu)x\in \ker(g-\mu)^{n-1}\cap E$, $gx\in \ker(g-\lambda)^{m}\cap E$ and $gy\in \ker (g-\mu)^m\cap E$.
If $m+n=k+1$,
We have
$$
(Ax,y)=(Agx,gy)=(A(g-\lambda)x,g y)+(A\lambda x,(g-\mu)y)+\lambda\bar{\mu}(Ax,y)=\lambda\bar\mu(Ax,y).
$$
Since $\lambda\bar\mu\neq 1$, we have $(Ax,y)=0$.
By induction, we have $(Ax,y)=0$ with $x\in {\cal H}_{\lambda}\cap E$ and $y\in {\cal H}_{\mu}\cap E$. This complete the proof.
\end{proof}
\begin{lem}\label{lm:ker_dom_decom}
Under the condition of Lemma \ref{lm:A_orthogonal}, then
$\ker A=\sum_{1\le i\le n}\ker A\cap {\cal H}_i$ and $E=\sum_{1\le i\le n}E\cap {\cal H}_{i}$.
\end{lem}
\begin{proof}
Note that $E$ is a invariant subspace of $g$.
Then $\Pi_{1\le i\le n}(g-\lambda_i)^m=0$ on $E$.
It follows that $E=\sum_{1\le i\le n}\ker (g|_E-\lambda_i)^m=\sum_{1\le i\le n}E\cap {\cal H}_{i}$.
We have $$g^*A g(\ker A)=A\ker A=0.$$ It follows that $g(\ker A)\subset \ker A$. So $\ker A$ is a invariant subspace of $g$. Similarly we have
$\ker A=\sum_{1\le i\le n}\ker A\cap {\cal H}_i$.
This complete the proof.
\end{proof}
For $A\in\cal{C}({\cal H})$, assume that ${\cal H}=\sum_{1\le i\le k}{\cal H}_i$ where
all of ${\cal H}_i$ are closed subspaces of ${\cal H}$.
Let $E$ be the domain of $A$ and
assume that $E=\sum_{1\le i\le k} E\cap {\cal H}_i$. ${\cal H}_i$, ${\cal H}_j$ are $A$-orthogonal if $i\neq j$.
Recall that we set \begin{eqnarray} F_\lambda=\left\{\begin{array}{cc} {\cal H}_\lambda, \quad if\quad \lambda\in \mathbb{U};
\\ {\cal H}_\lambda+{\cal H}_{\bar{\lambda}^{-1}}, \quad \quad if \quad \lambda \notin \mathbb{U}. \end{array}\right.\nonumber \end{eqnarray}
then we have
$ {\cal H}=\sum_{1\leq i\leq k}F_{\lambda_i} $
and $F_{\lambda_i}, F_{\lambda_j}$ are $A$-orthogonal if $i\neq j$.
Let $X=\bigoplus_{1\le i \le k} F_{\lambda_i}$,
we define an inner product on $X$:
$$
((x_1,x_2,\cdots,x_k),(y_1,y_2,\cdots,y_k))=\sum_{1\le i\le k} (x_i,y_i),
$$
where $(x_i,y_i)$ is the inner product in ${\cal H}$.
Then $X$ is a Hilbert space and the map $$ M: (x_1,x_2,\cdots,x_k)\to \sum_{1\le i \le k} x_i$$ is a homeomorphism from $X$ to ${\cal H}$.
Please note that $A|_{F_{\lambda_i}}$ is the map $M^*AM:M^{-1}F_{\lambda_i}\to M^{-1}( F_{\lambda_i})$. It is a self-adjoint Fredholm operator on $M^{-1}F_{\lambda_i}$ with domain $M^{-1}(E\cap F_{\lambda_i})$. It follows that
\begin{equation}
\ker(A|_{F_{\lambda_i}})=\ker(AM)\cap M^{-1}(F_{\lambda_i})=M^{-1}(\ker A\cap F_{\lambda_i}). \nonumber
\end{equation}
\begin{prop}\label{pro:decomposition}
Suppose $g\in \cal{M}^*({\cal H}) $, $A_s\in C([0,1], \cal{FS}({\cal H}))$ with fixed domain $E$ and $gE=E$. We assume $g^* A_sg=A_s$ of $s\in[0,1]$, then we have
\begin{eqnarray} sf(A_s)=sf(A_s|_{F_{\lambda_1}})+\cdots+sf(A_s|_{F_{\lambda_k}}). \end{eqnarray}
\end{prop}
\begin{proof}
By Theorem \ref{th:coginva}, we have $M^*A_sM\in C([0,1],\cal{FS}(X))$,
and
$$
sf(A_s)=sf(M^*A_sM).
$$
Note that $X=\bigoplus_{1\le i\le k} F_{\lambda_i}$ is an orthogonal decomposition.
By the Direct sum property of spectral flow, we have
\begin{equation*}
sf(A_s)=sf(M^*A_sM)=\sum_{1\le i\le k}sf(A_s|_{F_{\lambda_i}}).
\end{equation*} This complete the proof.
\end{proof}
\begin{lem} \label{lm:sf_notin_circle}
If $\lambda\notin\mathbb{U}$ then we have
\begin{eqnarray} sf(A_s|_{F_{\lambda}})=\frac{1}{2}(dim\ker (A_1|_{F_{\lambda}}) -dim\ker (A_0|_{F_{\lambda}}) ) . \end{eqnarray}
\end{lem}
\begin{proof}
Recall that $A_s|_{F_\lambda}$ is the operator
$M^*A_sM: M^{-1}(F_\lambda)\to M^{-1}(F_{\lambda}) $ and
$M^{-1}(F_{\lambda})=M^{-1}{\cal H}_\lambda+ M^{-1}{\cal H}_{\bar\lambda^{-1}}$.
We also have $M^{-1}{\cal H}_\lambda\perp M^{-1}{\cal H}_{\bar\lambda^{-1}} $.
Let $Q$ be the map $x+y\to -x+y$ with $x\in M^{-1}{\cal H}_\lambda, y\in M^{-1}{\cal H}_{\bar\lambda^{-1}}$. Then $Q$ is invertible and $Q^*=Q$.
Let $x_1,x_2\in M^{-1}(F_{\lambda}\cap E)$, $y_1,y_2\in M^{-1}(F_{\bar\lambda^{-1}}\cap E)$.
We have
\begin{equation*}
(QM^*A_sMQ(x_1+y_1),(x_2+y_2))=-(M^*A_sM(x_1+y_1),(x_2+y_2)).
\end{equation*}
It follows that $-A_s|_{F_\lambda}=Q(A_s|_{F_\lambda})Q$.
Then by Theorem \ref{th:coginva}, we have
$$2 sf(A_s|_{F_{\lambda}})=sf(A_s|_{F_{\lambda}})+sf(QA_s|_{F_\lambda}Q)=sf(A_s)+sf(-A_s)=dim \ker(A_1|_{F_{\lambda}})-dim \ker ( A_0|_{F_{\lambda}}).$$
The lemma then follows.
\end{proof}
\begin{proof}[Proof of Theorem \ref{th:decom}]
By Proposition \ref{pro:decomposition} and Lemma \ref{lm:sf_notin_circle} ,
we only need to show that
$$
\frac{1}{2}(dim\ker (A_1|_{\hat{F}}) -dim\ker (A_0|_{\hat{F}}))=\sum_{\lambda\notin \mathbb{U}}\frac{1}{2}(dim\ker (A_1|_{F_{\lambda}}) -dim\ker (A_0|_{F_{\lambda}}) ) .
$$
In fact $\ker A_1|_{\hat{F}}=\ker A_1\cap \hat{F}$.
By Lemma \ref{lm:ker_dom_decom}, we see that $\ker A_1\cap \hat{F}=\sum_{\lambda\notin \mathbb{U}}\ker(A_1)\cap F_{\lambda}$.
It follows that $dim\ker (A_1|_{\hat{F}})=\sum_{\lambda\notin \mathbb{U}}dim \ker(A_1|_{F_{\lambda}})$. It is also true for $A_0$.
The theorem then follows.
\end{proof}
\begin{cor}\label{cor2.1} Under the condition of Theorem \ref{th:decom}, if $\sigma(M)\cap \mathbb{U}=\emptyset$, then \begin{eqnarray} sf(A_s)=\frac{1}{2}(dim\ker (A_1) -dim\ker (A_0)). \end{eqnarray}
If the path is closed, then $$sf(A_s)=0.$$
\end{cor}
\begin{rem}\label{re3.8}
In the case $B$ is compact with respect to $A$, the spectral flow $A-sB$ is only depend on the end points, thus we define the relative Morse index by (follows \cite{ZL99})
\begin{eqnarray} I(A,A-B)=-sf(A-sB; s\in[0,1]). \end{eqnarray}
Especially, when $A$ is positive, then $I(A,A-B)=m^-(A-B)$ is just the Morse index of $A-B$, i.e. the total number of negative eigenvalues. It is obvious that Theorem \ref{th:decom} and Corollary \ref{cor2.1} give the decomposition formula of relative Morse index and Morse index.
\end{rem}
\section{Applications to Hamiltonian systems}
In this section, we will give the applications for Hamiltonian systems. We list $6$ cases which are common in applications.
For $\Lambda\in Lag(4n)$, we consider the solution of the flowing linear Hamiltonian systems
\begin{eqnarray} \dot{z}(t)= J B(t), \quad (z(0), z(T))\in\Lambda, \label{3.1}
\end{eqnarray}
where $B(t)\in C([0,T], \cal{S}({\mathbb R}^{2n}))$.
Recall that $A=-J{d\over dt}$ is self adjoint operator on
${\cal H}:=L^2([0,T],\mathbb C^{2n})$ with domain $$E_\Lambda=\{x\in W^{1,2}([0,T], \mathbb C^{2n}), (x(0), x(T))\in\Lambda \},$$
then $A, A-B\in \cal{FS}({\cal H})$. We will construct $g\in\mathcal{M}({\cal H})$ such that \begin{eqnarray} g^*Ag=A,\quad g^*Bg=B,\quad g E_\Lambda=E_\Lambda. \label{4.2} \end{eqnarray}
In order to make $g E_\Lambda=E_\Lambda$, $g$ is always assumed to preserve the boundary condition, that is $g\Lambda=\Lambda$ which means
\begin{eqnarray} ((gx)(0),(gx)(T))\in \Lambda \quad if \quad (x(0),x(T))\in\Lambda. \nonumber \end{eqnarray}
Hence we have \begin{eqnarray} g^*(A-sB)g=A-sB, \quad s\in{\mathbb R}.\nonumber \end{eqnarray}
and get the decomposition formula \eqref{decomf}.
It is well known that for $P\in{\mathrm {Sp}}(2n)$, if $\lambda\in\sigma(P)$, then $\bar{\lambda}, \lambda^{-1}, \bar{\lambda}^{-1}\in\sigma(P)$ and possess the same geometric and algebraic multiplicities \cite{Lon02}. Case 1 is given by symplectic matrix.
Case 1. For $P\in{\mathrm {Sp}}(2n)$, and satisfied $P\Lambda=\Lambda$
which means if $(x(0),x(T))\in\Lambda$, then $(Px(0),Px(T))\in\Lambda$.
Let \begin{eqnarray} (gx)(t)=Px(t),\label{g1} \end{eqnarray}
then it is obvious that $(g^*x)(t)=P^*x(t)$, $g^*Ag=A$, $g\Lambda=\Lambda$. Moreover, we assume $P^*B(t)P=B(t)$, then
$ g^*Bg=g$, hence we have \eqref{4.2}.
It is obvious that $g\in\mathcal{M}({\cal H})$ and \begin{eqnarray} \sigma(g)=\sigma(P). \nonumber\end{eqnarray}
Let $V_\lambda=\ker(P-\lambda)^{2n}$, then ${\cal H}_\lambda=L^2([0,T], V_\lambda)$.
Case 2.
For $S\in{\mathrm {Sp}}(2n)$, we consider the $S$-periodic solution of \eqref{3.1}, that is \begin{eqnarray} z(0)= S z(T), \label{bds} \end{eqnarray}
and moreover we assume \begin{eqnarray} S^*B(0)S=B(T). \label{Bc}\end{eqnarray}
We assume
\eqref{3.1} with $S$-periodic boundary conditions admits a $\mathbb Z_k$ symmetry. More precisely, let $P\in{\mathrm {Sp}}(2n)$ and $PS=SP$, the group generator $g$ is defined by
\begin{eqnarray} (gx)(t)=\left\{\begin{array}{cc} Px(t+\frac{T}{k}), \quad t\in[0,\frac{k-1}{k}T];
\\ PS^{-1}x(t+\frac{T}{k}-T), \quad t\in[\frac{k-1}{n}T,T]. \end{array}\right. \label{cird} \end{eqnarray}
Easy computation show that $g\in \cal{L}({\cal H})$ and $g E=E$. By direct computation, we get the adjoint operator $g^*$.
\begin{lem} The adjoint operator $g^*$ is given by \begin{eqnarray} (g^*x)(t)=\left\{\begin{array}{cc} (S^*)^{-1}P^*x(t+T-\frac{T}{k}), \quad t\in[0,\frac{T}{k}];
\\ P^*x(t-\frac{T}{k}), \quad t\in[\frac{T}{k},T]. \end{array}\right. \end{eqnarray}
\end{lem}
\begin{proof}
Let $y\in L^2([0,T],\mathbb{C}^{2n})$. We see that
$$
\int_{0}^{\frac{n-1}{n}T}(Px(t+T/k),y(t))dt= \int_{T/n}^{T}(x(t),P^*y(t-T/k))dt,
$$
and
$$
\int_{\frac{k-1}{k}T}^{T} (PS^{-1}x(t+T/k-T), y(t))dt=\int_{0}^{T/k}(x(t),(S^*)^{-1}P^*y(t+T-T/k)).
$$
Then we have checked $<gx,y>_{L^2}=<x,g^*y>_{L^2}$ for each $x,y\in L^2([0,T],\mathbb{C}^{2n})$.
\end{proof}
We assume $B(t)$ satisfied
\begin{eqnarray} B(t)=\left\{\begin{array}{cc} (S^*)^{-1}P^*B(t+T-\frac{T}{k})PS^{-1}, \quad t\in[0,\frac{T}{n}];
\\ P^*B(t-\frac{T}{k})P, \quad t\in[\frac{T}{n},T]. \end{array}\right. \label{BSy1} \end{eqnarray}
Please note that \eqref{BSy1} implies \eqref{Bc}, and \eqref{4.2} is satisfied.
Since $$ (g^kx)(t)=P^kS^{-1}x(t), $$
which is a multipliticity operator on ${\cal H}$. Then \begin{eqnarray} \sigma(g^k)=\sigma(P^kS^{-1}).\nonumber \end{eqnarray}
To simply the notation, for $\Omega\in\mathbb C$, we define
\begin{eqnarray} \Omega^{\frac{1}{k}} =\{z\in\mathbb C, z^k\in\Omega \}.\nonumber\end{eqnarray}
By this notation, we have $\sigma(g)\in(\sigma(P^kS^{-1}))^{\frac{1}{n}}$. For $\lambda\in\sigma(g)$, ${\cal H}_\lambda=\ker(g-\lambda)^{2n}$.
Case 3.
We consider the generalized brake symmetry.
We call a matrix $M$ anti-symplectic if it satisfied
\begin{eqnarray} M^*JM=-J. \end{eqnarray}
We denote by ${\mathrm {Sp}}_a(2n)$ the set of anti-symplectic matrices. For $M_1,M_2\in {\mathrm {Sp}}_a(2n)$ and $M_3\in{\mathrm {Sp}}(2n)$, then it is obvious that
$$ M_1M_2\in {\mathrm {Sp}}(2n) , \quad M_1M_3\in{\mathrm {Sp}}_a(2n) . $$
We list some basic property of ${\mathrm {Sp}}_a(2n)$ follows.
\begin{lem} If $M\in{\mathrm {Sp}}_a(2n)$, $\lambda\in\sigma(M)$, then $\bar{\lambda},-\lambda^{-1}, -\bar{\lambda}^{-1}\in\sigma(M)$ and posses the same geometric and algebraic multiplicities.
\end{lem}
\begin{proof}
Note that $M^*=-JM^{-1}J^{-1}$.
Let $\lambda\in \mathbb{C}\backslash\{0\}$.
It follows that $(M^*-\lambda)=-J(M^{-1}+\lambda)J^{-1}$.
Then we have
$$\hbox{\rm dim$\,$}\ker(M-\bar\lambda)=\hbox{\rm dim$\,$}\ker(M^*-\lambda)=\ker(M^{-1}+\lambda)=\hbox{\rm dim$\,$}\ker(M+\lambda^{-1}).$$
And we also have
$$
\overline{\det(M-\bar\lambda)}=\det(-M^{-1}-\lambda)=
\det(-M^{-1}\lambda)\det(M+\lambda^{-1}).
$$
It follows that
$\hbox{\rm dim$\,$}\ker(M-\bar\lambda)^{2n}=\hbox{\rm dim$\,$}\ker(M+\lambda^{-1})^{2n}$.
So $\lambda,-\lambda^{-1}\in \sigma(M)$ and posses the same geometric and algebraic multiplicities.
Specially, if $M$ is a real matrix, $\lambda,\bar\lambda,-\lambda^{-1},-\bar\lambda^{-1}\in \sigma(M)$ and posses the same geometric and algebraic multiplicities.
\end{proof}
Similar with the symplectic matrix, we have the following results.
\begin{lem} Let $M\in {\mathrm {Sp}}_a(2n)$ .
Let $\lambda,\mu \in \sigma(M)$.
Let $V_\lambda=\ker(M-\lambda)^{2n}$, $V_\mu=\ker(M-\mu)^{2n}$.
Then we have $(Jx,y)=0$ if $\lambda\bar{\mu}\neq-1$.
\end{lem}
\begin{proof}
Let $x\in \ker(M-\lambda)^p$,$y\in \ker(g-\mu)^q$ with $p,q\ge 0$.
We see that $(Jx,y)=0$ if $p+q=0$.
Assume that $(Jx,y)=0$ if $p+q\le k$.
Note that $(M-\lambda)x\in \ker(M-\lambda)^{p-1}$ ,$(M-\mu)x\in \ker(M-\mu)^{q-1}$.
If $p+q=k+1$,
We have
$$
(Jx,y)=-(JMx,My)=-(J(M-\lambda)x,M y)-(J\lambda x,(M-\mu)y)-\lambda\bar{\mu}(Jx,y)=-\lambda\bar\mu(Jx,y).
$$
Since $\lambda\bar\mu\neq - 1$, we have $(Ax,y)=0$.
By induction, we have $(Jx,y)=0$ with $x\in \ker(M-\lambda)^{2n}$ and $y\in \ker(M-\mu)^{2n} $. This complete the result.
\end{proof}
We assume \eqref{3.1} admits a generalized brake symmetry. More exactly, for $N\in{\mathrm {Sp}}_a(2n)$, let
\begin{eqnarray} (gx)(t)=Nx(T-t). \label{braked} \end{eqnarray}
We assume $g\Lambda=\Lambda$, that is \begin{eqnarray} (Nx(T),Nx(0))\in\Lambda, \, if \quad (x(0),x(T))\in\Lambda, \end{eqnarray}
then $gE=E$. Obviously,
$(g^*x)(t)=N^*x(T-t)$. We assume
\begin{eqnarray} N^*B(T-t)N=B(t), \end{eqnarray}
then \eqref{4.2} is satisfied.
Please note that for the $S$-periodic boundary conditions, $NS^{-1}=SN$ implies $g\Lambda=\Lambda$.
Separated boundary conditions is another kind of important boundary conditions. More preciselly, we consider solution of (\ref{3.1}) under the boundary conditions\begin{eqnarray} x(0)\in V_0,\quad x(T)\in V_1, \nonumber\end{eqnarray}
where $V_0,V_1\in Lag(2n)$. In this case $g$ is defined by \eqref{braked}, for $N\in{\mathrm {Sp}}_a(2n)$ which satisfied \begin{eqnarray} NV_0=V_1,\quad NV_1=V_0,\nonumber \end{eqnarray}
then $g\Lambda=\Lambda$.
Obviously,
we have $$ (g^2x)(t)=N^2x(t), $$ hence $g\in\mathcal{M}({\cal H})$ and
\begin{eqnarray} \sigma(g)=(\sigma(N^2))^{\frac{1}{2}}. \nonumber \end{eqnarray}
For $\lambda\in\sigma(g)$, ${\cal H}_\lambda=\ker(g-\lambda)^{2n}$.
From Theorem \ref{th:decom} we get the decomposition of spectral flow. Since on the finite interval $B$ is relative compact with respect to $A$, then from Remark \ref{re3.8}, we have
\begin{eqnarray} I(A,A-B)=\sum_{i=1}^m I(A|_{F_{\lambda_i}},A|_{F_{\lambda_i}}-B|_{F_{\lambda_i}})+\frac{1}{2}(dim\ker ((A-B)|_{\hat{F}}) -dim\ker (A|_{\hat{F}})). \label{decomrela} \end{eqnarray}
All the above discussions can be applied to Sturm-Liouville systems, so we not give the detail in all cases, instead we only consider the following two cases which have clearly background.
Case 4. We consider the one parameter family Sturm-Liouville system
\begin{eqnarray} -(G_s(t)\dot{x})'+R_s(t)x(t)=0, \quad x(0)=Sx(T),\, \dot{x}(0)=S\dot x(T), s\in[0,1]. \label{sl1} \end{eqnarray}
where $S\in\cal{L}^*({\mathbb R}^n)$. We suppose $G_s(t),R_s(t)\in \cal{S}(n)$, instead the Legender convex condition we only assume $G_s(t)$ is invertible. let $P\in\cal{L}^*({\mathbb R}^n)$ and $PS=SP$, the group generator $g$ is defined as same form of \eqref{cird}. We assume
\begin{eqnarray} G_s(t)=\left\{\begin{array}{cc} (S^*)^{-1}P^*G_s(t+T-\frac{T}{n})PS^{-1}, \quad t\in[0,\frac{T}{n}];
\\ P^*G_s(t-\frac{T}{n})P, \quad t\in[\frac{T}{n},T]. \end{array}\right. \end{eqnarray}
\begin{eqnarray} R_s(t)=\left\{\begin{array}{cc} (S^*)^{-1}P^*R_s(t+T-\frac{T}{n})PS^{-1}, \quad t\in[0,\frac{T}{n}];
\\ P^*R_s(t-\frac{T}{n})P, \quad t\in[\frac{T}{n},T]. \end{array}\right. \end{eqnarray}
Then \begin{eqnarray} g^*(-(G_s(t)\frac{d}{dt})'+R_s)g=-(G_s(t)\frac{d}{dt})'+R_s, \nonumber\end{eqnarray}
and we could give the decomposition of spectral flow from Theorem \ref{th:decom}.
This case include the Bott-type formula of Semi-Riemann manifold \cite{HPY17a}. Let $c$ be a space-like or time-like closed geodesic on $n+1$ dimension Semi-Riemann manifold $(M, \mathfrak{g})$ with period $T$. We choose a parallel $\mathfrak{g}$-orthonormal frame $e_i(t)$ alone $c$, and satisfied $\mathfrak{g}(e_i(t),\dot{c}(t))=0$. Assume
\begin{eqnarray} \mathfrak{g}(e_i,e_j)=\left\{\begin{array}{cc} 0, \quad i\neq j;
\\ 1,\quad 1\leq i=j\leq n-\nu; \\ -1,\quad n-\nu\leq i=j\leq n \end{array}\right. \nonumber \end{eqnarray}
and \begin{eqnarray} (e_1(0),\cdots,e_n(0))=(e_1(T),\cdots,e_n(T))P , \nonumber\end{eqnarray}
then $P^TGP=G$ with $G=diag(I_{n-\nu},-I_\nu)$.
Writing the $\dot{c}$ $\mathfrak{g}$-orthogonal Jacobi vectorfield alone $c$ as $J(t)=\sum_{i=1}^nu_i(t)e_i(t)$, then we get the linear second order system of ordinary differential equations
\begin{eqnarray} -G\ddot{u}+R(t)u(t)=0,\quad t\in[0,T], \end{eqnarray}
where $R$ is symmetry matrices which is get by the curvature. A period solution is satisfied
$$u(0)=Pu(T).$$
For $\omega\in\mathbb{U}$, let \begin{eqnarray} E^2_{\omega,T}:=\{u\in W^{2,2}([0,T],\mathbb{C}^n)|u(0)=\omega Pu(T), \dot{u}(0)=\omega P\dot{u}(T)\},\nonumber \end{eqnarray}
then \begin{eqnarray} A^\omega_{s,T}=-G\frac{d^2}{dt^2}+R(t)+sG \nonumber \end{eqnarray}
is self-adjoint Fredholm operators on $L^2([0,T],\mathbb{C}^n)$ with domain $E^2_{\omega,T}$.
It has proved in \cite{HPY17a} that there exist $s_0$ sufficiently large such that for $s\geq s_0$, $A^\omega_s$ is nondegenerate.
The $\omega$ spectral index of $c$ is defined by
\begin{eqnarray} i^\omega_{spec}(c) := sf(A^\omega_{s,T}; s\in [0, +\infty)). \nonumber\end{eqnarray}
Let $c^{(m)}$ be the $m$-th iteration of $c$, then
\begin{eqnarray} i^\omega_{spec}(c^{(m)}) := sf(A^\omega_{s,mT}; s\in [0, +\infty)).\nonumber \end{eqnarray}
Let $S=P^m$, $G_s=G$, $R_s=R(t)+sG$, $(gu)(t)=Pu(t+T)$, then from Case 4. we get the decomposition of spectral flow. Since $g^m=\omega$, then $$ \sigma(g)=\{\omega\}^{\frac{1}{m}}. $$
Let $\omega_j$ be the $m$-th root of $\omega$, then
\begin{eqnarray} {\cal H}_{\omega_j}= \ker(g-\omega_j)=\{ u(t)=\omega_j Pu(t+T) \}. \nonumber \end{eqnarray}
We have \begin{eqnarray} sf(A^\omega_{s,mT}; s\in [0, +\infty))=\sum_{\omega_j^m=\omega}sf(A^{\omega_j}_{s,T}; s\in [0, +\infty)). \nonumber \end{eqnarray}
Hence we get the Bott-type iteration formula \cite{HPY17a}
\begin{eqnarray} i^\omega_{spec}(c^{(m)}) =\sum_{\omega_j^m=\omega} i^{\omega_j}_{spec}(c). \end{eqnarray}
Obviously, we can consider the case of brake symmetry, since it is similar, we omit the detail.
Case 5.
Now we consider the case of heteroclinic orbits, for the one parameter family linear Hamiltonian system
\begin{eqnarray} \dot{x}=JB_s(t)x(t), \quad t\in{\mathbb R}, \quad s\in[0,1]. \label{case5f1} \end{eqnarray}
Let $B_s(\pm\infty)=\lim_{t\to\pm\infty}B_s(t) $ exist and satisfied the hyperbolic condition, i.e. $$\sigma(JB_s(\pm\infty))\cap i{\mathbb R}=\emptyset, \quad s\in[0,1].$$
Let ${\cal H}=L^2({\mathbb R}, \mathbb C^{2n})$, it is well known that $A-B_s \in\cal{FS}({\cal H})$ with domain $E=W^{1,2}({\mathbb R}, \mathbb{C}^{2n})$.
We assume
\begin{eqnarray} N^*B_s(-t)N=B_s(t), \nonumber \end{eqnarray}
then $g^*B_s g=g$. Easy computation show that $g^*Ag=A$, then we have
\begin{eqnarray} g^*(A-B_s)g=A-B_s, \quad s\in[0,1]. \nonumber \end{eqnarray}
Obviously, we have $$ (g^2x)(t)=N^2x(t), $$ hence $g\in\mathcal{M}({\cal H})$ and then \begin{eqnarray} \sigma(g)=(\sigma(N^2))^{\frac{1}{2}}. \nonumber \end{eqnarray}
For $\lambda\in\sigma(g)$, ${\cal H}_\lambda=\ker(g-\lambda)^{2n}$, then we get the decomposition formula from Theorem \ref{th:decom}.
In the case $N^2=I$, let \begin{eqnarray} {\cal H}_\pm=\ker{g\mp I}=\{x\in{\cal H},Nx(-t)=\pm x(t)\}, \label{hpm}\end{eqnarray}
we have
\begin{eqnarray} sf(A-B_s)=sf(A|_{{\cal H}_+}-B_s|_{{\cal H}_+} )+sf(A|_{{\cal H}_-}-B_s|_{{\cal H}_-} ). \label{brakehamil} \end{eqnarray}
Now we consider the case of Homoclinics. For the linear Hamiltonian system
\begin{eqnarray} \dot{x}(t)=JB(t)x(t),\quad t\in\mathbb{R}, \label{homo}\end{eqnarray}
assume $\lim_{t\to\pm\infty}B(t)=B_*$ and $JB_*$ is hyperbolic. In this case, $B-B_*$ is relative compact with respect to
$A-B_*$, where $A=-J\frac{d}{dt}$. The relative index is defined by \begin{eqnarray} I(A-B_*, A-B)=-sf(A-B_*+s(B-B_*)). \nonumber\end{eqnarray}
In the case \eqref{homo} is a linear system of Homoclinic orbits $z$, the index of $z$ is defined by \cite{CH07} $$i(z)= I(A-B_*, A-B).$$
Assume $ N^*B(-t)N=B(t)$ and $N^2=I$, from \eqref{brakehamil}, we have
\begin{eqnarray} I(A-B_*, A-B)= I(A|_{{\cal H}_+}-B_*|_{{\cal H}_+}, A|_{{\cal H}_+}-B|_{{\cal H}_+})+I(A|_{{\cal H}_-}-B_*|_{{\cal H}_-}, A|_{{\cal H}_-}-B|_{{\cal H}_-}).\label{dechomo}\end{eqnarray}
Case 6. We consider the one parameter Sturm-Liouville system on ${\mathbb R}$
\begin{eqnarray} -(G(t)\dot{x})'+R(t)x(t)=0,\quad t\in{\mathbb R} \end{eqnarray}
where $G(t),R(t)\in \cal{S}(n)$. We assume there exist $\delta>0$, such that $G(t)>\delta$ for $t\in{\mathbb R}$, and there exist $T, \delta_1, \delta_2>0$ such that
\begin{eqnarray} \delta_1<R(t)<\delta_2,\quad for\quad t\geq |T|. \label{slheter} \end{eqnarray}
The Morse index of $\mathcal{A}:=-(G(t)\frac{d}{dt})'+R(t) $ is defined by the maximum dimension of the subspace such that $\mathcal{A}$ restricted on it is negative definite.
It is well known that $m^-(\mathcal{A})$ is finite under the condition \eqref{slheter}. Obviously
\begin{eqnarray} m^-(\mathcal{A})=sf(\mathcal{A}+sG(0); s\in[0,+\infty)).\nonumber \end{eqnarray}
We assume there exist $N\in{\mathrm {Sp}}_a(n)$, such that
\begin{eqnarray} N^*R(-t)N=R(t),\quad N^*G(-t)N=G(t).\nonumber \end{eqnarray}
Let \begin{eqnarray} gx(t)=Nx(-t), \nonumber \end{eqnarray}
then $g^*(\mathcal{A}+sG(0))g=\mathcal{A}+sG(0)$. Since $g^2=N$, then $\sigma(g)=(\sigma(N))^{\frac{1}{2}}$, and we get the decomposition formula of Morse index.
\begin{eqnarray} m^-(\mathcal{A})=\sum_{i=1}^j m^-(\mathcal{A}|_{F_{\lambda_i}})-dim\ker \mathcal{A}|_{\hat{F}}. \nonumber \end{eqnarray}
In the case $N^2=I$, we have
\begin{eqnarray} m^-(\mathcal{A})=m^-(\mathcal{A}|_{{\cal H}_+})+m^-(\mathcal{A}|_{{\cal H}_-}) . \end{eqnarray}
\section{Relation with the Maslov index}
In this section, we will give some Bott-type iteration formulas of Maslov-index. In the what follows $g$ pointed as the Matrix-like operator appear in Case 2, 3, 5. To avoid discuss too many technique details, we only consider the case $$g^m=\omega I$$ for some $\omega\in\mathbb{U}$. Let $\omega_1,\cdots,\omega_m$ be the $m$-th roots of $\omega$, and let ${\cal H}_i=\ker(g-\omega_i)$.
In Case 2, 3, 5, ${\cal H}(\mathcal{I})=L^2(\mathcal{I}, \mathbb{C}^{2n})$ where $\mathcal{I}$ is some finite interval or $\mathbb{R}$, and $E$ is $W^{1,2}(\mathcal{I}, \mathbb{C}^{2n})$ which satisfied some boundary conditions. We choose a subinterval $\hat{\mathcal{I}}\subset \mathcal{I}$, and let $\mathcal{T}$ be the restricition map from ${\cal H}$ to ${\cal H}(\hat{\mathcal{I}}):=L^2(\hat{\mathcal{I}}, \mathbb{C}^{2n})$, that is
\begin{eqnarray} (\mathcal{T}f)(x)=f(x),\quad x\in \hat{\mathcal{I}}. \end{eqnarray}
$\hat{\mathcal{I}}$ is called a fundamental domain if for any $i=1,\cdots,m$, $\mathcal{T}$ is a bijection from ${\cal H}_i$ to $L^2(\hat{\mathcal{I}}, \mathbb{C}^{2n})$. Recall that $E_i=E\cap{\cal H}_i$ is domain of $A_s|_{{\cal H}_i}$, then $\mathcal{T}E_i$ is closed in the $W^{1,2}$ norm. Let $\hat{A}^{i}_s=-J\frac{d}{dt}-B_s$ be the operator on ${\cal H}(\hat{\mathcal{I}})$ with domain $\mathcal{T}E_i$.
%
\begin{lem} Suppose for $s\in[a,b]$, $\hat{A}^{i}_s$ is self-adjoint and $\hbox{\rm dim$\,$}\ker(A_s|_{{\cal H}_i})=\hbox{\rm dim$\,$}\ker(\hat{A}^{i}_s)$, then \begin{eqnarray} sf(A_s|_{{\cal H}_i}; s\in[a,b])=sf(\hat{A}^{i}_s; s\in[a,b]). \label{f5.1} \end{eqnarray}
\end{lem}
\begin{proof}
Note that $(A_s+t\Id)|_{{\cal H}_i}=P_{{\cal H}_i}(A_s+t\Id) P_{{\cal H}_i}=P_{{\cal H}_i}A_s P_{{\cal H}_i}+tP_{{\cal H}_i}$. Since $A_s\in \mathcal{FS}({\cal H})$, there is $\epsilon>0$, such that for each $t\in [0,\epsilon]$, $(A_s+t\Id)\in \mathcal{FS}({\cal H})$. Then $(A_s+t\Id)|_{{\cal H}_i}$ is a positive curve on $ \mathcal{FS}(H_i)$ with $t\in [0,\epsilon]$. Note that $-J\frac{d}{dt}-B_s+t\Id$ is also a positive curve on $\mathcal {FS}({\cal H}(\hat{I}))$, then \eqref{f5.1} is from Lemma \ref{abstract_decomposition}. This complete the proof.
\end{proof}
Now we consider Case 2. We assume $\omega S= P^m$, then \begin{eqnarray} {\cal H}_i=\ker(g-\omega_i)=\{x\in{\cal H}, \, \omega_i x(t)=Ps(t+\frac{T}{m}) \} .\nonumber \end{eqnarray}
We choose $\hat{\mathcal{I}}=[0,\frac{T}{m}]$ be the fundamental domain, then \begin{eqnarray} \mathcal{T}E_i=\{x\in W^{1,2}([0,\frac{T}{m}],\mathbb{C}^{2n}), \quad \omega_i x(0)=Ps(\frac{T}{m})\}. \nonumber\end{eqnarray}
From Corollary \ref{6.2}, we have \begin{eqnarray} -sf(A_s)=\mu(Gr(\omega S),Gr(\gamma(t));t\in[0,T]), \quad -sf(\hat{A}^{(i)}_s)=\mu(Gr(\omega_i), Gr(\gamma(t)); t\in[0,T/m]). \nonumber \end{eqnarray}
Then we have
\begin{eqnarray} \mu(Gr( S^{-1}),Gr(\gamma(t));t\in[0,T])=\sum_{i=1}^m\mu(Gr(\omega_iP^{-1}), Gr(\gamma(t));t\in[0,T/m]). \label{bt1} \end{eqnarray}
\begin{rem} In the case $P=I_{2n}$, \eqref{bt1} is the standard Bott-type iteration formula for Hamiltonian systems, please refer \cite{Lon99}, \cite{Lon02} for the detail. In the case $P\in{\mathrm {Sp}}(2n)\cap \mathbb{O}(2n)$, \eqref{bt1} is established by Hu and Sun \cite{HS09}, the general case is proved by Liu and Tang \cite{LT15}.
\end{rem}
For Case 3. We assume $N\in{\mathrm {Sp}}_a(2n)$ and $N^2=I$.
Since $N$ is anti-symplectic, we have $(Jx,x)=-(JNx,Nx)=-(Jx,x)=0$ with $x\in \ker(N-I)$.
So $\ker(N-I)$ is a Lagrange subspace of the symplectic space $(\mathbb{R}^{2n},J)$.
Recall that $(gx)(t)=Nx(T-t)$ and $g\Lambda=\Lambda$. Let \begin{eqnarray} {\cal H}_\pm=\ker{g\mp I}=\{x\in{\cal H},Nx(T-t)=\pm x(t)\},\nonumber \end{eqnarray}
and $\hat{\mathcal{I}}=[0,T/2]$ be the fundamental domain.
For the $S$-periodic boundary conditions, that is $x(0)=Sx(T)$, then
\begin{eqnarray} \mathcal{T}E_\pm =\{ x\in W^{1,2}([0,T/2], \mathbb{C}^{2n}), x(0)\in V^\pm(SN), x(T/2)\in V^\pm(N) \} .\nonumber \end{eqnarray}
We have
\begin{eqnarray} &&\mu(Gr(S),Gr(\gamma(t)); t\in[0,T])\nonumber \\ &&= \mu(V^+(N), \gamma(t)V^+(SN);t\in[0,T/2] )+\mu(V^-(N), \gamma(t)V^-(SN);t\in[0,T/2] ). \end{eqnarray}
Similarly if the boundary condition is given by $ x(0)\in V_0,\, x(T)\in V_1$ for
$V_0,V_1\in Lag(2n)$ and $ NV_0=V_1$, $NV_1=V_0$. Similar discussion with above, we have
\begin{eqnarray} \mu(V_1, \gamma(t)V_0; t\in[0,T]) = \mu(V^+(N), \gamma(t)V_0;t\in[0,T/2] )+\mu(V^-(N), \gamma(t)V_0;t\in[0,T/2] ). \label{bt2} \end{eqnarray}
\begin{rem} To our knowledge, in the case $S=I_{2n}$, $N^2=I$, \eqref{bt2} is first established by Long Zhang and Zhu \cite{LZZ06}. A deep study is given by Liu and Zhang \cite{LZ14a} \cite{LZ14b}. Hu and Sun had established the case of $S,N\in\mathbb{O}(2n)$, for the case of dihedral group please refer \cite{HPY17b}.
\end{rem}
Now we consider Case 5. For $\lambda\in[0,1]$, let $\gamma_\lambda(\tau,t)$ be the fundamental solution of \eqref{case5f1}, that is
\begin{eqnarray} \dot{\gamma}_\lambda(\tau,t)=JB_\lambda(t)\gamma_\lambda(\tau,t), \quad \gamma_\lambda(\tau,\tau)=I_{2n}. \end{eqnarray}
Let \begin{eqnarray} V^s_\lambda(\tau):=\{v\in\mathbb{R}^{2n}|\lim_{\tau\to\infty}\gamma_\lambda(\tau,t)v=0 \}, \quad V^u_\lambda(\tau):=\{v\in\mathbb{R}^{2n}|\lim_{\tau\to-\infty}\gamma_\lambda(\tau,t)v=0 \},\nonumber \end{eqnarray}
be the stable and unstable paths, then $V^s_\lambda(\tau), V^u_\lambda(\tau)\in Lag(2n)$.
Recall that in this case,
${\cal H}=L^2({\mathbb R},{\mathbb R}^{2n})$ and
$$
A_\lambda:= -J\frac{d}{dt}-B_s(t) : E=W^{1,2}(\mathbb{R},\mathbb{R}^{2n})\subset {\cal H}\to {\cal H}.
$$
Let $\mathbb{R}^-$ be the fundamental domain, then
\begin{eqnarray} \mathcal{T}E_\pm=\{x\in W^{1,2}(\mathbb{R}^-,\mathbb{R}^{2n}), x(0)\in V^\pm(N)\}. \nonumber \end{eqnarray}
Let $A^\pm_\lambda$ be the restricted operators on ${\cal H}(\mathbb{R}^-)$ with domain $ \mathcal{T}E_\pm$.
From Prop 3.7 of \cite{HP17}, we have
\begin{eqnarray} -sf (A_\lambda; \lambda\in[0,1])= \mu(V^s_\lambda(0), V^u_\lambda(0);\lambda\in[0,1]). \nonumber \end{eqnarray}
Similarly
\begin{eqnarray} -sf (A^\pm_\lambda; \lambda\in[0,1])= \mu(V^\pm(N), V^u_\lambda(0);\lambda\in[0,1]). \nonumber \end{eqnarray}
Then we have \begin{eqnarray} \mu(V^s_\lambda(0), V^u_\lambda(0);\lambda\in[0,1])=\mu(V^+(N), V^u_\lambda(0);\lambda\in[0,1])+\mu(V^-(N), V^u_\lambda(0);\lambda\in[0,1]). \end{eqnarray}
In the case of homoclinics, let $A_\lambda=A-B_*-\lambda(B-B_*)$, then from \cite{CH07} or \cite{HP17} the index satisfied
\begin{eqnarray} -sf(A_\lambda;\lambda\in[0,1] )=\mu(V^s(+\infty), V^u(t); t\in\mathbb{R})=-\mu(V^s(t), V^u(-t); t\in\mathbb{R}^+ ). \label{5.12} \end{eqnarray}
We have
\begin{eqnarray} -sf(A^\pm_\lambda;\lambda\in[0,1] )=\mu(V^\pm(N), V^u(t); t\in\mathbb{R}^-)=-\mu(V^\pm(N), V^u(-t); t\in\mathbb{R}^+ ). \label{5.13} \end{eqnarray}
Compare \eqref{5.12} and \eqref{5.13}, we have \begin{eqnarray} \mu(V^s(+\infty), V^u(t); t\in\mathbb{R})=\mu(V^+(N), V^u(t); t\in\mathbb{R}^-)+\mu(V^-(N), V^u(t); t\in\mathbb{R}^-),\end{eqnarray} or equivalently
\begin{eqnarray} \mu(V^s(t), V^u(-t); t\in\mathbb{R}^+ )=\mu(V^+(N), V^u(-t); t\in\mathbb{R}^+ )+\mu(V^-(N), V^u(-t); t\in\mathbb{R}^+ ). \end{eqnarray}
Obviously, we can use $\mathbb{R}^+$ as the fundamental domain, for reader's convenience, we list the formulas below.
Here we let $A_\lambda^{\pm}$ be the restricted operators on ${\cal H}(\mathbb{R}^+)$ with domain ${\cal T}(E_{\pm})$.
\begin{eqnarray} -sf (A^\pm_\lambda; \lambda\in[0,1])= \mu(V^s_\lambda(0),V^\pm(N) ;\lambda\in[0,1]). \end{eqnarray}
\begin{eqnarray} \mu(V^s_\lambda(0), V^u_\lambda(0);\lambda\in[0,1])=\mu(V^s_\lambda(0),V^+(N) ;\lambda\in[0,1])+\mu(V^s_\lambda(0),V^-(N);\lambda\in[0,1]). \end{eqnarray}
In case of the homoclinic,
\begin{eqnarray} \mu(V^s(t), V^u(-\infty); t\in\mathbb{R})&=&\mu(V^s(t), V^u(-t); t\in\mathbb{R}^+ )\nonumber \\ &=& \mu( V^s(t),V^+(N); t\in\mathbb{R}^+ )+\mu( V^s(t),V^-(N); t\in\mathbb{R}^+ ).\end{eqnarray}
In study the stability problem of homographic solution in planar $n$-body problem, Hu and Ou \cite{HO16} use the McGehee blow up method to get linear heteroclinic system, this system with brake symmetry if the corresponding central configurations with brake symmetry. Please refer \cite{HO16} for the detail.
\section{Appendix: Spectral flow and Maslov index}
Spectral flow was introduced by Atiyah, Patodi and Singer in their study of index theory on manifold with with boundary \cite{APS76}.
Let $\{A_t ,t\in [0,1]\}$ be a continuous path of self-adjoint Fredholm operators on a Hilbert space ${\cal H}$.
The spectral flow $sf\{A_t\}$ of $A_t$ counts the algebraic multiplicities of the spectral of $A_t$ cross the line
$\lambda=-\epsilon$ with some small positive number $\epsilon$. For reader's convenience, we list some basic properties of spectral flow.
{\bf (Stratum homotopy relative to the ends)\/} For $A_{s,\lambda}\in C([a,b]\times[0,1], \mathcal{FS}({\cal H}))$, such that $dim\ker A_{a,\lambda}$ and $dim\ker A_{b,\lambda}$ is constant, then
$$sf(A_{s,0};s\in[a,b])=sf(A_{s,1};s\in[a,b]) .$$
{\bf (Path additivity)\/} If $A^1,
A^2\in C\big([a,b];\mathcal{FS}({\cal H}))\big) $ are such that
$ A^1(b)= A^2(a)$, then
$$
sf( A^1_t * A^2_t; t \in [a,b]) = sf( A^1_t; t \in [a,b])+sf( A^2_t;
t \in [a,b])
$$
where $*$ denotes the usual catenation between the two paths.
{\bf (Direct sum)\/} If for $i=1,2$, ${\cal H}_i$ are Hilbert space, and $ A^i\in C\big([a,b];\mathcal{FS}({\cal H}_i))\big)$,
then $$
sf( A^1_t \oplus A^2_t; t \in [a,b]) = sf( A^1_t; t \in [a,b])+sf( A^2_t;
t \in [a,b]).
$$
{\bf (Nullity)\/} If $ A\in C\big([a,b];\mathcal{FS}({\cal H}))\big)$,
then $sf( A_t; t \in [a,b])=0$;
{\bf (Reversal)\/} Denote the
same path travelled in the reverse
direction in $\mathcal{FS}({\cal H})$ by $\widehat A(t)=A(-t)$.
Then
$$
sf( A_t; t \in [a,b])=-sf(\widehat A _t; t \in [-b,-a]).
$$
\\
The spectral flow is related to Maslov index in Hamiltonian systems.
We now briefly reviewing the Maslov index theory
\cite{Arn67,CLM94, RS93}. Let $(\mathbb{R}^{2n},\omega)$ be the standard
symplectic space and $Lag(2n)$ the Lagrangian Grassmanian.
For two
continuous paths $L_1(t), L_2(t)$, $t\in[a,b]$ in $Lag(2n)$,
the Maslov index $\mu(L_1, L_2)$ is an integer invariant.
Here we use the definition from \cite{CLM94}.
We list several properties of the Maslov index. The details could be found
in \cite{CLM94}.
{\bf(Reparametrization invariance)\/}
Let
$\phi:[c,d]\rightarrow [a,b]$ be a continuous and piecewise smooth
function with $\phi(c)=a$, $\phi(d)=b$, then \begin{eqnarray} \mu(L_1(t),
L_2(t))=\mu(L_1(\phi(\tau)), L_2(\phi(\tau))). \label{adp1.1} \end{eqnarray}
{\bf(Homotopy invariant with end points)\/}
For two continuous
family of Lagrangian path $L_1(s,t)$, $L_2(s,t)$, $0\leq s\leq 1$, $a\leq
t\leq b$, and satisfies $dim L_1(s,a)\cap L_2(s,a)$ and $dim
L_1(s,b)\cap L_2(s,b)$ is constant, then \begin{eqnarray} \mu(L_1(0,t),
L_2(0,t))=\mu(L_1(1,t),L_2(1,t)). \label{adp1.2} \end{eqnarray}
{\bf(Path additivity)\/}
If $a<c<b$, then then
\begin{eqnarray} \mu(L_1(t),L_2(t))=\mu(L_1(t),L_2(t)|_{[a,c]})+\mu(L_1(t),L_2(t)|_{[c,b]}).
\label{adp1.3} \end{eqnarray}
{\bf(Symplectic invariance)\/}
Let ${\gamma}(t)$, $t\in[a,b]$ is a
continuous path in ${\mathrm {Sp}}(2n)$, then \begin{eqnarray} \mu(L_1(t),L_2(t))=\mu({\gamma}(t)L_1(t), {\gamma}(t)L_2(t)). \label{adp1.4} \end{eqnarray}
{\bf(Symplectic additivity)\/}
Let $W_i$, $i=1,2$ be symplectic space,
$L_1,L_2\in C([a,b], Lag(W_1))$ and $\hat{L}_1,\hat{L}_2\in C([a,b], Lag(W_2))$,
then \begin{eqnarray} \mu(L_1(t)\oplus \hat{L}_1(t),L_2(t)\oplus \hat{L}_2(t))= \mu(L_1(t),L_2(t))+
\mu(\hat{L}_1(t),\hat{L}_2(t)).
\label{adp1.5add} \end{eqnarray}
The next Theorem give the relation of spectral flow and Maslov index.
\begin{thm}\label{th:6.1}
\begin{eqnarray} -sf(A_s-B_s)=\mu(\Lambda_s, Gr(\gamma_s(T)) ) \label{th:morma} \end{eqnarray}
\end{thm}
The above Theorem is well known \cite{HP17}, we like to give a direct proof here.
\begin{proof}
We use the idea of Lemma \ref{abstract_decomposition}.
We only need to prove the theorem locally.
Let $s_0\in [0,1]$. $A_{s_0}-B_{s_0}+rI,r\in [0,1]$ is a positive path in $\cal{FS}({\cal H})$.
There is $\epsilon>0$ such that $\ker(A_{s_0}-B_{s_0}+\epsilon I)={0}$.
Then there is $\delta>0$ such that $\ker(A_{s}-B_{s}+\epsilon I)={0}$, $\forall s\in [s_0-\delta,s_0+\delta]$.
Without loss of generality, we can assume that $$\ker(A_{s}-B_{s}+I)={0},\, \forall s\in [0,1].$$
Let $\gamma_{s,r}(t)$ be the fundamental solution of the equation
\begin{eqnarray} \dot{z}(t)= J (B_s(t)-rI)z(t), (s,t)\in[0,1]\times[0,T],\, r\in [0,1].
\end{eqnarray}
Recall that $\hbox{\rm dim$\,$}(\Graph(\gamma_{s,r}(T))\cap \Lambda_s)=\hbox{\rm dim$\,$}\ker(A_s-B_s+rI)$, then
we have $\mu(\Lambda_s,Gr(\gamma_{s,1}(T)))=0$.
Then use the homotopy invariant property of spectral flow and Maslov index, we have
\begin{eqnarray}
\begin{cases}\label{eq:sf_mas}
sf(A_s-B_s,s\in [0,1])=sf(A_{0}-B_{0}+rI)-sf(A_1-B_1+rI)\\
\mu(\Lambda_s,\Graph(\gamma_s(T)))=\mu(\Lambda_0,\Graph(\gamma_{0,r}(T)))-\mu(\Lambda_1,\Graph(\gamma_{0,r}(T)))
\end{cases}.
\end{eqnarray}
Note that $A_0-B_0+rI$ is a positive path in $\cal{FS}({\cal H})$,
then we have
$$
sf(A_0-B_0+rI)=\sum_{0<r\le 1} \hbox{\rm dim$\,$}\ker(A_0-B_0+rI).
$$
Let $Q_t((x,\gamma(t) x),(y,\gamma(t) y))=<-J\gamma(t)^{-1}\dot{\gamma}(t)x,y>$
which is a quadratic form on $\Graph(\gamma(t))$.
Recall that the crossing form of the Lagrangian pair $(\Lambda,\Graph(\gamma(t)))$ is given by
$Q_t|_{Gr(\gamma(t))\cap \Lambda}$.
Note that
\begin{eqnarray*}
\frac{\partial}{\partial t}(\gamma_s(t)^{-1}\frac{\partial}{\partial s}\gamma_s(t))=-\gamma_s(t)^{-1}\frac{\partial}{\partial t}\gamma_s(t)\gamma_s(t)^{-1}\frac{\partial}{\partial s}\gamma_s(t)
+\gamma_s(t)^{-1}\frac{\partial}{\partial s}(\frac{\partial}{\partial t}\gamma_s(t))\\
=-\gamma_s(t)^{-1}\frac{\partial}{\partial t}\gamma_s(t)\gamma_s(t)^{-1}\frac{\partial}{\partial s}\gamma_s(t)
+\gamma_s(t)^{-1}\frac{\partial}{\partial s}(JB_s)\gamma_s(t)+\gamma_s(t)^{-1}JB_s\frac{\partial}{\partial s}\gamma_s(t)\\
=\gamma_s(t)^{-1}\frac{\partial}{\partial s}(JB_s)\gamma_s(t).
\end{eqnarray*}
Then we have
$\frac{\partial}{\partial t}(-J\gamma_{0,r}(t)^{-1}\frac{\partial}{\partial r}\gamma_{0,r}(t))=-I$, and
it follows that
$-J\gamma_{0,r}(t)^{-1}\frac{\partial}{\partial r}\gamma_{0,r}(t)=-TI$.
Thus we have
\begin{eqnarray*}
\mu(\Lambda_0,\Graph(\gamma_{0,r}(T)))&&=-\sum_{0<r\le 1}\hbox{\rm dim$\,$}(\Lambda_0\cap \Graph(\gamma_{0,r}(T)))\\
&&=-\sum_{0<r\le 1}\hbox{\rm dim$\,$}(\ker(A_0-B_0+rI))=-sf(A_0-B_0+rI).
\end{eqnarray*}
Similarly
$$
\mu(\Lambda_1,\Graph(\gamma_{1,r}(T)))=-sf(A_1-B_1+rI).
$$
Then by \eqref{eq:sf_mas}, we get \eqref{th:morma}.
\end{proof}
From the homotopy invariance of Maslov index, we have
\begin{cor} \label{6.2} \begin{eqnarray} -sf(A-sB)=\mu(\Lambda,Gr(\gamma(t)), t\in[0,T]). \end{eqnarray}
\end{cor}
\noindent {\bf Acknowledgements.} We would like to thank Professor Alessandro Portaluri some helpful discussion with us for the spectral flow.
|
\section{Introduction}\label{sec:intro}
The fundamental connection~\cite{Barnett1935} between magnetic moment and spin angular momentum underlies the important role for magnets in nearly all spin-based concepts. An applied magnetic field provides the means to manipulate the state of a ferromagnet (FM), and thus the associated spin. Conversely, a spin-polarized current absorbed by the FM affects its magnetization~\cite{Slonczewski1996,Berger1996,Ralph2008,Brataas2012b}. Exploiting a related phenomenon, switching the state of an antiferromagnet (AFM) has also been achieved~\cite{Wadley2016}. Emboldened by this newly gained control, there has been an upsurge of interest in AFMs~\cite{Gomonay2014,Jungwirth2016,Baltz2018,Gomonay2018}, which offer several advantages over FMs. These include the absence of stray fields and a larger anisotropy-induced gap in the magnon spectrum. The two-sublattice nature of the AFMs further lends itself to phenomena distinct from FMs~\cite{Libor2018}.
Concurrently, ferrimagnets (FiMs) have been manifesting their niche in a wide range of phenomena such as ultrafast switching~\cite{Hansteen2005,Stanciu2007,Graves2013} and low-dissipation spin transport~\cite{Uchida2010,Adachi2013,Kruglyak2010,Chumak2015,Bauer2012,Cornelissen2015,Goennenwein2015,Weiler2013}. A class of FiMs exhibits the so-called compensation temperature~\cite{Rodrigue1960,Geschwind1959,Gepraegs2016,Cramer2017,Niklas2017,Kim2018}, at which the net magnetization vanishes, similar to the case of AFMs. Despite a vanishing magnetization in the compensated state, most properties remain distinct from that of AFMs~\cite{Gurevich1996}. Thus, these materials can be tuned to mimic FMs and AFMs via the temperature. In conjunction with the possibility of a separate angular-momentum compensation, when the magnetization does not vanish but the total spin does, FiMs provide a remarkably rich platform for physics and applications. An increased complexity in the theoretical description~\cite{Gurevich1996,Kamra2017B} hence accompanies these structurally complicated materials, and may be held responsible for comparatively fewer theoretical studies. Nevertheless, a two-sublattice model with distinct parameters for each sublattice qualitatively captures all the phenomena mentioned above.
Dissipation strongly influences the response of a magnet to a stimulus and is thus central to the study of magnetic phenomena such as switching, domain wall motion and spin transport. Nevertheless, magnetic damping has conventionally been investigated via the ferromagnetic resonance (FMR) linewidth. It is accounted for phenomenologically in the Landau-Lifshitz description of the magnetization dynamics via the so-called Gilbert damping term~\cite{Gilbert2004}, which produces a good agreement with experiments for a wide range of systems. The Gilbert damping represents the viscous contribution and may be `derived' within a Lagrangian formulation of classical field theory by including the Rayleigh dissipation functional~\cite{Gilbert2004}. While the magnetic damping for FMs has been studied in great detail~\cite{Gurevich1996,Akhiezer1968,Sparks1961,Gilbert2004,Tserkovnyak2002,Tserkovnyak2005}, from phenomenological descriptions to microscopic models, a systematic development of an analogous description for ferri- and antiferromagnets has been lacking in literature. Furthermore, recent theoretical results on spin pumping in two-sublattice magnets~\cite{Kamra2017} and damping in AFMs~\cite{Liu2017} suggest an important role for the previously disregarded~\cite{Gurevich1996} cross-sublattice terms in Gilbert damping, and thus set the stage for the present study. Yuan and co-workers have recently presented a step in this direction focussing on spin torques in AFMs~\cite{Yuan2018}.
Here, we formulate the magnetization dynamics equations in a general two-sublattice magnet following the classical Lagrangian approach that has previously been employed for FMs~\cite{Gilbert2004}. The Gilbert damping is included phenomenologically via a Rayleigh dissipation functional appropriately generalized to the two-sublattice system, which motivates intra- as well as cross-sublattice terms. The Gilbert damping parameter thus becomes a 2$\times$2 matrix, in contrast with its scalar form for a single-sublattice FM. Solving the system of equations for spatially homogeneous modes in a collinear ground state, we obtain the decay rates of the two eigenmodes finding direct pathways towards probing the dissipation mechanism and asymmetries in the system. Consistent with recent experiments~\cite{Hannes2017,Kim2018}, we find an enhancement in the decay rates~\cite{Hannes2017} close to the magnetization compensation in a FiM with an unaltered damping matrix~\cite{Kim2018}. The general description is found to be consistent with the spin pumping mediated damping in the magnet~\cite{Tserkovnyak2005,Kamra2017,Tserkovnyak2002}, and allows for relating the Gilbert damping matrix with the interfacial spin-mixing conductances. Focusing on AFMs, we express the magnetization dynamics in terms of the Neel variable thus clarifying the origin of the different damping terms in the corresponding dynamical equations~\cite{Hals2011,Yuan2018}. Apart from the usually considered terms, we find additional contributions for the case when sublattice-symmetry is broken in the AFM~\cite{Belashchenko2010,He2010,Kosub2017,Nogues1999,Kamra2017,Kamra2018}. Thus, FMR linewidth measurements offer a direct, parameter-free means of probing the sublattice asymmetry in AFMs, complementary to the spin pumping shot noise~\cite{Kamra2017}.
This paper is organized as follows. We derive the Landau-Lifshitz-Gilbert (LLG) equations for the two-sublattice model in Sec. \ref{sec:dynamics}. The ensuing equations are solved for the resonance frequencies and decay rates of the uniform modes in a collinear magnet in Sec. \ref{sec:uniform}. Section \ref{sec:special} presents the application of the phenomenology to describe a compensated ferrimagnet and spin pumping mediated Gilbert damping. The case of AFMs is discussed in Sec. \ref{sec:AFM}. We comment on the validity and possible generalizations of the theory in Sec. \ref{sec:discussion}. The paper is concluded with a summary in Sec. \ref{sec:summary}. The discussion of a generalized Rayleigh dissipation functional and properties of the damping matrix is deferred to the appendix.
\section{Magnetization dynamics and Gilbert damping}\label{sec:dynamics}
We consider a two-sublattice magnet described by classical magnetization fields $\pmb{M}_{A} \equiv \pmb{M}_{A}(\pmb{r},t)$ and $\pmb{M}_{B} \equiv \pmb{M}_{B}(\pmb{r},t)$ corresponding to the sublattices $A$ and $B$. The system is characterized by a magnetic free energy $F[\pmb{M}_A,\pmb{M}_B]$ with the magnetization fields assumed to be of constant magnitudes $M_{A0}$ and $M_{B0}$. Here, the notation $F[~]$ is employed to emphasize that the free energy is a functional over the magnetization fields, i.e. an integration of the free energy density over space.
The undamped magnetization dynamics is described by equating the time derivative of the spin angular momentum associated with the magnetization to the torque experienced by it. The resulting Landau-Lifshitz equations for the two fields may be written as:
\begin{align}\label{eq:angflow}
\frac{d}{dt} \left( \frac{\pmb{M}_{A,B}}{- |\gamma_{A,B}|} \right) = - \frac{\dot{\pmb{M}}_{A,B}}{ |\gamma_{A,B}|} = & ~\pmb{M}_{A,B} \times \mu_0 \pmb{H}_{A,B},
\end{align}
where $\gamma_{A,B} ~(< 0)$ are the gyromagnetic ratios for the two sublattices, and $\pmb{H}_{A,B}$ are the effective magnetic fields experienced by the respective magnetizations. This expression of angular momentum flow may be derived systematically within the Lagrangian classical field theory~\cite{Gilbert2004}. The same formalism also allows to account for a restricted form of damping via the so-called dissipation functional $R[\dot{\pmb{M}}_A,\dot{\pmb{M}}_B]$ in the generalized equations of motion:
\begin{align}\label{eq:lang}
\frac{d}{dt} \frac{\delta \mathcal{L}[\cdot]}{\delta \dot{\pmb{M}}_{A,B}} - \frac{\delta \mathcal{L}[\cdot]}{\delta \pmb{M}_{A,B}} = & - \frac{\delta R[\dot{\pmb{M}}_A,\dot{\pmb{M}}_B]}{\delta \dot{\pmb{M}}_{A,B}},
\end{align}
where $\mathcal{L}[\cdot] \equiv \mathcal{L}[\pmb{M}_A,\pmb{M}_B,\dot{\pmb{M}}_A,\dot{\pmb{M}}_B] $ is the Lagrangian of the magnetic system. Here, $\delta \mathcal{L}[\cdot] / \delta \pmb{M}_{A}$ represents the functional derivative of the Lagrangian with respect to the various components of $\pmb{M}_A$, and so on. The left hand side of Eq. (\ref{eq:lang}) above represents the conservative dynamics of the magnet and reproduces Eq. (\ref{eq:angflow}) with~\cite{Gilbert2004}
\begin{align}\label{eq:Hab}
\mu_0 \pmb{H}_{A,B} = & - \frac{\delta F[\pmb{M}_A,\pmb{M}_B]}{\delta \pmb{M}_{A,B}},
\end{align}
while the right hand side accounts for the damping.
The Gilbert damping is captured by a viscous Rayleigh dissipation functional parameterized by a symmetric matrix $\eta_{ij}$ with $\{ i,j\} = \{ A,B\}$:
\begin{align}
R[\dot{\pmb{M}}_{A},\dot{\pmb{M}}_{B}] = & \int_{V} d^3r \left( \frac{\eta_{AA}}{2} \dot{\pmb{M}}_{A} \cdot \dot{\pmb{M}}_{A} + \frac{\eta_{BB}}{2} \dot{\pmb{M}}_{B} \cdot \dot{\pmb{M}}_{B} + \eta_{AB} \dot{\pmb{M}}_{A} \cdot \dot{\pmb{M}}_{B} \right),
\end{align}
where $V$ is the volume of the magnet. The above form of the functional assumes the damping to be spatially homogeneous, isotropic, and independent of the equilibrium configuration. A more general form with a lower symmetry is discussed in appendix \ref{sec:RD}. Including the dissipation functional via Eq. (\ref{eq:lang}) leads to the following replacements in the equations of motion (\ref{eq:angflow}):
\begin{align}
\mu_0 \pmb{H}_{A} \to \mu_0 \pmb{H}_{A} - \eta_{AA} \dot{\pmb{M}}_A - \eta_{AB} \dot{\pmb{M}}_B, \\
\mu_0 \pmb{H}_{B} \to \mu_0 \pmb{H}_{B} - \eta_{BB} \dot{\pmb{M}}_B - \eta_{AB} \dot{\pmb{M}}_A.
\end{align}
Hence, the LLG equations for the two-sublattice magnet become:
\begin{align}
\dot{\pmb{M}}_{A} = & - |\gamma_A| \left( \pmb{M}_A \times \mu_0 \pmb{H}_{A} \right) + |\gamma_A| \eta_{AA} \left( \pmb{M}_A \times \dot{\pmb{M}}_A \right) + |\gamma_A| \eta_{AB} \left( \pmb{M}_A \times \dot{\pmb{M}}_B \right), \\
\dot{\pmb{M}}_{B} = & - |\gamma_B| \left( \pmb{M}_B \times \mu_0 \pmb{H}_{B} \right) + |\gamma_B| \eta_{AB} \left( \pmb{M}_B \times \dot{\pmb{M}}_A \right) + |\gamma_B| \eta_{BB} \left( \pmb{M}_B \times \dot{\pmb{M}}_B \right).
\end{align}
These can further be expressed in terms of the unit vectors $\hat{\pmb{m}}_{A,B} = \pmb{M}_{A,B} / M_{A0,B0}$:
\begin{align}
\dot{\hat{\pmb{m}}}_{A} = & - |\gamma_A| \left( \hat{\pmb{m}}_A \times \mu_0 \pmb{H}_{A} \right) + \alpha_{AA} \left( \pmb{m}_A \times \dot{\hat{\pmb{m}}}_A \right) + \alpha_{AB} \left( \hat{\pmb{m}}_A \times \dot{\hat{\pmb{m}}}_B \right), \label{eq:madot} \\
\dot{\hat{\pmb{m}}}_{B} = & - |\gamma_B| \left( \hat{\pmb{m}}_B \times \mu_0 \pmb{H}_{B} \right) + \alpha_{BA} \left( \hat{\pmb{m}}_B \times \dot{\hat{\pmb{m}}}_A \right) + \alpha_{BB} \left( \hat{\pmb{m}}_B \times \dot{\hat{\pmb{m}}}_B \right), \label{eq:mbdot}
\end{align}
thereby introducing the Gilbert damping matrix $\tilde{\alpha}$ for a two-sublattice system:
\begin{align}\label{eq:dampmat}
\tilde{\alpha} = \begin{pmatrix}
\alpha_{AA} & \alpha_{AB} \\
\alpha_{BA} & \alpha_{BB}
\end{pmatrix} = & \begin{pmatrix}
|\gamma_A| \eta_{AA} M_{A0} & |\gamma_A| \eta_{AB} M_{B0} \\
|\gamma_B| \eta_{AB} M_{A0} & |\gamma_B| \eta_{BB} M_{B0}
\end{pmatrix}, \\
\frac{\alpha_{AB}}{\alpha_{BA}} = & \frac{|\gamma_A| M_{B0}}{|\gamma_B| M_{A0}}. \label{eq:offratio}
\end{align}
As elaborated in appendix \ref{sec:DM}, the positivity of the dissipation functional implies that the eigenvalues and the determinant of $\tilde{\alpha}$ must be non-negative, which is equivalent to the following conditions:
\begin{align}\label{eq:det}
\eta_{AA}, \eta_{BB} \geq 0, \quad \eta_{AA} \eta_{BB} \geq \eta_{AB}^2 \implies \alpha_{AA}, \alpha_{BB} \geq 0, \quad \alpha_{AA} \alpha_{BB} \geq \alpha_{AB} \alpha_{BA} .
\end{align}
Thus, Eqs. (\ref{eq:madot}) and (\ref{eq:mbdot}) constitute the main result of this section, and introduce the damping matrix [Eq. (\ref{eq:dampmat})] along with the constraints imposed on it [Eq. (\ref{eq:offratio}) and (\ref{eq:det})] by the underlying formalism.
\section{Uniform modes in collinear ground state}\label{sec:uniform}
In this section, we employ the phenomenology introduced above to evaluate the resonance frequencies and the decay rates of the spatially homogeneous modes that can be probed in a typical FMR experiment. We thus work in the macrospin approximation, i.e. magnetizations are assumed to be spatially invariant. Considering an antiferromagnetic coupling $J~(> 0)$ between the two sublattices and parameterizing uniaxial easy-axis anisotropies via $K_{A,B} ~(>0)$, the free energy assumes the form:
\begin{align}\label{eq:free}
F[\pmb{M}_A,\pmb{M}_B] = & \int_{V} d^3r \left[ - \mu_0 H_0 (M_{Az} + M_{Bz}) - K_{A} M_{Az}^2 - K_{B} M_{Bz}^2 + J \pmb{M}_A \cdot \pmb{M}_{B} \right] ,
\end{align}
where $H_0 \hat{\pmb{z}}$ is the applied magnetic field. The magnet is assumed to be in a collinear ground state: $\pmb{M}_A = M_{A0} \hat{\pmb{z}}$ and $\pmb{M}_B = - M_{B0} \hat{\pmb{z}}$ with $M_{A0} > M_{B0}$. Employing Eq. (\ref{eq:Hab}) to evaluate the effective fields, the magnetization dynamics is expressed via the LLG equations (\ref{eq:madot}) and (\ref{eq:mbdot}). Considering $\pmb{M}_A = M_{Ax} \hat{\pmb{x}} + M_{Ay} \hat{\pmb{y}} + M_{A0} \hat{\pmb{z}}$, $\pmb{M}_B = M_{Bx} \hat{\pmb{x}} + M_{By} \hat{\pmb{y}} - M_{B0} \hat{\pmb{z}}$ with $|M_{Ax,Ay}| \ll M_{A0}$, $|M_{Bx,By}| \ll M_{B0}$, we linearize the resulting dynamical equations. Converting to Fourier space via $M_{Ax} = \mathcal{M}_{Ax} \exp{(i \omega t)}$ etc. and switching to circular basis via $\mathcal{M}_{A\pm (B\pm)} = \mathcal{M}_{Ax(Bx)} \pm i \mathcal{M}_{Ay(By)}$, we obtain two sets of coupled equations expressed succinctly as:
\begin{align}\label{eq:2x2}
\begin{pmatrix}
\pm \omega - \Omega_A - i \omega \alpha_{AA} & - \left( |\gamma_A| J M_{A0} + i \omega \alpha_{AB} \frac{M_{A0}}{M_{B0}} \right) \\
\left( |\gamma_B| J M_{B0} + i \omega \alpha_{BA} \frac{M_{B0}}{M_{A0}} \right) & \pm \omega + \Omega_B + i \omega \alpha_{BB}
\end{pmatrix} \begin{pmatrix}
\mathcal{M}_{A\pm} \\ \mathcal{M}_{B\pm}
\end{pmatrix} = & \begin{pmatrix}
0 \\ 0
\end{pmatrix},
\end{align}
where we define $\Omega_{A} \equiv |\gamma_A| (J M_{B0} + 2 K_A M_{A0} + \mu_0 H_0) $ and $\Omega_{B} \equiv |\gamma_B| (J M_{A0} + 2 K_B M_{B0} - \mu_0 H_0) $. Substituting $\omega = \omega_{r\pm} + i \omega_{i\pm}$ into the ensuing secular equation, we obtain the resonance frequencies $\omega_{r\pm}$ to the zeroth order and the corresponding decay rates $\omega_{i\pm}$ to the first order in the damping matrix elements:
\begin{align}
\omega_{r\pm} = & \frac{ \pm (\Omega_A - \Omega_B) + \sqrt{(\Omega_A + \Omega_B)^2 - 4 J^2 |\gamma_{A}| |\gamma_{B}| M_{A0} M_{B0}}}{2}, \label{eq:freqs} \\
\frac{\omega_{i \pm}}{\omega_{r\pm}} = & \frac{\pm \omega_{r\pm} (\alpha_{AA} - \alpha_{BB}) + \alpha_{AA} \Omega_B + \alpha_{BB} \Omega_A - 2 J |\gamma_B| M_{A0} \alpha_{AB} }{\omega_{r+} + \omega_{r-}} . \label{eq:decays}
\end{align}
In the expression above, Eq.\ (\ref{eq:freqs}) and Eq.\ (\ref{eq:decays}), we have chosen the positive solutions of the secular equations for the resonance frequencies. The negative solutions are equal in magnitude to the positive ones and physically represent the same two modes. The positive-polarized mode in our notation corresponds to the typical ferromagnetic resonance mode, while the negative-polarized solution is sometimes termed `antiferromagnetic resonance'~\cite{Gepraegs2016}. In order to avoid confusion with the ferromagnetic or antiferromagnetic nature of the underlying material, we call the two resonances as positive- and negative-polarized. The decay rates can further be expressed in the following form:
\begin{align}\label{eq:decay2}
\frac{\omega_{i \pm}}{\omega_{r\pm}} = & \frac{ \bar{\alpha} \left(\Omega_A + \Omega_B \right) - 2 J |\gamma_B| M_{A0} \alpha_{AB} }{\omega_{r+} + \omega_{r-}} \pm \Delta \bar{\alpha} ,
\end{align}
with $\bar{\alpha} \equiv \left(\alpha_{AA} + \alpha_{BB}\right)/2$ and $\Delta \bar{\alpha} \equiv \left(\alpha_{AA} - \alpha_{BB}\right)/2$. Eq.\ (\ref{eq:decay2}) constitutes the main result of this section and demonstrates that (i) asymmetric damping in the two sublattices is manifested directly in the normalized decay rates of the two modes (Figs. \ref{fig:field} and \ref{fig:comp}), and (ii) off-diagonal components of the damping matrix may reduce the decay rates (Fig. \ref{fig:comp}). Furthermore, it is consistent with and reproduces the mode-dependence of the decay rates observed in the numerical studies of some metallic AFMs~\cite{Liu2017}.
\begin{figure}[t]
\begin{center}
\includegraphics[height=100mm]{field.pdf}
\caption{Resonance frequencies and normalized decay rates vs. the applied field for a quasi-ferromagnet ($M_{A0} = 5 M_{B0}$). $|\gamma_A|/|\gamma_B| = 1, 1.5, 0.5$ correspond to solid, dashed and dash-dotted lines respectively. The curves in blue and red respectively depict the $+$ and $-$ modes. The damping parameters employed are $\alpha_{AA} = 0.06$, $\alpha_{BB} = 0.04$ and $\alpha_{AB} = 0$.}
\label{fig:field}
\end{center}
\end{figure}
To gain further insight into the results presented in Eqs. (\ref{eq:freqs}) and (\ref{eq:decay2}), we plot the resonance frequencies and the normalized decay rates vs. the applied magnetic field for a typical quasi-ferromagnet, such as yttrium iron garnet, in Fig. \ref{fig:field}. The parameters employed in the plot are $|\gamma_B| = 1.8 \times 10^{11}$, $M_{B0} = 10^5$, $K_{A} = K_{B} = 10^{-7}$, and $J = 10^{-5}$ in SI units, and have been chosen to represent the typical order of magnitude without pertaining to a specific material. The plus-polarized mode is lower in energy and is raised with an increasing applied magnetic field. The reverse is true for the minus-polarized mode whose relatively large frequency makes it inaccessible to typical ferromagnetic resonance experiments. As anticipated from Eq. (\ref{eq:decay2}), the normalized decay rates for the two modes differ when $\alpha_{AA} \neq \alpha_{BB}$. Furthermore, the normalized decay rates are independent of the applied field for symmetric gyromagnetic ratios for the two sublattices. Alternately, a measurement of the normalized decay rate for the plus-polarized mode is able to probe the sublattice asymmetry in the gyromagnetic ratios. Thus it provides essential information about the sublattices without requiring the measurement of the large frequency minus-polarized mode.
\section{Specific applications}\label{sec:special}
We now examine two cases of interest: (i) the mode decay rate in a ferrimagnet close to its compensation temperature, and (ii) the Gilbert damping matrix due to spin pumping into an adjacent conductor.
\subsection{Compensated ferrimagnets}
FMR experiments carried out on gadolinium iron garnet~\cite{Rodrigue1960,Hannes2017} find an enhancement in the linewidth, and hence the mode decay rate, as the temperature approaches the compensation condition, i.e. when the two effective~\footnote{The garnets have a complicated unit cell with several magnetic ions. Nevertheless, the two-sublattice model employed here captures the essential physics.} sublattices have equal saturation magnetizations. These experiments have conventionally been interpreted in terms of an effective single-sublattice model thereby ascribing the enhancement in the decay rate to an increase in the scalar Gilbert damping constant allowed within the single-sublattice model~\cite{Geschwind1959}. In contrast, experiments probing the Gilbert parameter in a different FiM via domain wall velocity find it to be essentially unchanged around compensation~\cite{Kim2018}. Here, we analyze FMR in a compensated FiM using the two-sublattice phenomenology developed above and thus address this apparent inconsistency.
\begin{figure}[t]
\begin{center}
\includegraphics[height=100mm]{mag.pdf}
\caption{Resonance frequencies and normalized decay rates vs. relative saturation magnetizations of the sublattices. The curves which are not labeled as $+$ or $-$ represent the common normalized decay rates for both modes. The parameters employed are the same as for Fig. \ref{fig:field} with $\gamma_A = \gamma_B$.}
\label{fig:comp}
\end{center}
\end{figure}
The compensation behavior of a FiM may be captured within our model by allowing $M_{A0}$ to vary while keeping $M_{B0}$ fixed. The mode frequencies and normalized decay rates are examined with respect to the saturation magnetization variation in Fig. \ref{fig:comp}. We find an enhancement in the normalized decay rate, consistent with the FMR experiments~\cite{Rodrigue1960,Hannes2017}, for a fixed Gilbert damping matrix. The single-sublattice interpretation ascribes this change to a modification of the effective Gilbert damping parameter~\cite{Geschwind1959}, which is equal to the normalized decay rate within that model. In contrast, the latter is given by Eq. (\ref{eq:decay2}) within the two-sublattice model and evolves with the magnetization without requiring a modification in the Gilbert damping matrix. Specifically, the enhancement in decay rate observed at the compensation point is analogous to the so-called exchange enhancement of damping in AFMs~\cite{Keffer1952}. Close to compensation, the FiM mimics an AFM to some extent.
We note that while the spherical samples employed in Ref. \onlinecite{Rodrigue1960} are captured well by our simple free energy expression [Eq. (\ref{eq:free})], the interfacial and shape anisotropies of the thin film sample employed in Ref. \onlinecite{Hannes2017} may result in additional contributions to decay rates. The similarity of the observed linewidth trends for the two kinds of samples suggests that these additional anisotropy effects may not underlie the observed damping enhancement. Quantitatively accounting for these thin film effects requires a numerical analysis, as discussed in Sec \ref{sec:discussion} below, and is beyond the scope of the present work. Furthermore, domain formation may result in additional damping contributions not captured within our single-domain model.
\subsection{Spin pumping mediated Gilbert damping}
Spin pumping~\cite{Tserkovnyak2002} from a FM into an adjacent conductor has been studied in great detail~\cite{Tserkovnyak2005} and has emerged as a key method for injecting pure spin currents into conductors~\cite{Maekawa2012}. The angular momentum thus lost into the conductor results in a contribution to the magnetic damping on top of the intrinsic dissipation in the bulk of the magnet. A variant of spin pumping has also been found to be the dominant cause of dissipation in metallic magnets~\cite{Liu2017}. Thus, we evaluate the Gilbert damping matrix arising due to spin pumping from a two-sublattice magnet~\cite{Kamra2017} into an adjacent conductor acting as an ideal spin sink.
Within the macrospin approximation, the total spin contained by the magnet is given by:
\begin{align}
\pmb{S} = - \frac{M_{A0} V \hat{\pmb{m}}_A}{|\gamma_A|} - \frac{M_{B0} V \hat{\pmb{m}}_B}{|\gamma_B|}.
\end{align}
The spin pumping current emitted by the two-sublattice magnet has the following general form~\cite{Kamra2017}:
\begin{align}
\pmb{I}_s = & \frac{\hbar}{e} \sum_{i,j = \{ A,B \}} G_{ij} \left( \hat{\pmb{m}}_i \times \dot{\hat{\pmb{m}}}_j \right),
\end{align}
with $G_{AB} = G_{BA}$, where the spin-mixing conductances $G_{ij}$ may be evaluated within different microscopic models~\cite{Cheng2014,Kamra2017,Sverre2018,Eirik2017}. Equating the spin pumping current to $- \dot{\pmb{S}}$ and employing Eqs. (\ref{eq:madot}) and (\ref{eq:mbdot}), the spin pumping contribution to the Gilbert damping matrix becomes:
\begin{align}
\alpha^\prime_{ij} & = \frac{\hbar G_{ij} |\gamma_i|}{e M_{i0} V},
\end{align}
which in turn implies
\begin{align}
\eta^\prime_{ij} = & \frac{\hbar G_{ij}}{e M_{i0} M_{j0} V},
\end{align}
for the corresponding dissipation functional. The resulting Gilbert damping matrix is found to be consistent with its general form and constraints formulated in Sec. \ref{sec:dynamics}. Thus, employing the phenomenology developed above, we are able to directly relate the magnetic damping in a two-sublattice magnet to the spin-mixing conductance of its interface with a conductor.
\section{Antiferromagnets}\label{sec:AFM}
Due to their special place with high symmetry in the two-sublattice model as well as the recent upsurge of interest~\cite{Gomonay2014,Jungwirth2016,Baltz2018,Gomonay2018,Johansen2018,Alireza2018,Alireza2018B}, we devote the present section to a focused discussion on AFMs in the context of the general results obtained above. It is often convenient to describe the AFM in terms of a different set of variables:
\begin{align}\label{eq:mn}
\pmb{m} = \frac{\hat{\pmb{m}}_A + \hat{\pmb{m}}_B}{2}, \quad \pmb{n} = \frac{\hat{\pmb{m}}_A - \hat{\pmb{m}}_B}{2}.
\end{align}
In contrast with $\hat{\pmb{m}}_A$ and $\hat{\pmb{m}}_B$, $\pmb{m}$ and $\pmb{n}$ are not unit vectors in general. The dynamical equations for $\pmb{m}$ and $\pmb{n}$ may be formulated by developing the entire field theory, starting with the free energy functional, in terms of $\pmb{m}$ and $\pmb{n}$. Such a formulation, including damping, has been accomplished by Hals and coworkers~\cite{Hals2011}. Here, we circumvent such a repetition and directly express the corresponding dynamical equations by employing Eqs. (\ref{eq:madot}) and (\ref{eq:mbdot}) into Eq. (\ref{eq:mn}):
\begin{align}
\dot{\pmb{m}} = & - \left(\pmb{m} \times \gamma_m \mu_0 \pmb{H}_m \right) - \left(\pmb{n} \times \gamma_n \mu_0 \pmb{H}_n \right) + \sum_{p,q = \{ m,n \} } \alpha^m_{pq} \left( \pmb{p} \times \dot{\pmb{q}} \right), \\
\dot{\pmb{n}} = & - \left(\pmb{m} \times \gamma_n \mu_0 \pmb{H}_n \right) - \left(\pmb{n} \times \gamma_m \mu_0 \pmb{H}_m \right) + \sum_{p,q = \{m,n\}} \alpha^n_{pq} \left( \pmb{p} \times \dot{\pmb{q}} \right),
\end{align}
with
\begin{align}
\gamma_m \mu_0 \pmb{H}_m \equiv & \frac{|\gamma_A| \mu_0 \pmb{H}_A + |\gamma_B| \mu_0 \pmb{H}_B}{2}, \\
\gamma_n \mu_0 \pmb{H}_n \equiv & \frac{|\gamma_A| \mu_0 \pmb{H}_A - |\gamma_B| \mu_0 \pmb{H}_B}{2}, \\
\alpha^{m}_{mm} = \alpha^{n}_{nm} = & \frac{\alpha_{AA} + \alpha_{BB} + \alpha_{AB} + \alpha_{BA}}{2}, \\
\alpha^{m}_{mn} = \alpha^{n}_{nn} = & \frac{\alpha_{AA} - \alpha_{BB} - \alpha_{AB} + \alpha_{BA}}{2}, \\
\alpha^{m}_{nn} = \alpha^{n}_{mn} = & \frac{\alpha_{AA} + \alpha_{BB} - \alpha_{AB} - \alpha_{BA}}{2}, \\
\alpha^{m}_{nm} = \alpha^{n}_{mm} = & \frac{\alpha_{AA} - \alpha_{BB} + \alpha_{AB} - \alpha_{BA}}{2}.
\end{align}
A general physical significance, analogous to $\gamma_{A,B}$, may not be associated with $\gamma_{m,n}$ which merely serve the purpose of notation here. The equations obtained above manifest new damping terms in addition to the ones that are typically considered in the description of AFMs. Accounting for the sublattice symmetry of the antiferromagnetic bulk while allowing for the damping to be asymmetric, we may assume $\gamma_A = \gamma_B$ and $M_{A0} = M_{B0}$, with $\bar{\alpha} \equiv \left(\alpha_{AA} + \alpha_{BB}\right)/2$, $\Delta \bar{\alpha} \equiv \left(\alpha_{AA} - \alpha_{BB}\right)/2$, and $\alpha_{AB} = \alpha_{BA} \equiv \alpha_{od}$. Thus, the damping parameters simplify to
\begin{align}
\alpha^{m}_{mm} = \alpha^{n}_{nm} = & \bar{\alpha} + \alpha_{od}, \\
\alpha^{m}_{mn} = \alpha^{n}_{nn} = & \Delta \bar{\alpha}, \\
\alpha^{m}_{nn} = \alpha^{n}_{mn} = & \bar{\alpha} - \alpha_{od}, \\
\alpha^{m}_{nm} = \alpha^{n}_{mm} = & \Delta \bar{\alpha},
\end{align}
thereby eliminating the ``new'' terms in the damping when $\alpha_{AA} = \alpha_{BB}$. However, the sublattice symmetry may not be applicable to AFMs, such as FeMn, with non-identical sublattices. Furthermore, the sublattice symmetry of the AFM may be broken at the interface~\cite{Belashchenko2010,He2010,Kosub2017} via, for example, spin mixing conductances~\cite{Kamra2017,Kamra2018,Bender2017} resulting in $\alpha_{AA} \neq \alpha_{BB}$.
The resonance frequencies and normalized decay rates [Eqs. (\ref{eq:freqs}) and (\ref{eq:decay2})] take a simpler form for AFMs. Substituting $K_A = K_B \equiv K$, $\gamma_A = \gamma_B \equiv \gamma$, and $M_{A0} = M_{B0} \equiv M_0$:
\begin{align}
\omega_{r\pm} = & \pm |\gamma| \mu_0 H_0 + 2 |\gamma| M_0 \sqrt{ (J + K) K }, \\
\frac{\omega_{i \pm}}{\omega_{r\pm}} = & \frac{ J (\bar{\alpha} - \alpha_{od}) + 2 K \bar{\alpha} }{2 \sqrt{ (J + K) K } } \pm \Delta \bar{\alpha} \approx \frac{(\bar{\alpha} - \alpha_{od})}{2} \sqrt{\frac{J}{K}} + \bar{\alpha} \sqrt{\frac{K}{J}} \pm \Delta \bar{\alpha},
\end{align}
where we have employed $J \gg K$ in the final simplification. The term $\propto \sqrt{K/J}$ has typically been disregarded on the grounds $K \ll J$. However, recent numerical studies of damping in several AFMs~\cite{Liu2017} find $\bar{\alpha} \gg \bar{\alpha} - \alpha_{od} > 0$ thus suggesting that this term should be comparable to the one proportional to $\sqrt{J/K}$ and hence may not be disregarded. The expression above also suggests measurement of the normalized decay rates as a means of detecting the sublattice asymmetry in damping. For AFMs symmetrical in the bulk, such an asymmetry may arise due to the corresponding asymmetry in the interfacial spin-mixing conductance~\cite{Kamra2017,Kamra2018,Bender2017}. Thus, decay rate measurements offer a method to detect and quantify such interfacial effects complementary to the spin pumping shot noise measurements suggested earlier~\cite{Kamra2017}.
\section{Discussion}\label{sec:discussion}
We have presented a phenomenological description of Gilbert damping in two-sublattice magnets and demonstrated how it can be exploited to describe and characterize the system effectively. We now comment on the limitations and possible generalizations of the formalism presented herein. To begin with, the two-sublattice model is the simplest description of ferri- and antiferromagnets. It has been successful in capturing a wide range of phenomenon. However, recent measurements of magnetization dynamics in nickel oxide could only be explained using an eight-sublattice model~\cite{Zhe2018}. The temperature dependence of the spin Seebeck effect in yttrium iron garnet also required accounting for more than two magnon bands~\cite{Barker2016}. A generalization of our formalism to a N-sublattice model is straightforward and can be achieved via a Rayleigh dissipation functional with $\mathrm{N}^2$ terms, counting $\eta_{ij}$ and $\eta_{ji}$ as separate terms. The ensuing Gilbert damping matrix will be N$\times$N while obeying the positive determinant constraint analogous to Eq. (\ref{eq:det}).
In our description of the collinear magnet [Eq. (\ref{eq:free})], we have disregarded contributions to the free energy which break the uniaxial symmetry of the system about the z-axis. Such terms arise due to spin-nonconserving interactions~\cite{Kamra2016A}, such as dipolar fields and magnetocrystalline anisotropies, and lead to a mixing between the plus- and minus-polarized modes~\cite{Kamra2017B}. Including these contributions converts the two uncoupled 2$\times$2 matrix equations [(\ref{eq:2x2})] into a single 4$\times$4 matrix equation rendering the solution analytically intractable. A detailed analysis of these contributions~\cite{Kamra2017B} shows that their effect is most prominent when the two modes are quasi-degenerate, and may be disregarded in a first approximation.
In evaluating the resonance frequencies and the decay rates [Eqs. (\ref{eq:freqs}) and (\ref{eq:decay2})], we have assumed the elements of the damping matrix to be small. A precise statement of the assumption employed is $\omega_{i} \ll \omega_{r}$, which simply translates to $\alpha \ll 1$ for a single-sublattice ferromagnet. In contrast, the constraint imposed on the damping matrix within the two-sublattice model by the assumption of small normalized decay rate is more stringent [Eq. (\ref{eq:decay2})]. For example, this assumption for an AFM with $\alpha_{AB} = \Delta \bar{\alpha} = 0$ requires $\bar{\alpha} \ll \sqrt{K/J} \ll 1$. This stringent constraint may not be satisfied in most AFMs~\cite{Liu2017}, thereby bringing the simple Lorentzian shape description of the FMR into question. It can also be seen from Fig. \ref{fig:comp} that the assumption of a small normalized decay rate is not very good for the chosen parameters.
\section{Summary}\label{sec:summary}
We have developed a systematic phenomenological description of the Gilbert damping in a two-sublattice magnet via inclusion of a Rayleigh dissipation functional within the Lagrangian formulation of the magnetization dynamics. Employing general expressions based on symmetry, we find cross-sublattice Gilbert damping terms in the LLG equations in consistence with other recent findings~\cite{Kamra2017,Liu2017,Yuan2018}. Exploiting the phenomenology, we explain the enhancement of damping~\cite{Rodrigue1960,Hannes2017} in a compensated ferrimagnet without requiring an increase in the damping parameters~\cite{Kim2018}. We also demonstrate approaches to probe the various forms of sublattice asymmetries. Our work provides a unified description of ferro- via ferri- to antiferromagnets and allows for understanding a broad range of materials and experiments that have emerged into focus in the recent years.
\section*{Acknowledgments}
A. K. thanks Hannes Maier-Flaig and Kathrin Ganzhorn for valuable discussions. We acknowledge financial support from the Alexander von Humboldt Foundation, the Research Council of Norway through its Centers of Excellence funding scheme, project 262633, ``QuSpin'', and the DFG through SFB 767 and SPP 1538.
|
\section{Introduction}
In statistical and condensed matter physics, the density of states (DOS) of a system
describes the number of states at each energy level. The DOS, which is independent of
temperature, represents a deep characterization of the system. In terms of
thermodynamics, knowledge of the DOS allows one to calculate the partition function
and hence all expectation values that can be derived from it, including the free and
internal energies as well as the specific heat, as a function of
temperature~\cite{sethna:06}. Alternatively, the DOS itself and closely related
quantities are the center of interest in an analysis of the thermodynamics of phase
transitions in the microcanonical ensemble~\cite{hueller:93}. DOS calculations can
also determine the spacing between energy bands in
semiconductors~\cite{solidState1}.
Whilst knowledge of the DOS, $\Omega(E)$, is extremely valuable, it cannot in
general be easily acquired. Exact calculations are only possible in a few special
cases such as Ising models on two-dimensional lattices
\cite{beale:96a,galluccio:00}. In general, the problem is exponentially hard as the
system size increases (in computational complexity terms it is a \#P-hard
problem)~\cite{DOScomplexity1,DOScomplexity2}. This difficulty notwithstanding, there
exist a number of approximation techniques, mostly based on Monte Carlo methods, that
allow one to estimate $\Omega(E)$. The most widely used approach of this type is the
Wang-Landau (WL) algorithm~\cite{wang:01a,wang:01b} and its variants
\cite{belardinelli:07,liang:07,vogel:13}, which is based on the multicanonical method
\cite{berg:92b}.
As stochastic approximation techniques, these approaches are affected by statistical
errors as well as systematic deviations (bias). Estimating statistical error is not
easily possible from a single WL simulation alone and normally requires statistics
over independent runs. The most relevant form of bias in WL simulations is that of a
false convergence, where the DOS estimate settles down on a deceptively smooth shape
that is however not a faithful representation of the actual DOS
\cite{schneider:17}. Naturally, such problems are notoriously hard to detect if no
independent information about the actual DOS is available.
Such difficulties apply in particular to systems with complex free-energy landscapes
that are typically accompanied by frustration in the interactions such as in the
protein-folding problem~\cite{bryngelson:95} and in the spin-glass systems that
result from a combination of frustration and quenched disorder~\cite{binder:86a}. The
latter may be viewed as prototypical classically-hard optimization problems, and they
are so challenging that specialized hardware has been built to simulate
them~\cite{belleti:09}. Recently, DOS estimation, which provides the relative
degeneracies of the energy levels of spin glasses, has become an indispensable tool
in the context of the benchmarking of experimental quantum
annealers~\cite{kadowaki_quantum_1998,farhi:01} and the attempts to demonstrate
speedups over classical devices. The currently available commercial realization of
this paradigm~\cite{Dwave} effectively samples low-lying energy configurations of
spin-glass samples that are coded into the couplers connecting an array of
superconducting flux qubits. The properties of the resulting samples and the question
of whether such devices indeed provide superior performance as compared to classical
algorithms for some problem classes has been the subject of much recent
debate~\cite{hen:15,katzgraber:15,boixo:16,perdomo,adachi,Amin:boltzmann,fairInUnfair,exponentiallyBiased,uncertainFate,marshallrieffelhen:2017}.
The questions of reliable Boltzmann sampling aided by quantum annealers for machine
learning applications~\cite{perdomo,adachi,Amin:boltzmann} as well as the advantages
that quantum annealing devices potentially hold when tasked with fairly sampling the
ground-state manifolds of spin glasses with multiple minimizing
configurations~\cite{fairInUnfair,exponentiallyBiased,uncertainFate} are now a topic
of considerable interest.
Our goal in this work is threefold. i) We devise techniques for \emph{verifiably}
benchmarking algorithms for sampling the DOS, designed to overcome the
pitfalls of misinterpreting false convergences of entropic samplers. ii) Employing
the above techniques, we demonstrate the difficulties in applying traditional
algorithms for sampling the DOS to spin-glass instances. iii) We
introduce a population annealing algorithm for estimating the DOS that allows for the
intrinsic control of statistical and systematical errors and demonstrate how it can
outperform the standard approach without the associated problems of choosing energy
windows and related parameters that occur for the latter.
\section{Verifiable benchmarking of entropic samplers\label{plantedSec}}
As mentioned above, approximation algorithms for the DOS are not always
reliable in converging to the correct answer, especially for the frustrated systems
considered here. While there are some general results regarding the convergence of
suitably modified WL type algorithms~\cite{liang:07}, convergence times can become
astronomically large and convergence hard to assess from intrinsic indicators
\cite{barash:17b}. As we shall see below, for disordered systems convergence times
can also fluctuate wildly between different realizations of the couplings. It is
hence highly desirable for benchmarking purposes to have at hand sets of samples that
are sufficiently challenging for the tested algorithms, but for which nevertheless
the (exact) DOS is known from other considerations. In general, such
samples are not readily available, but we present here two groups of examples that
are extremely useful in this respect: locally planar lattices and samples with
planted solutions.
For concreteness, we shall consider spin models of the Ising type, whose cost
function (or Hamiltonian) is of the form
\begin{equation}
\label{eq:Hp}
H = - \sum_{\langle i,j\rangle} J_{ij} s_i s_j - \sum_i h_i s_i,
\end{equation}
where $\langle i,j\rangle$ denotes the set of edges of the underlying graph. Here,
$s_i=\pm 1$ are the Ising spin variables and the quenched parameters $J_{ij}$ and
$h_i$ denote the exchange couplings and random fields, respectively. In the
following, we will focus on the zero-field case $h_i = 0$. While the problems of
computing the DOS (or partition function) and of finding a ground state
for this problem in at least three dimensions are NP hard~\cite{barahona:82a}, both
tasks can be completed in polynomial time on any set of graphs of a genus bounded by
a constant, which includes planar and toroidal two-dimensional lattices
\cite{galluccio:00}. For such cases there exist efficient algorithms to solve the
above problems. For ground states, these include minimum-weight perfect matching
\cite{bieche:80a,gregor:07,khoshbakht:17a}, while for the partition function or
DOS the usual approaches are based on the counting of dimer coverings
which can be achieved via an evaluation of Pfaffians
\cite{blackman:91,saul:93,galluccio:00}.
Unfortunately, such techniques are restricted to locally planar graphs and so they do
not apply, for example, to the Chimera graphs used in current implementations of
quantum annealers, which have a genus that grows linearly in the number of sites. For
such more general problems we propose here an approach that is based on generating
problem instances for which the values and degeneracies of the ground- and first
excited-states, $\Omega(E_0)$ and $\Omega(E_1)$, are {\em exactly\/}
computable. Since the states of lowest energies are usually the most difficult to
sample, the degeneracies of these two energy levels are the most difficult to
ascertain, and a correct estimation of these serves as a good indicator of true
convergence. We create such samples by considering problem instances with {\em
planted solutions}~\cite{hen:15,king:15,wang:17} --- an idea borrowed from
constraint satisfaction (SAT) problems~\cite{Barthel:2002tw,Krzakala:2009qo} where
the planted solution represents a ground-state configuration of Eq.~(\ref{eq:Hp})
that minimizes the energy and is known in advance. Following Ref.~\cite{hen:15}, the
Hamiltonian of a planted-solution spin glass is a sum of terms, each of which
consists of a small number of connected spins, namely, $H=\sum_j H_j$. Each term
$H_j$ is chosen such that one of its ground states is the planted solution. It
follows then that the planted solution is also a ground state of the total
Hamiltonian. This class of instances has two attractive properties: i) the
ground-state energies of the generated problems are known in advance, and ii) the
\emph{exact} degeneracies of the ground and first excited states are
computable~\cite{fairInUnfair,marshallrieffelhen:2017}. These in turn allow us to
check how close entropic samplers come to these exact values. The computation of
$\Omega(E_0)$ and $\Omega(E_1)$ is based on the fact that our generated instances
consist of a sum of terms, each of which has all minimizing configurations of $H$
as its ground state. To enumerate all ground states, we implement a form of the
`bucket' algorithm~\cite{dechter} designed to eliminate variables one at a time to
perform an exhaustive search efficiently (for a detailed description of the
algorithm, see the Supplementary Information of Ref.~\cite{fairInUnfair}). By noting
that the lowest energy excited states are those configurations that violate precisely
one clause, their degeneracy may also be calculated. We perform the same
exhaustive search as above, but where now the configurations tested will correspond
to first excited states of one of the $H_j$ (and are still minimizing for
$H_{i \neq j}$). This gives the number of configurations which are ground states of
$H_{i \neq j}$ and first excited states of $H_j$; by
performing this calculation for each of the $H_j$, we get the total number of first
excited states of $H$.
While this approach in principle works for arbitrary graphs, we focus here on Chimera
lattices, i.e., two-dimensional arrays of unit cells of eight spins with a $K_{4,4}$
bipartite connectivity~\cite{Choi1,Choi2}, see for example
Ref.~\cite{marshallrieffelhen:2017}. Our choice is motivated by the attention these
graphs have gained in recent years in the context of optimization as well as sampling
via quantum annealing as the quantum annealers currently commercially available
feature qubits connected with this topology~\cite{johnson:11,berkley:13,Bunyk:2014hb}.
While the Chimera graph is
two-dimensional in nature~\cite{weigel:15}, it is also non-planar and as such gives
rise to difficult spin-glass problems~\cite{barahona:82}. We generated 625
planted-solution instances of 501 spins each\footnote{The 501 spins correspond to
the graph considered in Ref.~\cite{marshallrieffelhen:2017}.}, following a technique
described in detail in Ref.~\cite{hen:15} wherein the clauses $H_j$ are chosen to be
`frustrated loops' along the Chimera graph. For each sample we employed the bucket
algorithm in order to obtain $\Omega(E_0)$ and $\Omega(E_1)$.
The combination of full exact DOS for samples on the square lattice and
toroidal boundary conditions and of exact values for $\Omega(E_0)$ and $\Omega(E_1)$
for the Chimera samples allows us to carefully examine the reliability and
performance of sampling schemes for estimating the DOS, avoiding the
pitfalls provided by badly converged estimates of stochastic approximation schemes.
\section{Sampling the DOS}
The common approximation algorithms for the DOS are based on Markov
chain Monte Carlo~\cite{berg:92b,wang:01a}. In the following, we will use the most
popular of these, the Wang-Landau algorithm~\cite{wang:01a,wang:01b}, in a variant
dubbed the WL-$1/t$ method~\cite{belardinelli:07} that in principle can be shown to
converge to the correct answer if given infinite run time~\cite{liang:07}, as a
reference and contender of the method introduced here, entropic population annealing.
\subsection{Wang-Landau sampling}
In Wang-Landau (WL) sampling as introduced in Refs.~\cite{wang:01a,wang:01b} a
running estimate $\hat{\Omega}(E)$ of the DOS (initialized as
$\hat{\Omega}(E)=1$ $\forall E$) is updated in a random walk through energy space by
multiplying $\hat{\Omega}(E)$ at the current energy $E$ by a modification factor $f$
(initially chosen to equal Euler's number $e$) at each step. A new configuration of
energy $E'$ is proposed according to the chosen move scheme (for a spin model
typically through single spin flips) and accepted with probability
\begin{equation}
p_\mathrm{acc} = \min\left[1,\frac{\hat{\Omega}(E)}{\hat{\Omega}(E')}\right].
\label{eq:wlprob}
\end{equation}
If, after some time, the histogram $\hat{H}(E)$ of all possible energies is found to
be ``sufficiently flat'' (typically interpreted as no histogram bin having less than
80\% of the average number of entries~\cite{wang:01a}, but see also
Ref.~\cite{gross:17} for a related discussion), the modification factor is reduced as
$f \to \sqrt{f}$, and the histogram $\hat{H}(E)$ is reset to an empty state. The
algorithm stops if $f$ is ``sufficiently small'', for example
$f=f_\mathrm{final}=\exp(10^{-8})$. While the approach was invented as a variant of
Markov chain Monte Carlo, the fact that the transition probabilities according to
Eq.~(\ref{eq:wlprob}) change constantly means that neither detailed nor global
balance are satisfied, and it is more useful to think of the method as a ``stochastic
approximation algorithm''~\cite{liang:07}.
It is well known that the original scheme of Ref.~\cite{wang:01a,wang:01b} does not
converge to the true DOS, but the error asymptotically saturates at a value
determined by the protocol used for reducing $f$~\cite{qiliang:03}. This shortcoming
is remedied by choosing a different modification protocol for $f$, leading to a
slower decay of $f$ at late times. The so-called $1/t$ algorithm proposed in
Ref.~\cite{belardinelli:07} uses two phases. In the first phase the standard WL
algorithm is used, with the only difference that the energy histograms are considered
to be sufficiently flat already if $\hat{H}(E)\ne 0$ for all $E$. Once $\ln f$ falls
below the moving threshold $N_E/t$, where $t$ is the simulation time measured in
spin-flip attempts and $N_E$ is the number of energy levels, the simulation enters
the second phase. There, the modification factor is adapted quasi continuously
according to $\ln f(t) = N_E/t$ and histogram flatness is no longer tested. The
simulation stops once $f(t) \le f_\mathrm{final}$.
While no saturation of error occurs in the $1/t$ algorithm
\cite{belardinelli:07,schneider:17}, it is still necessary to know the permissible
energy levels (including the ground state) beforehand to judge histogram flatness,
which is a major drawback of the method for disordered systems. In practice, we
therefore employ pre-runs of the WL type without any reduction of the modification
factor with the goal of discovering the available energy levels\footnote{We generally
choose the length of pre-runs so as to make sure that all levels are discovered,
but in a few cases the actual ground state is only found in the main run.}. For
large systems and problems with complex free-energy landscapes, it is usually
necessary to divide the total energy range into several windows for which the
algorithm is run separately to achieve convergence on realistic time scales for
interesting system sizes \cite{wang:01b}. The right choice of window sizes in such
schemes is a difficult problem especially for disordered and frustrated
systems~\cite{yin:12}, and we are not aware of any reliable systematic approach to
solve it. As a consequence, we had to spend considerable time with trial and error to
arrive at suitable setups for the problems studied below. A number of further
generalizations of the method have been proposed, for instance a combination with
parallel tempering~\cite{vogel:13} which uses progressively smaller windows at lower
energies, but here again there is no general algorithm for determining the
appropriate window sizes automatically.
\subsection{Entropic population annealing}
The new algorithm introduced here, which we call entropic population annealing (EPA),
is not based on Markov chains but on the sequential Monte Carlo method. Population
annealing (PA) was first studied in Refs.~\cite{iba:01,hukushima:03} and more recently
developed further in
Refs.~\cite{machta:10a,wang:15a,borovsky:16,barash:16,barash:17,amey:18,barzegar:17}. It
is based on the initialization of a population of replicas drawn from the equilibrium
distribution at high temperatures, which is then subsequently cooled to lower and
lower temperatures. During this process, a combination of population control and spin
flips is used to ensure that the ensemble remains in equilibrium. The simulation
entails the following steps~\cite{machta:10a,barash:16}:
\begin{enumerate}
\item Set up an equilibrium ensemble of $R_0 = R$ independent copies (replicas) of
the system at inverse temperature $\beta_0=1/k_BT_0$.
\item Take a step to inverse temperature $\beta_i > \beta_{i-1}$ by resampling the
configurations $j = 1,\ldots, R_{i-1}$ with their relative Boltzmann weight
$\hat{\tau}_i(E_j)$, leading to $R_i \ne R_{i-1}$ replicas, in general.
\item Update each replica by $\theta$ rounds of an MCMC algorithm at inverse
temperature $\beta_i$.
\item Goto step 2 unless the inverse target temperature $\beta_\mathrm{f}$ has been
reached.
\end{enumerate}
During resampling, the expected number of copies is
\begin{equation}
\hat{\tau}_i(E_j) = \frac{R}{R_{i-1}}\frac{e^{-(\beta_i-\beta_{i-1})E_j}}{Q(\beta_{i-1},\beta_i)},
\end{equation}
with a normalizing factor
\begin{equation}
Q(\beta_{i-1},\beta_i) = \frac{1}{R_{i-1}}
\sum_{j=1}^{R_{i-1}} e^{-(\beta_i-\beta_{i-1})E_j}.
\label{eq:Q}
\end{equation}
The actual number of copies is taken to be the integer part
$\lfloor\hat{\tau}_i(E_j)\rfloor$ plus an additional copy added with a probability
corresponding to the fractional part,
$\hat{\tau}_i(E_j) - \lfloor\hat{\tau}_i(E_j)\rfloor$. While initially constant
(inverse) temperature steps were used on increasing $\beta_i > \beta_{i-1}$
\cite{machta:10a}, it turns out that a better, parameter-free method consists of
choosing $\beta_i$ to ensure a certain overlap of the energy distributions between
the two temperatures~\cite{barash:16}. This overlap can be computed from the
resampling factors,
\[
\alpha(\beta_{i-1},\beta_i) = \frac{1}{R_{i-1}}\sum_{j=1}^{R_{i-1}} \min\left(1,
\frac{R\exp[-(\beta_i-\beta_{i-1})E_j]}{R_{i-1}Q(\beta_{i-1},\beta_i)}
\right),
\]
and $\beta_i$ is adapted using a bisection search such as to ensure an overlap
$\alpha^\ast$ of energy histograms. The method is not very sensitive to the precise
value of $\alpha^\ast$, and we choose $\alpha^\ast = 0.86$ in the runs below.
While the algorithm described above is just population annealing~\cite{machta:10a}
improved by adaptive temperature steps~\cite{barash:16,amey:18}, the possibility of
sampling the entropy arises from a combination of the method with multi-histogram
techniques~\cite{ferrenberg:89a}. An estimator of the free energy follows directly
from the resampling factors~\cite{machta:10a},
\begin{equation}
-\beta_i \hat{F}({\beta_i}) = \ln Z_{\beta_0} + \sum_{k=1}^{i} \ln Q_k,
\label{eq:free-energy}
\end{equation}
where $Z_{\beta_0}$ is the partition function at the initial temperature
$\beta_0$. In the following, we always choose $\beta_0 = 0$, such that simply
$Z_{\beta_0}=2^N$, where $N$ is the number of spins. We can then estimate the
DOS by combining the histograms at all temperature steps. As we show in
\ref{sec:dos-estimator}, a variance-optimized estimator is given by
\begin{equation}
\hat{\Omega}(E) = \frac{\sum\limits_{i = 1}^{N_\beta} \hat{H}_{\beta_i}(E) }{\sum\limits_{i = 1}^{N_\beta} R_i
\exp[\beta_i \hat{F}(\beta_i)-\beta_iE]}.
\label{eq:MHReq1}
\end{equation}
Here, $N_\beta$ is the total number of temperatures, and the energy histogram
$\hat{H}_{\beta_i}(E)$ at inverse temperature $\beta_i$ is normalized such that
$\sum_E \hat{H}_{\beta_i}(E) = R_i$. In Eq.~\eqref{eq:MHReq1}, the free-energy
estimate $\hat{F}(\beta_i)$ is taken from Eq.~\eqref{eq:free-energy}. More
sophisticated estimators that can lead to improved results in some cases are
discussed in \ref{sec:dos-estimator}.
This approach is naturally suited for (moderately or massively) parallel calculations
as the $R$ replicas are simulated independently of each other and the only
interaction occurs during resampling. An efficient GPU implementation was discussed
in Ref.~\cite{barash:16}. Importantly for our application, EPA does not require any
prior knowledge of the range of realized energies. Additionally, as we shall see
below, EPA performs better at estimating $\Omega(E)$ for hard spin-glass samples than
the WL-$1/t$ algorithm. A detailed analysis of systematic and statistical errors of
PA can be found in Ref.~\cite{weigel:17a}. Here it is worthwhile to note that
statistical errors can be estimated from a single run by a jackknife blocking
analysis over the population that is introduced in Ref.~\cite{weigel:17a}. This is
further discussed in \ref{sec:error-appendix}. Also, note that it is possible to
include histograms from independent runs in the overall estimate provided through
Eq.~\eqref{eq:MHReq1} by extending the sums over $i$ over the temperature steps of
all runs. The relative weight of these contributions is automatically taken into
account through the free-energy factors deduced from Eq.~\eqref{eq:free-energy} and
the population sizes $R_i$. This is a natural generalization of the weighted averages
first proposed for more basic observables in Ref.~\cite{machta:10a}. This scheme
makes it possible to determine the DOS with arbitrary accuracy in a fixed time given
sufficient parallel resources.
It should be noted that, as it stands, EPA only visits energies in the physical
region $E\lesssim 0$, which is in contrast to WL that naturally also explores
energies $E > 0$. Should one be interested in this unphysical regime, however, it is
possible to derive its DOS from EPA, too. For systems on bipartite graphs the DOS is
always symmetric, $\Omega(-E) = \Omega(E)$, so it is easy to construct the full DOS
while just actually sampling energies $E \le 0$. This is the case for all examples
discussed below. For more general situations, it is also possible to construct the
full DOS by running EPA twice, once for the Hamiltonian $H$ and once for $-H$ and
combining the results.
\section{Results}
In order to test the efficiency of EPA against the WL-$1/t$ algorithm in an objective
way that is unaffected by problems of false convergence, we applied the two
algorithms to the planted solutions on Chimera graphs as well as to the stochastic
$\pm J$ model on the square lattice with periodic boundaries, for both of which we
have exact results. As a baseline for the comparison, we tested both methods for the
case of the Ising ferromagnet on a square lattice for which extensive exact results
are available. There, we find similar performance of the two techniques, see the
discussion in~\ref{sec:ising}.
\subsection{Ising spin glasses on Chimera graphs}
\label{sec:planted}
We first considered the Chimera spin-glass instances with planted solutions and
$N=501$ spins. In order to be able to compare the two algorithms on an equal footing,
we could have directly considered them to be allowed the same runtime. This measure,
however, is implementation and platform specific. Since both algorithms spend most
of their time flipping spins, we compare them for simulations employing the same
number $2\times 10^{12}$ of spin-flip attempts. For the used (serial) code for WL-$1/t$ this
corresponded to a wallclock time of 37~h (on an Intel Xeon 2.4 GHz CPU), while for
EPA we used a massively parallel GPU program~\cite{barash:16} that took approximately
1.2~h (on an Nvidia Tesla K40 GPU) per realization.
As mentioned above, for WL it is required to know the allowed energy levels to decide
about the flatness of histograms. This knowledge was here acquired by a pre-run of
the WL type employing $2\times 10^{11}$ spin-flip attempts and with a fixed
modification factor $\ln f = 1$ to explore the energy landscape (the corresponding
run-time is included in the time estimate given above). This knowledge is not
required for the EPA runs. With a single window covering the full energy range,
WL-$1/t$ did not complete phase 1 for the vast majority of samples. We therefore used
two windows with energy ranges $[E_0,E_0+1200]$ and $[E_0+1100,50]$ in dimensionless
units, respectively. The spin-flips were divided evenly between the two energy windows.
Here, the disorder average of the ground state energy $E_0$ was found to be $-3635$.
The energy levels were determined during the
pre-run, which found the ground state in the vast majority of cases. It is clear that
for larger systems where it is much harder to find the ground state the determination
of suitable windows for WL-$1/t$ becomes much harder. The simulation was started in a
random configuration within the energy range of the window. With that restriction,
$567$ out of $625$ samples completed phase 1 in the first window
within the remaining $8\times 10^{11}$ flip attempts after the pre-run.
No range restriction was required for EPA, and we used a population of size
$R=3\,992\,000$ with $\theta=10$ rounds of spin flips per resampling step and a
histogram overlap $\alpha^\ast = 0.86$, resulting in typically 100 temperature steps
down to $\beta_\mathrm{f} = 5\times 10^4$\footnote{The unusual population size
results from the attempt of matching the average number of spin-flip attempts
between the two methods.}.
\begin{figure}[tb!]
\begin{center}
\includegraphics[width=0.7\textwidth]{Fig_Chimera501_r10}
\end{center}
\caption{ Scatter plot of the relative error of the WL-$1/t$ and the EPA algorithms
in estimating the ratio $r_{10} = \Omega(E_1)/\Omega(E_0)$ of degeneracies for
$N=501$ Chimera spin-glass samples with planted solutions. Both algorithms applied
a total of $2\times 10^{12}$ spin-flip attempts per sample, including an additional
pre-run for WL-$1/t$ to determine the allowed energy values.
For 54 out of 625 samples the deviation for WL-$1/t$ falls outside the
range of the plot, and these are shown at the right edge of the plot in red. }
\label{fig:chimeraPAvsWL}
\end{figure}
We note that both EPA and WL-$1/t$ intrinsically estimate entropy {\em
differences\/}, i.e., ratios of degeneracies for neighboring energy values, and the
absolute scale is only achieved through an additional normalization such as that
given by $Z_{\beta_0}$ in Eq.~(\ref{eq:free-energy}). It is therefore reasonable to
study their performance in estimating
\[
r_{10} = \frac{\Omega(E_1)}{\Omega(E_0)},
\]
the ratio of degeneracies of first excited and ground states. In
Fig.~\ref{fig:chimeraPAvsWL} we show the relative deviations of the ratios $r_{10}$
from the exact values known through the planting, as estimated from WL-$1/t$ and
EPA. WL-$1/t$ found the correct ground-state energy for $622$ of the $625$
samples. For some samples the relative deviations are so large that they exceed the
scale of the plot of Fig.~\ref{fig:chimeraPAvsWL}, some by many orders of
magnitude\footnote{This includes the three samples for which WL-$1/t$ does not find the
ground state, which effectively implies that $r_{10}$ and the relative deviation
are infinite.}. These samples are shown at the boundary of the box and in a
different color. It is clear that for most samples the deviations are substantially
smaller for EPA than for WL-$1/t$. In total, EPA outperforms WL-$1/t$ in 80\% of the
instances. The error of WL-$1/t$ is larger than 7\% for 25\% of samples and it is
difficult to distinguish between the accurate and inaccurate WL results. In
contrast, the EPA results are accurate to within 7\% for all of the $625$ samples.
\subsection{Ising spin glasses on toroidal graphs}
For planar or otherwise two-dimensional lattices of a fixed genus, a counting of
dimer coverings and the corresponding evaluation of Pfaffians can be used to
determine the full DOS in polynomial time
\cite{blackman:91,saul:93,galluccio:00}. We studied toroidal graphs, i.e., $L\times L$
patches of the square lattice with periodic boundary conditions using the
implementation proposed in Ref.~\cite{galluccio:00} which has an asymptotic run-time
scaling of O($L^5$). Using this approach, we evaluated $1000$ samples with a standard
$\pm J$ coupling distribution and $32\times 32$ spins, and also $500$ samples of size
$48\times 48$. An example of $\ln \Omega(E)$ as estimated for a single sample of size
$L=48$ from EPA is shown in the top panel of Fig.~\ref{fig:DOS-examples}. At this
scale, the data are completely indistinguishable from the exact result also shown for
comparison. As one reads off from the graph, the actual DOS $\Omega(E)$ spans about
700 orders of magnitude, and it is quite remarkable that it can be estimated so
accurately from the simulations. To systematically assess the accuracy of the
sampling for different disorder realizations, we considered the total deviation $\Delta$
of the simulation results from the exact DOS, where
\begin{equation}
\Delta = \frac{1}{N_E}\sum_{i=1}^{N_E} \left|\frac{\ln \Omega(E_i)-
\ln \Omega_{\mathrm{exact}}(E_i)}{\ln \Omega_{\mathrm{exact}}(E_i)} \right|.
\label{eq:delta}
\end{equation}
While for PA an absolute normalization of $\Omega(E)$~\cite{weigel:02a} follows from
the free-energy estimator Eq.~(\ref{eq:free-energy}) in combination with
Eq.~(\ref{eq:MHReq1}), WL-$1/t$ as described above only yields the DOS up to an
overall factor. To fix the latter we used the fact that the sum $\sum \Omega(E)$ over
all energy levels must equal the total number $2^N$ of states. Note that different
ways of normalization of WL-$1/t$ lead to quite different fluctuations of the final
DOS estimates, and the normalization via the total number of states used here leads
to the best results, see the discussion in~\ref{sec:normalization}. Since EPA
only samples states with energies $E \lesssim 0$, we restricted the energy range for
WL-$1/t$ to $E\le 50$ to ensure a fair comparison\footnote{For the normalization of
the DOS we hence completed $\Omega(E) = \Omega(-E)$ for the positive energies. For
consistency, we additionally applied the same normalization in EPA.}.
For $32\times 32$ samples, we used $1.8\times 10^{12}$ spin-flip attempts in the main
run of WL-$1/t$ employing a single energy window with $E\le 50$. Just as for the Chimera
samples, a pre-run was required to determine the range of possible energies, for
which an additional $2\times 10^{11}$ updates were used. For the EPA algorithm, we
used a population size $R=2\,340\,000$, and performed $\theta=19$ rounds of
spin-flips between two resampling steps. The imposed histogram overlap of
$\alpha^\ast = 0.55$ resulted in $N_\beta= 44$ temperature steps for most disorder
realizations down to $\beta_f = 5\times 10^4$. The total number of spin flips in
these EPA runs is hence also approximately $2\times 10^{12}$. For system size
$48\times 48$, WL-$1/t$ required two energy windows to converge; these were chosen as
$[E_0,E_0+64]$ and $[E_0+36,50]$. After the pre-run of length $6\times 10^{11}$
across both windows, we used $3.1\times 10^{12}$ spin-flip attempts in the main run
of the first window and $1.0\times 10^{12}$ updates in the second window. The two DOS
segments obtained by WL-$1/t$ were sewn together by matching the estimates at a point
in the intersection of the two windows. For the EPA algorithm, we used
$R=1\,019\,965$, $\theta=10$ and $\alpha^\ast = 0.86$, which resulted in
$N_\beta = 200$ ($\beta_\mathrm{f} = 3$) and hence the total number of spin flips is
$4.7\times 10^{12}$ as for the WL-$1/t$ runs.
\begin{figure}[tb!]
\begin{center}
\includegraphics[width=0.7\textwidth]{DOS-2D-L48}
\includegraphics[width=0.7\textwidth]{DOS-3D-L14}
\end{center}
\caption{ Examples of the DOS for the Edwards-Anderson spin glass. Top: Results from
EPA runs for a single $L=48$ toroidal lattice sample (points) as compared to the
exact DOS calculated from the Pfaffian method. Bottom: DOS estimated from EPA runs
for 20 $L=14$ 3D $\pm J$ samples. }
\label{fig:DOS-examples}
\end{figure}
The resulting values for the average relative deviation of the level entropies
according to Eq.~(\ref{eq:delta}) are shown in the scatter plots provided in
Fig.~\ref{fig:toroidalPAvsWL}. The top panel corresponds to 1000 samples of size
$32\times 32$. In some cases the ground states were not found or the first phase of
WL-$1/t$ did not complete, leading to extremely large or infinite deviations; the
corresponding samples are shown in red at the boundary of the plot. In this case we
only find a moderate advantage for EPA, which outperforms WL-$1/t$ on 516 of the
1000 samples. Considering the larger system size $L=48$, the advantage of EPA
increases, leading to a smaller value of $\Delta$ for 291 samples out of the 451
samples where both methods found all energy levels. This observation is in line with
a general trend of EPA faring relatively better for harder problems as compared to WL
that we shall see confirmed for other examples below.
\begin{figure}[tb!]
\begin{center}
\includegraphics[width=0.55\textwidth]{Fig_Toroidal_L32}
\includegraphics[width=0.55\textwidth]{Fig_Toroidal_L48}
\end{center}
\caption{Scatter plots of the average relative deviation of level entropies from the
exact result according to Eq.~(\ref{eq:delta}) for toroidal $\pm J$ spin-glass
samples of size $L\times L$ spins as estimated by the WL-$1/t$ and EPA methods. Top
panel: $L=32$. Both algorithms were run for a total of $2\times 10^{12}$ attempted
spin flips per sample. EPA outperforms WL-$1/t$ for 52\% of instances (516 out of
1000). For 79 samples the deviations for WL-$1/t$ fall outside the scale of the
plot and these samples are drawn at the right edge, in red. Bottom panel:
$L=48$. Both algorithms were run for a total of $4.7\times 10^{12}$ attempted spin
flips per sample. EPA outperforms WL-$1/t$ for 62\% of instances (312 out of
500).}
\label{fig:toroidalPAvsWL}
\end{figure}
\subsection{Three-dimensional Ising spin glasses}
\label{sec:3d_samples}
Concerning the trend of improving relative performance of EPA for harder problems, it
is interesting to see how the two samplers perform on spin-glass instances in three
dimensions, where the spin-glass problem is known to be NP-hard
\cite{barahona:82a}. To this end, we studied samples of the $\pm J$ Edwards-Anderson
model with $L=8$ and $L=14$. For $L=8$ we were able to employ a single energy window
for WL, and the parameters for EPA were $R=1\,992\,984$, $\theta = 10$,
$\alpha^\ast = 0.86$, and $\beta_\mathrm{f} = 3$, using $10^{12}$ spin-flip attempts
in both cases. For $L=14$, $5.5\times 10^{12}$ spin flips were applied in each
run. For the WL-$1/t$ method we used two windows, $[E_0, E_0 + 64]$ and
$[E_0 + 36, 50]$, with $3.9\times 10^{12}$ and $1.0\times 10^{12}$ in the main run,
respectively. The remaining $6\times 10^{11}$ spin-flip attempts were used in the
pre-run. The parameters for EPA were $R = 867\,694$, $\theta = 10$,
$\alpha^\ast = 0.86$, and $\beta_\mathrm{f} = 10$. In the bottom panel of
Fig.~\ref{fig:DOS-examples} we show $\ln \Omega(E)$ for the $L=14$ samples. The
sample-to-sample fluctuations in the DOS are in fact rather small and can only be
seen in the very low-energy part of the spectrum (as well as its mirror image for
large positive energies).
As the samples considered are neither planar nor planted, we do not have access to
the exact DOS, and we hence quantify the success of the two algorithms in estimating
the DOS by determining the level of fluctuation in the estimates of $\ln \Omega(E)$
between independent runs, both for the WL-$1/t$ and EPA methods. Specifically, we
estimated the relative standard deviation $\sigma[\ln \Omega(E)]/\ln \Omega(E)$ from
200 independent runs for each disorder sample. In Fig.~\ref{fig:3dsamples} we show
this quantity, averaged over all energy levels, for 20 $\pm J$ 3D spin-glass samples
of the two system sizes considered. While for $L=8$ the WL runs yield slightly
smaller error bars, for $L=14$ the situation is reversed, with EPA resulting in 5
times smaller error bars, on average, corresponding to saving a factor of 25 in
run-time.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.7\textwidth]{Fig_3D_L8}
\includegraphics[width=0.7\textwidth]{Fig_3D_L14}
\end{center}
\caption{Average relative standard deviations of level entropies,
$\overline{\sigma[\ln(\Omega(E))]/\ln \Omega(E)}$, resulting from WL (red dots) and
EPA (blue diamonds) simulations for $20$ 3D $\pm J$ Ising spin-glass samples of
sizes $L=8$ (top) and $L=14$ (bottom), respectively. Here, the average is over
energy levels. Note that for the WL result, we only included the 170 out of 200
runs where the first phase completed successfully for all samples.}
\label{fig:3dsamples}
\end{figure}
\subsection{Entropic sampling of problems of varying hardness}
\label{sec:varying_hardness}
Having ascertained that the EPA method can yield significantly better approximations
to the DOS of some hard problems in a given number of steps than the WL approach, it
is interesting to analyze more closely the actual distribution of performances of
these algorithms over the space of disorder realizations. The results above in
Figs.~\ref{fig:chimeraPAvsWL} and \ref{fig:toroidalPAvsWL} indicate the presence of
larger fluctuations in the quality of approximation for WL-$1/t$ as compared to EPA
across disorder realizations. To study this effect quantitatively, we used a set of
disorder samples classified according to their {\em thermal hardness\/}. A well
established measure of such hardness is the exponential autocorrelation or relaxation
time in parallel tempering simulations~\cite{hukushima:96a,sokal:97}. As very long
simulations are required to determine these time scales accurately, a number of proxy
quantities such as the so-called ``tunneling time'' are frequently used in practical
applications~\cite{katzgraber:06,bittner:08}. Here, we rely on a method developed in
the context of spin-glass simulations that analyses the dynamics of the random walk
of replicas in temperature space~\cite{banos:10} and extracts the corresponding
relaxation times $\tau$.
To benchmark the EPA algorithm against WL-$1/t$, we generated about $10^6$ random
instances on an $N=512$-spin Chimera graph (of which only 476 spins were used) and
measured the relaxation times $\tau$ of each instance with parallel
tempering\footnote{Specifically, we chose a temperature grid of the PT simulations
consisting of $N_T=30$ temperatures. Temperatures with indices $i=1,2,\ldots,12$
were uniformly distributed in the range $T_\mathrm{min}=0.045 \leq T_i \leq 0.2$,
while the temperatures $T_i$ with $i=13,14,\ldots, N_T$ were spread evenly in the
range $0.21 \leq T_i \leq T_\mathrm{max}=1.632$~\cite{scirep15:mm-hen}.}. Next, we
grouped together instances with similar classical hardness, i.e., similar relaxation
times $\tau$, $10^k \leq \tau \leq 3 \cdot 10^k$ for $k=3$, $4$, $5$, $6$ and
$7$. For each such `generation' of $\tau$, we randomly picked $100$ representative
instances for the benchmarking of the algorithm (only 14 instances with $k=7$ were
found). We then performed WL-$1/t$ simulations with a total of $10^{12}$ spin-flip
attempts for all samples, evenly split between one simulation each restricted to the
energy windows $[-900,-500]$ and $[-550,50]$, respectively (the ground-state energy
for these samples is roughly $E_0 \approx -800$). A pre-run of $2\times 10^{11}$
spin-flip attempts was again used to discover the range of possible energies for each
sample. All runs completed the first phase of the simulation here, owing to the use
of two energy windows. For the PA runs we used $R=2.1\times 10^6$, $\theta=10$,
$\alpha^\ast = 0.86$, corresponding to $N_\beta \approx 100$ temperature steps
($\beta_\mathrm{f} = 5$) and $10^{12}$ spin flip attempts. The DOS estimates from
both methods are only considered for $E \le 0$ and $\Omega(E) = \Omega(-E)$ is used
for $E > 0$. The resulting DOS estimate is normalized using the known total number of
states $2^N$.
As for the 3D samples, we compared the two algorithms by considering the relative
standard deviations $\sigma[\ln \Omega(E)]/\ln \Omega(E)$, averaged over all
energies. The resulting estimates are shown in Fig.~\ref{fig:isingHardness} for the
samples of the different hardness classes $k = 3$, $\ldots, 7$. It is clear that
EPA is less affected by sample hardness than WL-$1/t$, with the growth in fluctuation
with sample hardness being much steeper for WL-$1/t$ than for EPA. Note that this
quantity only covers the effect of statistical errors, whereas the data in
Figs.~\ref{fig:chimeraPAvsWL} and \ref{fig:toroidalPAvsWL} considered the total
deviation from the exact results that also includes bias effects. Note also that the
sample-to-sample fluctuations, represented in the error bars of the data points in
Fig.~\ref{fig:isingHardness}, are significantly larger for WL-$1/t$ than for EPA. We
find that WL-$1/t$ and EPA have rather different behavior in sampling the DOS in
different energy ranges, with WL-$1/t$ being more focused on higher energies, for
details see~\ref{sec:energy-dependence}.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.7\textwidth]{Fig_Varying_Hardness}
\end{center}
\caption{\label{fig:isingHardness} Performance of the EPA and WL-$1/t$ estimates of
the DOS on spin-glass samples of varying hardness. The plot shows the average standard
deviation of level entropies as a function of the hardness group in terms of the
parallel-tempering relaxation time $\tau$. WL-$1/t$ was run in each of the two
energy windows, and the obtained DOS was normalized to the total number of states.
The performance of WL-$1/t$ decreases sub-linearly with the problem difficulty,
$\overline{\sigma(\ln\Omega(E))/\ln\Omega(E)} \sim \tau^{0.35}$,
while for EPA $\overline{\sigma(\ln\Omega(E))/\ln\Omega(E)} \sim \tau^{0.19}$.
The results are averaged over 14 samples from each hardness class with the error
bars resulting from the sample-to-sample fluctuations. }
\end{figure}
While in the present demonstration we used relaxation times from parallel tempering
for classifying sample hardness, it is worthwhile to note that EPA can itself provide
a hardness measure and thereby differentiate easy and hard samples. A few such
quantities have been previously proposed for population annealing~\cite{wang:15a}. We
consider here in particular the (temperature dependent) mean-square family size
$\rho_t$, defined as~\cite{wang:15a}
\begin{equation}
\rho_t = R \sum_i{\frak n}_i^2,
\label{eq:rhot}
\end{equation}
where ${\frak n}_i$ is the fraction of the current population that descends from the
$i$-th member of the initial population at $\beta_0 = 0$, while $R$ corresponds to
the initial population size. The quantity $R/\rho_t$ can be understood as an
effective population size, corresponding to the number of statistically independent
replicas, such that $R/\rho_t = R$ corresponds to a perfectly uncorrelated
population, while $R/\rho_t \to 0$ for the strongest correlations. These two limits
hence represent the easiest and hardest samples, for which one would expect
$\tau \to 0$ and $\tau \to \infty$, respectively, for parallel tempering. A related
quantity that also takes the decorrelating effect of spin flips into account is the
effective population size $R_\mathrm{eff}$ defined in Ref.~\cite{weigel:17}. In
Fig.~\ref{fig:rho} we show a scatter plot of $\rho_t$ for 100 samples of each of the hardness
classes $k < 7$ and 14 samples for $k=7$, respectively. The disorder average of
$\rho_t$ at $\beta = 3$ is found to be $49$, $135$, $420$, $663$ and $840$ for $k=3$,
$4$, $5$, $6$, and $7$, respectively, indicating that while for the main part of the
distribution the hardnesses in EPA and parallel tempering are strongly correlated,
for the tails of the distribution the hardness in EPA increases more gently than that
found in parallel tempering. As is demonstrated elsewhere, these intrinsic hardness
measures can be used to make population annealing simulations adaptive to the sample
hardness~\cite{amey:18,weigel:17a}. We note that the planted samples of
Sec.~\ref{sec:planted} have an average $\rho_t$ of $\approx 2000$
(see~\ref{appendix:rhot_planted}), indicating that planted samples
of this type are much harder than random ones.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.7\textwidth]{rhot}
\end{center}
\caption{\label{fig:rho} Mean-square family size $\rho_t$ at the lowest temperature
in EPA for the varying-hardness samples (100 for $k=3$, $4$, $5$ and $6$ and 14 for
$k=7$) for each hardness group. Larger average family sizes result from stronger
correlations in the PA ensembles of replicas and hence indicate the difficulty of
the algorithm in equilibrating the system. The red diamonds mark the disorder
average of $\rho_t$ at $\beta=3$ for each of the hardness groups. }
\end{figure}
\section{Summary and discussion}
We have investigated the performance of sampling methods for estimating the DOS
for systems with complex free-energy landscapes, focusing on spin glasses
as the hardest problems among spin systems. We proposed a novel sampling technique
based on sequential Monte Carlo on a large population of copies and demonstrated that
it outperforms the most widely used entropic sampler, the Wang-Landau algorithm, in
the vast majority of cases. More importantly, the new approach shows better scaling
as the hardness of problems is increased (for example through considering larger
systems).
A notorious problem with benchmarking algorithms for estimating the DOS
lies in safely assessing convergence. Here we address this issue by considering
problems that are either of planar topology, in which case Pfaffian methods can be
used to determine the DOS exactly for systems with more than 1000
spins, or which have planted solutions such that the exact degeneracies of the ground
and first excited states can be calculated using a `bucket' algorithm. Both classes
provide hard optimization problems, implying a very non-trivial benchmark. In
addition, we also considered more general problems such as stochastic spin-glass
samples on Chimera graphs sorted by thermal hardness as well as the most challenging
case of three-dimensional spin-glass instances of up to $14\times 14\times 14$ spins.
One essential advantage of the approach based on population annealing is that it
does not require any prior knowledge about the energy spectrum, which in contrast
needs to be acquired in an additional pre-run for the Wang-Landau
method. Furthermore, the well-known and delicate problem of dividing the energy range
into windows that is the only way of making Wang-Landau simulations converge for the
more challenging cases, is completely absent for entropic population annealing. The
difficulty of premature and false convergence that plagues Wang-Landau and related
methods is not so much of an issue for the newly introduced technique, where a
re-distribution of weights can occur at all stages of the algorithm. In fact, for
the EPA method it is easily possible to monitor equilibration from intrinsic
properties and only output the DOS for energies where thermalization could be
ensured. The main advantage of the approach, however, lies in the ideal suitability
for massively parallel calculations, where given sufficient parallel resources the
accuracy of the approximation can be arbitrarily improved at a constant wall-clock
time by increasing the size of the population or combining the outcome of independent
runs in a weighted average.
The specific spin system we study, namely spin glasses, is very relevant as current
experimental quantum annealers attempt to solve precisely this type of problem. With
questions still lingering about which distribution these devices sample from
\cite{marshallrieffelhen:2017}, it is important to have an accurate tool to estimate
the DOS (for instance to understand
thermalization~\cite{marshallrieffelhen:2017}). For this problem we specifically
consider instances with vastly different hardnesses, confirming that the accuracy of
the technique proposed degrades significantly less for harder samples that previous
approaches. We note that the entropic population annealing approach is in no way
specific to the spin systems considered here, and it can be straightforwardly
generalized to other problems such as lattice polymers and, with the help of binning
or spectral methods \cite{farris:19}, to cases with continuous degrees of freedom.
Our results are very promising as we could clearly show that i) for a range of
problems with complex energy landscapes existing sampling methods for the DOS
are difficult to set up and do not converge very reliably, especially for hard
samples; ii) a dependable method for the benchmarking of entropic samplers on spin
glasses is now available, which we hope will drive research forward in finding even
better algorithms; and iii) the entropic population annealing (EPA) algorithm devised
here allows for the reliable sampling of large-scale frustrated systems. We therefore
trust that the algorithm will become a useful tool for DOS calculation
in condensed matter physics, quantum computing and other areas of research.
\ack
Part of the computing resources were provided by the USC Center for High Performance
Computing and Communications and the Oak Ridge Leadership Computing Facility at the
Oak Ridge National Laboratory, which is supported by the Office of Science of the
U.S. Department of Energy under Contract No. DE-AC05-00OR22725. The research is
based upon work (partially) supported by the Office of the Director of National
Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via
the U.S. Army Research Office contract W911NF-17-C-0050. The views and conclusions
contained herein are those of the authors and should not be interpreted as
necessarily representing the official policies or endorsements, either expressed or
implied, of the ODNI, IARPA, or the U.S. Government. This material is based on
research sponsored by AFRL under agreement number FA8750-18-1-0109. The
U.S. Government is authorized to reproduce and distribute reprints for Governmental
purposes notwithstanding any copyright annotation thereon. LB was supported by grant
No.\ 14-21-00158 from the Russian Science Foundation. LB and MW acknowledge support
by the European Commission through the IRSES network DIONICOS under Contract No.\
PIRSES-GA-2013-612707.
|
\section{Introduction}
\label{sec_in}
We consider a system of $N$ spherical-cap fluid droplets protruding from uniform, circular orifices on a flat surface. The droplets are connected through
straight fluid channels of uniform circular cross section and possibly variable length. The fluid is homogeneous and incompressible. The rheology of the fluid is governed by the Ostwald-de Waele power law, thus, in particular, permitting shear thinning which is often observed in complex (e.g.\,biological) fluids. The flow dynamics are dominated by surface tension acting on the droplets and viscous forces within the network channels. The pressure acting on each droplet is given by the Young-Laplace relation. Our physics-based model accounts for scavenging amongst neighboring spherical-cap droplets owing to the action of capillary pressures due to surface tension. Volume exchange arises by pressure (curvature) differences that drive liquid from one to another droplet along a network of interconnected channels. In this way, certain droplets gain volume at the expense of others. This mechanism, christened ``volume scavenging,'' leads to interesting dynamics, driving an initial droplet configuration to a stable equilibrium.
\begin{figure}[tbhp]
\centering
\includegraphics[width=0.7\textwidth]{array.jpg}
\caption{An array of $10\times 10$ fluid droplets protruding from holes of $0.5$ millimeter diameter and connected through a reservoir beneath, from \cite{VoSt}}
\label{drop-array}
\end{figure}
An array of fluid droplets connected to a common reservoir is displayed in \cref{drop-array}.
In the Newtonian case a small perturbation of an initial configuration of super-hemispherical drop\-lets of equal size is known to evolve over time into a configuration consisting of one ``winning'' super-hemi\-spher\-i\-cal droplet with all other droplets being sub-hemispherical of equal size. Since this evolution is driven by a minimization of the total surface area, volume scavenging is a coarsening process like Ostwald ripening \cite{RaVo}.
\begin{snugshade}
\noindent {\em Note to the reader}: Throughout this work we will highlight results of our physics-based model from different viewpoints, suggesting parallels to the socio-economic context among others. We will also offer ``asides."
\end{snugshade}
\begin{snugshade}
\noindent {\em Material science/fluid dynamics aside}:
Coarsening phenomena are observed in a large class of problems in material science and fluid dynamics, including binary mixtures and alloys, oil-water emulsions, epitaxial growth and re-crystallization.
\end{snugshade}
\begin{snugshade}\noindent {\em Socio-economic view}: Interestingly, coarsening has also been observed in the social sciences: Schelling's agent-based model of segregation, possibly the best known representative of segregation models, impressively demonstrates that the local interaction of individuals can lead to unexpected aggregate behavior (coarsening) \cite{Sc-art,Sc}.
\end{snugshade}
The mathematical model we study is a generalization of the Newtonian volume scavenging model introduced by van Lengerich, Vogel and Steen \cite{LeEA1,LeEA2}. That model was used to explain the grab-and-release mechanism for a
capillary adhesion device, described by Vogel and Steen \cite{VoSt}, and was motivated by a study of Eisner and Aneshansley \cite{EiAn} about the tarsal adhesion of the Florida tortoise beetle ({\em Hemisphaerota cyanea}) to defend against predators. In this work we consider the generalization of the previous model to power-law fluids. This generalization has far-reaching consequences, both for the analytical treatment of the governing equations and for the observed dynamics of solutions. Our analysis, while extending some of the previous results of van Lengerich et.\,al.\,\cite{LeEA1,LeEA2}, will do so without regard of the underlying conduit networks and without use of linearization techniques which are generally not available in the case of the power-law model. We will give a complete classification of possible equilibria and their stability with the dimensionless average droplet volume $\overline{V}$ serving as a bifurcation parameter. The {\em full} range
$\overline{V}>0$ will be discussed. One of our central results will be an {\emph {exhaustive}}, analytical identification of hierarchies of stationary droplet configurations, classified by the size and number of large droplets versus small ones.
Volume scavenging has been the focal point of several important studies. Let us mention a few: Wente \cite{We} analyzed the two- and three-droplet regime from the vantage point of catastrophe theory. His seminal work made use of an energy formulation to classify the stability of equilibria. Slater and Steen \cite{SlSt} discussed the inviscid case of $N$ spherical-cap droplets under the symmetry assumption of equivariance with respect to the permutation group $S_N$. Stability results were also given by van Lengerich et.\,al.\,\cite{LeEA1,LeEA2} for the Newtonian case with constant viscosity and large average volumes $\overline{V}>1$. This was achieved by linearization about hyperbolic equilibria for a conduit network where the linearized equations turned out to be particularly simple. The existence of a network-independent Lyapunov function made it then possible to extend the stability results to general networks. These works are the starting point of our own study. Some of the findings reported there will be included as special cases
in this article. Yet far from being narrowly tailored, our analysis provides a template of arguments which is likely to generalize and be of use beyond this specific problem.
We base our governing equations on the dimensionless model introduced in \cite{LeEA1}. In this framework the uniform radius $R$ of each circular opening is rescaled and non-di\-men\-sion\-al\-ized to be $1$. The same length scale is used to rescale the radius $r$ and the height $h$ of a spherical-cap droplet protruding from an orifice. Droplet volume $v$ is normalized such that the volume of a hemispherical droplet is $1$. Then radius $r$, height $h$ and volume $v$ of the droplet are related by
\begin{equation}\label{randv}
{2}\,r = h+\displaystyle{1\over h}\quad \text{and}\quad v = \displaystyle{1\over 4}\,h\,(h^2 +3).
\end{equation}
\begin{wrapfigure}{r}{0.46\textwidth}
\vspace*{-0.5cm}
\centering
\includegraphics[width=0.45\textwidth]{dropletpic2.jpg}
\vspace*{-0.7cm}
\caption{A spherical-cap droplet ($h>R$)}
\label{droplet}
\end{wrapfigure}
A schematic of the spherical-cap droplet geometry is given in Figure~\ref{droplet}. It is also worthwhile to record the corresponding normalized surface area $s_A=s_A(h)$ of the droplet in terms of its height $h$:
\begin{equation}\label{Adefn}
s_A={3\over 2}\,\left(h^2+1\right).
\end{equation}
Note that $v=v(h)$ is invertible for $h\in \mathbb R$ (with $h\geq 0$ being of physical importance in our situation) and that, due to the chosen scalings, $v=1$ if and only if $h=1$. Moreover, we have, of course,
\begin{equation}\label{vhasymp}
|h| = O\left(|v|^{1/3}\right)\quad \text{as $|v|\rightarrow \infty$.}
\end{equation}
The pressure $p$ acting on the spherical-cap droplet is caused by surface tension and given by the Young-Laplace law (after non-dimensionalization):
\begin{equation}\label{ph}
p = {2\over r}={{4\,h}\over {h^2+1}}.
\end{equation}
Hence the capillary pressure is the same for spherical-cap droplets of the same radius of curvature $r$. More precisely, if $h_0$ is a solution of the equation $p(h)=\lambda$ with $\lambda\in {\mathbb R}_+$, then by \eqref{randv}, \eqref{ph}, $h_0^{-1}$ is also a solution. In fact, $h_0$ and $h_0^{-1}$ are the only solutions. They are distinct if and only if $\lambda\not= 1$. Consequently, similar to the terminology used in \cite{LeEA1,LeEA2}, we call a spherical-cap droplet of height $h$ ``large'' if $h>1$ ($v>1$) and ``small'' if $0\leq h\leq 1$ ($0\leq v\leq 1$).
Because of the relations between radius $r$, height $h$ and volume $v$ given above, we can write the pressure $p=p(h)$ in terms of the droplet volume by setting
\begin{equation}
P(V) = p(h)\quad \text{whenever $V=v(h)$.}
\end{equation}
In this way droplet pressure $P$ is defined as a function of droplet volume $V$. In a similar way, we obtain the surface area of a spherical-cap droplet as a function of its volume via
\begin{equation}
S(V) = s_A(h)\quad \text{whenever $V=v(h)$.}
\end{equation}
\subsection{Networks of interaction}
\label{networks}
\begin{figure}[tbhp]
\centering
\vspace*{-0.3cm}
\subfloat[Complete network]{\label{graph1}
\includegraphics[width=0.25\textwidth]{complete.jpg}}
\subfloat[Star network]{\label{graph2}
\includegraphics[width=0.25\textwidth]{star.jpg}}
\subfloat[Linear Network]{\label{graph3}
\includegraphics[width=0.25\textwidth]{linear.jpg}}
\caption{Simple, connected graphs with 5 vertices}
\label{graphs}
\end{figure}
We now consider a network of conduits of circular cross section (with uniform radius) and possibly variable length connecting $N$ ($\geq 2$) spherical-cap fluid drop\-lets. The network is described by a simple, connected graph whose adjacency relation is given through a weighted adjacency matrix $(c_{i,j})$: For $1\leq i, j\leq N$,
\begin{align}
& c_{i,j} = c_{j,i}>0 \quad \text{if there is a channel connecting droplets $i$ and $j$ ($i\not= j$),}\\
& c_{i,j} = 0 \quad \text{in all other cases.}
\end{align}
The size of $c_{i,j}$ may be interpreted as a measure for the (inverse) length of the channel between droplets $i$ and $j$. In a uniform network with conduits of equal length we may assume $c_{i,j}\in \{0,1\}$ after rescaling. Examples of networks (graphs) with $N=5$ are given in Figure~\ref{graphs}. Important networks include the following simple, connected graphs:
\begin{itemize}
\item {\em Complete network:} Each vertex is adjacent to every other vertex.
\item {\em Star network:} Exactly one vertex (star center) is adjacent to every other vertex. No other edges are present.
\item {\em Linear network:} Each vertex is adjacent to no more than two other vertices. Exactly two vertices are adjacent to only one other vertex.
\end{itemize}
\begin{snugshade}
\noindent {\em Socio-economic view}: The network topology determines the level and efficiency of the $N$ individuals (droplet volumes) interacting with each other. Both local and global interaction of individuals can be modeled.
\end{snugshade}
\subsection{Rates of exchange}
\label{exchange}
We denote the volume of droplet $j$ at time $t$ by $V_j=V_j(t)$ and the volumetric flow rate from droplet $i$ to droplet $j$ at time $t$ by $q_{i,j}=q_{i,j}(t)$. Then we obtain the
change in volume $V_j$ at time $t$ from conservation of mass (volume):
\begin{equation}
{d\over {dt}} V_j = \sum_{i=1}^N q_{i,j},\quad 1\leq j\leq N.
\end{equation}
For a power-law fluid the volumetric flow rate $q_{i,j}$ is assumed to be given by the dimensionless closed-form flow rate--pressure change expression
\begin{equation}\label{fr-pd}
q_{i,j} = c_{i,j}\, \Delta_s P_{i,j},\quad 1\leq i, j\leq N,
\end{equation}
where
\begin{equation}\label{defDs}
\Delta_s P_{i,j} = |P(V_i)-P(V_j)|^s\,\text{sgn} \left(P(V_i)-P(V_j)\right)
\end{equation}
and the power-law parameter $s>0$ is the reciprocal of the power-law index of the fluid, see \cite{BiEA}. The case $s=1$ reduces this relationship to laminar pipe flow of
a Newtonian liquid. This is the situation studied in \cite{LeEA1,LeEA2}. For $s>1$ the fluid is shear thinning, i.e.\,the apparent viscosity decreases with increased stress. For $0<s<1$ the fluid is shear thickening where the opposite flow behavior is observed. Shear thinning is commonly seen for real fluids (including biological fluids, paints and foods), while shear thickening is rare. Observed values of the power-law index in the case of shear-thinning fluids are listed in \cite{BiEA,Ta}. The corresponding values of the
power-law parameter $s$ fall roughly in the interval $1\leq s\leq 5$. In contrast, rheological data for shear-thickening behavior are scarce. In fact, we will observe in the next section that the shear thickening parameter range $0<s<1$ is problematic in our model.
\begin{snugshade}
\noindent {\em Socio-economic view}: The liquid rheology determines the ``trading friction" or inefficiency of resource exchange between the individuals (apparent viscosity). In the Newtonian regime ($s=1$) the trading friction is constant. In the shear thinning regime ($s>1$) the trading friction is smaller when the resource exchange is faster.
It is worthwhile to note that all competitors perform according to a fixed set of rules (governing equations) and the results are deterministic. There are no choices or decisions. While this rigidity disallows individuals to change course, one may interpret the resulting behavior as being optimal under the given rules.
\end{snugshade}
\begin{snugshade}
\noindent {\em Socio-economic view}: It is insightful to relate volume scavenging to the socio-economic concept of the
``winner-take-all'' system (or market), a notion systematically explored by Frank and Cook in their seminal work \cite{FrCo-art, FrCo}. Such a system is
characterized by a concentration phenomenon where a disproportionately large reward falls to one or a few winners, even though other competitors start out with comparable (or even slightly more) resources and perform only marginally worse
than the winner. The losing competitors are not rewarded. Winner-take-all systems are observed in sports, the entertainment industry,
the rise of multi-national companies and elections. The ``egalitarian'' or ``all-share-evenly" system exhibits the opposite outcome. Here, comparable initial resources and similar performance leads to comparable outcomes. Remarkably, both winner-take-all and all-share-evenly behaviors can be observed within our physics-based model. This is an unusual feature since models which
follow the ``cumulative advantage principle" typically exhibit concentrated (coarsened) outcomes like the winner-take-all outcome only. In our model the distinction of winner-take-all outcome versus egalitarian outcome depends on the average resource per individual
$\overline{V}$ (more precisely: the average droplet volume $\overline{V}$).
\end{snugshade}
\section{The Competition: Motivation and Numerical Results}
\label{sec_mot}
The flow is modeled by the system
\begin{equation}\label{sys}
{d\over {dt}} V_j = \sum_{i=1}^N c_{i,j}\, \Delta_s P_{i,j},\quad 1\leq j\leq N
\end{equation}
for the unknown droplet volumes $V_j=V_j(t)$
together with the initial volume configuration
\begin{equation}\label{ini}
V_j(0) = v_j, \quad 1\leq j\leq N.
\end{equation}
The initial value problem \cref{sys,ini} has a local--in time solution for every $s>0$ (and every choice of initial data). Solutions are unique for every $s\geq 1$. In the case $0<s<1$, solutions might lose uniqueness at points where the right-hand side of \cref{sys} is not Lipschitz continuous. Indeed, the occurrence of non-unique solutions is readily confirmed when $0<s<1$ and $N=2$.
Non-uniqueness is clearly an unphysical artifact of our model. Instead of excluding the case $0<s<1$ outright, we carry it along as a mathematical curiosity and address some resultant challenges of analytical interest.
Let us give an elementary characterization of the equilibria of system \cref{sys}. The connectedness of the conduit network immediately implies:
\begin{proposition}\label{stat}
For\, ${\mathbf V}^* = \left(V^*_1,\ldots,V^*_N\right) \in {\mathbb R}^N$ to be an equilibrium of the system \cref{sys}, it is necessary and sufficient that
$
P\left(V^*_i\right) = P\left(V^*_j\right)$, $1\leq i, j\leq N.
$
\end{proposition}
\begin{snugshade}
\noindent {\em Socio-economic view}: At equilibrium the ``push and pull" by capillary pressures are the same for all competitors.
\end{snugshade}
For $0<s<1$, Lipschitz continuity of the right-hand side of \cref{sys} is lost at equilibria.
The model equations given by \cref{sys} are based on conservation of volume (mass). Indeed, we obtain immediately from the symmetry of the matrix $(c_{i,j})$:
\begin{proposition}\label{masscon}
Any solution ${\mathbf V}={\mathbf V}(t)=\left(V_1(t),\ldots,V_N(t)\right)$ of the system
\cref{sys} for $0\leq t\leq T$ satisfies
$\sum_{j=1}^N V_j(t) =\sum_{j=1}^N V_j(0)$.
\end{proposition}
Hence the average droplet volume
\begin{equation}
\overline{V} = {1\over N}\,\sum_{j=1}^N V_j(t)
\end{equation}
is an invariant of the system.
Let us explicitly state the underlying
(physically relevant)
\begin{equation}
\text{\it{Standing Assumption}:}\qquad \overline{V}>0
\end{equation}
and define:
\begin{definition}
For given $\overline{V}$, let
\begin{equation}
{\mathbb V}\left(\overline{V}\right) = \left\{{\mathbf v} = (v_1,\ldots,v_N)\in {\mathbb R}^N\,\Biggr|\, {1\over N}\,\sum_{j=1}^N v_j = \overline{V}\right\}.
\end{equation}
\end{definition}
By \cref{masscon}, the set ${\mathbb V}\left(\overline{V}\right)$ is forward invariant in the sense that any solution ${\mathbf V}={\mathbf V}(t)$
of \cref{sys,ini} for $0\leq t\leq T$ with initial data in ${\mathbb V}\left(\overline{V}\right)$ takes values in ${\mathbb V}\left(\overline{V}\right)$ for $0\leq t \leq T$.
\begin{figure}[tbhp]
\centering
\subfloat[$s=1.0$]{\label{s10}
\includegraphics[width=0.48\textwidth]{s1p0.pdf}}
\subfloat[$s=1.1$]{\label{s11}
\includegraphics[width=0.48\textwidth]{s1p1.pdf}}\vspace*{-0.4cm}
\newline
\centering
\subfloat{
\includegraphics[width=0.9\textwidth]{winner.pdf}}
\newline
\setcounter{subfigure}{2}
\subfloat[$s=1.2$]{\label{s12}
\includegraphics[width=0.48\textwidth]{s1p2.pdf}}
\subfloat[$s=1.3$]{\label{s13}
\includegraphics[width=0.48\textwidth]{s1p3.pdf}}
\caption{Winning droplets for (a) $s=1.0$, (b) $s=1.1$, (c) $s=1.2$, (d) $s=1.3$. Winning droplets with larger index $j$, $1\leq j\leq N$, are depicted by darker grays.}
\label{winners1}
\end{figure}
Having introduced the relevant parameters $N$, $\overline{V}$ and $s$, we now report some numerical experiments to document their impact on the flow dynamics.
First we consider solutions of the system \cref{sys} with $1<\overline{V}\leq 2.5$ for a uniform linear network with $3\leq N\leq 25$ for various values of $s\geq 1$. Specifically, we take $c_{i,i+1}=1=c_{i+1,i}$, $1\leq i\leq N-1$, and $c_{i,j}=0$ in all other cases.
As in \cite{LeEA1} we choose as initial data a small perturbation of the equilibrium $\left(\overline{V},\ldots,\overline{V}\right)$ consisting of droplets of equal size:
\begin{equation}
v_j = \overline{V}+ 10^{-3}\,{{2\,j-(N+1)}\over {N-1}}, \quad 1\leq j\leq N.
\end{equation}
The long-term behavior of solutions is computed by using standard numerical integrators. We start with a value of $\overline{V}=1.01$ and increase it successively to $\overline{V}=2.5$ in increments of $0.01$. Similarly to the Newtonian situation in \cite{LeEA1}, volume scavenging for $s>1$ leads to a depletion of volumes for all but one droplet that emerges as ``winner.'' In fact, $N-1$ small droplets remain with the winning droplet being large. As we will see later (and was shown in \cite{LeEA1} for the case $s=1$), this configuration corresponds to a stable equilibrium of the system. The winners of volume scavenging are displayed in \cref{winners1} for $s=1.0$, $s=1.1$, $s=1.2$ and $s=1.3$. Winning droplets with larger index $j$, $1\leq j\leq N$, are depicted by darker grays. \Cref{s10} reproduces the Newtonian ``change-of-winner" result of \cite{LeEA1} with more details.
\begin{snugshade}
\noindent {\em Physical mechanism}: The physical mechanism behind volume scavenging, independent of the particular fluid model, was made plausible in \cite{SlSt} in the case $N=2$: When two droplets of equal volumes are large, adding a small volume to one droplet removes the same amount from the other one (mass conservation). While the pressure in the first droplet decreases, it increases in the second droplet, pushing even more volume toward the first droplet. Hence this configuration is unstable.
Likewise, when both droplets are small of equal volumes less than 1, the configuration is stable. A similar behavior is also observed in soap films \cite{Bo} and is reminiscent of the ``two-balloon experiment" \cite{DrEA,WeBa}.
\end{snugshade}
\noindent When $N>2$, the situation is more complicated as we will demonstrate later in this work. Animations of volume scavenging can be found online at
\url{https://www.dropbox.com/s/0t4ytk5jkx5uxd0/VAllLog.mpg?dl=0} with a legend given at
\url{https://www.dropbox.com/s/8teepdlybd2j3op/Legend.pdf?dl=0}.
\begin{snugshade}
\noindent {\em Computational aside}: In our computation we have noticed a change in the dynamics with respect to the power-law parameter $s$: The transient times to equilibrium increase by several orders of magnitude as $s$ increases. This phenomenon occurs as the apparent viscosity (friction) of the fluid decreases with increasing shear rates.
Hence the longer transients are surprising, at least at first sight, since, na\"{i}vely, we anticipate a shear-thinning fluid ($s>1$) to become ``thinner'' than a Newtonian one ($s=1$). However, on second thought, one realizes that the apparent viscosity of shear-thinning fluids actually increases when shear rates become smaller. In volume scavenging an increase in the apparent viscosity occurs when pressure drops are small. Hence in low-pressure drop regions a shear-thinning fluid is expected to flow slower than a Newtonian one. This slow-down manifests itself in the flow dynamics, especially near equilibria.
To rationalize the observed transient behavior further, we refer to the similarity with the ``toy problem'' $x^\prime = -x^s$, $x(0)=x_0>0$. The solution $x(t)$ decays to the equilibrium $0$ for every $s\geq 1$. The decay is exponential when $s=1$. For $s>1$, however, the decay is only algebraic like $t^{-1/(s-1)}$.
\end{snugshade}
\begin{figure}[tbhp]
\centering
\subfloat[$s=2.0$]{ \label{graph20}
\includegraphics[width=0.48\textwidth]{s2p0.pdf}}
\subfloat[$s=3.0$]{ \label{graph30}
\includegraphics[width=0.48\textwidth]{s3p0.pdf}}\vspace*{-0.4cm}
\newline
\subfloat{\includegraphics[width=0.9\textwidth]{winner.pdf}}
\caption{Winning droplets for (a) $s=2.0$, (b) $s=3.0$}
\label{winners2}
\end{figure}
\Cref{winners1} shows that even small variations in $s$ lead to notable changes in the winning droplet configuration. This fact is particularly visible when $N\geq 15$. It appears that the basins of attraction of the corresponding stable equilibria not only vary drastically with $N$ and $\overline{V}$ (as indicated in \cite{LeEA1}), but also with $s$. These changes, it seems, become more pronounced for larger values of $s$ as seen in \cref{winners2}. The exact mechanisms underlying the selection of winning droplets remain largely unknown, although some explanations to this end are given in \cite{LeEA1} for networks with $N=3$ and $s=1$.
\begin{snugshade}
\noindent {\em Engineering aside}: As suggested by our model, it might be possible to identify (or narrow down) an unknown power-law parameter $s$ from a small number of
change-of-winner data. In this way such information could prove useful to replace shear stress versus shear rate measurements in conventional rheometers.
\end{snugshade}
\begin{figure}[tbhp]
\centering
\subfloat[$n=0$]{\label{n0}
\includegraphics[width=0.45\textwidth]{zerosn0.pdf}}
\subfloat[$n=1$]{\label{n1}
\includegraphics[width=0.45\textwidth]{zerosn1.pdf}}
\newline
\subfloat[$n=2$]{\label{n2}
\includegraphics[width=0.45\textwidth]{zerosn2.pdf}}
\subfloat[$n=3$]{\label{n3}
\includegraphics[width=0.45\textwidth]{zerosn3.pdf}}
\caption{Positive solutions $h$ of Equation \cref{n3eq} in dependence on $\overline{V}$}
\label{zerosh}\vspace*{-0.4cm}
\end{figure}
Our second concern is to highlight possible issues with the occurrence of equilibria in dependence on $\overline{V}>0$. We concentrate on the case $N=3$. In light of \cref{stat}, any equilibrium of the system \cref{sys} must consist of $n$ ($0\leq n\leq 3$) large droplets of height $h>1$ and $3-n$ small droplets of height $h^{-1}$ such that the capillary pressure in all droplets is equal. By mass conservation, for a given average droplet volume $\overline{V}>0$ we have
\begin{equation}
n\,v(h) + (3-n)\,v\left(h^{-1}\right) = 3\,\overline{V},
\end{equation}
or equivalently,
\begin{equation}\label{n3eq}
n\,h^6+3\,n\,h^4-12\,\overline{V}\,h^3+3\,(3-n)\,h^2+3-n=0.
\end{equation}
\Cref{zerosh} shows the positive solutions $h$ of this equation for varying values of $\overline{V}$ with $n\in\{0, 1, 2, 3\}$. We focus on solutions near $h=1$, $\overline{V}=1$ (intersection point of dashed lines). For given $n$, the solutions of interest are those with $h>1$. As expected, in \cref{n0,n3} we encounter equilibria consisting of droplets of the same size. They are the only equilibria which arise for $0<\overline{V}<1$ with $n=0$ (three small droplets of height $h^{-1}$ and zero large droplets) and for $\overline{V}>1$ with $n=3$ (three large droplets of height $h>1$ and zero small droplets). As seen in \cref{n2}, equilibria with exactly two large droplets occur for $\overline{V}>1$, but not for $0<\overline{V}<1$. The most interesting case is $n=1$. Here, \cref{n1} confirms that for $\overline{V}>1$, we have equilibria consisting of exactly one large droplet (and two small ones). If, however, $\overline{V}$ falls in the interval $(0.957,1)$, two types of equilibria consisting of exactly one large and two small droplets exist. For smaller values of $\overline{V}$, there are no equilibria with $n=1$. We illustrate all possible equilibria (up to permutation of the droplets) with one large and two small droplets for $\overline{V}=0.98$ by volume percentages of the total droplet volume $3\,\overline{V} = 2.94$ in \cref{pies}.
\begin{figure}[tbhp]\vspace*{-0.7cm}
\centering
\subfloat[Volumes $V_L$, $V_S$, $V_S$]{\label{LSS}
\includegraphics[width=0.32\textwidth]{LSSlarge.pdf}}
\subfloat[Volumes $V_l$, $V_s$, $V_s$]{\label{lss}
\includegraphics[width=0.32\textwidth]{lss.pdf}}
\subfloat[Volumes $\overline{V}$, $\overline{V}$, $\overline{V}$]{\label{sss}
\includegraphics[width=0.32\textwidth]{sss.pdf}}
\caption{Volume percentages (rounded) for equilibria with $N=3$, $\overline{V}=0.98$}
\label{pies}
\end{figure}
Here, the two possible heights of the large spherical-cap droplet are $h_L = 1.348$ and $h_l = 1.047$ with the height of the two small droplets given by the reciprocals $h_S=h_L^{-1}$ and $h_s=h_l^{-1}$, respectively. The corresponding volumes are denoted by $V_L=v(h_L)$, $V_S=v(h_S)$, $V_l=v(h_l)$ and $V_s=v(h_s)$. The only other possible equilibrium consists of three equal, small droplets of volume $\overline{V}$.
We will confirm in our work that for $\overline{V}>1$ only one type of stable equilibrium is present, as seen in the Newtonian case $s=1$ \cite{LeEA1} and in the inviscid, symmetric case \cite{SlSt}. If $\overline{V}$ is restricted to an interval of the form $l_0<\overline{V}<1$, two types of stable equilibria will be shown to arise side-by-side, resembling the inviscid case in \cite{SlSt}. The limiting value $l_0>0$ will turn out to depend on $N$. In the situation discussed above ($N=3$) $l_0\approx 0.957$, and the stable equilibria will be the ones described by \cref{LSS,sss}. If, however, $0<\overline{V}<l_0$, the only possible equilibrium will be the one consisting of droplets of equal size. This equilibrium is stable.
\begin{snugshade}
\noindent {\em Socio-economic view}: The parameter ranges $\overline{V}>1$ (``abundant resource") and $0<\overline{V}<l_0$ (``scarcest resource") correspond to the winner-take-all and egalitarian outcomes, respectively. The parameter range $l_0<\overline{V}<1$ (``scarce resource") exhibits both possible outcomes in dependence on initial conditions.
\end{snugshade}
We will study our physics-based model based on analytical techniques that are independent of the underlying network and of the power-law parameter $s$.
A key tool for our study of the system \cref{sys} will be provided by the following energy functional:
\begin{equation}\label{defW}
{\mathcal W}({\mathbf V}) = {1\over N}\,\sum_{j=1}^N \int_0^{V_j} P(V)\,dV\quad \text{for ${\mathbf V} = \left(V_1,\ldots,V_N\right)\in {\mathbb R}^N$.}
\end{equation}
$\mathcal W$ is a minor modification of the equivalent functional introduced in \cite{LeEA1}. It is a measure of the pressure-volume work due to the capillary pressure acting on each spher\-ical-cap droplet.
\begin{snugshade}
\noindent {\em Socio-economic view}: The objective functional $\mathcal W$ is the total cost of resource accumulation on the level of the population, which reflects the aggregate drive to accumulate.
\end{snugshade}
Since for any $V_0\in \mathbb R$ and $h_0 = v^{-1}\left(V_0\right)$
\begin{equation}\label{intP}
\int_0^{V_0} P(V)\,dV = \int_0^{h_0} p(h)\,v^\prime(h)\,dh = 3\,\int_0^{h_0} h\,dh={3\over 2}\,h_0^2,
\end{equation}
by Equation \cref{Adefn}, $\mathcal W$ is related to the total surface area $A({\mathbf V})=\sum_{j=1}^N S(V_j)$ through
\begin{equation}
{\mathcal W}({\mathbf V}) = {1\over N}\,A({\mathbf V})-{3\over 2}.
\end{equation}
Let us record:
\begin{proposition}\label{propA}
${\mathcal W}({\mathbf V})\geq 0$ for all\, ${\mathbf V} \in {\mathbb R}^N$, and ${\mathcal W}({\mathbf V})\rightarrow \infty$ if and only if\, $\|{\mathbf V}\|\rightarrow \infty$.
\end{proposition}
The following result states the key property of $\mathcal W$ for every $s>0$:
\begin{lemma}\label{Ader}
Suppose\, ${\mathbf V} = {\mathbf V}(t) = \left(V_1(t),\ldots,V_N(t)\right)$ is any solution of the system \cref{sys} on an open interval $I\subset \mathbb R$. Then
\begin{equation}
{d\over {dt}}\, {\mathcal W}({\mathbf V}(t))=-{1\over {2\,N}}\,\sum_{i,j=1}^N c_{i,j}\,\left|P(V_i(t))-P(V_j(t))\right|^{s+1}\quad \text{for all $t\in I$.}
\end{equation}
In particular,
$
{d\over {dt}}\, {\mathcal W}({\mathbf V}(t))\leq 0$ for all $t\in I$,
and
$
{d\over {dt}}\, {\mathcal W}({\mathbf V}(t_0))= 0$ for some $t_0\in I$
if and only if \, ${\mathbf V}(t_0)$ is an equilibrium of the system \cref{sys}.
\end{lemma}
\begin{snugshade}
\noindent {\em Socio-economic view}: Lowering the total cost $\mathcal W$ drives the dynamics of the competition.
\end{snugshade}
\begin{proof}
By direct calculation we obtain
\begin{align}
{d\over {dt}}\, {\mathcal W}({\mathbf V})&= {1\over {2\,N}}\,\sum_{j=1}^N P(V_j)\,V_j^\prime+{1\over {2\,N}}\,\sum_{i=1}^N P(V_i)\,V_i^\prime\\
&= {1\over {2\,N}}\,\sum_{i,j=1}^N c_{i,j}\, \left(P(V_j)\,\Delta_s P_{i,j} + P(V_i)\,\Delta_s P_{j,i}\right).
\end{align}
This implies the claim.
\end{proof}
Hence solutions stay bounded on ``any'' time interval of existence starting at $t=0$. Consequently,
we obtain a global existence and uniqueness result.
\begin{proposition}\label{gloex}
Solutions of the initial value problem \cref{sys,ini} exist for all time $t\geq 0$. They are unique if $s\geq 1$.
\end{proposition}
\section{Preserved During Competition: Forward Invariant Sets}
\label{sec_equ}
We have seen in \cref{masscon} that, by conservation of mass, the set ${\mathbb V}\left(\overline{V}\right)$ is forward invariant. In this section we study two more, physically relevant sets.
\begin{definition}
For given $\overline{V}$ and $k\in\mathbb N$ with $1\leq k\leq N$, define the sets
$
{\mathbb V}_+\left(\overline{V}\right) ={\mathbb V}\left(\overline{V}\right)\cap [0,\infty)^N$ and $
{{\mathbb V}}^k_+\left(\overline{V}\right) = \left\{{\mathbf v} = (v_1,\ldots,v_N)\in {\mathbb V}_+(\overline{V})\,\Biggr|\, v_k\leq 1\right\}$.
\end{definition}
Because of \cref{stat,masscon} we readily have:
\begin{proposition}\label{VV+}
Let ${\mathbf V}^* = \left(V^*_1,\dots, V^*_N\right)$ be an equilibrium of the system \cref{sys} with $\overline{V} = {1\over N}\,\sum_{j=1}^N V^*_j >0$. Then ${\mathbf V}^*$ is contained in ${\mathbb V}_+\left(\overline{V}\right)$. In fact, $V^*_j>0$ for all $1\leq j\leq N$.
\end{proposition}
Next we will show that the set ${\mathbb V}_+(\overline{V})$ is forward invariant when $s\geq 1$. To obtain this result, we choose $\epsilon>0$ and let
\begin{equation}\label{adeps}
c^\epsilon_{i,j} = \left\{\begin{matrix} c_{i,j} & \text{if $c_{i,j}>0$,}\\
\epsilon & \text{if $c_{i,j} = 0$ and $i\not=j$,}\\
0&\text{otherwise.} \end{matrix}\right.
\end{equation}
\begin{theorem}\label{invariant}
Suppose that $s\geq 1$, or that $s>0$ and the network is complete. Then
any solution ${\mathbf V}={\mathbf V}(t)$ of the system \cref{sys} with ${\mathbf V}(0)\in {\mathbb V}_+(\overline{V})$ satisfies
$
{\mathbf V}(t)\in {\mathbb V}_+(\overline{V})$ for all $t\geq 0$.
\end{theorem}
\begin{proof}
Let ${\mathbf V}={\mathbf V}(t)=\left(V_1(t),\ldots,V_N(t)\right)$ be a solution originating in ${\mathbb V}_+(\overline{V})$, and let us assume for the moment that the network is complete. It suffices to show that $V_j(t)\geq 0$ for all $t\geq 0$ and $1\leq j\leq N$. The map $V\mapsto P(V)$ is positive for $V>0$, increasing on $(0,1)$ and decreasing on $(1,\infty)$. It satisfies $P(0) = \lim_{V\rightarrow \infty} P(V) = 0$.
Suppose now that $t_0\geq 0$ is a fixed time and $j\in \{1,\ldots, N\}$ an index such that
$
0= V_j(t_0)\leq V_i(t_0)$, $1\leq i\leq N$.
Since $\overline{V}>0$,
there is $k\in \{1,\ldots, N\}$ such that
$V_k(t_0)>0$. Consequently,
$0 = P(V_j(t_0))< P(V_k(t_0))$ and $0=P(V_j(t_0))\leq P(V_i(t_0))$, $1\leq i\leq N$.
Hence we have from \cref{defDs} at $t_0$ that
$ \Delta_s P_{kj}>0$ and $\Delta_s P_{i,j}\geq 0$ for $1\leq i\leq N$.
Completeness of the network gives then
${d\over {dt}}\,V_j(t_0)> 0$.
However, this result implies that
$V_j(t)\geq 0$ for $t\geq 0$, $1\leq j\leq N$.
Together with \cref{masscon}, this proves the claim for the case of a complete network with any $s>0$.
Let us now turn to the case $s\geq 1$. For ${\mathbf v}\in {\mathbb V}_+(\overline{V})$ and $\epsilon>0$, denote the solution of the system \cref{sys} with initial data ${\mathbf V}^\epsilon(0)= {\mathbf v}$ and with adjacency matrix $\left(c^\epsilon_{i,j}\right)$, given in
\cref{adeps}, by ${\mathbf V}^\epsilon={\mathbf V}^\epsilon (t) = \left(V^\epsilon_1(t),\ldots, V^\epsilon_N(t)\right)$. The corresponding network is complete. Hence our argument above applies and proves that
$V^\epsilon_j(t)\geq 0$ for $t\geq 0$, $\leq j\leq N$.
As we pass to the limit $\epsilon\rightarrow 0^+$, continuous dependence of solutions on the parameter $\epsilon$ shows that
the pointwise limit ${\mathbf V}(t) = \lim_{\epsilon\rightarrow 0^+} {\mathbf V}^\epsilon(t)$ exists for every $t\geq 0$ and solves the initial value problem \cref{sys,ini} with initial data ${\mathbf V}(0) = \mathbf v$ and with the original adjacency matrix $\left(c_{i,j}\right)$. Hence
$V_j(t)\geq 0$ for $t\geq 0$, $1\leq j\leq N$.
\end{proof}
Finally let us turn to a physically important observation:
\begin{theorem}\label{invariant2}
Suppose that $s\geq 1$ and $k\in\{1,\ldots, N\}$. Then any solution ${\mathbf V}={\mathbf V}(t)$ of the system \cref{sys} with ${\mathbf V}(0)\in {\mathbb V}^k_+(\overline{V})$ satisfies
$
{\mathbf V}(t)\in {\mathbb V}^k_+(\overline{V})$ for all $t\geq 0$.
\end{theorem}
Hence for $s\geq 1$, a small droplet will remain small.
\begin{snugshade}
\noindent {\em Socio-economic view}: \Cref{invariant2} describes the ``once down-and-out, always down-and-out" outcome: When a competitor's resource falls below $1$, that competitor cannot recover. Hence upward mobility is impossible for individuals whose resource shrinks below this threshold.
\end{snugshade}
\begin{proof}
Suppose first that the network is complete. Given $k$, we may assume that there exists $t_0\geq 0$ and $j\in\{1,\ldots,N\}$ such that
$V_k(t_0)=1$ and $V_j(t_0)\not=1$.
For otherwise, we would either have $V_k(t)<1$ for all $t\geq 0$, or $V_i(t) =1$, $1\leq i\leq N$, whenever $V_k(t) = 1$. The latter situation implies that $V_i(t)=1$, $1\leq i\leq N$, for all $t\geq 0$ by uniqueness of solutions. In either case, the claim would follow.
Since for every $V>0$ with $V\not=1$, we have $P(V)<P(1)=2$, we obtain that at time $t_0$,
$ \Delta_s P_{j k}<0 $ and $\Delta_s P_{i k}\leq 0$ for all $1\leq i\leq N$.
Consequently, by completeness of the network,
${d\over {dt}} V_k(t_0)< 0$.
Hence as soon as $V_k$ reaches the value $1$, it must decrease again. In conclusion,
$V_k(t)\leq 1$ for all $t\geq 0$ and $k\in\{1,\dots,N\}$.
If the network is not complete, let us proceed as before: For ${\mathbf v}\in {\mathbb V}^k_+(\overline{V})$ and $\epsilon>0$, let ${\mathbf V}^\epsilon={\mathbf V}^\epsilon (t) = \left(V^\epsilon_1(t),\ldots, V^\epsilon_N(t)\right)$ be the solution with initial data ${\mathbf V}^\epsilon(0)= {\mathbf v}$ and with adjacency matrix $\left(c^\epsilon_{i,j}\right)$, introduced in
\cref{adeps}. Since the corresponding network is complete, our result above applies.
As we pass to the limit $\epsilon\rightarrow 0^+$, we obtain the sought-for estimate as before.
\end{proof}
The condition $s\geq 1$ was used in this proof in two instances: First, we appealed to the uniqueness of the solution of \cref{sys,ini} with initial data $V_i(0)=1$, $1\leq i\leq N$, in the case of a complete network. Second, we exploited the uniqueness of solutions for continuous dependence on parameters to obtain the claim for non-complete networks. When we have a complete network, it is thus sufficient to replace the condition $s\geq 1$ by the requirement that the initial value problem \cref{sys,ini} with initial data $V_i(0)=1$, $1\leq i\leq N$, only permit the solution $V_i(t)=1$, $1\leq i\le N$, $t\geq 0$. This is indeed satisfied for $0<s<1$ in the case $N=2$, as can be seen by direct calculation.
\section{Rest States and End States: Equilibria and Semiflows}
Let us now return to the study of equilibria of the system \cref{sys} in ${\mathbb V}_+(\overline{V})$ for fixed $\overline{V}>0$. In light of \cref{stat}, we consider the question whether an equilibrium can consist of $n$ large droplets (i.e. droplets of height $h>1$) and $N-n$ small droplets (i.e. droplets of height $h^{-1}\leq 1$). Noting the volume-height relation \cref{randv}, we can thus cast mass conservation, given by \cref{masscon}, in the form
${n\over N}\,v(h)+\left(1-{n\over N}\right)\,v\left(h^{-1}\right) = {n\over N}\,\left({{h^3+3\,h}\over 4}\right) + \left(1-{n\over N}\right)\,\left({{h^{-3}+3\,h^{-1}}\over 4}\right) =\overline{V}$, or
\begin{equation}\label{volconstr}
\alpha\,v(h)+(1-\alpha)\,v\left(h^{-1}\right)=\overline{V}\quad \text{with $\alpha={n\over N}$.}
\end{equation}
This equation has a simple symmetry:
\begin{equation}\label{symeq}
\text{$h>0$ solves \cref{volconstr} with $\alpha=\beta$ if and only if $h^{-1}$ solves \cref{volconstr} with $\alpha=1-\beta$.}
\end{equation}
It is convenient to rewrite Equation \cref{volconstr} as
$
{\mathcal P}_\alpha(h)=0
$
where the \emph{mass polynomial} ${\mathcal P}_\alpha = {\mathcal P}_\alpha(h;\overline{V})$ is given by
\begin{equation}\label{defpa}
{\mathcal P}_\alpha(h) \equiv \alpha\,h^6+3\,\alpha\,h^4-4\,\overline{V}\,h^3+3\,(1-\alpha)\,h^2+1-\alpha.
\end{equation}
\begin{snugshade}
\noindent {\em Applied mathematics aside}: We will allow the parameter $\alpha$ to take on all values in the interval $[0,1]$ instead of discrete values only. This parametrization makes it possible to study the zeros of ${\mathcal P}_\alpha$ in continuous dependence on $\alpha$ and without regard to particular values of $n$ and $N$.
\end{snugshade}
Let us now address the {\em uniform} equilibrium.
\begin{proposition}\label{unieq}
For every $\overline{V}>0$, there is an equilibrium consisting of $N$ droplets of equal height (and volume).
\end{proposition}
The uniform equilibrium arises for $\overline{V}>1$ with $n=N$ and for $0<\overline{V}\leq 1$ with $n=0$.
Before we continue with our analysis, let us briefly adapt an approach taken by Slater and Steen \cite{SlSt} in their study of $N$ inviscid spherical droplets with $S_N$ symmetry. To this end, we note that for constant $\overline{V}$, equilibria can be
identified with each other if they share the same number of large droplets of the same height (or, equivalently, the same number of small droplets of the same height). Because of the symmetry expressed in \cref{symeq}
we may assume that $ \alpha={n\over N}\leq {{\left\lfloor{{N/ 2}}\right\rfloor}\over N}$. Using ideas proposed in \cite{ThEA}, we now introduce the quantity
\begin{equation}\label{deftheta}
\theta = \theta(h) \equiv v(h)-v\left({1\over h}\right),
\end{equation}
defined for all positive $h$ such that ${\mathcal P}_\alpha(h)=0$ for some $\overline{V}>0$ and some $\alpha={n\over N}$, $1\leq n\leq {{\left\lfloor{{N/ 2}}\right\rfloor}}$. The function $\theta$ is clearly invertible on its domain.
The case $\theta>0$ arises exactly when there exists $h>1$ such that ${\mathcal P}_\alpha(h)=0$. Hence in this situation we have an equilibrium with exactly $n=\alpha\,N$ large droplets of height $h>1$. The negative case $\theta<0$ corresponds to exactly $n=\alpha\,N$ small droplets of height $0<h<1$ as is apparent from the symmetry \cref{symeq}. $\theta=0$ occurs exactly when $h=1$. Hence the corresponding equilibrium is uniform and $\overline{V}=1$. It will become evident from \cref{volconstr,symeq} and the later developments that, in this way, $\theta$ is defined for all $h>0$. To include all uniform equilibria, we extend the definition of $\theta$ by setting $\theta=0$ if there exists $h>0$ such that ${\mathcal P}_0(h)=0$ for some $\overline{V}>0$. Since ${\mathcal P}_0(h) = -4\,\overline{V}\,h^3+3\,h^2+1$, we have exactly one positive zero $h$ for every choice of $\overline{V}$. The zero $h$ is in $(0,1)$ if $\overline{V}>1$, it is $1$ if $\overline{V}=1$, and it lies in $(1,\infty)$ if $0<\overline{V}<1$. Therefore, as we appeal again to \cref{symeq}, we conclude that $\theta=0$ corresponds either to the uniform equilibrium with $N$ large droplets of height $h^{-1}$ if $\overline{V}>1$, or to the uniform equilibrium with $N$ small droplets of height $h^{-1}$ if $0<\overline{V}\leq 1$. Hence after this extended definition of $\theta$, $\theta=0$ represents an equilibrium in the $(\overline{V}, \theta)$-plane for every $\overline{V}>0$. Having such a constant equilibrium for all parameter values is commonly assumed in bifurcation theory \cite{GuHo}.
Now let $a\equiv h+h^{-1}$ and $b \equiv h-h^{-1}$. Then $a^2-b^2=4$. Since \cref{deftheta} becomes
\begin{equation}
\theta = {1\over 4}\,b\,(b^2+6),
\end{equation}
we have
\begin{equation}\label{defab}
b={\mathcal B}(\theta) \equiv { {
-2^{2/3}+2^{1/3}\,\left(\theta+\sqrt{2+\theta^2}\right)^{2/3} }\over {\left(\theta+\sqrt{2+\theta^2}\right)^{1/3}} }\quad\text{and}\quad
a = {\mathcal A}(\theta) \equiv \left(4+ {\mathcal B}(\theta)^2\right)^{1/2}.
\end{equation}
Equation \cref{volconstr} can now be written in the form
$\overline{V} = {1\over 8}\,a^3+{{2\,\alpha-1}\over 8}\,b\,(b^2+6)$.
Hence we recover a single equation relating $\theta$ and $\overline{V}$:
\begin{equation}\label{Aeq}
\overline{V} = {1\over 8}\,{\mathcal A}(\theta)^3+ \left(\alpha-{1\over 2}\right)\,\theta.
\end{equation}
Equation \cref{Aeq} is - up to a rescaling - the same as in \cite{SlSt}. This might seem curious since the work \cite{SlSt} is concerned with inviscid spherical-cap droplets subject to $S_N$ symmetry. On second thought, equilibria in \cite{SlSt} are determined by the same volume constraint together with the Young-Laplace relation as here. \Cref{ecurves} displays the generic situation: curves of equilibria with $n$ large droplets determined by Equation \cref{Aeq} plus the uniform equilibrium curve $\theta=0$ (labeled $n=0$ for $\overline{V}\leq 1$ and $n=N$ for $\overline{V}>1$).
\begin{figure}[tbhp]
\centering
\includegraphics[width=0.7\textwidth]{equilibcurves2.pdf}
\caption{Equilibrium curves in the $(\overline{V}, \theta)$-plane for equilibria with $n$ large droplets (odd $N$) }
\label{ecurves}
\end{figure}
Equation \cref{Aeq} allows an interesting first insight in the location of equilibria in dependence on $\overline{V}$. To obtain more detailed information about the location and nature of equilibria, we proceed by studying the zeros of the mass polynomial ${\mathcal P}_{\alpha}$ directly. As we will see, this approach will prove useful: Firstly, we gather details about the size of equilibria which have an immediate bearing on their stability. Secondly, we can easily identify the turning points of the equilibrium curves in \cref{ecurves}. Thirdly, our results lay the foundation of our study of hierarchies among the equilibria.
From here onwards it will be convenient to allow any real $\alpha$ with $0\leq \alpha\leq 1$ in the definition of ${\mathcal P}_\alpha$, keeping in mind that $\alpha={n\over N}$ describes the situation of $n$ large and $N-n$ small droplets.
Moreover, to find non-uniform equilibria (i.e.\,equilibria consisting of both large and small droplets), it suffices to assume that $1\leq n\leq N-1$ and to study the zeros of ${\mathcal P}_\alpha$ for $0<\alpha<1$.
It is tacitly understood in all later developments that zeros are counted according to their multiplicity. The key result on zeros of ${\mathcal P}_\alpha$ is the following:
\begin{theorem}\label{poszero}
Let $\overline{V}>0$. Then for every $0<\alpha<1$, the mass polynomial ${\mathcal P}_\alpha$, has either no positive zeros or exactly two. It has exactly two positive zeros $h_1=h_1(\alpha)$ and $h_2=h_2(\alpha)$ with $h_1\leq h_2$ if and only if
\begin{equation}\label{KVcon}
\overline{V}\geq \alpha^{3/4}\,\left(1-\alpha\right)^{1/4}+\alpha^{1/4}\,\left(1-\alpha\right)^{3/4}.
\end{equation}
In this case the zeros $h_1$ and $h_2$ are such that either
\begin{align}
& h_1=\left({1\over \alpha}-1\right)^{1/4}=h_2\quad \text{if\quad $\overline{V}=\alpha^{3/4}\,\left(1-\alpha\right)^{1/4}+\alpha^{1/4}\,\left(1-\alpha\right)^{3/4}$, or}\\
& h_1<\left({1\over \alpha}-1\right)^{1/4}<h_2\quad \text{if\quad $\overline{V}>\alpha^{3/4}\,\left(1-\alpha\right)^{1/4}+\alpha^{1/4}\,\left(1-\alpha\right)^{3/4}$.}
\end{align}
\end{theorem}
\begin{proof}
Consider the polynomial $q_\alpha$, defined by
\begin{multline}
q_\alpha(h) \equiv \alpha\,\biggl(h^4+2\,\left({1\over \alpha}-1\right)^{1/4}\,h^3+3\,\left(1+\left({1\over \alpha}-1\right)^{1/2}\right)\,h^2+\\
2\,\left({1\over \alpha}-1\right)^{1/4}\,h+\left({1\over \alpha}-1\right)^{1/2}\biggr).
\end{multline}
Evidently, $q_\alpha(h)>0$ for $h>0$.
Then for $h>0$,
\begin{equation}\label{gseq}
{\mathcal P}_\alpha(h)\gtreqqless
q_\alpha(h)\,\left(h-\left({1\over \alpha}-1\right)^{1/4}\right)^2\ \text{if $\overline{V}\lesseqqgtr \alpha^{3/4}\,\left(1-\alpha\right)^{1/4}+\alpha^{1/4}\,\left(1-\alpha\right)^{3/4}$,}\\
\end{equation}
respectively. The claim follows now immediately.
\end{proof}
Condition \cref{KVcon} in \cref{poszero} is both necessary and sufficient for positive zeros of the polynomial ${\mathcal P}_\alpha$. This surprising result hinges on the factorizability of ${\mathcal P}_\alpha$, a polynomial of degree 6, when $\overline{V}= \alpha^{3/4}\,\left(1-\alpha\right)^{1/4}+\alpha^{1/4}\,\left(1-\alpha\right)^{3/4}$. The symmetry \cref{symeq} hints at this result. For later use let us introduce
the \emph{limiting function}
\begin{equation}
\alpha\longmapsto L(\alpha)\equiv \alpha^{3/4}\,\left(1-\alpha\right)^{1/4}+\alpha^{1/4}\,\left(1-\alpha\right)^{3/4},
\end{equation}
appearing on the right of \cref{KVcon}. $L$ is strictly increasing on $(0,{1\over 2})$ and strictly decreasing on
$({1\over 2}, 1)$. It has the maximum 1 at $\alpha={1\over 2}$. The function is graphed in \cref{limfunc}.
\begin{figure}[tbhp]
\begin{center}
\includegraphics[width=0.35\textwidth]{limfun.pdf}
\end{center}
\vspace*{-0.3cm}
\caption{The limiting function $L$}
\label{limfunc}
\end{figure}
Next we record some additional properties of the mass polynomial ${\mathcal P}_\alpha$.
\begin{proposition}\label{propp} For every $\overline{V}>0$ and $0<\alpha< 1$,
\begin{align}
& \text{${\mathcal P}_\alpha(h)\rightarrow \infty$ as $h\rightarrow \infty$,}\quad
{\mathcal P}_\alpha(0) = 1-\alpha,\quad {\mathcal P}_\alpha(1) = 4\,\left(1-\overline{V}\right),\label{zeroone}\\
&{\mathcal P}_\alpha\left(\left({{1}\over \alpha}-1\right)^{1/4}\right) = {4}\,\left({1\over \alpha}-1\right)^{3/4}\,\left(L(\alpha)-\overline{V}\right).\label{critval}
\end{align}
\end{proposition}
With this information we obtain:
\begin{theorem}\label{pVL1}
Let $\overline{V}>1$. For every\, $0<\alpha<1$, the mass polynomial ${\mathcal P}_\alpha$ has exactly one zero $h_L=h_L(\alpha)$ larger than 1. In particular,
\begin{equation}\label{estLa1}
h_L>\left({1\over \alpha}-1\right)^{1/4}.
\end{equation}
\end{theorem}
Note that condition \cref{KVcon} is vacuously satisfied. The existence and uniqueness of $h_L>1$ is a consequence of \cref{zeroone} when $\overline{V}>1$.
The estimate \cref{estLa1} follows from \cref{critval} since for $\overline{V}>1$, $0<\alpha<1$, we have ${\mathcal P}_\alpha\left(\left({1\over \alpha}-1\right)^{1/4}\right)<0$. Of course, \cref{estLa1} is useful only if $0<\alpha<{1\over 2}$.
\begin{theorem}\label{pVE1}
Let $\overline{V} = 1$.
\begin{enumerate}
\item[\mbox{(a)}] If\, $0<\alpha<{1\over 2}$, the mass polynomial ${\mathcal P}_\alpha$ has exactly one zero $h_L=h_L(\alpha)$ larger than 1. In this case,
\begin{equation}\label{estE1}
h_L>\left({1\over \alpha}-1\right)^{1/4}.
\end{equation}
\item[\mbox{(b)}] If\, ${1\over 2}\leq \alpha<1$, the mass polynomial ${\mathcal P}_\alpha$ has no zero larger than 1.
\end{enumerate}
\end{theorem}
\begin{proof}
Since ${\mathcal P}_\alpha(1) = 0$ and ${\mathcal P}_\alpha^\prime(1) = 6\,(2\,\alpha-1)$, it is clear in light of \cref{zeroone} that ${\mathcal P}_{\alpha}$ has a zero larger than 1 if and only if $0<\alpha< {1\over 2}$. If so, this is the only zero larger than 1 since ${\mathcal P}_\alpha$ cannot have more than two positive zeros. The estimate \cref{estE1} follows again from \cref{critval} since ${\mathcal P}_\alpha\left(\left({1\over \alpha}-1\right)^{1/4}\right)<0$ if $\overline{V}=1$ and $0<\alpha<{1\over 2}$.
\end{proof}
\begin{theorem}\label{pVS1}
Let $0<\overline{V}<1$.
\begin{enumerate}
\item[\mbox{(a)}] Suppose\, $0<\alpha<{1\over 2}$ and $\overline{V}\geq L(\alpha)$.
Then the mass polynomial ${\mathcal P}_\alpha$ has exactly two zeros $h_l=h_l(\alpha)$ and $h_L=h_L(\alpha)$ such that
\begin{equation}\label{h12}
1<h_l\leq \left({1\over \alpha}-1\right)^{1/4}\leq h_L.
\end{equation}
The two zeros $h_l$ and $h_L$ are such that either
\begin{align}
& h_l=\left({1\over \alpha}-1\right)^{1/4}=h_L\quad \text{if\quad $\overline{V}=L(\alpha)$, or}\\
& h_l<\left({1\over \alpha}-1\right)^{1/4}<h_L\quad \text{if\quad $\overline{V}> L(\alpha)$.}
\end{align}
\item[\mbox{(b)}] If ${1\over 2}\leq \alpha<1$ or $\overline{V}< L(\alpha)$, the mass polynomial ${\mathcal P}_\alpha$ has no zero larger than 1.
\end{enumerate}
\end{theorem}
\begin{proof}
By \cref{poszero}, ${\mathcal P}_\alpha$, $0<\alpha<1$, has either no positive zeros or exactly two. The latter occurs if and only if \cref{KVcon} holds true. If so, since ${\mathcal P}_\alpha(0)>0$, ${\mathcal P}_\alpha(1)>0$ and ${\mathcal P}_\alpha\left(\left({1\over \alpha}-1\right)^{1/4}\right)\leq 0$, \cref{propp} implies that either the number $\left({1\over \alpha}-1\right)^{1/4}$ and the positive zeros of ${\mathcal P}_\alpha$ lie in the interval $(0,1)$, or they all lie in the interval $(1,\infty)$. Now $\left({1\over \alpha}-1\right)^{1/4}$ lies
in $(1,\infty)$ if and only if $0<\alpha<{1\over 2}$. The remaining claims follow from \cref{poszero}.
\end{proof}
The shaded area in \cref{limfunc} displays permissible pairs $(\alpha, \overline{V})$ to which part (a) of \cref{pVS1} applies.
All possible non-uniform equilibria are now completely described by \cref{pVL1,pVE1,pVS1}
when we note that each such equilibrium is -- up to permutation -- given by the number of its large droplets $n$ of height $h>1$ and the number of its small droplets $N-n$ of height $h^{-1}$.
We note, in particular, that, for $0<\overline{V}\leq 1$, the condition $0<\alpha<{1\over 2}$ translates to $N\geq 3$. Hence non-uniform equilibria with $0<\overline{V}\leq 1$ require $N\geq 3$. For $\overline{V}>1$, every equilibrium contains at least one large droplet. If $N\geq 2$ and $0<\overline{V}<L\left({1\over N}\right)$, then only the uniform equilibrium consisting of $N$ droplets of equal size is possible.
Finally, we end this section with some important remarks about the structure of the dynamical system, given by \cref{sys,ini}: In the case $s\geq 1$, \cref{gloex} together with \cref{masscon} shows that the initial value problem \cref{sys,ini} defines a semiflow on ${\mathbb V}(\overline{V})$.
By \cref{stat} and \cref{Ader}, this semiflow is a gradient dynamical system with Lyapunov function $\mathcal W$, see \cite{Ha}.
When $0<s<1$, the situation is more delicate. Here, \cref{stat} and \cref{Ader} imply that the initial value problem \cref{sys,ini} defines a generalized semiflow (in the sense of Ball) on ${\mathbb V}(\overline{V})$ with $\mathcal W$ serving again as a Lyapunov function, see \cite{Ba}.
Our discussion above shows that ${\mathbb V}_+(\overline{V})$ (and hence ${\mathbb V}(\overline{V})$ as well) contains only finitely many equilibria for fixed $\overline{V}$. Hence we can appeal to classical results about gradient dynamical systems in the case $s\geq 1$ and to analogous results for generalized semiflows with Lyapunov function in the case $0<s<1$ (see \cite{Ba,Ha}) to conclude:
\begin{proposition}
The (generalized) semiflow on ${\mathbb V}(\overline{V})$, given by the initial value problem \cref{sys,ini} with ${\mathbf V}(0)\in {{\mathbb V}(\overline{V})}$, has a global attractor which consists of all equilibria. Specifically, every semi-trajectory ${\mathbf V}={\mathbf V}(t)$ of the system \cref{sys} with ${\mathbf V}(0)\in {{\mathbb V}(\overline{V})}$ converges to an equilibrium of \cref{sys} in ${\mathbb V}_+(\overline{V})$ as $t\rightarrow\infty$.
\end{proposition}
\begin{snugshade}
\noindent {\em Socio-economic view}: While we have accounted for all possible rest states (equilibria) in the competition, only some of them will be end states, i.e.\,reachable rest states (stable equilibria). This observation motivates the following section.
\end{snugshade}
\section{Reachable Rest States: Stability of Equilibria}
\label{sec_stab}
The stability of equilibria was discussed for the Newtonian case $s=1$ with $\overline{V}>1$ in \cite{LeEA1}. The arguments there were based on the observation that
the Lyapunov function ${\mathcal W}$ is independent of the particular network of fluid channels. The authors exploited this observation by working with a network of their choice (star network) and
combining local information about equilibria (obtained by linearization) with the Lyapunov Stability Theorem. Instead of pursuing an approach as in \cite{LeEA1}, we will give an argument which is independent of any particular network (except the number $N$), and is also independent of the power-law parameter $s>0$ since the energy functional $\mathcal W$ is as well. Our discussion contains the stability findings in \cite{LeEA1}. At the core of our approach lies the following simple equivalence:
\begin{proposition}\label{Astab}
A point\, ${\mathbf V}^*\in {\mathbb V}(\overline{V})$ is a stable equilibrium of the system \cref{sys} if and only if it is a local minimizer of the Lyapunov function ${\mathcal W}$ on ${\mathbb V}(\overline{V})$.
\end{proposition}
Note that a local minimizer of $\mathcal W$ is actually a strict local minimizer by \cref{Ader}. Also, in light of \cref{VV+} we may replace ${\mathbb V}(\overline{V})$ by ${\mathbb V}_+(\overline{V})$ here.
\begin{snugshade}
\noindent {\em Applied mathematics aside}: \Cref{Astab} characterizes stable equilibria as minimizers of an optimization problem subject to constraints. A single Lagrange multiplier suffices for the volume constraint.
\end{snugshade}
Consider the Lagrangian ${\mathcal L}: {\mathbb R}^N\times {\mathbb R}\rightarrow {\mathbb R}$, given for ${\mathbf V} = (V_1,\ldots,V_N)$ by
\begin{equation}
{\mathcal L}({\mathbf V}, \lambda) \equiv{\mathcal W}({\mathbf V})-\lambda\,F({\mathbf V})\quad
\text{with}\quad
F({\mathbf V}) \equiv {1\over N}\,\sum_{k=1}^N V_k -\overline{V}.
\end{equation}
If ${\mathcal W}$ assumes a local minimum on ${\mathbb V}(\overline{V})$ at ${\mathbf V}^* = (V^*_1,\ldots,V^*_N)$, then there exists $\lambda^*\in \mathbb R$ such that $({\mathbf V}^*,\lambda^*)$ is a local minimizer of ${\mathcal L}$ on ${\mathbb R}^N\times {\mathbb R}$.
For ${\mathcal L}$ to have a local minimum on ${\mathbb R}^N\times {\mathbb R}$ at $({\mathbf V}^*, \lambda^*)$, it is necessary that
\begin{align}
& D_{\mathbf V} {\mathcal L}({\mathbf V}^*,\lambda^*) = {\mathbf 0},\quad F({\mathbf V}^*) = 0,\quad \text{and}\label{condmin1}\\
& {\boldsymbol \phi}\, D_{\mathbf V}^2 \,{\mathcal L}({\mathbf V}^*,\lambda^*)\, {{\boldsymbol \phi}}^T\geq 0\quad \text{for all ${\boldsymbol \phi}\in {\mathbb R}^N$ with $\left(D F({\mathbf V}^*)\right)\,{\boldsymbol \phi}^T = 0$.}\label{condmin2}
\end{align}
Here, $D F$ denotes the total derivative (row vector) of $F$, while $D_{\mathbf V} {\mathcal L}$ and $D_{\mathbf V}^2 \,{\mathcal L} $ denote the total derivative and the Hessian of $\mathcal L$ with regard to $\mathbf V$, respectively. Clearly, in our situation we have $D_{\mathbf V}^2 \,{\mathcal L} = D^2\,{\mathcal W}$, the Hessian of ${\mathcal W}$.
Conversely, if $({\mathbf V}^*, \lambda^*)\in {\mathbb V}(\overline{V})\times {\mathbb R}$ is such that the condition \cref{condmin1} holds and such that
\begin{equation}\label{condmin3}
{\boldsymbol \phi}\, D_{\mathbf V}^2 \,{\mathcal L}({\mathbf V}^*,\lambda^*)\, {{\boldsymbol \phi}}^T> 0\quad \text{for all nonzero $ {\boldsymbol \phi}\in {\mathbb R}^N$ with $\left(D F({\mathbf V}^*)\right)\,{ {\boldsymbol \phi}}^T = 0$,}
\end{equation}
${\mathcal W}$ assumes a strict local minimum on ${\mathbb V}(\overline{V})$ at ${\mathbf V}^*$.
Conditions \cref{condmin1,condmin2,condmin3}, are classical, see e.g.\,\cite{LuYe}. \Cref{condmin1} just states the requirement that ${\mathbf V}^*$ be an equilibrium of \cref{sys} in ${\mathbb V}(\overline{V})$.
If ${\mathbf V}^*=(V^*_1,\ldots,V^*_N) \in {\mathbb V}(\overline{V})$ is an equilibrium of \cref{sys}, ${\mathbf V}^*$ consists of $n$ entries for large droplets of volume $V_L>1$ (and height $h_L>1$) and $N-n$ entries for small droplets of corresponding volume $0<V_S\leq 1$ (and height $h_S$). Note that $V_S=1$ immediately requires $n=0$ and $\overline{V}=1$.
Let us define
\begin{equation}
\delta_S \equiv \left\{\begin{matrix} P^\prime(V_S) & \text{if $0\leq n\leq N-1$,}\\[1ex]
0 &{\text{if $n=N$}}\end{matrix}\right.\quad \text{and}\quad \delta_L \equiv \left\{\begin{matrix} P^\prime(V_L) & \text{if $1\leq n\leq N$,}\\[1ex]
0 & \text{if $n=0$.}\end{matrix}\right.
\end{equation}
Then the necessary condition \cref{condmin2} for a local minimizer can be cast in the form
\begin{equation}\label{conalt}
\delta_L\,\left(\sum_{k=1}^n \phi_k^2\right)+\delta_S\,\left(\sum_{k=n+1}^N \phi_k^2\right)\geq 0\quad \text{for all $(\phi_1,\ldots,\phi_N)$ with $\sum_{j=1}^N \phi_j = 0$.}
\end{equation}
The sufficient condition \cref{condmin3} for a strict local minimizer reduces to
\begin{equation}\label{conalt2}
\delta_L\,\left(\sum_{k=1}^n \phi_k^2\right)+\delta_S\,\left(\sum_{k=n+1}^N \phi_k^2\right)> 0\quad \text{for all nonzero $(\phi_1,\ldots,\phi_N)$ with $\sum_{j=1}^N \phi_j = 0$.}
\end{equation}
Next we have for $V=v(h)$,
\begin{equation}\label{Pder}
P^\prime(V) ={{16\,(1-h^2)}\over {3\,(h^2+1)^3}}.
\end{equation}
Hence $\delta_L<0$ if $1\leq n\leq N$; $\delta_S> 0$ if $1\leq n\leq N-1$, or if $n=0$ and $0<V<1$ ; and
\begin{equation}\label{pspl}
\delta_S=-h_L^4\,\delta_L\quad \text{if $1\leq n\leq N-1$.}
\end{equation}
Consequently, we have immediately from \cref{conalt} that
${\mathbf V}^*$ is unstable if there exist $i, j$, $1\leq i<j\leq N$ such that $V^*_i=V^*_j=V_L>1$. Hence we have shown:
\begin{theorem}\label{stabeq}
An equilibrium of the system \cref{sys} in ${\mathbb V}(\overline{V})$ is unstable if it contains two or more large droplets.
\end{theorem}
For
$\overline{V}>1$, any equilibrium ${\mathbf V}^*$ in ${\mathbb V}(\overline{V})$ must contain at least one entry $V^*_i>1$. Moreover, since $\mathcal W$ assumes a global minimum on ${{\mathbb V}(\overline{V})}$, the minimizer must be an equilibrium of the system \cref{sys} by \cref{condmin1}.
Hence we conclude:
\begin{corollary}\label{V>1stab}
An equilibrium of the system \cref{sys} in ${\mathbb V}(\overline{V})$ with $\overline{V}>1$ is stable if and only if it contains exactly one large droplet.
\end{corollary}
\begin{snugshade}
\noindent {\em Applied mathematics aside}:
\Cref{stabeq} is reminiscent of related results in \cite{We} for two and three spher\-i\-cal-cap droplets and in \cite{SlSt} for $N$ coupled inviscid droplet oscillators with $S_N$ symmetry, while \cref{V>1stab} was proved in \cite{LeEA1} for $s=1$, exploiting the hyperbolicity of equilibria in the Newtonian regime.
\end{snugshade}
Let us now focus on the case $n=1$. Since for $(\phi_1,\dots,\phi_N)$ with $\sum_{j=1}^N \phi_j = 0$
\begin{equation}
\phi_1^2=\left(\sum_{j=2}^N \phi_j\right)^2\leq (N-1)\,\sum_{j=2}^N \phi_j^2,
\end{equation}
we can use \cref{conalt2,pspl} to obtain the sufficient condition for a strict local minimizer
\begin{equation}
\delta_L\,\left(N-1-h_L^4\right)>0.
\end{equation}
\begin{theorem}\label{V<=1stab}
An equilibrium of the system \cref{sys} in ${\mathbb V}(\overline{V})$
is stable if it contains exactly one large droplet of height $h>\left(N-1\right)^{1/4}$.
\end{theorem}
It follows from \cref{pVE1,pVS1} with $\alpha={1\over N}$ that such equilibria arise for $\overline{V}=1$ if $N\geq 3$, and for $0<\overline{V}<1$ if $\overline{V}>L\left({1\over N}\right)$. For $\overline{V}>1$ this result is already contained in \cref{V>1stab} by \cref{pVL1}.
Next, in the case $n=1$, let us choose $\phi_1=1$ and $\phi_j = -(N-1)^{-1}$, $2\leq j\leq N$. Then \cref{conalt} reduces to the following necessary condition for a local minimizer
\begin{equation}
\delta_L\,\left(N-1-h_L^4\right)\geq 0.
\end{equation}
Hence an equilibrium is unstable if it contains a large droplet of height less than $\left(N-1\right)^{1/4}$. \Cref{pVS1} shows that such equilibria with $n=1$ occur for $0<\overline{V}<1$ if $\overline{V}>L\left({1\over N}\right)$. There the height of the large droplet for this unstable equilibrium was denoted by $h_l(\alpha)$ with $\alpha={1\over N}$.
The borderline case $n=1$ and $h_L = \left(N-1\right)^{1/4}$ with $0<V<1$ is left to be discussed. By \cref{pVS1}, this situation arises when $\overline{V}=L\left({1\over N}\right)$ with $N\geq 3$ .
To obtain a stability result in this case, let us argue directly. First
\begin{equation}\label{P2D}
P^{\prime\prime}(V) = {{256}\over 9}\,{{h\,(h^2-2)}\over {(h^2+1)^5}}\quad\text{where $V=v(h)$.}
\end{equation}
Next let ${\boldsymbol \phi}={\boldsymbol \phi}(t)=\left(\phi_1(t),\ldots,\phi_N(t)\right)$ be a smooth curve. Then
\begin{align}
&{{d^2}\over {dt^2}} {\mathcal W}\left({{\boldsymbol \phi}} \right)={1\over N}\,\sum_{k=1}^N \left(P^{\prime}(\phi_k)\,\left(\phi_k^{\prime}\right)^2+P(\phi_k)\,\phi_k^{\prime\prime}\right),\quad \text{and}\label{A2D}\\
&
{{d^3}\over {dt^3}}{\mathcal W}\left({{\boldsymbol \phi}} \right)={1\over N}\, \sum_{k=1}^N \left(P^{\prime\prime}(\phi_k)\,\left(\phi_k^{\prime}\right)^3+3\,P^\prime(\phi_k)\,\phi_k^\prime\,\phi_k^{\prime\prime}+P(\phi_k)\,\phi_k^{\prime\prime\prime}\right).\label{A3D}
\end{align}
Let $V_L=v(h_L)$ and $V_S = v\left(h_L^{-1}\right)$ with $h_L = \left(N-1\right)^{1/4}$. For fixed $\phi_0\in \mathbb R$, we set
\begin{equation}
\phi_k(t) \equiv \left\{\begin{matrix} V_L+(N-1)\,\phi_0\,t & \text{if $k=1$,}\\
V_S-\phi_0\,t & \text{otherwise.}\end{matrix}\right.
\end{equation}
Then ${{\boldsymbol \phi}}(t) = \left(\phi_1(t),\ldots,\phi_N(t)\right)$ is a smooth curve with trace in ${\mathbb V}({\overline{V}})$. Moreover,
\begin{align}
&{{d^2}\over {dt^2}} {\mathcal W}\left({\mathbf v}(t) \right)\Big|_{t=0} = 0,\quad \text{and}\\
&{{d^3}\over {dt^3}}{\mathcal W}\left({\mathbf v}(t) \right)\Big|_{t=0} =\left(P^{\prime\prime}(V_L)\,(N-1)^2-P^{\prime\prime}(V_S)\right)\,{{N-1}\over N}\,\phi_0^3\\
&\hspace*{2.35cm}={{256}\over 9}\,{{\left((N-1)^{1/2}-1\right)\,(N-1)^{7/4}}\over {\left((N-1)^{1/2}+1\right)^4}}\,{{N-1}\over N}\,\phi_0^3.
\end{align}
Here we have made use of \cref{P2D}. Since ${\mathbf V}^*={{\boldsymbol \phi}}(0)$ is an equilibrium of the desired form with $\phi_0$ arbitrary, ${\mathcal W}$ does not assume a local minimum at ${\mathbf V}^*$. Hence we have found:
\begin{theorem}\label{V<1instab}
An equilibrium of the system \cref{sys} in ${\mathbb V}(\overline{V})$
is unstable if it contains a large droplet of height $1<h\leq \left(N-1\right)^{1/4}$.
\end{theorem}
Finally, let us turn to the case $n=0$. The corresponding equilibria are uniform of height $0<h_S\leq 1$. Hence we have $0<\overline{V}\leq 1$. \Cref{conalt2} immediately gives the sufficient condition for a strict local minimizer in the form
\begin{equation}
\delta_S>0.
\end{equation}
Clearly, this condition holds true for the uniform equilibrium when $0<\overline{V}<1$. This condition fails, however, for $\overline{V}=1$ since then $\delta_S=0$.
To obtain a stability result for $n=0$, $\overline{V}=1$, we proceed differently. First, if $N=2$, ${\mathbf V}^* = \left(1,1\right)$ is the only possible equilibrium of \cref{sys} by \cref{unieq,pVE1}. Hence the only choice for a global minimizer of ${\mathcal W}$ on ${\mathbb V}(\overline{V})$ is ${\mathbf V}^*$ which must be stable.
Next assume $N\geq 3$. Let ${\mathbf V}^*=(1,\ldots,1)$ and consider ${{\boldsymbol \phi}}={{\boldsymbol \phi}}(t)\equiv {\mathbf V}^*+t\,\left(\phi_1,\ldots,\phi_N\right)$ where $\phi_1$, $\phi_2\in \mathbb R$ are fixed, $\phi_3 = -\left(\phi_1+\phi_2\right)$ and $\phi_k=0$ for $3<k\leq N$. Since $\sum_{k=1}^N \phi_k=0$, ${{\boldsymbol \phi}}(t)$ is a curve with trace in ${\mathbb V}(1)$.
From \cref{A2D,A3D} we obtain
\begin{align}
&{{d^2}\over {dt^2}} {\mathcal W}\left({{\boldsymbol \phi}}(t) \right)\Big|_{t=0} = 0,\quad\text{and}\\
& {{d^3}\over {dt^3}} {\mathcal W}\left({{\boldsymbol \phi}}(t) \right)\Big|_{t=0} = {1\over N}\,P^{\prime\prime}(1)\,\sum_{k=1}^N \phi_k^3 = -{3\over N}\,P^{\prime\prime}(1)\,\phi_1\,\phi_2\,\left(\phi_1+\phi_2\right).\label{indef}
\end{align}
Since $P^{\prime\prime}(1)\not=0$ by \cref{P2D} and since the right-hand side of Equation \cref{indef} can take positive and negative values for appropriate choices of $\phi_1$ and $\phi_2$, ${\mathbf V}^*$ is not a local minimizer of ${\mathcal W}$ on ${\mathbb V}(1)$. For $\overline{V}>1$ the uniform equilibrium in ${\mathbb V}(\overline{V})$ is always unstable by \cref{stabeq}. Hence we can summarize:
\begin{theorem}\label{uniV=1}
The uniform equilibrium of the system \cref{sys} in ${\mathbb V}({\overline{V}})$ is stable if and only if either $0<\overline{V}<1$ and $N\geq 3$, or $0<\overline{V}\leq 1$ and $N=2$.
\end{theorem}
An immediate consequence of this result is that for $N=2$ and any $s>0$, the initial value problem \cref{sys,ini} with initial data $V_1(0) = 1=V_2(0)$ has the unique solution $V_1(t) = 1=V_2(t)$, $t\geq 0$. Similarly, if $0<\overline{V}<1$, $N\geq 2$ and $s>0$, the uniform equilibrium is the unique solution of \cref{sys,ini} with initial data $V_i(0) = \overline{V}$, $1\leq i\leq N$. These findings for the case $N=2$ can be confirmed independently by direct arguments.
We have now determined the stability of all equilibria in ${\mathbb V}(\overline{V})$ and, by \cref{VV+}, in ${\mathbb V}_+(\overline{V})$ for every $\overline{V}>0$. The stability results are new for the case $s\not=1$. They are also new for the case $0<\overline{V}\leq 1$ when $s=1$.
\section{Rest State Orderings: Energy Hierarchies of Equilibria}
\label{sec_hier}
In the following our objective is to determine an ordering of equilibria according to the size of the energy functional $\mathcal W$.
\begin{snugshade}
\noindent {\em Quantum-mechanical aside}: The equilibria in volume scavenging take on discrete values of the energy functional $\mathcal W$ (or total surface area $A$). We expect them to be ordered from lowest energy (or lowest total surface area) to largest energy (or largest total surface area) in analogy with the discrete energy levels of electrons in an atom. Hence unstable equilibria in volume scavenging are equivalent to excited states in a quantum-mechanical system, while a stable equilibrium minimizing the energy functional corresponds to the ground state.
\end{snugshade}
\begin{snugshade}
\noindent {\em Socio-economic view}: The rest states of the competition are naturally ordered in hierarchies of
more versus less costly (or energetic) outcomes. An outcome is more costly (energetic) if its total cost (energy) ${\mathcal W}$ is larger.
\end{snugshade}
Let us first motivate the problem by pursuing an approach based on the variable $\theta$, introduced in \cref{deftheta}.
If $\mathbf V$ is an equilibrium of \cref{sys} consisting of $n$ droplets of height $h>0$ and $N-n$ droplets of height $h^{-1}$, $1\leq n\leq \lfloor {N\over 2}\rfloor$, then by \cref{defW,intP} with $\alpha={n\over N}$, we obtain
$ {\mathcal W}({\mathbf V}) = {3\over 2}\,\left(\alpha\,h^2+(1-\alpha)\,{1\over {h^2}}\right)$.
Using again $a=h+{1\over h}$ and $b=h-{1\over h}$, we find
$ {\mathcal W}({\mathbf V})= {3\over 2}+{3\over 4}\,b^2+{3\over 2}\,\left(\alpha-{1\over 2}\right)\,a\,b$.
Consequently, by \cref{defab}, we have a function in $\theta$
\begin{equation}\label{weq}
w(\theta) \equiv {3\over 2}+{3\over 4}\,{\mathcal B}(\theta)^2+{3\over 2}\,\left(\alpha-{1\over 2}\right)\,{\mathcal A}(\theta)\,{\mathcal B}(\theta)
\end{equation}
such that ${\mathcal W}({\mathbf V})=w(\theta)$ with $h={1\over 2}\left({\mathcal A}(\theta)+{\mathcal B}(\theta)\right)$.
This {\em reduced} energy functional is displayed in \cref{wred}.
\begin{figure}[tbhp]\vspace*{-0.7cm}
\begin{center}
\includegraphics[width=0.45\textwidth]{wreduced.pdf}
\caption{The reduced energy functional $w$ for various values of $\alpha$}
\label{wred}
\end{center}
\end{figure}
Let us consider an example. We choose $\overline{V}=1.2$ and determine $\theta$ and $w(\theta)$ from Equations \cref{Aeq,weq} for $\alpha\in \left\{{1\over {16}},{1\over 8},{1\over 4}\right\}$. Noting that Equation \cref{Aeq} has both a positive $\left(\oplus\right)$ and negative $\left(\ominus\right)$ solution for $\theta$, we display the results in the following tables:\\
\begin{center}
\begin{minipage}{0.45\textwidth}
\begin{tabular}{c|c|c|c}
$\oplus$ & $\alpha={1\over {16}}$ & $\alpha={1\over {8}}$ & $\alpha={1\over {4}}$\\[1ex] \hline
$\theta$ & 15.936 & 7.374 & 3.199\\[1ex] \hline
$w(\theta)$ & 1.427 & 1.646& 1.814
\end{tabular}
\end{minipage}
\begin{minipage}{0.45\textwidth}
\begin{tabular}{c|c|c|c}
$\ominus$ & $\alpha={1\over {16}}$ & $\alpha={1\over {8}}$ & $\alpha={1\over {4}}$\\[1ex] \hline
$\theta$ & -0.398 & -0.447 & -0.582\\[1ex] \hline
$w(\theta)$ & 1.899 & 1.898& 1.897
\end{tabular}
\end{minipage}
\end{center}
\medskip
For $\theta>0$ we have exactly $n=\alpha\,N$ large droplets, while for $\theta<0$ there are exactly $n=(1-\alpha)\,N$ large droplets. Hence the example above suggests (choosing $N=16$ for instance) that for the same average volume $\overline{V}$, the larger the energy functional $\mathcal W$ (or $w$), the more large droplets are contained in an equilibrium.
\begin{snugshade}
\noindent {\em Physics aside}: While this claim appears intuitive and obvious, it is neither: For simplicity take $\overline{V}>1$ and compare two different equilibria. One have $n_1$ large droplets, the other one $n_2$. If $n_1<n_2$, each of the $n_1$ large droplets in the first equilibrium is expected to have larger volume (hence larger surface area) than each of the $n_2$ large droplets in the second equilibrium. Yet there are only $n_1$ large droplets in the first equilibrium, while there are $n_2>n_1$ in the second. A similar picture arises for the small droplets. As the table with negative ($\ominus$) solutions for $\theta$ illustrates, the actual differences in the energy functional (or surface area) can be tiny.
\end{snugshade}
We will prove below that the claim above is indeed correct if $\overline{V}>1$. For $0<\overline{V}<1$, the situation is, however, more subtle since for constant $\alpha$, Equation \cref{Aeq} generally permits two distinct positive solutions for $\theta$ and no negative solution as seen in \cref{ecurves}.
To obtain more precise information about the ordering of equilibria in terms of $\mathcal W$, we put the approach above aside and instead examine the zeros of the mass polynomial ${\mathcal P}_\alpha$, given by \cref{defpa}, in more detail. For $0<\overline{V}\leq 1$, we define $\alpha^*\left(\overline{V}\right)$ to be the unique value of $\alpha\in \left(0,{1\over 2}\right]$ such that $\overline{V}=L(\alpha)$.
Now consider the map $\alpha\mapsto h_L(\alpha)$, defined on $(0,1)$ for $\overline{V}>1$ by \cref{pVL1} and on $\left(0,\alpha^*(\overline{V})\right)$ for $0<\overline{V}\leq 1$ by \cref{pVE1,pVS1} with $h_L(\alpha)>\left({1\over \alpha}-1\right)^{1/4}$. In each case, $h_L(\alpha)$ is a simple root of the equation ${\mathcal P}_\alpha(h) = 0$. Hence it is elementary to conclude that $\alpha\mapsto h_L(\alpha)$ is a smooth map and ${d\over {d\alpha}} h_L(\alpha)\not= 0$. Since $h_L(\alpha)>\left({1\over \alpha}-1\right)^{1/4}$, we have $h_L(\alpha)\rightarrow\infty$ as $\alpha\rightarrow 0^+$. Consequently, $ {d\over {d\alpha}} h_L(\alpha)<0$. Since $h_L(\alpha)$ is also bounded from below as $\alpha$ approaches the right endpoint of its interval of definition, we obtain:
\begin{proposition}\label{h_Lprop}
Let\, $\alpha_0 = 1$ if $\overline{V}>1$, and $\alpha_0= \alpha^*(\overline{V})$ if\, $0<\overline{V}\leq 1$. Then the map $\alpha\mapsto
h_L(\alpha)$ is smooth and strictly decreasing on $\left(0,\alpha_0\right)$. Moreover, it extends continuously to $\left(0,\alpha_0\right]$ such that
\begin{enumerate}
\item[\mbox{(a)}] in case\, $\overline{V}>1$, $h_L(\alpha_0)>1$ is the unique positive root of $h^3+3\,h-4\,\overline{V} = 0$,
\item[\mbox{(b)}] in case\, $\overline{V}=1$, $h_L(\alpha_0) = 1$, and
\item[\mbox{(c)}] in case\, $0<\overline{V}<1$, $h_L(\alpha_0) = \left({1\over \alpha_0} -1\right)^{1/4}>1$.
\end{enumerate}
\end{proposition}
Next let us investigate the map $\alpha\mapsto h_l(\alpha)$, defined on $\left(0,\alpha^*(\overline{V})\right)$ for $0<\overline{V}<1$ by \cref{pVS1} with $1<h_l(\alpha)<\left({1\over \alpha}-1\right)^{1/4}$. Again, $h_l(\alpha)$ is a simple root of the equation ${\mathcal P}_\alpha(h)=0$, and thus $\alpha\mapsto h_l(\alpha)$ is a smooth map and ${d\over {d\alpha}} h_l(\alpha)\not= 0$. Now let $h_l(0)$ be the unique positive root of the equation ${\mathcal P}_0(h) = -4\,\overline{V}\,h^3+3\,h^2+1=0$. It follows readily that this definition extends $h_l$ to a smooth function on $\left[0,\alpha^*(\overline{V})\right)$. Since
$
{\mathcal P}_0(1) = -4\,\overline{V}+4>0$, $
{\mathcal P}_0\left({1\over {2\,\overline{V}}}\right) = {{1\over {4\,\overline{V}^2}}}+1$, and
$ \lim_{h\rightarrow \infty} {\mathcal P}_0(h) = -\infty$,
we obtain
\begin{equation}
h_l(0)>\max\left\{1, {1\over {2\,\overline{V}}}\right\}.
\end{equation}
This estimate together with the equation ${d\over {d\alpha}} \left({\mathcal P}_\alpha(h_l(\alpha))\right)=0$ implies that
\begin{equation}
{d\over {d\alpha}} h_l (0) = {{h_l(0)^6+3\,h_l(0)^4-3\,h_l(0)^2-1}\over {6\,h_l(0)\,\left(2\,\overline{V}\,h_l(0)-1\right)}}>0.
\end{equation}
Hence by continuity, ${d\over {d\alpha}} h_l(\alpha)>0$ for all $\alpha$.
Since $h_l$ is increasing and $h_l(\alpha)<\left({1\over \alpha}-1\right)^{1/4}$ on $\left(0,\alpha^*(\overline{V})\right)$, the limit $\lim_{\alpha\rightarrow \alpha^*(\overline{V})^-} h_l(\alpha)$ exists. In fact, we obtain:
\begin{proposition}\label{h_lprop}
Let\, $0<\overline{V}<1$. Then the map $\alpha\mapsto
h_l(\alpha)$ is smooth and strictly increasing on $\left(0,\alpha^*(\overline{V})\right)$. It extends continuously to $\left[0,\alpha^*(\overline{V})\right]$ such that
$h_l(0)>1$ is the unique positive root of $-4\,\overline{V}\,h^3+3\,h^2+1=0$ and
$h_l\left(\alpha^*(\overline{V})\right)=\left({1\over {\alpha^*(\overline{V})}}-1\right)^{1/4}$.
\end{proposition}
Now we define the function $\kappa=\kappa(h)$ for $h>1$ by
\begin{equation}\label{defkappa}
\kappa(h) \equiv{{4\,\overline{V}\,h^3-3\,h^2-1}\over {h^6+3\,h^4-3\,h^2-1}}.
\end{equation}
If $\overline{V}=1$, this definition reduces to
\begin{equation}\label{defkappa2}
\kappa(h) = {{4\,h^2+h+1}\over {h^5+h^4+h^3+4\,h^2+h+1}},
\end{equation}
which extends $\kappa$ smoothly to all $h\geq 1$ with $\kappa(1) = {2\over 3}$.
The function $\kappa$ is trivially related to the polynomial ${\mathcal P}_\alpha$, given by \cref{defpa}, in the following way:
\begin{proposition}\label{kappap}
Suppose $h_0>1$ is a root of the equation ${\mathcal P}_\alpha(h) = 0$ for some $0\leq \alpha\leq 1$. Then
$\kappa(h_0) = \alpha$.
\end{proposition}
Finally we introduce the functional ${\mathcal S} = {\mathcal S}(h)$ for $h>1$ by
\begin{equation}
{\mathcal S}(h)\equiv\kappa(h)\,\int_0^{v(h)} P(V)\,dV+\left(1-\kappa(h)\right)\,\int_0^{v(1/h)} P(V)\,dV.
\end{equation}
Note that by Equation \cref{intP},
\begin{equation}\label{Sbetter}
{\mathcal S}(h) = {3\over 2}\,\left(\kappa(h)\,h^2+\left(1-\kappa(h)\right)\,{1\over {h^2}}\right).
\end{equation}
\begin{snugshade}
\noindent {\em Applied mathematics aside}:
For integers $N\geq 2$ and $0\leq n\leq N$, let $h_0>1$ be a root of the equation ${\mathcal P}_\alpha(h) = 0$ with $\alpha= {n\over N}$.
Then ${\mathcal S}(h_0) = {\mathcal W}({\mathbf V}_0)$ where ${\mathbf V}_0$ is any equilibrium consisting of exactly $n=\alpha\,N$ large droplets of height $h_0$. Hence $\mathcal S$ interpolates the values of $\mathcal W$ along the equilibria.
\end{snugshade}
Using \cref{defkappa}, we find
\begin{equation}
{{d\mathcal S}\over {dh}}(h) = -6\,{{\left(h^2-1\right)\,\left(\overline{V}\,h^4-h^3-h+\overline{V}\right)}\over
{\left(h^4+4\,h^2+1\right)^2}}.
\end{equation}
\begin{theorem}\label{incr}
Let $\alpha_0 = 1$ if $\overline{V}>1$, and $\alpha_0= \alpha^*(\overline{V})$ if\, $0<\overline{V}\leq 1$. Then the maps
$\alpha\mapsto {\mathcal S}(h_L(\alpha))$ and $\alpha\mapsto {\mathcal S}(h_l(\alpha))$ are strictly increasing on $(0,\alpha_0)$.
\end{theorem}
\begin{proof}
Let $h=h_L$ or $h=h_l$ on $(0,\alpha_0)$. Since $h(\alpha)>1$ and $\kappa(h(\alpha)) = \alpha>0$ on $(0,\alpha_0)$, the definition \cref{defkappa} of $\kappa$ shows that $4\,\overline{V}\,h(\alpha)^3-3\,h(\alpha)^2-1>0$. Hence by \cref{h_Lprop,h_lprop},
\begin{equation}
\left(4\,\overline{V}\,h^3-3\,h^2-1\right)\,{d\over {d\alpha}} h<0 \ \text{ if $h=h_L$,\ and}\ \left(4\,\overline{V}\,h^3-3\,h^2-1\right)\,{d\over {d\alpha}} h>0\ \text{if $h=h_l$.}
\end{equation}
Consequently, the map $\alpha\mapsto {\mathcal F}(\alpha)\equiv \overline{V}\,h(\alpha)^4-h(\alpha)^3-h(\alpha)+\overline{V}$ is strictly decreasing on $(0,\alpha_0)$ if $h=h_L$, and strictly increasing on $(0,\alpha_0)$ if $h=h_l$. In the first case, we obtain from \cref{h_Lprop} that $\lim_{\alpha\rightarrow \alpha_0^+} {\mathcal F}(\alpha)$ exists and is non-negative. Thus ${\mathcal F}(\alpha)>0$ on $(0,\alpha_0)$. In the second case, \cref{h_lprop} proves that $\lim_{\alpha\rightarrow \alpha_0^+} {\mathcal F}(\alpha)$ exists and equals $0$. Hence ${\mathcal F}(\alpha)<0$ on $(0,\alpha_0)$. Finally, the claim follows since $h(\alpha)>1$ on $(0,\alpha_0)$ and
\begin{equation}
{d\over {d\alpha}}\left({\mathcal S}(h(\alpha))\right) = -6\,{{\left(h(\alpha)^2-1\right)\,{\mathcal F}(\alpha)}\over
{\left(h(\alpha)^4+4\,h(\alpha)^2+1\right)^2}}\,{d\over {d\alpha}} h(\alpha).
\end{equation}
\end{proof}
Now, for fixed $N$, we let $N^*(\overline{V})$ be the largest integer $K$ such that $K\leq \alpha^*(\overline{V})\,N$ if $0<\overline{V}< 1$. We also define $N^*(1)$ to be the largest integer $K$ such that $K< {1\over 2}\,N$. Specifically, we have
\begin{equation}
N^*(\overline{V}) =\left\{\begin{matrix}
\left\lceil{{1\over 2}\,N}-1\right\rceil& \text{for $\overline{V}= 1$,}\\
\left\lfloor{\alpha^*(\overline{V})\,N}\right\rfloor& \text{for $0<\overline{V}< 1$.}
\end{matrix}\right.
\end{equation}
Note that for $0<\overline{V}<1$, the condition $N^*(\overline{V}) \geq 1$ is equivalent to $L\left({1\over N}\right)\leq \overline{V}<1$.
Now we set:
\begin{enumerate}
\item[\mbox{(i)}] If $\overline{V}>1$, let $\beta_n \equiv {\mathcal W}({\mathbf V}_n)$, $1\leq n\leq N$, where ${\mathbf V}_n$ is an equilibrium of the system \cref{sys} with exactly $n$ large droplets.
\item[\mbox{(ii)}] If $\overline{V}=1$, let $\gamma_n \equiv {\mathcal W}({\mathbf V}_n)$, $0\leq n\leq N^*(1)$,
where ${\mathbf V}_n$ is an equilibrium of the system \cref{sys} with exactly $n$ large droplets.
\item[\mbox{(iii)}] If $0<\overline{V}<1$ and $N^*(\overline{V})\geq 1$, let $\lambda_n \equiv {\mathcal W}({\mathbf V}_n)$, $1\leq n\leq N^*(\overline{V})$, where ${\mathbf V}_n$ is an equilibrium of the system \cref{sys} with exactly $n$ large droplets of height $h\geq {\left({N\over n}-1\right)^{1/4}}$.
\item[\mbox{(iv)}] If $0<\overline{V}<1$, let $\sigma_n \equiv {\mathcal W}({\mathbf V}_n)$, $0\leq n\leq N^*(\overline{V})$, where ${\mathbf V}_n$ is an equilibrium of the system \cref{sys} with exactly $n$ large droplets of height $h\leq {\left({N\over n}-1\right)^{1/4}}$.
\end{enumerate}
Observe, in particular, that $\beta_N$, $\gamma_0$ and $\sigma_0$ denote the value of the functional ${\mathcal W}$ at the uniform equilibrium when $\overline{V}>1$, $\overline{V} = 1$, and $0<\overline{V}<1$, respectively.
\begin{theorem}\label{hierarchies}
\mbox{\ }
\begin{enumerate}
\item[\mbox{(a)}] If\ \ $\overline{V}>1$, then $\beta_n<\beta_{n+1}$, $1\leq n\leq N-1$.
\item[\mbox{(b)}] If\ \ $\overline{V}=1$, then $\gamma_n<\gamma_{n+1}$, $1\leq n \leq N^*(1)-1$, and $\gamma_n<\gamma_0$, $1\leq n\leq N^*(1)$.
\item[\mbox{(c)}] If\ \ $0<\overline{V}<1$, then $\lambda_n<\lambda_{n+1}$, $1\leq n\leq N^*(\overline{V})-1$, and $\sigma_n<\sigma_{n+1}$, $0\leq n\leq N^*(\overline{V})-1$.
\end{enumerate}
\end{theorem}
\begin{proof}
Let $\alpha_0$ be given as in \cref{incr} and let $h=h_L$ or $h=h_l$ on $(0,\alpha_0)$. Then with $V_L = v(h(\alpha))$ and $V_S = v(1/h(\alpha))$,
${\mathcal S}(h(\alpha)) = \alpha\,\int_0^{V_L} P(V)\,dV + (1-\alpha)\,\int_0^{V_S} P(V)\,dV$.
Hence setting $\alpha={n\over N}$, we obtain
\begin{equation}
{\mathcal S}\left(h\left({n\over N}\right)\right) = \left\{\begin{matrix} \beta_n& \text{if $\overline{V}>1$, $h=h_L$, and $1\leq n\leq N-1$,}\\[1ex]
\gamma_n& \text{if $\overline{V}=1$, $h=h_L$, and $1\leq n\leq N^*(1)$,}\\[1ex]
\lambda_n& \text{if $0<\overline{V}<1$, $h=h_L$, and $1\leq n\leq N^*(\overline{V})-1$},\\[1ex]
\sigma_n& \text{if $0<\overline{V}<1$, $h=h_l$, and $1\leq n\leq N^*(\overline{V})-1$.}\end{matrix}\right.
\end{equation}
If either $\overline{V}>1$ or $0<\overline{V}<1$, $h_L$ extends to a continuous function on $(0,\alpha_0]$ such that
$h_L(\alpha)>1$ for $0<\alpha\leq \alpha_0$. Consequently,
${\mathcal S}\left(h_L(1)\right) = \beta_N$ if $\overline{V}>1$, and
${\mathcal S}\left(h_L\left({{N^*(\overline{V})}/ N}\right)\right) = \lambda_{N^*(\overline{V})}$ if $0<\overline{V}<1$,
provided that $N^*(\overline{V})\geq 1$.
If $\overline{V} = 1$, $h_L$ extends to a continuous function on $\left(0,{1\over 2}\right]$ such that $h_L(\alpha)>1$ for $0<\alpha<{1\over 2}$ and $h_L\left({1\over 2}\right) = 1$. Consequently, by \cref{defkappa2,Sbetter}, the map $\alpha\mapsto {\mathcal S}\left(h_L(\alpha)\right)$ extends continuously to $\left(0,{1\over 2}\right]$ such that
\begin{equation}
{\mathcal S}\left(h_L\left({1\over 2}\right)\right) = {\mathcal S}(1) = {2\over 3}\,\int_0^1 P(V)\,dV+ \left(1-{2\over 3}\right)\,\int_0^1 P(V)\,dV = \int_0^1 P(V)\,dV = \gamma_0.
\end{equation}
Finally, if $0<\overline{V}<1$, $h_l$ extends to a continuous function on $[0,\alpha_0]$ such that
$h_l(\alpha)>1$ for $0\leq \alpha\leq \alpha_0$. Hence
${\mathcal S}\left(h_l(0)\right) = \sigma_0$ and
${\mathcal S}\left(h_l\left({{N^*(\overline{V})}/N}\right)\right) = \sigma_{N^*(\overline{V})}$.
The ordering of the quantities $\beta_n$, $\gamma_n$, $\lambda_n$ and $\sigma_n$ follows now immediately from the strict monotonicity of the maps $\alpha\mapsto {\mathcal S}(h_L(\alpha))$ and
$\alpha\mapsto {\mathcal S}(h_l(\alpha))$.
\end{proof}
In the case $\overline{V}<1$ with $N^*(\overline{V})\geq 1$, the system \cref{sys} potentially exhibits two types of stable equilibria: the uniform equilibrium with droplets of height $h_0=h_l(0)$ and equilibria consisting of exactly one large droplet of height $h_N = h_L\left({1\over N}\right)$. It is natural to ask for which stable equilibrium the functional $\mathcal W$ is minimal.
Instead of a comprehensive answer we can easily give a partial one: Since $\kappa(h)\sim h^{-3}$ as $h\rightarrow \infty$ and $h_N=h_L\left({1\over N}\right)\rightarrow \infty$ as $N\rightarrow\infty$, we obtain from \cref{Sbetter} that
$
{\mathcal S}\left(h_N\right) \rightarrow 0$ as $N\rightarrow\infty$, while ${\mathcal S}\left(h_0\right) = {3\over 2}\,{{h_0^{-2}}}>0$.
Hence for sufficiently large $N$ (with constant $0<\overline{V}<1$), the stable equilibria with one large droplet are the global minimizers of $\mathcal W$. If, on the other hand,
$
\overline{V} =L\left({1\over N}\right)
$,
then $\alpha^*(\overline{V}) = {1\over N}$ and, by \cref{pVS1}, $h_l\left({1\over N}\right) = h_L\left({1\over N}\right)$. By \cref{V<1instab}, the corresponding equilibria with exactly one large droplet are unstable. Hence they cannot be minimizers of $\mathcal W$. Indeed, since
$
\sigma_0<\sigma_1= \lambda_1
$ by
\cref{hierarchies},
$\mathcal W$ takes its global minimum value at the uniform equilibrium. Clearly, there exists $\epsilon>0$ (depending on $N$) such that this conclusion remains valid if
$
L\left({1\over N}\right)\leq \overline{V}<L\left({1\over N}\right)+\epsilon$.
\begin{snugshade}
\noindent {\em Engineering aside}: The possible occurrence of two types of stable equilibria for $\overline{V}<1$ is an interesting observation which lends support to the switching mechanism of the adhesion device presented by Vogel and Steen in \cite{VoSt}. The device can switch between two states: attached and detached. Both states are equilibria of the droplet competition. The attached state corresponds to the uniform equilibrium where each droplet forms a liquid bridge with the flat substrate. The detached state arises when one droplet is much larger than the other ones. Liquid bridges between the small droplets and the substrate will be broken. Our idealized fluid flow model is able to describe and explain this grab-and-release mechanism in qualitative terms.
\end{snugshade}
\begin{snugshade}
\noindent{\em Material science aside}: Since for $L\left({1\over N}\right)<\overline{V}<1$ $(N\geq 3$) volume scavenging is ``bistable'' (i.e.\,it exhibits two different types of stable equilibria), an initial droplet configuration can evolve towards the uniform equilibrium. In this case, volume scavenging does not lead to Ostwald ripening. The uniform equilibrium is still a local minimizer of the energy functional $\mathcal W$, and yet no coarsening occurs.
\end{snugshade}
\section{Concluding Remarks}
\label{sec_concl}
In the preceding sections we concentrated on the occurrence of equilibria, their stability and their ordering for fixed $\overline{V}>0$. It is now instructive to shed light on the equilibria and their stability when $\overline{V}$ varies. We make use of the equilibrium curves obtained via Equation \cref{Aeq} and identify equilibria as stable or unstable. We classify as follows:
\begin{itemize}
\item $L^N$, $S^N$: uniform equilibria, consisting of $N$ large ($L$) or $N$ small ($S$) droplets, respectively
\item $L^k\,S^m$: non-uniform equilibria, consisting of $k$ large ($L$) droplets of height $h> \left({N\over k}-1\right)^{1/4}$ and $m=N-k$ small ($S$) droplets
\item $l^k\,s^m$: non-uniform equilibria, consisting of $k$ large ($l$) droplets of height $h\leq \left({N\over k}-1\right)^{1/4}$ and $m=N-k$ small ($s$) droplets
\end{itemize}
\begin{figure}[tbhp]
\centering
\subfloat[$N=2$]{
\includegraphics[width=0.48\textwidth]{N2BF.pdf}
\label{N2bf}}
\subfloat[$N=3$]{
\includegraphics[width=0.48\textwidth]{N3BFBR.pdf}
\label{N3bf}}
\caption{Bifurcation diagrams for $N=2, 3$}
\label{Bfd23}
\end{figure}
\begin{figure}[tbhp]
\centering
\includegraphics[width=0.95\textwidth]{NGenBF2BR.pdf}
\caption{Generic bifurcation diagram for odd $N$}
\label{Bfd}
\end{figure}
\Cref{Bfd23} displays the situation $N=2, 3$. A generic bifurcation diagram (for odd $N$) is shown in \cref{Bfd}:
Equilibria of the type $L^k\,S^m$ and $l^k\,s^m$ are both present on equilibrium curves with turning point in the half-plane $\overline{V}<1$. Such equilibria lie immediately above and below the turning point, respectively. The turning point itself occurs for $\overline{V} = L\left({n\over N}\right) = {1\over N}\,\left(\left({N\over n}-1\right)^{1/4}+\left({N\over n}-1\right)^{3/4}\right)$. Up to a rescaling, this value of $\overline{V}$ was given by Slater and Steen \cite{SlSt} in the case $n=1$ and has been reconfirmed here. At this turning point a saddle-node bifurcation occurs with equilibria of the type $L^1\,S^{N-1}$ stable and all other equilibria on the same solution curve unstable. Our results on the limiting function $L$ furnish the location of all turning points with $n\in \left\{1,\ldots,\left\lfloor{N\over 2}\right\rfloor\right\}$. Equilibria of the type $l^k\,s^m$ extend all the way to the point $(\overline{V},\theta) = (1,0)$ and then switch to the type $L^m S^k$. No equilibria of the type $l^k\,s^m$ are present on the equilibrium curve with $\alpha={1\over 2}$, i.e.\,for $N$ even and $n={N\over 2}$. This curve consists of two solution branches symmetric about the axis $\theta=0$. In the case $N=2$ this situation corresponds to a supercritical pitchfork bifurcation as indicated in \cref{N2bf}. Our results are in agreement with Wente's work for $N=2$ and $3$ \cite{We}. The {\em bistability range} $L\left({1\over N}\right)<\overline{V}<1$, i.e.\,the interval of $\overline{V}$-values where two different types of stable equilibria can occur, is indicated in \cref{N3bf,Bfd}.
When $\overline{V}>1$, the basins of attraction of stable equilibria are sensitive to changes in the initial data (and model parameters) as seen in \Cref{sec_mot}. A similar behavior is expected when $\overline{V}<1$. We leave it to future work to study the basins of attraction, especially for
$\overline{V}$ in the bistability range.
\begin{snugshade}
\noindent {\em Socio-economic view}: The volume scavenging competition of $N$ individuals exhibits three distinct behaviors: For scarcest resource, $0<\overline{V}<L\left({1\over N}\right)$, the only possible outcome is egalitarian. In contrast, for abundant resource, $\overline{V}>1$, the competition exhibits the winner-take-all outcome. Finally, for scarce resource, $L\left({1\over N}\right)<\overline{V}<1$, both the winner-take-all and the all-share-evenly outcomes are possible.
As $N\rightarrow \infty$,
$L\left({1\over N}\right)\rightarrow 0$. Hence, as the number of competitors grows, the range of resource where
bistability occurs increases. At the same time, the range where only the egalitarian outcome is possible becomes smaller.
\end{snugshade}
\begin{figure}[tbhp]
\centering
\subfloat[$N=10$]{
\includegraphics[width=0.31\textwidth]{SE10.pdf}
\label{SocEco10}}
\subfloat[$N=30$]{
\includegraphics[width=0.31\textwidth]{SE30.pdf}
\label{SocEco30}}
\subfloat[$N=100$]{
\includegraphics[width=0.31\textwidth]{SE100.pdf}
\label{SocEco100}}
\caption{Small droplet volume $V_S$ for stable equilibria as a function of ${\overline V}>1$}
\label{EcoSoc}
\end{figure}
\begin{snugshade}
\noindent {\em Socio-economic view}:
The physics-based model we studied motivates many questions of socio-eco\-nom\-ic interest.
Among other things we might ask whether the ``macroeconomic objective'' of increasing average resource per individual
$\overline{V}$ actually results in a better outcome for the majority of individuals in a population of size $N$.
Since, for abundant resource $\overline{V}>1$, individuals experience the winner-take-all outcome, all but one competitor will earn an equal, small share
of total resource, corresponding to the volume $V_S=V_S\left(\overline{V}\right)$ of a
small droplet in a stable equilibrium with exactly one large droplet. Hence $V_S\left(\overline{V}\right)<1$ is the end state resource for every non-winning individual in the winner-take-all market with abundant resource $\overline{V}>1$. Graphs of this function are depicted in \cref{EcoSoc} for populations of size $N=10$, $N=30$ and $N=100$. Surprisingly, it appears that an increase in average resource per individual $\overline{V}$ results in a smaller share $V_S\left(\overline{V}\right)$. Hence almost all individuals fare better when resource $\overline{V}$ is less abundant.
\end{snugshade}
We have now given a complete characterization of all equilibria and their stability for volume scavenging of Newtonian and power-law fluids. While our stability and bifurcation results bear resemblance with the case of $N$ inviscid droplets with $S_N$ symmetry \cite{SlSt}, there are notable differences: The inviscid flow in \cite{SlSt} exhibits periodic and quasi-periodic solutions as well as chaotic behavior. In the viscous situation discussed here, the fluid rheology (for any $s>0$) renders stable equilibria asymptotically stable and restricts $\omega$-limit sets to a finite number of equilibria. This viscous behavior for $\overline{V}<1$ allows us to rationalize the mechanism for the adhesion device discussed in \cite{VoSt}. It is, however, remarkable that, in turn, our results on the hierarchical ordering of equilibria in terms of the pressure-volume work functional $\mathcal W$ (and hence total surface area) apply to the inviscid case as well. Our description of the turning points of the equilibrium curves as special values of the limiting function $L$ carry over directly. Most of these findings are new, both for the viscous and inviscid case.
\section*{Acknowledgement}
The authors are grateful to the Institute of Mathematics and Its Applications at the University of Minnesota for its support. This work has its roots in scientific activities offered at the IMA during the thematic year on ``Complex Fluids and Complex Flows'' in 2009/2010.
The authors also thank Profs.\,\,Robert Frank (Cornell University) and Philip Cook (Duke University) for relevant references in sociology and economics.
\bibliographystyle{s-plain}
|
\section{Introduction}
\label{s:intro}
Estimation of population sizes and population dynamics over time is an important task in ecology and epidemiology.
Census population sizes can be difficult to estimate due to infeasible sampling requirements or study costs.
Genetic sequences are a growing source of information that can be used to infer past population sizes from the signatures of genetic diversity.
Phylodynamics is a discipline that uses genetic sequence data to estimate past population dynamics.
Many phylodynamic models draw on coalescent theory \citep{kingman1982, griffiths1994}, which provides a probabilistic framework that connects the branching times of a genealogical tree with the effective population size and other demographic variables, such as migration rates, of the population from which the genealogy was drawn.
Effective population size can be interpreted as a measure of genetic diversity in a population and is proportional to census population size if coalescent model assumptions are met.
When genetic diversity is high, the effective population size approaches the census population size, given random mating and no inbreeding or genetic drift, but is otherwise smaller than the census size.
In our work we concentrate on estimation of effective population sizes over evolutionary time, which can be short for rapidly evolving virus populations and longer (but still estimable with preserved ancient molecular sequence samples) for more slowly-evolving organisms.
Some examples of successful application of phylodynamics include describing seasonal trends of influenza virus spread around the world \citep{rambaut2008}, quantifying dynamics of outbreaks like hepatitis C \citep{pybus2003} and Ebola viruses \citep{alizon2014}, and assessing the effects of climate change on populations of large mammals during the ice ages using ancient DNA \citep{shapiro2004, lorenzen2011}.
\par
Some approaches to phylodynamics use parametric functional relationships to describe effective population size trajectories \citep[\emph{e.g.,}][]{pybus2003, rasmussen2014}, but nonparametric methods offer a flexible alternative when an accurate estimate of a complex population size trajectory is needed and knowledge of the mechanisms driving population size changes is incomplete.
Nonparametric models have a long history of use in inferring effective population size trajectories.
\cite{pybus2000} introduced a nonparametric method, called the skyline plot, that produced point-wise estimates of population size, where the number of estimates was equal to the number of sampled genetic sequences minus one.
The estimates from this method were highly variable, so a modification, referred to as the generalized skyline plot, created a set of discrete time interval groups that shared a single effective population size \citep{strimmer2001}.
These likelihood-based approaches were adapted to a Bayesian framework with the Bayesian skyline plot \citep{drummond2005} and the variable-knot spline approach of \cite{opgen2005}.
\cite{minin2008sky} provided an alternative to these change-point methods by introducing a Gaussian Markov random field (GMRF) smoothing prior that connected the piecewise-constant population size estimates between coalescent events without needing to specify or estimate knot locations.
\cite{palacios2012} and \cite{gill2013} extended the GMRF approach of \cite{minin2008sky} by constructing a GMRF prior on a discrete uniform grid.
A grid-free approach, introduced by \cite{palacios2013}, allowed the population size trajectories to vary continuously by using a Gaussian process (GP) prior.
\par
Modern nonparametric Bayesian methods offer the state-of-the-art for recovering effective population size trajectories of unknown form.
However, current methods cannot accurately recover trajectories that exhibit challenging features such as abrupt changes or varying levels of smoothness.
Such features may arise in populations in the form of bottlenecks, rapid population changes, or aperiodic fluctuations with varying amplitudes.
Accurate estimation of features like these can be important for understanding the demographic history of a population.
Outside of phylodynamics, various nonparametric statistical methods have been developed to deal with such nonstationary or locally-varying behavior under more standard likelihoods.
These methods include, but are not limited to, GPs with nonstationary covariance functions \citep{paciorek2006}, nonstationary process convolutions \citep{higdon1998, fuentes2002}, non-Gaussian Mat\'ern fields \citep{wallin2015}, and adaptive smoothing splines \citep{yue2012, yue2014}.
Each of these methods has good qualities and could potentially be adapted for inferring effective population sizes, but methods based on continuous random fields or process convolutions can be computationally challenging for large data sets, and some spline methods require selection or modeling of the number and location of knots.
\par
A recent method by \cite{faulk2017} uses shrinkage priors in combination with Markov random fields to perform nonparametric smoothing with locally-adaptive properties.
This is a fully Bayesian method that does not require the use of knots and avoids the costly computations of inverting dense covariance matrices.
Computations instead take advantage of the sparsity in the precision matrix of the Markov random field to avoid matrix inversion.
\cite{faulk2017} compared different specifications of their shrinkage prior Markov random field (SPMRF) models and found that putting a horseshoe prior on the $k$th order differences between successive function values had superior performance when applied to underlying functions with sharp breaks or varying levels of smoothness.
We refer to the model with the horseshoe prior as a horseshoe Markov random field (HSMRF).
\par
In this paper, we propose an adaptation of the HSMRF approach of \cite{faulk2017} for use in phylodynamic inference with coalescent priors.
We devise a new MCMC scheme for the model that uses efficient, tuning-parameter-free, high-dimensional block updates.
We provide an implementation of this MCMC in the program \texttt{RevBayes}, which allows us to target the joint distribution of genealogy, evolutionary model parameters, and effective population size parameters.
We also develop a method for setting the hyperparameter on the prior for the global shrinkage parameter for coalescent data.
We use simulations to compare the performance of the HSMRF model to that of a GMRF model and show that our model has lower bias and higher precision across a set of population trajectories that are difficult to estimate.
We then apply our model to two real data examples that are well-known in the phylodynamics literature and compare its performance to other popular nonparametric methods.
The first example reanalyzes epidemiological dynamics of hepatitis C virus in Egypt and the second looks at estimation of ancient bison population size changes from DNA data.
\section{Methods}
\label{s:methods}
\subsection{Sequence Data and Substitution Model }
Suppose we have a set of $n$ aligned RNA or DNA sequences for a set of $L$ sites within a gene.
We assume the sequences come from a random sample of $n$ individuals from a well-mixed population, where samples were collected potentially at different times.
Let $\mathbf{Y}$ be the $n \times L$ sequence alignment matrix.
We assume the sites are fully linked with no recombination possible between the sequences.
This allows us to assume the existence of a genealogy $\boldsymbol{g}$, which is a rooted bifurcating tree that describes the ancestral relationships among the sampled individuals.
\par
We assume that $\mathbf{Y}$ is generated by a continuous time Markov chain (CTMC) substitution model that models the evolution of the discrete states (\emph{e.g.}, A,C,T,G for DNA) along the genealogy $\boldsymbol{g}$ for each alignment site.
A variety of substitution models are available and are typically differentiated by the form of the transition matrix $M(\boldsymbol{\Omega})$, which controls the substitution rates in the CTMC for the nucleotide bases with a set of parameters $\boldsymbol{\Omega}$ (see \cite{yang2014} for examples).
Let the likelihood of the sequence data given the genealogy and substitution parameters be denoted by $p(\mathbf{Y}\mid \boldsymbol{g}, \boldsymbol{\Omega}).$
\subsection{Coalescent}
\label{ss:methods:coal}
Suppose that we now have a genealogy $\boldsymbol{g}$, where branch lengths of the genealogical tree are measured in units of clock time (\emph{e.g.}, years).
To build a Bayesian hierarchical model, we need a prior density for $\boldsymbol{g}$.
The times at which two lineages merge into a common ancestor on the tree are called coalescent times.
The coalescent model provides a probabilistic framework for relating the coalescent times in the sample to the effective size of the population.
\cite{kingman1982} developed the coalescent model for a constant effective population size and \cite{griffiths1994} extended it for varying effective population sizes.
\par
Let the $n-1$ coalescent times arising from genealogy $\boldsymbol{g}$ be denoted by $0 < t_{n-1} < \dots < t_1$, where 0 is the present and time is measured backward from there.
We will assume the general case where sampling of the genetic sequences occurs at different times (\emph{heterochronous sampling}), which will include the special case where all sampling occurs at time 0 (\emph{isochronous sampling}).
We denote the set of unique sampling times as $s_m = 0 < s_{m-1} < \dots < s_1 < t_1$ for samples of size $n_m, \dots, n_1$, respectively, where $n = \sum_{j=1}^{m}n_j$ and we assume no sample times are equal to coalescent times (Figure \ref{coalFig}).
We let $\boldsymbol{s}$ denote the vector of sampling times.
Further, we let the intervals that end with a coalescent event be denoted $I_{0,k} = (\text{max}\{t_{k+1}, s_j\} , t_k ], \, \text{for } s_j < t_k \text{and } k=1,\dots, n-1,$ and let the intervals that end with a sampling event be denoted $I_{i,k} = (\text{max}\{t_{k+1}, s_{j+i}\} , s_{j+i-1} ], \, \text{for } s_{j+i-1} > t_{k+1}\text{ and } s_{j+i} < t_k,\, k=1,\dots, n-1$.
For $k=n-1$, we substitute $t_{k+1}=0$.
We let $n_{i,k}$ be the number of lineages present in interval $I_{i,k}$ and let the vector of number of lineages be denoted $\boldsymbol{n}$.
Further, we denote the number of unique sampling times in interval $(t_{k+1}, t_k]$ as $m_k$, where $m=1+\sum_{k=1}^{n-1}m_k $.
The joint density of the coalescent times given $\boldsymbol{s}$ and the effective population size trajectory $N_e(t)$ can then be written as
\begin{equation}
\begin{split}
p(t_1, \dots, t_{n-1} \mid \boldsymbol{s},\boldsymbol{n}, N_e(t)) &= \prod_{k=1}^{n-1} p(t_k \mid t_{k+1}, \boldsymbol{s},\boldsymbol{n}, N_e(t)) \\
&=\prod_{k=1}^{n-1} \frac{C_{0,k}}{N_e(t_k)} e^{- \sum_{i=0}^{m_k} \int_{I_{i,k}}\frac{C_{i,k}}{N_e(t)} dt},
\end{split}
\label{coalLik}
\end{equation}
where $C_{i,k} = {n_{i,k} \choose 2}$ is the coalescent factor \citep{felsen1999}.
This model can be seen as an inhomogeneous Markov point process where the conditional intensity is $C_{i,k} [N_e(t)]^{-1}$ \citep{palacios2013}.
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{Figure_1.pdf}
\end{center}
\caption{Effective population size trajectory and associated genealogical tree under heterochronous sampling. The top panel shows a continuous effective population size trajectory (gray) and an associated piecewise constant approximation to it. Also shown are the relationships between the genealogy and sampling times $s_i$, coalescent times $t_i$, intervals $I_{i,k}$, number of lineages $n_{i,k}$, and the uniform grid points, $x_h$, used for approximating coalescent densities. \label{coalFig} }
\end{figure}
\par
Here we assume $N_e(t)$ is an unknown continuous function, so the integrals in equation (\ref{coalLik}) must be computed with numerical approximation techniques.
We follow \cite{palacios2012}, \cite{gill2013}, and \cite{lan2015} and use discrete approximations of the integrals over a finite grid.
We construct a regular grid, $\boldsymbol{x}=\{x_h\}_{h=1}^{H+1}$, and set the end points of the grid $\boldsymbol{x}$ such that $x_1 = 0$ and $x_{H+1} = t_1$ (Figure \ref{coalFig}).
This results in $H$ grid cells and $H+1$ cell boundaries.
Now for $t\in (x_h, x_{h+1} ]$, we have $N_e(t)\approx \exp[\theta_h]$, where $\theta_h$ is an unknown model parameter.
This implies that $\boldsymbol{\theta} = \{\theta_h\}_{h=1}^{H}$ is a piecewise-constant approximation to $f(t) = \ln[N_e(t)]$ for $t \in [s_m, t_1]$.
The piecewise constant population size can be integrated analytically, leading to a discrete approximation to the likelihood in equation \ref{coalLik}.
The details of this approximation are provided in Appendix \ref{sectionApprox}.
\subsection{Prior for Effective Population Size Trajectory}
\label{ss:methods:prior}
Next we develop a prior for the unknown function $N_{e}(t)$ that describes the effective population size trajectory over time.
Let $\boldsymbol{\theta} = (\theta_1, \dots, \theta_H )$ be a vector of parameters that govern the effective population size trajectory $N_e(t)$.
We propose using a SPMRF model \citep{faulk2017} for $\boldsymbol{\theta}$, which is a type of Markov model where the $p$th-order differences in the forward-time evolution of the sequence $\left\{\theta_h \right\}_{h=1}^{H}$ are independent and follow a shrinkage prior distribution.
We define the $p$th-order forward difference as $\Delta^p \theta_l \equiv (-1)^{p}\sum_{j=0}^{p}(-1)^j {{p}\choose {j}}\theta_{l + j - p + 1}$, for $l = p,\dots,H-1$, which is a discrete approximation to the $p$th derivative of $f(t)$ evaluated at $t$.
If we assume a horseshoe distribution \citep{carvalho2010horseshoe} as our shrinkage prior on the order-$p$ differences in $\boldsymbol{\theta}$, then
\begin{equation}
\Delta^p\theta_l \mid \gamma \sim \mathcal{HS}(\gamma),
\label{hsEqn}
\end{equation}
where the location parameter of the horseshoe distribution is zero and $\gamma$ is the scale parameter and controls how much $f(t)$ is allowed to vary \textit{a priori}.
Following \cite{carvalho2010horseshoe}, we put a half-Cauchy prior on $\gamma$ with scale hyperparameter $\zeta$, so that $\gamma \sim \mathcal{C}^+(0, \zeta)$.
We chose the half-Cauchy here because it has desirable properties as a prior on a scale parameter \citep{gelman2006prior, polson2012half} and its single hyperparameter simplifies implementation.
Depending on the order $p$ of the model, we also place proper priors on $\theta_1,\dots,\theta_p$.
To do this, we start by setting $\theta_1 \sim \mathcal{N}(\mu, \sigma^2)$, where $\mu$ and $\sigma$ are hyperparameters typically set to create a diffuse prior.
Then for $p \geq 2$ and $q = 1, \dots, p-1$, we let $\Delta^q\theta_q\,\vert\,\gamma \sim \mathcal{HS}(a_q\gamma)$, where $a_q = 2^{-(p-q)/2}$, which follows from the recursive property and independence of the order-$p$ differences.
For example, for $p = 2$, $a_1 = 2^{-1/2}$, and for $p=3$, $ a_2 = 2^{-1/2} $ and $a_1=4^{-1/2} $.
We will refer to this specific model formulation as a state-space formulation of a HSMRF.
\par
The horseshoe distribution is leptokurtic with an infinite spike in density at zero and Cauchy-like tails.
In our setting, this combination results in small $\theta$ differences being shrunk toward zero and larger differences being maintained, which corresponds to smoothing over smaller noisy signals while retaining the ability to adapt to rapid functional changes.
This is in contrast to the normal distribution, which has higher density around medium-sized values and normal tails.
These attributes result in noisier estimates and reduced ability to capture abrupt functional changes.
Different shrinkage priors will result in different levels of shrinkage and therefore different smoothing behavior.
\cite{faulk2017} found that the horseshoe prior performed better than the Laplace prior in terms of bias and precision for nonparametric smoothing with SPMRFs, but we do not investigate the effect of different shrinkage priors here.
\par
The horseshoe density does not have a closed form (although see \cite{faulk2017} for an approximation in closed form).
However, a horseshoe distribution can be represented hierarchically as a scale mixture of normal distributions by introducing a latent scale parameter that follows a half-Cauchy distribution \citep{carvalho2010horseshoe}.
That is, if $\tau_l \sim \mathcal{C}^+(0, \gamma)$ and $\Delta^p\theta_l\mid\tau_l \sim \mathcal{N}(0, \tau_l^2)$, then integrating over $\tau_l$ results in the marginal relationship in equation \ref{hsEqn}.
\par
The hierarchical HSMRF models are a type of $p$th-order normal random walk with separate variance parameters for each increment.
The inherent Markov properties and properties of the normal distribution allow the joint distribution of $\boldsymbol{\theta}$ conditional on the vector of scale parameters $\boldsymbol{\tau}$ to be expressed $p(\boldsymbol{\theta}\mid \boldsymbol{\tau}, \mu, \sigma^2) = p(\theta_1 \mid \mu, \sigma^2)p(\Delta^1\theta_1, \dots,\Delta^p\theta_p,\Delta^p\theta_{p+1},\dots \Delta^p\theta_{H-1} \mid \boldsymbol{\tau} )$, which results in a multivariate normal distribution with mean $\boldsymbol{\mu}$ and precision matrix $\mathbf{Q}(\boldsymbol{\tau})$.
Specifically, $\boldsymbol{\theta}$ follows a Gaussian Markov random field \citep[GMRF;][]{rue2005gaussian} conditional on $\boldsymbol{\tau}$, where the order $p$ of the differencing in $\theta$ determines the structure of the sparse $\mathbf{Q}(\boldsymbol{\tau})$.
For the models presented here, $\boldsymbol{\mu} = \mu\boldsymbol{1}$, where $\mu$ is a constant and $\boldsymbol{1}$ is a vector of ones.
We specify $p(\boldsymbol{\tau})$ by assuming that the $\tau$'s are independent $\mathcal{C}^+(0, \gamma)$-distributed random variables, where $\tau_l \sim \mathcal{C}^+(0, \gamma)$ for $l=p,\dots, H-1$ and $\tau_l \sim \mathcal{C}^+(0, a_l\gamma)$ for $l=1,\dots,p-1$ and $p\geq2$.
The marginal joint distribution of $\boldsymbol{\theta}$ that results from integrating over $\boldsymbol{\tau}$ is a HSMRF.
Note that a GMRF model results when a single scale parameter $\tau$ is used for all order-$p$ differences in $\boldsymbol{\theta}$.
For our GMRF models, we use $\tau \sim \mathcal{C}^+(0, \zeta)$, where $\zeta$ is a fixed hyperparameter.
The order of the HSMRF will determine the amount of smoothing, with higher orders resulting in more smoothing.
We only consider first-order and second-order models here.
In practice, we use the state-space formulation described previously but with the independent hierarchical representations of the horseshoe distributions for the individual order-$p$ differences, which improves computational efficiency over the conditional multivariate normal representation.
\subsection{Posterior Inference}
\label{ss:methods:post}
For the case where we have a fixed genealogical tree, $\boldsymbol{g}$, which consists of sampling times $\boldsymbol{s}$ and coalescent times $\boldsymbol{t}$, the posterior distribution of the parameters $\left\{ \boldsymbol{\theta},\boldsymbol{\tau}, \gamma \right\} $ can be written as
\begin{equation}
p(\boldsymbol{\theta},\boldsymbol{\tau},\gamma \mid \boldsymbol{g} ) \propto p(\boldsymbol{g} \mid \boldsymbol{\theta}) p(\boldsymbol{\theta} \mid \boldsymbol{\tau})p(\boldsymbol{\tau}\mid \gamma) p(\gamma).
\label{fixedPost}
\end{equation}
Here $\boldsymbol{g}$ is considered data and we assume the coalescent times are known.
Then $p(\boldsymbol{g} \mid \boldsymbol{\theta})$ is the coalescent likelihood and $p(\boldsymbol{\theta} \mid \boldsymbol{\tau})p(\boldsymbol{\tau}\mid \gamma) p(\gamma)$ is the HSMRF prior described in Section \ref{ss:methods:prior}.
For our GMRF models, the righthand side of equation \ref{fixedPost} becomes $p(\boldsymbol{g} \mid \boldsymbol{\theta}) p(\boldsymbol{\theta} \mid \tau)p(\tau)$.
\par
For our analyses with fixed genealogical trees, we follow \cite{faulk2017} and \cite{lan2015} and use Hamiltonian Monte Carlo \citep[HMC;][]{neal2011mcmc} for posterior inference.
HMC performs joint proposals for the parameters that are typically far from the current parameter state and have high acceptance rates, resulting in efficient posterior sampling.
We used the \texttt{Stan} computing environment \citep{carpenter2016} for implementing HMC.
Specifically, we used the open source package \texttt{rstan} \citep{rstan-software:2017}, which provides a platform for fitting models using HMC in the \texttt{R} computing environment \citep{rmanual}.
Our \texttt{R} package titled \texttt{spmrf} allows for easy implementation of our models for use on fixed genealogical trees via a wrapper to the \texttt{rstan} tools.
A link to the package code is provided in the Supporting Information section.
We present a method for objectively setting the scale hyperparameter $\zeta$ of the prior distribution of the global smoothing parameter $\gamma$ in Appendix \ref{sectionGscale}.
\par
When there are genetic sequence data available and we want to jointly estimate evolutionary parameters, coalescent times, and population size trajectories, our posterior can be written as
\begin{equation}
p(\boldsymbol{g},\boldsymbol{\Omega},\boldsymbol{\theta},\boldsymbol{\tau},\gamma \mid \mathbf{Y} ) \propto p(\mathbf{Y}\mid \boldsymbol{g},\boldsymbol{\Omega})p(\boldsymbol{g} \mid \boldsymbol{\theta}) p(\boldsymbol{\Omega} )
p(\boldsymbol{\theta} \mid \boldsymbol{\tau})p(\boldsymbol{\tau}\mid \gamma) p(\gamma) ,
\end{equation}
where $\mathbf{Y}$ are the sequence data and $\boldsymbol{\Omega}$ are the parameters related to the DNA substitution model.
The likelihood of the sequence data given the parameters is $p(\mathbf{Y}\mid \boldsymbol{g},\boldsymbol{\Omega})$, and now $p(\boldsymbol{g} \mid \boldsymbol{\theta})$ is a prior for the genealogy given the population sizes and is proportional to $p(\boldsymbol{g} \mid \boldsymbol{\theta})$ in equation \ref{fixedPost}.
The remaining components are the prior for the evolution parameters $p(\boldsymbol{\Omega} )$ and the HSMRF prior as in equation \ref{fixedPost}.
\par
HMC requires the calculation of gradients over continuous parameter space and therefore cannot be used for inference on discrete parameters.
Therefore, we developed a custom MCMC algorithm that uses a combination of Gibbs sampling, elliptical slice sampling, and the Metropolis-Hastings (MH) algorithm to sample from the joint posterior of the evolution parameters and the effective population size parameters.
In particular, elliptical slice sampling \citep{murray2010} was used to sample from the joint field of log effective population sizes conditional on the latent scale parameters, a Gibbs sampler based on an approach developed by \citep{makalic2016} for horseshoe random variables was used to sample the latent scale parameters conditional on the field parameters, and standard phylogenetic MH steps were used to update the genealogy and substitution model parameters.
We implemented our custom MCMC in \texttt{RevBayes} --- a statistical computing environment geared primarily for phylogenetic inference \citep{hohna2016}.
The standard phylogenetic MH updates mentioned above were already implemented in \texttt{RevBayes}, so we contributed a heterochronous coalescent likelihood calculator, elliptical slice sampling, and Gibbs updates of our model parameters to the \texttt{RevBayes} source code.
The details of the sampling scheme are provided in the Appendix \ref{sectionESS} and a link to the code for implementing our methods for analyzing sequence data is provided in the Supporting Information section.
\section{Results}
\label{s:results}
\subsection{Simulated Data}
\label{ss:results:sim}
We used simulated data to assess the performance of the HSMRF model relative to the GMRF model.
We investigated four scenarios with different trajectories for $N_e(t)$: (1) Bottleneck (BN), (2) Boom-Bust (BB), (3) Broken Exponential (BE), and (4) Nonstationary Gaussian Process (NGP) realization.
The trajectory shapes are shown at the top of Figure \ref{simsumFig}.
For each scenario, we generated 100 data sets of coalescent times and fit GMRF and HSMRF models of first and second order using the fixed-tree approach.
The scenario descriptions and further methodological details of the simulations are provided in Appendix \ref{apx:sim}.
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{Figure_2.pdf}
\end{center}
\caption{Effective population size trajectories used in simulations and simulation results by model and scenario.
Models are GMRF of order 1 (G1) and order 2 (G2) and HSMRF of order 1 (H1) and order 2 (H2).
Top row shows true effective population size trajectories used to simulate coalescent data.
Remaining rows show mean absolute deviation (MAD), mean credible interval width (MCIW), mean absolute sequential variation (MASV), and credible interval Envelope.
Horizontal dashed lines in the third row plots indicate the true mean absolute sequential variation (TMASV) values.
Shown for each model are standard boxplots of the performance metrics (left) and mean values with 95\% frequentist confidence intervals (right).
Also shown for Envelope are the number of simulations with Envelope equal to 1.0.\label{simsumFig} }
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{Figure_3.pdf}
\end{center}
\caption{Example fits of first- and second-order Gaussian Markov random field (GMRF) and horseshoe Markov random field (HSMRF) models for four different simulation scenarios. Scenarios are a) Bottleneck (BN), b) Boom-Bust (BB), c) Broken Exponential (BE), and d) Nonstationary Gaussian Process (NGP). Results for all models within a particular scenario are for the same set of simulated data. Shown are the true effective population size trajectories that generated the data (dashed line), posterior medians of estimated trajectories (solid line) and associated 95\% Bayesian credible intervals (shaded band). \label{simfitsFig} }
\end{figure}
\par
We assessed the relative performance of the models using a set of summary statistics.
As a measure of bias, we used the mean absolute deviation (MAD) to compare the posterior medians of the trend parameters ($\hat{\theta}_i$) to the true trend values ($\theta_i$): $\text{MAD} = \frac{1}{H}\sum_{i=1}^{H} \vert \hat{\theta}_i - \theta_i \vert .$
\noindent We assessed the width of the 95\% Bayesian credible intervals (BCIs) using the mean credible interval width (MCIW):
$\text{MCIW} = \frac{1}{H}\sum_{i=1}^{H}\left( \hat{\theta}_{97.5, i} - \hat{\theta}_{2.5, i} \right)$, where $\hat{\theta}_{97.5, i}$ and $\hat{\theta}_{2.5, i}$ are the 97.5\% and 2.5\% quantiles of the posterior distribution for $\theta_i$.
We assessed the coverage of BCIs using $\text{Envelope}= \frac{1}{H}\sum_{i=1}^{H}I(\theta_i\in [\hat{\theta}_{97.5, i}, \hat{\theta}_{2.5, i}] )$, where $I(\cdot)$ is the indicator function.
To measure local variability in the estimated population trend, we used the mean absolute sequential variation (MASV) of $ \boldsymbol{\hat{\theta} } $, which was computed as $\text{MASV} = \frac{1}{H-1} \sum_{i=1}^{H-1} \vert \hat{\theta}_{i+1} - \hat{\theta}_i \vert .$
We compared the observed MASV to the true MASV (TMASV) in the underlying trend function, which is calculated by substituting true $\theta$'s into the equation for MASV.
For a measure of model complexity, we estimated the effective number of parameters $p_{eff}$ using an approach suggested by \cite{raftery2006}: $p_{eff} = \frac{2}{R-1}\sum_{r=1}^{R}(\mathcal{L}_r - \bar{\mathcal{L}})^2$,
where $\mathcal{L}_r$ is the log-likelihood evaluated at the parameter values for the $r$th of $R$ samples from the posterior, and $\bar{\mathcal{L}}$ is the mean value of $\mathcal{L}$ across the $R$ samples.
We used the Watanabe-Akaike information criterion \citep[WAIC;][]{watanabe2010} to calculate model weights and rank model performance.
The weight for model $m$ was calculated as $w_m = \exp\left(-0.5\Delta\text{W}_m\right)/\sum_{j=1}^{M}\exp\left(-0.5\Delta\text{W}_j\right)$ for a set of $M$ models, where $\Delta\text{W}_m = \text{WAIC}_m - \min\limits_{j \in M}\text{WAIC}_j$.
We utilized the \texttt{loo} package \citep{vehtari2017} to calculate WAIC.
For a measure of computational efficiency, we calculated the mean effective sample size (ESS) of the posterior samples across parameters for each model and simulated data set and used those with the total sampling times to calculate the mean ESS per second of sampling time.
\begin{table}
\caption{Summary of model selection criteria across 100 simulations by scenario and model set.
WAIC weights were calculated and the best model (greatest WAIC weight) was determined for each simulated data set within each scenario and model set. Metrics shown are the percentage of simulations each model was determined best and the mean model weight across simulations.
Values for each metric are compared among models within each scenario and model set.
Highest percentage of best models is in bold within each scenario and model set.
Scenarios are Bottleneck (BN), Boom-Bust (BB), Broken Exponential (BE), and Nonstationary Gaussian Process (NGP). Models are GMRF of order 1 (G1) and order 2 (G2) and HSMRF of order 1 (H1) and order 2 (H2). } \label{simTabWAIC}
\centering
\begin{tabular}{lllrrrr}
\toprule
\textbf{Metric} & \textbf{Model Set} & \textbf{Model} & \textbf{BN} & \textbf{BB} & \textbf{BE} & \textbf{NGP} \\
\midrule
Best Model (\%) & All Models & G1 & 1 & 9 & 13 & 1 \\
& & H1 & \textbf{93} & 14 & 34 & 9 \\
& & G2 & 0 & 3 & 1 & 24 \\
& & H2 & 6 & \textbf{74} & \textbf{52} & \textbf{66} \\
\cmidrule{2-7}
& Order 1 & G1 & 1 & \textbf{51} & 29 & 50 \\
& & H1 & \textbf{99} & 49 & \textbf{71} & 50 \\
\cmidrule{2-7}
& Order 2 & G2 & 9 & 9 & 5 & 27 \\
& & H2 & \textbf{91} & \textbf{91} & \textbf{95} & \textbf{73} \\
\midrule
Mean Weight & All Models & G1 & 0.03 & 0.11 & 0.14 & 0.04 \\
& & H1 & 0.89 & 0.15 & 0.35 & 0.09 \\
& & G2 & 0.01 & 0.10 & 0.07 & 0.26 \\
& & H2 & 0.08 & 0.63 & 0.44 & 0.61 \\
\cmidrule{2-7}
& Order 1 & G1 & 0.03 & 0.48 & 0.24 & 0.46 \\
& & H1 & 0.97 & 0.52 & 0.76 & 0.54 \\
\cmidrule{2-7}
& Order 2 & G2 & 0.12 & 0.11 & 0.11 & 0.43 \\
& & H2 & 0.88 & 0.89 & 0.89 & 0.57 \\
\bottomrule
\end{tabular}
\end{table}
\par For the BN scenario, the HSMRF model clearly had better performance than the GMRF model for the main performance metrics for both model orders (Figure \ref{simsumFig}, Table \ref{simTabWAIC}, and Table \ref{simTab} in Appendix \ref{apx:sim}).
Example model fits from each scenario provide some intuition for the simulation results (Figure \ref{simfitsFig}).
First order models did better than second order models within model types for the BN scenario.
Differences between model types were not as strong for the other scenarios.
The second-order HSMRF performed the best in terms of MAD, MCIW, and WAIC for the remaining scenarios.
Among second-order models, the HSMRF was clearly favored over the GMRF in terms of WAIC across all scenarios.
However, the HSMRF models were not noticeably different from the second-order GMRF in terms of MASV for the BB and BE scenarios.
The second-order GMRF had mean MASV closer to TMASV than did the second-order HSMRF for the NGP scenario.
Although the GMRF tended to estimate excess variation in the middle section of the trend for the NGP scenario, it did capture the peaks and troughs a little better than the HSMRF in other parts of the trend (see Figure \ref{simfitsFig} for an example).
In all scenarios, the HSMRF had lower $p_{eff}$ compared to the GMRF of the same order.
The GMRF was consistently more computationally efficient than the HSMRF, with mean ESS/sec approximately 1.5 to 6 times higher for models of the same order.
These differences are due to the additional parameters in the HSMRF models.
The second-order models were relatively slow for both model types, but the HSMRF was always slower.
As we show in the following data examples, however, the differences in computational speed between the HSMRF and GMRF models is negligible when genealogies and effective population size trajectories are jointly estimated.
\subsection{Egyptian Hepatitis C Virus}
\label{ss:results:HCV}
The hepatitis C virus (HCV) is a blood-borne RNA virus that exclusively infects humans.
HCV infection is often asymptomatic, but can lead to liver disease and liver failure.
HCV infections have historically had high prevalence in Egypt \citep{miller2010hcv}.
This is thought to be due to past widespread use of unsanitary medical practices in the region.
Of particular interest is a treatment for the parasite disease schistosomiasis known as parenteral antischistosomal therapy (PAT), which uses intravenous injections.
PAT was practiced from the 1920's to 1980's in Egypt and is thought to have contributed to the spread of HCV during that period due to unsterilized injection equipment \citep{frank2000}.
\par
We analyze 63 RNA sequences of type 4 with 411 base pairs from the E1 region of the HCV genome that were collected in 1993 in Egypt \citep{ray2000}.
\cite{pybus2003} used a piecewise demographic model for effective population size with a period of exponential growth between two periods of constant population size and concluded that the HCV population grew exponentially during the period of PAT treatment.
Other authors have applied nonparametric methods to estimate the effective population size trajectory for these data \citep[\emph{e.g.},][]{drummond2005, minin2008sky, palacios2013}.
Different nonparametric methods lead to different estimated trajectories and different levels of uncertainty.
We are interested in estimating the rapid change of HCV effective population size during the epidemic.
\par
We fit six different nonparametric models to these data: 1) Bayesian Skyline --- a piecewise constant/linear model with estimable locations of change-points \citep[SkyLine;][]{drummond2005}, 2) Bayesian Skyride \citep[SkyRide;][]{minin2008sky} 3) GMRF-1 (similar to Bayesian Skygrid, \cite{gill2013}), 4) GMRF-2, 5) HSMRF-1, and 6) HSMRF-2.
We note that the SkyRide model is also a type of GMRF model where the non-uniform grid cell boundaries are determined by coalescent events.
For all six models we jointly estimated the evolutionary model parameters, genealogies, and effective population size parameters.
We used the program \texttt{BEAST} implementation of the SkyLine and SkyRide models \citep{drummond2012}, and used our own \texttt{RevBayes} implementation of the GMRF and HSMRF models.
Although the Skygrid implementation of the GMRF-1 model is available in \texttt{BEAST}, the GMRF-2 and the HSMRF models are not, so we decided to use common software for the GMRF and HSMRF models.
For the GMRF and HSMRF models we used 100 equally-spaced grid cells where the first 99 ended at a fixed boundary of 227 years before 1993, and the final cell captured any coalescent events beyond the boundary (see Appendix \ref{sectionGrids} for discussion on setting grids).
The SkyLine model requires specification of the number of discrete population intervals, where each interval describes a piecewise constant population size between two coalescent events.
We used 20 population intervals to allow fair flexibility to capture sharp features in the population trajectory.
Further details about the MCMC implementation and computation times are provided in Appendix \ref{sectionImp}.
For model comparison, we calculated posterior model probabilities using marginal likelihood estimates calculated with steppingstone sampling \citep{xie2010}.
See Appendix \ref{sectionPostProb} for details on calculation of posterior model probabilities.
\begin{figure}[!t]
\begin{center}
\includegraphics[width=\textwidth]{Figure_4.pdf}
\end{center}
\caption{Posterior medians (solid black lines) of effective population sizes and associated 95\% credible intervals (grey shaded areas) for the HCV data for the Bayesian Skyline (SkyLine), Bayesian Skyride (SkyRide), Gaussian Markov random field of order 1 (GMRF-1) and order 2 (GMRF-2), and horseshoe Markov random field of order 1 (HSMRF-1) and order 2 (HSMRF-2).
Also shown for each model are posterior model probabilities ($\text{Pr}(\text{M}\mid\text{D})$) and heat maps of mean posterior frequencies of coalescent times.
A vertical reference line is shown at year 1918, which is the year PAT was introduced. \label{hcvFig} }
\end{figure}
\par
While the broad pattern of the demographic trajectory was similar among the six models, they differed in the estimated rate of change in effective population size and in the uncertainty around the effective population size estimates (Figure \ref{hcvFig}).
The SkyLine and HSMRF-1 models had the highest posterior model probabilities, with the SkyLine favored a little over the HSMRF-1 (Figure \ref{hcvFig}).
The shape of the median trajectory from the HSMRF-1 model was similar to that of the SkyLine model, yet the HSMRF-1 model showed a very rapid increase in population between 1925 and 1945, while the SkyLine and other models showed more gradual increases that started earlier and ended later.
The increase estimated by the SkyRide model lasted the longest, starting near 1900 and ending near 1970.
The HSMRF and the SkyLine also showed relatively constant population size following the increase in the mid 20th century, while the SkyRide and GMRF-1 models showed a decrease after 1970.
\par
In addition to differing in the rate of population growth after the epidemic began, the models differed in their estimates of when the epidemic began.
The posterior mean densities of frequencies of coalescent times provide an indication of when the HCV epidemic started (Figure \ref{hcvFig}).
The results of the HSMRF-1 support the idea that HCV epidemic started after PAT was introduced and suggest that early PAT campaigns may have used less sanitary practices and contributed more to the spread of HCV than the major PAT campaigns started in the 1950's.
Plots of the effective population trajectories covering the entire span of the coalescent times are provided with further discussion in Appendix \ref{sectionHCVapx}.
\subsection{Beringian Steppe Bison}
\label{ss:results:bison}
Modern molecular methods have allowed the recovery of DNA samples from specimens that lived hundreds to hundreds of thousands of years ago \citep{paabo2004, shapiro2014}.
Large mammals that lived in the Northern Hemisphere during the Pleistocene and Holocene epochs have been a valuable source of this ancient DNA due to conditions favorable for specimen preservation in the northern latitudes \citep[\emph{e.g.,}][]{shapiro2004, lorenzen2011}.
We focus on bison (\emph{Bison} spp.) that lived on the steppe-tundra of Northern Asia and Europe and crossed into North America over the Bering land bridge during the middle to late Pleistocene \citep{froese2017}.
Interest has been in determining whether human impact or climate and related habitat change instigated the decline of bison across their range during the late Pleistocene.
\cite{shapiro2004} used a parametric piecewise-exponential model for the bison effective population size and estimated that the time of transition from population growth to decline was 37 thousand years ago (kya).
\cite{drummond2005} used the more flexible SkyLine model, which indicated a more rounded and prolonged peak in population size followed by a rapid decline and bottleneck around 10 kya.
Here we use a modified version of the bison data described by \cite{shapiro2004} and fit coalescent models directly to the sequence data as with the HCV data.
We make qualitative comparisons among the resulting estimated population trajectories and in relation to some benchmark times describing the arrival of humans and the period of the Last Glacial Maximum (LGM).
\par
We analyze 152 sequences (135 ancient and 17 modern) of mitrochondrial DNA with 602 base pairs from the mitochondrial control region.
DNA was extracted from bison fossils from Alaska (68), Canada (46), Siberia (13), the lower 48 United States (6), and China (2).
Sample dates were estimated for the ancient samples using radiocarbon dating, with dates ranging up to 59k years.
We treat the calibrated radiocarbon dates as known in the following analyses.
These data are the same as those used by \cite{gill2013}, and are slightly modified from the data first described by \cite{shapiro2004} to remove sequences identified as potentially contaminated with young radiocarbon \citep{shapiro2010} and include additional sequences generated since generation of the initial data set.
In this data set, radiocarbon dates are calibrated to calendar time using the IntCal09 calibration curve \citep{reimer2009}.
\begin{figure}[!t]
\begin{center}
\includegraphics[width=\textwidth]{Figure_5.pdf}
\end{center}
\caption{Posterior medians of effective population sizes and associated 95\% credible intervals obtained from the bison DNA sequence data using the Bayesian Skyline (SkyLine), Bayesian Skyride (SkyRide), and GMRF and HSMRF models of order 1 and order 2.
Also shown for each model are posterior model probabilities ($\text{Pr}(\text{M}\mid\text{D})$) and posterior median and 95\% credible intervals for the time of peak effective population size.
The period of the Last Glacial Maximum and timing of first human settlement in North America are shown for reference.
\label{bisonFigB} }
\end{figure}
\par
The LGM in the Northern Hemisphere is estimated to have occurred between 26.5 to 19 kya \citep{clark2009}.
A small, isolated population of humans existed in central Beringia, including, potentially, the land bridge that connected the continents during the LGM \citep{llamas2016}.
Humans may have ventured into eastern Beringia (Alaska and Yukon) as early as 26 kya \citep{bourgeon2017}, but there is as yet no evidence of continuous occupation until 14 kya \citep{easton2011, holmes2011}.
Humans probably first entered continental North America via a western coastal route that became available close to 16 kya \citep{llamas2016, heintzman2016}, where they would have encountered the population of steppe bison that were isolated in the south with the coalescence of the Laurentide and Cordilleran glaciers \citep{shapiro2004, heintzman2016}.
Because the majority of our bison samples were collected in North America, we used 16-14 kya as the time of first human occupation.
\par
We used methods similar to those used in the HCV example.
We also calculated posterior distributions for the time of the peak in population size.
Method details can be found in Appendices \ref{sectionImp} and \ref{sectionPostProb}.
\par
While the broad pattern of an increase followed by a decrease in effective population size was recovered by all six models, the timing and nature of the population size change differed considerably between them (Figure \ref{bisonFigB}).
The HSMRF-1 model had the highest posterior model probability among the six models.
The posterior median trajectory from the HSMRF-1 model was most similar to the SkyLine model, but the credible intervals for the HSMRF-1 model were most similar to the GMRF-1 model.
The second-order models both produced strongly piecewise-linear trajectories with relatively narrow credible intervals, but had low posterior probability and smoothed over some of the local features displayed by other models.
The HSMRF-1 model displayed a more complex descent from the peak size to the present in comparison to the other models, and the areas of rapid descent are coincident with the arrival of humans in eastern Beringia and ice-free North America and the initial retreat of the glaciers, both of which are coincident with changes in habitat.
All models suggested that the overall decline in population size started before the LGM, and all had median time of population peak between 41.6 and 47.3 kya, but uncertainty in the time of peak population size varied widely across the models.
\section{Discussion}
We introduced a novel and fully Bayesian method for nonparametric inference of changes in effective population size that we call the HSMRF.
This method utilizes a shrinkage prior known as the horseshoe distribution, which allows more flexibility to respond to rapid changes in effective population size trajectories, yet also generates smoother trajectories in comparison to standard GMRF methods.
Our simulations demonstrated that the HSMRF had lower bias and higher precision than the GMRF and was able to recover the underlying true trajectories better in most cases.
\par
There are many situations where the local adaptivity of the HSMRF models would provide advantages over the GMRF and other models.
In infectious disease dynamics, examples that could lead to rapid changes in effective population sizes include sudden changes in contact rates due to behavioral changes or quarantine, or sudden changes in the infection rate due to introduction of treatment or vaccine.
At a macro-evolutionary scale, sudden changes in effective population size could be brought on by sudden population collapse (e.g., extinction) or rapid expansions due to dispersals or ecological release.
As we have demonstrated, in situations like these the GMRF and other models tend to smooth over the sharp changes that the HSMRF can capture.
\par
Our results from both data examples indicated that the properties of the population size trajectories estimated by the HSMRF-1 model were somewhere between those from the GMRF-1 model and the SkyLine model.
The SkyLine model is a type of change-point model, which suggests the HSMRF-1 can produce behavior of change-point models without explicitly needing to specify number or location of change points.
\par
We demonstrated in our simulations that second-order models for either the HSMRF or GMRF formulations can perform better than first-order models in many cases.
Although the second-order models did not perform as well as the first-order models in our particular data examples, they would likely do well in other examples with smoother trajectories.
Among the second-order models, the HSMRF did as well or better than the GMRF for the simulated examples and had higher posterior model probabilities for both of the data examples.
\par
Second-order models have not been used much for estimating effective population sizes previously.
\cite{palacios2013}, whose method assumes a fixed and known genealogy, tested an integrated Brownian motion (IBM) prior for their GP model for the purpose of testing prior sensitivity but did not use the prior beyond that.
The IBM prior is equivalent to the second-order GMRF in continuous time.
Our use of second-order GMRF model for jointly estimating genealogy and effective population size trajectory is the first we are aware of in the literature.
The second-order GMRF and HSMRF can have similar performance in many cases, but HSMRF has the advantage of added flexibility when needed, so it is a reasonable default choice over the GMRF.
We suggest that researchers fit both orders and use a metric such as Bayes factors to select the best order of model for the data.
\section*{Acknowledgments}
J.R.F. and V.N.M. were supported by the NIH grant U54 GM111274.
V.N.M. was supported by the NIH grant R01 AI107034.
A.F.M. was supported by ARCS Foundation Fellowship.
B.S. was supported by the NSF grant DEB-1754461.
We thank the Associate Editor and Reviewers for their constructive criticism and helpful suggestions.
\section*{Supporting Information}
Our \texttt{R} package titled \texttt{spmrf} can be used to fit our models to fixed genealogical trees and is available at {\tt https://github.com/jrfaulkner/spmrf}.
The data and \texttt{RevBayes} code for fitting our models to the molecular sequence data described in Sections \ref{ss:results:HCV} and \ref{ss:results:bison} is available at {\tt https://github.com/jrfaulkner/phylocode}.
\bibliographystyle{biom}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.